summaryrefslogtreecommitdiffstats
path: root/fs/fs-writeback.c
AgeCommit message (Collapse)AuthorFilesLines
2016-12-12fs/fs-writeback.c: remove redundant if checkTahsin Erdogan1-9/+7
b_more_io non-empty check is already preceded by an opposite check. Link: http://lkml.kernel.org/r/1478591249-30641-1-git-send-email-tahsin@google.com Signed-off-by: Tahsin Erdogan <tahsin@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-09mm, writeback: flush plugged IO in wakeup_flusher_threads()Konstantin Khlebnikov1-0/+6
I've found funny live-lock between raid10 barriers during resync and memory controller hard limits. Inside mpage_readpages() task holds on to its plug bio which blocks the barrier in raid10. Its memory cgroup have no free memory thus the task goes into reclaimer but all reclaimable pages are dirty and cannot be written because raid10 is rebuilding and stuck on the barrier. Common flush of such IO in schedule() never happens, because the caller doesn't go to sleep. Lock is 'live' because changing memory limit or killing tasks which holds that stuck bio unblock whole progress. That was what happened in 3.18.x but I see no difference in upstream logic. Theoretically this might happen even without memory cgroup. Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-08-04writeback: Write dirty times for WB_SYNC_ALL writebackJan Kara1-0/+1
Currently we take care to handle I_DIRTY_TIME in vfs_fsync() and queue_io() so that inodes which have only dirty timestamps are properly written on fsync(2) and sync(2). However there are other call sites - most notably going through write_inode_now() - which expect inode to be clean after WB_SYNC_ALL writeback. This is not currently true as we do not clear I_DIRTY_TIME in __writeback_single_inode() even for WB_SYNC_ALL writeback in all the cases. This then resulted in the following oops because bdev_write_inode() did not clean the inode and writeback code later stumbled over a dirty inode with detached wb. general protection fault: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN Modules linked in: CPU: 3 PID: 32 Comm: kworker/u10:1 Not tainted 4.6.0-rc3+ #349 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Workqueue: writeback wb_workfn (flush-11:0) task: ffff88006ccf1840 ti: ffff88006cda8000 task.ti: ffff88006cda8000 RIP: 0010:[<ffffffff818884d2>] [<ffffffff818884d2>] locked_inode_to_wb_and_lock_list+0xa2/0x750 RSP: 0018:ffff88006cdaf7d0 EFLAGS: 00010246 RAX: 0000000000000000 RBX: 0000000000000000 RCX: ffff88006ccf2050 RDX: 0000000000000000 RSI: 000000114c8a8484 RDI: 0000000000000286 RBP: ffff88006cdaf820 R08: ffff88006ccf1840 R09: 0000000000000000 R10: 000229915090805f R11: 0000000000000001 R12: ffff88006a72f5e0 R13: dffffc0000000000 R14: ffffed000d4e5eed R15: ffffffff8830cf40 FS: 0000000000000000(0000) GS:ffff88006d500000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000000003301bf8 CR3: 000000006368f000 CR4: 00000000000006e0 DR0: 0000000000001ec9 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000600 Stack: ffff88006a72f680 ffff88006a72f768 ffff8800671230d8 03ff88006cdaf948 ffff88006a72f668 ffff88006a72f5e0 ffff8800671230d8 ffff88006cdaf948 ffff880065b90cc8 ffff880067123100 ffff88006cdaf970 ffffffff8188e12e Call Trace: [< inline >] inode_to_wb_and_lock_list fs/fs-writeback.c:309 [<ffffffff8188e12e>] writeback_sb_inodes+0x4de/0x1250 fs/fs-writeback.c:1554 [<ffffffff8188efa4>] __writeback_inodes_wb+0x104/0x1e0 fs/fs-writeback.c:1600 [<ffffffff8188f9ae>] wb_writeback+0x7ce/0xc90 fs/fs-writeback.c:1709 [< inline >] wb_do_writeback fs/fs-writeback.c:1844 [<ffffffff81891079>] wb_workfn+0x2f9/0x1000 fs/fs-writeback.c:1884 [<ffffffff813bcd1e>] process_one_work+0x78e/0x15c0 kernel/workqueue.c:2094 [<ffffffff813bdc2b>] worker_thread+0xdb/0xfc0 kernel/workqueue.c:2228 [<ffffffff813cdeef>] kthread+0x23f/0x2d0 drivers/block/aoe/aoecmd.c:1303 [<ffffffff867bc5d2>] ret_from_fork+0x22/0x50 arch/x86/entry/entry_64.S:392 Code: 05 94 4a a8 06 85 c0 0f 85 03 03 00 00 e8 07 15 d0 ff 41 80 3e 00 0f 85 64 06 00 00 49 8b 9c 24 88 01 00 00 48 89 d8 48 c1 e8 03 <42> 80 3c 28 00 0f 85 17 06 00 00 48 8b 03 48 83 c0 50 48 39 c3 RIP [< inline >] wb_get include/linux/backing-dev-defs.h:212 RIP [<ffffffff818884d2>] locked_inode_to_wb_and_lock_list+0xa2/0x750 fs/fs-writeback.c:281 RSP <ffff88006cdaf7d0> ---[ end trace 986a4d314dcb2694 ]--- Fix the problem by making sure __writeback_single_inode() writes inode only with dirty times in WB_SYNC_ALL mode. Reported-by: Dmitry Vyukov <dvyukov@google.com> Tested-by: Laurent Dufour <ldufour@linux.vnet.ibm.com> Signed-off-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-07-28mm: move most file-based accounting to the nodeMel Gorman1-2/+2
There are now a number of accounting oddities such as mapped file pages being accounted for on the node while the total number of file pages are accounted on the zone. This can be coped with to some extent but it's confusing so this patch moves the relevant file-based accounted. Due to throttling logic in the page allocator for reliable OOM detection, it is still necessary to track dirty and writeback pages on a per-zone basis. [mgorman@techsingularity.net: fix NR_ZONE_WRITE_PENDING accounting] Link: http://lkml.kernel.org/r/1468404004-5085-5-git-send-email-mgorman@techsingularity.net Link: http://lkml.kernel.org/r/1467970510-21195-20-git-send-email-mgorman@techsingularity.net Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Acked-by: Vlastimil Babka <vbabka@suse.cz> Acked-by: Michal Hocko <mhocko@suse.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-26fs/fs-writeback.c: inode writeback list tracking tracepointsBrian Foster1-2/+7
The per-sb inode writeback list tracks inodes currently under writeback to facilitate efficient sync processing. In particular, it ensures that sync only needs to walk through a list of inodes that were cleaned by the sync. Add a couple tracepoints to help identify when inodes are added/removed to and from the writeback lists. Piggyback off of the writeback lazytime tracepoint template as it already tracks the relevant inode information. Link: http://lkml.kernel.org/r/1466594593-6757-3-git-send-email-bfoster@redhat.com Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Dave Chinner <dchinner@redhat.com> cc: Josef Bacik <jbacik@fb.com> Cc: Holger Hoffstätte <holger.hoffstaette@applied-asynchrony.com> Cc: Al Viro <viro@ZenIV.linux.org.uk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-26fs/fs-writeback.c: add a new writeback list for syncDave Chinner1-25/+81
