aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorMark Fasheh <mark.fasheh@oracle.com>2007-02-09 23:24:12 -0500
committerMark Fasheh <mark.fasheh@oracle.com>2007-04-26 18:02:08 -0400
commit9517bac6cc7a7aa4fee63cb38a32cb6014e264c7 (patch)
tree3cac0c18d0cacc316e0e8a60f483282d6f991779
parent89488984ac23b0580f959b9ee549f2fcb1c2f194 (diff)
ocfs2: teach ocfs2_file_aio_write() about sparse files
Unfortunately, ocfs2 can no longer make use of generic_file_aio_write_nlock() because allocating writes will require zeroing of pages adjacent to the I/O for cluster sizes greater than page size. Implement a custom file write here, which can order page locks for zeroing. This also has the advantage that cluster locks can easily be ordered outside of the page locks. Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
-rw-r--r--fs/ocfs2/aops.c679
-rw-r--r--fs/ocfs2/aops.h38
-rw-r--r--fs/ocfs2/extent_map.c4
-rw-r--r--fs/ocfs2/extent_map.h2
-rw-r--r--fs/ocfs2/file.c374
-rw-r--r--fs/ocfs2/file.h4
-rw-r--r--fs/ocfs2/ocfs2.h32
7 files changed, 1076 insertions, 57 deletions
diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index f3b0cc5cba1a..5ffb3702b5e9 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -24,6 +24,7 @@
24#include <linux/highmem.h> 24#include <linux/highmem.h>
25#include <linux/pagemap.h> 25#include <linux/pagemap.h>
26#include <asm/byteorder.h> 26#include <asm/byteorder.h>
27#include <linux/swap.h>
27 28
28#define MLOG_MASK_PREFIX ML_FILE_IO 29#define MLOG_MASK_PREFIX ML_FILE_IO
29#include <cluster/masklog.h> 30#include <cluster/masklog.h>
@@ -37,6 +38,7 @@
37#include "file.h" 38#include "file.h"
38#include "inode.h" 39#include "inode.h"
39#include "journal.h" 40#include "journal.h"
41#include "suballoc.h"
40#include "super.h" 42#include "super.h"
41#include "symlink.h" 43#include "symlink.h"
42 44
@@ -645,23 +647,27 @@ static ssize_t ocfs2_direct_IO(int rw,
645 647
646 mlog_entry_void(); 648 mlog_entry_void();
647 649
648 /* 650 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
649 * We get PR data locks even for O_DIRECT. This allows 651 /*
650 * concurrent O_DIRECT I/O but doesn't let O_DIRECT with 652 * We get PR data locks even for O_DIRECT. This
651 * extending and buffered zeroing writes race. If they did 653 * allows concurrent O_DIRECT I/O but doesn't let
652 * race then the buffered zeroing could be written back after 654 * O_DIRECT with extending and buffered zeroing writes
653 * the O_DIRECT I/O. It's one thing to tell people not to mix 655 * race. If they did race then the buffered zeroing
654 * buffered and O_DIRECT writes, but expecting them to 656 * could be written back after the O_DIRECT I/O. It's
655 * understand that file extension is also an implicit buffered 657 * one thing to tell people not to mix buffered and
656 * write is too much. By getting the PR we force writeback of 658 * O_DIRECT writes, but expecting them to understand
657 * the buffered zeroing before proceeding. 659 * that file extension is also an implicit buffered
658 */ 660 * write is too much. By getting the PR we force
659 ret = ocfs2_data_lock(inode, 0); 661 * writeback of the buffered zeroing before
660 if (ret < 0) { 662 * proceeding.
661 mlog_errno(ret); 663 */
662 goto out; 664 ret = ocfs2_data_lock(inode, 0);
665 if (ret < 0) {
666 mlog_errno(ret);
667 goto out;
668 }
669 ocfs2_data_unlock(inode, 0);
663 } 670 }
664 ocfs2_data_unlock(inode, 0);
665 671
666 ret = blockdev_direct_IO_no_locking(rw, iocb, inode, 672 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
667 inode->i_sb->s_bdev, iov, offset, 673 inode->i_sb->s_bdev, iov, offset,
@@ -673,6 +679,647 @@ out:
673 return ret; 679 return ret;
674} 680}
675 681
682static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
683 u32 cpos,
684 unsigned int *start,
685 unsigned int *end)
686{
687 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
688
689 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
690 unsigned int cpp;
691
692 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
693
694 cluster_start = cpos % cpp;
695 cluster_start = cluster_start << osb->s_clustersize_bits;
696
697 cluster_end = cluster_start + osb->s_clustersize;
698 }
699
700 BUG_ON(cluster_start > PAGE_SIZE);
701 BUG_ON(cluster_end > PAGE_SIZE);
702
703 if (start)
704 *start = cluster_start;
705 if (end)
706 *end = cluster_end;
707}
708
709/*
710 * 'from' and 'to' are the region in the page to avoid zeroing.
711 *
712 * If pagesize > clustersize, this function will avoid zeroing outside
713 * of the cluster boundary.
714 *
715 * from == to == 0 is code for "zero the entire cluster region"
716 */
717static void ocfs2_clear_page_regions(struct page *page,
718 struct ocfs2_super *osb, u32 cpos,
719 unsigned from, unsigned to)
720{
721 void *kaddr;
722 unsigned int cluster_start, cluster_end;
723
724 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
725
726 kaddr = kmap_atomic(page, KM_USER0);
727
728 if (from || to) {
729 if (from > cluster_start)
730 memset(kaddr + cluster_start, 0, from - cluster_start);
731 if (to < cluster_end)
732 memset(kaddr + to, 0, cluster_end - to);
733 } else {
734 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
735 }
736
737 kunmap_atomic(kaddr, KM_USER0);
738}
739
740/*
741 * Some of this taken from block_prepare_write(). We already have our
742 * mapping by now though, and the entire write will be allocating or
743 * it won't, so not much need to use BH_New.
744 *
745 * This will also skip zeroing, which is handled externally.
