aboutsummaryrefslogtreecommitdiffstats
path: root/mm
diff options
context:
space:
mode:
Diffstat (limited to 'mm')
-rw-r--r--mm/Kconfig2
-rw-r--r--mm/Makefile2
-rw-r--r--mm/allocpercpu.c28
-rw-r--r--mm/backing-dev.c90
-rw-r--r--mm/kmemleak-test.c6
-rw-r--r--mm/page-writeback.c27
-rw-r--r--mm/percpu.c1420
-rw-r--r--mm/quicklist.c2
-rw-r--r--mm/slub.c6
-rw-r--r--mm/vmalloc.c338
10 files changed, 1503 insertions, 418 deletions
diff --git a/mm/Kconfig b/mm/Kconfig
index fe5f674d7a7d..3aa519f52e18 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -153,7 +153,7 @@ config MEMORY_HOTREMOVE
153# 153#
154config PAGEFLAGS_EXTENDED 154config PAGEFLAGS_EXTENDED
155 def_bool y 155 def_bool y
156 depends on 64BIT || SPARSEMEM_VMEMMAP || !NUMA || !SPARSEMEM 156 depends on 64BIT || SPARSEMEM_VMEMMAP || !SPARSEMEM
157 157
158# Heavily threaded applications may benefit from splitting the mm-wide 158# Heavily threaded applications may benefit from splitting the mm-wide
159# page_table_lock, so that faults on different parts of the user address 159# page_table_lock, so that faults on different parts of the user address
diff --git a/mm/Makefile b/mm/Makefile
index 147a7a7873c4..ea4b18bd3960 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -33,7 +33,7 @@ obj-$(CONFIG_FAILSLAB) += failslab.o
33obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o 33obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
34obj-$(CONFIG_FS_XIP) += filemap_xip.o 34obj-$(CONFIG_FS_XIP) += filemap_xip.o
35obj-$(CONFIG_MIGRATION) += migrate.o 35obj-$(CONFIG_MIGRATION) += migrate.o
36ifdef CONFIG_HAVE_DYNAMIC_PER_CPU_AREA 36ifndef CONFIG_HAVE_LEGACY_PER_CPU_AREA
37obj-$(CONFIG_SMP) += percpu.o 37obj-$(CONFIG_SMP) += percpu.o
38else 38else
39obj-$(CONFIG_SMP) += allocpercpu.o 39obj-$(CONFIG_SMP) += allocpercpu.o
diff --git a/mm/allocpercpu.c b/mm/allocpercpu.c
index dfdee6a47359..df34ceae0c67 100644
--- a/mm/allocpercpu.c
+++ b/mm/allocpercpu.c
@@ -5,6 +5,8 @@
5 */ 5 */
6#include <linux/mm.h> 6#include <linux/mm.h>
7#include <linux/module.h> 7#include <linux/module.h>
8#include <linux/bootmem.h>
9#include <asm/sections.h>
8 10
9#ifndef cache_line_size 11#ifndef cache_line_size
10#define cache_line_size() L1_CACHE_BYTES 12#define cache_line_size() L1_CACHE_BYTES
@@ -147,3 +149,29 @@ void free_percpu(void *__pdata)
147 kfree(__percpu_disguise(__pdata)); 149 kfree(__percpu_disguise(__pdata));
148} 150}
149EXPORT_SYMBOL_GPL(free_percpu); 151EXPORT_SYMBOL_GPL(free_percpu);
152
153/*
154 * Generic percpu area setup.
155 */
156#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
157unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
158
159EXPORT_SYMBOL(__per_cpu_offset);
160
161void __init setup_per_cpu_areas(void)
162{
163 unsigned long size, i;
164 char *ptr;
165 unsigned long nr_possible_cpus = num_possible_cpus();
166
167 /* Copy section for each CPU (we discard the original) */
168 size = ALIGN(PERCPU_ENOUGH_ROOM, PAGE_SIZE);
169 ptr = alloc_bootmem_pages(size * nr_possible_cpus);
170
171 for_each_possible_cpu(i) {
172 __per_cpu_offset[i] = ptr - __per_cpu_start;
173 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
174 ptr += size;
175 }
176}
177#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
diff --git a/mm/backing-dev.c b/mm/backing-dev.c
index d3ca0dac1111..3d3accb1f800 100644
--- a/mm/backing-dev.c
+++ b/mm/backing-dev.c
@@ -26,6 +26,12 @@ struct backing_dev_info default_backing_dev_info = {
26EXPORT_SYMBOL_GPL(default_backing_dev_info); 26EXPORT_SYMBOL_GPL(default_backing_dev_info);
27 27
28static struct class *bdi_class; 28static struct class *bdi_class;
29
30/*
31 * bdi_lock protects updates to bdi_list and bdi_pending_list, as well as
32 * reader side protection for bdi_pending_list. bdi_list has RCU reader side
33 * locking.
34 */
29DEFINE_SPINLOCK(bdi_lock); 35DEFINE_SPINLOCK(bdi_lock);
30LIST_HEAD(bdi_list); 36LIST_HEAD(bdi_list);
31LIST_HEAD(bdi_pending_list); 37LIST_HEAD(bdi_pending_list);
@@ -284,9 +290,9 @@ static int bdi_start_fn(void *ptr)
284 /* 290 /*
285 * Add us to the active bdi_list 291 * Add us to the active bdi_list
286 */ 292 */
287 spin_lock(&bdi_lock); 293 spin_lock_bh(&bdi_lock);
288 list_add(&bdi->bdi_list, &bdi_list); 294 list_add_rcu(&bdi->bdi_list, &bdi_list);
289 spin_unlock(&bdi_lock); 295 spin_unlock_bh(&bdi_lock);
290 296
291 bdi_task_init(bdi, wb); 297 bdi_task_init(bdi, wb);
292 298
@@ -389,7 +395,7 @@ static int bdi_forker_task(void *ptr)
389 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) 395 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list))
390 wb_do_writeback(me, 0); 396 wb_do_writeback(me, 0);
391 397
392 spin_lock(&bdi_lock); 398 spin_lock_bh(&bdi_lock);
393 399
394 /* 400 /*
395 * Check if any existing bdi's have dirty data without 401 * Check if any existing bdi's have dirty data without
@@ -410,7 +416,7 @@ static int bdi_forker_task(void *ptr)
410 if (list_empty(&bdi_pending_list)) { 416 if (list_empty(&bdi_pending_list)) {
411 unsigned long wait; 417 unsigned long wait;
412 418
413 spin_unlock(&bdi_lock); 419 spin_unlock_bh(&bdi_lock);
414 wait = msecs_to_jiffies(dirty_writeback_interval * 10); 420 wait = msecs_to_jiffies(dirty_writeback_interval * 10);
415 schedule_timeout(wait); 421 schedule_timeout(wait);
416 try_to_freeze(); 422 try_to_freeze();
@@ -426,7 +432,7 @@ static int bdi_forker_task(void *ptr)
426 bdi = list_entry(bdi_pending_list.next, struct backing_dev_info, 432 bdi = list_entry(bdi_pending_list.next, struct backing_dev_info,
427 bdi_list); 433 bdi_list);
428 list_del_init(&bdi->bdi_list); 434 list_del_init(&bdi->bdi_list);
429 spin_unlock(&bdi_lock); 435 spin_unlock_bh(&bdi_lock);
430 436
431 wb = &bdi->wb; 437 wb = &bdi->wb;
432 wb->task = kthread_run(bdi_start_fn, wb, "flush-%s", 438 wb->task = kthread_run(bdi_start_fn, wb, "flush-%s",
@@ -445,9 +451,9 @@ static int bdi_forker_task(void *ptr)
445 * a chance to flush other bdi's to free 451 * a chance to flush other bdi's to free
446 * memory. 452 * memory.
447 */ 453 */
448 spin_lock(&bdi_lock); 454 spin_lock_bh(&bdi_lock);
449 list_add_tail(&bdi->bdi_list, &bdi_pending_list); 455 list_add_tail(&bdi->bdi_list, &bdi_pending_list);
450 spin_unlock(&bdi_lock); 456 spin_unlock_bh(&bdi_lock);
451 457
452 bdi_flush_io(bdi); 458 bdi_flush_io(bdi);
453 } 459 }
@@ -456,6 +462,24 @@ static int bdi_forker_task(void *ptr)
456 return 0; 462 return 0;
457} 463}
458 464
465static void bdi_add_to_pending(struct rcu_head *head)
466{
467 struct backing_dev_info *bdi;
468
469 bdi = container_of(head, struct backing_dev_info, rcu_head);
470 INIT_LIST_HEAD(&bdi->bdi_list);
471
472 spin_lock(&bdi_lock);
473 list_add_tail(&bdi->bdi_list, &bdi_pending_list);
474 spin_unlock(&bdi_lock);
475
476 /*
477 * We are now on the pending list, wake up bdi_forker_task()
478 * to finish the job and add us back to the active bdi_list
479 */
480 wake_up_process(default_backing_dev_info.wb.task);
481}
482
459/* 483/*
460 * Add the default flusher task that gets created for any bdi 484 * Add the default flusher task that gets created for any bdi
461 * that has dirty data pending writeout 485 * that has dirty data pending writeout
@@ -478,16 +502,29 @@ void static bdi_add_default_flusher_task(struct backing_dev_info *bdi)
478 * waiting for previous additions to finish. 502 * waiting for previous additions to finish.
479 */ 503 */
480 if (!test_and_set_bit(BDI_pending, &bdi->state)) { 504 if (!test_and_set_bit(BDI_pending, &bdi->state)) {
481 list_move_tail(&bdi->bdi_list, &bdi_pending_list); 505 list_del_rcu(&bdi->bdi_list);
482 506
483 /* 507 /*
484 * We are now on the pending list, wake up bdi_forker_task() 508 * We must wait for the current RCU period to end before
485 * to finish the job and add us back to the active bdi_list 509 * moving to the pending list. So schedule that operation
510 * from an RCU callback.
486 */ 511 */
487 wake_up_process(default_backing_dev_info.wb.task); 512 call_rcu(&bdi->rcu_head, bdi_add_to_pending);
488 } 513 }
489} 514}
490 515
516/*
517 * Remove bdi from bdi_list, and ensure that it is no longer visible
518 */
519static void bdi_remove_from_list(struct backing_dev_info *bdi)
520{
521 spin_lock_bh(&bdi_lock);
522 list_del_rcu(&bdi->bdi_list);
523 spin_unlock_bh(&bdi_lock);
524
525 synchronize_rcu();
526}
527
491int bdi_register(struct backing_dev_info *bdi, struct device *parent, 528int bdi_register(struct backing_dev_info *bdi, struct device *parent,
492 const char *fmt, ...) 529 const char *fmt, ...)