wait_sb_inodes() currently does a walk of all inodes in the filesystem to find dirty one to wait on during sync. This is highly inefficient and wastes a lot of CPU when there are lots of clean cached inodes that we don't need to wait on. To avoid this "all inode" walk, we need to track inodes that are currently under writeback that we need to wait for. We do this by adding inodes to a writeback list on the sb when the mapping is first tagged as having pages under writeback. wait_sb_inodes() can then walk this list of "inodes under IO" and wait specifically just for the inodes that the current sync(2) needs to wait for. Define a couple helpers to add/remove an inode from the writeback list and call them when the overall mapping is tagged for or cleared from writeback. Update wait_sb_inodes() to walk only the inodes under writeback due to the sync. With this change, filesystem sync times are significantly reduced for fs' with largely populated inode caches and otherwise no other work to do. For example, on a 16xcpu 2GHz x86-64 server, 10TB XFS filesystem with a ~10m entry inode cache, sync times are reduced from ~7.3s to less than 0.1s when the filesystem is fully clean. Link: http://lkml.kernel.org/r/1466594593-6757-2-git-send-email-bfoster@redhat.com Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Josef Bacik <jbacik@fb.com> Signed-off-by: Brian Foster <bfoster@redhat.com> Reviewed-by: Jan Kara <jack@suse.cz> Tested-by: Holger Hoffstätte <holger.hoffstaette@applied-asynchrony.com> Cc: Al Viro <viro@ZenIV.linux.org.uk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-06-30writeback: inode cgroup wb switch should not call ihold()Tahsin Erdogan1-1/+1
Asynchronous wb switching of inodes takes an additional ref count on an inode to make sure inode remains valid until switchover is completed. However, anyone calling ihold() must already have a ref count on inode, but in this case inode->i_count may already be zero: ------------[ cut here ]------------ WARNING: CPU: 1 PID: 917 at fs/inode.c:397 ihold+0x2b/0x30 CPU: 1 PID: 917 Comm: kworker/u4:5 Not tainted 4.7.0-rc2+ #49 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 Workqueue: writeback wb_workfn (flush-8:16) 0000000000000000 ffff88007ca0fb58 ffffffff805990af 0000000000000000 0000000000000000 ffff88007ca0fb98 ffffffff80268702 0000018d000004e2 ffff88007cef40e8 ffff88007c9b89a8 ffff880079e3a740 0000000000000003 Call Trace: [<ffffffff805990af>] dump_stack+0x4d/0x6e [<ffffffff80268702>] __warn+0xc2/0xe0 [<ffffffff802687d8>] warn_slowpath_null+0x18/0x20 [<ffffffff8035b4ab>] ihold+0x2b/0x30 [<ffffffff80367ecc>] inode_switch_wbs+0x11c/0x180 [<ffffffff80369110>] wbc_detach_inode+0x170/0x1a0 [<ffffffff80369abc>] writeback_sb_inodes+0x21c/0x530 [<ffffffff80369f7e>] wb_writeback+0xee/0x1e0 [<ffffffff8036a147>] wb_workfn+0xd7/0x280 [<ffffffff80287531>] ? try_to_wake_up+0x1b1/0x2b0 [<ffffffff8027bb09>] process_one_work+0x129/0x300 [<ffffffff8027be06>] worker_thread+0x126/0x480 [<ffffffff8098cde7>] ? __schedule+0x1c7/0x561 [<ffffffff8027bce0>] ? process_one_work+0x300/0x300 [<ffffffff80280ff4>] kthread+0xc4/0xe0 [<ffffffff80335578>] ? kfree+0xc8/0x100 [<ffffffff809903cf>] ret_from_fork+0x1f/0x40 [<ffffffff80280f30>] ? __kthread_parkme+0x70/0x70 ---[ end trace aaefd2fd9f306bc4 ]--- Signed-off-by: Tahsin Erdogan <tahsin@google.com> Acked-by: Tejun Heo <tj@kernel.org> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-05-20mm,writeback: don't use memory reserves for wb_start_writebackTetsuo Handa1-1/+2
When writeback operation cannot make forward progress because memory allocation requests needed for doing I/O cannot be satisfied (e.g. under OOM-livelock situation), we can observe flood of order-0 page allocation failure messages caused by complete depletion of memory reserves. This is caused by unconditionally allocating "struct wb_writeback_work" objects using GFP_ATOMIC from PF_MEMALLOC context. __alloc_pages_nodemask() { __alloc_pages_slowpath() { __alloc_pages_direct_reclaim() { __perform_reclaim() { current->flags |= PF_MEMALLOC; try_to_free_pages() { do_try_to_free_pages() { wakeup_flusher_threads() { wb_start_writeback() { kzalloc(sizeof(*work), GFP_ATOMIC) { /* ALLOC_NO_WATERMARKS via PF_MEMALLOC */ } } } } } current->flags &= ~PF_MEMALLOC; } } } } Since I/O is stalling, allocating writeback requests forever shall deplete memory reserves. Fortunately, since wb_start_writeback() can fall back to wb_wakeup() when allocating "struct wb_writeback_work" failed, we don't need to allow wb_start_writeback() to use memory reserves. Mem-Info: active_anon:289393 inactive_anon:2093 isolated_anon:29 active_file:10838 inactive_file:113013 isolated_file:859 unevictable:0 dirty:108531 writeback:5308 unstable:0 slab_reclaimable:5526 slab_unreclaimable:7077 mapped:9970 shmem:2159 pagetables:2387 bounce:0 free:3042 free_pcp:0 free_cma:0 Node 0 DMA free:6968kB min:44kB low:52kB high:64kB active_anon:6056kB inactive_anon:176kB active_file:712kB inactive_file:744kB unevictable:0kB isolated(anon):0kB isolated(file):0kB present:15988kB managed:15904kB mlocked:0kB dirty:756kB writeback:0kB mapped:736kB shmem:184kB slab_reclaimable:48kB slab_unreclaimable:208kB kernel_stack:160kB pagetables:144kB unstable:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:9708 all_unreclaimable? yes lowmem_reserve[]: 0 1732 1732 1732 Node 0 DMA32 free:5200kB min:5200kB low:6500kB high:7800kB active_anon:1151516kB inactive_anon:8196kB active_file:42640kB inactive_file:451076kB unevictable:0kB isolated(anon):116kB isolated(file):3564kB present:2080640kB managed:1775332kB mlocked:0kB dirty:433368kB writeback:21232kB mapped:39144kB shmem:8452kB slab_reclaimable:22056kB slab_unreclaimable:28100kB kernel_stack:20976kB pagetables:9404kB unstable:0kB bounce:0kB free_pcp:120kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:2701604 all_unreclaimable? no lowmem_reserve[]: 0 0 0 0 Node 0 DMA: 25*4kB (UME) 16*8kB (UME) 3*16kB (UE) 5*32kB (UME) 2*64kB (UM) 2*128kB (ME) 2*256kB (ME) 1*512kB (E) 1*1024kB (E) 2*2048kB (ME) 0*4096kB = 6964kB Node 0 DMA32: 925*4kB (UME) 140*8kB (UME) 5*16kB (ME) 5*32kB (M) 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 5060kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=1048576kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=2048kB 126847 total pagecache pages 0 pages in swap cache Swap cache stats: add 0, delete 0, find 0/0 Free swap = 0kB Total swap = 0kB 524157 pages RAM 0 pages HighMem/MovableOnly 76348 pages reserved 0 pages hwpoisoned Out of memory: Kill process 4450 (file_io.00) score 998 or sacrifice child Killed process 4450 (file_io.00) total-vm:4308kB, anon-rss:100kB, file-rss:1184kB, shmem-rss:0kB kthreadd: page allocation failure: order:0, mode:0x2200020 file_io.00: page allocation failure: order:0, mode:0x2200020 CPU: 0 PID: 4457 Comm: file_io.00 Not tainted 4.5.0-rc7+ #45 Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 07/31/2013 Call Trace: warn_alloc_failed+0xf7/0x150 __alloc_pages_nodemask+0x23f/0xa60 alloc_pages_current+0x87/0x110 new_slab+0x3a1/0x440 ___slab_alloc+0x3cf/0x590 __slab_alloc.isra.64+0x18/0x1d kmem_cache_alloc+0x11c/0x150 wb_start_writeback+0x39/0x90 wakeup_flusher_threads+0x7f/0xf0 do_try_to_free_pages+0x1f9/0x410 try_to_free_pages+0x94/0xc0 __alloc_pages_nodemask+0x566/0xa60 alloc_pages_current+0x87/0x110 __page_cache_alloc+0xaf/0xc0 pagecache_get_page+0x88/0x260 grab_cache_page_write_begin+0x21/0x40 xfs_vm_write_begin+0x2f/0xf0 generic_perform_write+0xca/0x1c0 xfs_file_buffered_aio_write+0xcc/0x1f0 xfs_file_write_iter+0x84/0x140 __vfs_write+0xc7/0x100 vfs_write+0x9d/0x190 SyS_write+0x50/0xc0 entry_SYSCALL_64_fastpath+0x12/0x6a Mem-Info: active_anon:293335 inactive_anon:2093 isolated_anon:0 active_file:10829 inactive_file:110045 isolated_file:32 unevictable:0 dirty:109275 writeback:822 unstable:0 slab_reclaimable:5489 slab_unreclaimable:10070 mapped:9999 shmem:2159 pagetables:2420 bounce:0 free:3 free_pcp:0 free_cma:0 Node 0 DMA free:12kB min:44kB low:52kB high:64kB active_anon:6060kB inactive_anon:176kB active_file:708kB inactive_file:756kB unevictable:0kB isolated(anon):0kB