aboutsummaryrefslogtreecommitdiffstats
path: root/fs/ceph
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2015-07-02 14:35:00 -0400
committerLinus Torvalds <torvalds@linux-foundation.org>2015-07-02 14:35:00 -0400
commit0c76c6ba246043bbc5c0f9620a0645ae78217421 (patch)
tree644a4db58706c4e97478951f0a3a0087ddf26e5e /fs/ceph
parent8688d9540cc6e17df4cba71615e27f04e0378fe6 (diff)
parent5a60e87603c4c533492c515b7f62578189b03c9c (diff)
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client
Pull Ceph updates from Sage Weil: "We have a pile of bug fixes from Ilya, including a few patches that sync up the CRUSH code with the latest from userspace. There is also a long series from Zheng that fixes various issues with snapshots, inline data, and directory fsync, some simplification and improvement in the cap release code, and a rework of the caching of directory contents. To top it off there are a few small fixes and cleanups from Benoit and Hong" * 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client: (40 commits) rbd: use GFP_NOIO in rbd_obj_request_create() crush: fix a bug in tree bucket decode libceph: Fix ceph_tcp_sendpage()'s more boolean usage libceph: Remove spurious kunmap() of the zero page rbd: queue_depth map option rbd: store rbd_options in rbd_device rbd: terminate rbd_opts_tokens with Opt_err ceph: fix ceph_writepages_start() rbd: bump queue_max_segments ceph: rework dcache readdir crush: sync up with userspace crush: fix crash from invalid 'take' argument ceph: switch some GFP_NOFS memory allocation to GFP_KERNEL ceph: pre-allocate data structure that tracks caps flushing ceph: re-send flushing caps (which are revoked) in reconnect stage ceph: send TID of the oldest pending caps flush to MDS ceph: track pending caps flushing globally ceph: track pending caps flushing accurately libceph: fix wrong name "Ceph filesystem for Linux" ceph: fix directory fsync ...
Diffstat (limited to 'fs/ceph')
-rw-r--r--fs/ceph/acl.c4
-rw-r--r--fs/ceph/addr.c308
-rw-r--r--fs/ceph/caps.c836
-rw-r--r--fs/ceph/dir.c383
-rw-r--r--fs/ceph/file.c61
-rw-r--r--fs/ceph/inode.c155
-rw-r--r--fs/ceph/mds_client.c425
-rw-r--r--fs/ceph/mds_client.h23
-rw-r--r--fs/ceph/snap.c173
-rw-r--r--fs/ceph/super.c25
-rw-r--r--fs/ceph/super.h125
-rw-r--r--fs/ceph/xattr.c65
12 files changed, 1689 insertions, 894 deletions
diff --git a/fs/ceph/acl.c b/fs/ceph/acl.c
index 64fa248343f6..8f84646f10e9 100644
--- a/fs/ceph/acl.c
+++ b/fs/ceph/acl.c
@@ -187,10 +187,10 @@ int ceph_pre_init_acls(struct inode *dir, umode_t *mode,
187 val_size2 = posix_acl_xattr_size(default_acl->a_count); 187 val_size2 = posix_acl_xattr_size(default_acl->a_count);
188 188
189 err = -ENOMEM; 189 err = -ENOMEM;
190 tmp_buf = kmalloc(max(val_size1, val_size2), GFP_NOFS); 190 tmp_buf = kmalloc(max(val_size1, val_size2), GFP_KERNEL);
191 if (!tmp_buf) 191 if (!tmp_buf)
192 goto out_err; 192 goto out_err;
193 pagelist = kmalloc(sizeof(struct ceph_pagelist), GFP_NOFS); 193 pagelist = kmalloc(sizeof(struct ceph_pagelist), GFP_KERNEL);
194 if (!pagelist) 194 if (!pagelist)
195 goto out_err; 195 goto out_err;
196 ceph_pagelist_init(pagelist); 196 ceph_pagelist_init(pagelist);
diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
index e162bcd105ee..890c50971a69 100644
--- a/fs/ceph/addr.c
+++ b/fs/ceph/addr.c
@@ -87,17 +87,21 @@ static int ceph_set_page_dirty(struct page *page)
87 inode = mapping->host; 87 inode = mapping->host;
88 ci = ceph_inode(inode); 88 ci = ceph_inode(inode);
89 89
90 /*
91 * Note that we're grabbing a snapc ref here without holding
92 * any locks!
93 */
94 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
95
96 /* dirty the head */ 90 /* dirty the head */
97 spin_lock(&ci->i_ceph_lock); 91 spin_lock(&ci->i_ceph_lock);
98 if (ci->i_head_snapc == NULL) 92 BUG_ON(ci->i_wr_ref == 0); // caller should hold Fw reference
99 ci->i_head_snapc = ceph_get_snap_context(snapc); 93 if (__ceph_have_pending_cap_snap(ci)) {
100 ++ci->i_wrbuffer_ref_head; 94 struct ceph_cap_snap *capsnap =
95 list_last_entry(&ci->i_cap_snaps,
96 struct ceph_cap_snap,
97 ci_item);
98 snapc = ceph_get_snap_context(capsnap->context);
99 capsnap->dirty_pages++;
100 } else {
101 BUG_ON(!ci->i_head_snapc);
102 snapc = ceph_get_snap_context(ci->i_head_snapc);
103 ++ci->i_wrbuffer_ref_head;
104 }
101 if (ci->i_wrbuffer_ref == 0) 105 if (ci->i_wrbuffer_ref == 0)
102 ihold(inode); 106 ihold(inode);
103 ++ci->i_wrbuffer_ref; 107 ++ci->i_wrbuffer_ref;
@@ -346,7 +350,7 @@ static int start_read(struct inode *inode, struct list_head *page_list, int max)
346 350
347 /* build page vector */ 351 /* build page vector */
348 nr_pages = calc_pages_for(0, len); 352 nr_pages = calc_pages_for(0, len);
349 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS); 353 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_KERNEL);
350 ret = -ENOMEM; 354 ret = -ENOMEM;
351 if (!pages) 355 if (!pages)
352 goto out; 356 goto out;
@@ -358,7 +362,7 @@ static int start_read(struct inode *inode, struct list_head *page_list, int max)
358 dout("start_read %p adding %p idx %lu\n", inode, page, 362 dout("start_read %p adding %p idx %lu\n", inode, page,
359 page->index); 363 page->index);
360 if (add_to_page_cache_lru(page, &inode->i_data, page->index, 364 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
361 GFP_NOFS)) { 365 GFP_KERNEL)) {
362 ceph_fscache_uncache_page(inode, page); 366 ceph_fscache_uncache_page(inode, page);
363 page_cache_release(page); 367 page_cache_release(page);
364 dout("start_read %p add_to_page_cache failed %p\n", 368 dout("start_read %p add_to_page_cache failed %p\n",
@@ -436,7 +440,7 @@ out:
436 * only snap context we are allowed to write back. 440 * only snap context we are allowed to write back.
437 */ 441 */
438static struct ceph_snap_context *get_oldest_context(struct inode *inode, 442static struct ceph_snap_context *get_oldest_context(struct inode *inode,
439 u64 *snap_size) 443 loff_t *snap_size)
440{ 444{
441 struct ceph_inode_info *ci = ceph_inode(inode); 445 struct ceph_inode_info *ci = ceph_inode(inode);
442 struct ceph_snap_context *snapc = NULL; 446 struct ceph_snap_context *snapc = NULL;
@@ -476,8 +480,9 @@ static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
476 struct ceph_osd_client *osdc; 480 struct ceph_osd_client *osdc;
477 struct ceph_snap_context *snapc, *oldest; 481 struct ceph_snap_context *snapc, *oldest;
478 loff_t page_off = page_offset(page); 482 loff_t page_off = page_offset(page);
483 loff_t snap_size = -1;
479 long writeback_stat; 484 long writeback_stat;
480 u64 truncate_size, snap_size = 0; 485 u64 truncate_size;
481 u32 truncate_seq; 486 u32 truncate_seq;
482 int err = 0, len = PAGE_CACHE_SIZE; 487 int err = 0, len = PAGE_CACHE_SIZE;
483 488
@@ -512,7 +517,7 @@ static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
512 spin_lock(&ci->i_ceph_lock); 517 spin_lock(&ci->i_ceph_lock);
513 truncate_seq = ci->i_truncate_seq; 518 truncate_seq = ci->i_truncate_seq;
514 truncate_size = ci->i_truncate_size; 519 truncate_size = ci->i_truncate_size;
515 if (!snap_size) 520 if (snap_size == -1)
516 snap_size = i_size_read(inode); 521 snap_size = i_size_read(inode);
517 spin_unlock(&ci->i_ceph_lock); 522 spin_unlock(&ci->i_ceph_lock);
518 523
@@ -695,7 +700,8 @@ static int ceph_writepages_start(struct address_space *mapping,
695 unsigned wsize = 1 << inode->i_blkbits; 700 unsigned wsize = 1 << inode->i_blkbits;
696 struct ceph_osd_request *req = NULL; 701 struct ceph_osd_request *req = NULL;
697 int do_sync = 0; 702 int do_sync = 0;
698 u64 truncate_size, snap_size; 703 loff_t snap_size, i_size;
704 u64 truncate_size;
699 u32 truncate_seq; 705 u32 truncate_seq;
700 706
701 /* 707 /*
@@ -741,7 +747,7 @@ static int ceph_writepages_start(struct address_space *mapping,
741retry: 747retry:
742 /* find oldest snap context with dirty data */ 748 /* find oldest snap context with dirty data */
743 ceph_put_snap_context(snapc); 749 ceph_put_snap_context(snapc);
744 snap_size = 0; 750 snap_size = -1;
745 snapc = get_oldest_context(inode, &snap_size); 751 snapc = get_oldest_context(inode, &snap_size);
746 if (!snapc) { 752 if (!snapc) {
747 /* hmm, why does writepages get called when there 753 /* hmm, why does writepages get called when there
@@ -749,16 +755,13 @@ retry:
749 dout(" no snap context with dirty data?\n"); 755 dout(" no snap context with dirty data?\n");
750 goto out; 756 goto out;
751 } 757 }
752 if (snap_size == 0)
753 snap_size = i_size_read(inode);
754 dout(" oldest snapc is %p seq %lld (%d snaps)\n", 758 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
755 snapc, snapc->seq, snapc->num_snaps); 759 snapc, snapc->seq, snapc->num_snaps);
756 760
757 spin_lock(&ci->i_ceph_lock); 761 spin_lock(&ci->i_ceph_lock);
758 truncate_seq = ci->i_truncate_seq; 762 truncate_seq = ci->i_truncate_seq;
759 truncate_size = ci->i_truncate_size; 763 truncate_size = ci->i_truncate_size;
760 if (!snap_size) 764 i_size = i_size_read(inode);
761 snap_size = i_size_read(inode);
762 spin_unlock(&ci->i_ceph_lock); 765 spin_unlock(&ci->i_ceph_lock);
763 766
764 if (last_snapc && snapc != last_snapc) { 767 if (last_snapc && snapc != last_snapc) {
@@ -828,8 +831,10 @@ get_more_pages:
828 dout("waiting on writeback %p\n", page); 831 dout("waiting on writeback %p\n", page);
829 wait_on_page_writeback(page); 832 wait_on_page_writeback(page);
830 } 833 }
831 if (page_offset(page) >= snap_size) { 834 if (page_offset(page) >=
832 dout("%p page eof %llu\n", page, snap_size); 835 (snap_size == -1 ? i_size : snap_size)) {
836 dout("%p page eof %llu\n", page,
837 (snap_size == -1 ? i_size : snap_size));
833 done = 1; 838 done = 1;
834 unlock_page(page); 839 unlock_page(page);
835 break; 840 break;
@@ -884,7 +889,8 @@ get_more_pages:
884 } 889 }
885 890
886 if (do_sync) 891 if (do_sync)
887 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 892 osd_req_op_init(req, 1,
893 CEPH_OSD_OP_STARTSYNC, 0);
888 894
889 req->r_callback = writepages_finish; 895 req->r_callback = writepages_finish;
890 req->r_inode = inode; 896 req->r_inode = inode;
@@ -944,10 +950,18 @@ get_more_pages:
944 } 950 }
945 951
946 /* Format the osd request message and submit the write */ 952 /* Format the osd request message and submit the write */
947
948 offset = page_offset(pages[0]); 953 offset = page_offset(pages[0]);
949 len = min(snap_size - offset, 954 len = (u64)locked_pages << PAGE_CACHE_SHIFT;
950 (u64)locked_pages << PAGE_CACHE_SHIFT); 955 if (snap_size == -1) {
956 len = min(len, (u64)i_size_read(inode) - offset);
957 /* writepages_finish() clears writeback pages
958 * according to the data length, so make sure
959 * data length covers all locked pages */
960 len = max(len, 1 +
961 ((u64)(locked_pages - 1) << PAGE_CACHE_SHIFT));
962 } else {
963 len = min(len, snap_size - offset);
964 }
951 dout("writepages got %d pages at %llu~%llu\n", 965 dout("writepages got %d pages at %llu~%llu\n",
952 locked_pages, offset, len); 966 locked_pages, offset, len);
953 967
@@ -1032,7 +1046,6 @@ static int ceph_update_writeable_page(struct file *file,
1032{ 1046{
1033 struct inode *inode = file_inode(file); 1047 struct inode *inode = file_inode(file);
1034 struct ceph_inode_info *ci = ceph_inode(inode); 1048 struct ceph_inode_info *ci = ceph_inode(inode);
1035 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1036 loff_t page_off = pos & PAGE_CACHE_MASK; 1049 loff_t page_off = pos & PAGE_CACHE_MASK;
1037 int pos_in_page = pos & ~PAGE_CACHE_MASK; 1050 int pos_in_page = pos & ~PAGE_CACHE_MASK;
1038 int end_in_page = pos_in_page + len; 1051 int end_in_page = pos_in_page + len;
@@ -1044,10 +1057,6 @@ retry_locked:
1044 /* writepages currently holds page lock, but if we change that later, */ 1057 /* writepages currently holds page lock, but if we change that later, */
1045 wait_on_page_writeback(page); 1058 wait_on_page_writeback(page);
1046 1059
1047 /* check snap context */
1048 BUG_ON(!ci->i_snap_realm);
1049 down_read(&mdsc->snap_rwsem);
1050 BUG_ON(!ci->i_snap_realm->cached_context);
1051 snapc = page_snap_context(page); 1060 snapc = page_snap_context(page);
1052 if (snapc && snapc != ci->i_head_snapc) { 1061 if (snapc && snapc != ci->i_head_snapc) {
1053 /* 1062 /*
@@ -1055,7 +1064,6 @@ retry_locked:
1055 * context! is it writeable now? 1064 * context! is it writeable now?
1056 */ 1065 */
1057 oldest = get_oldest_context(inode, NULL); 1066 oldest = get_oldest_context(inode, NULL);
1058 up_read(&mdsc->snap_rwsem);
1059 1067
1060 if (snapc->seq > oldest->seq) { 1068 if (snapc->seq > oldest->seq) {
1061 ceph_put_snap_context(oldest); 1069 ceph_put_snap_context(oldest);
@@ -1112,7 +1120,6 @@ retry_locked:
1112 } 1120 }
1113 1121
1114 /* we need to read it. */ 1122 /* we need to read it. */
1115 up_read(&mdsc->snap_rwsem);
1116 r = readpage_nounlock(file, page); 1123 r = readpage_nounlock(file, page);
1117 if (r < 0) 1124 if (r < 0)
1118 goto fail_nosnap; 1125 goto fail_nosnap;
@@ -1157,16 +1164,13 @@ static int ceph_write_begin(struct file *file, struct address_space *mapping,
1157 1164
1158/* 1165/*
1159 * we don't do anything in here that simple_write_end doesn't do 1166 * we don't do anything in here that simple_write_end doesn't do
1160 * except adjust dirty page accounting and drop read lock on 1167 * except adjust dirty page accounting
1161 * mdsc->snap_rwsem.
1162 */ 1168 */
1163static int ceph_write_end(struct file *file, struct address_space *mapping, 1169static int ceph_write_end(struct file *file, struct address_space *mapping,
1164 loff_t pos, unsigned len, unsigned copied, 1170 loff_t pos, unsigned len, unsigned copied,
1165 struct page *page, void *fsdata) 1171 struct page *page, void *fsdata)
1166{ 1172{
1167 struct inode *inode = file_inode(file); 1173 struct inode *inode = file_inode(file);
1168 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1169 struct ceph_mds_client *mdsc = fsc->mdsc;
1170 unsigned from = pos & (PAGE_CACHE_SIZE - 1); 1174 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1171 int check_cap = 0; 1175 int check_cap = 0;
1172 1176
@@ -1188,7 +1192,6 @@ static int ceph_write_end(struct file *file, struct address_space *mapping,
1188 set_page_dirty(page); 1192 set_page_dirty(page);
1189 1193
1190 unlock_page(page); 1194 unlock_page(page);
1191 up_read(&mdsc->snap_rwsem);
1192 page_cache_release(page); 1195 page_cache_release(page);
1193 1196
1194 if (check_cap) 1197 if (check_cap)
@@ -1314,13 +1317,17 @@ static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1314 struct inode *inode = file_inode(vma->vm_file); 1317 struct inode *inode = file_inode(vma->vm_file);
1315 struct ceph_inode_info *ci = ceph_inode(inode); 1318 struct ceph_inode_info *ci = ceph_inode(inode);
1316 struct ceph_file_info *fi = vma->vm_file->private_data; 1319 struct ceph_file_info *fi = vma->vm_file->private_data;
1317 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc; 1320 struct ceph_cap_flush *prealloc_cf;
1318 struct page *page = vmf->page; 1321 struct page *page = vmf->page;
1319 loff_t off = page_offset(page); 1322 loff_t off = page_offset(page);
1320 loff_t size = i_size_read(inode); 1323 loff_t size = i_size_read(inode);
1321 size_t len; 1324 size_t len;
1322 int want, got, ret; 1325 int want, got, ret;
1323 1326
1327 prealloc_cf = ceph_alloc_cap_flush();
1328 if (!prealloc_cf)
1329 return VM_FAULT_SIGBUS;
1330
1324 if (ci->i_inline_version != CEPH_INLINE_NONE) { 1331 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1325 struct page *locked_page = NULL; 1332 struct page *locked_page = NULL;
1326 if (off == 0) { 1333 if (off == 0) {
@@ -1330,8 +1337,10 @@ static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1330 ret = ceph_uninline_data(vma->vm_file, locked_page); 1337 ret = ceph_uninline_data(vma->vm_file, locked_page);
1331 if (locked_page) 1338 if (locked_page)
1332 unlock_page(locked_page); 1339 unlock_page(locked_page);
1333 if (ret < 0) 1340 if (ret < 0) {
1334 return VM_FAULT_SIGBUS; 1341 ret = VM_FAULT_SIGBUS;
1342 goto out_free;
1343 }
1335 } 1344 }
1336 1345
1337 if (off + PAGE_CACHE_SIZE <= size) 1346 if (off + PAGE_CACHE_SIZE <= size)
@@ -1353,7 +1362,8 @@ static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1353 break; 1362 break;
1354 if (ret != -ERESTARTSYS) { 1363 if (ret != -ERESTARTSYS) {
1355 WARN_ON(1); 1364 WARN_ON(1);
1356 return VM_FAULT_SIGBUS; 1365 ret = VM_FAULT_SIGBUS;
1366 goto out_free;
1357 } 1367 }
1358 } 1368 }
1359 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n", 1369 dout("page_mkwrite %p %llu~%zd got cap refs on %s\n",
@@ -1373,7 +1383,6 @@ static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1373 if (ret == 0) { 1383 if (ret == 0) {
1374 /* success. we'll keep the page locked. */ 1384 /* success. we'll keep the page locked. */
1375 set_page_dirty(page); 1385 set_page_dirty(page);
1376 up_read(&mdsc->snap_rwsem);
1377 ret = VM_FAULT_LOCKED; 1386 ret = VM_FAULT_LOCKED;
1378 } else { 1387 } else {
1379 if (ret == -ENOMEM) 1388 if (ret == -ENOMEM)
@@ -1389,7 +1398,8 @@ out:
1389 int dirty; 1398 int dirty;
1390 spin_lock(&ci->i_ceph_lock); 1399 spin_lock(&ci->i_ceph_lock);
1391 ci->i_inline_version = CEPH_INLINE_NONE; 1400 ci->i_inline_version = CEPH_INLINE_NONE;
1392 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1401 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1402 &prealloc_cf);
1393 spin_unlock(&ci->i_ceph_lock); 1403 spin_unlock(&ci->i_ceph_lock);
1394 if (dirty) 1404 if (dirty)
1395 __mark_inode_dirty(inode, dirty); 1405 __mark_inode_dirty(inode, dirty);
@@ -1398,6 +1408,8 @@ out:
1398 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n", 1408 dout("page_mkwrite %p %llu~%zd dropping cap refs on %s ret %d\n",
1399 inode, off, len, ceph_cap_string(got), ret); 1409 inode, off, len, ceph_cap_string(got), ret);
1400 ceph_put_cap_refs(ci, got); 1410 ceph_put_cap_refs(ci, got);
1411out_free:
1412 ceph_free_cap_flush(prealloc_cf);
1401 1413
1402 return ret; 1414 return ret;
1403} 1415}
@@ -1509,8 +1521,7 @@ int ceph_uninline_data(struct file *filp, struct page *locked_page)
1509 ceph_vino(inode), 0, &len, 0, 1, 1521 ceph_vino(inode), 0, &len, 0, 1,
1510 CEPH_OSD_OP_CREATE, 1522 CEPH_OSD_OP_CREATE,
1511 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 1523 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1512 ci->i_snap_realm->cached_context, 1524 ceph_empty_snapc, 0, 0, false);
1513 0, 0, false);
1514 if (IS_ERR(req)) { 1525 if (IS_ERR(req)) {
1515 err = PTR_ERR(req); 1526 err = PTR_ERR(req);
1516 goto out; 1527 goto out;
@@ -1528,7 +1539,7 @@ int ceph_uninline_data(struct file *filp, struct page *locked_page)
1528 ceph_vino(inode), 0, &len, 1, 3, 1539 ceph_vino(inode), 0, &len, 1, 3,
1529 CEPH_OSD_OP_WRITE, 1540 CEPH_OSD_OP_WRITE,
1530 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE, 1541 CEPH_OSD_FLAG_ONDISK | CEPH_OSD_FLAG_WRITE,
1531 ci->i_snap_realm->cached_context, 1542 ceph_empty_snapc,
1532 ci->i_truncate_seq, ci->i_truncate_size, 1543 ci->i_truncate_seq, ci->i_truncate_size,
1533 false); 1544 false);
1534 if (IS_ERR(req)) { 1545 if (IS_ERR(req)) {
@@ -1597,3 +1608,206 @@ int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1597 vma->vm_ops = &ceph_vmops; 1608 vma->vm_ops = &ceph_vmops;
1598 return 0; 1609 return 0;
1599} 1610}
1611
1612enum {
1613 POOL_READ = 1,
1614 POOL_WRITE = 2,
1615};
1616
1617static int __ceph_pool_perm_get(struct ceph_inode_info *ci, u32 pool)
1618{
1619 struct ceph_fs_client *fsc = ceph_inode_to_client(&ci->vfs_inode);
1620 struct ceph_mds_client *mdsc = fsc->mdsc;
1621 struct ceph_osd_request *rd_req = NULL, *wr_req = NULL;
1622 struct rb_node **p, *parent;
1623 struct ceph_pool_perm *perm;
1624 struct page **pages;
1625 int err = 0, err2 = 0, have = 0;
1626
1627 down_read(&mdsc->pool_perm_rwsem);
1628 p = &mdsc->pool_perm_tree.rb_node;
1629 while (*p) {
1630 perm = rb_entry(*p, struct ceph_pool_perm, node);
1631 if (pool < perm->pool)
1632 p = &(*p)->rb_left;
1633 else if (pool > perm->pool)
1634 p = &(*p)->rb_right;
1635 else {
1636 have = perm->perm;
1637 break;
1638 }
1639 }
1640 up_read(&mdsc->pool_perm_rwsem);
1641 if (*p)
1642 goto out;
1643
1644 dout("__ceph_pool_perm_get pool %u no perm cached\n", pool);
1645
1646 down_write(&mdsc->pool_perm_rwsem);
1647 parent = NULL;
1648 while (*p) {
1649 parent = *p;
1650 perm = rb_entry(parent, struct ceph_pool_perm, node);
1651 if (pool < perm->pool)
1652 p = &(*p)->rb_left;
1653 else if (pool > perm->pool)
1654 p = &(*p)->rb_right;
1655 else {
1656 have = perm->perm;
1657 break;
1658 }
1659 }
1660 if (*p) {
1661 up_write(&mdsc->pool_perm_rwsem);
1662 goto out;
1663 }
1664
1665 rd_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1666 ceph_empty_snapc,
1667 1, false, GFP_NOFS);
1668 if (!rd_req) {
1669 err = -ENOMEM;
1670 goto out_unlock;
1671 }
1672
1673 rd_req->r_flags = CEPH_OSD_FLAG_READ;
1674 osd_req_op_init(rd_req, 0, CEPH_OSD_OP_STAT, 0);
1675 rd_req->r_base_oloc.pool = pool;
1676 snprintf(rd_req->r_base_oid.name, sizeof(rd_req->r_base_oid.name),
1677 "%llx.00000000", ci->i_vino.ino);
1678 rd_req->r_base_oid.name_len = strlen(rd_req->r_base_oid.name);
1679
1680 wr_req = ceph_osdc_alloc_request(&fsc->client->osdc,
1681 ceph_empty_snapc,
1682 1, false, GFP_NOFS);
1683 if (!wr_req) {
1684 err = -ENOMEM;
1685 goto out_unlock;
1686 }
1687
1688 wr_req->r_flags = CEPH_OSD_FLAG_WRITE |
1689 CEPH_OSD_FLAG_ACK | CEPH_OSD_FLAG_ONDISK;
1690 osd_req_op_init(wr_req, 0, CEPH_OSD_OP_CREATE, CEPH_OSD_OP_FLAG_EXCL);
1691 wr_req->r_base_oloc.pool = pool;
1692 wr_req->r_base_oid = rd_req->r_base_oid;
1693
1694 /* one page should be large enough for STAT data */
1695 pages = ceph_alloc_page_vector(1, GFP_KERNEL);
1696 if (IS_ERR(pages)) {
1697 err = PTR_ERR(pages);
1698 goto out_unlock;
1699 }
1700
1701 osd_req_op_raw_data_in_pages(rd_req, 0, pages, PAGE_SIZE,
1702 0, false, true);
1703 ceph_osdc_build_request(rd_req, 0, NULL, CEPH_NOSNAP,
1704 &ci->vfs_inode.i_mtime);
1705 err = ceph_osdc_start_request(&fsc->client->osdc, rd_req, false);
1706
1707 ceph_osdc_build_request(wr_req, 0, NULL, CEPH_NOSNAP,
1708 &ci->vfs_inode.i_mtime);
1709 err2 = ceph_osdc_start_request(&fsc->client->osdc, wr_req, false);
1710
1711 if (!err)
1712 err = ceph_osdc_wait_request(&fsc->client->osdc, rd_req);
1713 if (!err2)
1714 err2 = ceph_osdc_wait_request(&fsc->client->osdc, wr_req);
1715
1716 if (err >= 0 || err == -ENOENT)
1717 have |= POOL_READ;
1718 else if (err != -EPERM)
1719 goto out_unlock;
1720
1721 if (err2 == 0 || err2 == -EEXIST)
1722 have |= POOL_WRITE;
1723 else if (err2 != -EPERM) {
1724 err = err2;
1725 goto out_unlock;
1726 }
1727
1728 perm = kmalloc(sizeof(*perm), GFP_NOFS);
1729 if (!perm) {
1730 err = -ENOMEM;
1731 goto out_unlock;
1732 }
1733
1734 perm->pool = pool;
1735 perm->perm = have;
1736 rb_link_node(&perm->node, parent, p);
1737 rb_insert_color(&perm->node, &mdsc->pool_perm_tree);
1738 err = 0;
1739out_unlock:
1740 up_write(&mdsc->pool_perm_rwsem);
1741
1742 if (rd_req)
1743 ceph_osdc_put_request(rd_req);
1744 if (wr_req)
1745 ceph_osdc_put_request(wr_req);
1746out:
1747 if (!err)
1748 err = have;
1749 dout("__ceph_pool_perm_get pool %u result = %d\n", pool, err);
1750 return err;
1751}
1752
1753int ceph_pool_perm_check(struct ceph_inode_info *ci, int need)
1754{
1755 u32 pool;
1756 int ret, flags;
1757
1758 if (ceph_test_mount_opt(ceph_inode_to_client(&ci->vfs_inode),
1759 NOPOOLPERM))
1760 return 0;
1761
1762 spin_lock(&ci->i_ceph_lock);
1763 flags = ci->i_ceph_flags;
1764 pool = ceph_file_layout_pg_pool(ci->i_layout);
1765 spin_unlock(&ci->i_ceph_lock);
1766check:
1767 if (flags & CEPH_I_POOL_PERM) {
1768 if ((need & CEPH_CAP_FILE_RD) && !(flags & CEPH_I_POOL_RD)) {
1769 dout("ceph_pool_perm_check pool %u no read perm\n",
1770 pool);
1771 return -EPERM;
1772 }
1773 if ((need & CEPH_CAP_FILE_WR) && !(flags & CEPH_I_POOL_WR)) {
1774 dout("ceph_pool_perm_check pool %u no write perm\n",
1775 pool);
1776 return -EPERM;
1777 }
1778 return 0;
1779 }
1780
1781 ret = __ceph_pool_perm_get(ci, pool);
1782 if (ret < 0)
1783 return ret;
1784
1785 flags = CEPH_I_POOL_PERM;
1786 if (ret & POOL_READ)
1787 flags |= CEPH_I_POOL_RD;
1788 if (ret & POOL_WRITE)
1789 flags |= CEPH_I_POOL_WR;
1790
1791 spin_lock(&ci->i_ceph_lock);
1792 if (pool == ceph_file_layout_pg_pool(ci->i_layout)) {
1793 ci->i_ceph_flags = flags;
1794 } else {
1795 pool = ceph_file_layout_pg_pool(ci->i_layout);
1796 flags = ci->i_ceph_flags;
1797 }
1798 spin_unlock(&ci->i_ceph_lock);
1799 goto check;
1800}
1801
1802void ceph_pool_perm_destroy(struct ceph_mds_client *mdsc)
1803{
1804 struct ceph_pool_perm *perm;
1805 struct rb_node *n;
1806
1807 while (!RB_EMPTY_ROOT(&mdsc->pool_perm_tree)) {
1808 n = rb_first(&mdsc->pool_perm_tree);
1809 perm = rb_entry(n, struct ceph_pool_perm, node);
1810 rb_erase(n, &mdsc->pool_perm_tree);
1811 kfree(perm);
1812 }
1813}
diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
index be5ea6af8366..dc10c9dd36c1 100644
--- a/fs/ceph/caps.c
+++ b/fs/ceph/caps.c
@@ -833,7 +833,9 @@ int __ceph_caps_used(struct ceph_inode_info *ci)
833 used |= CEPH_CAP_PIN; 833 used |= CEPH_CAP_PIN;
834 if (ci->i_rd_ref) 834 if (ci->i_rd_ref)
835 used |= CEPH_CAP_FILE_RD; 835 used |= CEPH_CAP_FILE_RD;
836 if (ci->i_rdcache_ref || ci->vfs_inode.i_data.nrpages) 836 if (ci->i_rdcache_ref ||
837 (!S_ISDIR(ci->vfs_inode.i_mode) && /* ignore readdir cache */
838 ci->vfs_inode.i_data.nrpages))
837 used |= CEPH_CAP_FILE_CACHE; 839 used |= CEPH_CAP_FILE_CACHE;
838 if (ci->i_wr_ref) 840 if (ci->i_wr_ref)
839 used |= CEPH_CAP_FILE_WR; 841 used |= CEPH_CAP_FILE_WR;
@@ -926,16 +928,6 @@ void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
926 928
927 /* remove from session list */ 929 /* remove from session list */
928 spin_lock(&session->s_cap_lock); 930 spin_lock(&session->s_cap_lock);
929 /*
930 * s_cap_reconnect is protected by s_cap_lock. no one changes
931 * s_cap_gen while session is in the reconnect state.
