This is food for thought. On XFS, the only difference between sync and freeze is that freeze stops incoming writers: $ dd if=/dev/zero of=/mnt/scratch/test bs=1024k count=8000 & [1] 2565 $ sleep 5; time (sudo xfs_freeze -f /mnt/scratch ; sudo xfs_freeze -u /mnt/scratch) real 0m2.536s user 0m0.000s sys 0m0.020s $ sleep 5; time (sudo xfs_freeze -f /mnt/scratch ; sudo xfs_freeze -u /mnt/scratch) real 0m2.251s user 0m0.004s sys 0m0.012s $ sleep 5; time (sudo xfs_freeze -f /mnt/scratch ; sudo xfs_freeze -u /mnt/scratch) real 0m1.985s user 0m0.004s sys 0m0.024s $ sleep 5; time sync 8000+0 records in 8000+0 records out 8388608000 bytes (8.4 GB) copied, 80.822 s, 104 MB/s [1]+ Done dd if=/dev/zero of=/mnt/scratch/test bs=1024k count=8000 real 0m47.237s user 0m0.036s sys 0m18.769s $ Freezing the filesystem and immediately unfreezing is much, much faster than running sync, yet it gives us exactly the same data integrity guarantees without the endless blocking problems. Is it time to make sure every filesystem implements freeze and thaw, and start using them for sync instead of the current code? Cheers, Dave. On Mon, Apr 19, 2010 at 05:04:58PM +1000, Dave Chinner wrote: > On Mon, Apr 19, 2010 at 11:37:02AM +1000, Dave Chinner wrote: > > On Thu, Apr 08, 2010 at 11:28:50AM +0200, Jan Kara wrote: > > > > SUPERPROXY ~ # cat /proc/1753/stack > > > > [<c019a93c>] bdi_sched_wait+0x8/0xc > > > > [<c019a807>] wait_on_bit+0x20/0x2c > > > > [<c019a9af>] sync_inodes_sb+0x6f/0x10a > > > > [<c019dd53>] __sync_filesystem+0x28/0x49 > > > > [<c019ddf3>] sync_filesystems+0x7f/0xc0 > > > > [<c019de7a>] sys_sync+0x1b/0x2d > > > > [<c02f7a25>] syscall_call+0x7/0xb > > > > [<ffffffff>] 0xffffffff > > > Hmm, I guess you are observing the problem reported in > > > https://bugzilla.kernel.org/show_bug.cgi?id=14830 > > > There seem to be several issues in the per-bdi writeback code that > > > cause sync on a busy filesystem to last almost forever. To that bug are > > > attached two patches that fix two issues but apparently it's not all. > > > I'm still looking into it... > > > > So Jen's writeback tracing shows this for a normal cycle during a > > large dd: > > > > <...>-6030 [005] 604446.696454: writeback_sched: work=38c0, task=task > > flush-253:16-6029 [002] 604446.696492: writeback_exec: work=38c0 pages=9223372036854775807, sb=0, kupdate=0, range_cyclic=-1 for_background=-1 > > flush-253:16-6029 [002] 604446.696493: writeback_clear: work=ffff88011f1a38c0, refs=1 > > flush-253:16-6029 [003] 604446.784240: writeback_pages_written: 1024 > > > > There were 100 of these traces (102400 pages (400MB) which is exactly 10% of > > RAM) before this: > > > > <...>-6030 [000] 604462.346329: writeback_sched: work=6c0, task=task > > flush-253:16-6029 [002] 604462.267098: writeback_exec: work=6c0 pages=9223372036854775807, sb=0, kupdate=0, range_cyclic=-1 for_background=-1 > > flush-253:16-6029 [002] 604462.267101: writeback_clear: work=ffff88011e8006c0, refs=1 > > flush-253:16-6029 [001] 604465.406331: writeback_pages_written: 160771 > > > > which shows 160771 pages written in a single iteration (650MB). I suspect some > > interaction between foreground and background writeout is occurring here. > > > > The flusher thread appears to be getting stuck on congestion - the wchan > > it is sitting on indicates it is blocking in get_request(). I'm going to > > extend this tracing further down into the writeback code so that what is > > happening is clearer... > > Ok, I moved to a VM with less memory (1GB vs 4GB) and slower block > devices (100MB/s vs 500MB/s) and now it's pretty clear what is > happening. i'll go through the traces. > > To start with, sync() queues up several tasks to the bdi flush daemon: > > SYSCALL_DEFINE0(sync) > { > wakeup_flusher_threads(0); > sync_filesystems(0); > sync_filesystems(1); > if (unlikely(laptop_mode)) > laptop_sync_completion(); > return 0; > } > > > First is via wakeup_flusher_threads() - an