isolated(file):0kB present:15988kB managed:15904kB mlocked:0kB dirty:756kB writeback:0kB mapped:736kB shmem:184kB slab_reclaimable:48kB slab_unreclaimable:7160kB kernel_stack:160kB pagetables:144kB unstable:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:9844 all_unreclaimable? yes lowmem_reserve[]: 0 1732 1732 1732 Node 0 DMA32 free:0kB min:5200kB low:6500kB high:7800kB active_anon:1167280kB inactive_anon:8196kB active_file:42608kB inactive_file:439424kB unevictable:0kB isolated(anon):0kB isolated(file):128kB present:2080640kB managed:1775332kB mlocked:0kB dirty:436344kB writeback:3288kB mapped:39260kB shmem:8452kB slab_reclaimable:21908kB slab_unreclaimable:33120kB kernel_stack:20976kB pagetables:9536kB unstable:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:11073180 all_unreclaimable? yes lowmem_reserve[]: 0 0 0 0 Node 0 DMA: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 0kB Node 0 DMA32: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 0kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=1048576kB Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=2048kB 123086 total pagecache pages 0 pages in swap cache Swap cache stats: add 0, delete 0, find 0/0 Free swap = 0kB Total swap = 0kB 524157 pages RAM 0 pages HighMem/MovableOnly 76348 pages reserved 0 pages hwpoisoned SLUB: Unable to allocate memory on node -1 (gfp=0x2088020) cache: kmalloc-64, object size: 64, buffer size: 64, default order: 0, min order: 0 node 0: slabs: 3218, objs: 205952, free: 0 file_io.00: page allocation failure: order:0, mode:0x2200020 CPU: 0 PID: 4457 Comm: file_io.00 Not tainted 4.5.0-rc7+ #45 Assuming that somebody will find a better solution, let's apply this patch for now to stop bleeding, for this problem frequently prevents me from testing OOM livelock condition. Link: http://lkml.kernel.org/r/20160318131136.GE7152@quack.suse.cz Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Acked-by: Michal Hocko <mhocko@suse.com> Cc: Jan Kara <jack@suse.cz> Cc: Tejun Heo <tj@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-04mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macrosKirill A. Shutemov1-1/+1
PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-20writeback, cgroup: fix use of the wrong bdi_writeback which mismatches the inodeTejun Heo1-10/+19
When cgroup writeback is in use, there can be multiple wb's (bdi_writeback's) per bdi and an inode may switch among them dynamically. In a couple places, the wrong wb was used leading to performing operations on the wrong list under the wrong lock corrupting the io lists. * writeback_single_inode() was taking @wb parameter and used it to remove the inode from io lists if it becomes clean after writeback. The callers of this function were always passing in the root wb regardless of the actual wb that the inode was associated with, which could also change while writeback is in progress. Fix it by dropping the @wb parameter and using inode_to_wb_and_lock_list() to determine and lock the associated wb. * After writeback_sb_inodes() writes out an inode, it re-locks @wb and inode to remove it from or move it to the right io list. It assumes that the inode is still associated with @wb; however, the inode may have switched to another wb while writeback was in progress. Fix it by using inode_to_wb_and_lock_list() to determine and lock the associated wb after writeback is complete. As the function requires the original @wb->list_lock locked for the next iteration, in the unlikely case where the inode has changed association, switch the locks. Kudos to Tahsin for pinpointing these subtle breakages. Signed-off-by: Tejun Heo <tj@kernel.org> Fixes: d10c80955265 ("writeback: implement foreign cgroup inode bdi_writeback switching") Link: http://lkml.kernel.org/g/CAAeU0aMYeM_39Y2+PaRvyB1nqAPYZSNngJ1eBRmrxn7gKAt2Mg@mail.gmail.com Reported-and-diagnosed-by: Tahsin Erdogan <tahsin@google.com> Tested-by: Tahsin Erdogan <tahsin@google.com> Cc: stable@vger.kernel.org # v4.2+ Signed-off-by: Jens Axboe <axboe@fb.com>
2016-03-20writeback, cgroup: fix premature wb_put() in locked_inode_to_wb_and_lock_list()Tejun Heo1-3/+5
locked_inode_to_wb_and_lock_list() wb_get()'s the wb associated with the target inode, unlocks inode, locks the wb's list_lock and verifies that the inode is still associated with the wb. To prevent the wb going away between dropping inode lock and acquiring list_lock, the wb is pinned while inode lock is held. The wb reference is put right after acquiring list_lock citing that the wb won't be dereferenced anymore. This isn't true. If the inode is still associated with the wb, the inode has reference and it's safe to return the wb; however, if inode has been switched, the wb still needs to be unlocked which is a dereference and can lead to use-after-free if it it races with wb destruction. Fix it by putting the reference after releasing list_lock. Signed-off-by: Tejun Heo <tj@kernel.org> Fixes: 87e1d789bf55 ("writeback: implement [locked_]inode_to_wb_and_lock_list()") Cc: stable@vger.kernel.org # v4.2+ Tested-by: Tahsin Erdogan <tahsin@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-03-03writeback: flush inode cgroup wb switches instead of pinning super_blockTejun Heo1-13/+41
If cgroup writeback is in use, inodes can be scheduled for asynchronous wb switching. Before 5ff8eaac1636 ("writeback: keep superblock pinned during cgroup writeback association switches"), this could race with umount leading to super_block being destroyed while inodes are pinned for wb switching. 5ff8eaac1636 fixed it by bumping s_active while wb switches are in flight; however, this allowed in-flight wb switches to make umounts asynchronous when the userland expected synchronosity - e.g. fsck immediately following umount may fail because the device is still busy. This patch removes the problematic super_block pinning and instead makes generic_shutdown_super() flush in-flight wb switches. wb switches are now executed on a dedicated isw_wq so that they can be flushed and isw_nr_in_flight keeps track of the number of in-flight wb switches so that flushing can be avoided in most cases. v2: Move cgroup_writeback_umount() further below and add MS_ACTIVE check in inode_switch_wbs() as Jan an Al suggested. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Tahsin Erdogan <tahsin@google.com> Cc: Jan Kara <jack@suse.cz> Cc: Al Viro <viro@ZenIV.linux.org.uk> Link: http://lkml.kernel.org/g/CAAeU0aNCq7LGODvVGRU-oU_o-6enii5ey0p1c26D1ZzYwkDc5A@mail.gmail.com Fixes: 5ff8eaac1636 ("writeback: keep superblock pinned during cgroup writeback association switches") Cc: stable@vger.kernel.org #v4.5 Reviewed-by: Jan Kara <jack@suse.cz> Tested-by: Tahsin Erdogan <tahsin@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2016-02-16writeback: keep superblock pinned during cgroup writeback association switchesTejun Heo1-4/+11