932 */
933 if (queue_release &&
934 (!session->s_cap_reconnect ||
935 cap->cap_gen == session->s_cap_gen))
936 __queue_cap_release(session, ci->i_vino.ino, cap->cap_id,
937 cap->mseq, cap->issue_seq);
938
939 if (session->s_cap_iterator == cap) { 931 if (session->s_cap_iterator == cap) {
940 /* not yet, we are iterating over this very cap */ 932 /* not yet, we are iterating over this very cap */
941 dout("__ceph_remove_cap delaying %p removal from session %p\n", 933 dout("__ceph_remove_cap delaying %p removal from session %p\n",
@@ -948,6 +940,25 @@ void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
948 } 940 }
949 /* protect backpointer with s_cap_lock: see iterate_session_caps */ 941 /* protect backpointer with s_cap_lock: see iterate_session_caps */
950 cap->ci = NULL; 942 cap->ci = NULL;
943
944 /*
945 * s_cap_reconnect is protected by s_cap_lock. no one changes
946 * s_cap_gen while session is in the reconnect state.
947 */
948 if (queue_release &&
949 (!session->s_cap_reconnect || cap->cap_gen == session->s_cap_gen)) {
950 cap->queue_release = 1;
951 if (removed) {
952 list_add_tail(&cap->session_caps,
953 &session->s_cap_releases);
954 session->s_num_cap_releases++;
955 removed = 0;
956 }
957 } else {
958 cap->queue_release = 0;
959 }
960 cap->cap_ino = ci->i_vino.ino;
961
951 spin_unlock(&session->s_cap_lock); 962 spin_unlock(&session->s_cap_lock);
952 963
953 /* remove from inode list */ 964 /* remove from inode list */
@@ -977,8 +988,8 @@ void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
977static int send_cap_msg(struct ceph_mds_session *session, 988static int send_cap_msg(struct ceph_mds_session *session,
978 u64 ino, u64 cid, int op, 989 u64 ino, u64 cid, int op,
979 int caps, int wanted, int dirty, 990 int caps, int wanted, int dirty,
980 u32 seq, u64 flush_tid, u32 issue_seq, u32 mseq, 991 u32 seq, u64 flush_tid, u64 oldest_flush_tid,
981 u64 size, u64 max_size, 992 u32 issue_seq, u32 mseq, u64 size, u64 max_size,
982 struct timespec *mtime, struct timespec *atime, 993 struct timespec *mtime, struct timespec *atime,
983 u64 time_warp_seq, 994 u64 time_warp_seq,
984 kuid_t uid, kgid_t gid, umode_t mode, 995 kuid_t uid, kgid_t gid, umode_t mode,
@@ -992,20 +1003,23 @@ static int send_cap_msg(struct ceph_mds_session *session,
992 size_t extra_len; 1003 size_t extra_len;
993 1004
994 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s" 1005 dout("send_cap_msg %s %llx %llx caps %s wanted %s dirty %s"
995 " seq %u/%u mseq %u follows %lld size %llu/%llu" 1006 " seq %u/%u tid %llu/%llu mseq %u follows %lld size %llu/%llu"
996 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op), 1007 " xattr_ver %llu xattr_len %d\n", ceph_cap_op_name(op),
997 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted), 1008 cid, ino, ceph_cap_string(caps), ceph_cap_string(wanted),
998 ceph_cap_string(dirty), 1009 ceph_cap_string(dirty),
999 seq, issue_seq, mseq, follows, size, max_size, 1010 seq, issue_seq, flush_tid, oldest_flush_tid,
1011 mseq, follows, size, max_size,
1000 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0); 1012 xattr_version, xattrs_buf ? (int)xattrs_buf->vec.iov_len : 0);
1001 1013
1002 /* flock buffer size + inline version + inline data size */ 1014 /* flock buffer size + inline version + inline data size +
1003 extra_len = 4 + 8 + 4; 1015 * osd_epoch_barrier + oldest_flush_tid */
1016 extra_len = 4 + 8 + 4 + 4 + 8;
1004 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len, 1017 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
1005 GFP_NOFS, false); 1018 GFP_NOFS, false);
1006 if (!msg) 1019 if (!msg)
1007 return -ENOMEM; 1020 return -ENOMEM;
1008 1021
1022 msg->hdr.version = cpu_to_le16(6);
1009 msg->hdr.tid = cpu_to_le64(flush_tid); 1023 msg->hdr.tid = cpu_to_le64(flush_tid);
1010 1024
1011 fc = msg->front.iov_base; 1025 fc = msg->front.iov_base;
@@ -1041,6 +1055,10 @@ static int send_cap_msg(struct ceph_mds_session *session,
1041 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE); 1055 ceph_encode_64(&p, inline_data ? 0 : CEPH_INLINE_NONE);
1042 /* inline data size */ 1056 /* inline data size */
1043 ceph_encode_32(&p, 0); 1057 ceph_encode_32(&p, 0);
1058 /* osd_epoch_barrier */
1059 ceph_encode_32(&p, 0);
1060 /* oldest_flush_tid */
1061 ceph_encode_64(&p, oldest_flush_tid);
1044 1062
1045 fc->xattr_version = cpu_to_le64(xattr_version); 1063 fc->xattr_version = cpu_to_le64(xattr_version);
1046 if (xattrs_buf) { 1064 if (xattrs_buf) {
@@ -1053,44 +1071,6 @@ static int send_cap_msg(struct ceph_mds_session *session,
1053 return 0; 1071 return 0;
1054} 1072}
1055 1073
1056void __queue_cap_release(struct ceph_mds_session *session,
1057 u64 ino, u64 cap_id, u32 migrate_seq,
1058 u32 issue_seq)
1059{
1060 struct ceph_msg *msg;
1061 struct ceph_mds_cap_release *head;
1062 struct ceph_mds_cap_item *item;
1063
1064 BUG_ON(!session->s_num_cap_releases);
1065 msg = list_first_entry(&session->s_cap_releases,
1066 struct ceph_msg, list_head);
1067
1068 dout(" adding %llx release to mds%d msg %p (%d left)\n",
1069 ino, session->s_mds, msg, session->s_num_cap_releases);
1070
1071 BUG_ON(msg->front.iov_len + sizeof(*item) > PAGE_CACHE_SIZE);
1072 head = msg->front.iov_base;
1073 le32_add_cpu(&head->num, 1);
1074 item = msg->front.iov_base + msg->front.iov_len;
1075 item->ino = cpu_to_le64(ino);
1076 item->cap_id = cpu_to_le64(cap_id);
1077 item->migrate_seq = cpu_to_le32(migrate_seq);
1078 item->seq = cpu_to_le32(issue_seq);
1079
1080 session->s_num_cap_releases--;
1081
1082 msg->front.iov_len += sizeof(*item);
1083 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1084 dout(" release msg %p full\n", msg);
1085 list_move_tail(&msg->list_head, &session->s_cap_releases_done);
1086 } else {
1087 dout(" release msg %p at %d/%d (%d)\n", msg,
1088 (int)le32_to_cpu(head->num),
1089 (int)CEPH_CAPS_PER_RELEASE,
1090 (int)msg->front.iov_len);
1091 }
1092}
1093
1094/* 1074/*
1095 * Queue cap releases when an inode is dropped from our cache. Since 1075 * Queue cap releases when an inode is dropped from our cache. Since
1096 * inode is about to be destroyed, there is no need for i_ceph_lock. 1076 * inode is about to be destroyed, there is no need for i_ceph_lock.
@@ -1127,7 +1107,7 @@ void ceph_queue_caps_release(struct inode *inode)
1127 */ 1107 */
1128static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap, 1108static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1129 int op, int used, int want, int retain, int flushing, 1109 int op, int used, int want, int retain, int flushing,
1130 unsigned *pflush_tid) 1110 u64 flush_tid, u64 oldest_flush_tid)
1131 __releases(cap->ci->i_ceph_lock) 1111 __releases(cap->ci->i_ceph_lock)
1132{ 1112{
1133 struct ceph_inode_info *ci = cap->ci; 1113 struct ceph_inode_info *ci = cap->ci;
@@ -1145,8 +1125,6 @@ static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1145 u64 xattr_version = 0; 1125 u64 xattr_version = 0;
1146 struct ceph_buffer *xattr_blob = NULL; 1126 struct ceph_buffer *xattr_blob = NULL;
1147 int delayed = 0; 1127 int delayed = 0;
1148 u64 flush_tid = 0;
1149 int i;
1150 int ret; 1128 int ret;
1151 bool inline_data; 1129 bool inline_data;
1152 1130
@@ -1190,26 +1168,7 @@ static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1190 cap->implemented &= cap->issued | used; 1168 cap->implemented &= cap->issued | used;
1191 cap->mds_wanted = want; 1169 cap->mds_wanted = want;
1192 1170
1193 if (flushing) { 1171 follows = flushing ? ci->i_head_snapc->seq : 0;
1194 /*
1195 * assign a tid for flush operations so we can avoid
1196 * flush1 -> dirty1 -> flush2 -> flushack1 -> mark
1197 * clean type races. track latest tid for every bit
1198 * so we can handle flush AxFw, flush Fw, and have the
1199 * first ack clean Ax.
1200 */
1201 flush_tid = ++ci->i_cap_flush_last_tid;
1202 if (pflush_tid)
1203 *pflush_tid = flush_tid;
1204 dout(" cap_flush_tid %d\n", (int)flush_tid);
1205 for (i = 0; i < CEPH_CAP_BITS; i++)
1206 if (flushing & (1 << i))
1207 ci->i_cap_flush_tid[i] = flush_tid;
1208
1209 follows = ci->i_head_snapc->seq;
1210 } else {
1211 follows = 0;
1212 }
1213 1172
1214 keep = cap->implemented; 1173 keep = cap->implemented;
1215 seq = cap->seq; 1174 seq = cap->seq;
@@ -1237,7 +1196,8 @@ static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1237 spin_unlock(&ci->i_ceph_lock); 1196 spin_unlock(&ci->i_ceph_lock);
1238 1197
1239 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id, 1198 ret = send_cap_msg(session, ceph_vino(inode).ino, cap_id,
1240 op, keep, want, flushing, seq, flush_tid, issue_seq, mseq, 1199 op, keep, want, flushing, seq,
1200 flush_tid, oldest_flush_tid, issue_seq, mseq,
1241 size, max_size, &mtime, &atime, time_warp_seq, 1201 size, max_size, &mtime, &atime, time_warp_seq,
1242 uid, gid, mode, xattr_version, xattr_blob, 1202 uid, gid, mode, xattr_version, xattr_blob,
1243 follows, inline_data); 1203 follows, inline_data);
@@ -1259,14 +1219,14 @@ static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
1259 * asynchronously back to the MDS once sync writes complete and dirty 1219 * asynchronously back to the MDS once sync writes complete and dirty
1260 * data is written out. 1220 * data is written out.
1261 * 1221 *
1262 * Unless @again is true, skip cap_snaps that were already sent to 1222 * Unless @kick is true, skip cap_snaps that were already sent to
1263 * the MDS (i.e., during this session). 1223 * the MDS (i.e., during this session).
1264 * 1224 *
1265 * Called under i_ceph_lock. Takes s_mutex as needed. 1225 * Called under i_ceph_lock. Takes s_mutex as needed.
1266 */ 1226 */
1267void __ceph_flush_snaps(struct ceph_inode_info *ci, 1227void __ceph_flush_snaps(struct ceph_inode_info *ci,
1268 struct ceph_mds_session **psession, 1228 struct ceph_mds_session **psession,
1269 int again) 1229 int kick)
1270 __releases(ci->i_ceph_lock) 1230 __releases(ci->i_ceph_lock)
1271 __acquires(ci->i_ceph_lock) 1231 __acquires(ci->i_ceph_lock)
1272{ 1232{
@@ -1297,11 +1257,8 @@ retry:
1297 if (capsnap->dirty_pages || capsnap->writing) 1257 if (capsnap->dirty_pages || capsnap->writing)
1298 break; 1258 break;
1299 1259
1300 /* 1260 /* should be removed by ceph_try_drop_cap_snap() */
1301 * if cap writeback already occurred, we should have dropped 1261 BUG_ON(!capsnap->need_flush);
1302 * the capsnap in ceph_put_wrbuffer_cap_refs.
1303 */
1304 BUG_ON(capsnap->dirty == 0);
1305 1262
1306 /* pick mds, take s_mutex */ 1263 /* pick mds, take s_mutex */
1307 if (ci->i_auth_cap == NULL) { 1264 if (ci->i_auth_cap == NULL) {
@@ -1310,7 +1267,7 @@ retry:
1310 } 1267 }
1311 1268
1312 /* only flush each capsnap once */ 1269 /* only flush each capsnap once */
1313 if (!again && !list_empty(&capsnap->flushing_item)) { 1270 if (!kick && !list_empty(&capsnap->flushing_item)) {
1314 dout("already flushed %p, skipping\n", capsnap); 1271 dout("already flushed %p, skipping\n", capsnap);
1315 continue; 1272 continue;
1316 } 1273 }
@@ -1320,6 +1277,9 @@ retry:
1320 1277
1321 if (session && session->s_mds != mds) { 1278 if (session && session->s_mds != mds) {
1322 dout("oops, wrong session %p mutex\n", session); 1279 dout("oops, wrong session %p mutex\n", session);
1280 if (kick)
1281 goto out;
1282
1323 mutex_unlock(&session->s_mutex); 1283 mutex_unlock(&session->s_mutex);
1324 ceph_put_mds_session(session); 1284 ceph_put_mds_session(session);
1325 session = NULL; 1285 session = NULL;
@@ -1343,20 +1303,22 @@ retry:
1343 goto retry; 1303 goto retry;
1344 } 1304 }
1345 1305
1346 capsnap->flush_tid = ++ci->i_cap_flush_last_tid; 1306 spin_lock(&mdsc->cap_dirty_lock);
1307 capsnap->flush_tid = ++mdsc->last_cap_flush_tid;
1308 spin_unlock(&mdsc->cap_dirty_lock);
1309
1347 atomic_inc(&capsnap->nref); 1310 atomic_inc(&capsnap->nref);
1348 if (!list_empty(&capsnap->flushing_item)) 1311 if (list_empty(&capsnap->flushing_item))
1349 list_del_init(&capsnap->flushing_item); 1312 list_add_tail(&capsnap->flushing_item,
1350 list_add_tail(&capsnap->flushing_item, 1313 &session->s_cap_snaps_flushing);
1351 &session->s_cap_snaps_flushing);
1352 spin_unlock(&ci->i_ceph_lock); 1314 spin_unlock(&ci->i_ceph_lock);
1353 1315
1354 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n", 1316 dout("flush_snaps %p cap_snap %p follows %lld tid %llu\n",
1355 inode, capsnap, capsnap->follows, capsnap->flush_tid); 1317 inode, capsnap, capsnap->follows, capsnap->flush_tid);
1356 send_cap_msg(session, ceph_vino(inode).ino, 0, 1318 send_cap_msg(session, ceph_vino(inode).ino, 0,
1357 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0, 1319 CEPH_CAP_OP_FLUSHSNAP, capsnap->issued, 0,
1358 capsnap->dirty, 0, capsnap->flush_tid, 0, mseq, 1320 capsnap->dirty, 0, capsnap->flush_tid, 0,
1359 capsnap->size, 0, 1321 0, mseq, capsnap->size, 0,
1360 &capsnap->mtime, &capsnap->atime, 1322 &capsnap->mtime, &capsnap->atime,
1361 capsnap->time_warp_seq, 1323 capsnap->time_warp_seq,
1362 capsnap->uid, capsnap->gid, capsnap->mode, 1324 capsnap->uid, capsnap->gid, capsnap->mode,
@@ -1396,7 +1358,8 @@ static void ceph_flush_snaps(struct ceph_inode_info *ci)
1396 * Caller is then responsible for calling __mark_inode_dirty with the 1358 * Caller is then responsible for calling __mark_inode_dirty with the
1397 * returned flags value. 1359 * returned flags value.
1398 */ 1360 */
1399int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask) 1361int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
1362 struct ceph_cap_flush **pcf)
1400{ 1363{
1401 struct ceph_mds_client *mdsc = 1364 struct ceph_mds_client *mdsc =
1402 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc; 1365 ceph_sb_to_client(ci->vfs_inode.i_sb)->mdsc;
@@ -1416,9 +1379,14 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1416 ceph_cap_string(was | mask)); 1379 ceph_cap_string(was | mask));
1417 ci->i_dirty_caps |= mask; 1380 ci->i_dirty_caps |= mask;
1418 if (was == 0) { 1381 if (was == 0) {
1419 if (!ci->i_head_snapc) 1382 WARN_ON_ONCE(ci->i_prealloc_cap_flush);
1383 swap(ci->i_prealloc_cap_flush, *pcf);
1384
1385 if (!ci->i_head_snapc) {
1386 WARN_ON_ONCE(!rwsem_is_locked(&mdsc->snap_rwsem));
1420 ci->i_head_snapc = ceph_get_snap_context( 1387 ci->i_head_snapc = ceph_get_snap_context(
1421 ci->i_snap_realm->cached_context); 1388 ci->i_snap_realm->cached_context);
1389 }
1422 dout(" inode %p now dirty snapc %p auth cap %p\n", 1390 dout(" inode %p now dirty snapc %p auth cap %p\n",
1423 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap); 1391 &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
1424 BUG_ON(!list_empty(&ci->i_dirty_item)); 1392 BUG_ON(!list_empty(&ci->i_dirty_item));
@@ -1429,6 +1397,8 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1429 ihold(inode); 1397 ihold(inode);
1430 dirty |= I_DIRTY_SYNC; 1398 dirty |= I_DIRTY_SYNC;
1431 } 1399 }
1400 } else {
1401 WARN_ON_ONCE(!ci->i_prealloc_cap_flush);
1432 } 1402 }
1433 BUG_ON(list_empty(&ci->i_dirty_item)); 1403 BUG_ON(list_empty(&ci->i_dirty_item));
1434 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) && 1404 if (((was | ci->i_flushing_caps) & CEPH_CAP_FILE_BUFFER) &&
@@ -1438,6 +1408,74 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1438 return dirty; 1408 return dirty;
1439} 1409}
1440 1410
1411static void __add_cap_flushing_to_inode(struct ceph_inode_info *ci,
1412 struct ceph_cap_flush *cf)
1413{
1414 struct rb_node **p = &ci->i_cap_flush_tree.rb_node;
1415 struct rb_node *parent = NULL;
1416 struct ceph_cap_flush *other = NULL;
1417
1418 while (*p) {
1419 parent = *p;
1420 other = rb_entry(parent, struct ceph_cap_flush, i_node);
1421
1422 if (cf->tid < other->tid)
1423 p = &(*p)->rb_left;
1424 else if (cf->tid > other->tid)
1425 p = &(*p)->rb_right;
1426 else
1427 BUG();
1428 }
1429
1430 rb_link_node(&cf->i_node, parent, p);
1431 rb_insert_color(&cf->i_node, &ci->i_cap_flush_tree);
1432}
1433
1434static void __add_cap_flushing_to_mdsc(struct ceph_mds_client *mdsc,
1435 struct ceph_cap_flush *cf)
1436{
1437 struct rb_node **p = &mdsc->cap_flush_tree.rb_node;
1438 struct rb_node *parent = NULL;
1439 struct ceph_cap_flush *other = NULL;
1440
1441 while (*p) {
1442 parent = *p;
1443 other = rb_entry(parent, struct ceph_cap_flush, g_node);
1444
1445 if (cf->tid < other->tid)
1446 p = &(*p)->rb_left;
1447 else if (cf->tid > other->tid)
1448 p = &(*p)->rb_right;
1449 else
1450 BUG();
1451 }
1452
1453 rb_link_node(&cf->g_node, parent, p);
1454 rb_insert_color(&cf->g_node, &mdsc->cap_flush_tree);
1455}
1456
1457struct ceph_cap_flush *ceph_alloc_cap_flush(void)
1458{
1459 return kmem_cache_alloc(ceph_cap_flush_cachep, GFP_KERNEL);
1460}
1461
1462void ceph_free_cap_flush(struct ceph_cap_flush *cf)
1463{
1464 if (cf)
1465 kmem_cache_free(ceph_cap_flush_cachep, cf);
1466}
1467
1468static u64 __get_oldest_flush_tid(struct ceph_mds_client *mdsc)
1469{
1470 struct rb_node *n = rb_first(&mdsc->cap_flush_tree);
1471 if (n) {
1472 struct ceph_cap_flush *cf =
1473 rb_entry(n, struct ceph_cap_flush, g_node);
1474 return cf->tid;
1475 }
1476 return 0;
1477}
1478
1441/* 1479/*
1442 * Add dirty inode to the flushing list. Assigned a seq number so we 1480 * Add dirty inode to the flushing list. Assigned a seq number so we
1443 * can wait for caps to flush without starving. 1481 * can wait for caps to flush without starving.
@@ -1445,14 +1483,17 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask)
1445 * Called under i_ceph_lock. 1483 * Called under i_ceph_lock.
1446 */ 1484 */
1447static int __mark_caps_flushing(struct inode *inode, 1485static int __mark_caps_flushing(struct inode *inode,
1448 struct ceph_mds_session *session) 1486 struct ceph_mds_session *session,
1487 u64 *flush_tid, u64 *oldest_flush_tid)
1449{ 1488{
1450 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 1489 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1451 struct ceph_inode_info *ci = ceph_inode(inode); 1490 struct ceph_inode_info *ci = ceph_inode(inode);
1491 struct ceph_cap_flush *cf = NULL;
1452 int flushing; 1492 int flushing;
1453 1493
1454 BUG_ON(ci->i_dirty_caps == 0); 1494 BUG_ON(ci->i_dirty_caps == 0);
1455 BUG_ON(list_empty(&ci->i_dirty_item)); 1495 BUG_ON(list_empty(&ci->i_dirty_item));
1496 BUG_ON(!ci->i_prealloc_cap_flush);
1456 1497
1457 flushing = ci->i_dirty_caps; 1498 flushing = ci->i_dirty_caps;
1458 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n", 1499 dout("__mark_caps_flushing flushing %s, flushing_caps %s -> %s\n",
@@ -1463,22 +1504,31 @@ static int __mark_caps_flushing(struct inode *inode,
1463 ci->i_dirty_caps = 0; 1504 ci->i_dirty_caps = 0;
1464 dout(" inode %p now !dirty\n", inode); 1505 dout(" inode %p now !dirty\n", inode);
1465 1506
1507 swap(cf, ci->i_prealloc_cap_flush);
1508 cf->caps = flushing;
1509 cf->kick = false;
1510
1466 spin_lock(&mdsc->cap_dirty_lock); 1511 spin_lock(&mdsc->cap_dirty_lock);
1467 list_del_init(&ci->i_dirty_item); 1512 list_del_init(&ci->i_dirty_item);
1468 1513
1514 cf->tid = ++mdsc->last_cap_flush_tid;
1515 __add_cap_flushing_to_mdsc(mdsc, cf);
1516 *oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1517
1469 if (list_empty(&ci->i_flushing_item)) { 1518 if (list_empty(&ci->i_flushing_item)) {
1470 ci->i_cap_flush_seq = ++mdsc->cap_flush_seq;
1471 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1519 list_add_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1472 mdsc->num_cap_flushing++; 1520 mdsc->num_cap_flushing++;
1473 dout(" inode %p now flushing seq %lld\n", inode, 1521 dout(" inode %p now flushing tid %llu\n", inode, cf->tid);
1474 ci->i_cap_flush_seq);
1475 } else { 1522 } else {
1476 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing); 1523 list_move_tail(&ci->i_flushing_item, &session->s_cap_flushing);
1477 dout(" inode %p now flushing (more) seq %lld\n", inode, 1524 dout(" inode %p now flushing (more) tid %llu\n",
1478 ci->i_cap_flush_seq); 1525 inode, cf->tid);
1479 } 1526 }
1480 spin_unlock(&mdsc->cap_dirty_lock); 1527 spin_unlock(&mdsc->cap_dirty_lock);
1481 1528
1529 __add_cap_flushing_to_inode(ci, cf);
1530
1531 *flush_tid = cf->tid;
1482 return flushing; 1532 return flushing;
1483} 1533}
1484 1534
@@ -1524,6 +1574,7 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
1524 struct ceph_mds_client *mdsc = fsc->mdsc; 1574 struct ceph_mds_client *mdsc = fsc->mdsc;
1525 struct inode *inode = &ci->vfs_inode; 1575 struct inode *inode = &ci->vfs_inode;
1526 struct ceph_cap *cap; 1576 struct ceph_cap *cap;
1577 u64 flush_tid, oldest_flush_tid;
1527 int file_wanted, used, cap_used; 1578 int file_wanted, used, cap_used;
1528 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */ 1579 int took_snap_rwsem = 0; /* true if mdsc->snap_rwsem held */
1529 int issued, implemented, want, retain, revoking, flushing = 0; 1580 int issued, implemented, want, retain, revoking, flushing = 0;
@@ -1553,13 +1604,13 @@ retry:
1553retry_locked: 1604retry_locked:
1554 file_wanted = __ceph_caps_file_wanted(ci); 1605 file_wanted = __ceph_caps_file_wanted(ci);
1555 used = __ceph_caps_used(ci); 1606 used = __ceph_caps_used(ci);
1556 want = file_wanted | used;
1557 issued = __ceph_caps_issued(ci, &implemented); 1607 issued = __ceph_caps_issued(ci, &implemented);
1558 revoking = implemented & ~issued; 1608 revoking = implemented & ~issued;
1559 1609
1560 retain = want | CEPH_CAP_PIN; 1610 want = file_wanted;
1611 retain = file_wanted | used | CEPH_CAP_PIN;
1561 if (!mdsc->stopping && inode->i_nlink > 0) { 1612 if (!mdsc->stopping && inode->i_nlink > 0) {
1562 if (want) { 1613 if (file_wanted) {
1563 retain |= CEPH_CAP_ANY; /* be greedy */ 1614 retain |= CEPH_CAP_ANY; /* be greedy */
1564 } else if (S_ISDIR(inode->i_mode) && 1615 } else if (S_ISDIR(inode->i_mode) &&
1565 (issued & CEPH_CAP_FILE_SHARED) && 1616 (issued & CEPH_CAP_FILE_SHARED) &&
@@ -1602,9 +1653,10 @@ retry_locked:
1602 * If we fail, it's because pages are locked.... try again later. 1653 * If we fail, it's because pages are locked.... try again later.
1603 */ 1654 */
1604 if ((!is_delayed || mdsc->stopping) && 1655 if ((!is_delayed || mdsc->stopping) &&
1605 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */ 1656 !S_ISDIR(inode->i_mode) && /* ignore readdir cache */
1606 inode->i_data.nrpages && /* have cached pages */ 1657 ci->i_wrbuffer_ref == 0 && /* no dirty pages... */
1607 (file_wanted == 0 || /* no open files */ 1658 inode->i_data.nrpages && /* have cached pages */
1659 (file_wanted == 0 || /* no open files */
1608 (revoking & (CEPH_CAP_FILE_CACHE| 1660 (revoking & (CEPH_CAP_FILE_CACHE|
1609 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */ 1661 CEPH_CAP_FILE_LAZYIO))) && /* or revoking cache */
1610 !tried_invalidate) { 1662 !tried_invalidate) {
@@ -1742,17 +1794,25 @@ ack:
1742 took_snap_rwsem = 1; 1794 took_snap_rwsem = 1;
1743 } 1795 }
1744 1796
1745 if (cap == ci->i_auth_cap && ci->i_dirty_caps) 1797 if (cap == ci->i_auth_cap && ci->i_dirty_caps) {
1746 flushing = __mark_caps_flushing(inode, session); 1798 flushing = __mark_caps_flushing(inode, session,
1747 else 1799 &flush_tid,
1800 &oldest_flush_tid);
1801 } else {
1748 flushing = 0; 1802 flushing = 0;
1803 flush_tid = 0;
1804 spin_lock(&mdsc->cap_dirty_lock);
1805 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
1806 spin_unlock(&mdsc->cap_dirty_lock);
1807 }
1749 1808
1750 mds = cap->mds; /* remember mds, so we don't repeat */ 1809 mds = cap->mds; /* remember mds, so we don't repeat */
1751 sent++; 1810 sent++;
1752 1811
1753 /* __send_cap drops i_ceph_lock */ 1812 /* __send_cap drops i_ceph_lock */
1754 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used, 1813 delayed += __send_cap(mdsc, cap, CEPH_CAP_OP_UPDATE, cap_used,
1755 want, retain, flushing, NULL); 1814 want, retain, flushing,
1815 flush_tid, oldest_flush_tid);
1756 goto retry; /* retake i_ceph_lock and restart our cap scan. */ 1816 goto retry; /* retake i_ceph_lock and restart our cap scan. */
1757 } 1817 }
1758 1818
@@ -1781,12 +1841,13 @@ ack:
1781/* 1841/*
1782 * Try to flush dirty caps back to the auth mds. 1842 * Try to flush dirty caps back to the auth mds.
1783 */ 1843 */
1784static int try_flush_caps(struct inode *inode, unsigned *flush_tid) 1844static int try_flush_caps(struct inode *inode, u64 *ptid)
1785{ 1845{
1786 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 1846 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
1787 struct ceph_inode_info *ci = ceph_inode(inode); 1847 struct ceph_inode_info *ci = ceph_inode(inode);
1788 int flushing = 0;
1789 struct ceph_mds_session *session = NULL; 1848 struct ceph_mds_session *session = NULL;
1849 int flushing = 0;
1850 u64 flush_tid = 0, oldest_flush_tid = 0;
1790 1851
1791retry: 1852retry:
1792 spin_lock(&ci->i_ceph_lock); 1853 spin_lock(&ci->i_ceph_lock);
@@ -1811,42 +1872,54 @@ retry:
1811 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN) 1872 if (cap->session->s_state < CEPH_MDS_SESSION_OPEN)
1812 goto out; 1873 goto out;
1813 1874
1814 flushing = __mark_caps_flushing(inode, session); 1875 flushing = __mark_caps_flushing(inode, session, &flush_tid,
1876 &oldest_flush_tid);
1815 1877
1816 /* __send_cap drops i_ceph_lock */ 1878 /* __send_cap drops i_ceph_lock */
1817 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want, 1879 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, used, want,
1818 cap->issued | cap->implemented, flushing, 1880 (cap->issued | cap->implemented),
1819 flush_tid); 1881 flushing, flush_tid, oldest_flush_tid);
1820 if (!delayed)
1821 goto out_unlocked;
1822 1882
1823 spin_lock(&ci->i_ceph_lock); 1883 if (delayed) {
1824 __cap_delay_requeue(mdsc, ci); 1884 spin_lock(&ci->i_ceph_lock);
1885 __cap_delay_requeue(mdsc, ci);
1886 spin_unlock(&ci->i_ceph_lock);
1887 }
1888 } else {
1889 struct rb_node *n = rb_last(&ci->i_cap_flush_tree);
1890 if (n) {
1891 struct ceph_cap_flush *cf =
1892 rb_entry(n, struct ceph_cap_flush, i_node);
1893 flush_tid = cf->tid;
1894 }
1895 flushing = ci->i_flushing_caps;
1896 spin_unlock(&ci->i_ceph_lock);
1825 } 1897 }
1826out: 1898out:
1827 spin_unlock(&ci->i_ceph_lock);
1828out_unlocked:
1829 if (session) 1899 if (session)
1830 mutex_unlock(&session->s_mutex); 1900 mutex_unlock(&session->s_mutex);
1901
1902 *ptid = flush_tid;
1831 return flushing; 1903 return flushing;
1832} 1904}
1833 1905
1834/* 1906/*
1835 * Return true if we've flushed caps through the given flush_tid. 1907 * Return true if we've flushed caps through the given flush_tid.