493{ 530{
@@ -506,9 +543,9 @@ int bdi_register(struct backing_dev_info *bdi, struct device *parent,
506 goto exit; 543 goto exit;
507 } 544 }
508 545
509 spin_lock(&bdi_lock); 546 spin_lock_bh(&bdi_lock);
510 list_add_tail(&bdi->bdi_list, &bdi_list); 547 list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
511 spin_unlock(&bdi_lock); 548 spin_unlock_bh(&bdi_lock);
512 549
513 bdi->dev = dev; 550 bdi->dev = dev;
514 551
@@ -526,9 +563,7 @@ int bdi_register(struct backing_dev_info *bdi, struct device *parent,
526 wb->task = NULL; 563 wb->task = NULL;
527 ret = -ENOMEM; 564 ret = -ENOMEM;
528 565
529 spin_lock(&bdi_lock); 566 bdi_remove_from_list(bdi);
530 list_del(&bdi->bdi_list);
531 spin_unlock(&bdi_lock);
532 goto exit; 567 goto exit;
533 } 568 }
534 } 569 }
@@ -565,9 +600,7 @@ static void bdi_wb_shutdown(struct backing_dev_info *bdi)
565 /* 600 /*
566 * Make sure nobody finds us on the bdi_list anymore 601 * Make sure nobody finds us on the bdi_list anymore
567 */ 602 */
568 spin_lock(&bdi_lock); 603 bdi_remove_from_list(bdi);
569 list_del(&bdi->bdi_list);
570 spin_unlock(&bdi_lock);
571 604
572 /* 605 /*
573 * Finally, kill the kernel threads. We don't need to be RCU 606 * Finally, kill the kernel threads. We don't need to be RCU
@@ -599,6 +632,7 @@ int bdi_init(struct backing_dev_info *bdi)
599 bdi->max_ratio = 100; 632 bdi->max_ratio = 100;
600 bdi->max_prop_frac = PROP_FRAC_BASE; 633 bdi->max_prop_frac = PROP_FRAC_BASE;
601 spin_lock_init(&bdi->wb_lock); 634 spin_lock_init(&bdi->wb_lock);
635 INIT_RCU_HEAD(&bdi->rcu_head);
602 INIT_LIST_HEAD(&bdi->bdi_list); 636 INIT_LIST_HEAD(&bdi->bdi_list);
603 INIT_LIST_HEAD(&bdi->wb_list); 637 INIT_LIST_HEAD(&bdi->wb_list);
604 INIT_LIST_HEAD(&bdi->work_list); 638 INIT_LIST_HEAD(&bdi->work_list);
@@ -634,7 +668,19 @@ void bdi_destroy(struct backing_dev_info *bdi)
634{ 668{
635 int i; 669 int i;
636 670
637 WARN_ON(bdi_has_dirty_io(bdi)); 671 /*
672 * Splice our entries to the default_backing_dev_info, if this
673 * bdi disappears
674 */
675 if (bdi_has_dirty_io(bdi)) {
676 struct bdi_writeback *dst = &default_backing_dev_info.wb;
677
678 spin_lock(&inode_lock);
679 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
680 list_splice(&bdi->wb.b_io, &dst->b_io);
681 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
682 spin_unlock(&inode_lock);
683 }
638 684
639 bdi_unregister(bdi); 685 bdi_unregister(bdi);
640 686
diff --git a/mm/kmemleak-test.c b/mm/kmemleak-test.c
index d5292fc6f523..177a5169bbde 100644
--- a/mm/kmemleak-test.c
+++ b/mm/kmemleak-test.c
@@ -36,7 +36,7 @@ struct test_node {
36}; 36};
37 37
38static LIST_HEAD(test_list); 38static LIST_HEAD(test_list);
39static DEFINE_PER_CPU(void *, test_pointer); 39static DEFINE_PER_CPU(void *, kmemleak_test_pointer);
40 40
41/* 41/*
42 * Some very simple testing. This function needs to be extended for 42 * Some very simple testing. This function needs to be extended for
@@ -86,9 +86,9 @@ static int __init kmemleak_test_init(void)
86 } 86 }
87 87
88 for_each_possible_cpu(i) { 88 for_each_possible_cpu(i) {
89 per_cpu(test_pointer, i) = kmalloc(129, GFP_KERNEL); 89 per_cpu(kmemleak_test_pointer, i) = kmalloc(129, GFP_KERNEL);
90 pr_info("kmemleak: kmalloc(129) = %p\n", 90 pr_info("kmemleak: kmalloc(129) = %p\n",
91 per_cpu(test_pointer, i)); 91 per_cpu(kmemleak_test_pointer, i));
92 } 92 }
93 93
94 return 0; 94 return 0;
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 25e7770309b8..1eea4fa0d410 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -315,7 +315,7 @@ int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
315{ 315{
316 int ret = 0; 316 int ret = 0;
317 317
318 spin_lock(&bdi_lock); 318 spin_lock_bh(&bdi_lock);
319 if (min_ratio > bdi->max_ratio) { 319 if (min_ratio > bdi->max_ratio) {
320 ret = -EINVAL; 320 ret = -EINVAL;
321 } else { 321 } else {
@@ -327,7 +327,7 @@ int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
327 ret = -EINVAL; 327 ret = -EINVAL;
328 } 328 }
329 } 329 }
330 spin_unlock(&bdi_lock); 330 spin_unlock_bh(&bdi_lock);
331 331
332 return ret; 332 return ret;
333} 333}
@@ -339,14 +339,14 @@ int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
339 if (max_ratio > 100) 339 if (max_ratio > 100)
340 return -EINVAL; 340 return -EINVAL;
341 341
342 spin_lock(&bdi_lock); 342 spin_lock_bh(&bdi_lock);
343 if (bdi->min_ratio > max_ratio) { 343 if (bdi->min_ratio > max_ratio) {
344 ret = -EINVAL; 344 ret = -EINVAL;
345 } else { 345 } else {
346 bdi->max_ratio = max_ratio; 346 bdi->max_ratio = max_ratio;
347 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100; 347 bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348 } 348 }
349 spin_unlock(&bdi_lock); 349 spin_unlock_bh(&bdi_lock);
350 350
351 return ret; 351 return ret;
352} 352}
@@ -582,16 +582,8 @@ static void balance_dirty_pages(struct address_space *mapping)
582 if ((laptop_mode && pages_written) || 582 if ((laptop_mode && pages_written) ||
583 (!laptop_mode && ((nr_writeback = global_page_state(NR_FILE_DIRTY) 583 (!laptop_mode && ((nr_writeback = global_page_state(NR_FILE_DIRTY)
584 + global_page_state(NR_UNSTABLE_NFS)) 584 + global_page_state(NR_UNSTABLE_NFS))
585 > background_thresh))) { 585 > background_thresh)))
586 struct writeback_control wbc = { 586 bdi_start_writeback(bdi, nr_writeback);
587 .bdi = bdi,
588 .sync_mode = WB_SYNC_NONE,
589 .nr_to_write = nr_writeback,
590 };
591
592
593 bdi_start_writeback(&wbc);
594 }
595} 587}
596 588
597void set_page_dirty_balance(struct page *page, int page_mkwrite) 589void set_page_dirty_balance(struct page *page, int page_mkwrite)
@@ -604,6 +596,8 @@ void set_page_dirty_balance(struct page *page, int page_mkwrite)
604 } 596 }
605} 597}
606 598
599static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
600
607/** 601/**
608 * balance_dirty_pages_ratelimited_nr - balance dirty memory state 602 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
609 * @mapping: address_space which was dirtied 603 * @mapping: address_space which was dirtied
@@ -621,7 +615,6 @@ void set_page_dirty_balance(struct page *page, int page_mkwrite)
621void balance_dirty_pages_ratelimited_nr(struct address_space *mapping, 615void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
622 unsigned long nr_pages_dirtied) 616 unsigned long nr_pages_dirtied)
623{ 617{
624 static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
625 unsigned long ratelimit; 618 unsigned long ratelimit;
626 unsigned long *p; 619 unsigned long *p;
627 620
@@ -634,7 +627,7 @@ void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
634 * tasks in balance_dirty_pages(). Period. 627 * tasks in balance_dirty_pages(). Period.
635 */ 628 */
636 preempt_disable(); 629 preempt_disable();
637 p = &__get_cpu_var(ratelimits); 630 p = &__get_cpu_var(bdp_ratelimits);
638 *p += nr_pages_dirtied; 631 *p += nr_pages_dirtied;
639 if (unlikely(*p >= ratelimit)) { 632 if (unlikely(*p >= ratelimit)) {
640 *p = 0; 633 *p = 0;
@@ -1019,12 +1012,10 @@ int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1019 1012
1020 if (wbc->nr_to_write <= 0) 1013 if (wbc->nr_to_write <= 0)
1021 return 0; 1014 return 0;
1022 wbc->for_writepages = 1;
1023 if (mapping->a_ops->writepages) 1015 if (mapping->a_ops->writepages)
1024 ret = mapping->a_ops->writepages(mapping, wbc); 1016 ret = mapping->a_ops->writepages(mapping, wbc);
1025 else 1017 else
1026 ret = generic_writepages(mapping, wbc); 1018 ret = generic_writepages(mapping, wbc);
1027 wbc->for_writepages = 0;
1028 return ret; 1019 return ret;
1029} 1020}
1030 1021
diff --git a/mm/percpu.c b/mm/percpu.c
index 3311c8919f37..43d8cacfdaa5 100644
--- a/mm/percpu.c
+++ b/mm/percpu.c
@@ -8,12 +8,13 @@
8 * 8 *
9 * This is percpu allocator which can handle both static and dynamic 9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks in vmalloc area. Each 10 * areas. Percpu areas are allocated in chunks in vmalloc area. Each
11 * chunk is consisted of nr_cpu_ids units and the first chunk is used 11 * chunk is consisted of boot-time determined number of units and the
12 * for static percpu variables in the kernel image (special boot time 12 * first chunk is used for static percpu variables in the kernel image
13 * alloc/init handling necessary as these areas need to be brought up 13 * (special boot time alloc/init handling necessary as these areas
14 * before allocation services are running). Unit grows as necessary 14 * need to be brought up before allocation services are running).
15 * and all units grow or shrink in unison. When a chunk is filled up, 15 * Unit grows as necessary and all units grow or shrink in unison.
16 * another chunk is allocated. ie. in vmalloc area 16 * When a chunk is filled up, another chunk is allocated. ie. in
17 * vmalloc area
17 * 18 *
18 * c0 c1 c2 19 * c0 c1 c2
19 * ------------------- ------------------- ------------ 20 * ------------------- ------------------- ------------
@@ -22,11 +23,13 @@
22 * 23 *
23 * Allocation is done in offset-size areas of single unit space. Ie, 24 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, 25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. Percpu access can be done by configuring 26 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * percpu base registers pcpu_unit_size apart. 27 * cpus. On NUMA, the mapping can be non-linear and even sparse.
28 * Percpu access can be done by configuring percpu base registers
29 * according to cpu to unit mapping and pcpu_unit_size.
27 * 30 *
28 * There are usually many small percpu allocations many of them as 31 * There are usually many small percpu allocations many of them being
29 * small as 4 bytes. The allocator organizes chunks into lists 32 * as small as 4 bytes. The allocator organizes chunks into lists
30 * according to free size and tries to allocate from the fullest one. 33 * according to free size and tries to allocate from the fullest one.
31 * Each chunk keeps the maximum contiguous area size hint which is 34 * Each chunk keeps the maximum contiguous area size hint which is
32 * guaranteed to be eqaul to or larger than the maximum contiguous 35 * guaranteed to be eqaul to or larger than the maximum contiguous
@@ -43,7 +46,7 @@
43 * 46 *
44 * To use this allocator, arch code should do the followings. 47 * To use this allocator, arch code should do the followings.
45 * 48 *
46 * - define CONFIG_HAVE_DYNAMIC_PER_CPU_AREA 49 * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA
47 * 50 *
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate 51 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be 52 * regular address to percpu pointer and back if they need to be
@@ -55,7 +58,9 @@
55 58
56#include <linux/bitmap.h> 59#include <linux/bitmap.h>
57#include <linux/bootmem.h> 60#include <linux/bootmem.h>
61#include <linux/err.h>
58#include <linux/list.h> 62#include <linux/list.h>
63#include <linux/log2.h>
59#include <linux/mm.h> 64#include <linux/mm.h>
60#include <linux/module.h> 65#include <linux/module.h>
61#include <linux/mutex.h> 66#include <linux/mutex.h>
@@ -89,25 +94,38 @@ struct pcpu_chunk {
89 struct list_head list; /* linked to pcpu_slot lists */ 94 struct list_head list; /* linked to pcpu_slot lists */
90 int free_size; /* free bytes in the chunk */ 95 int free_size; /* free bytes in the chunk */
91 int contig_hint; /* max contiguous size hint */ 96 int contig_hint; /* max contiguous size hint */
92 struct vm_struct *vm; /* mapped vmalloc region */ 97 void *base_addr; /* base address of this chunk */
93 int map_used; /* # of map entries used */ 98 int map_used; /* # of map entries used */
94 int map_alloc; /* # of map entries allocated */ 99 int map_alloc; /* # of map entries allocated */
95 int *map; /* allocation map */ 100 int *map; /* allocation map */
101 struct vm_struct **vms; /* mapped vmalloc regions */
96 bool immutable; /* no [de]population allowed */ 102 bool immutable; /* no [de]population allowed */
97 struct page **page; /* points to page array */ 103 unsigned long populated[]; /* populated bitmap */
98 struct page *page_ar[]; /* #cpus * UNIT_PAGES */
99}; 104};
100 105
101static int pcpu_unit_pages __read_mostly; 106static int pcpu_unit_pages __read_mostly;
102static int pcpu_unit_size __read_mostly; 107static int pcpu_unit_size __read_mostly;
103static int pcpu_chunk_size __read_mostly; 108static int pcpu_nr_units __read_mostly;
109static int pcpu_atom_size __read_mostly;
104static int pcpu_nr_slots __read_mostly; 110static int pcpu_nr_slots __read_mostly;
105static size_t pcpu_chunk_struct_size __read_mostly; 111static size_t pcpu_chunk_struct_size __read_mostly;
106 112
113/* cpus with the lowest and highest unit numbers */
114static unsigned int pcpu_first_unit_cpu __read_mostly;
115static unsigned int pcpu_last_unit_cpu __read_mostly;
116
107/* the address of the first chunk which starts with the kernel static area */ 117/* the address of the first chunk which starts with the kernel static area */
108void *pcpu_base_addr __read_mostly; 118void *pcpu_base_addr __read_mostly;
109EXPORT_SYMBOL_GPL(pcpu_base_addr); 119EXPORT_SYMBOL_GPL(pcpu_base_addr);
110 120
121static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
122const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
123
124/* group information, used for vm allocation */
125static int pcpu_nr_groups __read_mostly;
126static const unsigned long *pcpu_group_offsets __read_mostly;
127static const size_t *pcpu_group_sizes __read_mostly;
128
111/* 129/*
112 * The first chunk which always exists. Note that unlike other 130 * The first chunk which always exists. Note that unlike other
113 * chunks, this one can be allocated and mapped in several different 131 * chunks, this one can be allocated and mapped in several different
@@ -129,9 +147,9 @@ static int pcpu_reserved_chunk_limit;
129 * Synchronization rules. 147 * Synchronization rules.