If cgroup writeback is in use, an inode is associated with a cgroup for writeback. If the inode's main dirtier changes to another cgroup, the association gets updated asynchronously. Nothing was pinning the superblock while such switches are in progress and superblock could go away while async switching is pending or in progress leading to crashes like the following. kernel BUG at fs/jbd2/transaction.c:319! invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC CPU: 1 PID: 29158 Comm: kworker/1:10 Not tainted 4.5.0-rc3 #51 Hardware name: Google Google, BIOS Google 01/01/2011 Workqueue: events inode_switch_wbs_work_fn task: ffff880213dbbd40 ti: ffff880209264000 task.ti: ffff880209264000 RIP: 0010:[<ffffffff803e6922>] [<ffffffff803e6922>] start_this_handle+0x382/0x3e0 RSP: 0018:ffff880209267c30 EFLAGS: 00010202 ... Call Trace: [<ffffffff803e6be4>] jbd2__journal_start+0xf4/0x190 [<ffffffff803cfc7e>] __ext4_journal_start_sb+0x4e/0x70 [<ffffffff803b31ec>] ext4_evict_inode+0x12c/0x3d0 [<ffffffff8035338b>] evict+0xbb/0x190 [<ffffffff80354190>] iput+0x130/0x190 [<ffffffff80360223>] inode_switch_wbs_work_fn+0x343/0x4c0 [<ffffffff80279819>] process_one_work+0x129/0x300 [<ffffffff80279b16>] worker_thread+0x126/0x480 [<ffffffff8027ed14>] kthread+0xc4/0xe0 [<ffffffff809771df>] ret_from_fork+0x3f/0x70 Fix it by bumping s_active while cgroup association switching is in flight. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-tested-by: Tahsin Erdogan <tahsin@google.com> Link: http://lkml.kernel.org/g/CAAeU0aNCq7LGODvVGRU-oU_o-6enii5ey0p1c26D1ZzYwkDc5A@mail.gmail.com Fixes: d10c80955265 ("writeback: implement foreign cgroup inode bdi_writeback switching") Cc: stable@vger.kernel.org #v4.5+ Signed-off-by: Jens Axboe <axboe@fb.com>
2016-01-15cgroup, memcg, writeback: drop spurious rcu locking around ↵Tejun Heo1-2/+0
mem_cgroup_css_from_page() In earlier versions, mem_cgroup_css_from_page() could return non-root css on a legacy hierarchy which can go away and required rcu locking; however, the eventual version simply returns the root cgroup if memcg is on a legacy hierarchy and thus doesn't need rcu locking around or in it. Remove spurious rcu lockings. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Jens Axboe <axboe@kernel.dk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-09fs/writeback.c: fix kernel-doc warningsRandy Dunlap1-1/+2
Fix kernel-doc warnings in fs/fs-writeback.c by moving a #define macro to after the function's opening brace. Also #undef this macro at the end of the function. ../fs/fs-writeback.c:1984: warning: Excess function parameter 'inode' description in 'I_DIRTY_INODE' ../fs/fs-writeback.c:1984: warning: Excess function parameter 'flags' description in 'I_DIRTY_INODE' Signed-off-by: Randy Dunlap <rdunlap@infradead.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05mm/filemap.c: make global sync not clear error status of individual inodesJunichi Nomura1-1/+6
filemap_fdatawait() is a function to wait for on-going writeback to complete but also consume and clear error status of the mapping set during writeback. The latter functionality is critical for applications to detect writeback error with system calls like fsync(2)/fdatasync(2). However filemap_fdatawait() is also used by sync(2) or FIFREEZE ioctl, which don't check error status of individual mappings. As a result, fsync() may not be able to detect writeback error if events happen in the following order: Application System admin ---------------------------------------------------------- write data on page cache Run sync command writeback completes with error filemap_fdatawait() clears error fsync returns success (but the data is not on disk) This patch adds filemap_fdatawait_keep_errors() for call sites where writeback error is not handled so that they don't clear error status. Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com> Acked-by: Andi Kleen <ak@linux.intel.com> Reviewed-by: Tejun Heo <tj@kernel.org> Cc: Fengguang Wu <fengguang.wu@gmail.com> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-10-28fs/writeback, rcu: Don't use list_entry_rcu() for pointer offsetting in ↵Tejun Heo1-2/+2
bdi_split_work_to_wbs() bdi_split_work_to_wbs() uses list_for_each_entry_rcu_continue() to walk @bdi->wb_list. To set up the initial iteration condition, it uses list_entry_rcu() to calculate the entry pointer corresponding to the list head; however, this isn't an actual RCU dereference and using list_entry_rcu() for it ended up breaking a proposed list_entry_rcu() change because it was feeding an non-lvalue pointer into the macro. Don't use the RCU variant for simple pointer offsetting. Use list_entry() instead. Reported-by: Ingo Molnar <mingo@kernel.org> Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Darren Hart <dvhart@linux.intel.com> Cc: David Howells <dhowells@redhat.com> Cc: Dipankar Sarma <dipankar@in.ibm.com> Cc: Eric Dumazet <edumazet@google.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Lai Jiangshan <jiangshanlai@gmail.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Patrick Marlier <patrick.marlier@gmail.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: pranith kumar <bobby.prani@gmail.com> Link: http://lkml.kernel.org/r/20151027051939.GA19355@mtj.duckdns.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-10-12writeback: bdi_writeback iteration must not skip dying onesTejun Heo1-9/+22
bdi_for_each_wb() is used in several places to wake up or issue writeback work items to all wb's (bdi_writeback's) on a given bdi. The iteration is performed by walking bdi->cgwb_tree; however, the tree only indexes wb's which are currently active. For example, when a memcg gets associated with a different blkcg, the old wb is removed from the tree so that the new one can be indexed. The old wb starts dying from then on but will linger till all its inodes are drained. As these dying wb's may still host dirty inodes, writeback operations which affect all wb's must include them. bdi_for_each_wb() skipping dying wb's led to sync(2) missing and failing to sync the inodes belonging to those wb's. This patch adds a RCU protected @bdi->wb_list which lists all wb's beloinging to that bdi. wb's are added on creation and removed on release rather than on the start of destruction. bdi_for_each_wb() usages are replaced with list_for_each[_continue]_rcu() iterations over @bdi->wb_list and bdi_for_each_wb() and its helpers are removed. v2: Updated as per Jan. last_wb ref leak in bdi_split_work_to_wbs() fixed and unnecessary list head severing in cgwb_bdi_destroy() removed. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-and-tested-by: Artem Bityutskiy <dedekind1@gmail.com> Fixes: ebe41ab0c79d ("writeback: implement bdi_for_each_wb()") Link: http://lkml.kernel.org/g/1443012552.19983.209.camel@gmail.com Cc: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-10-12writeback: fix bdi_writeback iteration in wakeup_dirtytime_writeback()Tejun Heo1-2/+2
wakeup_dirtytime_writeback() walks and wakes up all wb's of all bdi's; unfortunately, it was always waking up bdi->wb instead of the wb being walked. Fix it. Signed-off-by: Tejun Heo <tj@kernel.org> Fixes: 001fe6f617b1 ("writeback: make wakeup_dirtytime_writeback() handle multiple bdi_writeback's") Reviewed-by: Jan Kara <jack@suse.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-09-19fs-writeback: unplug before cond_resched in writeback_sb_inodesChris Mason1-1/+16