1836 */ 1908 */
1837static int caps_are_flushed(struct inode *inode, unsigned tid) 1909static int caps_are_flushed(struct inode *inode, u64 flush_tid)
1838{ 1910{
1839 struct ceph_inode_info *ci = ceph_inode(inode); 1911 struct ceph_inode_info *ci = ceph_inode(inode);
1840 int i, ret = 1; 1912 struct ceph_cap_flush *cf;
1913 struct rb_node *n;
1914 int ret = 1;
1841 1915
1842 spin_lock(&ci->i_ceph_lock); 1916 spin_lock(&ci->i_ceph_lock);
1843 for (i = 0; i < CEPH_CAP_BITS; i++) 1917 n = rb_first(&ci->i_cap_flush_tree);
1844 if ((ci->i_flushing_caps & (1 << i)) && 1918 if (n) {
1845 ci->i_cap_flush_tid[i] <= tid) { 1919 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1846 /* still flushing this bit */ 1920 if (cf->tid <= flush_tid)
1847 ret = 0; 1921 ret = 0;
1848 break; 1922 }
1849 }
1850 spin_unlock(&ci->i_ceph_lock); 1923 spin_unlock(&ci->i_ceph_lock);
1851 return ret; 1924 return ret;
1852} 1925}
@@ -1864,13 +1937,16 @@ static void sync_write_wait(struct inode *inode)
1864 struct ceph_osd_request *req; 1937 struct ceph_osd_request *req;
1865 u64 last_tid; 1938 u64 last_tid;
1866 1939
1940 if (!S_ISREG(inode->i_mode))
1941 return;
1942
1867 spin_lock(&ci->i_unsafe_lock); 1943 spin_lock(&ci->i_unsafe_lock);
1868 if (list_empty(head)) 1944 if (list_empty(head))
1869 goto out; 1945 goto out;
1870 1946
1871 /* set upper bound as _last_ entry in chain */ 1947 /* set upper bound as _last_ entry in chain */
1872 req = list_entry(head->prev, struct ceph_osd_request, 1948 req = list_last_entry(head, struct ceph_osd_request,
1873 r_unsafe_item); 1949 r_unsafe_item);
1874 last_tid = req->r_tid; 1950 last_tid = req->r_tid;
1875 1951
1876 do { 1952 do {
@@ -1888,18 +1964,64 @@ static void sync_write_wait(struct inode *inode)
1888 */ 1964 */
1889 if (list_empty(head)) 1965 if (list_empty(head))
1890 break; 1966 break;
1891 req = list_entry(head->next, struct ceph_osd_request, 1967 req = list_first_entry(head, struct ceph_osd_request,
1892 r_unsafe_item); 1968 r_unsafe_item);
1893 } while (req->r_tid < last_tid); 1969 } while (req->r_tid < last_tid);
1894out: 1970out:
1895 spin_unlock(&ci->i_unsafe_lock); 1971 spin_unlock(&ci->i_unsafe_lock);
1896} 1972}
1897 1973
1974/*
1975 * wait for any uncommitted directory operations to commit.
1976 */
1977static int unsafe_dirop_wait(struct inode *inode)
1978{
1979 struct ceph_inode_info *ci = ceph_inode(inode);
1980 struct list_head *head = &ci->i_unsafe_dirops;
1981 struct ceph_mds_request *req;
1982 u64 last_tid;
1983 int ret = 0;
1984
1985 if (!S_ISDIR(inode->i_mode))
1986 return 0;
1987
1988 spin_lock(&ci->i_unsafe_lock);
1989 if (list_empty(head))
1990 goto out;
1991
1992 req = list_last_entry(head, struct ceph_mds_request,
1993 r_unsafe_dir_item);
1994 last_tid = req->r_tid;
1995
1996 do {
1997 ceph_mdsc_get_request(req);
1998 spin_unlock(&ci->i_unsafe_lock);
1999
2000 dout("unsafe_dirop_wait %p wait on tid %llu (until %llu)\n",
2001 inode, req->r_tid, last_tid);
2002 ret = !wait_for_completion_timeout(&req->r_safe_completion,
2003 ceph_timeout_jiffies(req->r_timeout));
2004 if (ret)
2005 ret = -EIO; /* timed out */
2006
2007 ceph_mdsc_put_request(req);
2008
2009 spin_lock(&ci->i_unsafe_lock);
2010 if (ret || list_empty(head))
2011 break;
2012 req = list_first_entry(head, struct ceph_mds_request,
2013 r_unsafe_dir_item);
2014 } while (req->r_tid < last_tid);
2015out:
2016 spin_unlock(&ci->i_unsafe_lock);
2017 return ret;
2018}
2019
1898int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync) 2020int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1899{ 2021{
1900 struct inode *inode = file->f_mapping->host; 2022 struct inode *inode = file->f_mapping->host;
1901 struct ceph_inode_info *ci = ceph_inode(inode); 2023 struct ceph_inode_info *ci = ceph_inode(inode);
1902 unsigned flush_tid; 2024 u64 flush_tid;
1903 int ret; 2025 int ret;
1904 int dirty; 2026 int dirty;
1905 2027
@@ -1908,25 +2030,30 @@ int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1908 2030
1909 ret = filemap_write_and_wait_range(inode->i_mapping, start, end); 2031 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1910 if (ret < 0) 2032 if (ret < 0)
1911 return ret; 2033 goto out;
2034
2035 if (datasync)
2036 goto out;
2037
1912 mutex_lock(&inode->i_mutex); 2038 mutex_lock(&inode->i_mutex);
1913 2039
1914 dirty = try_flush_caps(inode, &flush_tid); 2040 dirty = try_flush_caps(inode, &flush_tid);
1915 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty)); 2041 dout("fsync dirty caps are %s\n", ceph_cap_string(dirty));
1916 2042
2043 ret = unsafe_dirop_wait(inode);
2044
1917 /* 2045 /*
1918 * only wait on non-file metadata writeback (the mds 2046 * only wait on non-file metadata writeback (the mds
1919 * can recover size and mtime, so we don't need to 2047 * can recover size and mtime, so we don't need to
1920 * wait for that) 2048 * wait for that)
1921 */ 2049 */
1922 if (!datasync && (dirty & ~CEPH_CAP_ANY_FILE_WR)) { 2050 if (!ret && (dirty & ~CEPH_CAP_ANY_FILE_WR)) {
1923 dout("fsync waiting for flush_tid %u\n", flush_tid);
1924 ret = wait_event_interruptible(ci->i_cap_wq, 2051 ret = wait_event_interruptible(ci->i_cap_wq,
1925 caps_are_flushed(inode, flush_tid)); 2052 caps_are_flushed(inode, flush_tid));
1926 } 2053 }
1927
1928 dout("fsync %p%s done\n", inode, datasync ? " datasync" : "");
1929 mutex_unlock(&inode->i_mutex); 2054 mutex_unlock(&inode->i_mutex);
2055out:
2056 dout("fsync %p%s result=%d\n", inode, datasync ? " datasync" : "", ret);
1930 return ret; 2057 return ret;
1931} 2058}
1932 2059
@@ -1939,7 +2066,7 @@ int ceph_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1939int ceph_write_inode(struct inode *inode, struct writeback_control *wbc) 2066int ceph_write_inode(struct inode *inode, struct writeback_control *wbc)
1940{ 2067{
1941 struct ceph_inode_info *ci = ceph_inode(inode); 2068 struct ceph_inode_info *ci = ceph_inode(inode);
1942 unsigned flush_tid; 2069 u64 flush_tid;
1943 int err = 0; 2070 int err = 0;
1944 int dirty; 2071 int dirty;
1945 int wait = wbc->sync_mode == WB_SYNC_ALL; 2072 int wait = wbc->sync_mode == WB_SYNC_ALL;
@@ -1994,6 +2121,104 @@ static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
1994 } 2121 }
1995} 2122}
1996 2123
2124static int __kick_flushing_caps(struct ceph_mds_client *mdsc,
2125 struct ceph_mds_session *session,
2126 struct ceph_inode_info *ci,
2127 bool kick_all)
2128{
2129 struct inode *inode = &ci->vfs_inode;
2130 struct ceph_cap *cap;
2131 struct ceph_cap_flush *cf;
2132 struct rb_node *n;
2133 int delayed = 0;
2134 u64 first_tid = 0;
2135 u64 oldest_flush_tid;
2136
2137 spin_lock(&mdsc->cap_dirty_lock);
2138 oldest_flush_tid = __get_oldest_flush_tid(mdsc);
2139 spin_unlock(&mdsc->cap_dirty_lock);
2140
2141 while (true) {
2142 spin_lock(&ci->i_ceph_lock);
2143 cap = ci->i_auth_cap;
2144 if (!(cap && cap->session == session)) {
2145 pr_err("%p auth cap %p not mds%d ???\n", inode,
2146 cap, session->s_mds);
2147 spin_unlock(&ci->i_ceph_lock);
2148 break;
2149 }
2150
2151 for (n = rb_first(&ci->i_cap_flush_tree); n; n = rb_next(n)) {
2152 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2153 if (cf->tid < first_tid)
2154 continue;
2155 if (kick_all || cf->kick)
2156 break;
2157 }
2158 if (!n) {
2159 spin_unlock(&ci->i_ceph_lock);
2160 break;
2161 }
2162
2163 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2164 cf->kick = false;
2165
2166 first_tid = cf->tid + 1;
2167
2168 dout("kick_flushing_caps %p cap %p tid %llu %s\n", inode,
2169 cap, cf->tid, ceph_cap_string(cf->caps));
2170 delayed |= __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2171 __ceph_caps_used(ci),
2172 __ceph_caps_wanted(ci),
2173 cap->issued | cap->implemented,
2174 cf->caps, cf->tid, oldest_flush_tid);
2175 }
2176 return delayed;
2177}
2178
2179void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
2180 struct ceph_mds_session *session)
2181{
2182 struct ceph_inode_info *ci;
2183 struct ceph_cap *cap;
2184 struct ceph_cap_flush *cf;
2185 struct rb_node *n;
2186
2187 dout("early_kick_flushing_caps mds%d\n", session->s_mds);
2188 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2189 spin_lock(&ci->i_ceph_lock);
2190 cap = ci->i_auth_cap;
2191 if (!(cap && cap->session == session)) {
2192 pr_err("%p auth cap %p not mds%d ???\n",
2193 &ci->vfs_inode, cap, session->s_mds);
2194 spin_unlock(&ci->i_ceph_lock);
2195 continue;
2196 }
2197
2198
2199 /*
2200 * if flushing caps were revoked, we re-send the cap flush
2201 * in client reconnect stage. This guarantees MDS * processes
2202 * the cap flush message before issuing the flushing caps to
2203 * other client.
2204 */
2205 if ((cap->issued & ci->i_flushing_caps) !=
2206 ci->i_flushing_caps) {
2207 spin_unlock(&ci->i_ceph_lock);
2208 if (!__kick_flushing_caps(mdsc, session, ci, true))
2209 continue;
2210 spin_lock(&ci->i_ceph_lock);
2211 }
2212
2213 for (n = rb_first(&ci->i_cap_flush_tree); n; n = rb_next(n)) {
2214 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2215 cf->kick = true;
2216 }
2217
2218 spin_unlock(&ci->i_ceph_lock);
2219 }
2220}
2221
1997void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 2222void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
1998 struct ceph_mds_session *session) 2223 struct ceph_mds_session *session)
1999{ 2224{
@@ -2003,28 +2228,10 @@ void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
2003 2228
2004 dout("kick_flushing_caps mds%d\n", session->s_mds); 2229 dout("kick_flushing_caps mds%d\n", session->s_mds);
2005 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) { 2230 list_for_each_entry(ci, &session->s_cap_flushing, i_flushing_item) {
2006 struct inode *inode = &ci->vfs_inode; 2231 int delayed = __kick_flushing_caps(mdsc, session, ci, false);
2007 struct ceph_cap *cap; 2232 if (delayed) {
2008 int delayed = 0; 2233 spin_lock(&ci->i_ceph_lock);
2009 2234 __cap_delay_requeue(mdsc, ci);
2010 spin_lock(&ci->i_ceph_lock);
2011 cap = ci->i_auth_cap;
2012 if (cap && cap->session == session) {
2013 dout("kick_flushing_caps %p cap %p %s\n", inode,
2014 cap, ceph_cap_string(ci->i_flushing_caps));
2015 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH,
2016 __ceph_caps_used(ci),
2017 __ceph_caps_wanted(ci),
2018 cap->issued | cap->implemented,
2019 ci->i_flushing_caps, NULL);
2020 if (delayed) {
2021 spin_lock(&ci->i_ceph_lock);
2022 __cap_delay_requeue(mdsc, ci);
2023 spin_unlock(&ci->i_ceph_lock);
2024 }
2025 } else {
2026 pr_err("%p auth cap %p not mds%d ???\n", inode,
2027 cap, session->s_mds);
2028 spin_unlock(&ci->i_ceph_lock); 2235 spin_unlock(&ci->i_ceph_lock);
2029 } 2236 }
2030 } 2237 }
@@ -2036,26 +2243,25 @@ static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2036{ 2243{
2037 struct ceph_inode_info *ci = ceph_inode(inode); 2244 struct ceph_inode_info *ci = ceph_inode(inode);
2038 struct ceph_cap *cap; 2245 struct ceph_cap *cap;
2039 int delayed = 0;
2040 2246
2041 spin_lock(&ci->i_ceph_lock); 2247 spin_lock(&ci->i_ceph_lock);
2042 cap = ci->i_auth_cap; 2248 cap = ci->i_auth_cap;
2043 dout("kick_flushing_inode_caps %p flushing %s flush_seq %lld\n", inode, 2249 dout("kick_flushing_inode_caps %p flushing %s\n", inode,
2044 ceph_cap_string(ci->i_flushing_caps), ci->i_cap_flush_seq); 2250 ceph_cap_string(ci->i_flushing_caps));
2045 2251
2046 __ceph_flush_snaps(ci, &session, 1); 2252 __ceph_flush_snaps(ci, &session, 1);
2047 2253
2048 if (ci->i_flushing_caps) { 2254 if (ci->i_flushing_caps) {
2255 int delayed;
2256
2049 spin_lock(&mdsc->cap_dirty_lock); 2257 spin_lock(&mdsc->cap_dirty_lock);
2050 list_move_tail(&ci->i_flushing_item, 2258 list_move_tail(&ci->i_flushing_item,
2051 &cap->session->s_cap_flushing); 2259 &cap->session->s_cap_flushing);
2052 spin_unlock(&mdsc->cap_dirty_lock); 2260 spin_unlock(&mdsc->cap_dirty_lock);
2053 2261
2054 delayed = __send_cap(mdsc, cap, CEPH_CAP_OP_FLUSH, 2262 spin_unlock(&ci->i_ceph_lock);
2055 __ceph_caps_used(ci), 2263
2056 __ceph_caps_wanted(ci), 2264 delayed = __kick_flushing_caps(mdsc, session, ci, true);
2057 cap->issued | cap->implemented,
2058 ci->i_flushing_caps, NULL);
2059 if (delayed) { 2265 if (delayed) {
2060 spin_lock(&ci->i_ceph_lock); 2266 spin_lock(&ci->i_ceph_lock);
2061 __cap_delay_requeue(mdsc, ci); 2267 __cap_delay_requeue(mdsc, ci);
@@ -2073,7 +2279,8 @@ static void kick_flushing_inode_caps(struct ceph_mds_client *mdsc,
2073 * 2279 *
2074 * Protected by i_ceph_lock. 2280 * Protected by i_ceph_lock.
2075 */ 2281 */
2076static void __take_cap_refs(struct ceph_inode_info *ci, int got) 2282static void __take_cap_refs(struct ceph_inode_info *ci, int got,
2283 bool snap_rwsem_locked)
2077{ 2284{
2078 if (got & CEPH_CAP_PIN) 2285 if (got & CEPH_CAP_PIN)
2079 ci->i_pin_ref++; 2286 ci->i_pin_ref++;
@@ -2081,8 +2288,14 @@ static void __take_cap_refs(struct ceph_inode_info *ci, int got)
2081 ci->i_rd_ref++; 2288 ci->i_rd_ref++;
2082 if (got & CEPH_CAP_FILE_CACHE) 2289 if (got & CEPH_CAP_FILE_CACHE)
2083 ci->i_rdcache_ref++; 2290 ci->i_rdcache_ref++;
2084 if (got & CEPH_CAP_FILE_WR) 2291 if (got & CEPH_CAP_FILE_WR) {
2292 if (ci->i_wr_ref == 0 && !ci->i_head_snapc) {
2293 BUG_ON(!snap_rwsem_locked);
2294 ci->i_head_snapc = ceph_get_snap_context(
2295 ci->i_snap_realm->cached_context);
2296 }
2085 ci->i_wr_ref++; 2297 ci->i_wr_ref++;
2298 }
2086 if (got & CEPH_CAP_FILE_BUFFER) { 2299 if (got & CEPH_CAP_FILE_BUFFER) {
2087 if (ci->i_wb_ref == 0) 2300 if (ci->i_wb_ref == 0)
2088 ihold(&ci->vfs_inode); 2301 ihold(&ci->vfs_inode);
@@ -2100,16 +2313,19 @@ static void __take_cap_refs(struct ceph_inode_info *ci, int got)
2100 * requested from the MDS. 2313 * requested from the MDS.
2101 */ 2314 */
2102static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want, 2315static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2103 loff_t endoff, int *got, int *check_max, int *err) 2316 loff_t endoff, bool nonblock, int *got, int *err)
2104{ 2317{
2105 struct inode *inode = &ci->vfs_inode; 2318 struct inode *inode = &ci->vfs_inode;
2319 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
2106 int ret = 0; 2320 int ret = 0;
2107 int have, implemented; 2321 int have, implemented;
2108 int file_wanted; 2322 int file_wanted;
2323 bool snap_rwsem_locked = false;
2109 2324
2110 dout("get_cap_refs %p need %s want %s\n", inode, 2325 dout("get_cap_refs %p need %s want %s\n", inode,
2111 ceph_cap_string(need), ceph_cap_string(want)); 2326 ceph_cap_string(need), ceph_cap_string(want));
2112 2327
2328again:
2113 spin_lock(&ci->i_ceph_lock); 2329 spin_lock(&ci->i_ceph_lock);
2114 2330
2115 /* make sure file is actually open */ 2331 /* make sure file is actually open */
@@ -2125,6 +2341,10 @@ static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2125 /* finish pending truncate */ 2341 /* finish pending truncate */
2126 while (ci->i_truncate_pending) { 2342 while (ci->i_truncate_pending) {
2127 spin_unlock(&ci->i_ceph_lock); 2343 spin_unlock(&ci->i_ceph_lock);
2344 if (snap_rwsem_locked) {
2345 up_read(&mdsc->snap_rwsem);
2346 snap_rwsem_locked = false;
2347 }
2128 __ceph_do_pending_vmtruncate(inode); 2348 __ceph_do_pending_vmtruncate(inode);
2129 spin_lock(&ci->i_ceph_lock); 2349 spin_lock(&ci->i_ceph_lock);
2130 } 2350 }
@@ -2136,7 +2356,7 @@ static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2136 dout("get_cap_refs %p endoff %llu > maxsize %llu\n", 2356 dout("get_cap_refs %p endoff %llu > maxsize %llu\n",
2137 inode, endoff, ci->i_max_size); 2357 inode, endoff, ci->i_max_size);
2138 if (endoff > ci->i_requested_max_size) { 2358 if (endoff > ci->i_requested_max_size) {
2139 *check_max = 1; 2359 *err = -EAGAIN;
2140 ret = 1; 2360 ret = 1;
2141 } 2361 }
2142 goto out_unlock; 2362 goto out_unlock;
@@ -2164,8 +2384,29 @@ static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2164 inode, ceph_cap_string(have), ceph_cap_string(not), 2384 inode, ceph_cap_string(have), ceph_cap_string(not),
2165 ceph_cap_string(revoking)); 2385 ceph_cap_string(revoking));
2166 if ((revoking & not) == 0) { 2386 if ((revoking & not) == 0) {
2387 if (!snap_rwsem_locked &&
2388 !ci->i_head_snapc &&
2389 (need & CEPH_CAP_FILE_WR)) {
2390 if (!down_read_trylock(&mdsc->snap_rwsem)) {
2391 /*
2392 * we can not call down_read() when
2393 * task isn't in TASK_RUNNING state
2394 */
2395 if (nonblock) {
2396 *err = -EAGAIN;
2397 ret = 1;
2398 goto out_unlock;
2399 }
2400
2401 spin_unlock(&ci->i_ceph_lock);
2402 down_read(&mdsc->snap_rwsem);
2403 snap_rwsem_locked = true;
2404 goto again;
2405 }
2406 snap_rwsem_locked = true;
2407 }
2167 *got = need | (have & want); 2408 *got = need | (have & want);
2168 __take_cap_refs(ci, *got); 2409 __take_cap_refs(ci, *got, true);
2169 ret = 1; 2410 ret = 1;
2170 } 2411 }
2171 } else { 2412 } else {
@@ -2189,6 +2430,8 @@ static int try_get_cap_refs(struct ceph_inode_info *ci, int need, int want,
2189 } 2430 }
2190out_unlock: 2431out_unlock:
2191 spin_unlock(&ci->i_ceph_lock); 2432 spin_unlock(&ci->i_ceph_lock);
2433 if (snap_rwsem_locked)
2434 up_read(&mdsc->snap_rwsem);
2192 2435
2193 dout("get_cap_refs %p ret %d got %s\n", inode, 2436 dout("get_cap_refs %p ret %d got %s\n", inode,
2194 ret, ceph_cap_string(*got)); 2437 ret, ceph_cap_string(*got));
@@ -2231,50 +2474,70 @@ static void check_max_size(struct inode *inode, loff_t endoff)
2231int ceph_get_caps(struct ceph_inode_info *ci, int need, int want, 2474int ceph_get_caps(struct ceph_inode_info *ci, int need, int want,
2232 loff_t endoff, int *got, struct page **pinned_page) 2475 loff_t endoff, int *got, struct page **pinned_page)
2233{ 2476{
2234 int _got, check_max, ret, err = 0; 2477 int _got, ret, err = 0;
2235 2478
2236retry: 2479 ret = ceph_pool_perm_check(ci, need);
2237 if (endoff > 0)
2238 check_max_size(&ci->vfs_inode, endoff);
2239 _got = 0;
2240 check_max = 0;
2241 ret = wait_event_interruptible(ci->i_cap_wq,
2242 try_get_cap_refs(ci, need, want, endoff,
2243 &_got, &check_max, &err));
2244 if (err)
2245 ret = err;
2246 if (ret < 0) 2480 if (ret < 0)
2247 return ret; 2481 return ret;
2248 2482
2249 if (check_max) 2483 while (true) {
2250 goto retry; 2484 if (endoff > 0)
2485 check_max_size(&ci->vfs_inode, endoff);
2251 2486
2252 if (ci->i_inline_version != CEPH_INLINE_NONE && 2487 err = 0;
2253 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2254 i_size_read(&ci->vfs_inode) > 0) {
2255 struct page *page = find_get_page(ci->vfs_inode.i_mapping, 0);
2256 if (page) {
2257 if (PageUptodate(page)) {
2258 *pinned_page = page;
2259 goto out;
2260 }
2261 page_cache_release(page);
2262 }
2263 /*
2264 * drop cap refs first because getattr while holding
2265 * caps refs can cause deadlock.
2266 */
2267 ceph_put_cap_refs(ci, _got);
2268 _got = 0; 2488 _got = 0;
2489 ret = try_get_cap_refs(ci, need, want, endoff,
2490 false, &_got, &err);
2491 if (ret) {
2492 if (err == -EAGAIN)
2493 continue;
2494 if (err < 0)
2495 return err;
2496 } else {
2497 ret = wait_event_interruptible(ci->i_cap_wq,
2498 try_get_cap_refs(ci, need, want, endoff,
2499 true, &_got, &err));
2500 if (err == -EAGAIN)
2501 continue;
2502 if (err < 0)
2503 ret = err;
2504 if (ret < 0)
2505 return ret;
2506 }
2269 2507
2270 /* getattr request will bring inline data into page cache */ 2508 if (ci->i_inline_version != CEPH_INLINE_NONE &&
2271 ret = __ceph_do_getattr(&ci->vfs_inode, NULL, 2509 (_got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) &&
2272 CEPH_STAT_CAP_INLINE_DATA, true); 2510 i_size_read(&ci->vfs_inode) > 0) {
2273 if (ret < 0) 2511 struct page *page =
2274 return ret; 2512 find_get_page(ci->vfs_inode.i_mapping, 0);
2275 goto retry; 2513 if (page) {
2514 if (PageUptodate(page)) {
2515 *pinned_page = page;
2516 break;
2517 }
2518 page_cache_release(page);
2519 }
2520 /*
2521 * drop cap refs first because getattr while
2522 * holding * caps refs can cause deadlock.
2523 */
2524 ceph_put_cap_refs(ci, _got);
2525 _got = 0;
2526
2527 /*
2528 * getattr request will bring inline data into
2529 * page cache
2530 */
2531 ret = __ceph_do_getattr(&ci->vfs_inode, NULL,
2532 CEPH_STAT_CAP_INLINE_DATA,
2533 true);
2534 if (ret < 0)
2535 return ret;
2536 continue;
2537 }
2538 break;
2276 } 2539 }
2277out: 2540
2278 *got = _got; 2541 *got = _got;
2279 return 0; 2542 return 0;
2280} 2543}
@@ -2286,10 +2549,31 @@ out:
2286void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps) 2549void ceph_get_cap_refs(struct ceph_inode_info *ci, int caps)
2287{ 2550{
2288 spin_lock(&ci->i_ceph_lock); 2551 spin_lock(&ci->i_ceph_lock);
2289 __take_cap_refs(ci, caps); 2552 __take_cap_refs(ci, caps, false);
2290 spin_unlock(&ci->i_ceph_lock); 2553 spin_unlock(&ci->i_ceph_lock);
2291} 2554}
2292 2555
2556
2557/*
2558 * drop cap_snap that is not associated with any snapshot.
2559 * we don't need to send FLUSHSNAP message for it.
2560 */
2561static int ceph_try_drop_cap_snap(struct ceph_cap_snap *capsnap)
2562{
2563 if (!capsnap->need_flush &&
2564 !capsnap->writing && !capsnap->dirty_pages) {
2565
2566 dout("dropping cap_snap %p follows %llu\n",
2567 capsnap, capsnap->follows);
2568 ceph_put_snap_context(capsnap->context);
2569 list_del(&capsnap->ci_item);
2570 list_del(&capsnap->flushing_item);
2571 ceph_put_cap_snap(capsnap);
2572 return 1;
2573 }
2574 return 0;
2575}
2576
2293/* 2577/*
2294 * Release cap refs. 2578 * Release cap refs.
2295 * 2579 *
@@ -2303,7 +2587,6 @@ void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2303{ 2587{
2304 struct inode *inode = &ci->vfs_inode; 2588 struct inode *inode = &ci->vfs_inode;
2305 int last = 0, put = 0, flushsnaps = 0, wake = 0; 2589 int last = 0, put = 0, flushsnaps = 0, wake = 0;
2306 struct ceph_cap_snap *capsnap;
2307 2590
2308 spin_lock(&ci->i_ceph_lock); 2591 spin_lock(&ci->i_ceph_lock);
2309 if (had & CEPH_CAP_PIN) 2592 if (had & CEPH_CAP_PIN)
@@ -2325,17 +2608,24 @@ void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2325 if (had & CEPH_CAP_FILE_WR) 2608 if (had & CEPH_CAP_FILE_WR)
2326 if (--ci->i_wr_ref == 0) { 2609 if (--ci->i_wr_ref == 0) {
2327 last++; 2610 last++;
2328 if (!list_empty(&ci->i_cap_snaps)) { 2611 if (__ceph_have_pending_cap_snap(ci)) {
2329 capsnap = list_first_entry(&ci->i_cap_snaps, 2612 struct ceph_cap_snap *capsnap =
2330 struct ceph_cap_snap, 2613 list_last_entry(&ci->i_cap_snaps,
2331 ci_item); 2614 struct ceph_cap_snap,
2332 if (capsnap->writing) { 2615 ci_item);
2333 capsnap->writing = 0; 2616 capsnap->writing = 0;
2334 flushsnaps = 2617 if (ceph_try_drop_cap_snap(capsnap))
2335 __ceph_finish_cap_snap(ci, 2618 put++;
2336 capsnap); 2619 else if (__ceph_finish_cap_snap(ci, capsnap))
2337 wake = 1; 2620 flushsnaps = 1;
2338 } 2621 wake = 1;
2622 }
2623 if (ci->i_wrbuffer_ref_head == 0 &&
2624 ci->i_dirty_caps == 0 &&
2625 ci->i_flushing_caps == 0) {
2626 BUG_ON(!ci->i_head_snapc);
2627 ceph_put_snap_context(ci->i_head_snapc);
2628 ci->i_head_snapc = NULL;
2339 } 2629 }
2340 /* see comment in __ceph_remove_cap() */ 2630 /* see comment in __ceph_remove_cap() */
2341 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm) 2631 if (!__ceph_is_any_caps(ci) && ci->i_snap_realm)
@@ -2352,7 +2642,7 @@ void ceph_put_cap_refs(struct ceph_inode_info *ci, int had)
2352 ceph_flush_snaps(ci); 2642 ceph_flush_snaps(ci);
2353 if (wake) 2643 if (wake)
2354 wake_up_all(&ci->i_cap_wq); 2644 wake_up_all(&ci->i_cap_wq);
2355 if (put) 2645 while (put-- > 0)
2356 iput(inode); 2646 iput(inode);
2357} 2647}
2358 2648
@@ -2380,7 +2670,9 @@ void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2380 if (ci->i_head_snapc == snapc) { 2670 if (ci->i_head_snapc == snapc) {
2381 ci->i_wrbuffer_ref_head -= nr; 2671 ci->i_wrbuffer_ref_head -= nr;
2382 if (ci->i_wrbuffer_ref_head == 0 && 2672 if (ci->i_wrbuffer_ref_head == 0 &&
2383 ci->i_dirty_caps == 0 && ci->i_flushing_caps == 0) { 2673 ci->i_wr_ref == 0 &&
2674 ci->i_dirty_caps == 0 &&
2675 ci->i_flushing_caps == 0) {
2384 BUG_ON(!ci->i_head_snapc); 2676 BUG_ON(!ci->i_head_snapc);
2385 ceph_put_snap_context(ci->i_head_snapc); 2677 ceph_put_snap_context(ci->i_head_snapc);
2386 ci->i_head_snapc = NULL; 2678 ci->i_head_snapc = NULL;
@@ -2401,25 +2693,15 @@ void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
2401 capsnap->dirty_pages -= nr; 2693 capsnap->dirty_pages -= nr;
2402 if (capsnap->dirty_pages == 0) { 2694 if (capsnap->dirty_pages == 0) {
2403 complete_capsnap = 1; 2695 complete_capsnap = 1;
2404 if (capsnap->dirty == 0) 2696 drop_capsnap = ceph_try_drop_cap_snap(capsnap);
2405 /* cap writeback completed before we created
2406 * the cap_snap; no FLUSHSNAP is needed */
2407 drop_capsnap = 1;
2408 } 2697 }
2409 dout("put_wrbuffer_cap_refs on %p cap_snap %p " 2698 dout("put_wrbuffer_cap_refs on %p cap_snap %p "
2410 " snap %lld %d/%d -> %d/%d %s%s%s\n", 2699 " snap %lld %d/%d -> %d/%d %s%s\n",
2411 inode, capsnap, capsnap->context->seq, 2700 inode, capsnap, capsnap->context->seq,
2412 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr, 2701 ci->i_wrbuffer_ref+nr, capsnap->dirty_pages + nr,
2413 ci->i_wrbuffer_ref, capsnap->dirty_pages, 2702 ci->i_wrbuffer_ref, capsnap->dirty_pages,
2414 last ? " (wrbuffer last)" : "", 2703 last ? " (wrbuffer last)" : "",
2415 complete_capsnap ? " (complete capsnap)" : "", 2704 complete_capsnap ? " (complete capsnap)" : "");
2416 drop_capsnap ? " (drop capsnap)" : "");
2417 if (drop_capsnap) {
2418 ceph_put_snap_context(capsnap->context);
2419 list_del(&capsnap->ci_item);
2420 list_del(&capsnap->flushing_item);
2421 ceph_put_cap_snap(capsnap);
2422 }
2423 } 2705 }
2424 2706
2425 spin_unlock(&ci->i_ceph_lock); 2707 spin_unlock(&ci->i_ceph_lock);
@@ -2526,7 +2808,8 @@ static void handle_cap_grant(struct ceph_mds_client *mdsc,
2526 * try to invalidate (once). (If there are dirty buffers, we 2808 * try to invalidate (once). (If there are dirty buffers, we
2527 * will invalidate _after_ writeback.) 2809 * will invalidate _after_ writeback.)