130 * 148 *
131 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former 149 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
132 * protects allocation/reclaim paths, chunks and chunk->page arrays. 150 * protects allocation/reclaim paths, chunks, populated bitmap and
133 * The latter is a spinlock and protects the index data structures - 151 * vmalloc mapping. The latter is a spinlock and protects the index
134 * chunk slots, chunks and area maps in chunks. 152 * data structures - chunk slots, chunks and area maps in chunks.
135 * 153 *
136 * During allocation, pcpu_alloc_mutex is kept locked all the time and 154 * During allocation, pcpu_alloc_mutex is kept locked all the time and
137 * pcpu_lock is grabbed and released as necessary. All actual memory 155 * pcpu_lock is grabbed and released as necessary. All actual memory
@@ -178,31 +196,23 @@ static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
178 196
179static int pcpu_page_idx(unsigned int cpu, int page_idx) 197static int pcpu_page_idx(unsigned int cpu, int page_idx)
180{ 198{
181 return cpu * pcpu_unit_pages + page_idx; 199 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
182}
183
184static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk,
185 unsigned int cpu, int page_idx)
186{
187 return &chunk->page[pcpu_page_idx(cpu, page_idx)];
188} 200}
189 201
190static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, 202static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
191 unsigned int cpu, int page_idx) 203 unsigned int cpu, int page_idx)
192{ 204{
193 return (unsigned long)chunk->vm->addr + 205 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
194 (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); 206 (page_idx << PAGE_SHIFT);
195} 207}
196 208
197static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk, 209static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
198 int page_idx) 210 unsigned int cpu, int page_idx)
199{ 211{
200 /* 212 /* must not be used on pre-mapped chunk */
201 * Any possible cpu id can be used here, so there's no need to 213 WARN_ON(chunk->immutable);
202 * worry about preemption or cpu hotplug. 214
203 */ 215 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
204 return *pcpu_chunk_pagep(chunk, raw_smp_processor_id(),
205 page_idx) != NULL;
206} 216}
207 217
208/* set the pointer to a chunk in a page struct */ 218/* set the pointer to a chunk in a page struct */
@@ -217,6 +227,34 @@ static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
217 return (struct pcpu_chunk *)page->index; 227 return (struct pcpu_chunk *)page->index;
218} 228}
219 229
230static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
231{
232 *rs = find_next_zero_bit(chunk->populated, end, *rs);
233 *re = find_next_bit(chunk->populated, end, *rs + 1);
234}
235
236static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
237{
238 *rs = find_next_bit(chunk->populated, end, *rs);
239 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
240}
241
242/*
243 * (Un)populated page region iterators. Iterate over (un)populated
244 * page regions betwen @start and @end in @chunk. @rs and @re should
245 * be integer variables and will be set to start and end page index of
246 * the current region.
247 */
248#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
249 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
250 (rs) < (re); \
251 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
252
253#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
254 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
255 (rs) < (re); \
256 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
257
220/** 258/**
221 * pcpu_mem_alloc - allocate memory 259 * pcpu_mem_alloc - allocate memory
222 * @size: bytes to allocate 260 * @size: bytes to allocate
@@ -292,10 +330,10 @@ static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
292 */ 330 */
293static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) 331static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
294{ 332{
295 void *first_start = pcpu_first_chunk->vm->addr; 333 void *first_start = pcpu_first_chunk->base_addr;
296 334
297 /* is it in the first chunk? */ 335 /* is it in the first chunk? */
298 if (addr >= first_start && addr < first_start + pcpu_chunk_size) { 336 if (addr >= first_start && addr < first_start + pcpu_unit_size) {
299 /* is it in the reserved area? */ 337 /* is it in the reserved area? */
300 if (addr < first_start + pcpu_reserved_chunk_limit) 338 if (addr < first_start + pcpu_reserved_chunk_limit)
301 return pcpu_reserved_chunk; 339 return pcpu_reserved_chunk;
@@ -309,7 +347,7 @@ static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
309 * space. Note that any possible cpu id can be used here, so 347 * space. Note that any possible cpu id can be used here, so
310 * there's no need to worry about preemption or cpu hotplug. 348 * there's no need to worry about preemption or cpu hotplug.
311 */ 349 */
312 addr += raw_smp_processor_id() * pcpu_unit_size; 350 addr += pcpu_unit_offsets[raw_smp_processor_id()];
313 return pcpu_get_page_chunk(vmalloc_to_page(addr)); 351 return pcpu_get_page_chunk(vmalloc_to_page(addr));
314} 352}
315 353
@@ -558,125 +596,327 @@ static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
558} 596}
559 597
560/** 598/**
561 * pcpu_unmap - unmap pages out of a pcpu_chunk 599 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
562 * @chunk: chunk of interest 600 * @chunk: chunk of interest
563 * @page_start: page index of the first page to unmap 601 * @bitmapp: output parameter for bitmap
564 * @page_end: page index of the last page to unmap + 1 602 * @may_alloc: may allocate the array
565 * @flush_tlb: whether to flush tlb or not
566 * 603 *
567 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. 604 * Returns pointer to array of pointers to struct page and bitmap,
568 * If @flush is true, vcache is flushed before unmapping and tlb 605 * both of which can be indexed with pcpu_page_idx(). The returned
569 * after. 606 * array is cleared to zero and *@bitmapp is copied from
607 * @chunk->populated. Note that there is only one array and bitmap
608 * and access exclusion is the caller's responsibility.
609 *
610 * CONTEXT:
611 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
612 * Otherwise, don't care.
613 *
614 * RETURNS:
615 * Pointer to temp pages array on success, NULL on failure.
570 */ 616 */
571static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end, 617static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
572 bool flush_tlb) 618 unsigned long **bitmapp,
619 bool may_alloc)
573{ 620{
574 unsigned int last = nr_cpu_ids - 1; 621 static struct page **pages;
575 unsigned int cpu; 622 static unsigned long *bitmap;
623 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
624 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
625 sizeof(unsigned long);
626
627 if (!pages || !bitmap) {
628 if (may_alloc && !pages)
629 pages = pcpu_mem_alloc(pages_size);
630 if (may_alloc && !bitmap)
631 bitmap = pcpu_mem_alloc(bitmap_size);
632 if (!pages || !bitmap)
633 return NULL;
634 }
576 635
577 /* unmap must not be done on immutable chunk */ 636 memset(pages, 0, pages_size);
578 WARN_ON(chunk->immutable); 637 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
579 638
580 /* 639 *bitmapp = bitmap;
581 * Each flushing trial can be very expensive, issue flush on 640 return pages;
582 * the whole region at once rather than doing it for each cpu. 641}
583 * This could be an overkill but is more scalable.
584 */
585 flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start),
586 pcpu_chunk_addr(chunk, last, page_end));
587 642
588 for_each_possible_cpu(cpu) 643/**
589 unmap_kernel_range_noflush( 644 * pcpu_free_pages - free pages which were allocated for @chunk
590 pcpu_chunk_addr(chunk, cpu, page_start), 645 * @chunk: chunk pages were allocated for
591 (page_end - page_start) << PAGE_SHIFT); 646 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
592 647 * @populated: populated bitmap
593 /* ditto as flush_cache_vunmap() */ 648 * @page_start: page index of the first page to be freed
594 if (flush_tlb) 649 * @page_end: page index of the last page to be freed + 1
595 flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start), 650 *
596 pcpu_chunk_addr(chunk, last, page_end)); 651 * Free pages [@page_start and @page_end) in @pages for all units.
652 * The pages were allocated for @chunk.
653 */
654static void pcpu_free_pages(struct pcpu_chunk *chunk,
655 struct page **pages, unsigned long *populated,
656 int page_start, int page_end)
657{
658 unsigned int cpu;
659 int i;
660
661 for_each_possible_cpu(cpu) {
662 for (i = page_start; i < page_end; i++) {
663 struct page *page = pages[pcpu_page_idx(cpu, i)];
664
665 if (page)
666 __free_page(page);
667 }
668 }
597} 669}
598 670
599/** 671/**
600 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk 672 * pcpu_alloc_pages - allocates pages for @chunk
601 * @chunk: chunk to depopulate 673 * @chunk: target chunk
602 * @off: offset to the area to depopulate 674 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
603 * @size: size of the area to depopulate in bytes 675 * @populated: populated bitmap
604 * @flush: whether to flush cache and tlb or not 676 * @page_start: page index of the first page to be allocated
605 * 677 * @page_end: page index of the last page to be allocated + 1
606 * For each cpu, depopulate and unmap pages [@page_start,@page_end) 678 *
607 * from @chunk. If @flush is true, vcache is flushed before unmapping 679 * Allocate pages [@page_start,@page_end) into @pages for all units.
608 * and tlb after. 680 * The allocation is for @chunk. Percpu core doesn't care about the
609 * 681 * content of @pages and will pass it verbatim to pcpu_map_pages().
610 * CONTEXT:
611 * pcpu_alloc_mutex.
612 */ 682 */
613static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size, 683static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
614 bool flush) 684 struct page **pages, unsigned long *populated,
685 int page_start, int page_end)
615{ 686{
616 int page_start = PFN_DOWN(off); 687 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
617 int page_end = PFN_UP(off + size);
618 int unmap_start = -1;
619 int uninitialized_var(unmap_end);
620 unsigned int cpu; 688 unsigned int cpu;
621 int i; 689 int i;
622 690
623 for (i = page_start; i < page_end; i++) { 691 for_each_possible_cpu(cpu) {
624 for_each_possible_cpu(cpu) { 692 for (i = page_start; i < page_end; i++) {
625 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); 693 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
694
695 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
696 if (!*pagep) {
697 pcpu_free_pages(chunk, pages, populated,
698 page_start, page_end);
699 return -ENOMEM;
700 }
701 }
702 }
703 return 0;
704}
626 705
627 if (!*pagep) 706/**
628 continue; 707 * pcpu_pre_unmap_flush - flush cache prior to unmapping
708 * @chunk: chunk the regions to be flushed belongs to
709 * @page_start: page index of the first page to be flushed
710 * @page_end: page index of the last page to be flushed + 1
711 *
712 * Pages in [@page_start,@page_end) of @chunk are about to be
713 * unmapped. Flush cache. As each flushing trial can be very
714 * expensive, issue flush on the whole region at once rather than
715 * doing it for each cpu. This could be an overkill but is more
716 * scalable.
717 */
718static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
719 int page_start, int page_end)
720{
721 flush_cache_vunmap(
722 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
723 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
724}
725
726static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
727{
728 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
729}
629 730
630 __free_page(*pagep); 731/**
732 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
733 * @chunk: chunk of interest
734 * @pages: pages array which can be used to pass information to free
735 * @populated: populated bitmap
736 * @page_start: page index of the first page to unmap
737 * @page_end: page index of the last page to unmap + 1
738 *
739 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
740 * Corresponding elements in @pages were cleared by the caller and can
741 * be used to carry information to pcpu_free_pages() which will be
742 * called after all unmaps are finished. The caller should call
743 * proper pre/post flush functions.
744 */
745static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
746 struct page **pages, unsigned long *populated,
747 int page_start, int page_end)
748{
749 unsigned int cpu;
750 int i;
631 751
632 /* 752 for_each_possible_cpu(cpu) {
633 * If it's partial depopulation, it might get 753 for (i = page_start; i < page_end; i++) {
634 * populated or depopulated again. Mark the 754 struct page *page;
635 * page gone.
636 */
637 *pagep = NULL;
638 755
639 unmap_start = unmap_start < 0 ? i : unmap_start; 756 page = pcpu_chunk_page(chunk, cpu, i);
640 unmap_end = i + 1; 757 WARN_ON(!page);
758 pages[pcpu_page_idx(cpu, i)] = page;
641 } 759 }
760 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
761 page_end - page_start);
642 } 762 }
643 763
644 if (unmap_start >= 0) 764 for (i = page_start; i < page_end; i++)
645 pcpu_unmap(chunk, unmap_start, unmap_end, flush); 765 __clear_bit(i, populated);
766}
767
768/**
769 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
770 * @chunk: pcpu_chunk the regions to be flushed belong to
771 * @page_start: page index of the first page to be flushed
772 * @page_end: page index of the last page to be flushed + 1
773 *
774 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
775 * TLB for the regions. This can be skipped if the area is to be
776 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
777 *
778 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
779 * for the whole region.