Commit 505a666ee3fc ("writeback: plug writeback in wb_writeback() and writeback_inodes_wb()") has us holding a plug during writeback_sb_inodes, which increases the merge rate when relatively contiguous small files are written by the filesystem. It helps both on flash and spindles. For an fs_mark workload creating 4K files in parallel across 8 drives, this commit improves performance ~9% more by unplugging before calling cond_resched(). cond_resched() doesn't trigger an implicit unplug, so explicitly getting the IO down to the device before scheduling reduces latencies for anyone waiting on clean pages. It also cuts down on how often we use kblockd to unplug, which means less work bouncing from one workqueue to another. Many more details about how we got here: https://lkml.org/lkml/2015/9/11/570 Signed-off-by: Chris Mason <clm@fb.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-12writeback: plug writeback in wb_writeback() and writeback_inodes_wb()Linus Torvalds1-0/+6
We had to revert the pluggin in writeback_sb_inodes() because the wb->list_lock is held, but we could easily plug at a higher level before taking that lock, and unplug after releasing it. This does that. Chris will run performance numbers, just to verify that this approach is comparable to the alternative (we could just drop and re-take the lock around the blk_finish_plug() rather than these two commits. I'd have preferred waiting for actual performance numbers before picking one approach over the other, but I don't want to release rc1 with the known "sleeping function called from invalid context" issue, so I'll pick this cleanup version for now. But if the numbers show that we really want to plug just at the writeback_sb_inodes() level, and we should just play ugly games with the spinlock, we'll switch to that. Cc: Chris Mason <clm@fb.com> Cc: Josef Bacik <jbacik@fb.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Neil Brown <neilb@suse.de> Cc: Jan Kara <jack@suse.cz> Cc: Christoph Hellwig <hch@lst.de> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-11Revert "writeback: plug writeback at a high level"Linus Torvalds1-3/+4
This reverts commit d353d7587d02116b9732d5c06615aed75a4d3a47. Doing the block layer plug/unplug inside writeback_sb_inodes() is broken, because that function is actually called with a spinlock held: wb->list_lock, as pointed out by Chris Mason. Chris suggested just dropping and re-taking the spinlock around the blk_finish_plug() call (the plgging itself can happen under the spinlock), and that would technically work, but is just disgusting. We do something fairly similar - but not quite as disgusting because we at least have a better reason for it - in writeback_single_inode(), so it's not like the caller can depend on the lock being held over the call, but in this case there just isn't any good reason for that "release and re-take the lock" pattern. [ In general, we should really strive to avoid the "release and retake" pattern for locks, because in the general case it can easily cause subtle bugs when the caller caches any state around the call that might be invalidated by dropping the lock even just temporarily. ] But in this case, the plugging should be easy to just move up to the callers before the spinlock is taken, which should even improve the effectiveness of the plug. So there is really no good reason to play games with locking here. I'll send off a test-patch so that Dave Chinner can verify that that plug movement works. In the meantime this just reverts the problematic commit and adds a comment to the function so that we hopefully don't make this mistake again. Reported-by: Chris Mason <clm@fb.com> Cc: Josef Bacik <jbacik@fb.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Neil Brown <neilb@suse.de> Cc: Jan Kara <jack@suse.cz> Cc: Christoph Hellwig <hch@lst.de> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-10Merge branch 'for-4.3/blkcg' of git://git.kernel.dk/linux-blockLinus Torvalds1-96/+43
Pull blk-cg updates from Jens Axboe: "A bit later in the cycle, but this has been in the block tree for a a while. This is basically four patchsets from Tejun, that improve our buffered cgroup writeback. It was dependent on the other cgroup changes, but they went in earlier in this cycle. Series 1 is set of 5 patches that has cgroup writeback updates: - bdi_writeback iteration fix which could lead to some wb's being skipped or repeated during e.g. sync under memory pressure. - Simplification of wb work wait mechanism. - Writeback tracepoints updated to report cgroup. Series 2 is is a set of updates for the CFQ cgroup writeback handling: cfq has always charged all async IOs to the root cgroup. It didn't have much choice as writeback didn't know about cgroups and there was no way to tell who to blame for a given writeback IO. writeback finally grew support for cgroups and now tags each writeback IO with the appropriate cgroup to charge it against. This patchset updates cfq so that it follows the blkcg each bio is tagged with. Async cfq_queues are now shared across cfq_group, which is per-cgroup, instead of per-request_queue cfq_data. This makes all IOs follow the weight based IO resource distribution implemented by cfq. - Switched from GFP_ATOMIC to GFP_NOWAIT as suggested by Jeff. - Other misc review points addressed, acks added and rebased. Series 3 is the blkcg policy cleanup patches: This patchset contains assorted cleanups for blkcg_policy methods and blk[c]g_policy_data handling. - alloc/free added for blkg_policy_data. exit dropped. - alloc/free added for blkcg_policy_data. - blk-throttle's async percpu allocation is replaced with direct allocation. - all methods now take blk[c]g_policy_data instead of blkcg_gq or blkcg. And finally, series 4 is a set of patches cleaning up the blkcg stats handling: blkcg's stats have always been somwhat of a mess. This patchset tries to improve the situation a bit. - The following patches added to consolidate blkcg entry point and blkg creation. This is in itself is an improvement and helps colllecting common stats on bio issue. - per-blkg stats now accounted on bio issue rather than request completion so that bio based and request based drivers can behave the same way. The issue was spotted by Vivek. - cfq-iosched implements custom recursive stats and blk-throttle implements custom per-cpu stats. This patchset make blkcg core support both by default. - cfq-iosched and blk-throttle keep track of the same stats multiple times. Unify them" * 'for-4.3/blkcg' of git://git.kernel.dk/linux-block: (45 commits) blkcg: use CGROUP_WEIGHT_* scale for io.weight on the unified hierarchy blkcg: s/CFQ_WEIGHT_*/CFQ_WEIGHT_LEGACY_*/ blkcg: implement interface for the unified hierarchy blkcg: misc preparations for unified hierarchy interface blkcg: separate out tg_conf_updated() from tg_set_conf() blkcg: move body parsing from blkg_conf_prep() to its callers blkcg: mark existing cftypes as legacy blkcg: rename subsystem name from blkio to io blkcg: refine error codes returned during blkcg configuration blkcg: remove unnecessary NULL checks from __cfqg_set_weight_device() blkcg: reduce stack usage of blkg_rwstat_recursive_sum() blkcg: remove cfqg_stats->sectors blkcg: move io_service_bytes and io_serviced stats into blkcg_gq blkcg: make blkg_[rw]stat_recursive_sum() to be able to index into blkcg_gq blkcg: make blkcg_[rw]stat per-cpu blkcg: add blkg_[rw]stat->aux_cnt and replace cfq_group->dead_stats with it blkcg: consolidate blkg creation in blkcg_bio_issue_check() blk-throttle: improve queue bypass handling blkcg: move root blkg lookup optimization from throtl_lookup_tg() to __blkg_lookup() blkcg: inline [__]blkg_lookup() ...