2528 */ 2810 */
2529 if (((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) && 2811 if (!S_ISDIR(inode->i_mode) && /* don't invalidate readdir cache */
2812 ((cap->issued & ~newcaps) & CEPH_CAP_FILE_CACHE) &&
2530 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 && 2813 (newcaps & CEPH_CAP_FILE_LAZYIO) == 0 &&
2531 !ci->i_wrbuffer_ref) { 2814 !ci->i_wrbuffer_ref) {
2532 if (try_nonblocking_invalidate(inode)) { 2815 if (try_nonblocking_invalidate(inode)) {
@@ -2732,16 +3015,29 @@ static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2732{ 3015{
2733 struct ceph_inode_info *ci = ceph_inode(inode); 3016 struct ceph_inode_info *ci = ceph_inode(inode);
2734 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc; 3017 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
3018 struct ceph_cap_flush *cf;
3019 struct rb_node *n;
3020 LIST_HEAD(to_remove);
2735 unsigned seq = le32_to_cpu(m->seq); 3021 unsigned seq = le32_to_cpu(m->seq);
2736 int dirty = le32_to_cpu(m->dirty); 3022 int dirty = le32_to_cpu(m->dirty);
2737 int cleaned = 0; 3023 int cleaned = 0;
2738 int drop = 0; 3024 int drop = 0;
2739 int i;
2740 3025
2741 for (i = 0; i < CEPH_CAP_BITS; i++) 3026 n = rb_first(&ci->i_cap_flush_tree);
2742 if ((dirty & (1 << i)) && 3027 while (n) {
2743 (u16)flush_tid == ci->i_cap_flush_tid[i]) 3028 cf = rb_entry(n, struct ceph_cap_flush, i_node);
2744 cleaned |= 1 << i; 3029 n = rb_next(&cf->i_node);
3030 if (cf->tid == flush_tid)
3031 cleaned = cf->caps;
3032 if (cf->tid <= flush_tid) {
3033 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
3034 list_add_tail(&cf->list, &to_remove);
3035 } else {
3036 cleaned &= ~cf->caps;
3037 if (!cleaned)
3038 break;
3039 }
3040 }
2745 3041
2746 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s," 3042 dout("handle_cap_flush_ack inode %p mds%d seq %d on %s cleaned %s,"
2747 " flushing %s -> %s\n", 3043 " flushing %s -> %s\n",
@@ -2749,12 +3045,23 @@ static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2749 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps), 3045 ceph_cap_string(cleaned), ceph_cap_string(ci->i_flushing_caps),
2750 ceph_cap_string(ci->i_flushing_caps & ~cleaned)); 3046 ceph_cap_string(ci->i_flushing_caps & ~cleaned));
2751 3047
2752 if (ci->i_flushing_caps == (ci->i_flushing_caps & ~cleaned)) 3048 if (list_empty(&to_remove) && !cleaned)
2753 goto out; 3049 goto out;
2754 3050
2755 ci->i_flushing_caps &= ~cleaned; 3051 ci->i_flushing_caps &= ~cleaned;
2756 3052
2757 spin_lock(&mdsc->cap_dirty_lock); 3053 spin_lock(&mdsc->cap_dirty_lock);
3054
3055 if (!list_empty(&to_remove)) {
3056 list_for_each_entry(cf, &to_remove, list)
3057 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
3058
3059 n = rb_first(&mdsc->cap_flush_tree);
3060 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
3061 if (!cf || cf->tid > flush_tid)
3062 wake_up_all(&mdsc->cap_flushing_wq);
3063 }
3064
2758 if (ci->i_flushing_caps == 0) { 3065 if (ci->i_flushing_caps == 0) {
2759 list_del_init(&ci->i_flushing_item); 3066 list_del_init(&ci->i_flushing_item);
2760 if (!list_empty(&session->s_cap_flushing)) 3067 if (!list_empty(&session->s_cap_flushing))
@@ -2764,14 +3071,14 @@ static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2764 struct ceph_inode_info, 3071 struct ceph_inode_info,
2765 i_flushing_item)->vfs_inode); 3072 i_flushing_item)->vfs_inode);
2766 mdsc->num_cap_flushing--; 3073 mdsc->num_cap_flushing--;
2767 wake_up_all(&mdsc->cap_flushing_wq);
2768 dout(" inode %p now !flushing\n", inode); 3074 dout(" inode %p now !flushing\n", inode);
2769 3075
2770 if (ci->i_dirty_caps == 0) { 3076 if (ci->i_dirty_caps == 0) {
2771 dout(" inode %p now clean\n", inode); 3077 dout(" inode %p now clean\n", inode);
2772 BUG_ON(!list_empty(&ci->i_dirty_item)); 3078 BUG_ON(!list_empty(&ci->i_dirty_item));
2773 drop = 1; 3079 drop = 1;
2774 if (ci->i_wrbuffer_ref_head == 0) { 3080 if (ci->i_wr_ref == 0 &&
3081 ci->i_wrbuffer_ref_head == 0) {
2775 BUG_ON(!ci->i_head_snapc); 3082 BUG_ON(!ci->i_head_snapc);
2776 ceph_put_snap_context(ci->i_head_snapc); 3083 ceph_put_snap_context(ci->i_head_snapc);
2777 ci->i_head_snapc = NULL; 3084 ci->i_head_snapc = NULL;
@@ -2785,6 +3092,13 @@ static void handle_cap_flush_ack(struct inode *inode, u64 flush_tid,
2785 3092
2786out: 3093out:
2787 spin_unlock(&ci->i_ceph_lock); 3094 spin_unlock(&ci->i_ceph_lock);
3095
3096 while (!list_empty(&to_remove)) {
3097 cf = list_first_entry(&to_remove,
3098 struct ceph_cap_flush, list);
3099 list_del(&cf->list);
3100 ceph_free_cap_flush(cf);
3101 }
2788 if (drop) 3102 if (drop)
2789 iput(inode); 3103 iput(inode);
2790} 3104}
@@ -2800,6 +3114,7 @@ static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2800 struct ceph_mds_session *session) 3114 struct ceph_mds_session *session)
2801{ 3115{
2802 struct ceph_inode_info *ci = ceph_inode(inode); 3116 struct ceph_inode_info *ci = ceph_inode(inode);
3117 struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
2803 u64 follows = le64_to_cpu(m->snap_follows); 3118 u64 follows = le64_to_cpu(m->snap_follows);
2804 struct ceph_cap_snap *capsnap; 3119 struct ceph_cap_snap *capsnap;
2805 int drop = 0; 3120 int drop = 0;
@@ -2823,6 +3138,7 @@ static void handle_cap_flushsnap_ack(struct inode *inode, u64 flush_tid,
2823 list_del(&capsnap->ci_item); 3138 list_del(&capsnap->ci_item);
2824 list_del(&capsnap->flushing_item); 3139 list_del(&capsnap->flushing_item);
2825 ceph_put_cap_snap(capsnap); 3140 ceph_put_cap_snap(capsnap);
3141 wake_up_all(&mdsc->cap_flushing_wq);
2826 drop = 1; 3142 drop = 1;
2827 break; 3143 break;
2828 } else { 3144 } else {
@@ -2971,7 +3287,6 @@ retry:
2971 mutex_lock_nested(&session->s_mutex, 3287 mutex_lock_nested(&session->s_mutex,
2972 SINGLE_DEPTH_NESTING); 3288 SINGLE_DEPTH_NESTING);
2973 } 3289 }
2974 ceph_add_cap_releases(mdsc, tsession);
2975 new_cap = ceph_get_cap(mdsc, NULL); 3290 new_cap = ceph_get_cap(mdsc, NULL);
2976 } else { 3291 } else {
2977 WARN_ON(1); 3292 WARN_ON(1);
@@ -3167,16 +3482,20 @@ void ceph_handle_caps(struct ceph_mds_session *session,
3167 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq, 3482 dout(" mds%d seq %lld cap seq %u\n", session->s_mds, session->s_seq,
3168 (unsigned)seq); 3483 (unsigned)seq);
3169 3484
3170 if (op == CEPH_CAP_OP_IMPORT)
3171 ceph_add_cap_releases(mdsc, session);
3172
3173 if (!inode) { 3485 if (!inode) {
3174 dout(" i don't have ino %llx\n", vino.ino); 3486 dout(" i don't have ino %llx\n", vino.ino);
3175 3487
3176 if (op == CEPH_CAP_OP_IMPORT) { 3488 if (op == CEPH_CAP_OP_IMPORT) {
3489 cap = ceph_get_cap(mdsc, NULL);
3490 cap->cap_ino = vino.ino;
3491 cap->queue_release = 1;
3492 cap->cap_id = cap_id;
3493 cap->mseq = mseq;
3494 cap->seq = seq;
3177 spin_lock(&session->s_cap_lock); 3495 spin_lock(&session->s_cap_lock);
3178 __queue_cap_release(session, vino.ino, cap_id, 3496 list_add_tail(&cap->session_caps,
3179 mseq, seq); 3497 &session->s_cap_releases);
3498 session->s_num_cap_releases++;
3180 spin_unlock(&session->s_cap_lock); 3499 spin_unlock(&session->s_cap_lock);
3181 } 3500 }
3182 goto flush_cap_releases; 3501 goto flush_cap_releases;
@@ -3252,11 +3571,10 @@ void ceph_handle_caps(struct ceph_mds_session *session,
3252 3571
3253flush_cap_releases: 3572flush_cap_releases:
3254 /* 3573 /*
3255 * send any full release message to try to move things 3574 * send any cap release message to try to move things
3256 * along for the mds (who clearly thinks we still have this 3575 * along for the mds (who clearly thinks we still have this
3257 * cap). 3576 * cap).
3258 */ 3577 */
3259 ceph_add_cap_releases(mdsc, session);
3260 ceph_send_cap_releases(mdsc, session); 3578 ceph_send_cap_releases(mdsc, session);
3261 3579
3262done: 3580done:
diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
index 4248307fea90..9314b4ea2375 100644
--- a/fs/ceph/dir.c
+++ b/fs/ceph/dir.c
@@ -38,7 +38,7 @@ int ceph_init_dentry(struct dentry *dentry)
38 if (dentry->d_fsdata) 38 if (dentry->d_fsdata)
39 return 0; 39 return 0;
40 40
41 di = kmem_cache_alloc(ceph_dentry_cachep, GFP_NOFS | __GFP_ZERO); 41 di = kmem_cache_alloc(ceph_dentry_cachep, GFP_KERNEL | __GFP_ZERO);
42 if (!di) 42 if (!di)
43 return -ENOMEM; /* oh well */ 43 return -ENOMEM; /* oh well */
44 44
@@ -107,6 +107,27 @@ static int fpos_cmp(loff_t l, loff_t r)
107} 107}
108 108
109/* 109/*
110 * make note of the last dentry we read, so we can
111 * continue at the same lexicographical point,
112 * regardless of what dir changes take place on the
113 * server.
114 */
115static int note_last_dentry(struct ceph_file_info *fi, const char *name,
116 int len, unsigned next_offset)
117{
118 char *buf = kmalloc(len+1, GFP_KERNEL);
119 if (!buf)
120 return -ENOMEM;
121 kfree(fi->last_name);
122 fi->last_name = buf;
123 memcpy(fi->last_name, name, len);
124 fi->last_name[len] = 0;
125 fi->next_offset = next_offset;
126 dout("note_last_dentry '%s'\n", fi->last_name);
127 return 0;
128}
129
130/*
110 * When possible, we try to satisfy a readdir by peeking at the 131 * When possible, we try to satisfy a readdir by peeking at the
111 * dcache. We make this work by carefully ordering dentries on 132 * dcache. We make this work by carefully ordering dentries on
112 * d_child when we initially get results back from the MDS, and 133 * d_child when we initially get results back from the MDS, and
@@ -123,123 +144,113 @@ static int __dcache_readdir(struct file *file, struct dir_context *ctx,
123 struct ceph_file_info *fi = file->private_data; 144 struct ceph_file_info *fi = file->private_data;
124 struct dentry *parent = file->f_path.dentry; 145 struct dentry *parent = file->f_path.dentry;
125 struct inode *dir = d_inode(parent); 146 struct inode *dir = d_inode(parent);
126 struct list_head *p; 147 struct dentry *dentry, *last = NULL;
127 struct dentry *dentry, *last;
128 struct ceph_dentry_info *di; 148 struct ceph_dentry_info *di;
149 unsigned nsize = PAGE_CACHE_SIZE / sizeof(struct dentry *);
129 int err = 0; 150 int err = 0;
151 loff_t ptr_pos = 0;
152 struct ceph_readdir_cache_control cache_ctl = {};
130 153
131 /* claim ref on last dentry we returned */ 154 dout("__dcache_readdir %p v%u at %llu\n", dir, shared_gen, ctx->pos);
132 last = fi->dentry;
133 fi->dentry = NULL;
134
135 dout("__dcache_readdir %p v%u at %llu (last %p)\n",
136 dir, shared_gen, ctx->pos, last);
137 155
138 spin_lock(&parent->d_lock); 156 /* we can calculate cache index for the first dirfrag */
139 157 if (ceph_frag_is_leftmost(fpos_frag(ctx->pos))) {
140 /* start at beginning? */ 158 cache_ctl.index = fpos_off(ctx->pos) - 2;
141 if (ctx->pos == 2 || last == NULL || 159 BUG_ON(cache_ctl.index < 0);
142 fpos_cmp(ctx->pos, ceph_dentry(last)->offset) < 0) { 160 ptr_pos = cache_ctl.index * sizeof(struct dentry *);
143 if (list_empty(&parent->d_subdirs))
144 goto out_unlock;
145 p = parent->d_subdirs.prev;
146 dout(" initial p %p/%p\n", p->prev, p->next);
147 } else {
148 p = last->d_child.prev;
149 } 161 }
150 162
151more: 163 while (true) {
152 dentry = list_entry(p, struct dentry, d_child); 164 pgoff_t pgoff;
153 di = ceph_dentry(dentry); 165 bool emit_dentry;
154 while (1) { 166
155 dout(" p %p/%p %s d_subdirs %p/%p\n", p->prev, p->next, 167 if (ptr_pos >= i_size_read(dir)) {
156 d_unhashed(dentry) ? "!hashed" : "hashed",
157 parent->d_subdirs.prev, parent->d_subdirs.next);
158 if (p == &parent->d_subdirs) {
159 fi->flags |= CEPH_F_ATEND; 168 fi->flags |= CEPH_F_ATEND;
160 goto out_unlock; 169 err = 0;
170 break;
171 }
172
173 err = -EAGAIN;
174 pgoff = ptr_pos >> PAGE_CACHE_SHIFT;
175 if (!cache_ctl.page || pgoff != page_index(cache_ctl.page)) {
176 ceph_readdir_cache_release(&cache_ctl);
177 cache_ctl.page = find_lock_page(&dir->i_data, pgoff);
178 if (!cache_ctl.page) {
179 dout(" page %lu not found\n", pgoff);
180 break;
181 }
182 /* reading/filling the cache are serialized by
183 * i_mutex, no need to use page lock */
184 unlock_page(cache_ctl.page);
185 cache_ctl.dentries = kmap(cache_ctl.page);
161 } 186 }
162 spin_lock_nested(&dentry->d_lock, DENTRY_D_LOCK_NESTED); 187
188 rcu_read_lock();
189 spin_lock(&parent->d_lock);
190 /* check i_size again here, because empty directory can be
191 * marked as complete while not holding the i_mutex. */
192 if (ceph_dir_is_complete_ordered(dir) &&
193 ptr_pos < i_size_read(dir))
194 dentry = cache_ctl.dentries[cache_ctl.index % nsize];
195 else
196 dentry = NULL;
197 spin_unlock(&parent->d_lock);
198 if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
199 dentry = NULL;
200 rcu_read_unlock();
201 if (!dentry)
202 break;
203
204 emit_dentry = false;
205 di = ceph_dentry(dentry);
206 spin_lock(&dentry->d_lock);
163 if (di->lease_shared_gen == shared_gen && 207 if (di->lease_shared_gen == shared_gen &&
164 !d_unhashed(dentry) && d_really_is_positive(dentry) && 208 d_really_is_positive(dentry) &&
165 ceph_snap(d_inode(dentry)) != CEPH_SNAPDIR && 209 ceph_snap(d_inode(dentry)) != CEPH_SNAPDIR &&
166 ceph_ino(d_inode(dentry)) != CEPH_INO_CEPH && 210 ceph_ino(d_inode(dentry)) != CEPH_INO_CEPH &&
167 fpos_cmp(ctx->pos, di->offset) <= 0) 211 fpos_cmp(ctx->pos, di->offset) <= 0) {
168 break; 212 emit_dentry = true;
169 dout(" skipping %p %pd at %llu (%llu)%s%s\n", dentry, 213 }
170 dentry, di->offset,
171 ctx->pos, d_unhashed(dentry) ? " unhashed" : "",
172 !d_inode(dentry) ? " null" : "");
173 spin_unlock(&dentry->d_lock); 214 spin_unlock(&dentry->d_lock);
174 p = p->prev;
175 dentry = list_entry(p, struct dentry, d_child);
176 di = ceph_dentry(dentry);
177 }
178
179 dget_dlock(dentry);
180 spin_unlock(&dentry->d_lock);
181 spin_unlock(&parent->d_lock);
182 215
183 /* make sure a dentry wasn't dropped while we didn't have parent lock */ 216 if (emit_dentry) {
184 if (!ceph_dir_is_complete_ordered(dir)) { 217 dout(" %llu (%llu) dentry %p %pd %p\n", di->offset, ctx->pos,
185 dout(" lost dir complete on %p; falling back to mds\n", dir); 218 dentry, dentry, d_inode(dentry));
186 dput(dentry); 219 ctx->pos = di->offset;
187 err = -EAGAIN; 220 if (!dir_emit(ctx, dentry->d_name.name,
188 goto out; 221 dentry->d_name.len,
189 } 222 ceph_translate_ino(dentry->d_sb,
223 d_inode(dentry)->i_ino),
224 d_inode(dentry)->i_mode >> 12)) {
225 dput(dentry);
226 err = 0;
227 break;
228 }
229 ctx->pos++;
190 230
191 dout(" %llu (%llu) dentry %p %pd %p\n", di->offset, ctx->pos, 231 if (last)
192 dentry, dentry, d_inode(dentry)); 232 dput(last);
193 if (!dir_emit(ctx, dentry->d_name.name, 233 last = dentry;
194 dentry->d_name.len, 234 } else {
195 ceph_translate_ino(dentry->d_sb, d_inode(dentry)->i_ino), 235 dput(dentry);
196 d_inode(dentry)->i_mode >> 12)) {
197 if (last) {
198 /* remember our position */
199 fi->dentry = last;
200 fi->next_offset = fpos_off(di->offset);
201 } 236 }
202 dput(dentry);
203 return 0;
204 }
205
206 ctx->pos = di->offset + 1;
207 237
208 if (last) 238 cache_ctl.index++;
209 dput(last); 239 ptr_pos += sizeof(struct dentry *);
210 last = dentry; 240 }
211 241 ceph_readdir_cache_release(&cache_ctl);
212 spin_lock(&parent->d_lock); 242 if (last) {
213 p = p->prev; /* advance to next dentry */ 243 int ret;
214 goto more; 244 di = ceph_dentry(last);
215 245 ret = note_last_dentry(fi, last->d_name.name, last->d_name.len,
216out_unlock: 246 fpos_off(di->offset) + 1);
217 spin_unlock(&parent->d_lock); 247 if (ret < 0)
218out: 248 err = ret;
219 if (last)
220 dput(last); 249 dput(last);
250 }
221 return err; 251 return err;
222} 252}
223 253
224/*
225 * make note of the last dentry we read, so we can
226 * continue at the same lexicographical point,
227 * regardless of what dir changes take place on the
228 * server.
229 */
230static int note_last_dentry(struct ceph_file_info *fi, const char *name,
231 int len)
232{
233 kfree(fi->last_name);
234 fi->last_name = kmalloc(len+1, GFP_NOFS);
235 if (!fi->last_name)
236 return -ENOMEM;
237 memcpy(fi->last_name, name, len);
238 fi->last_name[len] = 0;
239 dout("note_last_dentry '%s'\n", fi->last_name);
240 return 0;
241}
242
243static int ceph_readdir(struct file *file, struct dir_context *ctx) 254static int ceph_readdir(struct file *file, struct dir_context *ctx)
244{ 255{
245 struct ceph_file_info *fi = file->private_data; 256 struct ceph_file_info *fi = file->private_data;
@@ -280,8 +291,7 @@ static int ceph_readdir(struct file *file, struct dir_context *ctx)
280 291
281 /* can we use the dcache? */ 292 /* can we use the dcache? */
282 spin_lock(&ci->i_ceph_lock); 293 spin_lock(&ci->i_ceph_lock);
283 if ((ctx->pos == 2 || fi->dentry) && 294 if (ceph_test_mount_opt(fsc, DCACHE) &&
284 ceph_test_mount_opt(fsc, DCACHE) &&
285 !ceph_test_mount_opt(fsc, NOASYNCREADDIR) && 295 !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
286 ceph_snap(inode) != CEPH_SNAPDIR && 296 ceph_snap(inode) != CEPH_SNAPDIR &&
287 __ceph_dir_is_complete_ordered(ci) && 297 __ceph_dir_is_complete_ordered(ci) &&
@@ -296,24 +306,8 @@ static int ceph_readdir(struct file *file, struct dir_context *ctx)
296 } else { 306 } else {
297 spin_unlock(&ci->i_ceph_lock); 307 spin_unlock(&ci->i_ceph_lock);
298 } 308 }
299 if (fi->dentry) {
300 err = note_last_dentry(fi, fi->dentry->d_name.name,
301 fi->dentry->d_name.len);
302 if (err)
303 return err;
304 dput(fi->dentry);
305 fi->dentry = NULL;
306 }
307 309
308 /* proceed with a normal readdir */ 310 /* proceed with a normal readdir */
309
310 if (ctx->pos == 2) {
311 /* note dir version at start of readdir so we can tell
312 * if any dentries get dropped */
313 fi->dir_release_count = atomic_read(&ci->i_release_count);
314 fi->dir_ordered_count = ci->i_ordered_count;
315 }
316
317more: 311more:
318 /* do we have the correct frag content buffered? */ 312 /* do we have the correct frag content buffered? */
319 if (fi->frag != frag || fi->last_readdir == NULL) { 313 if (fi->frag != frag || fi->last_readdir == NULL) {
@@ -342,12 +336,15 @@ more:
342 req->r_direct_hash = ceph_frag_value(frag); 336 req->r_direct_hash = ceph_frag_value(frag);
343 req->r_direct_is_hash = true; 337 req->r_direct_is_hash = true;
344 if (fi->last_name) { 338 if (fi->last_name) {
345 req->r_path2 = kstrdup(fi->last_name, GFP_NOFS); 339 req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL);
346 if (!req->r_path2) { 340 if (!req->r_path2) {
347 ceph_mdsc_put_request(req); 341 ceph_mdsc_put_request(req);
348 return -ENOMEM; 342 return -ENOMEM;
349 } 343 }
350 } 344 }
345 req->r_dir_release_cnt = fi->dir_release_count;
346 req->r_dir_ordered_cnt = fi->dir_ordered_count;
347 req->r_readdir_cache_idx = fi->readdir_cache_idx;
351 req->r_readdir_offset = fi->next_offset; 348 req->r_readdir_offset = fi->next_offset;
352 req->r_args.readdir.frag = cpu_to_le32(frag); 349 req->r_args.readdir.frag = cpu_to_le32(frag);
353 350
@@ -364,26 +361,38 @@ more:
364 (int)req->r_reply_info.dir_end, 361 (int)req->r_reply_info.dir_end,
365 (int)req->r_reply_info.dir_complete); 362 (int)req->r_reply_info.dir_complete);
366 363
367 if (!req->r_did_prepopulate) {
368 dout("readdir !did_prepopulate");
369 /* preclude from marking dir complete */
370 fi->dir_release_count--;
371 }
372 364
373 /* note next offset and last dentry name */ 365 /* note next offset and last dentry name */
374 rinfo = &req->r_reply_info; 366 rinfo = &req->r_reply_info;
375 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) { 367 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
376 frag = le32_to_cpu(rinfo->dir_dir->frag); 368 frag = le32_to_cpu(rinfo->dir_dir->frag);
377 if (ceph_frag_is_leftmost(frag)) 369 off = req->r_readdir_offset;
378 fi->next_offset = 2; 370 fi->next_offset = off;
379 else
380 fi->next_offset = 0;
381 off = fi->next_offset;
382 } 371 }
372
383 fi->frag = frag; 373 fi->frag = frag;
384 fi->offset = fi->next_offset; 374 fi->offset = fi->next_offset;
385 fi->last_readdir = req; 375 fi->last_readdir = req;
386 376
377 if (req->r_did_prepopulate) {
378 fi->readdir_cache_idx = req->r_readdir_cache_idx;
379 if (fi->readdir_cache_idx < 0) {
380 /* preclude from marking dir ordered */
381 fi->dir_ordered_count = 0;
382 } else if (ceph_frag_is_leftmost(frag) && off == 2) {
383 /* note dir version at start of readdir so
384 * we can tell if any dentries get dropped */
385 fi->dir_release_count = req->r_dir_release_cnt;
386 fi->dir_ordered_count = req->r_dir_ordered_cnt;
387 }
388 } else {
389 dout("readdir !did_prepopulate");
390 /* disable readdir cache */
391 fi->readdir_cache_idx = -1;
392 /* preclude from marking dir complete */
393 fi->dir_release_count = 0;
394 }
395
387 if (req->r_reply_info.dir_end) { 396 if (req->r_reply_info.dir_end) {
388 kfree(fi->last_name); 397 kfree(fi->last_name);
389 fi->last_name = NULL; 398 fi->last_name = NULL;
@@ -394,10 +403,10 @@ more:
394 } else { 403 } else {
395 err = note_last_dentry(fi, 404 err = note_last_dentry(fi,
396 rinfo->dir_dname[rinfo->dir_nr-1], 405 rinfo->dir_dname[rinfo->dir_nr-1],
397 rinfo->dir_dname_len[rinfo->dir_nr-1]); 406 rinfo->dir_dname_len[rinfo->dir_nr-1],
407 fi->next_offset + rinfo->dir_nr);
398 if (err) 408 if (err)
399 return err; 409 return err;
400 fi->next_offset += rinfo->dir_nr;
401 } 410 }
402 } 411 }
403 412
@@ -453,16 +462,22 @@ more:
453 * were released during the whole readdir, and we should have 462 * were released during the whole readdir, and we should have
454 * the complete dir contents in our cache. 463 * the complete dir contents in our cache.
455 */ 464 */
456 spin_lock(&ci->i_ceph_lock); 465 if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) {
457 if (atomic_read(&ci->i_release_count) == fi->dir_release_count) { 466 spin_lock(&ci->i_ceph_lock);
458 if (ci->i_ordered_count == fi->dir_ordered_count) 467 if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) {
459 dout(" marking %p complete and ordered\n", inode); 468 dout(" marking %p complete and ordered\n", inode);
460 else 469 /* use i_size to track number of entries in
470 * readdir cache */
471 BUG_ON(fi->readdir_cache_idx < 0);
472 i_size_write(inode, fi->readdir_cache_idx *
473 sizeof(struct dentry*));
474 } else {
461 dout(" marking %p complete\n", inode); 475 dout(" marking %p complete\n", inode);
476 }
462 __ceph_dir_set_complete(ci, fi->dir_release_count, 477 __ceph_dir_set_complete(ci, fi->dir_release_count,
463 fi->dir_ordered_count); 478 fi->dir_ordered_count);
479 spin_unlock(&ci->i_ceph_lock);
464 } 480 }
465 spin_unlock(&ci->i_ceph_lock);
466 481
467 dout("readdir %p file %p done.\n", inode, file); 482 dout("readdir %p file %p done.\n", inode, file);
468 return 0; 483 return 0;
@@ -476,14 +491,12 @@ static void reset_readdir(struct ceph_file_info *fi, unsigned frag)
476 } 491 }
477 kfree(fi->last_name); 492 kfree(fi->last_name);
478 fi->last_name = NULL; 493 fi->last_name = NULL;
494 fi->dir_release_count = 0;
495 fi->readdir_cache_idx = -1;
479 if (ceph_frag_is_leftmost(frag)) 496 if (ceph_frag_is_leftmost(frag))
480 fi->next_offset = 2; /* compensate for . and .. */ 497 fi->next_offset = 2; /* compensate for . and .. */
481 else 498 else
482 fi->next_offset = 0; 499 fi->next_offset = 0;
483 if (fi->dentry) {
484 dput(fi->dentry);
485 fi->dentry = NULL;
486 }
487 fi->flags &= ~CEPH_F_ATEND; 500 fi->flags &= ~CEPH_F_ATEND;
488} 501}
489 502
@@ -497,13 +510,12 @@ static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
497 mutex_lock(&inode->i_mutex); 510 mutex_lock(&inode->i_mutex);
498 retval = -EINVAL; 511 retval = -EINVAL;
499 switch (whence) { 512 switch (whence) {
500 case SEEK_END:
501 offset += inode->i_size + 2; /* FIXME */
502 break;
503 case SEEK_CUR: 513 case SEEK_CUR:
504 offset += file->f_pos; 514 offset += file->f_pos;
505 case SEEK_SET: 515 case SEEK_SET:
506 break; 516 break;
517 case SEEK_END:
518 retval = -EOPNOTSUPP;
507 default: 519 default:
508 goto out; 520 goto out;
509 } 521 }
@@ -516,20 +528,18 @@ static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
516 } 528 }
517 retval = offset; 529 retval = offset;
518 530
519 /*
520 * discard buffered readdir content on seekdir(0), or
521 * seek to new frag, or seek prior to current chunk.
522 */
523 if (offset == 0 || 531 if (offset == 0 ||
524 fpos_frag(offset) != fi->frag || 532 fpos_frag(offset) != fi->frag ||
525 fpos_off(offset) < fi->offset) { 533 fpos_off(offset) < fi->offset) {
534 /* discard buffered readdir content on seekdir(0), or
535 * seek to new frag, or seek prior to current chunk */
526 dout("dir_llseek dropping %p content\n", file); 536 dout("dir_llseek dropping %p content\n", file);
527 reset_readdir(fi, fpos_frag(offset)); 537 reset_readdir(fi, fpos_frag(offset));
538 } else if (fpos_cmp(offset, old_offset) > 0) {
539 /* reset dir_release_count if we did a forward seek */
540 fi->dir_release_count = 0;
541 fi->readdir_cache_idx = -1;
528 } 542 }
529
530 /* bump dir_release_count if we did a forward seek */
531 if (fpos_cmp(offset, old_offset) > 0)
532 fi->dir_release_count--;
533 } 543 }
534out: 544out:
535 mutex_unlock(&inode->i_mutex); 545 mutex_unlock(&inode->i_mutex);
@@ -764,7 +774,7 @@ static int ceph_symlink(struct inode *dir, struct dentry *dentry,
764 err = PTR_ERR(req); 774 err = PTR_ERR(req);
765 goto out; 775 goto out;
766 } 776 }
767 req->r_path2 = kstrdup(dest, GFP_NOFS); 777 req->r_path2 = kstrdup(dest, GFP_KERNEL);
768 if (!req->r_path2) { 778 if (!req->r_path2) {
769 err = -ENOMEM; 779 err = -ENOMEM;
770 ceph_mdsc_put_request(req); 780 ceph_mdsc_put_request(req);
@@ -985,16 +995,15 @@ static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
985 * to do it here. 995 * to do it here.