780 */
781static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
782 int page_start, int page_end)
783{
784 flush_tlb_kernel_range(
785 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
786 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
787}
788
789static int __pcpu_map_pages(unsigned long addr, struct page **pages,
790 int nr_pages)
791{
792 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
793 PAGE_KERNEL, pages);
646} 794}
647 795
648/** 796/**
649 * pcpu_map - map pages into a pcpu_chunk 797 * pcpu_map_pages - map pages into a pcpu_chunk
650 * @chunk: chunk of interest 798 * @chunk: chunk of interest
799 * @pages: pages array containing pages to be mapped
800 * @populated: populated bitmap
651 * @page_start: page index of the first page to map 801 * @page_start: page index of the first page to map
652 * @page_end: page index of the last page to map + 1 802 * @page_end: page index of the last page to map + 1
653 * 803 *
654 * For each cpu, map pages [@page_start,@page_end) into @chunk. 804 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
655 * vcache is flushed afterwards. 805 * caller is responsible for calling pcpu_post_map_flush() after all
806 * mappings are complete.
807 *
808 * This function is responsible for setting corresponding bits in
809 * @chunk->populated bitmap and whatever is necessary for reverse
810 * lookup (addr -> chunk).
656 */ 811 */
657static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end) 812static int pcpu_map_pages(struct pcpu_chunk *chunk,
813 struct page **pages, unsigned long *populated,
814 int page_start, int page_end)
658{ 815{
659 unsigned int last = nr_cpu_ids - 1; 816 unsigned int cpu, tcpu;
660 unsigned int cpu; 817 int i, err;
661 int err;
662
663 /* map must not be done on immutable chunk */
664 WARN_ON(chunk->immutable);
665 818
666 for_each_possible_cpu(cpu) { 819 for_each_possible_cpu(cpu) {
667 err = map_kernel_range_noflush( 820 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
668 pcpu_chunk_addr(chunk, cpu, page_start), 821 &pages[pcpu_page_idx(cpu, page_start)],
669 (page_end - page_start) << PAGE_SHIFT, 822 page_end - page_start);
670 PAGE_KERNEL,
671 pcpu_chunk_pagep(chunk, cpu, page_start));
672 if (err < 0) 823 if (err < 0)
673 return err; 824 goto err;
825 }
826
827 /* mapping successful, link chunk and mark populated */
828 for (i = page_start; i < page_end; i++) {
829 for_each_possible_cpu(cpu)
830 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
831 chunk);
832 __set_bit(i, populated);
674 } 833 }
675 834
676 /* flush at once, please read comments in pcpu_unmap() */
677 flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start),
678 pcpu_chunk_addr(chunk, last, page_end));
679 return 0; 835 return 0;
836
837err:
838 for_each_possible_cpu(tcpu) {
839 if (tcpu == cpu)
840 break;
841 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
842 page_end - page_start);
843 }
844 return err;
845}
846
847/**
848 * pcpu_post_map_flush - flush cache after mapping
849 * @chunk: pcpu_chunk the regions to be flushed belong to
850 * @page_start: page index of the first page to be flushed
851 * @page_end: page index of the last page to be flushed + 1
852 *
853 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
854 * cache.
855 *
856 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
857 * for the whole region.
858 */
859static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
860 int page_start, int page_end)
861{
862 flush_cache_vmap(
863 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
864 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
865}
866
867/**
868 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
869 * @chunk: chunk to depopulate
870 * @off: offset to the area to depopulate
871 * @size: size of the area to depopulate in bytes
872 * @flush: whether to flush cache and tlb or not
873 *
874 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
875 * from @chunk. If @flush is true, vcache is flushed before unmapping
876 * and tlb after.
877 *
878 * CONTEXT:
879 * pcpu_alloc_mutex.
880 */
881static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
882{
883 int page_start = PFN_DOWN(off);
884 int page_end = PFN_UP(off + size);
885 struct page **pages;
886 unsigned long *populated;
887 int rs, re;
888
889 /* quick path, check whether it's empty already */
890 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
891 if (rs == page_start && re == page_end)
892 return;
893 break;
894 }
895
896 /* immutable chunks can't be depopulated */
897 WARN_ON(chunk->immutable);
898
899 /*
900 * If control reaches here, there must have been at least one
901 * successful population attempt so the temp pages array must
902 * be available now.
903 */
904 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
905 BUG_ON(!pages);
906
907 /* unmap and free */
908 pcpu_pre_unmap_flush(chunk, page_start, page_end);
909
910 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
911 pcpu_unmap_pages(chunk, pages, populated, rs, re);
912
913 /* no need to flush tlb, vmalloc will handle it lazily */
914
915 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
916 pcpu_free_pages(chunk, pages, populated, rs, re);
917
918 /* commit new bitmap */
919 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
680} 920}
681 921
682/** 922/**
@@ -693,58 +933,68 @@ static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
693 */ 933 */
694static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) 934static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
695{ 935{
696 const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
697 int page_start = PFN_DOWN(off); 936 int page_start = PFN_DOWN(off);
698 int page_end = PFN_UP(off + size); 937 int page_end = PFN_UP(off + size);
699 int map_start = -1; 938 int free_end = page_start, unmap_end = page_start;
700 int uninitialized_var(map_end); 939 struct page **pages;
940 unsigned long *populated;
701 unsigned int cpu; 941 unsigned int cpu;
702 int i; 942 int rs, re, rc;
703 943
704 for (i = page_start; i < page_end; i++) { 944 /* quick path, check whether all pages are already there */
705 if (pcpu_chunk_page_occupied(chunk, i)) { 945 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) {
706 if (map_start >= 0) { 946 if (rs == page_start && re == page_end)
707 if (pcpu_map(chunk, map_start, map_end)) 947 goto clear;
708 goto err; 948 break;
709 map_start = -1; 949 }
710 }
711 continue;
712 }
713 950
714 map_start = map_start < 0 ? i : map_start; 951 /* need to allocate and map pages, this chunk can't be immutable */
715 map_end = i + 1; 952 WARN_ON(chunk->immutable);
716 953
717 for_each_possible_cpu(cpu) { 954 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
718 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); 955 if (!pages)
956 return -ENOMEM;
719 957
720 *pagep = alloc_pages_node(cpu_to_node(cpu), 958 /* alloc and map */
721 alloc_mask, 0); 959 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
722 if (!*pagep) 960 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
723 goto err; 961 if (rc)
724 pcpu_set_page_chunk(*pagep, chunk); 962 goto err_free;
725 } 963 free_end = re;
726 } 964 }
727 965
728 if (map_start >= 0 && pcpu_map(chunk, map_start, map_end)) 966 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
729 goto err; 967 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
968 if (rc)
969 goto err_unmap;
970 unmap_end = re;
971 }
972 pcpu_post_map_flush(chunk, page_start, page_end);
730 973
974 /* commit new bitmap */
975 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
976clear:
731 for_each_possible_cpu(cpu) 977 for_each_possible_cpu(cpu)
732 memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0, 978 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
733 size);
734
735 return 0; 979 return 0;
736err: 980
737 /* likely under heavy memory pressure, give memory back */ 981err_unmap:
738 pcpu_depopulate_chunk(chunk, off, size, true); 982 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
739 return -ENOMEM; 983 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
984 pcpu_unmap_pages(chunk, pages, populated, rs, re);
985 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
986err_free:
987 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
988 pcpu_free_pages(chunk, pages, populated, rs, re);
989 return rc;
740} 990}
741 991
742static void free_pcpu_chunk(struct pcpu_chunk *chunk) 992static void free_pcpu_chunk(struct pcpu_chunk *chunk)
743{ 993{
744 if (!chunk) 994 if (!chunk)
745 return; 995 return;
746 if (chunk->vm) 996 if (chunk->vms)
747 free_vm_area(chunk->vm); 997 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
748 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); 998 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
749 kfree(chunk); 999 kfree(chunk);
750} 1000}
@@ -760,10 +1010,11 @@ static struct pcpu_chunk *alloc_pcpu_chunk(void)
760 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); 1010 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
761 chunk->map_alloc = PCPU_DFL_MAP_ALLOC; 1011 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
762 chunk->map[chunk->map_used++] = pcpu_unit_size; 1012 chunk->map[chunk->map_used++] = pcpu_unit_size;
763 chunk->page = chunk->page_ar;
764 1013
765 chunk->vm = get_vm_area(pcpu_chunk_size, VM_ALLOC); 1014 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
766 if (!chunk->vm) { 1015 pcpu_nr_groups, pcpu_atom_size,
1016 GFP_KERNEL);
1017 if (!chunk->vms) {
767 free_pcpu_chunk(chunk); 1018 free_pcpu_chunk(chunk);
768 return NULL; 1019 return NULL;
769 } 1020 }
@@ -771,6 +1022,7 @@ static struct pcpu_chunk *alloc_pcpu_chunk(void)
771 INIT_LIST_HEAD(&chunk->list); 1022 INIT_LIST_HEAD(&chunk->list);
772 chunk->free_size = pcpu_unit_size; 1023 chunk->free_size = pcpu_unit_size;
773 chunk->contig_hint = pcpu_unit_size; 1024 chunk->contig_hint = pcpu_unit_size;
1025 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
774 1026
775 return chunk; 1027 return chunk;
776} 1028}
@@ -860,7 +1112,8 @@ area_found:
860 1112
861 mutex_unlock(&pcpu_alloc_mutex); 1113 mutex_unlock(&pcpu_alloc_mutex);
862 1114
863 return __addr_to_pcpu_ptr(chunk->vm->addr + off); 1115 /* return address relative to base address */
1116 return __addr_to_pcpu_ptr(chunk->base_addr + off);
864 1117
865fail_unlock: 1118fail_unlock:
866 spin_unlock_irq(&pcpu_lock); 1119 spin_unlock_irq(&pcpu_lock);
@@ -938,12 +1191,13 @@ static void pcpu_reclaim(struct work_struct *work)
938 } 1191 }
939 1192
940 spin_unlock_irq(&pcpu_lock); 1193 spin_unlock_irq(&pcpu_lock);
941 mutex_unlock(&pcpu_alloc_mutex);
942 1194
943 list_for_each_entry_safe(chunk, next, &todo, list) { 1195 list_for_each_entry_safe(chunk, next, &todo, list) {
944 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false); 1196 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
945 free_pcpu_chunk(chunk); 1197 free_pcpu_chunk(chunk);
946 } 1198 }
1199
1200 mutex_unlock(&pcpu_alloc_mutex);
947} 1201}
948 1202
949/** 1203/**
@@ -968,7 +1222,7 @@ void free_percpu(void *ptr)
968 spin_lock_irqsave(&pcpu_lock, flags); 1222 spin_lock_irqsave(&pcpu_lock, flags);
969 1223
970 chunk = pcpu_chunk_addr_search(addr); 1224 chunk = pcpu_chunk_addr_search(addr);
971 off = addr - chunk->vm->addr; 1225 off = addr - chunk->base_addr;
972 1226
973 pcpu_free_area(chunk, off); 1227 pcpu_free_area(chunk, off);
974 1228
@@ -987,30 +1241,295 @@ void free_percpu(void *ptr)
987} 1241}
988EXPORT_SYMBOL_GPL(free_percpu); 1242EXPORT_SYMBOL_GPL(free_percpu);
989 1243
1244static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1245 size_t reserved_size,
1246 ssize_t *dyn_sizep)
1247{
1248 size_t size_sum;
1249
1250 size_sum = PFN_ALIGN(static_size + reserved_size +
1251 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1252 if (*dyn_sizep != 0)
1253 *dyn_sizep = size_sum - static_size - reserved_size;
1254
1255 return size_sum;
1256}
1257
990/** 1258/**
991 * pcpu_setup_first_chunk - initialize the first percpu chunk 1259 * pcpu_alloc_alloc_info - allocate percpu allocation info
992 * @get_page_fn: callback to fetch page pointer 1260 * @nr_groups: the number of groups
993 * @static_size: the size of static percpu area in bytes 1261 * @nr_units: the number of units
1262 *
1263 * Allocate ai which is large enough for @nr_groups groups containing
1264 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1265 * cpu_map array which is long enough for @nr_units and filled with
1266 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1267 * pointer of other groups.