2015-09-05Merge branch 'for-linus' of ↵Linus Torvalds1-29/+43
git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs Pull vfs updates from Al Viro: "In this one: - d_move fixes (Eric Biederman) - UFS fixes (me; locking is mostly sane now, a bunch of bugs in error handling ought to be fixed) - switch of sb_writers to percpu rwsem (Oleg Nesterov) - superblock scalability (Josef Bacik and Dave Chinner) - swapon(2) race fix (Hugh Dickins)" * 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (65 commits) vfs: Test for and handle paths that are unreachable from their mnt_root dcache: Reduce the scope of i_lock in d_splice_alias dcache: Handle escaped paths in prepend_path mm: fix potential data race in SyS_swapon inode: don't softlockup when evicting inodes inode: rename i_wb_list to i_io_list sync: serialise per-superblock sync operations inode: convert inode_sb_list_lock to per-sb inode: add hlist_fake to avoid the inode hash lock in evict writeback: plug writeback at a high level change sb_writers to use percpu_rw_semaphore shift percpu_counter_destroy() into destroy_super_work() percpu-rwsem: kill CONFIG_PERCPU_RWSEM percpu-rwsem: introduce percpu_rwsem_release() and percpu_rwsem_acquire() percpu-rwsem: introduce percpu_down_read_trylock() document rwsem_release() in sb_wait_write() fix the broken lockdep logic in __sb_start_write() introduce __sb_writers_{acquired,release}() helpers ufs_inode_get{frag,block}(): get rid of 'phys' argument ufs_getfrag_block(): tidy up a bit ...
2015-08-25writeback: sync_inodes_sb() must write out I_DIRTY_TIME inodes and always ↵Tejun Heo1-9/+13
call wait_sb_inodes() e79729123f63 ("writeback: don't issue wb_writeback_work if clean") updated writeback path to avoid kicking writeback work items if there are no inodes to be written out; unfortunately, the avoidance logic was too aggressive and broke sync_inodes_sb(). * sync_inodes_sb() must write out I_DIRTY_TIME inodes but I_DIRTY_TIME inodes dont't contribute to bdi/wb_has_dirty_io() tests and were being skipped over. * inodes are taken off wb->b_dirty/io/more_io lists after writeback starts on them. sync_inodes_sb() skipping wait_sb_inodes() when bdi_has_dirty_io() breaks it by making it return while writebacks are in-flight. This patch fixes the breakages by * Removing bdi_has_dirty_io() shortcut from bdi_split_work_to_wbs(). The callers are already testing the condition. * Removing bdi_has_dirty_io() shortcut from sync_inodes_sb() so that it always calls into bdi_split_work_to_wbs() and wait_sb_inodes(). * Making bdi_split_work_to_wbs() consider the b_dirty_time list for WB_SYNC_ALL writebacks. Kudos to Eryu, Dave and Jan for tracking down the issue. Signed-off-by: Tejun Heo <tj@kernel.org> Fixes: e79729123f63 ("writeback: don't issue wb_writeback_work if clean") Link: http://lkml.kernel.org/g/20150812101204.GE17933@dhcp-13-216.nay.redhat.com Reported-and-bisected-by: Eryu Guan <eguan@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Jan Kara <jack@suse.com> Cc: Ted Ts'o <tytso@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-08-18writeback: update writeback tracepoints to report cgroupTejun Heo1-7/+7
The following tracepoints are updated to report the cgroup used during cgroup writeback. * writeback_write_inode[_start] * writeback_queue * writeback_exec * writeback_start * writeback_written * writeback_wait * writeback_nowork * writeback_wake_background * wbc_writepage * writeback_queue_io * bdi_dirty_ratelimit * balance_dirty_pages * writeback_sb_inodes_requeue * writeback_single_inode[_start] Note that writeback_bdi_register is separated out from writeback_class as reporting cgroup doesn't make sense to it. Tracepoints which take bdi are updated to take bdi_writeback instead. Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Jan Kara <jack@suse.cz> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-08-18writeback: explain why @inode is allowed to be NULL for inode_congested()Tejun Heo1-1/+4
Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-08-18writeback: remove wb_writeback_work->single_wait/doneTejun Heo1-86/+30
wb_writeback_work->single_wait/done are used for the wait mechanism for synchronous wb_work (wb_writeback_work) items which are issued when bdi_split_work_to_wbs() fails to allocate memory for asynchronous wb_work items; however, there's no reason to use a separate wait mechanism for this. bdi_split_work_to_wbs() can simply use on-stack fallback wb_work item and separate wb_completion to wait for it. This patch removes wb_work->single_wait/done and the related code and make bdi_split_work_to_wbs() use on-stack fallback wb_work and wb_completion instead. Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Jan Kara <jack@suse.cz> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-08-18writeback: bdi_for_each_wb() iteration is memcg ID based not blkcgTejun Heo1-3/+3
wb's (bdi_writeback's) are currently keyed by memcg ID; however, in an earlier implementation, wb's were keyed by blkcg ID. bdi_for_each_wb() walks bdi->cgwb_tree in the ascending ID order and allows iterations to start from an arbitrary ID which is used to interrupt and resume iterations. Unfortunately, while changing wb to be keyed by memcg ID instead of blkcg, bdi_for_each_wb() was missed and is still assuming that wb's are keyed by blkcg ID. This doesn't affect iterations which don't get interrupted but bdi_split_work_to_wbs() makes use of iteration resuming on allocation failures and thus may incorrectly skip or repeat wb's. Fix it by changing bdi_for_each_wb() to take memcg IDs instead of blkcg IDs and updating bdi_split_work_to_wbs() accordingly. Signed-off-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-08-17inode: rename i_wb_list to i_io_listDave Chinner1-23/+23
There's a small consistency problem between the inode and writeback naming. Writeback calls the "for IO" inode queues b_io and b_more_io, but the inode calls these the "writeback list" or i_wb_list. This makes it hard to an new "under writeback" list to the inode, or call it an "under IO" list on the bdi because either way we'll have writeback on IO and IO on writeback and it'll just be confusing. I'm getting confused just writing this! So, rename the inode "for IO" list variable to i_io_list so we can add a new "writeback list" in a subsequent patch. Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Josef Bacik <jbacik@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Dave Chinner <dchinner@redhat.com>
2015-08-17sync: serialise per-superblock sync operationsDave Chinner1-0/+11
When competing sync(2) calls walk the same filesystem, they need to walk the list of inodes on the superblock to find all the inodes that we need to wait for IO completion on. However, when multiple wait_sb_inodes() calls do this at the same time, they contend on the the inode_sb_list_lock and the contention causes system wide slowdowns. In effect, concurrent sync(2) calls can take longer and burn more CPU than if they were serialised. Stop the worst of the contention by adding a per-sb mutex to wrap around wait_sb_inodes() so that we only execute one sync(2) IO completion walk per superblock superblock at a time and hence avoid contention being triggered by concurrent sync(2) calls. Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Josef Bacik <jbacik@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Dave Chinner <dchinner@redhat.com>
2015-08-17inode: convert inode_sb_list_lock to per-sbDave Chinner1-6/+6
The process of reducing contention on per-superblock inode lists starts with moving the locking to match the per-superblock inode list. This takes the global lock out of the picture and reduces the contention problems to within a single filesystem. This doesn't get rid of contention as the locks still have global CPU scope, but it does isolate operations on different superblocks form each other. Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Josef Bacik <jbacik@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Dave Chinner <dchinner@redhat.com>
2015-08-17writeback: plug writeback at a high levelDave Chinner1-0/+3