746 */
747static int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
748 struct inode *inode, unsigned int from,
749 unsigned int to, int new)
750{
751 int ret = 0;
752 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
753 unsigned int block_end, block_start;
754 unsigned int bsize = 1 << inode->i_blkbits;
755
756 if (!page_has_buffers(page))
757 create_empty_buffers(page, bsize, 0);
758
759 head = page_buffers(page);
760 for (bh = head, block_start = 0; bh != head || !block_start;
761 bh = bh->b_this_page, block_start += bsize) {
762 block_end = block_start + bsize;
763
764 /*
765 * Ignore blocks outside of our i/o range -
766 * they may belong to unallocated clusters.
767 */
768 if (block_start >= to ||
769 (block_start + bsize) <= from) {
770 if (PageUptodate(page))
771 set_buffer_uptodate(bh);
772 continue;
773 }
774
775 /*
776 * For an allocating write with cluster size >= page
777 * size, we always write the entire page.
778 */
779
780 if (buffer_new(bh))
781 clear_buffer_new(bh);
782
783 if (!buffer_mapped(bh)) {
784 map_bh(bh, inode->i_sb, *p_blkno);
785 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
786 }
787
788 if (PageUptodate(page)) {
789 if (!buffer_uptodate(bh))
790 set_buffer_uptodate(bh);
791 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
792 (block_start < from || block_end > to)) {
793 ll_rw_block(READ, 1, &bh);
794 *wait_bh++=bh;
795 }
796
797 *p_blkno = *p_blkno + 1;
798 }
799
800 /*
801 * If we issued read requests - let them complete.
802 */
803 while(wait_bh > wait) {
804 wait_on_buffer(*--wait_bh);
805 if (!buffer_uptodate(*wait_bh))
806 ret = -EIO;
807 }
808
809 if (ret == 0 || !new)
810 return ret;
811
812 /*
813 * If we get -EIO above, zero out any newly allocated blocks
814 * to avoid exposing stale data.
815 */
816 bh = head;
817 block_start = 0;
818 do {
819 void *kaddr;
820
821 block_end = block_start + bsize;
822 if (block_end <= from)
823 goto next_bh;
824 if (block_start >= to)
825 break;
826
827 kaddr = kmap_atomic(page, KM_USER0);
828 memset(kaddr+block_start, 0, bh->b_size);
829 flush_dcache_page(page);
830 kunmap_atomic(kaddr, KM_USER0);
831 set_buffer_uptodate(bh);
832 mark_buffer_dirty(bh);
833
834next_bh:
835 block_start = block_end;
836 bh = bh->b_this_page;
837 } while (bh != head);
838
839 return ret;
840}
841
842/*
843 * This will copy user data from the iovec in the buffered write
844 * context.
845 */
846int ocfs2_map_and_write_user_data(struct inode *inode,
847 struct ocfs2_write_ctxt *wc, u64 *p_blkno,
848 unsigned int *ret_from, unsigned int *ret_to)
849{
850 int ret;
851 unsigned int to, from, cluster_start, cluster_end;
852 unsigned long bytes, src_from;
853 char *dst;
854 struct ocfs2_buffered_write_priv *bp = wc->w_private;
855 const struct iovec *cur_iov = bp->b_cur_iov;
856 char __user *buf;
857 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
858
859 ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
860 &cluster_end);
861
862 buf = cur_iov->iov_base + bp->b_cur_off;
863 src_from = (unsigned long)buf & ~PAGE_CACHE_MASK;
864
865 from = wc->w_pos & (PAGE_CACHE_SIZE - 1);
866
867 /*
868 * This is a lot of comparisons, but it reads quite
869 * easily, which is important here.
870 */
871 /* Stay within the src page */
872 bytes = PAGE_SIZE - src_from;
873 /* Stay within the vector */
874 bytes = min(bytes,
875 (unsigned long)(cur_iov->iov_len - bp->b_cur_off));
876 /* Stay within count */
877 bytes = min(bytes, (unsigned long)wc->w_count);
878 /*
879 * For clustersize > page size, just stay within
880 * target page, otherwise we have to calculate pos
881 * within the cluster and obey the rightmost
882 * boundary.
883 */
884 if (wc->w_large_pages) {
885 /*
886 * For cluster size < page size, we have to
887 * calculate pos within the cluster and obey
888 * the rightmost boundary.
889 */
890 bytes = min(bytes, (unsigned long)(osb->s_clustersize
891 - (wc->w_pos & (osb->s_clustersize - 1))));
892 } else {
893 /*
894 * cluster size > page size is the most common
895 * case - we just stay within the target page
896 * boundary.
897 */
898 bytes = min(bytes, PAGE_CACHE_SIZE - from);
899 }
900
901 to = from + bytes;
902
903 if (wc->w_this_page_new)
904 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
905 cluster_start, cluster_end, 1);
906 else
907 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
908 from, to, 0);
909 if (ret) {
910 mlog_errno(ret);
911 goto out;
912 }
913
914 BUG_ON(from > PAGE_CACHE_SIZE);
915 BUG_ON(to > PAGE_CACHE_SIZE);
916 BUG_ON(from > osb->s_clustersize);
917 BUG_ON(to > osb->s_clustersize);
918
919 dst = kmap(wc->w_this_page);
920 memcpy(dst + from, bp->b_src_buf + src_from, bytes);
921 kunmap(wc->w_this_page);
922
923 /*
924 * XXX: This is slow, but simple. The caller of
925 * ocfs2_buffered_write_cluster() is responsible for
926 * passing through the iovecs, so it's difficult to
927 * predict what our next step is in here after our
928 * initial write. A future version should be pushing
929 * that iovec manipulation further down.
930 *
931 * By setting this, we indicate that a copy from user
932 * data was done, and subsequent calls for this
933 * cluster will skip copying more data.
934 */
935 wc->w_finished_copy = 1;
936
937 *ret_from = from;
938 *ret_to = to;
939out:
940
941 return bytes ? (unsigned int)bytes : ret;
942}
943
944/*
945 * Map, fill and write a page to disk.
946 *
947 * The work of copying data is done via callback. Newly allocated
948 * pages which don't take user data will be zero'd (set 'new' to
949 * indicate an allocating write)
950 *
951 * Returns a negative error code or the number of bytes copied into
952 * the page.