986 */ 996 */
987 997
998 /* d_move screws up sibling dentries' offsets */
999 ceph_dir_clear_complete(old_dir);
1000 ceph_dir_clear_complete(new_dir);
1001
988 d_move(old_dentry, new_dentry); 1002 d_move(old_dentry, new_dentry);
989 1003
990 /* ensure target dentry is invalidated, despite 1004 /* ensure target dentry is invalidated, despite
991 rehashing bug in vfs_rename_dir */ 1005 rehashing bug in vfs_rename_dir */
992 ceph_invalidate_dentry_lease(new_dentry); 1006 ceph_invalidate_dentry_lease(new_dentry);
993
994 /* d_move screws up sibling dentries' offsets */
995 ceph_dir_clear_complete(old_dir);
996 ceph_dir_clear_complete(new_dir);
997
998 } 1007 }
999 ceph_mdsc_put_request(req); 1008 ceph_mdsc_put_request(req);
1000 return err; 1009 return err;
@@ -1189,7 +1198,7 @@ static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1189 return -EISDIR; 1198 return -EISDIR;
1190 1199
1191 if (!cf->dir_info) { 1200 if (!cf->dir_info) {
1192 cf->dir_info = kmalloc(bufsize, GFP_NOFS); 1201 cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1193 if (!cf->dir_info) 1202 if (!cf->dir_info)
1194 return -ENOMEM; 1203 return -ENOMEM;
1195 cf->dir_info_len = 1204 cf->dir_info_len =
@@ -1224,66 +1233,6 @@ static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1224} 1233}
1225 1234
1226/* 1235/*
1227 * an fsync() on a dir will wait for any uncommitted directory
1228 * operations to commit.
1229 */
1230static int ceph_dir_fsync(struct file *file, loff_t start, loff_t end,
1231 int datasync)
1232{
1233 struct inode *inode = file_inode(file);
1234 struct ceph_inode_info *ci = ceph_inode(inode);
1235 struct list_head *head = &ci->i_unsafe_dirops;
1236 struct ceph_mds_request *req;
1237 u64 last_tid;
1238 int ret = 0;
1239
1240 dout("dir_fsync %p\n", inode);
1241 ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1242 if (ret)
1243 return ret;
1244 mutex_lock(&inode->i_mutex);
1245
1246 spin_lock(&ci->i_unsafe_lock);
1247 if (list_empty(head))
1248 goto out;
1249
1250 req = list_entry(head->prev,
1251 struct ceph_mds_request, r_unsafe_dir_item);
1252 last_tid = req->r_tid;
1253
1254 do {
1255 ceph_mdsc_get_request(req);
1256 spin_unlock(&ci->i_unsafe_lock);
1257
1258 dout("dir_fsync %p wait on tid %llu (until %llu)\n",
1259 inode, req->r_tid, last_tid);
1260 if (req->r_timeout) {
1261 unsigned long time_left = wait_for_completion_timeout(
1262 &req->r_safe_completion,
1263 req->r_timeout);
1264 if (time_left > 0)
1265 ret = 0;
1266 else
1267 ret = -EIO; /* timed out */
1268 } else {
1269 wait_for_completion(&req->r_safe_completion);
1270 }
1271 ceph_mdsc_put_request(req);
1272
1273 spin_lock(&ci->i_unsafe_lock);
1274 if (ret || list_empty(head))
1275 break;
1276 req = list_entry(head->next,
1277 struct ceph_mds_request, r_unsafe_dir_item);
1278 } while (req->r_tid < last_tid);
1279out:
1280 spin_unlock(&ci->i_unsafe_lock);
1281 mutex_unlock(&inode->i_mutex);
1282
1283 return ret;
1284}
1285
1286/*
1287 * We maintain a private dentry LRU. 1236 * We maintain a private dentry LRU.
1288 * 1237 *
1289 * FIXME: this needs to be changed to a per-mds lru to be useful. 1238 * FIXME: this needs to be changed to a per-mds lru to be useful.
@@ -1353,7 +1302,7 @@ const struct file_operations ceph_dir_fops = {
1353 .open = ceph_open, 1302 .open = ceph_open,
1354 .release = ceph_release, 1303 .release = ceph_release,
1355 .unlocked_ioctl = ceph_ioctl, 1304 .unlocked_ioctl = ceph_ioctl,
1356 .fsync = ceph_dir_fsync, 1305 .fsync = ceph_fsync,
1357}; 1306};
1358 1307
1359const struct file_operations ceph_snapdir_fops = { 1308const struct file_operations ceph_snapdir_fops = {
diff --git a/fs/ceph/file.c b/fs/ceph/file.c
index 3b6b522b4b31..faf92095e105 100644
--- a/fs/ceph/file.c
+++ b/fs/ceph/file.c
@@ -89,13 +89,14 @@ static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
89 case S_IFDIR: 89 case S_IFDIR:
90 dout("init_file %p %p 0%o (regular)\n", inode, file, 90 dout("init_file %p %p 0%o (regular)\n", inode, file,
91 inode->i_mode); 91 inode->i_mode);
92 cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO); 92 cf = kmem_cache_alloc(ceph_file_cachep, GFP_KERNEL | __GFP_ZERO);
93 if (cf == NULL) { 93 if (cf == NULL) {
94 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */ 94 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
95 return -ENOMEM; 95 return -ENOMEM;
96 } 96 }
97 cf->fmode = fmode; 97 cf->fmode = fmode;
98 cf->next_offset = 2; 98 cf->next_offset = 2;
99 cf->readdir_cache_idx = -1;
99 file->private_data = cf; 100 file->private_data = cf;
100 BUG_ON(inode->i_fop->release != ceph_release); 101 BUG_ON(inode->i_fop->release != ceph_release);
101 break; 102 break;
@@ -324,7 +325,6 @@ int ceph_release(struct inode *inode, struct file *file)
324 ceph_mdsc_put_request(cf->last_readdir); 325 ceph_mdsc_put_request(cf->last_readdir);
325 kfree(cf->last_name); 326 kfree(cf->last_name);
326 kfree(cf->dir_info); 327 kfree(cf->dir_info);
327 dput(cf->dentry);
328 kmem_cache_free(ceph_file_cachep, cf); 328 kmem_cache_free(ceph_file_cachep, cf);
329 329
330 /* wake up anyone waiting for caps on this inode */ 330 /* wake up anyone waiting for caps on this inode */
@@ -483,7 +483,7 @@ static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
483 } 483 }
484 } else { 484 } else {
485 num_pages = calc_pages_for(off, len); 485 num_pages = calc_pages_for(off, len);
486 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 486 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
487 if (IS_ERR(pages)) 487 if (IS_ERR(pages))
488 return PTR_ERR(pages); 488 return PTR_ERR(pages);
489 ret = striped_read(inode, off, len, pages, 489 ret = striped_read(inode, off, len, pages,
@@ -557,13 +557,13 @@ static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
557 * objects, rollback on failure, etc.) 557 * objects, rollback on failure, etc.)
558 */ 558 */
559static ssize_t 559static ssize_t
560ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos) 560ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
561 struct ceph_snap_context *snapc)
561{ 562{
562 struct file *file = iocb->ki_filp; 563 struct file *file = iocb->ki_filp;
563 struct inode *inode = file_inode(file); 564 struct inode *inode = file_inode(file);
564 struct ceph_inode_info *ci = ceph_inode(inode); 565 struct ceph_inode_info *ci = ceph_inode(inode);
565 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 566 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
566 struct ceph_snap_context *snapc;
567 struct ceph_vino vino; 567 struct ceph_vino vino;
568 struct ceph_osd_request *req; 568 struct ceph_osd_request *req;
569 struct page **pages; 569 struct page **pages;
@@ -600,7 +600,6 @@ ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
600 size_t start; 600 size_t start;
601 ssize_t n; 601 ssize_t n;
602 602
603 snapc = ci->i_snap_realm->cached_context;
604 vino = ceph_vino(inode); 603 vino = ceph_vino(inode);
605 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 604 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
606 vino, pos, &len, 0, 605 vino, pos, &len, 0,
@@ -614,7 +613,7 @@ ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
614 break; 613 break;
615 } 614 }
616 615
617 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC); 616 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
618 617
619 n = iov_iter_get_pages_alloc(from, &pages, len, &start); 618 n = iov_iter_get_pages_alloc(from, &pages, len, &start);
620 if (unlikely(n < 0)) { 619 if (unlikely(n < 0)) {
@@ -674,13 +673,13 @@ ceph_sync_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
674 * objects, rollback on failure, etc.) 673 * objects, rollback on failure, etc.)
675 */ 674 */
676static ssize_t 675static ssize_t
677ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos) 676ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
677 struct ceph_snap_context *snapc)
678{ 678{
679 struct file *file = iocb->ki_filp; 679 struct file *file = iocb->ki_filp;
680 struct inode *inode = file_inode(file); 680 struct inode *inode = file_inode(file);
681 struct ceph_inode_info *ci = ceph_inode(inode); 681 struct ceph_inode_info *ci = ceph_inode(inode);
682 struct ceph_fs_client *fsc = ceph_inode_to_client(inode); 682 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
683 struct ceph_snap_context *snapc;
684 struct ceph_vino vino; 683 struct ceph_vino vino;
685 struct ceph_osd_request *req; 684 struct ceph_osd_request *req;
686 struct page **pages; 685 struct page **pages;
@@ -717,7 +716,6 @@ ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
717 size_t left; 716 size_t left;
718 int n; 717 int n;
719 718
720 snapc = ci->i_snap_realm->cached_context;
721 vino = ceph_vino(inode); 719 vino = ceph_vino(inode);
722 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, 720 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
723 vino, pos, &len, 0, 1, 721 vino, pos, &len, 0, 1,
@@ -736,7 +734,7 @@ ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
736 */ 734 */
737 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; 735 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
738 736
739 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS); 737 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
740 if (IS_ERR(pages)) { 738 if (IS_ERR(pages)) {
741 ret = PTR_ERR(pages); 739 ret = PTR_ERR(pages);
742 goto out; 740 goto out;
@@ -860,7 +858,7 @@ again:
860 struct page *page = NULL; 858 struct page *page = NULL;
861 loff_t i_size; 859 loff_t i_size;
862 if (retry_op == READ_INLINE) { 860 if (retry_op == READ_INLINE) {
863 page = __page_cache_alloc(GFP_NOFS); 861 page = __page_cache_alloc(GFP_KERNEL);
864 if (!page) 862 if (!page)
865 return -ENOMEM; 863 return -ENOMEM;
866 } 864 }
@@ -941,6 +939,7 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
941 struct ceph_inode_info *ci = ceph_inode(inode); 939 struct ceph_inode_info *ci = ceph_inode(inode);
942 struct ceph_osd_client *osdc = 940 struct ceph_osd_client *osdc =
943 &ceph_sb_to_client(inode->i_sb)->client->osdc; 941 &ceph_sb_to_client(inode->i_sb)->client->osdc;
942 struct ceph_cap_flush *prealloc_cf;
944 ssize_t count, written = 0; 943 ssize_t count, written = 0;
945 int err, want, got; 944 int err, want, got;
946 loff_t pos; 945 loff_t pos;
@@ -948,6 +947,10 @@ static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
948 if (ceph_snap(inode) != CEPH_NOSNAP) 947 if (ceph_snap(inode) != CEPH_NOSNAP)
949 return -EROFS; 948 return -EROFS;
950 949
950 prealloc_cf = ceph_alloc_cap_flush();
951 if (!prealloc_cf)
952 return -ENOMEM;
953
951 mutex_lock(&inode->i_mutex); 954 mutex_lock(&inode->i_mutex);
952 955
953 /* We can write back this queue in page reclaim */ 956 /* We can write back this queue in page reclaim */
@@ -996,14 +999,30 @@ retry_snap:
996 999
997 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 || 1000 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
998 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) { 1001 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1002 struct ceph_snap_context *snapc;
999 struct iov_iter data; 1003 struct iov_iter data;
1000 mutex_unlock(&inode->i_mutex); 1004 mutex_unlock(&inode->i_mutex);
1005
1006 spin_lock(&ci->i_ceph_lock);
1007 if (__ceph_have_pending_cap_snap(ci)) {
1008 struct ceph_cap_snap *capsnap =
1009 list_last_entry(&ci->i_cap_snaps,
1010 struct ceph_cap_snap,
1011 ci_item);
1012 snapc = ceph_get_snap_context(capsnap->context);
1013 } else {
1014 BUG_ON(!ci->i_head_snapc);
1015 snapc = ceph_get_snap_context(ci->i_head_snapc);
1016 }
1017 spin_unlock(&ci->i_ceph_lock);
1018
1001 /* we might need to revert back to that point */ 1019 /* we might need to revert back to that point */
1002 data = *from; 1020 data = *from;
1003 if (iocb->ki_flags & IOCB_DIRECT) 1021 if (iocb->ki_flags & IOCB_DIRECT)
1004 written = ceph_sync_direct_write(iocb, &data, pos); 1022 written = ceph_sync_direct_write(iocb, &data, pos,
1023 snapc);
1005 else 1024 else
1006 written = ceph_sync_write(iocb, &data, pos); 1025 written = ceph_sync_write(iocb, &data, pos, snapc);
1007 if (written == -EOLDSNAPC) { 1026 if (written == -EOLDSNAPC) {
1008 dout("aio_write %p %llx.%llx %llu~%u" 1027 dout("aio_write %p %llx.%llx %llu~%u"
1009 "got EOLDSNAPC, retrying\n", 1028 "got EOLDSNAPC, retrying\n",
@@ -1014,6 +1033,7 @@ retry_snap:
1014 } 1033 }
1015 if (written > 0) 1034 if (written > 0)
1016 iov_iter_advance(from, written); 1035 iov_iter_advance(from, written);
1036 ceph_put_snap_context(snapc);
1017 } else { 1037 } else {
1018 loff_t old_size = inode->i_size; 1038 loff_t old_size = inode->i_size;
1019 /* 1039 /*
@@ -1035,7 +1055,8 @@ retry_snap:
1035 int dirty; 1055 int dirty;
1036 spin_lock(&ci->i_ceph_lock); 1056 spin_lock(&ci->i_ceph_lock);
1037 ci->i_inline_version = CEPH_INLINE_NONE; 1057 ci->i_inline_version = CEPH_INLINE_NONE;
1038 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1058 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1059 &prealloc_cf);
1039 spin_unlock(&ci->i_ceph_lock); 1060 spin_unlock(&ci->i_ceph_lock);
1040 if (dirty) 1061 if (dirty)
1041 __mark_inode_dirty(inode, dirty); 1062 __mark_inode_dirty(inode, dirty);
@@ -1059,6 +1080,7 @@ retry_snap:
1059out: 1080out:
1060 mutex_unlock(&inode->i_mutex); 1081 mutex_unlock(&inode->i_mutex);
1061out_unlocked: 1082out_unlocked:
1083 ceph_free_cap_flush(prealloc_cf);
1062 current->backing_dev_info = NULL; 1084 current->backing_dev_info = NULL;
1063 return written ? written : err; 1085 return written ? written : err;
1064} 1086}
@@ -1255,6 +1277,7 @@ static long ceph_fallocate(struct file *file, int mode,
1255 struct ceph_inode_info *ci = ceph_inode(inode); 1277 struct ceph_inode_info *ci = ceph_inode(inode);
1256 struct ceph_osd_client *osdc = 1278 struct ceph_osd_client *osdc =
1257 &ceph_inode_to_client(inode)->client->osdc; 1279 &ceph_inode_to_client(inode)->client->osdc;
1280 struct ceph_cap_flush *prealloc_cf;
1258 int want, got = 0; 1281 int want, got = 0;
1259 int dirty; 1282 int dirty;
1260 int ret = 0; 1283 int ret = 0;
@@ -1267,6 +1290,10 @@ static long ceph_fallocate(struct file *file, int mode,
1267 if (!S_ISREG(inode->i_mode)) 1290 if (!S_ISREG(inode->i_mode))
1268 return -EOPNOTSUPP; 1291 return -EOPNOTSUPP;
1269 1292
1293 prealloc_cf = ceph_alloc_cap_flush();
1294 if (!prealloc_cf)
1295 return -ENOMEM;
1296
1270 mutex_lock(&inode->i_mutex); 1297 mutex_lock(&inode->i_mutex);
1271 1298
1272 if (ceph_snap(inode) != CEPH_NOSNAP) { 1299 if (ceph_snap(inode) != CEPH_NOSNAP) {
@@ -1313,7 +1340,8 @@ static long ceph_fallocate(struct file *file, int mode,
1313 if (!ret) { 1340 if (!ret) {
1314 spin_lock(&ci->i_ceph_lock); 1341 spin_lock(&ci->i_ceph_lock);
1315 ci->i_inline_version = CEPH_INLINE_NONE; 1342 ci->i_inline_version = CEPH_INLINE_NONE;
1316 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR); 1343 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1344 &prealloc_cf);
1317 spin_unlock(&ci->i_ceph_lock); 1345 spin_unlock(&ci->i_ceph_lock);
1318 if (dirty) 1346 if (dirty)
1319 __mark_inode_dirty(inode, dirty); 1347 __mark_inode_dirty(inode, dirty);
@@ -1322,6 +1350,7 @@ static long ceph_fallocate(struct file *file, int mode,
1322 ceph_put_cap_refs(ci, got); 1350 ceph_put_cap_refs(ci, got);
1323unlock: 1351unlock:
1324 mutex_unlock(&inode->i_mutex); 1352 mutex_unlock(&inode->i_mutex);
1353 ceph_free_cap_flush(prealloc_cf);
1325 return ret; 1354 return ret;
1326} 1355}
1327 1356
diff --git a/fs/ceph/inode.c b/fs/ceph/inode.c
index 571acd88606c..96d2bd829902 100644
--- a/fs/ceph/inode.c
+++ b/fs/ceph/inode.c
@@ -389,9 +389,10 @@ struct inode *ceph_alloc_inode(struct super_block *sb)
389 ci->i_inline_version = 0; 389 ci->i_inline_version = 0;
390 ci->i_time_warp_seq = 0; 390 ci->i_time_warp_seq = 0;
391 ci->i_ceph_flags = 0; 391 ci->i_ceph_flags = 0;
392 ci->i_ordered_count = 0; 392 atomic64_set(&ci->i_ordered_count, 1);
393 atomic_set(&ci->i_release_count, 1); 393 atomic64_set(&ci->i_release_count, 1);
394 atomic_set(&ci->i_complete_count, 0); 394 atomic64_set(&ci->i_complete_seq[0], 0);
395 atomic64_set(&ci->i_complete_seq[1], 0);
395 ci->i_symlink = NULL; 396 ci->i_symlink = NULL;
396 397
397 memset(&ci->i_dir_layout, 0, sizeof(ci->i_dir_layout)); 398 memset(&ci->i_dir_layout, 0, sizeof(ci->i_dir_layout));
@@ -415,9 +416,8 @@ struct inode *ceph_alloc_inode(struct super_block *sb)
415 ci->i_flushing_caps = 0; 416 ci->i_flushing_caps = 0;
416 INIT_LIST_HEAD(&ci->i_dirty_item); 417 INIT_LIST_HEAD(&ci->i_dirty_item);
417 INIT_LIST_HEAD(&ci->i_flushing_item); 418 INIT_LIST_HEAD(&ci->i_flushing_item);
418 ci->i_cap_flush_seq = 0; 419 ci->i_prealloc_cap_flush = NULL;
419 ci->i_cap_flush_last_tid = 0; 420 ci->i_cap_flush_tree = RB_ROOT;
420 memset(&ci->i_cap_flush_tid, 0, sizeof(ci->i_cap_flush_tid));
421 init_waitqueue_head(&ci->i_cap_wq); 421 init_waitqueue_head(&ci->i_cap_wq);
422 ci->i_hold_caps_min = 0; 422 ci->i_hold_caps_min = 0;
423 ci->i_hold_caps_max = 0; 423 ci->i_hold_caps_max = 0;
@@ -752,7 +752,10 @@ static int fill_inode(struct inode *inode, struct page *locked_page,
752 752
753 if (new_version || 753 if (new_version ||
754 (new_issued & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR))) { 754 (new_issued & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
755 if (ci->i_layout.fl_pg_pool != info->layout.fl_pg_pool)
756 ci->i_ceph_flags &= ~CEPH_I_POOL_PERM;
755 ci->i_layout = info->layout; 757 ci->i_layout = info->layout;
758
756 queue_trunc = ceph_fill_file_size(inode, issued, 759 queue_trunc = ceph_fill_file_size(inode, issued,
757 le32_to_cpu(info->truncate_seq), 760 le32_to_cpu(info->truncate_seq),
758 le64_to_cpu(info->truncate_size), 761 le64_to_cpu(info->truncate_size),
@@ -858,9 +861,10 @@ static int fill_inode(struct inode *inode, struct page *locked_page,
858 (issued & CEPH_CAP_FILE_EXCL) == 0 && 861 (issued & CEPH_CAP_FILE_EXCL) == 0 &&
859 !__ceph_dir_is_complete(ci)) { 862 !__ceph_dir_is_complete(ci)) {
860 dout(" marking %p complete (empty)\n", inode); 863 dout(" marking %p complete (empty)\n", inode);
864 i_size_write(inode, 0);
861 __ceph_dir_set_complete(ci, 865 __ceph_dir_set_complete(ci,
862 atomic_read(&ci->i_release_count), 866 atomic64_read(&ci->i_release_count),
863 ci->i_ordered_count); 867 atomic64_read(&ci->i_ordered_count));
864 } 868 }
865 869
866 wake = true; 870 wake = true;
@@ -1212,6 +1216,10 @@ retry_lookup:
1212 dout("fill_trace doing d_move %p -> %p\n", 1216 dout("fill_trace doing d_move %p -> %p\n",
1213 req->r_old_dentry, dn); 1217 req->r_old_dentry, dn);
1214 1218
1219 /* d_move screws up sibling dentries' offsets */
1220 ceph_dir_clear_ordered(dir);
1221 ceph_dir_clear_ordered(olddir);
1222
1215 d_move(req->r_old_dentry, dn); 1223 d_move(req->r_old_dentry, dn);
1216 dout(" src %p '%pd' dst %p '%pd'\n", 1224 dout(" src %p '%pd' dst %p '%pd'\n",
1217 req->r_old_dentry, 1225 req->r_old_dentry,
@@ -1222,10 +1230,6 @@ retry_lookup:
1222 rehashing bug in vfs_rename_dir */ 1230 rehashing bug in vfs_rename_dir */
1223 ceph_invalidate_dentry_lease(dn); 1231 ceph_invalidate_dentry_lease(dn);
1224 1232
1225 /* d_move screws up sibling dentries' offsets */
1226 ceph_dir_clear_ordered(dir);
1227 ceph_dir_clear_ordered(olddir);
1228
1229 dout("dn %p gets new offset %lld\n", req->r_old_dentry, 1233 dout("dn %p gets new offset %lld\n", req->r_old_dentry,
1230 ceph_dentry(req->r_old_dentry)->offset); 1234 ceph_dentry(req->r_old_dentry)->offset);
1231 1235
@@ -1333,6 +1337,49 @@ static int readdir_prepopulate_inodes_only(struct ceph_mds_request *req,
1333 return err; 1337 return err;
1334} 1338}
1335 1339
1340void ceph_readdir_cache_release(struct ceph_readdir_cache_control *ctl)
1341{
1342 if (ctl->page) {
1343 kunmap(ctl->page);
1344 page_cache_release(ctl->page);
1345 ctl->page = NULL;
1346 }
1347}
1348
1349static int fill_readdir_cache(struct inode *dir, struct dentry *dn,
1350 struct ceph_readdir_cache_control *ctl,
1351 struct ceph_mds_request *req)
1352{
1353 struct ceph_inode_info *ci = ceph_inode(dir);
1354 unsigned nsize = PAGE_CACHE_SIZE / sizeof(struct dentry*);
1355 unsigned idx = ctl->index % nsize;
1356 pgoff_t pgoff = ctl->index / nsize;
1357
1358 if (!ctl->page || pgoff != page_index(ctl->page)) {
1359 ceph_readdir_cache_release(ctl);
1360 ctl->page = grab_cache_page(&dir->i_data, pgoff);
1361 if (!ctl->page) {
1362 ctl->index = -1;
1363 return -ENOMEM;
1364 }
1365 /* reading/filling the cache are serialized by
1366 * i_mutex, no need to use page lock */
1367 unlock_page(ctl->page);
1368 ctl->dentries = kmap(ctl->page);
1369 }
1370
1371 if (req->r_dir_release_cnt == atomic64_read(&ci->i_release_count) &&
1372 req->r_dir_ordered_cnt == atomic64_read(&ci->i_ordered_count)) {
1373 dout("readdir cache dn %p idx %d\n", dn, ctl->index);
1374 ctl->dentries[idx] = dn;
1375 ctl->index++;
1376 } else {
1377 dout("disable readdir cache\n");
1378 ctl->index = -1;
1379 }
1380 return 0;
1381}
1382
1336int ceph_readdir_prepopulate(struct ceph_mds_request *req, 1383int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1337 struct ceph_mds_session *session) 1384 struct ceph_mds_session *session)
1338{ 1385{
@@ -1345,8 +1392,11 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1345 struct inode *snapdir = NULL; 1392 struct inode *snapdir = NULL;
1346 struct ceph_mds_request_head *rhead = req->r_request->front.iov_base; 1393 struct ceph_mds_request_head *rhead = req->r_request->front.iov_base;
1347 struct ceph_dentry_info *di; 1394 struct ceph_dentry_info *di;
1348 u64 r_readdir_offset = req->r_readdir_offset;
1349 u32 frag = le32_to_cpu(rhead->args.readdir.frag); 1395 u32 frag = le32_to_cpu(rhead->args.readdir.frag);
1396 struct ceph_readdir_cache_control cache_ctl = {};
1397
1398 if (req->r_aborted)
1399 return readdir_prepopulate_inodes_only(req, session);
1350 1400
1351 if (rinfo->dir_dir && 1401 if (rinfo->dir_dir &&
1352 le32_to_cpu(rinfo->dir_dir->frag) != frag) { 1402 le32_to_cpu(rinfo->dir_dir->frag) != frag) {
@@ -1354,14 +1404,11 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1354 frag, le32_to_cpu(rinfo->dir_dir->frag)); 1404 frag, le32_to_cpu(rinfo->dir_dir->frag));
1355 frag = le32_to_cpu(rinfo->dir_dir->frag); 1405 frag = le32_to_cpu(rinfo->dir_dir->frag);
1356 if (ceph_frag_is_leftmost(frag)) 1406 if (ceph_frag_is_leftmost(frag))
1357 r_readdir_offset = 2; 1407 req->r_readdir_offset = 2;
1358 else 1408 else
1359 r_readdir_offset = 0; 1409 req->r_readdir_offset = 0;
1360 } 1410 }
1361 1411
1362 if (req->r_aborted)
1363 return readdir_prepopulate_inodes_only(req, session);
1364
1365 if (le32_to_cpu(rinfo->head->op) == CEPH_MDS_OP_LSSNAP) { 1412 if (le32_to_cpu(rinfo->head->op) == CEPH_MDS_OP_LSSNAP) {
1366 snapdir = ceph_get_snapdir(d_inode(parent)); 1413 snapdir = ceph_get_snapdir(d_inode(parent));
1367 parent = d_find_alias(snapdir); 1414 parent = d_find_alias(snapdir);
@@ -1374,6 +1421,17 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
1374 ceph_fill_dirfrag(d_inode(parent), rinfo->dir_dir); 1421 ceph_fill_dirfrag(d_inode(parent), rinfo->dir_dir);
1375 } 1422 }
1376 1423
1424 if (ceph_frag_is_leftmost(frag) && req->r_readdir_offset == 2) {
1425 /* note dir version at start of readdir so we can tell
1426 * if any dentries get dropped */
1427 struct ceph_inode_info *ci = ceph_inode(d_inode(parent));
1428 req->r_dir_release_cnt = atomic64_read(&ci->i_release_count);
1429 req->r_dir_ordered_cnt = atomic64_read(&ci->i_ordered_count);
1430 req->r_readdir_cache_idx = 0;
1431 }
1432