Doing writeback on lots of little files causes terrible IOPS storms because of the per-mapping writeback plugging we do. This essentially causes imeediate dispatch of IO for each mapping, regardless of the context in which writeback is occurring. IOWs, running a concurrent write-lots-of-small 4k files using fsmark on XFS results in a huge number of IOPS being issued for data writes. Metadata writes are sorted and plugged at a high level by XFS, so aggregate nicely into large IOs. However, data writeback IOs are dispatched in individual 4k IOs, even when the blocks of two consecutively written files are adjacent. Test VM: 8p, 8GB RAM, 4xSSD in RAID0, 100TB sparse XFS filesystem, metadata CRCs enabled. Kernel: 3.10-rc5 + xfsdev + my 3.11 xfs queue (~70 patches) Test: $ ./fs_mark -D 10000 -S0 -n 10000 -s 4096 -L 120 -d /mnt/scratch/0 -d /mnt/scratch/1 -d /mnt/scratch/2 -d /mnt/scratch/3 -d /mnt/scratch/4 -d /mnt/scratch/5 -d /mnt/scratch/6 -d /mnt/scratch/7 Result: wall sys create rate Physical write IO time CPU (avg files/s) IOPS Bandwidth ----- ----- ------------ ------ --------- unpatched 6m56s 15m47s 24,000+/-500 26,000 130MB/s patched 5m06s 13m28s 32,800+/-600 1,500 180MB/s improvement -26.44% -14.68% +36.67% -94.23% +38.46% If I use zero length files, this workload at about 500 IOPS, so plugging drops the data IOs from roughly 25,500/s to 1000/s. 3 lines of code, 35% better throughput for 15% less CPU. The benefits of plugging at this layer are likely to be higher for spinning media as the IO patterns for this workload are going make a much bigger difference on high IO latency devices..... Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Josef Bacik <jbacik@fb.com> Reviewed-by: Jan Kara <jack@suse.cz> Tested-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de>
2015-07-23block: export bio_associate_*() and wbc_account_io()Tejun Heo1-0/+1
bio_associate_blkcg(), bio_associate_current() and wbc_account_io() are used to implement cgroup writeback support for filesystems and thus need to be exported. Export them. Signed-off-by: Tejun Heo <tj@kernel.org> Reported-by: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-17writeback: do foreign inode detection iff cgroup writeback is enabledTejun Heo1-3/+13
Currently, even when a filesystem doesn't set the FS_CGROUP_WRITEBACK flag, if the filesystem uses wbc_init_bio() and wbc_account_io(), the foreign inode detection and migration logic still ends up activating cgroup writeback which is unexpected. This patch ensures that the foreign inode detection logic stays disabled when inode_cgwb_enabled() is false by not associating writeback_control's with bdi_writeback's. This also avoids unnecessary operations in wbc_init_bio(), wbc_account_io() and wbc_detach_inode() for filesystems which don't support cgroup writeback. Signed-off-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: disassociate inodes from dying bdi_writebacksTejun Heo1-0/+7
For the purpose of foreign inode detection, wb's (bdi_writeback's) are identified by the associated memcg ID. As we create a separate wb for each memcg, this is enough to identify the active wb's; however, when blkcg is enabled or disabled higher up in the hierarchy, the mapping between memcg and blkcg changes which in turn creates a new wb to service the new mapping. The old wb is unlinked from index and released after all references are drained. The foreign inode detection logic can't detect this condition because both the old and new wb's point to the same memcg and thus never decides to move inodes attached to the old wb to the new one. This patch adds logic to initiate switching immediately in wbc_attach_and_unlock_inode() if the associated wb is dying. We can make the usual foreign detection logic to distinguish the different wb's mapped to the memcg but the dying wb is never gonna be in active service again and there's no point in tracking the usage history and reaching the switch verdict after enough data points are collected. It's already known that the wb has to be switched. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: implement foreign cgroup inode bdi_writeback switchingTejun Heo1-2/+84
As concurrent write sharing of an inode is expected to be very rare and memcg only tracks page ownership on first-use basis severely confining the usefulness of such sharing, cgroup writeback tracks ownership per-inode. While the support for concurrent write sharing of an inode is deemed unnecessary, an inode being written to by different cgroups at different points in time is a lot more common, and, more importantly, charging only by first-use can too readily lead to grossly incorrect behaviors (single foreign page can lead to gigabytes of writeback to be incorrectly attributed). To resolve this issue, cgroup writeback detects the majority dirtier of an inode and transfers the ownership to it. The previous patches implemented the foreign condition detection mechanism and laid the groundwork. This patch implements the actual switching. With the previously implemented [unlocked_]inode_to_wb_and_list_lock() and wb stat transaction, grabbing wb->list_lock, inode->i_lock and mapping->tree_lock gives us full exclusion against all wb operations on the target inode. inode_switch_wb_work_fn() grabs all the locks and transfers the inode atomically along with its RECLAIMABLE and WRITEBACK stats. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: add lockdep annotation to inode_to_wb()Tejun Heo1-2/+3
With the previous three patches, all operations which acquire wb from inode are either under one of inode->i_lock, mapping->tree_lock or wb->list_lock or protected by unlocked_inode_to_wb transaction. This will be depended upon by foreign inode wb switching. This patch adds lockdep assertion to inode_to_wb() so that usages outside the above list locks can be caught easily. There are three exceptions. * locked_inode_to_wb_and_lock_list() is holding wb->list_lock but the wb may not be the inode's. Ensuring that is the function's role after all. Updated to deref inode->i_wb directly. * inode_wb_stat_unlocked_begin() is usually protected by combination of !I_WB_SWITCH and rcu_read_lock(). Updated to deref inode->i_wb directly. * inode_congested() wants to test whether inode->i_wb is set before starting the transaction. Added inode_to_wb_is_valid() which tests inode->i_wb directly. v5: might_lock() removed. It annotates that the lock is grabbed w/ irq enabled which isn't the case and triggering lockdep warning spuriously. v4: might_lock() added to unlocked_inode_to_wb_begin(). v3: inode_congested() conversion added. v2: locked_inode_to_wb_and_lock_list() was missing in the first version. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: use unlocked_inode_to_wb transaction in inode_congested()Tejun Heo1-4/+12
Similar to wb stat updates, inode_congested() accesses the associated wb of an inode locklessly, which will break with foreign inode wb switching. This path updates inode_congested() to use unlocked inode wb access transaction introduced by the previous patch. Combined with the previous two patches, this makes all wb list and access operations to be protected by either of inode->i_lock, wb->list_lock, or mapping->tree_lock while wb switching is in progress. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: implement unlocked_inode_to_wb transaction and use it for stat ↵Tejun Heo1-6/+111
updates The mechanism for detecting whether an inode should switch its wb (bdi_writeback) association is now in place. This patch build the framework for the actual switching. This patch adds a new inode flag I_WB_SWITCHING, which has two functions. First, the easy one, it ensures that there's only one switching in progress for a give inode. Second, it's used as a mechanism to synchronize wb stat updates. The two stats, WB_RECLAIMABLE and WB_WRITEBACK, aren't event counters but track the current number of dirty pages and pages under writeback respectively. As such, when an inode is moved from one wb to another, the inode's portion of those stats have to be transferred together; unfortunately, this is a bit tricky as those stat updates are percpu operations which are performed without holding any lock in some places. This patch solves the problem in a similar way as memcg. Each such lockless stat updates are wrapped in transaction surrounded by unlocked_inode_to_wb_begin/end(). During normal operation, they map to rcu_read_lock/unlock(); however, if I_WB_SWITCHING is asserted, mapping->tree_lock is grabbed across the transaction. In turn, the switching path sets I_WB_SWITCHING and waits for a RCU grace period to pass before actually starting to switch, which guarantees that all stat update paths are synchronizing against mapping->tree_lock. This patch still doesn't implement the actual switching. v3: Updated on top of the recent cancel_dirty_page() updates. unlocked_inode_to_wb_begin() now nests inside mem_cgroup_begin_page_stat() to match the locking order. v2: The i_wb access transaction will be used for !stat accesses too. Function names and comments updated accordingly. s/inode_wb_stat_unlocked_{begin|end}/unlocked_inode_to_wb_{begin|end}/ s/switch_wb/switch_wbs/ Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: implement [locked_]inode_to_wb_and_lock_list()Tejun Heo1-5/+75
cgroup writeback currently assumes that inode to wb association doesn't change; however, with the planned foreign inode wb switching mechanism, the association will change dynamically. When an inode needs to be put on one of the IO lists of its wb, the current code simply calls inode_to_wb() and locks the returned wb; however, with the planned wb switching, the association may change before locking the wb and may even get released. This patch implements [locked_]inode_to_wb_and_lock_list() which pins the associated wb while holding i_lock, releases it, acquires wb->list_lock and verifies that the association hasn't changed inbetween. As the association will be protected by both locks among other things, this guarantees that the wb is the inode's associated wb until the list_lock is released. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: implement foreign cgroup inode detectionTejun Heo1-2/+175