953 */
954int ocfs2_write_data_page(struct inode *inode, handle_t *handle,
955 u64 *p_blkno, struct page *page,
956 struct ocfs2_write_ctxt *wc, int new)
957{
958 int ret, copied = 0;
959 unsigned int from = 0, to = 0;
960 unsigned int cluster_start, cluster_end;
961 unsigned int zero_from = 0, zero_to = 0;
962
963 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), wc->w_cpos,
964 &cluster_start, &cluster_end);
965
966 if ((wc->w_pos >> PAGE_CACHE_SHIFT) == page->index
967 && !wc->w_finished_copy) {
968
969 wc->w_this_page = page;
970 wc->w_this_page_new = new;
971 ret = wc->w_write_data_page(inode, wc, p_blkno, &from, &to);
972 if (ret < 0) {
973 mlog_errno(ret);
974 goto out;
975 }
976
977 copied = ret;
978
979 zero_from = from;
980 zero_to = to;
981 if (new) {
982 from = cluster_start;
983 to = cluster_end;
984 }
985 } else {
986 /*
987 * If we haven't allocated the new page yet, we
988 * shouldn't be writing it out without copying user
989 * data. This is likely a math error from the caller.
990 */
991 BUG_ON(!new);
992
993 from = cluster_start;
994 to = cluster_end;
995
996 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
997 cluster_start, cluster_end, 1);
998 if (ret) {
999 mlog_errno(ret);
1000 goto out;
1001 }
1002 }
1003
1004 /*
1005 * Parts of newly allocated pages need to be zero'd.
1006 *
1007 * Above, we have also rewritten 'to' and 'from' - as far as
1008 * the rest of the function is concerned, the entire cluster
1009 * range inside of a page needs to be written.
1010 *
1011 * We can skip this if the page is up to date - it's already
1012 * been zero'd from being read in as a hole.
1013 */
1014 if (new && !PageUptodate(page))
1015 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1016 wc->w_cpos, zero_from, zero_to);
1017
1018 flush_dcache_page(page);
1019
1020 if (ocfs2_should_order_data(inode)) {
1021 ret = walk_page_buffers(handle,
1022 page_buffers(page),
1023 from, to, NULL,
1024 ocfs2_journal_dirty_data);
1025 if (ret < 0)
1026 mlog_errno(ret);
1027 }
1028
1029 /*
1030 * We don't use generic_commit_write() because we need to
1031 * handle our own i_size update.
1032 */
1033 ret = block_commit_write(page, from, to);
1034 if (ret)
1035 mlog_errno(ret);
1036out:
1037
1038 return copied ? copied : ret;
1039}
1040
1041/*
1042 * Do the actual write of some data into an inode. Optionally allocate
1043 * in order to fulfill the write.
1044 *
1045 * cpos is the logical cluster offset within the file to write at
1046 *
1047 * 'phys' is the physical mapping of that offset. a 'phys' value of
1048 * zero indicates that allocation is required. In this case, data_ac
1049 * and meta_ac should be valid (meta_ac can be null if metadata
1050 * allocation isn't required).
1051 */
1052static ssize_t ocfs2_write(struct file *file, u32 phys, handle_t *handle,
1053 struct buffer_head *di_bh,
1054 struct ocfs2_alloc_context *data_ac,
1055 struct ocfs2_alloc_context *meta_ac,
1056 struct ocfs2_write_ctxt *wc)
1057{
1058 int ret, i, numpages = 1, new;
1059 unsigned int copied = 0;
1060 u32 tmp_pos;
1061 u64 v_blkno, p_blkno;
1062 struct address_space *mapping = file->f_mapping;
1063 struct inode *inode = mapping->host;
1064 unsigned int cbits = OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1065 unsigned long index, start;
1066 struct page **cpages;
1067
1068 new = phys == 0 ? 1 : 0;
1069
1070 /*
1071 * Figure out how many pages we'll be manipulating here. For
1072 * non-allocating write, or any writes where cluster size is
1073 * less than page size, we only need one page. Otherwise,
1074 * allocating writes of cluster size larger than page size
1075 * need cluster size pages.
1076 */
1077 if (new && !wc->w_large_pages)
1078 numpages = (1 << cbits) / PAGE_SIZE;
1079
1080 cpages = kzalloc(sizeof(*cpages) * numpages, GFP_NOFS);
1081 if (!cpages) {
1082 ret = -ENOMEM;
1083 mlog_errno(ret);
1084 return ret;
1085 }
1086
1087 /*
1088 * Fill our page array first. That way we've grabbed enough so
1089 * that we can zero and flush if we error after adding the
1090 * extent.
1091 */
1092 if (new) {
1093 start = ocfs2_align_clusters_to_page_index(inode->i_sb,
1094 wc->w_cpos);
1095 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, wc->w_cpos);
1096 } else {
1097 start = wc->w_pos >> PAGE_CACHE_SHIFT;
1098 v_blkno = wc->w_pos >> inode->i_sb->s_blocksize_bits;
1099 }
1100
1101 for(i = 0; i < numpages; i++) {
1102 index = start + i;
1103
1104 cpages[i] = grab_cache_page(mapping, index);
1105 if (!cpages[i]) {
1106 ret = -ENOMEM;
1107 mlog_errno(ret);
1108 goto out;
1109 }
1110 }
1111
1112 if (new) {
1113 /*
1114 * This is safe to call with the page locks - it won't take
1115 * any additional semaphores or cluster locks.
1116 */
1117 tmp_pos = wc->w_cpos;
1118 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1119 &tmp_pos, 1, di_bh, handle,
1120 data_ac, meta_ac, NULL);
1121 /*
1122 * This shouldn't happen because we must have already
1123 * calculated the correct meta data allocation required. The
1124 * internal tree allocation code should know how to increase
1125 * transaction credits itself.
1126 *
1127 * If need be, we could handle -EAGAIN for a
1128 * RESTART_TRANS here.
1129 */
1130 mlog_bug_on_msg(ret == -EAGAIN,
1131 "Inode %llu: EAGAIN return during allocation.\n",
1132 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1133 if (ret < 0) {
1134 mlog_errno(ret);
1135 goto out;
1136 }
1137 }
1138
1139 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL);
1140 if (ret < 0) {
1141
1142 /*
1143 * XXX: Should we go readonly here?