1268 *
1269 * RETURNS:
1270 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1271 * failure.
1272 */
1273struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1274 int nr_units)
1275{
1276 struct pcpu_alloc_info *ai;
1277 size_t base_size, ai_size;
1278 void *ptr;
1279 int unit;
1280
1281 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1282 __alignof__(ai->groups[0].cpu_map[0]));
1283 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1284
1285 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1286 if (!ptr)
1287 return NULL;
1288 ai = ptr;
1289 ptr += base_size;
1290
1291 ai->groups[0].cpu_map = ptr;
1292
1293 for (unit = 0; unit < nr_units; unit++)
1294 ai->groups[0].cpu_map[unit] = NR_CPUS;
1295
1296 ai->nr_groups = nr_groups;
1297 ai->__ai_size = PFN_ALIGN(ai_size);
1298
1299 return ai;
1300}
1301
1302/**
1303 * pcpu_free_alloc_info - free percpu allocation info
1304 * @ai: pcpu_alloc_info to free
1305 *
1306 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1307 */
1308void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1309{
1310 free_bootmem(__pa(ai), ai->__ai_size);
1311}
1312
1313/**
1314 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
994 * @reserved_size: the size of reserved percpu area in bytes 1315 * @reserved_size: the size of reserved percpu area in bytes
995 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1316 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
996 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto 1317 * @atom_size: allocation atom size
997 * @base_addr: mapped address, NULL for auto 1318 * @cpu_distance_fn: callback to determine distance between cpus, optional
998 * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary 1319 *
1320 * This function determines grouping of units, their mappings to cpus
1321 * and other parameters considering needed percpu size, allocation
1322 * atom size and distances between CPUs.
1323 *
1324 * Groups are always mutliples of atom size and CPUs which are of
1325 * LOCAL_DISTANCE both ways are grouped together and share space for
1326 * units in the same group. The returned configuration is guaranteed
1327 * to have CPUs on different nodes on different groups and >=75% usage
1328 * of allocated virtual address space.
1329 *
1330 * RETURNS:
1331 * On success, pointer to the new allocation_info is returned. On
1332 * failure, ERR_PTR value is returned.
1333 */
1334struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1335 size_t reserved_size, ssize_t dyn_size,
1336 size_t atom_size,
1337 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1338{
1339 static int group_map[NR_CPUS] __initdata;
1340 static int group_cnt[NR_CPUS] __initdata;
1341 const size_t static_size = __per_cpu_end - __per_cpu_start;
1342 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
1343 size_t size_sum, min_unit_size, alloc_size;
1344 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1345 int last_allocs, group, unit;
1346 unsigned int cpu, tcpu;
1347 struct pcpu_alloc_info *ai;
1348 unsigned int *cpu_map;
1349
1350 /*
1351 * Determine min_unit_size, alloc_size and max_upa such that
1352 * alloc_size is multiple of atom_size and is the smallest
1353 * which can accomodate 4k aligned segments which are equal to
1354 * or larger than min_unit_size.
1355 */
1356 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
1357 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1358
1359 alloc_size = roundup(min_unit_size, atom_size);
1360 upa = alloc_size / min_unit_size;
1361 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1362 upa--;
1363 max_upa = upa;
1364
1365 /* group cpus according to their proximity */
1366 for_each_possible_cpu(cpu) {
1367 group = 0;
1368 next_group:
1369 for_each_possible_cpu(tcpu) {
1370 if (cpu == tcpu)
1371 break;
1372 if (group_map[tcpu] == group && cpu_distance_fn &&
1373 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1374 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1375 group++;
1376 nr_groups = max(nr_groups, group + 1);
1377 goto next_group;
1378 }
1379 }
1380 group_map[cpu] = group;
1381 group_cnt[group]++;
1382 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1383 }
1384
1385 /*
1386 * Expand unit size until address space usage goes over 75%
1387 * and then as much as possible without using more address
1388 * space.
1389 */
1390 last_allocs = INT_MAX;
1391 for (upa = max_upa; upa; upa--) {
1392 int allocs = 0, wasted = 0;
1393
1394 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1395 continue;
1396
1397 for (group = 0; group < nr_groups; group++) {
1398 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1399 allocs += this_allocs;
1400 wasted += this_allocs * upa - group_cnt[group];
1401 }
1402
1403 /*
1404 * Don't accept if wastage is over 25%. The
1405 * greater-than comparison ensures upa==1 always
1406 * passes the following check.
1407 */
1408 if (wasted > num_possible_cpus() / 3)
1409 continue;
1410
1411 /* and then don't consume more memory */
1412 if (allocs > last_allocs)
1413 break;
1414 last_allocs = allocs;
1415 best_upa = upa;
1416 }
1417 upa = best_upa;
1418
1419 /* allocate and fill alloc_info */
1420 for (group = 0; group < nr_groups; group++)
1421 nr_units += roundup(group_cnt[group], upa);
1422
1423 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1424 if (!ai)
1425 return ERR_PTR(-ENOMEM);
1426 cpu_map = ai->groups[0].cpu_map;
1427
1428 for (group = 0; group < nr_groups; group++) {
1429 ai->groups[group].cpu_map = cpu_map;
1430 cpu_map += roundup(group_cnt[group], upa);
1431 }
1432
1433 ai->static_size = static_size;
1434 ai->reserved_size = reserved_size;
1435 ai->dyn_size = dyn_size;
1436 ai->unit_size = alloc_size / upa;
1437 ai->atom_size = atom_size;
1438 ai->alloc_size = alloc_size;
1439
1440 for (group = 0, unit = 0; group_cnt[group]; group++) {
1441 struct pcpu_group_info *gi = &ai->groups[group];
1442
1443 /*
1444 * Initialize base_offset as if all groups are located
1445 * back-to-back. The caller should update this to
1446 * reflect actual allocation.
1447 */
1448 gi->base_offset = unit * ai->unit_size;
1449
1450 for_each_possible_cpu(cpu)
1451 if (group_map[cpu] == group)
1452 gi->cpu_map[gi->nr_units++] = cpu;
1453 gi->nr_units = roundup(gi->nr_units, upa);
1454 unit += gi->nr_units;
1455 }
1456 BUG_ON(unit != nr_units);
1457
1458 return ai;
1459}
1460
1461/**
1462 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1463 * @lvl: loglevel
1464 * @ai: allocation info to dump
1465 *
1466 * Print out information about @ai using loglevel @lvl.
1467 */
1468static void pcpu_dump_alloc_info(const char *lvl,
1469 const struct pcpu_alloc_info *ai)
1470{
1471 int group_width = 1, cpu_width = 1, width;
1472 char empty_str[] = "--------";
1473 int alloc = 0, alloc_end = 0;
1474 int group, v;
1475 int upa, apl; /* units per alloc, allocs per line */
1476
1477 v = ai->nr_groups;
1478 while (v /= 10)
1479 group_width++;
1480
1481 v = num_possible_cpus();
1482 while (v /= 10)
1483 cpu_width++;
1484 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1485
1486 upa = ai->alloc_size / ai->unit_size;
1487 width = upa * (cpu_width + 1) + group_width + 3;
1488 apl = rounddown_pow_of_two(max(60 / width, 1));
1489
1490 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1491 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1492 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1493
1494 for (group = 0; group < ai->nr_groups; group++) {
1495 const struct pcpu_group_info *gi = &ai->groups[group];
1496 int unit = 0, unit_end = 0;
1497
1498 BUG_ON(gi->nr_units % upa);
1499 for (alloc_end += gi->nr_units / upa;
1500 alloc < alloc_end; alloc++) {
1501 if (!(alloc % apl)) {
1502 printk("\n");
1503 printk("%spcpu-alloc: ", lvl);
1504 }
1505 printk("[%0*d] ", group_width, group);
1506
1507 for (unit_end += upa; unit < unit_end; unit++)
1508 if (gi->cpu_map[unit] != NR_CPUS)
1509 printk("%0*d ", cpu_width,
1510 gi->cpu_map[unit]);
1511 else
1512 printk("%s ", empty_str);
1513 }
1514 }
1515 printk("\n");
1516}
1517
1518/**
1519 * pcpu_setup_first_chunk - initialize the first percpu chunk
1520 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1521 * @base_addr: mapped address
999 * 1522 *
1000 * Initialize the first percpu chunk which contains the kernel static 1523 * Initialize the first percpu chunk which contains the kernel static
1001 * perpcu area. This function is to be called from arch percpu area 1524 * perpcu area. This function is to be called from arch percpu area
1002 * setup path. The first two parameters are mandatory. The rest are 1525 * setup path.
1003 * optional. 1526 *
1004 * 1527 * @ai contains all information necessary to initialize the first
1005 * @get_page_fn() should return pointer to percpu page given cpu 1528 * chunk and prime the dynamic percpu allocator.
1006 * number and page number. It should at least return enough pages to 1529 *
1007 * cover the static area. The returned pages for static area should 1530 * @ai->static_size is the size of static percpu area.
1008 * have been initialized with valid data. If @unit_size is specified, 1531 *
1009 * it can also return pages after the static area. NULL return 1532 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1010 * indicates end of pages for the cpu. Note that @get_page_fn() must
1011 * return the same number of pages for all cpus.
1012 *
1013 * @reserved_size, if non-zero, specifies the amount of bytes to
1014 * reserve after the static area in the first chunk. This reserves 1533 * reserve after the static area in the first chunk. This reserves
1015 * the first chunk such that it's available only through reserved 1534 * the first chunk such that it's available only through reserved
1016 * percpu allocation. This is primarily used to serve module percpu 1535 * percpu allocation. This is primarily used to serve module percpu
@@ -1018,22 +1537,29 @@ EXPORT_SYMBOL_GPL(free_percpu);
1018 * limited offset range for symbol relocations to guarantee module 1537 * limited offset range for symbol relocations to guarantee module
1019 * percpu symbols fall inside the relocatable range. 1538 * percpu symbols fall inside the relocatable range.
1020 * 1539 *
1021 * @dyn_size, if non-negative, determines the number of bytes 1540 * @ai->dyn_size determines the number of bytes available for dynamic
1022 * available for dynamic allocation in the first chunk. Specifying 1541 * allocation in the first chunk. The area between @ai->static_size +
1023 * non-negative value makes percpu leave alone the area beyond 1542 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1024 * @static_size + @reserved_size + @dyn_size.
1025 * 1543 *
1026 * @unit_size, if non-negative, specifies unit size and must be 1544 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1027 * aligned to PAGE_SIZE and equal to or larger than @static_size + 1545 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1028 * @reserved_size + if non-negative, @dyn_size. 1546 * @ai->dyn_size.
1029 * 1547 *
1030 * Non-null @base_addr means that the caller already allocated virtual 1548 * @ai->atom_size is the allocation atom size and used as alignment
1031 * region for the first chunk and mapped it. percpu must not mess 1549 * for vm areas.
1032 * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL
1033 * @populate_pte_fn doesn't make any sense.
1034 * 1550 *
1035 * @populate_pte_fn is used to populate the pagetable. NULL means the 1551 * @ai->alloc_size is the allocation size and always multiple of
1036 * caller already populated the pagetable. 1552 * @ai->atom_size. This is larger than @ai->atom_size if
1553 * @ai->unit_size is larger than @ai->atom_size.
1554 *
1555 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1556 * percpu areas. Units which should be colocated are put into the
1557 * same group. Dynamic VM areas will be allocated according to these
1558 * groupings. If @ai->nr_groups is zero, a single group containing
1559 * all units is assumed.
1560 *
1561 * The caller should have mapped the first chunk at @base_addr and
1562 * copied static data to each unit.
1037 * 1563 *
1038 * If the first chunk ends up with both reserved and dynamic areas, it 1564 * If the first chunk ends up with both reserved and dynamic areas, it
1039 * is served by two chunks - one to serve the core static and reserved 1565 * is served by two chunks - one to serve the core static and reserved
@@ -1043,49 +1569,83 @@ EXPORT_SYMBOL_GPL(free_percpu);
1043 * and available for dynamic allocation like any other chunks. 1569 * and available for dynamic allocation like any other chunks.
1044 * 1570 *
1045 * RETURNS: 1571 * RETURNS:
1046 * The determined pcpu_unit_size which can be used to initialize 1572 * 0 on success, -errno on failure.
1047 * percpu access.