As concurrent write sharing of an inode is expected to be very rare and memcg only tracks page ownership on first-use basis severely confining the usefulness of such sharing, cgroup writeback tracks ownership per-inode. While the support for concurrent write sharing of an inode is deemed unnecessary, an inode being written to by different cgroups at different points in time is a lot more common, and, more importantly, charging only by first-use can too readily lead to grossly incorrect behaviors (single foreign page can lead to gigabytes of writeback to be incorrectly attributed). To resolve this issue, cgroup writeback detects the majority dirtier of an inode and will transfer the ownership to it. To avoid unnnecessary oscillation, the detection mechanism keeps track of history and gives out the switch verdict only if the foreign usage pattern is stable over a certain amount of time and/or writeback attempts. The detection mechanism has fairly low space and computation overhead. It adds 8 bytes to struct inode (one int and two u16's) and minimal amount of calculation per IO. The detection mechanism converges to the correct answer usually in several seconds of IO time when there's a clear majority dirtier. Even when there isn't, it can reach an acceptable answer fairly quickly under most circumstances. Please see wb_detach_inode() for more details. This patch only implements detection. Following patches will implement actual switching. v2: wbc_account_io() now checks whether the wbc is associated with a wb before dereferencing it. This can happen when pageout() is writing pages directly without going through the usual writeback path. As pageout() path is single-threaded, we don't want it to be blocked behind a slow cgroup and ultimately want it to delegate actual writing to the usual writeback path. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: make writeback_control track the inode being written backTejun Heo1-2/+35
Currently, for cgroup writeback, the IO submission paths directly associate the bio's with the blkcg from inode_to_wb_blkcg_css(); however, it'd be necessary to keep more writeback context to implement foreign inode writeback detection. wbc (writeback_control) is the natural fit for the extra context - it persists throughout the writeback of each inode and is passed all the way down to IO submission paths. This patch adds wbc_attach_and_unlock_inode(), wbc_detach_inode(), and wbc_attach_fdatawrite_inode() which are used to associate wbc with the inode being written back. IO submission paths now use wbc_init_bio() instead of directly associating bio's with blkcg themselves. This leaves inode_to_wb_blkcg_css() w/o any user. The function is removed. wbc currently only tracks the associated wb (bdi_writeback). Future patches will add more for foreign inode detection. The association is established under i_lock which will be depended upon when migrating foreign inodes to other wb's. As currently, once established, inode to wb association never changes, going through wbc when initializing bio's doesn't cause any behavior changes. v2: submit_blk_blkcg() now checks whether the wbc is associated with a wb before dereferencing it. This can happen when pageout() is writing pages directly without going through the usual writeback path. As pageout() path is single-threaded, we don't want it to be blocked behind a slow cgroup and ultimately want it to delegate actual writing to the usual writeback path. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: relocate wb[_try]_get(), wb_put(), inode_{attach|detach}_wb()Tejun Heo1-0/+31
Currently, majority of cgroup writeback support including all the above functions are implemented in include/linux/backing-dev.h and mm/backing-dev.c; however, the portion closely related to writeback logic implemented in include/linux/writeback.h and mm/page-writeback.c will expand to support foreign writeback detection and correction. This patch moves wb[_try]_get() and wb_put() to include/linux/backing-dev-defs.h so that they can be used from writeback.h and inode_{attach|detach}_wb() to writeback.h and page-writeback.c. This is pure reorganization and doesn't introduce any functional changes. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: move over_bground_thresh() to mm/page-writeback.cTejun Heo1-18/+2
and rename it to wb_over_bg_thresh(). The function is closely tied to the dirty throttling mechanism implemented in page-writeback.c. This relocation will allow future updates necessary for cgroup writeback support. While at it, add function comment. This is pure reorganization and doesn't introduce any behavioral changes. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: move global_dirty_limit into wb_domainTejun Heo1-1/+1
This patch is a part of the series to define wb_domain which represents a domain that wb's (bdi_writeback's) belong to and are measured against each other in. This will enable IO backpressure propagation for cgroup writeback. global_dirty_limit exists to regulate the global dirty threshold which is a property of the wb_domain. This patch moves hard_dirty_limit, dirty_lock, and update_time into wb_domain. This is pure reorganization and doesn't introduce any behavioral changes. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: reorganize [__]wb_update_bandwidth()Tejun Heo1-10/+0
__wb_update_bandwidth() is called from two places - fs/fs-writeback.c::balance_dirty_pages() and mm/page-writeback.c::wb_writeback(). The latter updates only the write bandwidth while the former also deals with the dirty ratelimit. The two callsites are distinguished by whether @thresh parameter is zero or not, which is cryptic. In addition, the two files define their own different versions of wb_update_bandwidth() on top of __wb_update_bandwidth(), which is confusing to say the least. This patch cleans up [__]wb_update_bandwidth() in the following ways. * __wb_update_bandwidth() now takes explicit @update_ratelimit parameter to gate dirty ratelimit handling. * mm/page-writeback.c::wb_update_bandwidth() is flattened into its caller - balance_dirty_pages(). * fs/fs-writeback.c::wb_update_bandwidth() is moved to mm/page-writeback.c and __wb_update_bandwidth() is made static. * While at it, add a lockdep assertion to __wb_update_bandwidth(). Except for the lockdep addition, this is pure reorganization and doesn't introduce any behavioral changes. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: clean up wb_dirty_limit()Tejun Heo1-1/+1
The function name wb_dirty_limit(), its argument @dirty and the local variable @wb_dirty are mortally confusing given that the function calculates per-wb threshold value not dirty pages, especially given that @dirty and @wb_dirty are used elsewhere for dirty pages. Let's rename the function to wb_calc_thresh() and wb_dirty to wb_thresh. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Cc: Wu Fengguang <fengguang.wu@intel.com> Cc: Greg Thelen <gthelen@google.com> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: dirty inodes against their matching cgroup bdi_writeback'sTejun Heo1-12/+11
__mark_inode_dirty() always dirtied the inode against the root wb (bdi_writeback). The previous patches added all the infrastructure necessary to attribute an inode against the wb of the dirtying cgroup. This patch updates __mark_inode_dirty() so that it uses the wb associated with the inode instead of unconditionally using the root one. Currently, none of the filesystems has FS_CGROUP_WRITEBACK and all pages will keep being dirtied against the root wb. v2: Updated for per-inode wb association. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>
2015-06-02writeback: make writeback initiation functions handle multiple bdi_writeback'sTejun Heo1-5/+91
[try_]writeback_inodes_sb[_nr]() and sync_inodes_sb() currently only handle dirty inodes on the root wb (bdi_writeback) of the target bdi. This patch implements bdi_split_work_to_wbs() and use it to make these functions handle multiple wb's. bdi_split_work_to_wbs() takes a base wb_writeback_work and create clones of it and issue them to the wb's of the target bdi. The base work's nr_pages is distributed using wb_split_bdi_pages() - ie. according to each wb's write bandwidth's proportion in the bdi. Cloning a bdi involves memory allocation which may fail. In such cases, bdi_split_work_to_wbs() issues the base work directly and waits for its completion before proceeding to the next wb to guarantee forward progress and correctness under memory pressure. Signed-off-by: Tejun Heo <tj@kernel.org> Cc: Jens Axboe <axboe@kernel.dk> Cc: Jan Kara <jack@suse.cz> Signed-off-by: Jens Axboe <axboe@fb.com>