1144 */
1145
1146 mlog_errno(ret);
1147 goto out;
1148 }
1149
1150 BUG_ON(p_blkno == 0);
1151
1152 for(i = 0; i < numpages; i++) {
1153 ret = ocfs2_write_data_page(inode, handle, &p_blkno, cpages[i],
1154 wc, new);
1155 if (ret < 0) {
1156 mlog_errno(ret);
1157 goto out;
1158 }
1159
1160 copied += ret;
1161 }
1162
1163out:
1164 for(i = 0; i < numpages; i++) {
1165 unlock_page(cpages[i]);
1166 mark_page_accessed(cpages[i]);
1167 page_cache_release(cpages[i]);
1168 }
1169 kfree(cpages);
1170
1171 return copied ? copied : ret;
1172}
1173
1174static void ocfs2_write_ctxt_init(struct ocfs2_write_ctxt *wc,
1175 struct ocfs2_super *osb, loff_t pos,
1176 size_t count, ocfs2_page_writer *cb,
1177 void *cb_priv)
1178{
1179 wc->w_count = count;
1180 wc->w_pos = pos;
1181 wc->w_cpos = wc->w_pos >> osb->s_clustersize_bits;
1182 wc->w_finished_copy = 0;
1183
1184 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1185 wc->w_large_pages = 1;
1186 else
1187 wc->w_large_pages = 0;
1188
1189 wc->w_write_data_page = cb;
1190 wc->w_private = cb_priv;
1191}
1192
1193/*
1194 * Write a cluster to an inode. The cluster may not be allocated yet,
1195 * in which case it will be. This only exists for buffered writes -
1196 * O_DIRECT takes a more "traditional" path through the kernel.
1197 *
1198 * The caller is responsible for incrementing pos, written counts, etc
1199 *
1200 * For file systems that don't support sparse files, pre-allocation
1201 * and page zeroing up until cpos should be done prior to this
1202 * function call.
1203 *
1204 * Callers should be holding i_sem, and the rw cluster lock.
1205 *
1206 * Returns the number of user bytes written, or less than zero for
1207 * error.
1208 */
1209ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos,
1210 size_t count, ocfs2_page_writer *actor,
1211 void *priv)
1212{
1213 int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1214 ssize_t written = 0;
1215 u32 phys;
1216 struct inode *inode = file->f_mapping->host;
1217 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1218 struct buffer_head *di_bh = NULL;
1219 struct ocfs2_dinode *di;
1220 struct ocfs2_alloc_context *data_ac = NULL;
1221 struct ocfs2_alloc_context *meta_ac = NULL;
1222 handle_t *handle;
1223 struct ocfs2_write_ctxt wc;
1224
1225 ocfs2_write_ctxt_init(&wc, osb, pos, count, actor, priv);
1226
1227 ret = ocfs2_meta_lock(inode, &di_bh, 1);
1228 if (ret) {
1229 mlog_errno(ret);
1230 goto out;
1231 }
1232 di = (struct ocfs2_dinode *)di_bh->b_data;
1233
1234 /*
1235 * Take alloc sem here to prevent concurrent lookups. That way
1236 * the mapping, zeroing and tree manipulation within
1237 * ocfs2_write() will be safe against ->readpage(). This
1238 * should also serve to lock out allocation from a shared
1239 * writeable region.
1240 */
1241 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1242
1243 ret = ocfs2_get_clusters(inode, wc.w_cpos, &phys, NULL);
1244 if (ret) {
1245 mlog_errno(ret);
1246 goto out_meta;
1247 }
1248
1249 /* phys == 0 means that allocation is required. */
1250 if (phys == 0) {
1251 ret = ocfs2_lock_allocators(inode, di, 1, &data_ac, &meta_ac);
1252 if (ret) {
1253 mlog_errno(ret);
1254 goto out_meta;
1255 }
1256
1257 credits = ocfs2_calc_extend_credits(inode->i_sb, di, 1);
1258 }
1259
1260 ret = ocfs2_data_lock(inode, 1);
1261 if (ret) {
1262 mlog_errno(ret);
1263 goto out_meta;
1264 }
1265
1266 handle = ocfs2_start_trans(osb, credits);
1267 if (IS_ERR(handle)) {
1268 ret = PTR_ERR(handle);
1269 mlog_errno(ret);
1270 goto out_data;
1271 }
1272
1273 written = ocfs2_write(file, phys, handle, di_bh, data_ac,
1274 meta_ac, &wc);
1275 if (written < 0) {
1276 ret = written;
1277 mlog_errno(ret);
1278 goto out_commit;
1279 }
1280
1281 ret = ocfs2_journal_access(handle, inode, di_bh,
1282 OCFS2_JOURNAL_ACCESS_WRITE);
1283 if (ret) {
1284 mlog_errno(ret);
1285 goto out_commit;
1286 }
1287
1288 pos += written;
1289 if (pos > inode->i_size) {
1290 i_size_write(inode, pos);
1291 mark_inode_dirty(inode);
1292 }
1293 inode->i_blocks = ocfs2_align_bytes_to_sectors((u64)(i_size_read(inode)));
1294 di->i_size = cpu_to_le64((u64)i_size_read(inode));
1295 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1296 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1297 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1298
1299 ret = ocfs2_journal_dirty(handle, di_bh);
1300 if (ret)
1301 mlog_errno(ret);
1302
1303out_commit:
1304 ocfs2_commit_trans(osb, handle);
1305
1306out_data:
1307 ocfs2_data_unlock(inode, 1);
1308
1309out_meta:
1310 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1311 ocfs2_meta_unlock(inode, 1);
1312
1313out:
1314 brelse(di_bh);
1315 if (data_ac)
1316 ocfs2_free_alloc_context(data_ac);
1317 if (meta_ac)
1318 ocfs2_free_alloc_context(meta_ac);
1319
1320 return written ? written : ret;
1321}
1322
676const struct address_space_operations ocfs2_aops = { 1323const struct address_space_operations ocfs2_aops = {