1048 */ 1573 */
1049size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn, 1574int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1050 size_t static_size, size_t reserved_size, 1575 void *base_addr)
1051 ssize_t dyn_size, ssize_t unit_size,
1052 void *base_addr,
1053 pcpu_populate_pte_fn_t populate_pte_fn)
1054{ 1576{
1055 static struct vm_struct first_vm;
1056 static int smap[2], dmap[2]; 1577 static int smap[2], dmap[2];
1057 size_t size_sum = static_size + reserved_size + 1578 size_t dyn_size = ai->dyn_size;
1058 (dyn_size >= 0 ? dyn_size : 0); 1579 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1059 struct pcpu_chunk *schunk, *dchunk = NULL; 1580 struct pcpu_chunk *schunk, *dchunk = NULL;
1581 unsigned long *group_offsets;
1582 size_t *group_sizes;
1583 unsigned long *unit_off;
1060 unsigned int cpu; 1584 unsigned int cpu;
1061 int nr_pages; 1585 int *unit_map;
1062 int err, i; 1586 int group, unit, i;
1063 1587
1064 /* santiy checks */ 1588 /* sanity checks */
1065 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || 1589 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1066 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); 1590 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
1067 BUG_ON(!static_size); 1591 BUG_ON(ai->nr_groups <= 0);
1068 if (unit_size >= 0) { 1592 BUG_ON(!ai->static_size);
1069 BUG_ON(unit_size < size_sum); 1593 BUG_ON(!base_addr);
1070 BUG_ON(unit_size & ~PAGE_MASK); 1594 BUG_ON(ai->unit_size < size_sum);
1071 BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE); 1595 BUG_ON(ai->unit_size & ~PAGE_MASK);
1072 } else 1596 BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1073 BUG_ON(base_addr); 1597
1074 BUG_ON(base_addr && populate_pte_fn); 1598 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1075 1599
1076 if (unit_size >= 0) 1600 /* process group information and build config tables accordingly */
1077 pcpu_unit_pages = unit_size >> PAGE_SHIFT; 1601 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1078 else 1602 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
1079 pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT, 1603 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
1080 PFN_UP(size_sum)); 1604 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
1605
1606 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1607 unit_map[cpu] = NR_CPUS;
1608 pcpu_first_unit_cpu = NR_CPUS;
1609
1610 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1611 const struct pcpu_group_info *gi = &ai->groups[group];
1612
1613 group_offsets[group] = gi->base_offset;
1614 group_sizes[group] = gi->nr_units * ai->unit_size;
1615
1616 for (i = 0; i < gi->nr_units; i++) {
1617 cpu = gi->cpu_map[i];
1618 if (cpu == NR_CPUS)
1619 continue;
1081 1620
1082 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; 1621 BUG_ON(cpu > nr_cpu_ids || !cpu_possible(cpu));
1083 pcpu_chunk_size = nr_cpu_ids * pcpu_unit_size; 1622 BUG_ON(unit_map[cpu] != NR_CPUS);
1084 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
1085 + nr_cpu_ids * pcpu_unit_pages * sizeof(struct page *);
1086 1623
1087 if (dyn_size < 0) 1624 unit_map[cpu] = unit + i;
1088 dyn_size = pcpu_unit_size - static_size - reserved_size; 1625 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1626
1627 if (pcpu_first_unit_cpu == NR_CPUS)
1628 pcpu_first_unit_cpu = cpu;
1629 }
1630 }
1631 pcpu_last_unit_cpu = cpu;
1632 pcpu_nr_units = unit;
1633
1634 for_each_possible_cpu(cpu)
1635 BUG_ON(unit_map[cpu] == NR_CPUS);
1636
1637 pcpu_nr_groups = ai->nr_groups;
1638 pcpu_group_offsets = group_offsets;
1639 pcpu_group_sizes = group_sizes;
1640 pcpu_unit_map = unit_map;
1641 pcpu_unit_offsets = unit_off;
1642
1643 /* determine basic parameters */
1644 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1645 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1646 pcpu_atom_size = ai->atom_size;
1647 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1648 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1089 1649
1090 /* 1650 /*
1091 * Allocate chunk slots. The additional last slot is for 1651 * Allocate chunk slots. The additional last slot is for
@@ -1105,189 +1665,351 @@ size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn,
1105 */ 1665 */
1106 schunk = alloc_bootmem(pcpu_chunk_struct_size); 1666 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1107 INIT_LIST_HEAD(&schunk->list); 1667 INIT_LIST_HEAD(&schunk->list);
1108 schunk->vm = &first_vm; 1668 schunk->base_addr = base_addr;
1109 schunk->map = smap; 1669 schunk->map = smap;
1110 schunk->map_alloc = ARRAY_SIZE(smap); 1670 schunk->map_alloc = ARRAY_SIZE(smap);
1111 schunk->page = schunk->page_ar; 1671 schunk->immutable = true;
1672 bitmap_fill(schunk->populated, pcpu_unit_pages);
1112 1673
1113 if (reserved_size) { 1674 if (ai->reserved_size) {
1114 schunk->free_size = reserved_size; 1675 schunk->free_size = ai->reserved_size;
1115 pcpu_reserved_chunk = schunk; 1676 pcpu_reserved_chunk = schunk;
1116 pcpu_reserved_chunk_limit = static_size + reserved_size; 1677 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1117 } else { 1678 } else {
1118 schunk->free_size = dyn_size; 1679 schunk->free_size = dyn_size;
1119 dyn_size = 0; /* dynamic area covered */ 1680 dyn_size = 0; /* dynamic area covered */
1120 } 1681 }
1121 schunk->contig_hint = schunk->free_size; 1682 schunk->contig_hint = schunk->free_size;
1122 1683
1123 schunk->map[schunk->map_used++] = -static_size; 1684 schunk->map[schunk->map_used++] = -ai->static_size;
1124 if (schunk->free_size) 1685 if (schunk->free_size)
1125 schunk->map[schunk->map_used++] = schunk->free_size; 1686 schunk->map[schunk->map_used++] = schunk->free_size;
1126 1687
1127 /* init dynamic chunk if necessary */ 1688 /* init dynamic chunk if necessary */
1128 if (dyn_size) { 1689 if (dyn_size) {
1129 dchunk = alloc_bootmem(sizeof(struct pcpu_chunk)); 1690 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
1130 INIT_LIST_HEAD(&dchunk->list); 1691 INIT_LIST_HEAD(&dchunk->list);
1131 dchunk->vm = &first_vm; 1692 dchunk->base_addr = base_addr;
1132 dchunk->map = dmap; 1693 dchunk->map = dmap;
1133 dchunk->map_alloc = ARRAY_SIZE(dmap); 1694 dchunk->map_alloc = ARRAY_SIZE(dmap);
1134 dchunk->page = schunk->page_ar; /* share page map with schunk */ 1695 dchunk->immutable = true;
1696 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1135 1697
1136 dchunk->contig_hint = dchunk->free_size = dyn_size; 1698 dchunk->contig_hint = dchunk->free_size = dyn_size;
1137 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; 1699 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1138 dchunk->map[dchunk->map_used++] = dchunk->free_size; 1700 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1139 } 1701 }
1140 1702
1141 /* allocate vm address */
1142 first_vm.flags = VM_ALLOC;
1143 first_vm.size = pcpu_chunk_size;
1144
1145 if (!base_addr)
1146 vm_area_register_early(&first_vm, PAGE_SIZE);
1147 else {
1148 /*
1149 * Pages already mapped. No need to remap into
1150 * vmalloc area. In this case the first chunks can't
1151 * be mapped or unmapped by percpu and are marked
1152 * immutable.
1153 */
1154 first_vm.addr = base_addr;
1155 schunk->immutable = true;
1156 if (dchunk)
1157 dchunk->immutable = true;
1158 }
1159
1160 /* assign pages */
1161 nr_pages = -1;
1162 for_each_possible_cpu(cpu) {
1163 for (i = 0; i < pcpu_unit_pages; i++) {
1164 struct page *page = get_page_fn(cpu, i);
1165
1166 if (!page)
1167 break;
1168 *pcpu_chunk_pagep(schunk, cpu, i) = page;
1169 }
1170
1171 BUG_ON(i < PFN_UP(static_size));
1172
1173 if (nr_pages < 0)
1174 nr_pages = i;
1175 else
1176 BUG_ON(nr_pages != i);
1177 }
1178
1179 /* map them */
1180 if (populate_pte_fn) {
1181 for_each_possible_cpu(cpu)
1182 for (i = 0; i < nr_pages; i++)
1183 populate_pte_fn(pcpu_chunk_addr(schunk,
1184 cpu, i));
1185
1186 err = pcpu_map(schunk, 0, nr_pages);
1187 if (err)
1188 panic("failed to setup static percpu area, err=%d\n",
1189 err);
1190 }
1191
1192 /* link the first chunk in */ 1703 /* link the first chunk in */
1193 pcpu_first_chunk = dchunk ?: schunk; 1704 pcpu_first_chunk = dchunk ?: schunk;
1194 pcpu_chunk_relocate(pcpu_first_chunk, -1); 1705 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1195 1706
1196 /* we're done */ 1707 /* we're done */
1197 pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0); 1708 pcpu_base_addr = base_addr;
1198 return pcpu_unit_size; 1709 return 0;
1199} 1710}
1200 1711
1201/* 1712const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1202 * Embedding first chunk setup helper. 1713 [PCPU_FC_AUTO] = "auto",
1203 */ 1714 [PCPU_FC_EMBED] = "embed",
1204static void *pcpue_ptr __initdata; 1715 [PCPU_FC_PAGE] = "page",
1205static size_t pcpue_size __initdata; 1716};
1206static size_t pcpue_unit_size __initdata;
1207 1717
1208static struct page * __init pcpue_get_page(unsigned int cpu, int pageno) 1718enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1209{
1210 size_t off = (size_t)pageno << PAGE_SHIFT;
1211 1719
1212 if (off >= pcpue_size) 1720static int __init percpu_alloc_setup(char *str)
1213 return NULL; 1721{
1722 if (0)
1723 /* nada */;
1724#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1725 else if (!strcmp(str, "embed"))
1726 pcpu_chosen_fc = PCPU_FC_EMBED;
1727#endif
1728#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1729 else if (!strcmp(str, "page"))
1730 pcpu_chosen_fc = PCPU_FC_PAGE;
1731#endif
1732 else
1733 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1214 1734
1215 return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size + off); 1735 return 0;
1216} 1736}
1737early_param("percpu_alloc", percpu_alloc_setup);
1217 1738
1739#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1740 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1218/** 1741/**
1219 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem 1742 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1220 * @static_size: the size of static percpu area in bytes
1221 * @reserved_size: the size of reserved percpu area in bytes 1743 * @reserved_size: the size of reserved percpu area in bytes
1222 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto 1744 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1223 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto 1745 * @atom_size: allocation atom size
1746 * @cpu_distance_fn: callback to determine distance between cpus, optional
1747 * @alloc_fn: function to allocate percpu page
1748 * @free_fn: funtion to free percpu page
1224 * 1749 *
1225 * This is a helper to ease setting up embedded first percpu chunk and 1750 * This is a helper to ease setting up embedded first percpu chunk and
1226 * can be called where pcpu_setup_first_chunk() is expected. 1751 * can be called where pcpu_setup_first_chunk() is expected.
1227 * 1752 *
1228 * If this function is used to setup the first chunk, it is allocated 1753 * If this function is used to setup the first chunk, it is allocated
1229 * as a contiguous area using bootmem allocator and used as-is without 1754 * by calling @alloc_fn and used as-is without being mapped into
1230 * being mapped into vmalloc area. This enables the first chunk to 1755 * vmalloc area. Allocations are always whole multiples of @atom_size
1231 * piggy back on the linear physical mapping which often uses larger 1756 * aligned to @atom_size.
1232 * page size. 1757 *
1758 * This enables the first chunk to piggy back on the linear physical
1759 * mapping which often uses larger page size. Please note that this
1760 * can result in very sparse cpu->unit mapping on NUMA machines thus
1761 * requiring large vmalloc address space. Don't use this allocator if
1762 * vmalloc space is not orders of magnitude larger than distances
1763 * between node memory addresses (ie. 32bit NUMA machines).
1233 * 1764 *
1234 * When @dyn_size is positive, dynamic area might be larger than 1765 * When @dyn_size is positive, dynamic area might be larger than
1235 * specified to fill page alignment. Also, when @dyn_size is auto, 1766 * specified to fill page alignment. When @dyn_size is auto,
1236 * @dyn_size does not fill the whole first chunk but only what's 1767 * @dyn_size is just big enough to fill page alignment after static
1237 * necessary for page alignment after static and reserved areas. 1768 * and reserved areas.
1238 * 1769 *
1239 * If the needed size is smaller than the minimum or specified unit 1770 * If the needed size is smaller than the minimum or specified unit
1240 * size, the leftover is returned to the bootmem allocator. 1771 * size, the leftover is returned using @free_fn.
1241 * 1772 *
1242 * RETURNS: 1773 * RETURNS:
1243 * The determined pcpu_unit_size which can be used to initialize 1774 * 0 on success, -errno on failure.