1433 cache_ctl.index = req->r_readdir_cache_idx;
1434
1377 /* FIXME: release caps/leases if error occurs */ 1435 /* FIXME: release caps/leases if error occurs */
1378 for (i = 0; i < rinfo->dir_nr; i++) { 1436 for (i = 0; i < rinfo->dir_nr; i++) {
1379 struct ceph_vino vino; 1437 struct ceph_vino vino;
@@ -1413,13 +1471,6 @@ retry_lookup:
1413 d_delete(dn); 1471 d_delete(dn);
1414 dput(dn); 1472 dput(dn);
1415 goto retry_lookup; 1473 goto retry_lookup;
1416 } else {
1417 /* reorder parent's d_subdirs */
1418 spin_lock(&parent->d_lock);
1419 spin_lock_nested(&dn->d_lock, DENTRY_D_LOCK_NESTED);
1420 list_move(&dn->d_child, &parent->d_subdirs);
1421 spin_unlock(&dn->d_lock);
1422 spin_unlock(&parent->d_lock);
1423 } 1474 }
1424 1475
1425 /* inode */ 1476 /* inode */
@@ -1436,13 +1487,15 @@ retry_lookup:
1436 } 1487 }
1437 } 1488 }
1438 1489
1439 if (fill_inode(in, NULL, &rinfo->dir_in[i], NULL, session, 1490 ret = fill_inode(in, NULL, &rinfo->dir_in[i], NULL, session,
1440 req->r_request_started, -1, 1491 req->r_request_started, -1,
1441 &req->r_caps_reservation) < 0) { 1492 &req->r_caps_reservation);
1493 if (ret < 0) {
1442 pr_err("fill_inode badness on %p\n", in); 1494 pr_err("fill_inode badness on %p\n", in);
1443 if (d_really_is_negative(dn)) 1495 if (d_really_is_negative(dn))
1444 iput(in); 1496 iput(in);
1445 d_drop(dn); 1497 d_drop(dn);
1498 err = ret;
1446 goto next_item; 1499 goto next_item;
1447 } 1500 }
1448 1501
@@ -1458,19 +1511,28 @@ retry_lookup:
1458 } 1511 }
1459 1512
1460 di = dn->d_fsdata; 1513 di = dn->d_fsdata;
1461 di->offset = ceph_make_fpos(frag, i + r_readdir_offset); 1514 di->offset = ceph_make_fpos(frag, i + req->r_readdir_offset);
1462 1515
1463 update_dentry_lease(dn, rinfo->dir_dlease[i], 1516 update_dentry_lease(dn, rinfo->dir_dlease[i],
1464 req->r_session, 1517 req->r_session,
1465 req->r_request_started); 1518 req->r_request_started);
1519
1520 if (err == 0 && cache_ctl.index >= 0) {
1521 ret = fill_readdir_cache(d_inode(parent), dn,
1522 &cache_ctl, req);
1523 if (ret < 0)
1524 err = ret;
1525 }
1466next_item: 1526next_item:
1467 if (dn) 1527 if (dn)
1468 dput(dn); 1528 dput(dn);
1469 } 1529 }
1470 if (err == 0)
1471 req->r_did_prepopulate = true;
1472
1473out: 1530out:
1531 if (err == 0) {
1532 req->r_did_prepopulate = true;
1533 req->r_readdir_cache_idx = cache_ctl.index;
1534 }
1535 ceph_readdir_cache_release(&cache_ctl);
1474 if (snapdir) { 1536 if (snapdir) {
1475 iput(snapdir); 1537 iput(snapdir);
1476 dput(parent); 1538 dput(parent);
@@ -1712,11 +1774,13 @@ int ceph_setattr(struct dentry *dentry, struct iattr *attr)
1712 const unsigned int ia_valid = attr->ia_valid; 1774 const unsigned int ia_valid = attr->ia_valid;
1713 struct ceph_mds_request *req; 1775 struct ceph_mds_request *req;
1714 struct ceph_mds_client *mdsc = ceph_sb_to_client(dentry->d_sb)->mdsc; 1776 struct ceph_mds_client *mdsc = ceph_sb_to_client(dentry->d_sb)->mdsc;
1777 struct ceph_cap_flush *prealloc_cf;
1715 int issued; 1778 int issued;
1716 int release = 0, dirtied = 0; 1779 int release = 0, dirtied = 0;
1717 int mask = 0; 1780 int mask = 0;
1718 int err = 0; 1781 int err = 0;
1719 int inode_dirty_flags = 0; 1782 int inode_dirty_flags = 0;
1783 bool lock_snap_rwsem = false;
1720 1784
1721 if (ceph_snap(inode) != CEPH_NOSNAP) 1785 if (ceph_snap(inode) != CEPH_NOSNAP)
1722 return -EROFS; 1786 return -EROFS;
@@ -1725,13 +1789,31 @@ int ceph_setattr(struct dentry *dentry, struct iattr *attr)
1725 if (err != 0) 1789 if (err != 0)
1726 return err; 1790 return err;
1727 1791
1792 prealloc_cf = ceph_alloc_cap_flush();
1793 if (!prealloc_cf)
1794 return -ENOMEM;
1795
1728 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETATTR, 1796 req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SETATTR,
1729 USE_AUTH_MDS); 1797 USE_AUTH_MDS);
1730 if (IS_ERR(req)) 1798 if (IS_ERR(req)) {
1799 ceph_free_cap_flush(prealloc_cf);
1731 return PTR_ERR(req); 1800 return PTR_ERR(req);
1801 }
1732 1802
1733 spin_lock(&ci->i_ceph_lock); 1803 spin_lock(&ci->i_ceph_lock);
1734 issued = __ceph_caps_issued(ci, NULL); 1804 issued = __ceph_caps_issued(ci, NULL);
1805
1806 if (!ci->i_head_snapc &&
1807 (issued & (CEPH_CAP_ANY_EXCL | CEPH_CAP_FILE_WR))) {
1808 lock_snap_rwsem = true;
1809 if (!down_read_trylock(&mdsc->snap_rwsem)) {
1810 spin_unlock(&ci->i_ceph_lock);
1811 down_read(&mdsc->snap_rwsem);
1812 spin_lock(&ci->i_ceph_lock);
1813 issued = __ceph_caps_issued(ci, NULL);
1814 }
1815 }
1816
1735 dout("setattr %p issued %s\n", inode, ceph_cap_string(issued)); 1817 dout("setattr %p issued %s\n", inode, ceph_cap_string(issued));
1736 1818
1737 if (ia_valid & ATTR_UID) { 1819 if (ia_valid & ATTR_UID) {
@@ -1874,12 +1956,15 @@ int ceph_setattr(struct dentry *dentry, struct iattr *attr)
1874 dout("setattr %p ATTR_FILE ... hrm!\n", inode); 1956 dout("setattr %p ATTR_FILE ... hrm!\n", inode);
1875 1957
1876 if (dirtied) { 1958 if (dirtied) {
1877 inode_dirty_flags = __ceph_mark_dirty_caps(ci, dirtied); 1959 inode_dirty_flags = __ceph_mark_dirty_caps(ci, dirtied,
1960 &prealloc_cf);
1878 inode->i_ctime = CURRENT_TIME; 1961 inode->i_ctime = CURRENT_TIME;
1879 } 1962 }
1880 1963
1881 release &= issued; 1964 release &= issued;
1882 spin_unlock(&ci->i_ceph_lock); 1965 spin_unlock(&ci->i_ceph_lock);
1966 if (lock_snap_rwsem)
1967 up_read(&mdsc->snap_rwsem);
1883 1968
1884 if (inode_dirty_flags) 1969 if (inode_dirty_flags)
1885 __mark_inode_dirty(inode, inode_dirty_flags); 1970 __mark_inode_dirty(inode, inode_dirty_flags);
@@ -1904,9 +1989,11 @@ int ceph_setattr(struct dentry *dentry, struct iattr *attr)
1904 ceph_mdsc_put_request(req); 1989 ceph_mdsc_put_request(req);
1905 if (mask & CEPH_SETATTR_SIZE) 1990 if (mask & CEPH_SETATTR_SIZE)
1906 __ceph_do_pending_vmtruncate(inode); 1991 __ceph_do_pending_vmtruncate(inode);
1992 ceph_free_cap_flush(prealloc_cf);
1907 return err; 1993 return err;
1908out_put: 1994out_put:
1909 ceph_mdsc_put_request(req); 1995 ceph_mdsc_put_request(req);
1996 ceph_free_cap_flush(prealloc_cf);
1910 return err; 1997 return err;
1911} 1998}
1912 1999
diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index 84f37f34f9aa..6aa07af67603 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -8,6 +8,7 @@
8#include <linux/debugfs.h> 8#include <linux/debugfs.h>
9#include <linux/seq_file.h> 9#include <linux/seq_file.h>
10#include <linux/utsname.h> 10#include <linux/utsname.h>
11#include <linux/ratelimit.h>
11 12
12#include "super.h" 13#include "super.h"
13#include "mds_client.h" 14#include "mds_client.h"
@@ -458,7 +459,6 @@ static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
458 s->s_cap_reconnect = 0; 459 s->s_cap_reconnect = 0;
459 s->s_cap_iterator = NULL; 460 s->s_cap_iterator = NULL;
460 INIT_LIST_HEAD(&s->s_cap_releases); 461 INIT_LIST_HEAD(&s->s_cap_releases);
461 INIT_LIST_HEAD(&s->s_cap_releases_done);
462 INIT_LIST_HEAD(&s->s_cap_flushing); 462 INIT_LIST_HEAD(&s->s_cap_flushing);
463 INIT_LIST_HEAD(&s->s_cap_snaps_flushing); 463 INIT_LIST_HEAD(&s->s_cap_snaps_flushing);
464 464
@@ -629,6 +629,9 @@ static void __register_request(struct ceph_mds_client *mdsc,
629 req->r_uid = current_fsuid(); 629 req->r_uid = current_fsuid();
630 req->r_gid = current_fsgid(); 630 req->r_gid = current_fsgid();
631 631
632 if (mdsc->oldest_tid == 0 && req->r_op != CEPH_MDS_OP_SETFILELOCK)
633 mdsc->oldest_tid = req->r_tid;
634
632 if (dir) { 635 if (dir) {
633 struct ceph_inode_info *ci = ceph_inode(dir); 636 struct ceph_inode_info *ci = ceph_inode(dir);
634 637
@@ -644,6 +647,21 @@ static void __unregister_request(struct ceph_mds_client *mdsc,
644 struct ceph_mds_request *req) 647 struct ceph_mds_request *req)
645{ 648{
646 dout("__unregister_request %p tid %lld\n", req, req->r_tid); 649 dout("__unregister_request %p tid %lld\n", req, req->r_tid);
650
651 if (req->r_tid == mdsc->oldest_tid) {
652 struct rb_node *p = rb_next(&req->r_node);
653 mdsc->oldest_tid = 0;
654 while (p) {
655 struct ceph_mds_request *next_req =
656 rb_entry(p, struct ceph_mds_request, r_node);
657 if (next_req->r_op != CEPH_MDS_OP_SETFILELOCK) {
658 mdsc->oldest_tid = next_req->r_tid;
659 break;
660 }
661 p = rb_next(p);
662 }
663 }
664
647 rb_erase(&req->r_node, &mdsc->request_tree); 665 rb_erase(&req->r_node, &mdsc->request_tree);
648 RB_CLEAR_NODE(&req->r_node); 666 RB_CLEAR_NODE(&req->r_node);
649 667
@@ -998,27 +1016,25 @@ void ceph_mdsc_open_export_target_sessions(struct ceph_mds_client *mdsc,
998 * session caps 1016 * session caps
999 */ 1017 */
1000 1018
1001/* 1019/* caller holds s_cap_lock, we drop it */
1002 * Free preallocated cap messages assigned to this session 1020static void cleanup_cap_releases(struct ceph_mds_client *mdsc,
1003 */ 1021 struct ceph_mds_session *session)
1004static void cleanup_cap_releases(struct ceph_mds_session *session) 1022 __releases(session->s_cap_lock)
1005{ 1023{
1006 struct ceph_msg *msg; 1024 LIST_HEAD(tmp_list);
1025 list_splice_init(&session->s_cap_releases, &tmp_list);
1026 session->s_num_cap_releases = 0;
1027 spin_unlock(&session->s_cap_lock);
1007 1028
1008 spin_lock(&session->s_cap_lock); 1029 dout("cleanup_cap_releases mds%d\n", session->s_mds);
1009 while (!list_empty(&session->s_cap_releases)) { 1030 while (!list_empty(&tmp_list)) {
1010 msg = list_first_entry(&session->s_cap_releases, 1031 struct ceph_cap *cap;
1011 struct ceph_msg, list_head); 1032 /* zero out the in-progress message */
1012 list_del_init(&msg->list_head); 1033 cap = list_first_entry(&tmp_list,
1013 ceph_msg_put(msg); 1034 struct ceph_cap, session_caps);
1014 } 1035 list_del(&cap->session_caps);
1015 while (!list_empty(&session->s_cap_releases_done)) { 1036 ceph_put_cap(mdsc, cap);
1016 msg = list_first_entry(&session->s_cap_releases_done,
1017 struct ceph_msg, list_head);
1018 list_del_init(&msg->list_head);
1019 ceph_msg_put(msg);
1020 } 1037 }
1021 spin_unlock(&session->s_cap_lock);
1022} 1038}
1023 1039
1024static void cleanup_session_requests(struct ceph_mds_client *mdsc, 1040static void cleanup_session_requests(struct ceph_mds_client *mdsc,
@@ -1033,7 +1049,8 @@ static void cleanup_session_requests(struct ceph_mds_client *mdsc,
1033 req = list_first_entry(&session->s_unsafe, 1049 req = list_first_entry(&session->s_unsafe,
1034 struct ceph_mds_request, r_unsafe_item); 1050 struct ceph_mds_request, r_unsafe_item);
1035 list_del_init(&req->r_unsafe_item); 1051 list_del_init(&req->r_unsafe_item);
1036 pr_info(" dropping unsafe request %llu\n", req->r_tid); 1052 pr_warn_ratelimited(" dropping unsafe request %llu\n",
1053 req->r_tid);
1037 __unregister_request(mdsc, req); 1054 __unregister_request(mdsc, req);
1038 } 1055 }
1039 /* zero r_attempts, so kick_requests() will re-send requests */ 1056 /* zero r_attempts, so kick_requests() will re-send requests */
@@ -1095,10 +1112,16 @@ static int iterate_session_caps(struct ceph_mds_session *session,
1095 dout("iterate_session_caps finishing cap %p removal\n", 1112 dout("iterate_session_caps finishing cap %p removal\n",
1096 cap); 1113 cap);
1097 BUG_ON(cap->session != session); 1114 BUG_ON(cap->session != session);
1115 cap->session = NULL;
1098 list_del_init(&cap->session_caps); 1116 list_del_init(&cap->session_caps);
1099 session->s_nr_caps--; 1117 session->s_nr_caps--;
1100 cap->session = NULL; 1118 if (cap->queue_release) {
1101 old_cap = cap; /* put_cap it w/o locks held */ 1119 list_add_tail(&cap->session_caps,
1120 &session->s_cap_releases);
1121 session->s_num_cap_releases++;
1122 } else {
1123 old_cap = cap; /* put_cap it w/o locks held */
1124 }
1102 } 1125 }
1103 if (ret < 0) 1126 if (ret < 0)
1104 goto out; 1127 goto out;
@@ -1119,6 +1142,7 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1119 void *arg) 1142 void *arg)
1120{ 1143{
1121 struct ceph_inode_info *ci = ceph_inode(inode); 1144 struct ceph_inode_info *ci = ceph_inode(inode);
1145 LIST_HEAD(to_remove);
1122 int drop = 0; 1146 int drop = 0;
1123 1147
1124 dout("removing cap %p, ci is %p, inode is %p\n", 1148 dout("removing cap %p, ci is %p, inode is %p\n",
@@ -1126,12 +1150,27 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1126 spin_lock(&ci->i_ceph_lock); 1150 spin_lock(&ci->i_ceph_lock);
1127 __ceph_remove_cap(cap, false); 1151 __ceph_remove_cap(cap, false);
1128 if (!ci->i_auth_cap) { 1152 if (!ci->i_auth_cap) {
1153 struct ceph_cap_flush *cf;
1129 struct ceph_mds_client *mdsc = 1154 struct ceph_mds_client *mdsc =
1130 ceph_sb_to_client(inode->i_sb)->mdsc; 1155 ceph_sb_to_client(inode->i_sb)->mdsc;
1131 1156
1157 while (true) {
1158 struct rb_node *n = rb_first(&ci->i_cap_flush_tree);
1159 if (!n)
1160 break;
1161 cf = rb_entry(n, struct ceph_cap_flush, i_node);
1162 rb_erase(&cf->i_node, &ci->i_cap_flush_tree);
1163 list_add(&cf->list, &to_remove);
1164 }
1165
1132 spin_lock(&mdsc->cap_dirty_lock); 1166 spin_lock(&mdsc->cap_dirty_lock);
1167
1168 list_for_each_entry(cf, &to_remove, list)
1169 rb_erase(&cf->g_node, &mdsc->cap_flush_tree);
1170
1133 if (!list_empty(&ci->i_dirty_item)) { 1171 if (!list_empty(&ci->i_dirty_item)) {
1134 pr_info(" dropping dirty %s state for %p %lld\n", 1172 pr_warn_ratelimited(
1173 " dropping dirty %s state for %p %lld\n",
1135 ceph_cap_string(ci->i_dirty_caps), 1174 ceph_cap_string(ci->i_dirty_caps),
1136 inode, ceph_ino(inode)); 1175 inode, ceph_ino(inode));
1137 ci->i_dirty_caps = 0; 1176 ci->i_dirty_caps = 0;
@@ -1139,7 +1178,8 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1139 drop = 1; 1178 drop = 1;
1140 } 1179 }
1141 if (!list_empty(&ci->i_flushing_item)) { 1180 if (!list_empty(&ci->i_flushing_item)) {
1142 pr_info(" dropping dirty+flushing %s state for %p %lld\n", 1181 pr_warn_ratelimited(
1182 " dropping dirty+flushing %s state for %p %lld\n",
1143 ceph_cap_string(ci->i_flushing_caps), 1183 ceph_cap_string(ci->i_flushing_caps),
1144 inode, ceph_ino(inode)); 1184 inode, ceph_ino(inode));
1145 ci->i_flushing_caps = 0; 1185 ci->i_flushing_caps = 0;
@@ -1148,8 +1188,20 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
1148 drop = 1; 1188 drop = 1;
1149 } 1189 }
1150 spin_unlock(&mdsc->cap_dirty_lock); 1190 spin_unlock(&mdsc->cap_dirty_lock);
1191
1192 if (!ci->i_dirty_caps && ci->i_prealloc_cap_flush) {
1193 list_add(&ci->i_prealloc_cap_flush->list, &to_remove);
1194 ci->i_prealloc_cap_flush = NULL;
1195 }
1151 } 1196 }
1152 spin_unlock(&ci->i_ceph_lock); 1197 spin_unlock(&ci->i_ceph_lock);
1198 while (!list_empty(&to_remove)) {
1199 struct ceph_cap_flush *cf;
1200 cf = list_first_entry(&to_remove,
1201 struct ceph_cap_flush, list);
1202 list_del(&cf->list);
1203 ceph_free_cap_flush(cf);
1204 }
1153 while (drop--) 1205 while (drop--)
1154 iput(inode); 1206 iput(inode);
1155 return 0; 1207 return 0;
@@ -1191,11 +1243,12 @@ static void remove_session_caps(struct ceph_mds_session *session)
1191 spin_lock(&session->s_cap_lock); 1243 spin_lock(&session->s_cap_lock);
1192 } 1244 }
1193 } 1245 }
1194 spin_unlock(&session->s_cap_lock); 1246
1247 // drop cap expires and unlock s_cap_lock
1248 cleanup_cap_releases(session->s_mdsc, session);
1195 1249
1196 BUG_ON(session->s_nr_caps > 0); 1250 BUG_ON(session->s_nr_caps > 0);
1197 BUG_ON(!list_empty(&session->s_cap_flushing)); 1251 BUG_ON(!list_empty(&session->s_cap_flushing));
1198 cleanup_cap_releases(session);
1199} 1252}
1200 1253
1201/* 1254/*
@@ -1371,7 +1424,8 @@ static int trim_caps_cb(struct inode *inode, struct ceph_cap *cap, void *arg)
1371 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued), 1424 inode, cap, ceph_cap_string(mine), ceph_cap_string(oissued),
1372 ceph_cap_string(used), ceph_cap_string(wanted)); 1425 ceph_cap_string(used), ceph_cap_string(wanted));
1373 if (cap == ci->i_auth_cap) { 1426 if (cap == ci->i_auth_cap) {
1374 if (ci->i_dirty_caps | ci->i_flushing_caps) 1427 if (ci->i_dirty_caps || ci->i_flushing_caps ||
1428 !list_empty(&ci->i_cap_snaps))
1375 goto out; 1429 goto out;
1376 if ((used | wanted) & CEPH_CAP_ANY_WR) 1430 if ((used | wanted) & CEPH_CAP_ANY_WR)
1377 goto out; 1431 goto out;
@@ -1417,121 +1471,80 @@ static int trim_caps(struct ceph_mds_client *mdsc,
1417 session->s_trim_caps = 0; 1471 session->s_trim_caps = 0;
1418 } 1472 }
1419 1473
1420 ceph_add_cap_releases(mdsc, session);
1421 ceph_send_cap_releases(mdsc, session); 1474 ceph_send_cap_releases(mdsc, session);
1422 return 0; 1475 return 0;
1423} 1476}
1424 1477
1425/* 1478static int check_capsnap_flush(struct ceph_inode_info *ci,
1426 * Allocate cap_release messages. If there is a partially full message 1479 u64 want_snap_seq)
1427 * in the queue, try to allocate enough to cover it's remainder, so that
1428 * we can send it immediately.
1429 *
1430 * Called under s_mutex.
1431 */
1432int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
1433 struct ceph_mds_session *session)
1434{ 1480{
1435 struct ceph_msg *msg, *partial = NULL; 1481 int ret = 1;
1436 struct ceph_mds_cap_release *head; 1482 spin_lock(&ci->i_ceph_lock);
1437 int err = -ENOMEM; 1483 if (want_snap_seq > 0 && !list_empty(&ci->i_cap_snaps)) {
1438 int extra = mdsc->fsc->mount_options->cap_release_safety; 1484 struct ceph_cap_snap *capsnap =
1439 int num; 1485 list_first_entry(&ci->i_cap_snaps,
1440 1486 struct ceph_cap_snap, ci_item);
1441 dout("add_cap_releases %p mds%d extra %d\n", session, session->s_mds, 1487 ret = capsnap->follows >= want_snap_seq;
1442 extra);
1443
1444 spin_lock(&session->s_cap_lock);
1445
1446 if (!list_empty(&session->s_cap_releases)) {
1447 msg = list_first_entry(&session->s_cap_releases,
1448 struct ceph_msg,
1449 list_head);
1450 head = msg->front.iov_base;
1451 num = le32_to_cpu(head->num);
1452 if (num) {
1453 dout(" partial %p with (%d/%d)\n", msg, num,
1454 (int)CEPH_CAPS_PER_RELEASE);
1455 extra += CEPH_CAPS_PER_RELEASE - num;
1456 partial = msg;
1457 }
1458 }
1459 while (session->s_num_cap_releases < session->s_nr_caps + extra) {
1460 spin_unlock(&session->s_cap_lock);
1461 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE, PAGE_CACHE_SIZE,
1462 GFP_NOFS, false);
1463 if (!msg)
1464 goto out_unlocked;
1465 dout("add_cap_releases %p msg %p now %d\n", session, msg,
1466 (int)msg->front.iov_len);
1467 head = msg->front.iov_base;
1468 head->num = cpu_to_le32(0);
1469 msg->front.iov_len = sizeof(*head);
1470 spin_lock(&session->s_cap_lock);
1471 list_add(&msg->list_head, &session->s_cap_releases);
1472 session->s_num_cap_releases += CEPH_CAPS_PER_RELEASE;
1473 }
1474
1475 if (partial) {
1476 head = partial->front.iov_base;
1477 num = le32_to_cpu(head->num);
1478 dout(" queueing partial %p with %d/%d\n", partial, num,
1479 (int)CEPH_CAPS_PER_RELEASE);
1480 list_move_tail(&partial->list_head,
1481 &session->s_cap_releases_done);
1482 session->s_num_cap_releases -= CEPH_CAPS_PER_RELEASE - num;
1483 } 1488 }
1484 err = 0; 1489 spin_unlock(&ci->i_ceph_lock);
1485 spin_unlock(&session->s_cap_lock); 1490 return ret;
1486out_unlocked:
1487 return err;
1488} 1491}
1489 1492
1490static int check_cap_flush(struct inode *inode, u64 want_flush_seq) 1493static int check_caps_flush(struct ceph_mds_client *mdsc,
1494 u64 want_flush_tid)
1491{ 1495{
1492 struct ceph_inode_info *ci = ceph_inode(inode); 1496 struct rb_node *n;
1493 int ret; 1497 struct ceph_cap_flush *cf;
1494 spin_lock(&ci->i_ceph_lock); 1498 int ret = 1;
1495 if (ci->i_flushing_caps) 1499
1496 ret = ci->i_cap_flush_seq >= want_flush_seq; 1500 spin_lock(&mdsc->cap_dirty_lock);
1497 else 1501 n = rb_first(&mdsc->cap_flush_tree);
1498 ret = 1; 1502 cf = n ? rb_entry(n, struct ceph_cap_flush, g_node) : NULL;
1499 spin_unlock(&ci->i_ceph_lock); 1503 if (cf && cf->tid <= want_flush_tid) {
1504 dout("check_caps_flush still flushing tid %llu <= %llu\n",
1505 cf->tid, want_flush_tid);
1506 ret = 0;
1507 }
1508 spin_unlock(&mdsc->cap_dirty_lock);
1500 return ret; 1509 return ret;
1501} 1510}
1502 1511
1503/* 1512/*
1504 * flush all dirty inode data to disk. 1513 * flush all dirty inode data to disk.
1505 * 1514 *
1506 * returns true if we've flushed through want_flush_seq 1515 * returns true if we've flushed through want_flush_tid
1507 */ 1516 */
1508static void wait_caps_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq) 1517static void wait_caps_flush(struct ceph_mds_client *mdsc,
1518 u64 want_flush_tid, u64 want_snap_seq)
1509{ 1519{
1510 int mds; 1520 int mds;
1511 1521
1512 dout("check_cap_flush want %lld\n", want_flush_seq); 1522 dout("check_caps_flush want %llu snap want %llu\n",
1523 want_flush_tid, want_snap_seq);
1513 mutex_lock(&mdsc->mutex); 1524 mutex_lock(&mdsc->mutex);
1514 for (mds = 0; mds < mdsc->max_sessions; mds++) { 1525 for (mds = 0; mds < mdsc->max_sessions; ) {
1515 struct ceph_mds_session *session = mdsc->sessions[mds]; 1526 struct ceph_mds_session *session = mdsc->sessions[mds];
1516 struct inode *inode = NULL; 1527 struct inode *inode = NULL;
1517 1528
1518 if (!session) 1529 if (!session) {
1530 mds++;
1519 continue; 1531 continue;
1532 }
1520 get_session(session); 1533 get_session(session);
1521 mutex_unlock(&mdsc->mutex); 1534 mutex_unlock(&mdsc->mutex);
1522 1535
1523 mutex_lock(&session->s_mutex); 1536 mutex_lock(&session->s_mutex);
1524 if (!list_empty(&session->s_cap_flushing)) { 1537 if (!list_empty(&session->s_cap_snaps_flushing)) {
1525 struct ceph_inode_info *ci = 1538 struct ceph_cap_snap *capsnap =
1526 list_entry(session->s_cap_flushing.next, 1539 list_first_entry(&session->s_cap_snaps_flushing,
1527 struct ceph_inode_info, 1540 struct ceph_cap_snap,
1528 i_flushing_item); 1541 flushing_item);
1529 1542 struct ceph_inode_info *ci = capsnap->ci;
1530 if (!check_cap_flush(&ci->vfs_inode, want_flush_seq)) { 1543 if (!check_capsnap_flush(ci, want_snap_seq)) {
1531 dout("check_cap_flush still flushing %p " 1544 dout("check_cap_flush still flushing snap %p "
1532 "seq %lld <= %lld to mds%d\n", 1545 "follows %lld <= %lld to mds%d\n",
1533 &ci->vfs_inode, ci->i_cap_flush_seq, 1546 &ci->vfs_inode, capsnap->follows,
1534 want_flush_seq, session->s_mds); 1547 want_snap_seq, mds);
1535 inode = igrab(&ci->vfs_inode); 1548 inode = igrab(&ci->vfs_inode);
1536 } 1549 }
1537 } 1550 }
@@ -1540,15 +1553,21 @@ static void wait_caps_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1540 1553
1541 if (inode) { 1554 if (inode) {
1542 wait_event(mdsc->cap_flushing_wq, 1555 wait_event(mdsc->cap_flushing_wq,
1543 check_cap_flush(inode, want_flush_seq)); 1556 check_capsnap_flush(ceph_inode(inode),
1557 want_snap_seq));
1544 iput(inode); 1558 iput(inode);
1559 } else {
1560 mds++;
1545 } 1561 }
1546 1562
1547 mutex_lock(&mdsc->mutex); 1563 mutex_lock(&mdsc->mutex);
1548 } 1564 }
1549
1550 mutex_unlock(&mdsc->mutex); 1565 mutex_unlock(&mdsc->mutex);
1551 dout("check_cap_flush ok, flushed thru %lld\n", want_flush_seq); 1566
1567 wait_event(mdsc->cap_flushing_wq,
1568 check_caps_flush(mdsc, want_flush_tid));
1569
1570 dout("check_caps_flush ok, flushed thru %llu\n", want_flush_tid);
1552} 1571}
1553 1572
1554/* 1573/*
@@ -1557,60 +1576,74 @@ static void wait_caps_flush(struct ceph_mds_client *mdsc, u64 want_flush_seq)
1557void ceph_send_cap_releases(struct ceph_mds_client *mdsc, 1576void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
1558 struct ceph_mds_session *session) 1577 struct ceph_mds_session *session)
1559{ 1578{
1560 struct ceph_msg *msg; 1579 struct ceph_msg *msg = NULL;
1580 struct ceph_mds_cap_release *head;
1581 struct ceph_mds_cap_item *item;
1582 struct ceph_cap *cap;
1583 LIST_HEAD(tmp_list);
1584 int num_cap_releases;
1561 1585
1562 dout("send_cap_releases mds%d\n", session->s_mds);