677 .readpage = ocfs2_readpage, 1324 .readpage = ocfs2_readpage,
678 .writepage = ocfs2_writepage, 1325 .writepage = ocfs2_writepage,
diff --git a/fs/ocfs2/aops.h b/fs/ocfs2/aops.h
index f446a15eab88..eeb2c42483e8 100644
--- a/fs/ocfs2/aops.h
+++ b/fs/ocfs2/aops.h
@@ -30,6 +30,44 @@ handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
30 unsigned from, 30 unsigned from,
31 unsigned to); 31 unsigned to);
32 32
33struct ocfs2_write_ctxt;
34typedef int (ocfs2_page_writer)(struct inode *, struct ocfs2_write_ctxt *,
35 u64 *, unsigned int *, unsigned int *);
36
37ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos,
38 size_t count, ocfs2_page_writer *actor,
39 void *priv);
40
41struct ocfs2_write_ctxt {
42 size_t w_count;
43 loff_t w_pos;
44 u32 w_cpos;
45 unsigned int w_finished_copy;
46
47 /* This is true if page_size > cluster_size */
48 unsigned int w_large_pages;
49
50 /* Filler callback and private data */
51 ocfs2_page_writer *w_write_data_page;
52 void *w_private;
53
54 /* Only valid for the filler callback */
55 struct page *w_this_page;
56 unsigned int w_this_page_new;
57};
58
59struct ocfs2_buffered_write_priv {
60 char *b_src_buf;
61 const struct iovec *b_cur_iov; /* Current iovec */
62 size_t b_cur_off; /* Offset in the
63 * current iovec */
64};
65int ocfs2_map_and_write_user_data(struct inode *inode,
66 struct ocfs2_write_ctxt *wc,
67 u64 *p_blkno,
68 unsigned int *ret_from,
69 unsigned int *ret_to);
70
33/* all ocfs2_dio_end_io()'s fault */ 71/* all ocfs2_dio_end_io()'s fault */
34#define ocfs2_iocb_is_rw_locked(iocb) \ 72#define ocfs2_iocb_is_rw_locked(iocb) \
35 test_bit(0, (unsigned long *)&iocb->private) 73 test_bit(0, (unsigned long *)&iocb->private)
diff --git a/fs/ocfs2/extent_map.c b/fs/ocfs2/extent_map.c
index 3b4322fd369a..937c2722b753 100644
--- a/fs/ocfs2/extent_map.c
+++ b/fs/ocfs2/extent_map.c
@@ -67,8 +67,8 @@ static int ocfs2_search_extent_list(struct ocfs2_extent_list *el,
67 return ret; 67 return ret;
68} 68}
69 69
70static int ocfs2_get_clusters(struct inode *inode, u32 v_cluster, 70int ocfs2_get_clusters(struct inode *inode, u32 v_cluster,
71 u32 *p_cluster, u32 *num_clusters) 71 u32 *p_cluster, u32 *num_clusters)
72{ 72{
73 int ret, i; 73 int ret, i;
74 struct buffer_head *di_bh = NULL; 74 struct buffer_head *di_bh = NULL;
diff --git a/fs/ocfs2/extent_map.h b/fs/ocfs2/extent_map.h
index 036e23251448..625d0ee5e04a 100644
--- a/fs/ocfs2/extent_map.h
+++ b/fs/ocfs2/extent_map.h
@@ -25,6 +25,8 @@
25#ifndef _EXTENT_MAP_H 25#ifndef _EXTENT_MAP_H
26#define _EXTENT_MAP_H 26#define _EXTENT_MAP_H
27 27
28int ocfs2_get_clusters(struct inode *inode, u32 v_cluster, u32 *p_cluster,
29 u32 *num_clusters);
28int ocfs2_extent_map_get_blocks(struct inode *inode, u64 v_blkno, u64 *p_blkno, 30int ocfs2_extent_map_get_blocks(struct inode *inode, u64 v_blkno, u64 *p_blkno,
29 int *ret_count); 31 int *ret_count);
30 32
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
index 3bcf3629265e..667e5a869bf5 100644
--- a/fs/ocfs2/file.c
+++ b/fs/ocfs2/file.c
@@ -33,6 +33,7 @@
33#include <linux/sched.h> 33#include <linux/sched.h>
34#include <linux/pipe_fs_i.h> 34#include <linux/pipe_fs_i.h>
35#include <linux/mount.h> 35#include <linux/mount.h>
36#include <linux/writeback.h>
36 37
37#define MLOG_MASK_PREFIX ML_INODE 38#define MLOG_MASK_PREFIX ML_INODE
38#include <cluster/masklog.h> 39#include <cluster/masklog.h>
@@ -485,13 +486,13 @@ leave:
485 * accessed, and lock them, reserving the appropriate number of bits. 486 * accessed, and lock them, reserving the appropriate number of bits.
486 * 487 *
487 * Called from ocfs2_extend_allocation() for file systems which don't 488 * Called from ocfs2_extend_allocation() for file systems which don't
488 * support holes, and from ocfs2_prepare_write() for file systems 489 * support holes, and from ocfs2_write() for file systems which
489 * which understand sparse inodes. 490 * understand sparse inodes.
490 */ 491 */
491static int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di, 492int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
492 u32 clusters_to_add, 493 u32 clusters_to_add,
493 struct ocfs2_alloc_context **data_ac, 494 struct ocfs2_alloc_context **data_ac,
494 struct ocfs2_alloc_context **meta_ac) 495 struct ocfs2_alloc_context **meta_ac)
495{ 496{
496 int ret, num_free_extents; 497 int ret, num_free_extents;
497 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 498 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
@@ -518,7 +519,7 @@ static int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
518 * a cluster lock (because we ran out of room for another 519 * a cluster lock (because we ran out of room for another
519 * extent) will violate ordering rules. 520 * extent) will violate ordering rules.
520 * 521 *
521 * Most of the time we'll only be seeing this 1 page at a time 522 * Most of the time we'll only be seeing this 1 cluster at a time
522 * anyway. 523 * anyway.