1244 * percpu access on success, -errno on failure.
1245 */ 1775 */
1246ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size, 1776int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1247 ssize_t dyn_size, ssize_t unit_size) 1777 size_t atom_size,
1778 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1779 pcpu_fc_alloc_fn_t alloc_fn,
1780 pcpu_fc_free_fn_t free_fn)
1248{ 1781{
1249 size_t chunk_size; 1782 void *base = (void *)ULONG_MAX;
1250 unsigned int cpu; 1783 void **areas = NULL;
1784 struct pcpu_alloc_info *ai;
1785 size_t size_sum, areas_size;
1786 int group, i, rc;
1787
1788 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1789 cpu_distance_fn);
1790 if (IS_ERR(ai))
1791 return PTR_ERR(ai);
1792
1793 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1794 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1795
1796 areas = alloc_bootmem_nopanic(areas_size);
1797 if (!areas) {
1798 rc = -ENOMEM;
1799 goto out_free;
1800 }
1251 1801
1252 /* determine parameters and allocate */ 1802 /* allocate, copy and determine base address */
1253 pcpue_size = PFN_ALIGN(static_size + reserved_size + 1803 for (group = 0; group < ai->nr_groups; group++) {
1254 (dyn_size >= 0 ? dyn_size : 0)); 1804 struct pcpu_group_info *gi = &ai->groups[group];
1255 if (dyn_size != 0) 1805 unsigned int cpu = NR_CPUS;
1256 dyn_size = pcpue_size - static_size - reserved_size; 1806 void *ptr;
1257 1807
1258 if (unit_size >= 0) { 1808 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1259 BUG_ON(unit_size < pcpue_size); 1809 cpu = gi->cpu_map[i];
1260 pcpue_unit_size = unit_size; 1810 BUG_ON(cpu == NR_CPUS);
1261 } else 1811
1262 pcpue_unit_size = max_t(size_t, pcpue_size, PCPU_MIN_UNIT_SIZE); 1812 /* allocate space for the whole group */
1263 1813 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1264 chunk_size = pcpue_unit_size * nr_cpu_ids; 1814 if (!ptr) {
1265 1815 rc = -ENOMEM;
1266 pcpue_ptr = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE, 1816 goto out_free_areas;
1267 __pa(MAX_DMA_ADDRESS)); 1817 }
1268 if (!pcpue_ptr) { 1818 areas[group] = ptr;
1269 pr_warning("PERCPU: failed to allocate %zu bytes for " 1819
1270 "embedding\n", chunk_size); 1820 base = min(ptr, base);
1271 return -ENOMEM; 1821
1822 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1823 if (gi->cpu_map[i] == NR_CPUS) {
1824 /* unused unit, free whole */
1825 free_fn(ptr, ai->unit_size);
1826 continue;
1827 }
1828 /* copy and return the unused part */
1829 memcpy(ptr, __per_cpu_load, ai->static_size);
1830 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1831 }
1272 } 1832 }
1273 1833
1274 /* return the leftover and copy */ 1834 /* base address is now known, determine group base offsets */
1275 for (cpu = 0; cpu < nr_cpu_ids; cpu++) { 1835 for (group = 0; group < ai->nr_groups; group++)
1276 void *ptr = pcpue_ptr + cpu * pcpue_unit_size; 1836 ai->groups[group].base_offset = areas[group] - base;
1837
1838 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1839 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1840 ai->dyn_size, ai->unit_size);
1841
1842 rc = pcpu_setup_first_chunk(ai, base);
1843 goto out_free;
1844
1845out_free_areas:
1846 for (group = 0; group < ai->nr_groups; group++)
1847 free_fn(areas[group],
1848 ai->groups[group].nr_units * ai->unit_size);
1849out_free:
1850 pcpu_free_alloc_info(ai);
1851 if (areas)
1852 free_bootmem(__pa(areas), areas_size);
1853 return rc;
1854}
1855#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1856 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
1857
1858#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1859/**
1860 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1861 * @reserved_size: the size of reserved percpu area in bytes
1862 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1863 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1864 * @populate_pte_fn: function to populate pte
1865 *
1866 * This is a helper to ease setting up page-remapped first percpu
1867 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1868 *
1869 * This is the basic allocator. Static percpu area is allocated
1870 * page-by-page into vmalloc area.
1871 *
1872 * RETURNS:
1873 * 0 on success, -errno on failure.
1874 */
1875int __init pcpu_page_first_chunk(size_t reserved_size,
1876 pcpu_fc_alloc_fn_t alloc_fn,
1877 pcpu_fc_free_fn_t free_fn,
1878 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1879{
1880 static struct vm_struct vm;
1881 struct pcpu_alloc_info *ai;
1882 char psize_str[16];
1883 int unit_pages;
1884 size_t pages_size;
1885 struct page **pages;
1886 int unit, i, j, rc;
1887
1888 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1889
1890 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1891 if (IS_ERR(ai))
1892 return PTR_ERR(ai);
1893 BUG_ON(ai->nr_groups != 1);
1894 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1895
1896 unit_pages = ai->unit_size >> PAGE_SHIFT;
1897
1898 /* unaligned allocations can't be freed, round up to page size */
1899 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1900 sizeof(pages[0]));
1901 pages = alloc_bootmem(pages_size);
1902
1903 /* allocate pages */
1904 j = 0;
1905 for (unit = 0; unit < num_possible_cpus(); unit++)
1906 for (i = 0; i < unit_pages; i++) {
1907 unsigned int cpu = ai->groups[0].cpu_map[unit];
1908 void *ptr;
1909
1910 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1911 if (!ptr) {
1912 pr_warning("PERCPU: failed to allocate %s page "
1913 "for cpu%u\n", psize_str, cpu);
1914 goto enomem;
1915 }
1916 pages[j++] = virt_to_page(ptr);
1917 }
1918
1919 /* allocate vm area, map the pages and copy static data */
1920 vm.flags = VM_ALLOC;
1921 vm.size = num_possible_cpus() * ai->unit_size;
1922 vm_area_register_early(&vm, PAGE_SIZE);
1923
1924 for (unit = 0; unit < num_possible_cpus(); unit++) {
1925 unsigned long unit_addr =
1926 (unsigned long)vm.addr + unit * ai->unit_size;
1927
1928 for (i = 0; i < unit_pages; i++)
1929 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1930
1931 /* pte already populated, the following shouldn't fail */
1932 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1933 unit_pages);
1934 if (rc < 0)
1935 panic("failed to map percpu area, err=%d\n", rc);
1277 1936
1278 if (cpu_possible(cpu)) { 1937 /*
1279 free_bootmem(__pa(ptr + pcpue_size), 1938 * FIXME: Archs with virtual cache should flush local
1280 pcpue_unit_size - pcpue_size); 1939 * cache for the linear mapping here - something
1281 memcpy(ptr, __per_cpu_load, static_size); 1940 * equivalent to flush_cache_vmap() on the local cpu.
1282 } else 1941 * flush_cache_vmap() can't be used as most supporting
1283 free_bootmem(__pa(ptr), pcpue_unit_size); 1942 * data structures are not set up yet.
1943 */
1944
1945 /* copy static data */
1946 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1284 } 1947 }
1285 1948
1286 /* we're ready, commit */ 1949 /* we're ready, commit */
1287 pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n", 1950 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1288 pcpue_size >> PAGE_SHIFT, pcpue_ptr, static_size); 1951 unit_pages, psize_str, vm.addr, ai->static_size,
1952 ai->reserved_size, ai->dyn_size);
1953
1954 rc = pcpu_setup_first_chunk(ai, vm.addr);
1955 goto out_free_ar;
1956
1957enomem:
1958 while (--j >= 0)
1959 free_fn(page_address(pages[j]), PAGE_SIZE);
1960 rc = -ENOMEM;
1961out_free_ar:
1962 free_bootmem(__pa(pages), pages_size);
1963 pcpu_free_alloc_info(ai);
1964 return rc;
1965}
1966#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
1967
1968/*
1969 * Generic percpu area setup.
1970 *
1971 * The embedding helper is used because its behavior closely resembles
1972 * the original non-dynamic generic percpu area setup. This is
1973 * important because many archs have addressing restrictions and might
1974 * fail if the percpu area is located far away from the previous
1975 * location. As an added bonus, in non-NUMA cases, embedding is
1976 * generally a good idea TLB-wise because percpu area can piggy back
1977 * on the physical linear memory mapping which uses large page
1978 * mappings on applicable archs.
1979 */
1980#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1981unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1982EXPORT_SYMBOL(__per_cpu_offset);
1983
1984static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1985 size_t align)
1986{
1987 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
1988}
1289 1989
1290 return pcpu_setup_first_chunk(pcpue_get_page, static_size, 1990static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1291 reserved_size, dyn_size, 1991{
1292 pcpue_unit_size, pcpue_ptr, NULL); 1992 free_bootmem(__pa(ptr), size);
1993}
1994
1995void __init setup_per_cpu_areas(void)
1996{
1997 unsigned long delta;
1998 unsigned int cpu;
1999 int rc;
2000
2001 /*
2002 * Always reserve area for module percpu variables. That's
2003 * what the legacy allocator did.
2004 */
2005 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2006 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2007 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2008 if (rc < 0)
2009 panic("Failed to initialized percpu areas.");
2010
2011 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2012 for_each_possible_cpu(cpu)
2013 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1293} 2014}
2015#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
diff --git a/mm/quicklist.c b/mm/quicklist.c
index e66d07d1b4ff..6eedf7e473d1 100644
--- a/mm/quicklist.c
+++ b/mm/quicklist.c
@@ -19,7 +19,7 @@
19#include <linux/module.h> 19#include <linux/module.h>
20#include <linux/quicklist.h> 20#include <linux/quicklist.h>
21 21
22DEFINE_PER_CPU(struct quicklist, quicklist)[CONFIG_NR_QUICK]; 22DEFINE_PER_CPU(struct quicklist [CONFIG_NR_QUICK], quicklist);
23 23
24#define FRACTION_OF_NODE_MEM 16 24#define FRACTION_OF_NODE_MEM 16
25 25
diff --git a/mm/slub.c b/mm/slub.c
index 417ed843b251..0a216aae227e 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1071,6 +1071,8 @@ static inline unsigned long kmem_cache_flags(unsigned long objsize,
1071} 1071}
1072#define slub_debug 0 1072#define slub_debug 0
1073 1073
1074#define disable_higher_order_debug 0
1075
1074static inline unsigned long slabs_node(struct kmem_cache *s, int node) 1076static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1075 { return 0; } 1077 { return 0; }
1076static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) 1078static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
@@ -2111,8 +2113,8 @@ init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2111 */ 2113 */
2112#define NR_KMEM_CACHE_CPU 100 2114#define NR_KMEM_CACHE_CPU 100
2113 2115
2114static DEFINE_PER_CPU(struct kmem_cache_cpu, 2116static DEFINE_PER_CPU(struct kmem_cache_cpu [NR_KMEM_CACHE_CPU],
2115 kmem_cache_cpu)[NR_KMEM_CACHE_CPU]; 2117 kmem_cache_cpu);
2116 2118
2117static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free); 2119static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2118static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS); 2120static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index f8189a4b3e13..204b8243d8ab 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -265,6 +265,7 @@ struct vmap_area {
265static DEFINE_SPINLOCK(vmap_area_lock); 265static DEFINE_SPINLOCK(vmap_area_lock);
266static struct rb_root vmap_area_root = RB_ROOT; 266static struct rb_root vmap_area_root = RB_ROOT;
267static LIST_HEAD(vmap_area_list); 267static LIST_HEAD(vmap_area_list);
268static unsigned long vmap_area_pcpu_hole;
268 269
269static struct vmap_area *__find_vmap_area(unsigned long addr) 270static struct vmap_area *__find_vmap_area(unsigned long addr)
270{ 271{
@@ -431,6 +432,15 @@ static void __free_vmap_area(struct vmap_area *va)
431 RB_CLEAR_NODE(&va->rb_node); 432 RB_CLEAR_NODE(&va->rb_node);
432 list_del_rcu(&va->list); 433 list_del_rcu(&va->list);
433 434
435 /*
436 * Track the highest possible candidate for pcpu area
437 * allocation. Areas outside of vmalloc area can be returned
438 * here too, consider only end addresses which fall inside
439 * vmalloc area proper.