1563 spin_lock(&session->s_cap_lock); 1586 spin_lock(&session->s_cap_lock);
1564 while (!list_empty(&session->s_cap_releases_done)) { 1587again:
1565 msg = list_first_entry(&session->s_cap_releases_done, 1588 list_splice_init(&session->s_cap_releases, &tmp_list);
1566 struct ceph_msg, list_head); 1589 num_cap_releases = session->s_num_cap_releases;
1567 list_del_init(&msg->list_head); 1590 session->s_num_cap_releases = 0;
1568 spin_unlock(&session->s_cap_lock);
1569 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1570 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1571 ceph_con_send(&session->s_con, msg);
1572 spin_lock(&session->s_cap_lock);
1573 }
1574 spin_unlock(&session->s_cap_lock); 1591 spin_unlock(&session->s_cap_lock);
1575}
1576 1592
1577static void discard_cap_releases(struct ceph_mds_client *mdsc, 1593 while (!list_empty(&tmp_list)) {
1578 struct ceph_mds_session *session) 1594 if (!msg) {
1579{ 1595 msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPRELEASE,
1580 struct ceph_msg *msg; 1596 PAGE_CACHE_SIZE, GFP_NOFS, false);
1581 struct ceph_mds_cap_release *head; 1597 if (!msg)
1582 unsigned num; 1598 goto out_err;
1583 1599 head = msg->front.iov_base;
1584 dout("discard_cap_releases mds%d\n", session->s_mds); 1600 head->num = cpu_to_le32(0);
1601 msg->front.iov_len = sizeof(*head);
1602 }
1603 cap = list_first_entry(&tmp_list, struct ceph_cap,
1604 session_caps);
1605 list_del(&cap->session_caps);
1606 num_cap_releases--;
1585 1607
1586 if (!list_empty(&session->s_cap_releases)) {
1587 /* zero out the in-progress message */
1588 msg = list_first_entry(&session->s_cap_releases,
1589 struct ceph_msg, list_head);
1590 head = msg->front.iov_base; 1608 head = msg->front.iov_base;
1591 num = le32_to_cpu(head->num); 1609 le32_add_cpu(&head->num, 1);
1592 dout("discard_cap_releases mds%d %p %u\n", 1610 item = msg->front.iov_base + msg->front.iov_len;
1593 session->s_mds, msg, num); 1611 item->ino = cpu_to_le64(cap->cap_ino);
1594 head->num = cpu_to_le32(0); 1612 item->cap_id = cpu_to_le64(cap->cap_id);
1595 msg->front.iov_len = sizeof(*head); 1613 item->migrate_seq = cpu_to_le32(cap->mseq);
1596 session->s_num_cap_releases += num; 1614 item->seq = cpu_to_le32(cap->issue_seq);
1615 msg->front.iov_len += sizeof(*item);
1616
1617 ceph_put_cap(mdsc, cap);
1618
1619 if (le32_to_cpu(head->num) == CEPH_CAPS_PER_RELEASE) {
1620 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1621 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1622 ceph_con_send(&session->s_con, msg);
1623 msg = NULL;
1624 }
1597 } 1625 }
1598 1626
1599 /* requeue completed messages */ 1627 BUG_ON(num_cap_releases != 0);
1600 while (!list_empty(&session->s_cap_releases_done)) {
1601 msg = list_first_entry(&session->s_cap_releases_done,
1602 struct ceph_msg, list_head);
1603 list_del_init(&msg->list_head);
1604 1628
1605 head = msg->front.iov_base; 1629 spin_lock(&session->s_cap_lock);
1606 num = le32_to_cpu(head->num); 1630 if (!list_empty(&session->s_cap_releases))
1607 dout("discard_cap_releases mds%d %p %u\n", session->s_mds, msg, 1631 goto again;
1608 num); 1632 spin_unlock(&session->s_cap_lock);
1609 session->s_num_cap_releases += num; 1633
1610 head->num = cpu_to_le32(0); 1634 if (msg) {
1611 msg->front.iov_len = sizeof(*head); 1635 msg->hdr.front_len = cpu_to_le32(msg->front.iov_len);
1612 list_add(&msg->list_head, &session->s_cap_releases); 1636 dout("send_cap_releases mds%d %p\n", session->s_mds, msg);
1637 ceph_con_send(&session->s_con, msg);
1613 } 1638 }
1639 return;
1640out_err:
1641 pr_err("send_cap_releases mds%d, failed to allocate message\n",
1642 session->s_mds);
1643 spin_lock(&session->s_cap_lock);
1644 list_splice(&tmp_list, &session->s_cap_releases);
1645 session->s_num_cap_releases += num_cap_releases;
1646 spin_unlock(&session->s_cap_lock);
1614} 1647}
1615 1648
1616/* 1649/*
@@ -1635,7 +1668,8 @@ int ceph_alloc_readdir_reply_buffer(struct ceph_mds_request *req,
1635 1668
1636 order = get_order(size * num_entries); 1669 order = get_order(size * num_entries);
1637 while (order >= 0) { 1670 while (order >= 0) {
1638 rinfo->dir_in = (void*)__get_free_pages(GFP_NOFS | __GFP_NOWARN, 1671 rinfo->dir_in = (void*)__get_free_pages(GFP_KERNEL |
1672 __GFP_NOWARN,
1639 order); 1673 order);
1640 if (rinfo->dir_in) 1674 if (rinfo->dir_in)
1641 break; 1675 break;
@@ -1697,13 +1731,9 @@ static struct ceph_mds_request *__get_oldest_req(struct ceph_mds_client *mdsc)
1697 struct ceph_mds_request, r_node); 1731 struct ceph_mds_request, r_node);
1698} 1732}
1699 1733
1700static u64 __get_oldest_tid(struct ceph_mds_client *mdsc) 1734static inline u64 __get_oldest_tid(struct ceph_mds_client *mdsc)
1701{ 1735{
1702 struct ceph_mds_request *req = __get_oldest_req(mdsc); 1736 return mdsc->oldest_tid;
1703
1704 if (req)
1705 return req->r_tid;
1706 return 0;
1707} 1737}
1708 1738
1709/* 1739/*
@@ -2267,15 +2297,18 @@ int ceph_mdsc_do_request(struct ceph_mds_client *mdsc,
2267 /* wait */ 2297 /* wait */
2268 mutex_unlock(&mdsc->mutex); 2298 mutex_unlock(&mdsc->mutex);
2269 dout("do_request waiting\n"); 2299 dout("do_request waiting\n");
2270 if (req->r_timeout) { 2300 if (!req->r_timeout && req->r_wait_for_completion) {
2271 err = (long)wait_for_completion_killable_timeout(
2272 &req->r_completion, req->r_timeout);
2273 if (err == 0)
2274 err = -EIO;
2275 } else if (req->r_wait_for_completion) {
2276 err = req->r_wait_for_completion(mdsc, req); 2301 err = req->r_wait_for_completion(mdsc, req);
2277 } else { 2302 } else {
2278 err = wait_for_completion_killable(&req->r_completion); 2303 long timeleft = wait_for_completion_killable_timeout(
2304 &req->r_completion,
2305 ceph_timeout_jiffies(req->r_timeout));
2306 if (timeleft > 0)
2307 err = 0;
2308 else if (!timeleft)
2309 err = -EIO; /* timed out */
2310 else
2311 err = timeleft; /* killed */
2279 } 2312 }
2280 dout("do_request waited, got %d\n", err); 2313 dout("do_request waited, got %d\n", err);
2281 mutex_lock(&mdsc->mutex); 2314 mutex_lock(&mdsc->mutex);
@@ -2496,7 +2529,6 @@ out_err:
2496 } 2529 }
2497 mutex_unlock(&mdsc->mutex); 2530 mutex_unlock(&mdsc->mutex);
2498 2531
2499 ceph_add_cap_releases(mdsc, req->r_session);
2500 mutex_unlock(&session->s_mutex); 2532 mutex_unlock(&session->s_mutex);
2501 2533
2502 /* kick calling process */ 2534 /* kick calling process */
@@ -2888,8 +2920,7 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2888 */ 2920 */
2889 session->s_cap_reconnect = 1; 2921 session->s_cap_reconnect = 1;
2890 /* drop old cap expires; we're about to reestablish that state */ 2922 /* drop old cap expires; we're about to reestablish that state */
2891 discard_cap_releases(mdsc, session); 2923 cleanup_cap_releases(mdsc, session);
2892 spin_unlock(&session->s_cap_lock);
2893 2924
2894 /* trim unused caps to reduce MDS's cache rejoin time */ 2925 /* trim unused caps to reduce MDS's cache rejoin time */
2895 if (mdsc->fsc->sb->s_root) 2926 if (mdsc->fsc->sb->s_root)
@@ -2956,6 +2987,9 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
2956 2987
2957 reply->hdr.data_len = cpu_to_le32(pagelist->length); 2988 reply->hdr.data_len = cpu_to_le32(pagelist->length);
2958 ceph_msg_data_add_pagelist(reply, pagelist); 2989 ceph_msg_data_add_pagelist(reply, pagelist);
2990
2991 ceph_early_kick_flushing_caps(mdsc, session);
2992
2959 ceph_con_send(&session->s_con, reply); 2993 ceph_con_send(&session->s_con, reply);
2960 2994
2961 mutex_unlock(&session->s_mutex); 2995 mutex_unlock(&session->s_mutex);
@@ -3352,7 +3386,6 @@ static void delayed_work(struct work_struct *work)
3352 send_renew_caps(mdsc, s); 3386 send_renew_caps(mdsc, s);
3353 else 3387 else
3354 ceph_con_keepalive(&s->s_con); 3388 ceph_con_keepalive(&s->s_con);
3355 ceph_add_cap_releases(mdsc, s);
3356 if (s->s_state == CEPH_MDS_SESSION_OPEN || 3389 if (s->s_state == CEPH_MDS_SESSION_OPEN ||
3357 s->s_state == CEPH_MDS_SESSION_HUNG) 3390 s->s_state == CEPH_MDS_SESSION_HUNG)
3358 ceph_send_cap_releases(mdsc, s); 3391 ceph_send_cap_releases(mdsc, s);
@@ -3390,11 +3423,13 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
3390 atomic_set(&mdsc->num_sessions, 0); 3423 atomic_set(&mdsc->num_sessions, 0);
3391 mdsc->max_sessions = 0; 3424 mdsc->max_sessions = 0;
3392 mdsc->stopping = 0; 3425 mdsc->stopping = 0;
3426 mdsc->last_snap_seq = 0;
3393 init_rwsem(&mdsc->snap_rwsem); 3427 init_rwsem(&mdsc->snap_rwsem);
3394 mdsc->snap_realms = RB_ROOT; 3428 mdsc->snap_realms = RB_ROOT;
3395 INIT_LIST_HEAD(&mdsc->snap_empty); 3429 INIT_LIST_HEAD(&mdsc->snap_empty);
3396 spin_lock_init(&mdsc->snap_empty_lock); 3430 spin_lock_init(&mdsc->snap_empty_lock);
3397 mdsc->last_tid = 0; 3431 mdsc->last_tid = 0;
3432 mdsc->oldest_tid = 0;
3398 mdsc->request_tree = RB_ROOT; 3433 mdsc->request_tree = RB_ROOT;
3399 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work); 3434 INIT_DELAYED_WORK(&mdsc->delayed_work, delayed_work);
3400 mdsc->last_renew_caps = jiffies; 3435 mdsc->last_renew_caps = jiffies;
@@ -3402,7 +3437,8 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
3402 spin_lock_init(&mdsc->cap_delay_lock); 3437 spin_lock_init(&mdsc->cap_delay_lock);
3403 INIT_LIST_HEAD(&mdsc->snap_flush_list); 3438 INIT_LIST_HEAD(&mdsc->snap_flush_list);
3404 spin_lock_init(&mdsc->snap_flush_lock); 3439 spin_lock_init(&mdsc->snap_flush_lock);
3405 mdsc->cap_flush_seq = 0; 3440 mdsc->last_cap_flush_tid = 1;
3441 mdsc->cap_flush_tree = RB_ROOT;
3406 INIT_LIST_HEAD(&mdsc->cap_dirty); 3442 INIT_LIST_HEAD(&mdsc->cap_dirty);
3407 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating); 3443 INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
3408 mdsc->num_cap_flushing = 0; 3444 mdsc->num_cap_flushing = 0;
@@ -3414,6 +3450,9 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
3414 ceph_caps_init(mdsc); 3450 ceph_caps_init(mdsc);
3415 ceph_adjust_min_caps(mdsc, fsc->min_caps); 3451 ceph_adjust_min_caps(mdsc, fsc->min_caps);
3416 3452
3453 init_rwsem(&mdsc->pool_perm_rwsem);
3454 mdsc->pool_perm_tree = RB_ROOT;
3455
3417 return 0; 3456 return 0;
3418} 3457}
3419 3458
@@ -3423,8 +3462,8 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
3423 */ 3462 */
3424static void wait_requests(struct ceph_mds_client *mdsc) 3463static void wait_requests(struct ceph_mds_client *mdsc)
3425{ 3464{
3465 struct ceph_options *opts = mdsc->fsc->client->options;
3426 struct ceph_mds_request *req; 3466 struct ceph_mds_request *req;
3427 struct ceph_fs_client *fsc = mdsc->fsc;
3428 3467
3429 mutex_lock(&mdsc->mutex); 3468 mutex_lock(&mdsc->mutex);
3430 if (__get_oldest_req(mdsc)) { 3469 if (__get_oldest_req(mdsc)) {
@@ -3432,7 +3471,7 @@ static void wait_requests(struct ceph_mds_client *mdsc)
3432 3471
3433 dout("wait_requests waiting for requests\n"); 3472 dout("wait_requests waiting for requests\n");
3434 wait_for_completion_timeout(&mdsc->safe_umount_waiters, 3473 wait_for_completion_timeout(&mdsc->safe_umount_waiters,
3435 fsc->client->options->mount_timeout * HZ); 3474 ceph_timeout_jiffies(opts->mount_timeout));
3436 3475
3437 /* tear down remaining requests */ 3476 /* tear down remaining requests */
3438 mutex_lock(&mdsc->mutex); 3477 mutex_lock(&mdsc->mutex);
@@ -3485,7 +3524,8 @@ restart:
3485 nextreq = rb_entry(n, struct ceph_mds_request, r_node); 3524 nextreq = rb_entry(n, struct ceph_mds_request, r_node);
3486 else 3525 else
3487 nextreq = NULL; 3526 nextreq = NULL;
3488 if ((req->r_op & CEPH_MDS_OP_WRITE)) { 3527 if (req->r_op != CEPH_MDS_OP_SETFILELOCK &&
3528 (req->r_op & CEPH_MDS_OP_WRITE)) {
3489 /* write op */ 3529 /* write op */
3490 ceph_mdsc_get_request(req); 3530 ceph_mdsc_get_request(req);
3491 if (nextreq) 3531 if (nextreq)
@@ -3513,7 +3553,7 @@ restart:
3513 3553
3514void ceph_mdsc_sync(struct ceph_mds_client *mdsc) 3554void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3515{ 3555{
3516 u64 want_tid, want_flush; 3556 u64 want_tid, want_flush, want_snap;
3517 3557
3518 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN) 3558 if (mdsc->fsc->mount_state == CEPH_MOUNT_SHUTDOWN)
3519 return; 3559 return;
@@ -3525,13 +3565,18 @@ void ceph_mdsc_sync(struct ceph_mds_client *mdsc)
3525 3565
3526 ceph_flush_dirty_caps(mdsc); 3566 ceph_flush_dirty_caps(mdsc);
3527 spin_lock(&mdsc->cap_dirty_lock); 3567 spin_lock(&mdsc->cap_dirty_lock);
3528 want_flush = mdsc->cap_flush_seq; 3568 want_flush = mdsc->last_cap_flush_tid;
3529 spin_unlock(&mdsc->cap_dirty_lock); 3569 spin_unlock(&mdsc->cap_dirty_lock);
3530 3570
3531 dout("sync want tid %lld flush_seq %lld\n", want_tid, want_flush); 3571 down_read(&mdsc->snap_rwsem);
3572 want_snap = mdsc->last_snap_seq;
3573 up_read(&mdsc->snap_rwsem);
3574
3575 dout("sync want tid %lld flush_seq %lld snap_seq %lld\n",
3576 want_tid, want_flush, want_snap);
3532 3577
3533 wait_unsafe_requests(mdsc, want_tid); 3578 wait_unsafe_requests(mdsc, want_tid);
3534 wait_caps_flush(mdsc, want_flush); 3579 wait_caps_flush(mdsc, want_flush, want_snap);
3535} 3580}
3536 3581
3537/* 3582/*
@@ -3549,10 +3594,9 @@ static bool done_closing_sessions(struct ceph_mds_client *mdsc)
3549 */ 3594 */
3550void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc) 3595void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3551{ 3596{
3597 struct ceph_options *opts = mdsc->fsc->client->options;
3552 struct ceph_mds_session *session; 3598 struct ceph_mds_session *session;
3553 int i; 3599 int i;
3554 struct ceph_fs_client *fsc = mdsc->fsc;
3555 unsigned long timeout = fsc->client->options->mount_timeout * HZ;
3556 3600
3557 dout("close_sessions\n"); 3601 dout("close_sessions\n");
3558 3602
@@ -3573,7 +3617,7 @@ void ceph_mdsc_close_sessions(struct ceph_mds_client *mdsc)
3573 3617
3574 dout("waiting for sessions to close\n"); 3618 dout("waiting for sessions to close\n");
3575 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc), 3619 wait_event_timeout(mdsc->session_close_wq, done_closing_sessions(mdsc),
3576 timeout); 3620 ceph_timeout_jiffies(opts->mount_timeout));
3577 3621
3578 /* tear down remaining sessions */ 3622 /* tear down remaining sessions */
3579 mutex_lock(&mdsc->mutex); 3623 mutex_lock(&mdsc->mutex);
@@ -3607,6 +3651,7 @@ static void ceph_mdsc_stop(struct ceph_mds_client *mdsc)
3607 ceph_mdsmap_destroy(mdsc->mdsmap); 3651 ceph_mdsmap_destroy(mdsc->mdsmap);
3608 kfree(mdsc->sessions); 3652 kfree(mdsc->sessions);
3609 ceph_caps_finalize(mdsc); 3653 ceph_caps_finalize(mdsc);
3654 ceph_pool_perm_destroy(mdsc);
3610} 3655}
3611 3656
3612void ceph_mdsc_destroy(struct ceph_fs_client *fsc) 3657void ceph_mdsc_destroy(struct ceph_fs_client *fsc)
diff --git a/fs/ceph/mds_client.h b/fs/ceph/mds_client.h
index 1875b5d985c6..762757e6cebf 100644
--- a/fs/ceph/mds_client.h
+++ b/fs/ceph/mds_client.h
@@ -139,7 +139,6 @@ struct ceph_mds_session {
139 int s_cap_reconnect; 139 int s_cap_reconnect;
140 int s_readonly; 140 int s_readonly;
141 struct list_head s_cap_releases; /* waiting cap_release messages */ 141 struct list_head s_cap_releases; /* waiting cap_release messages */
142 struct list_head s_cap_releases_done; /* ready to send */
143 struct ceph_cap *s_cap_iterator; 142 struct ceph_cap *s_cap_iterator;
144 143
145 /* protected by mutex */ 144 /* protected by mutex */
@@ -228,7 +227,7 @@ struct ceph_mds_request {
228 int r_err; 227 int r_err;
229 bool r_aborted; 228 bool r_aborted;
230 229
231 unsigned long r_timeout; /* optional. jiffies */ 230 unsigned long r_timeout; /* optional. jiffies, 0 is "wait forever" */
232 unsigned long r_started; /* start time to measure timeout against */ 231 unsigned long r_started; /* start time to measure timeout against */
233 unsigned long r_request_started; /* start time for mds request only, 232 unsigned long r_request_started; /* start time for mds request only,
234 used to measure lease durations */ 233 used to measure lease durations */
@@ -254,12 +253,21 @@ struct ceph_mds_request {
254 bool r_got_unsafe, r_got_safe, r_got_result; 253 bool r_got_unsafe, r_got_safe, r_got_result;
255 254
256 bool r_did_prepopulate; 255 bool r_did_prepopulate;
256 long long r_dir_release_cnt;
257 long long r_dir_ordered_cnt;
258 int r_readdir_cache_idx;
257 u32 r_readdir_offset; 259 u32 r_readdir_offset;
258 260
259 struct ceph_cap_reservation r_caps_reservation; 261 struct ceph_cap_reservation r_caps_reservation;
260 int r_num_caps; 262 int r_num_caps;
261}; 263};
262 264
265struct ceph_pool_perm {
266 struct rb_node node;
267 u32 pool;
268 int perm;
269};
270
263/* 271/*
264 * mds client state 272 * mds client state
265 */ 273 */
@@ -284,12 +292,15 @@ struct ceph_mds_client {
284 * references (implying they contain no inodes with caps) that 292 * references (implying they contain no inodes with caps) that
285 * should be destroyed. 293 * should be destroyed.
286 */ 294 */
295 u64 last_snap_seq;
287 struct rw_semaphore snap_rwsem; 296 struct rw_semaphore snap_rwsem;
288 struct rb_root snap_realms; 297 struct rb_root snap_realms;
289 struct list_head snap_empty; 298 struct list_head snap_empty;
290 spinlock_t snap_empty_lock; /* protect snap_empty */ 299 spinlock_t snap_empty_lock; /* protect snap_empty */
291 300
292 u64 last_tid; /* most recent mds request */ 301 u64 last_tid; /* most recent mds request */
302 u64 oldest_tid; /* oldest incomplete mds request,
303 excluding setfilelock requests */
293 struct rb_root request_tree; /* pending mds requests */ 304 struct rb_root request_tree; /* pending mds requests */
294 struct delayed_work delayed_work; /* delayed work */ 305 struct delayed_work delayed_work; /* delayed work */
295 unsigned long last_renew_caps; /* last time we renewed our caps */ 306 unsigned long last_renew_caps; /* last time we renewed our caps */
@@ -298,7 +309,8 @@ struct ceph_mds_client {
298 struct list_head snap_flush_list; /* cap_snaps ready to flush */ 309 struct list_head snap_flush_list; /* cap_snaps ready to flush */
299 spinlock_t snap_flush_lock; 310 spinlock_t snap_flush_lock;
300 311
301 u64 cap_flush_seq; 312 u64 last_cap_flush_tid;
313 struct rb_root cap_flush_tree;
302 struct list_head cap_dirty; /* inodes with dirty caps */ 314 struct list_head cap_dirty; /* inodes with dirty caps */
303 struct list_head cap_dirty_migrating; /* ...that are migration... */ 315 struct list_head cap_dirty_migrating; /* ...that are migration... */
304 int num_cap_flushing; /* # caps we are flushing */ 316 int num_cap_flushing; /* # caps we are flushing */
@@ -328,6 +340,9 @@ struct ceph_mds_client {
328 spinlock_t dentry_lru_lock; 340 spinlock_t dentry_lru_lock;
329 struct list_head dentry_lru; 341 struct list_head dentry_lru;
330 int num_dentry; 342 int num_dentry;
343
344 struct rw_semaphore pool_perm_rwsem;
345 struct rb_root pool_perm_tree;
331}; 346};
332 347
333extern const char *ceph_mds_op_name(int op); 348extern const char *ceph_mds_op_name(int op);
@@ -379,8 +394,6 @@ static inline void ceph_mdsc_put_request(struct ceph_mds_request *req)
379 kref_put(&req->r_kref, ceph_mdsc_release_request); 394 kref_put(&req->r_kref, ceph_mdsc_release_request);
380} 395}
381 396
382extern int ceph_add_cap_releases(struct ceph_mds_client *mdsc,
383 struct ceph_mds_session *session);
384extern void ceph_send_cap_releases(struct ceph_mds_client *mdsc, 397extern void ceph_send_cap_releases(struct ceph_mds_client *mdsc,
385 struct ceph_mds_session *session); 398 struct ceph_mds_session *session);
386 399
diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c
index a97e39f09ba6..233d906aec02 100644
--- a/fs/ceph/snap.c
+++ b/fs/ceph/snap.c
@@ -296,7 +296,7 @@ static int cmpu64_rev(const void *a, const void *b)
296} 296}
297 297
298 298
299static struct ceph_snap_context *empty_snapc; 299struct ceph_snap_context *ceph_empty_snapc;
300 300
301/* 301/*
302 * build the snap context for a given realm. 302 * build the snap context for a given realm.
@@ -338,9 +338,9 @@ static int build_snap_context(struct ceph_snap_realm *realm)
338 return 0; 338 return 0;
339 } 339 }
340 340
341 if (num == 0 && realm->seq == empty_snapc->seq) { 341 if (num == 0 && realm->seq == ceph_empty_snapc->seq) {
342 ceph_get_snap_context(empty_snapc); 342 ceph_get_snap_context(ceph_empty_snapc);
343 snapc = empty_snapc; 343 snapc = ceph_empty_snapc;
344 goto done; 344 goto done;
345 } 345 }
346 346
@@ -436,6 +436,14 @@ static int dup_array(u64 **dst, __le64 *src, u32 num)
436 return 0; 436 return 0;
437} 437}
438 438
439static bool has_new_snaps(struct ceph_snap_context *o,
440 struct ceph_snap_context *n)
441{
442 if (n->num_snaps == 0)
443 return false;
444 /* snaps are in descending order */
445 return n->snaps[0] > o->seq;
446}
439 447
440/* 448/*
441 * When a snapshot is applied, the size/mtime inode metadata is queued 449 * When a snapshot is applied, the size/mtime inode metadata is queued
@@ -455,6 +463,7 @@ void ceph_queue_cap_snap(struct ceph_inode_info *ci)
455{ 463{
456 struct inode *inode = &ci->vfs_inode; 464 struct inode *inode = &ci->vfs_inode;
457 struct ceph_cap_snap *capsnap; 465 struct ceph_cap_snap *capsnap;
466 struct ceph_snap_context *old_snapc, *new_snapc;
458 int used, dirty; 467 int used, dirty;
459 468
460 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS); 469 capsnap = kzalloc(sizeof(*capsnap), GFP_NOFS);
@@ -467,6 +476,9 @@ void ceph_queue_cap_snap(struct ceph_inode_info *ci)
467 used = __ceph_caps_used(ci); 476 used = __ceph_caps_used(ci);
468 dirty = __ceph_caps_dirty(ci); 477 dirty = __ceph_caps_dirty(ci);
469 478
479 old_snapc = ci->i_head_snapc;
480 new_snapc = ci->i_snap_realm->cached_context;
481
470 /* 482 /*
471 * If there is a write in progress, treat that as a dirty Fw, 483 * If there is a write in progress, treat that as a dirty Fw,
472 * even though it hasn't completed yet; by the time we finish 484 * even though it hasn't completed yet; by the time we finish
@@ -481,76 +493,95 @@ void ceph_queue_cap_snap(struct ceph_inode_info *ci)
481 writes in progress now were started before the previous 493 writes in progress now were started before the previous
482 cap_snap. lucky us. */ 494 cap_snap. lucky us. */
483 dout("queue_cap_snap %p already pending\n", inode); 495 dout("queue_cap_snap %p already pending\n", inode);
484 kfree(capsnap); 496 goto update_snapc;
485 } else if (ci->i_snap_realm->cached_context == empty_snapc) { 497 }
486 dout("queue_cap_snap %p empty snapc\n", inode); 498 if (ci->i_wrbuffer_ref_head == 0 &&
487 kfree(capsnap); 499 !(dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))) {
488 } else if (dirty & (CEPH_CAP_AUTH_EXCL|CEPH_CAP_XATTR_EXCL| 500 dout("queue_cap_snap %p nothing dirty|writing\n", inode);
489 CEPH_CAP_FILE_EXCL|CEPH_CAP_FILE_WR)) { 501 goto update_snapc;
490 struct ceph_snap_context *snapc = ci->i_head_snapc; 502 }
491
492 /*
493 * if we are a sync write, we may need to go to the snaprealm
494 * to get the current snapc.
495 */
496 if (!snapc)
497 snapc = ci->i_snap_realm->cached_context;
498 503
499 dout("queue_cap_snap %p cap_snap %p queuing under %p %s\n", 504 BUG_ON(!old_snapc);
500 inode, capsnap, snapc, ceph_cap_string(dirty));
501 ihold(inode);
502 505
503 atomic_set(&capsnap->nref, 1); 506 /*
504 capsnap->ci = ci; 507 * There is no need to send FLUSHSNAP message to MDS if there is
505 INIT_LIST_HEAD(&capsnap->ci_item); 508 * no new snapshot. But when there is dirty pages or on-going
506 INIT_LIST_HEAD(&capsnap->flushing_item); 509 * writes, we still need to create cap_snap. cap_snap is needed
507 510 * by the write path and page writeback path.