523 */ 524 */
524 if (!num_free_extents || 525 if (!num_free_extents ||
@@ -596,13 +597,6 @@ static int ocfs2_extend_allocation(struct inode *inode,
596restart_all: 597restart_all:
597 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters); 598 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
598 599
599 status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
600 &meta_ac);
601 if (status) {
602 mlog_errno(status);
603 goto leave;
604 }
605
606 /* blocks peope in read/write from reading our allocation 600 /* blocks peope in read/write from reading our allocation
607 * until we're done changing it. We depend on i_mutex to block 601 * until we're done changing it. We depend on i_mutex to block
608 * other extend/truncate calls while we're here. Ordering wrt 602 * other extend/truncate calls while we're here. Ordering wrt
@@ -610,6 +604,13 @@ restart_all:
610 down_write(&OCFS2_I(inode)->ip_alloc_sem); 604 down_write(&OCFS2_I(inode)->ip_alloc_sem);
611 drop_alloc_sem = 1; 605 drop_alloc_sem = 1;
612 606
607 status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
608 &meta_ac);
609 if (status) {
610 mlog_errno(status);
611 goto leave;
612 }
613
613 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add); 614 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
614 handle = ocfs2_start_trans(osb, credits); 615 handle = ocfs2_start_trans(osb, credits);
615 if (IS_ERR(handle)) { 616 if (IS_ERR(handle)) {
@@ -1088,10 +1089,49 @@ out:
1088 return ret; 1089 return ret;
1089} 1090}
1090 1091
1092/*
1093 * Will look for holes and unwritten extents in the range starting at
1094 * pos for count bytes (inclusive).
1095 */
1096static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1097 size_t count)
1098{
1099 int ret = 0;
1100 unsigned int extent_flags;
1101 u32 cpos, clusters, extent_len, phys_cpos;
1102 struct super_block *sb = inode->i_sb;
1103
1104 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1105 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1106
1107 while (clusters) {
1108 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1109 &extent_flags);
1110 if (ret < 0) {
1111 mlog_errno(ret);
1112 goto out;
1113 }
1114
1115 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1116 ret = 1;
1117 break;
1118 }
1119
1120 if (extent_len > clusters)
1121 extent_len = clusters;
1122
1123 clusters -= extent_len;
1124 cpos += extent_len;
1125 }
1126out:
1127 return ret;
1128}
1129
1091static int ocfs2_prepare_inode_for_write(struct dentry *dentry, 1130static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1092 loff_t *ppos, 1131 loff_t *ppos,
1093 size_t count, 1132 size_t count,
1094 int appending) 1133 int appending,
1134 int *direct_io)
1095{ 1135{
1096 int ret = 0, meta_level = appending; 1136 int ret = 0, meta_level = appending;
1097 struct inode *inode = dentry->d_inode; 1137 struct inode *inode = dentry->d_inode;
@@ -1143,12 +1183,47 @@ static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1143 saved_pos = *ppos; 1183 saved_pos = *ppos;
1144 } 1184 }
1145 1185
1186 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1187 loff_t end = saved_pos + count;
1188
1189 /*
1190 * Skip the O_DIRECT checks if we don't need
1191 * them.
1192 */
1193 if (!direct_io || !(*direct_io))
1194 break;
1195
1196 /*
1197 * Allowing concurrent direct writes means
1198 * i_size changes wouldn't be synchronized, so
1199 * one node could wind up truncating another
1200 * nodes writes.
1201 */
1202 if (end > i_size_read(inode)) {
1203 *direct_io = 0;
1204 break;
1205 }
1206
1207 /*
1208 * We don't fill holes during direct io, so
1209 * check for them here. If any are found, the
1210 * caller will have to retake some cluster
1211 * locks and initiate the io as buffered.
1212 */
1213 ret = ocfs2_check_range_for_holes(inode, saved_pos,
1214 count);
1215 if (ret == 1) {
1216 *direct_io = 0;
1217 ret = 0;
1218 } else if (ret < 0)
1219 mlog_errno(ret);
1220 break;
1221 }
1222
1146 /* 1223 /*
1147 * The rest of this loop is concerned with legacy file 1224 * The rest of this loop is concerned with legacy file
1148 * systems which don't support sparse files. 1225 * systems which don't support sparse files.
1149 */ 1226 */
1150 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1151 break;
1152 1227
1153 newsize = count + saved_pos; 1228 newsize = count + saved_pos;
1154 1229
@@ -1202,55 +1277,264 @@ out:
1202 return ret; 1277 return ret;
1203} 1278}
1204 1279
1280static inline void
1281ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1282{
1283 const struct iovec *iov = *iovp;
1284 size_t base = *basep;
1285
1286 do {
1287 int copy = min(bytes, iov->iov_len - base);
1288
1289 bytes -= copy;
1290 base += copy;
1291 if (iov->iov_len == base) {
1292 iov++;
1293 base = 0;
1294 }
1295 } while (bytes);
1296 *iovp = iov;
1297 *basep = base;
1298}
1299
1300static struct page * ocfs2_get_write_source(struct ocfs2_buffered_write_priv *bp,
1301 const struct iovec *cur_iov,
1302 size_t iov_offset)
1303{
1304 int ret;
1305 char *buf;
1306 struct page *src_page = NULL;
1307
1308 buf = cur_iov->iov_base + iov_offset;
1309
1310 if (!segment_eq(get_fs(), KERNEL_DS)) {
1311 /*
1312 * Pull in the user page. We want to do this outside
1313 * of the meta data locks in order to preserve locking
1314 * order in case of page fault.
1315 */
1316 ret = get_user_pages(current, current->mm,
1317 (unsigned long)buf & PAGE_CACHE_MASK, 1,
1318 0, 0, &src_page, NULL);
1319 if (ret == 1)
1320 bp->b_src_buf = kmap(src_page);
1321 else
1322 src_page = ERR_PTR(-EFAULT);
1323 } else {
1324 bp->b_src_buf = buf;
1325 }
1326
1327 return src_page;
1328}
1329
1330static void ocfs2_put_write_source(struct ocfs2_buffered_write_priv *bp,
1331 struct page *page)
1332{
1333 if (page) {
1334 kunmap(page);
1335 page_cache_release(page);
1336 }
1337}
1338
1339static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1340 const struct iovec *iov,
1341 unsigned long nr_segs,
1342 size_t count,
1343 ssize_t o_direct_written)
1344{
1345 int ret = 0;
1346 ssize_t copied, total = 0;
1347 size_t iov_offset = 0;
1348 const struct iovec *cur_iov = iov;
1349 struct ocfs2_buffered_write_priv bp;
1350 struct page *page;
1351
1352 /*
1353 * handle partial DIO write. Adjust cur_iov if needed.