440 */
441 if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
442 vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
443
434 call_rcu(&va->rcu_head, rcu_free_va); 444 call_rcu(&va->rcu_head, rcu_free_va);
435} 445}
436 446
@@ -1038,6 +1048,9 @@ void __init vmalloc_init(void)
1038 va->va_end = va->va_start + tmp->size; 1048 va->va_end = va->va_start + tmp->size;
1039 __insert_vmap_area(va); 1049 __insert_vmap_area(va);
1040 } 1050 }
1051
1052 vmap_area_pcpu_hole = VMALLOC_END;
1053
1041 vmap_initialized = true; 1054 vmap_initialized = true;
1042} 1055}
1043 1056
@@ -1122,13 +1135,34 @@ EXPORT_SYMBOL_GPL(map_vm_area);
1122DEFINE_RWLOCK(vmlist_lock); 1135DEFINE_RWLOCK(vmlist_lock);
1123struct vm_struct *vmlist; 1136struct vm_struct *vmlist;
1124 1137
1138static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1139 unsigned long flags, void *caller)
1140{
1141 struct vm_struct *tmp, **p;
1142
1143 vm->flags = flags;
1144 vm->addr = (void *)va->va_start;
1145 vm->size = va->va_end - va->va_start;
1146 vm->caller = caller;
1147 va->private = vm;
1148 va->flags |= VM_VM_AREA;
1149
1150 write_lock(&vmlist_lock);
1151 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1152 if (tmp->addr >= vm->addr)
1153 break;
1154 }
1155 vm->next = *p;
1156 *p = vm;
1157 write_unlock(&vmlist_lock);
1158}
1159
1125static struct vm_struct *__get_vm_area_node(unsigned long size, 1160static struct vm_struct *__get_vm_area_node(unsigned long size,
1126 unsigned long flags, unsigned long start, unsigned long end, 1161 unsigned long flags, unsigned long start, unsigned long end,
1127 int node, gfp_t gfp_mask, void *caller) 1162 int node, gfp_t gfp_mask, void *caller)
1128{ 1163{
1129 static struct vmap_area *va; 1164 static struct vmap_area *va;
1130 struct vm_struct *area; 1165 struct vm_struct *area;
1131 struct vm_struct *tmp, **p;
1132 unsigned long align = 1; 1166 unsigned long align = 1;
1133 1167
1134 BUG_ON(in_interrupt()); 1168 BUG_ON(in_interrupt());
@@ -1147,7 +1181,7 @@ static struct vm_struct *__get_vm_area_node(unsigned long size,
1147 if (unlikely(!size)) 1181 if (unlikely(!size))
1148 return NULL; 1182 return NULL;
1149 1183
1150 area = kmalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node); 1184 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1151 if (unlikely(!area)) 1185 if (unlikely(!area))
1152 return NULL; 1186 return NULL;
1153 1187
@@ -1162,25 +1196,7 @@ static struct vm_struct *__get_vm_area_node(unsigned long size,
1162 return NULL; 1196 return NULL;
1163 } 1197 }
1164 1198
1165 area->flags = flags; 1199 insert_vmalloc_vm(area, va, flags, caller);
1166 area->addr = (void *)va->va_start;
1167 area->size = size;
1168 area->pages = NULL;
1169 area->nr_pages = 0;
1170 area->phys_addr = 0;
1171 area->caller = caller;
1172 va->private = area;
1173 va->flags |= VM_VM_AREA;
1174
1175 write_lock(&vmlist_lock);
1176 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1177 if (tmp->addr >= area->addr)
1178 break;
1179 }
1180 area->next = *p;
1181 *p = area;
1182 write_unlock(&vmlist_lock);
1183
1184 return area; 1200 return area;
1185} 1201}
1186 1202
@@ -1818,6 +1834,286 @@ void free_vm_area(struct vm_struct *area)
1818} 1834}
1819EXPORT_SYMBOL_GPL(free_vm_area); 1835EXPORT_SYMBOL_GPL(free_vm_area);
1820 1836
1837static struct vmap_area *node_to_va(struct rb_node *n)
1838{
1839 return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
1840}
1841
1842/**
1843 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
1844 * @end: target address
1845 * @pnext: out arg for the next vmap_area
1846 * @pprev: out arg for the previous vmap_area
1847 *
1848 * Returns: %true if either or both of next and prev are found,
1849 * %false if no vmap_area exists
1850 *
1851 * Find vmap_areas end addresses of which enclose @end. ie. if not
1852 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
1853 */
1854static bool pvm_find_next_prev(unsigned long end,
1855 struct vmap_area **pnext,
1856 struct vmap_area **pprev)
1857{
1858 struct rb_node *n = vmap_area_root.rb_node;
1859 struct vmap_area *va = NULL;
1860
1861 while (n) {
1862 va = rb_entry(n, struct vmap_area, rb_node);
1863 if (end < va->va_end)
1864 n = n->rb_left;
1865 else if (end > va->va_end)
1866 n = n->rb_right;
1867 else
1868 break;
1869 }
1870
1871 if (!va)
1872 return false;
1873
1874 if (va->va_end > end) {
1875 *pnext = va;
1876 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
1877 } else {
1878 *pprev = va;
1879 *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
1880 }
1881 return true;
1882}
1883
1884/**
1885 * pvm_determine_end - find the highest aligned address between two vmap_areas
1886 * @pnext: in/out arg for the next vmap_area
1887 * @pprev: in/out arg for the previous vmap_area
1888 * @align: alignment
1889 *
1890 * Returns: determined end address
1891 *
1892 * Find the highest aligned address between *@pnext and *@pprev below
1893 * VMALLOC_END. *@pnext and *@pprev are adjusted so that the aligned
1894 * down address is between the end addresses of the two vmap_areas.
1895 *
1896 * Please note that the address returned by this function may fall
1897 * inside *@pnext vmap_area. The caller is responsible for checking
1898 * that.
1899 */
1900static unsigned long pvm_determine_end(struct vmap_area **pnext,
1901 struct vmap_area **pprev,
1902 unsigned long align)
1903{
1904 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
1905 unsigned long addr;
1906
1907 if (*pnext)
1908 addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
1909 else
1910 addr = vmalloc_end;
1911
1912 while (*pprev && (*pprev)->va_end > addr) {
1913 *pnext = *pprev;
1914 *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
1915 }
1916
1917 return addr;
1918}
1919
1920/**
1921 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
1922 * @offsets: array containing offset of each area
1923 * @sizes: array containing size of each area
1924 * @nr_vms: the number of areas to allocate
1925 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
1926 * @gfp_mask: allocation mask
1927 *
1928 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
1929 * vm_structs on success, %NULL on failure
1930 *
1931 * Percpu allocator wants to use congruent vm areas so that it can
1932 * maintain the offsets among percpu areas. This function allocates
1933 * congruent vmalloc areas for it. These areas tend to be scattered
1934 * pretty far, distance between two areas easily going up to
1935 * gigabytes. To avoid interacting with regular vmallocs, these areas
1936 * are allocated from top.
1937 *
1938 * Despite its complicated look, this allocator is rather simple. It
1939 * does everything top-down and scans areas from the end looking for
1940 * matching slot. While scanning, if any of the areas overlaps with
1941 * existing vmap_area, the base address is pulled down to fit the
1942 * area. Scanning is repeated till all the areas fit and then all
1943 * necessary data structres are inserted and the result is returned.
1944 */
1945struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
1946 const size_t *sizes, int nr_vms,
1947 size_t align, gfp_t gfp_mask)
1948{
1949 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
1950 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
1951 struct vmap_area **vas, *prev, *next;
1952 struct vm_struct **vms;
1953 int area, area2, last_area, term_area;
1954 unsigned long base, start, end, last_end;
1955 bool purged = false;
1956
1957 gfp_mask &= GFP_RECLAIM_MASK;
1958
1959 /* verify parameters and allocate data structures */
1960 BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
1961 for (last_area = 0, area = 0; area < nr_vms; area++) {
1962 start = offsets[area];
1963 end = start + sizes[area];
1964
1965 /* is everything aligned properly? */
1966 BUG_ON(!IS_ALIGNED(offsets[area], align));
1967 BUG_ON(!IS_ALIGNED(sizes[area], align));
1968
1969 /* detect the area with the highest address */
1970 if (start > offsets[last_area])
1971 last_area = area;
1972
1973 for (area2 = 0; area2 < nr_vms; area2++) {
1974 unsigned long start2 = offsets[area2];
1975 unsigned long end2 = start2 + sizes[area2];
1976
1977 if (area2 == area)
1978 continue;
1979
1980 BUG_ON(start2 >= start && start2 < end);
1981 BUG_ON(end2 <= end && end2 > start);
1982 }
1983 }
1984 last_end = offsets[last_area] + sizes[last_area];
1985
1986 if (vmalloc_end - vmalloc_start < last_end) {
1987 WARN_ON(true);
1988 return NULL;
1989 }
1990
1991 vms = kzalloc(sizeof(vms[0]) * nr_vms, gfp_mask);
1992 vas = kzalloc(sizeof(vas[0]) * nr_vms, gfp_mask);
1993 if (!vas || !vms)
1994 goto err_free;
1995
1996 for (area = 0; area < nr_vms; area++) {
1997 vas[area] = kzalloc(sizeof(struct vmap_area), gfp_mask);
1998 vms[area] = kzalloc(sizeof(struct vm_struct), gfp_mask);
1999 if (!vas[area] || !vms[area])
2000 goto err_free;
2001 }
2002retry:
2003 spin_lock(&vmap_area_lock);
2004
2005 /* start scanning - we scan from the top, begin with the last area */
2006 area = term_area = last_area;
2007 start = offsets[area];
2008 end = start + sizes[area];
2009
2010 if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2011 base = vmalloc_end - last_end;
2012 goto found;
2013 }
2014 base = pvm_determine_end(&next, &prev, align) - end;
2015
2016 while (true) {
2017 BUG_ON(next && next->va_end <= base + end);
2018 BUG_ON(prev && prev->va_end > base + end);
2019
2020 /*
2021 * base might have underflowed, add last_end before
2022 * comparing.
2023 */
2024 if (base + last_end < vmalloc_start + last_end) {
2025 spin_unlock(&vmap_area_lock);
2026 if (!purged) {
2027 purge_vmap_area_lazy();
2028 purged = true;
2029 goto retry;
2030 }
2031 goto err_free;
2032 }
2033
2034 /*
2035 * If next overlaps, move base downwards so that it's
2036 * right below next and then recheck.
2037 */
2038 if (next && next->va_start < base + end) {
2039 base = pvm_determine_end(&next, &prev, align) - end;
2040 term_area = area;
2041 continue;
2042 }
2043
2044 /*
2045 * If prev overlaps, shift down next and prev and move
2046 * base so that it's right below new next and then
2047 * recheck.
2048 */
2049 if (prev && prev->va_end > base + start) {
2050 next = prev;
2051 prev = node_to_va(rb_prev(&next->rb_node));
2052 base = pvm_determine_end(&next, &prev, align) - end;
2053 term_area = area;
2054 continue;
2055 }
2056
2057 /*
2058 * This area fits, move on to the previous one. If
2059 * the previous one is the terminal one, we're done.
2060 */
2061 area = (area + nr_vms - 1) % nr_vms;
2062 if (area == term_area)
2063 break;
2064 start = offsets[area];
2065 end = start + sizes[area];
2066 pvm_find_next_prev(base + end, &next, &prev);
2067 }
2068found:
2069 /* we've found a fitting base, insert all va's */
2070 for (area = 0; area < nr_vms; area++) {
2071 struct vmap_area *va = vas[area];
2072
2073 va->va_start = base + offsets[area];
2074 va->va_end = va->va_start + sizes[area];
2075 __insert_vmap_area(va);
2076 }
2077
2078 vmap_area_pcpu_hole = base + offsets[last_area];
2079
2080 spin_unlock(&vmap_area_lock);
2081
2082 /* insert all vm's */
2083 for (area = 0; area < nr_vms; area++)
2084 insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2085 pcpu_get_vm_areas);
2086
2087 kfree(vas);
2088 return vms;
2089
2090err_free:
2091 for (area = 0; area < nr_vms; area++) {
2092 if (vas)
2093 kfree(vas[area]);
2094 if (vms)
2095 kfree(vms[area]);
2096 }
2097 kfree(vas);
2098 kfree(vms);
2099 return NULL;
2100}
2101
2102/**
2103 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2104 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2105 * @nr_vms: the number of allocated areas
2106 *
2107 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2108 */
2109void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2110{
2111 int i;
2112
2113 for (i = 0; i < nr_vms; i++)
2114 free_vm_area(vms[i]);
2115 kfree(vms);
2116}
1821 2117
1822#ifdef CONFIG_PROC_FS 2118#ifdef CONFIG_PROC_FS
1823static void *s_start(struct seq_file *m, loff_t *pos) 2119static void *s_start(struct seq_file *m, loff_t *pos)
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-26 04:40:55 -0500