508 capsnap->follows = snapc->seq; 511 *
509 capsnap->issued = __ceph_caps_issued(ci, NULL); 512 * also see ceph_try_drop_cap_snap()
510 capsnap->dirty = dirty; 513 */
511 514 if (has_new_snaps(old_snapc, new_snapc)) {
512 capsnap->mode = inode->i_mode; 515 if (dirty & (CEPH_CAP_ANY_EXCL|CEPH_CAP_FILE_WR))
513 capsnap->uid = inode->i_uid; 516 capsnap->need_flush = true;
514 capsnap->gid = inode->i_gid; 517 } else {
515 518 if (!(used & CEPH_CAP_FILE_WR) &&
516 if (dirty & CEPH_CAP_XATTR_EXCL) { 519 ci->i_wrbuffer_ref_head == 0) {
517 __ceph_build_xattrs_blob(ci); 520 dout("queue_cap_snap %p "
518 capsnap->xattr_blob = 521 "no new_snap|dirty_page|writing\n", inode);
519 ceph_buffer_get(ci->i_xattrs.blob); 522 goto update_snapc;
520 capsnap->xattr_version = ci->i_xattrs.version;
521 } else {
522 capsnap->xattr_blob = NULL;
523 capsnap->xattr_version = 0;
524 } 523 }
524 }
525 525
526 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE; 526 dout("queue_cap_snap %p cap_snap %p queuing under %p %s %s\n",
527 527 inode, capsnap, old_snapc, ceph_cap_string(dirty),
528 /* dirty page count moved from _head to this cap_snap; 528 capsnap->need_flush ? "" : "no_flush");
529 all subsequent writes page dirties occur _after_ this 529 ihold(inode);
530 snapshot. */ 530
531 capsnap->dirty_pages = ci->i_wrbuffer_ref_head; 531 atomic_set(&capsnap->nref, 1);
532 ci->i_wrbuffer_ref_head = 0; 532 capsnap->ci = ci;
533 capsnap->context = snapc; 533 INIT_LIST_HEAD(&capsnap->ci_item);
534 ci->i_head_snapc = 534 INIT_LIST_HEAD(&capsnap->flushing_item);
535 ceph_get_snap_context(ci->i_snap_realm->cached_context); 535
536 dout(" new snapc is %p\n", ci->i_head_snapc); 536 capsnap->follows = old_snapc->seq;
537 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps); 537 capsnap->issued = __ceph_caps_issued(ci, NULL);
538 538 capsnap->dirty = dirty;
539 if (used & CEPH_CAP_FILE_WR) { 539
540 dout("queue_cap_snap %p cap_snap %p snapc %p" 540 capsnap->mode = inode->i_mode;
541 " seq %llu used WR, now pending\n", inode, 541 capsnap->uid = inode->i_uid;
542 capsnap, snapc, snapc->seq); 542 capsnap->gid = inode->i_gid;
543 capsnap->writing = 1; 543
544 } else { 544 if (dirty & CEPH_CAP_XATTR_EXCL) {
545 /* note mtime, size NOW. */ 545 __ceph_build_xattrs_blob(ci);
546 __ceph_finish_cap_snap(ci, capsnap); 546 capsnap->xattr_blob =
547 } 547 ceph_buffer_get(ci->i_xattrs.blob);
548 capsnap->xattr_version = ci->i_xattrs.version;
548 } else { 549 } else {
549 dout("queue_cap_snap %p nothing dirty|writing\n", inode); 550 capsnap->xattr_blob = NULL;
550 kfree(capsnap); 551 capsnap->xattr_version = 0;
551 } 552 }
552 553
554 capsnap->inline_data = ci->i_inline_version != CEPH_INLINE_NONE;
555
556 /* dirty page count moved from _head to this cap_snap;
557 all subsequent writes page dirties occur _after_ this
558 snapshot. */
559 capsnap->dirty_pages = ci->i_wrbuffer_ref_head;
560 ci->i_wrbuffer_ref_head = 0;
561 capsnap->context = old_snapc;
562 list_add_tail(&capsnap->ci_item, &ci->i_cap_snaps);
563 old_snapc = NULL;
564
565 if (used & CEPH_CAP_FILE_WR) {
566 dout("queue_cap_snap %p cap_snap %p snapc %p"
567 " seq %llu used WR, now pending\n", inode,
568 capsnap, old_snapc, old_snapc->seq);
569 capsnap->writing = 1;
570 } else {
571 /* note mtime, size NOW. */
572 __ceph_finish_cap_snap(ci, capsnap);
573 }
574 capsnap = NULL;
575
576update_snapc:
577 if (ci->i_head_snapc) {
578 ci->i_head_snapc = ceph_get_snap_context(new_snapc);
579 dout(" new snapc is %p\n", new_snapc);
580 }
553 spin_unlock(&ci->i_ceph_lock); 581 spin_unlock(&ci->i_ceph_lock);
582
583 kfree(capsnap);
584 ceph_put_snap_context(old_snapc);
554} 585}
555 586
556/* 587/*
@@ -699,6 +730,8 @@ more:
699 730
700 /* queue realm for cap_snap creation */ 731 /* queue realm for cap_snap creation */
701 list_add(&realm->dirty_item, &dirty_realms); 732 list_add(&realm->dirty_item, &dirty_realms);
733 if (realm->seq > mdsc->last_snap_seq)
734 mdsc->last_snap_seq = realm->seq;
702 735
703 invalidate = 1; 736 invalidate = 1;
704 } else if (!realm->cached_context) { 737 } else if (!realm->cached_context) {
@@ -964,14 +997,14 @@ out:
964 997
965int __init ceph_snap_init(void) 998int __init ceph_snap_init(void)
966{ 999{
967 empty_snapc = ceph_create_snap_context(0, GFP_NOFS); 1000 ceph_empty_snapc = ceph_create_snap_context(0, GFP_NOFS);
968 if (!empty_snapc) 1001 if (!ceph_empty_snapc)
969 return -ENOMEM; 1002 return -ENOMEM;
970 empty_snapc->seq = 1; 1003 ceph_empty_snapc->seq = 1;
971 return 0; 1004 return 0;
972} 1005}
973 1006
974void ceph_snap_exit(void) 1007void ceph_snap_exit(void)
975{ 1008{
976 ceph_put_snap_context(empty_snapc); 1009 ceph_put_snap_context(ceph_empty_snapc);
977} 1010}
diff --git a/fs/ceph/super.c b/fs/ceph/super.c
index 4e9905374078..d1c833c321b9 100644
--- a/fs/ceph/super.c
+++ b/fs/ceph/super.c
@@ -134,10 +134,12 @@ enum {
134 Opt_noino32, 134 Opt_noino32,
135 Opt_fscache, 135 Opt_fscache,
136 Opt_nofscache, 136 Opt_nofscache,
137 Opt_poolperm,
138 Opt_nopoolperm,
137#ifdef CONFIG_CEPH_FS_POSIX_ACL 139#ifdef CONFIG_CEPH_FS_POSIX_ACL
138 Opt_acl, 140 Opt_acl,
139#endif 141#endif
140 Opt_noacl 142 Opt_noacl,
141}; 143};
142 144
143static match_table_t fsopt_tokens = { 145static match_table_t fsopt_tokens = {
@@ -165,6 +167,8 @@ static match_table_t fsopt_tokens = {
165 {Opt_noino32, "noino32"}, 167 {Opt_noino32, "noino32"},
166 {Opt_fscache, "fsc"}, 168 {Opt_fscache, "fsc"},
167 {Opt_nofscache, "nofsc"}, 169 {Opt_nofscache, "nofsc"},
170 {Opt_poolperm, "poolperm"},
171 {Opt_nopoolperm, "nopoolperm"},
168#ifdef CONFIG_CEPH_FS_POSIX_ACL 172#ifdef CONFIG_CEPH_FS_POSIX_ACL
169 {Opt_acl, "acl"}, 173 {Opt_acl, "acl"},
170#endif 174#endif
@@ -268,6 +272,13 @@ static int parse_fsopt_token(char *c, void *private)
268 case Opt_nofscache: 272 case Opt_nofscache:
269 fsopt->flags &= ~CEPH_MOUNT_OPT_FSCACHE; 273 fsopt->flags &= ~CEPH_MOUNT_OPT_FSCACHE;
270 break; 274 break;
275 case Opt_poolperm:
276 fsopt->flags &= ~CEPH_MOUNT_OPT_NOPOOLPERM;
277 printk ("pool perm");
278 break;
279 case Opt_nopoolperm:
280 fsopt->flags |= CEPH_MOUNT_OPT_NOPOOLPERM;
281 break;
271#ifdef CONFIG_CEPH_FS_POSIX_ACL 282#ifdef CONFIG_CEPH_FS_POSIX_ACL
272 case Opt_acl: 283 case Opt_acl:
273 fsopt->sb_flags |= MS_POSIXACL; 284 fsopt->sb_flags |= MS_POSIXACL;
@@ -436,6 +447,8 @@ static int ceph_show_options(struct seq_file *m, struct dentry *root)
436 seq_puts(m, ",nodcache"); 447 seq_puts(m, ",nodcache");
437 if (fsopt->flags & CEPH_MOUNT_OPT_FSCACHE) 448 if (fsopt->flags & CEPH_MOUNT_OPT_FSCACHE)
438 seq_puts(m, ",fsc"); 449 seq_puts(m, ",fsc");
450 if (fsopt->flags & CEPH_MOUNT_OPT_NOPOOLPERM)
451 seq_puts(m, ",nopoolperm");
439 452
440#ifdef CONFIG_CEPH_FS_POSIX_ACL 453#ifdef CONFIG_CEPH_FS_POSIX_ACL
441 if (fsopt->sb_flags & MS_POSIXACL) 454 if (fsopt->sb_flags & MS_POSIXACL)
@@ -609,6 +622,7 @@ static void destroy_fs_client(struct ceph_fs_client *fsc)
609 */ 622 */
610struct kmem_cache *ceph_inode_cachep; 623struct kmem_cache *ceph_inode_cachep;
611struct kmem_cache *ceph_cap_cachep; 624struct kmem_cache *ceph_cap_cachep;
625struct kmem_cache *ceph_cap_flush_cachep;
612struct kmem_cache *ceph_dentry_cachep; 626struct kmem_cache *ceph_dentry_cachep;
613struct kmem_cache *ceph_file_cachep; 627struct kmem_cache *ceph_file_cachep;
614 628
@@ -634,6 +648,10 @@ static int __init init_caches(void)
634 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD); 648 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD);
635 if (ceph_cap_cachep == NULL) 649 if (ceph_cap_cachep == NULL)
636 goto bad_cap; 650 goto bad_cap;
651 ceph_cap_flush_cachep = KMEM_CACHE(ceph_cap_flush,
652 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD);
653 if (ceph_cap_flush_cachep == NULL)
654 goto bad_cap_flush;
637 655
638 ceph_dentry_cachep = KMEM_CACHE(ceph_dentry_info, 656 ceph_dentry_cachep = KMEM_CACHE(ceph_dentry_info,
639 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD); 657 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD);
@@ -652,6 +670,8 @@ static int __init init_caches(void)
652bad_file: 670bad_file:
653 kmem_cache_destroy(ceph_dentry_cachep); 671 kmem_cache_destroy(ceph_dentry_cachep);
654bad_dentry: 672bad_dentry:
673 kmem_cache_destroy(ceph_cap_flush_cachep);
674bad_cap_flush:
655 kmem_cache_destroy(ceph_cap_cachep); 675 kmem_cache_destroy(ceph_cap_cachep);
656bad_cap: 676bad_cap:
657 kmem_cache_destroy(ceph_inode_cachep); 677 kmem_cache_destroy(ceph_inode_cachep);
@@ -668,6 +688,7 @@ static void destroy_caches(void)
668 688
669 kmem_cache_destroy(ceph_inode_cachep); 689 kmem_cache_destroy(ceph_inode_cachep);
670 kmem_cache_destroy(ceph_cap_cachep); 690 kmem_cache_destroy(ceph_cap_cachep);
691 kmem_cache_destroy(ceph_cap_flush_cachep);
671 kmem_cache_destroy(ceph_dentry_cachep); 692 kmem_cache_destroy(ceph_dentry_cachep);
672 kmem_cache_destroy(ceph_file_cachep); 693 kmem_cache_destroy(ceph_file_cachep);
673 694
@@ -729,7 +750,7 @@ static struct dentry *open_root_dentry(struct ceph_fs_client *fsc,
729 req->r_ino1.ino = CEPH_INO_ROOT; 750 req->r_ino1.ino = CEPH_INO_ROOT;
730 req->r_ino1.snap = CEPH_NOSNAP; 751 req->r_ino1.snap = CEPH_NOSNAP;
731 req->r_started = started; 752 req->r_started = started;
732 req->r_timeout = fsc->client->options->mount_timeout * HZ; 753 req->r_timeout = fsc->client->options->mount_timeout;
733 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INODE); 754 req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INODE);
734 req->r_num_caps = 2; 755 req->r_num_caps = 2;
735 err = ceph_mdsc_do_request(mdsc, NULL, req); 756 err = ceph_mdsc_do_request(mdsc, NULL, req);
diff --git a/fs/ceph/super.h b/fs/ceph/super.h
index fa20e1318939..860cc016e70d 100644
--- a/fs/ceph/super.h
+++ b/fs/ceph/super.h
@@ -35,6 +35,7 @@
35#define CEPH_MOUNT_OPT_INO32 (1<<8) /* 32 bit inos */ 35#define CEPH_MOUNT_OPT_INO32 (1<<8) /* 32 bit inos */
36#define CEPH_MOUNT_OPT_DCACHE (1<<9) /* use dcache for readdir etc */ 36#define CEPH_MOUNT_OPT_DCACHE (1<<9) /* use dcache for readdir etc */
37#define CEPH_MOUNT_OPT_FSCACHE (1<<10) /* use fscache */ 37#define CEPH_MOUNT_OPT_FSCACHE (1<<10) /* use fscache */
38#define CEPH_MOUNT_OPT_NOPOOLPERM (1<<11) /* no pool permission check */
38 39
39#define CEPH_MOUNT_OPT_DEFAULT (CEPH_MOUNT_OPT_RBYTES | \ 40#define CEPH_MOUNT_OPT_DEFAULT (CEPH_MOUNT_OPT_RBYTES | \
40 CEPH_MOUNT_OPT_DCACHE) 41 CEPH_MOUNT_OPT_DCACHE)
@@ -121,11 +122,21 @@ struct ceph_cap {
121 struct rb_node ci_node; /* per-ci cap tree */ 122 struct rb_node ci_node; /* per-ci cap tree */
122 struct ceph_mds_session *session; 123 struct ceph_mds_session *session;
123 struct list_head session_caps; /* per-session caplist */ 124 struct list_head session_caps; /* per-session caplist */
124 int mds;
125 u64 cap_id; /* unique cap id (mds provided) */ 125 u64 cap_id; /* unique cap id (mds provided) */
126 int issued; /* latest, from the mds */ 126 union {
127 int implemented; /* implemented superset of issued (for revocation) */ 127 /* in-use caps */
128 int mds_wanted; 128 struct {
129 int issued; /* latest, from the mds */
130 int implemented; /* implemented superset of
131 issued (for revocation) */
132 int mds, mds_wanted;
133 };
134 /* caps to release */
135 struct {
136 u64 cap_ino;
137 int queue_release;
138 };
139 };
129 u32 seq, issue_seq, mseq; 140 u32 seq, issue_seq, mseq;
130 u32 cap_gen; /* active/stale cycle */ 141 u32 cap_gen; /* active/stale cycle */
131 unsigned long last_used; 142 unsigned long last_used;
@@ -163,6 +174,7 @@ struct ceph_cap_snap {
163 int writing; /* a sync write is still in progress */ 174 int writing; /* a sync write is still in progress */
164 int dirty_pages; /* dirty pages awaiting writeback */ 175 int dirty_pages; /* dirty pages awaiting writeback */
165 bool inline_data; 176 bool inline_data;
177 bool need_flush;
166}; 178};
167 179
168static inline void ceph_put_cap_snap(struct ceph_cap_snap *capsnap) 180static inline void ceph_put_cap_snap(struct ceph_cap_snap *capsnap)
@@ -174,6 +186,17 @@ static inline void ceph_put_cap_snap(struct ceph_cap_snap *capsnap)
174 } 186 }
175} 187}
176 188
189struct ceph_cap_flush {
190 u64 tid;
191 int caps;
192 bool kick;
193 struct rb_node g_node; // global
194 union {
195 struct rb_node i_node; // inode
196 struct list_head list;
197 };
198};
199
177/* 200/*
178 * The frag tree describes how a directory is fragmented, potentially across 201 * The frag tree describes how a directory is fragmented, potentially across
179 * multiple metadata servers. It is also used to indicate points where 202 * multiple metadata servers. It is also used to indicate points where
@@ -259,9 +282,9 @@ struct ceph_inode_info {
259 u32 i_time_warp_seq; 282 u32 i_time_warp_seq;
260 283
261 unsigned i_ceph_flags; 284 unsigned i_ceph_flags;
262 int i_ordered_count; 285 atomic64_t i_release_count;
263 atomic_t i_release_count; 286 atomic64_t i_ordered_count;
264 atomic_t i_complete_count; 287 atomic64_t i_complete_seq[2];
265 288
266 struct ceph_dir_layout i_dir_layout; 289 struct ceph_dir_layout i_dir_layout;
267 struct ceph_file_layout i_layout; 290 struct ceph_file_layout i_layout;
@@ -283,11 +306,11 @@ struct ceph_inode_info {
283 struct ceph_cap *i_auth_cap; /* authoritative cap, if any */ 306 struct ceph_cap *i_auth_cap; /* authoritative cap, if any */
284 unsigned i_dirty_caps, i_flushing_caps; /* mask of dirtied fields */ 307 unsigned i_dirty_caps, i_flushing_caps; /* mask of dirtied fields */
285 struct list_head i_dirty_item, i_flushing_item; 308 struct list_head i_dirty_item, i_flushing_item;
286 u64 i_cap_flush_seq;
287 /* we need to track cap writeback on a per-cap-bit basis, to allow 309 /* we need to track cap writeback on a per-cap-bit basis, to allow
288 * overlapping, pipelined cap flushes to the mds. we can probably 310 * overlapping, pipelined cap flushes to the mds. we can probably
289 * reduce the tid to 8 bits if we're concerned about inode size. */ 311 * reduce the tid to 8 bits if we're concerned about inode size. */
290 u16 i_cap_flush_last_tid, i_cap_flush_tid[CEPH_CAP_BITS]; 312 struct ceph_cap_flush *i_prealloc_cap_flush;
313 struct rb_root i_cap_flush_tree;
291 wait_queue_head_t i_cap_wq; /* threads waiting on a capability */ 314 wait_queue_head_t i_cap_wq; /* threads waiting on a capability */
292 unsigned long i_hold_caps_min; /* jiffies */ 315 unsigned long i_hold_caps_min; /* jiffies */
293 unsigned long i_hold_caps_max; /* jiffies */ 316 unsigned long i_hold_caps_max; /* jiffies */
@@ -438,36 +461,46 @@ static inline struct inode *ceph_find_inode(struct super_block *sb,
438/* 461/*
439 * Ceph inode. 462 * Ceph inode.
440 */ 463 */
441#define CEPH_I_DIR_ORDERED 1 /* dentries in dir are ordered */ 464#define CEPH_I_DIR_ORDERED (1 << 0) /* dentries in dir are ordered */
442#define CEPH_I_NODELAY 4 /* do not delay cap release */ 465#define CEPH_I_NODELAY (1 << 1) /* do not delay cap release */
443#define CEPH_I_FLUSH 8 /* do not delay flush of dirty metadata */ 466#define CEPH_I_FLUSH (1 << 2) /* do not delay flush of dirty metadata */
444#define CEPH_I_NOFLUSH 16 /* do not flush dirty caps */ 467#define CEPH_I_NOFLUSH (1 << 3) /* do not flush dirty caps */
468#define CEPH_I_POOL_PERM (1 << 4) /* pool rd/wr bits are valid */
469#define CEPH_I_POOL_RD (1 << 5) /* can read from pool */
470#define CEPH_I_POOL_WR (1 << 6) /* can write to pool */
471
445 472
446static inline void __ceph_dir_set_complete(struct ceph_inode_info *ci, 473static inline void __ceph_dir_set_complete(struct ceph_inode_info *ci,
447 int release_count, int ordered_count) 474 long long release_count,
475 long long ordered_count)
448{ 476{
449 atomic_set(&ci->i_complete_count, release_count); 477 smp_mb__before_atomic();
450 if (ci->i_ordered_count == ordered_count) 478 atomic64_set(&ci->i_complete_seq[0], release_count);
451 ci->i_ceph_flags |= CEPH_I_DIR_ORDERED; 479 atomic64_set(&ci->i_complete_seq[1], ordered_count);
452 else
453 ci->i_ceph_flags &= ~CEPH_I_DIR_ORDERED;
454} 480}
455 481
456static inline void __ceph_dir_clear_complete(struct ceph_inode_info *ci) 482static inline void __ceph_dir_clear_complete(struct ceph_inode_info *ci)
457{ 483{
458 atomic_inc(&ci->i_release_count); 484 atomic64_inc(&ci->i_release_count);
485}
486
487static inline void __ceph_dir_clear_ordered(struct ceph_inode_info *ci)
488{
489 atomic64_inc(&ci->i_ordered_count);
459} 490}
460 491
461static inline bool __ceph_dir_is_complete(struct ceph_inode_info *ci) 492static inline bool __ceph_dir_is_complete(struct ceph_inode_info *ci)
462{ 493{
463 return atomic_read(&ci->i_complete_count) == 494 return atomic64_read(&ci->i_complete_seq[0]) ==
464 atomic_read(&ci->i_release_count); 495 atomic64_read(&ci->i_release_count);
465} 496}
466 497
467static inline bool __ceph_dir_is_complete_ordered(struct ceph_inode_info *ci) 498static inline bool __ceph_dir_is_complete_ordered(struct ceph_inode_info *ci)
468{ 499{
469 return __ceph_dir_is_complete(ci) && 500 return atomic64_read(&ci->i_complete_seq[0]) ==
470 (ci->i_ceph_flags & CEPH_I_DIR_ORDERED); 501 atomic64_read(&ci->i_release_count) &&
502 atomic64_read(&ci->i_complete_seq[1]) ==
503 atomic64_read(&ci->i_ordered_count);
471} 504}
472 505
473static inline void ceph_dir_clear_complete(struct inode *inode) 506static inline void ceph_dir_clear_complete(struct inode *inode)
@@ -477,20 +510,13 @@ static inline void ceph_dir_clear_complete(struct inode *inode)
477 510
478static inline void ceph_dir_clear_ordered(struct inode *inode) 511static inline void ceph_dir_clear_ordered(struct inode *inode)
479{ 512{
480 struct ceph_inode_info *ci = ceph_inode(inode); 513 __ceph_dir_clear_ordered(ceph_inode(inode));
481 spin_lock(&ci->i_ceph_lock);
482 ci->i_ordered_count++;
483 ci->i_ceph_flags &= ~CEPH_I_DIR_ORDERED;
484 spin_unlock(&ci->i_ceph_lock);
485} 514}
486 515
487static inline bool ceph_dir_is_complete_ordered(struct inode *inode) 516static inline bool ceph_dir_is_complete_ordered(struct inode *inode)
488{ 517{
489 struct ceph_inode_info *ci = ceph_inode(inode); 518 bool ret = __ceph_dir_is_complete_ordered(ceph_inode(inode));
490 bool ret; 519 smp_rmb();
491 spin_lock(&ci->i_ceph_lock);
492 ret = __ceph_dir_is_complete_ordered(ci);
493 spin_unlock(&ci->i_ceph_lock);
494 return ret; 520 return ret;
495} 521}
496 522
@@ -552,7 +578,10 @@ static inline int __ceph_caps_dirty(struct ceph_inode_info *ci)
552{ 578{
553 return ci->i_dirty_caps | ci->i_flushing_caps; 579 return ci->i_dirty_caps | ci->i_flushing_caps;
554} 580}
555extern int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask); 581extern struct ceph_cap_flush *ceph_alloc_cap_flush(void);
582extern void ceph_free_cap_flush(struct ceph_cap_flush *cf);
583extern int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
584 struct ceph_cap_flush **pcf);
556 585
557extern int __ceph_caps_revoking_other(struct ceph_inode_info *ci, 586extern int __ceph_caps_revoking_other(struct ceph_inode_info *ci,
558 struct ceph_cap *ocap, int mask); 587 struct ceph_cap *ocap, int mask);
@@ -606,16 +635,20 @@ struct ceph_file_info {
606 unsigned offset; /* offset of last chunk, adjusted for . and .. */ 635 unsigned offset; /* offset of last chunk, adjusted for . and .. */
607 unsigned next_offset; /* offset of next chunk (last_name's + 1) */ 636 unsigned next_offset; /* offset of next chunk (last_name's + 1) */
608 char *last_name; /* last entry in previous chunk */ 637 char *last_name; /* last entry in previous chunk */
609 struct dentry *dentry; /* next dentry (for dcache readdir) */ 638 long long dir_release_count;
610 int dir_release_count; 639 long long dir_ordered_count;
611 int dir_ordered_count; 640 int readdir_cache_idx;
612 641
613 /* used for -o dirstat read() on directory thing */ 642 /* used for -o dirstat read() on directory thing */
614 char *dir_info; 643 char *dir_info;
615 int dir_info_len; 644 int dir_info_len;
616}; 645};
617 646
618 647struct ceph_readdir_cache_control {
648 struct page *page;
649 struct dentry **dentries;
650 int index;
651};
619 652
620/* 653/*
621 * A "snap realm" describes a subset of the file hierarchy sharing 654 * A "snap realm" describes a subset of the file hierarchy sharing
@@ -687,6 +720,7 @@ static inline int default_congestion_kb(void)
687 720
688 721
689/* snap.c */ 722/* snap.c */
723extern struct ceph_snap_context *ceph_empty_snapc;
690struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc, 724struct ceph_snap_realm *ceph_lookup_snap_realm(struct ceph_mds_client *mdsc,
691 u64 ino); 725 u64 ino);
692extern void ceph_get_snap_realm(struct ceph_mds_client *mdsc, 726extern void ceph_get_snap_realm(struct ceph_mds_client *mdsc,
@@ -713,8 +747,8 @@ extern void ceph_snap_exit(void);
713static inline bool __ceph_have_pending_cap_snap(struct ceph_inode_info *ci) 747static inline bool __ceph_have_pending_cap_snap(struct ceph_inode_info *ci)
714{ 748{
715 return !list_empty(&ci->i_cap_snaps) && 749 return !list_empty(&ci->i_cap_snaps) &&
716 list_entry(ci->i_cap_snaps.prev, struct ceph_cap_snap, 750 list_last_entry(&ci->i_cap_snaps, struct ceph_cap_snap,
717 ci_item)->writing; 751 ci_item)->writing;
718} 752}
719 753
720/* inode.c */ 754/* inode.c */
@@ -838,12 +872,12 @@ extern void ceph_put_cap(struct ceph_mds_client *mdsc,
838 struct ceph_cap *cap); 872 struct ceph_cap *cap);
839extern int ceph_is_any_caps(struct inode *inode); 873extern int ceph_is_any_caps(struct inode *inode);
840 874
841extern void __queue_cap_release(struct ceph_mds_session *session, u64 ino,
842 u64 cap_id, u32 migrate_seq, u32 issue_seq);
843extern void ceph_queue_caps_release(struct inode *inode); 875extern void ceph_queue_caps_release(struct inode *inode);
844extern int ceph_write_inode(struct inode *inode, struct writeback_control *wbc); 876extern int ceph_write_inode(struct inode *inode, struct writeback_control *wbc);
845extern int ceph_fsync(struct file *file, loff_t start, loff_t end, 877extern int ceph_fsync(struct file *file, loff_t start, loff_t end,
846 int datasync); 878 int datasync);
879extern void ceph_early_kick_flushing_caps(struct ceph_mds_client *mdsc,
880 struct ceph_mds_session *session);
847extern void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc, 881extern void ceph_kick_flushing_caps(struct ceph_mds_client *mdsc,
848 struct ceph_mds_session *session); 882 struct ceph_mds_session *session);
849extern struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci, 883extern struct ceph_cap *ceph_get_cap_for_mds(struct ceph_inode_info *ci,
@@ -879,6 +913,9 @@ extern void ceph_put_fmode(struct ceph_inode_info *ci, int mode);
879/* addr.c */ 913/* addr.c */
880extern const struct address_space_operations ceph_aops; 914extern const struct address_space_operations ceph_aops;
881extern int ceph_mmap(struct file *file, struct vm_area_struct *vma); 915extern int ceph_mmap(struct file *file, struct vm_area_struct *vma);
916extern int ceph_uninline_data(struct file *filp, struct page *locked_page);
917extern int ceph_pool_perm_check(struct ceph_inode_info *ci, int need);
918extern void ceph_pool_perm_destroy(struct ceph_mds_client* mdsc);
882 919
883/* file.c */ 920/* file.c */
884extern const struct file_operations ceph_file_fops; 921extern const struct file_operations ceph_file_fops;
@@ -890,7 +927,6 @@ extern int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
890extern int ceph_release(struct inode *inode, struct file *filp); 927extern int ceph_release(struct inode *inode, struct file *filp);
891extern void ceph_fill_inline_data(struct inode *inode, struct page *locked_page, 928extern void ceph_fill_inline_data(struct inode *inode, struct page *locked_page,
892 char *data, size_t len); 929 char *data, size_t len);
893int ceph_uninline_data(struct file *filp, struct page *locked_page);
894/* dir.c */ 930/* dir.c */
895extern const struct file_operations ceph_dir_fops; 931extern const struct file_operations ceph_dir_fops;
896extern const struct file_operations ceph_snapdir_fops; 932extern const struct file_operations ceph_snapdir_fops;
@@ -911,6 +947,7 @@ extern void ceph_dentry_lru_del(struct dentry *dn);
911extern void ceph_invalidate_dentry_lease(struct dentry *dentry); 947extern void ceph_invalidate_dentry_lease(struct dentry *dentry);
912extern unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn); 948extern unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn);
913extern struct inode *ceph_get_dentry_parent_inode(struct dentry *dentry); 949extern struct inode *ceph_get_dentry_parent_inode(struct dentry *dentry);
950extern void ceph_readdir_cache_release(struct ceph_readdir_cache_control *ctl);
914 951
915/* 952/*
916 * our d_ops vary depending on whether the inode is live, 953 * our d_ops vary depending on whether the inode is live,
diff --git a/fs/ceph/xattr.c b/fs/ceph/xattr.c
index cd7ffad4041d..819163d8313b 100644
--- a/fs/ceph/xattr.c
+++ b/fs/ceph/xattr.c
@@ -911,6 +911,8 @@ int __ceph_setxattr(struct dentry *dentry, const char *name,
911 struct inode *inode = d_inode(dentry); 911 struct inode *inode = d_inode(dentry);
912 struct ceph_vxattr *vxattr; 912 struct ceph_vxattr *vxattr;
913 struct ceph_inode_info *ci = ceph_inode(inode); 913 struct ceph_inode_info *ci = ceph_inode(inode);
914 struct ceph_mds_client *mdsc = ceph_sb_to_client(dentry->d_sb)->mdsc;
915 struct ceph_cap_flush *prealloc_cf = NULL;
914 int issued; 916 int issued;
915 int err; 917 int err;
916 int dirty = 0; 918 int dirty = 0;
@@ -920,6 +922,7 @@ int __ceph_setxattr(struct dentry *dentry, const char *name,
920 char *newval = NULL; 922 char *newval = NULL;
921 struct ceph_inode_xattr *xattr = NULL; 923 struct ceph_inode_xattr *xattr = NULL;
922 int required_blob_size; 924 int required_blob_size;
925 bool lock_snap_rwsem = false;
923 926
924 if (!ceph_is_valid_xattr(name)) 927 if (!ceph_is_valid_xattr(name))
925 return -EOPNOTSUPP; 928 return -EOPNOTSUPP;
@@ -948,12 +951,27 @@ int __ceph_setxattr(struct dentry *dentry, const char *name,
948 if (!xattr) 951 if (!xattr)
949 goto out; 952 goto out;
950 953
954 prealloc_cf = ceph_alloc_cap_flush();
955 if (!prealloc_cf)
956 goto out;
957
951 spin_lock(&ci->i_ceph_lock); 958 spin_lock(&ci->i_ceph_lock);
952retry: 959retry:
953 issued = __ceph_caps_issued(ci, NULL); 960 issued = __ceph_caps_issued(ci, NULL);
954 dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
955 if (ci->i_xattrs.version == 0 || !(issued & CEPH_CAP_XATTR_EXCL)) 961 if (ci->i_xattrs.version == 0 || !(issued & CEPH_CAP_XATTR_EXCL))
956 goto do_sync; 962 goto do_sync;
963
964 if (!lock_snap_rwsem && !ci->i_head_snapc) {
965 lock_snap_rwsem = true;
966 if (!down_read_trylock(&mdsc->snap_rwsem)) {
967 spin_unlock(&ci->i_ceph_lock);
968 down_read(&mdsc->snap_rwsem);
969 spin_lock(&ci->i_ceph_lock);
970 goto retry;
971 }
972 }
973
974 dout("setxattr %p issued %s\n", inode, ceph_cap_string(issued));
957 __build_xattrs(inode); 975 __build_xattrs(inode);
958 976
959 required_blob_size = __get_required_blob_size(ci, name_len, val_len); 977 required_blob_size = __get_required_blob_size(ci, name_len, val_len);
@@ -966,7 +984,7 @@ retry:
966 dout(" preaallocating new blob size=%d\n", required_blob_size); 984 dout(" preaallocating new blob size=%d\n", required_blob_size);
967 blob = ceph_buffer_new(required_blob_size, GFP_NOFS); 985 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
968 if (!blob) 986 if (!blob)
969 goto out; 987 goto do_sync_unlocked;
970 spin_lock(&ci->i_ceph_lock); 988 spin_lock(&ci->i_ceph_lock);
971 if (ci->i_xattrs.prealloc_blob) 989 if (ci->i_xattrs.prealloc_blob)
972 ceph_buffer_put(ci->i_xattrs.prealloc_blob); 990 ceph_buffer_put(ci->i_xattrs.prealloc_blob);
@@ -978,21 +996,28 @@ retry:
978 flags, value ? 1 : -1, &xattr); 996 flags, value ? 1 : -1, &xattr);
979 997
980 if (!err) { 998 if (!err) {
981 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL); 999 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL,
1000 &prealloc_cf);
982 ci->i_xattrs.dirty = true; 1001 ci->i_xattrs.dirty = true;
983 inode->i_ctime = CURRENT_TIME; 1002 inode->i_ctime = CURRENT_TIME;
984 } 1003 }
985 1004
986 spin_unlock(&ci->i_ceph_lock); 1005 spin_unlock(&ci->i_ceph_lock);
1006 if (lock_snap_rwsem)
1007 up_read(&mdsc->snap_rwsem);
987 if (dirty) 1008 if (dirty)
988 __mark_inode_dirty(inode, dirty); 1009 __mark_inode_dirty(inode, dirty);
1010 ceph_free_cap_flush(prealloc_cf);
989 return err; 1011 return err;
990 1012
991do_sync: 1013do_sync:
992 spin_unlock(&ci->i_ceph_lock); 1014 spin_unlock(&ci->i_ceph_lock);
993do_sync_unlocked: 1015do_sync_unlocked:
1016 if (lock_snap_rwsem)
1017 up_read(&mdsc->snap_rwsem);
994 err = ceph_sync_setxattr(dentry, name, value, size, flags); 1018 err = ceph_sync_setxattr(dentry, name, value, size, flags);
995out: 1019out:
1020 ceph_free_cap_flush(prealloc_cf);
996 kfree(newname); 1021 kfree(newname);
997 kfree(newval); 1022 kfree(newval);
998 kfree(xattr); 1023 kfree(xattr);
@@ -1044,10 +1069,13 @@ int __ceph_removexattr(struct dentry *dentry, const char *name)
1044 struct inode *inode = d_inode(dentry); 1069 struct inode *inode = d_inode(dentry);
1045 struct ceph_vxattr *vxattr; 1070 struct ceph_vxattr *vxattr;
1046 struct ceph_inode_info *ci = ceph_inode(inode); 1071 struct ceph_inode_info *ci = ceph_inode(inode);
1072 struct ceph_mds_client *mdsc = ceph_sb_to_client(dentry->d_sb)->mdsc;
1073 struct ceph_cap_flush *prealloc_cf = NULL;
1047 int issued; 1074 int issued;
1048 int err; 1075 int err;
1049 int required_blob_size; 1076 int required_blob_size;
1050 int dirty; 1077 int dirty;
1078 bool lock_snap_rwsem = false;
1051 1079
1052 if (!ceph_is_valid_xattr(name)) 1080 if (!ceph_is_valid_xattr(name))
1053 return -EOPNOTSUPP; 1081 return -EOPNOTSUPP;
@@ -1060,14 +1088,29 @@ int __ceph_removexattr(struct dentry *dentry, const char *name)
1060 if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN)) 1088 if (!strncmp(name, XATTR_CEPH_PREFIX, XATTR_CEPH_PREFIX_LEN))
1061 goto do_sync_unlocked; 1089 goto do_sync_unlocked;
1062 1090
1091 prealloc_cf = ceph_alloc_cap_flush();
1092 if (!prealloc_cf)
1093 return -ENOMEM;
1094
1063 err = -ENOMEM; 1095 err = -ENOMEM;
1064 spin_lock(&ci->i_ceph_lock); 1096 spin_lock(&ci->i_ceph_lock);
1065retry: 1097retry:
1066 issued = __ceph_caps_issued(ci, NULL); 1098 issued = __ceph_caps_issued(ci, NULL);
1067 dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
1068
1069 if (ci->i_xattrs.version == 0 || !(issued & CEPH_CAP_XATTR_EXCL)) 1099 if (ci->i_xattrs.version == 0 || !(issued & CEPH_CAP_XATTR_EXCL))
1070 goto do_sync; 1100 goto do_sync;
1101
1102 if (!lock_snap_rwsem && !ci->i_head_snapc) {
1103 lock_snap_rwsem = true;
1104 if (!down_read_trylock(&mdsc->snap_rwsem)) {
1105 spin_unlock(&ci->i_ceph_lock);
1106 down_read(&mdsc->snap_rwsem);
1107 spin_lock(&ci->i_ceph_lock);
1108 goto retry;
1109 }
1110 }
1111
1112 dout("removexattr %p issued %s\n", inode, ceph_cap_string(issued));
1113
1071 __build_xattrs(inode); 1114 __build_xattrs(inode);
1072 1115
1073 required_blob_size = __get_required_blob_size(ci, 0, 0); 1116 required_blob_size = __get_required_blob_size(ci, 0, 0);
@@ -1080,7 +1123,7 @@ retry:
1080 dout(" preaallocating new blob size=%d\n", required_blob_size); 1123 dout(" preaallocating new blob size=%d\n", required_blob_size);
1081 blob = ceph_buffer_new(required_blob_size, GFP_NOFS); 1124 blob = ceph_buffer_new(required_blob_size, GFP_NOFS);
1082 if (!blob) 1125 if (!blob)
1083 goto out; 1126 goto do_sync_unlocked;
1084 spin_lock(&ci->i_ceph_lock); 1127 spin_lock(&ci->i_ceph_lock);
1085 if (ci->i_xattrs.prealloc_blob) 1128 if (ci->i_xattrs.prealloc_blob)
1086 ceph_buffer_put(ci->i_xattrs.prealloc_blob); 1129 ceph_buffer_put(ci->i_xattrs.prealloc_blob);
@@ -1090,18 +1133,24 @@ retry:
1090 1133
1091 err = __remove_xattr_by_name(ceph_inode(inode), name); 1134 err = __remove_xattr_by_name(ceph_inode(inode), name);
1092 1135
1093 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL); 1136 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_XATTR_EXCL,
1137 &prealloc_cf);
1094 ci->i_xattrs.dirty = true; 1138 ci->i_xattrs.dirty = true;
1095 inode->i_ctime = CURRENT_TIME; 1139 inode->i_ctime = CURRENT_TIME;
1096 spin_unlock(&ci->i_ceph_lock); 1140 spin_unlock(&ci->i_ceph_lock);
1141 if (lock_snap_rwsem)
1142 up_read(&mdsc->snap_rwsem);
1097 if (dirty) 1143 if (dirty)
1098 __mark_inode_dirty(inode, dirty); 1144 __mark_inode_dirty(inode, dirty);
1145 ceph_free_cap_flush(prealloc_cf);
1099 return err; 1146 return err;
1100do_sync: 1147do_sync:
1101 spin_unlock(&ci->i_ceph_lock); 1148 spin_unlock(&ci->i_ceph_lock);
1102do_sync_unlocked: 1149do_sync_unlocked:
1150 if (lock_snap_rwsem)
1151 up_read(&mdsc->snap_rwsem);
1152 ceph_free_cap_flush(prealloc_cf);
1103 err = ceph_send_removexattr(dentry, name); 1153 err = ceph_send_removexattr(dentry, name);
1104out:
1105 return err; 1154 return err;
1106} 1155}
1107 1156
generated by cgit v1.2.2 (git 2.25.0) at 2025-04-13 17:08:22 -0400