1354 */
1355 ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1356
1357 do {
1358 bp.b_cur_off = iov_offset;
1359 bp.b_cur_iov = cur_iov;
1360
1361 page = ocfs2_get_write_source(&bp, cur_iov, iov_offset);
1362 if (IS_ERR(page)) {
1363 ret = PTR_ERR(page);
1364 goto out;
1365 }
1366
1367 copied = ocfs2_buffered_write_cluster(file, *ppos, count,
1368 ocfs2_map_and_write_user_data,
1369 &bp);
1370
1371 ocfs2_put_write_source(&bp, page);
1372
1373 if (copied < 0) {
1374 mlog_errno(copied);
1375 ret = copied;
1376 goto out;
1377 }
1378
1379 total += copied;
1380 *ppos = *ppos + copied;
1381 count -= copied;
1382
1383 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1384 } while(count);
1385
1386out:
1387 return total ? total : ret;
1388}
1389
1390static int ocfs2_check_iovec(const struct iovec *iov, size_t *counted,
1391 unsigned long *nr_segs)
1392{
1393 size_t ocount; /* original count */
1394 unsigned long seg;
1395
1396 ocount = 0;
1397 for (seg = 0; seg < *nr_segs; seg++) {
1398 const struct iovec *iv = &iov[seg];
1399
1400 /*
1401 * If any segment has a negative length, or the cumulative
1402 * length ever wraps negative then return -EINVAL.
1403 */
1404 ocount += iv->iov_len;
1405 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
1406 return -EINVAL;
1407 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
1408 continue;
1409 if (seg == 0)
1410 return -EFAULT;
1411 *nr_segs = seg;
1412 ocount -= iv->iov_len; /* This segment is no good */
1413 break;
1414 }
1415
1416 *counted = ocount;
1417 return 0;
1418}
1419
1205static ssize_t ocfs2_file_aio_write(struct kiocb *iocb, 1420static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1206 const struct iovec *iov, 1421 const struct iovec *iov,
1207 unsigned long nr_segs, 1422 unsigned long nr_segs,
1208 loff_t pos) 1423 loff_t pos)
1209{ 1424{
1210 int ret, rw_level, have_alloc_sem = 0; 1425 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1211 struct file *filp = iocb->ki_filp; 1426 int can_do_direct, sync = 0;
1212 struct inode *inode = filp->f_path.dentry->d_inode; 1427 ssize_t written = 0;
1213 int appending = filp->f_flags & O_APPEND ? 1 : 0; 1428 size_t ocount; /* original count */
1214 1429 size_t count; /* after file limit checks */
1215 mlog_entry("(0x%p, %u, '%.*s')\n", filp, 1430 loff_t *ppos = &iocb->ki_pos;
1431 struct file *file = iocb->ki_filp;
1432 struct inode *inode = file->f_path.dentry->d_inode;
1433
1434 mlog_entry("(0x%p, %u, '%.*s')\n", file,
1216 (unsigned int)nr_segs, 1435 (unsigned int)nr_segs,
1217 filp->f_path.dentry->d_name.len, 1436 file->f_path.dentry->d_name.len,
1218 filp->f_path.dentry->d_name.name); 1437 file->f_path.dentry->d_name.name);
1219 1438
1220 /* happy write of zero bytes */
1221 if (iocb->ki_left == 0) 1439 if (iocb->ki_left == 0)
1222 return 0; 1440 return 0;
1223 1441
1442 ret = ocfs2_check_iovec(iov, &ocount, &nr_segs);
1443 if (ret)
1444 return ret;
1445
1446 count = ocount;
1447
1448 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1449
1450 appending = file->f_flags & O_APPEND ? 1 : 0;
1451 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1452
1224 mutex_lock(&inode->i_mutex); 1453 mutex_lock(&inode->i_mutex);
1454
1455relock:
1225 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */ 1456 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1226 if (filp->f_flags & O_DIRECT) { 1457 if (direct_io) {
1227 have_alloc_sem = 1;
1228 down_read(&inode->i_alloc_sem); 1458 down_read(&inode->i_alloc_sem);
1459 have_alloc_sem = 1;
1229 } 1460 }
1230 1461
1231 /* concurrent O_DIRECT writes are allowed */ 1462 /* concurrent O_DIRECT writes are allowed */
1232 rw_level = (filp->f_flags & O_DIRECT) ? 0 : 1; 1463 rw_level = !direct_io;
1233 ret = ocfs2_rw_lock(inode, rw_level); 1464 ret = ocfs2_rw_lock(inode, rw_level);
1234 if (ret < 0) { 1465 if (ret < 0) {
1235 rw_level = -1;
1236 mlog_errno(ret); 1466 mlog_errno(ret);
1237 goto out; 1467 goto out_sems;
1238 } 1468 }
1239 1469
1240 ret = ocfs2_prepare_inode_for_write(filp->f_path.dentry, &iocb->ki_pos, 1470 can_do_direct = direct_io;
1241 iocb->ki_left, appending); 1471 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1472 iocb->ki_left, appending,
1473 &can_do_direct);
1242 if (ret < 0) { 1474 if (ret < 0) {
1243 mlog_errno(ret); 1475 mlog_errno(ret);
1244 goto out; 1476 goto out;
1245 } 1477 }
1246 1478
1479 /*
1480 * We can't complete the direct I/O as requested, fall back to
1481 * buffered I/O.
1482 */
1483 if (direct_io && !can_do_direct) {
1484 ocfs2_rw_unlock(inode, rw_level);
1485 up_read(&inode->i_alloc_sem);
1486
1487 have_alloc_sem = 0;
1488 rw_level = -1;
1489
1490 direct_io = 0;
1491 sync = 1;
1492 goto relock;
1493 }
1494
1495 if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1496 sync = 1;
1497
1498 /*
1499 * XXX: Is it ok to execute these checks a second time?
1500 */
1501 ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1502 if (ret)
1503 goto out;
1504
1505 /*
1506 * Set pos so that sync_page_range_nolock() below understands
1507 * where to start from. We might've moved it around via the
1508 * calls above. The range we want to actually sync starts from
1509 * *ppos here.
1510 *
1511 */
1512 pos = *ppos;
1513
1247 /* communicate with ocfs2_dio_end_io */ 1514 /* communicate with ocfs2_dio_end_io */
1248 ocfs2_iocb_set_rw_locked(iocb); 1515 ocfs2_iocb_set_rw_locked(iocb);
1249 1516
1250 ret = generic_file_aio_write_nolock(iocb, iov, nr_segs, iocb->ki_pos); 1517 if (direct_io) {
1518 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1519 ppos, count, ocount);
1520 if (written < 0) {
1521 ret = written;
1522 goto out_dio;
1523 }
1524 } else {
1525 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1526 count, written);
1527 if (written < 0) {
1528 ret = written;
1529 if (ret != -EFAULT || ret != -ENOSPC)
1530 mlog_errno(ret);
1531 goto out;
1532 }
1533 }
1251 1534
1535out_dio:
1252 /* buffered aio wouldn't have proper lock coverage today */ 1536 /* buffered aio wouldn't have proper lock coverage today */
1253 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT)); 1537 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1254 1538
1255 /* 1539 /*
1256 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io 1540 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
@@ -1268,14 +1552,25 @@ static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1268 } 1552 }
1269 1553
1270out: 1554out:
1555 if (rw_level != -1)
1556 ocfs2_rw_unlock(inode, rw_level);
1557
1558out_sems:
1271 if (have_alloc_sem) 1559 if (have_alloc_sem)
1272 up_read(&inode->i_alloc_sem); 1560 up_read(&inode->i_alloc_sem);
1273 if (rw_level != -1) 1561
1274 ocfs2_rw_unlock(inode, rw_level); 1562 if (written > 0 && sync) {
1563 ssize_t err;
1564
1565 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1566 if (err < 0)
1567 written = err;
1568 }
1569
1275 mutex_unlock(&inode->i_mutex); 1570 mutex_unlock(&inode->i_mutex);
1276 1571
1277 mlog_exit(ret); 1572 mlog_exit(ret);
1278 return ret; 1573 return written ? written : ret;
1279} 1574}
1280 1575
1281static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe, 1576static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
@@ -1300,7 +1595,8 @@ static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1300 goto out; 1595 goto out;
1301 } 1596 }
1302 1597
1303 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0); 1598 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1599 NULL);
1304 if (ret < 0) { 1600 if (ret < 0) {
1305 mlog_errno(ret); 1601 mlog_errno(ret);
1306 goto out_unlock; 1602 goto out_unlock;
diff --git a/fs/ocfs2/file.h b/fs/ocfs2/file.h
index e2f6551604d0..2c4460fced52 100644
--- a/fs/ocfs2/file.h
+++ b/fs/ocfs2/file.h
@@ -46,6 +46,10 @@ int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
46 struct ocfs2_alloc_context *data_ac, 46 struct ocfs2_alloc_context *data_ac,
47 struct ocfs2_alloc_context *meta_ac, 47 struct ocfs2_alloc_context *meta_ac,
48 enum ocfs2_alloc_restarted *reason); 48 enum ocfs2_alloc_restarted *reason);
49int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
50 u32 clusters_to_add,
51 struct ocfs2_alloc_context **data_ac,
52 struct ocfs2_alloc_context **meta_ac);
49int ocfs2_setattr(struct dentry *dentry, struct iattr *attr); 53int ocfs2_setattr(struct dentry *dentry, struct iattr *attr);
50int ocfs2_getattr(struct vfsmount *mnt, struct dentry *dentry, 54int ocfs2_getattr(struct vfsmount *mnt, struct dentry *dentry,
51 struct kstat *stat); 55 struct kstat *stat);
diff --git a/fs/ocfs2/ocfs2.h b/fs/ocfs2/ocfs2.h
index faeb53f2eecf..2699f7cac21a 100644
--- a/fs/ocfs2/ocfs2.h
+++ b/fs/ocfs2/ocfs2.h
@@ -463,6 +463,38 @@ static inline unsigned long ocfs2_align_bytes_to_sectors(u64 bytes)
463 return (unsigned long)((bytes + 511) >> 9); 463 return (unsigned long)((bytes + 511) >> 9);
464} 464}
465 465
466static inline unsigned int ocfs2_page_index_to_clusters(struct super_block *sb,
467 unsigned long pg_index)
468{
469 u32 clusters = pg_index;
470 unsigned int cbits = OCFS2_SB(sb)->s_clustersize_bits;
471
472 if (unlikely(PAGE_CACHE_SHIFT > cbits))
473 clusters = pg_index << (PAGE_CACHE_SHIFT - cbits);
474 else if (PAGE_CACHE_SHIFT < cbits)
475 clusters = pg_index >> (cbits - PAGE_CACHE_SHIFT);
476
477 return clusters;
478}
479
480/*
481 * Find the 1st page index which covers the given clusters.
482 */
483static inline unsigned long ocfs2_align_clusters_to_page_index(struct super_block *sb,
484 u32 clusters)
485{
486 unsigned int cbits = OCFS2_SB(sb)->s_clustersize_bits;
487 unsigned long index = clusters;
488
489 if (PAGE_CACHE_SHIFT > cbits) {
490 index = clusters >> (PAGE_CACHE_SHIFT - cbits);
491 } else if (PAGE_CACHE_SHIFT < cbits) {
492 index = clusters << (cbits - PAGE_CACHE_SHIFT);
493 }
494
495 return index;
496}
497
466#define ocfs2_set_bit ext2_set_bit 498#define ocfs2_set_bit ext2_set_bit
467#define ocfs2_clear_bit ext2_clear_bit 499#define ocfs2_clear_bit ext2_clear_bit
468#define ocfs2_test_bit ext2_test_bit 500#define ocfs2_test_bit ext2_test_bit