async flush: > > sync-2499 [000] 616072.710212: writeback_queue: 253:16: pages=13818, sb=0, kupdate=0, range_cyclic=0 for_background=0 > sync-2499 [000] 616072.710217: writeback_sched: work=13c0, task=task > ^^^^ > second is a per-sb async flush via sync_filesystems(0): > > sync-2499 [000] 616072.717181: writeback_queue: 253:16: pages=15761, sb=1, kupdate=0, range_cyclic=-1 for_background=0 > sync-2499 [000] 616072.717191: writeback_sched: work=3840, task=task > ^^^^ > And finally there is the sync flush via sync_filesystems(1): > > sync-2499 [000] 616072.737836: writeback_queue: 253:16: pages=9223372036854775807, sb=1, kupdate=0, range_cyclic=0 for_background=0 > sync-2499 [000] 616072.737840: writeback_sched: work=3e58, task=task > ^^^^ > > The first async flush does: > vvvv > flush-253:16-2497 [000] 616072.897747: writeback_exec: work=13c0 pages=13818, sb=0, kupdate=0, range_cyclic=0 for_background=0 > flush-253:16-2497 [000] 616072.897748: writeback_clear: work=ffff88003d8813c0, refs=1 > flush-253:16-2497 [000] 616072.897753: wbc_writeback_start: dev 253:16 wbc=9d20 towrt=1024 skip=0 sb=0 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=0 more=0 older=0x0 start=0x0 end=0x7fffffffffffffff > flush-253:16-2497 [000] 616072.897768: wbc_writeback_written: dev 253:16 wbc=9d20 towrt=1024 skip=0 sb=0 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=0 more=0 older=0x0 start=0x0 end=0x7fffffffffffffff > > Nothing - it does not write any pages towrt (nr_to_write) is > unchanged by the attempted flush. > > The second async flush: > vvvv > flush-253:16-2497 [000] 616072.897769: writeback_exec: work=3840 pages=15761, sb=1, kupdate=0, range_cyclic=-1 for_background=0 > flush-253:16-2497 [000] 616072.897770: writeback_clear: work=ffff88003fb53840, refs=1 > flush-253:16-2497 [000] 616072.897771: wbc_writeback_start: dev 253:16 wbc=9d20 towrt=1024 skip=0 sb=1 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > flush-253:16-2497 [000] 616072.897783: wbc_writeback_written: dev 253:16 wbc=9d20 towrt=1024 skip=0 sb=1 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > > Differs in setup only by range_cyclic=1 instead of zero, and it also > fails to write anything. > > The third flush - the sync one - does: > vvvv > flush-253:16-2497 [000] 616072.897784: writeback_exec: work=3e58 pages=9223372036854775807, sb=1, kupdate=0, range_cyclic=0 for_background=0 > flush-253:16-2497 [000] 616072.897785: wbc_writeback_start: dev 253:16 wbc=9d20 towrt=1024 skip=0 sb=1 mode=1 kupd=0 bgrd=0 reclm=0 cyclic=0 more=0 older=0x0 start=0x0 end=0x7fffffffffffffff > > some 75 seconds later having written only 1024 pages. In the mean > time, the traces show dd blocked in balance_dirty_pages(): > > dd-2498 [000] 616072.908675: wbc_balance_dirty_start: dev 253:16 wbc=fb68 towrt=1536 skip=0 sb=0 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > dd-2498 [000] 616072.908679: wbc_balance_dirty_wait: dev 253:16 wbc=fb68 towrt=1536 skip=0 sb=0 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > dd-2498 [000] 616073.238785: wbc_balance_dirty_start: dev 253:16 wbc=fb68 towrt=1536 skip=0 sb=0 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > dd-2498 [000] 616073.238788: wbc_balance_dirty_wait: dev 253:16 wbc=fb68 towrt=1536 skip=0 sb=0 mode=0 kupd=0 bgrd=0 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > > And it appears to stay blocked there without doing any writeback at > all - there are no wbc_balance_dirty_pages_written traces at all. > That is, it is blocking until the number of dirty pages is dropping > below the dirty threshold, then continuing to write and dirty more > pages. > > This continues for another 75 seconds, until the dd completes and > then the sync flush thread completes: > > flush-253:16-2497 [000] 616145.763145: wbc_writeback_written: dev 253:16 wbc=9d20 towrt=0 skip=0 sb=1 mode=1 kupd=0 bgrd=0 reclm=0 cyclic=0 more=1 older=0x0 start=0x0 end=0x7fffffffffffffff > flush-253:16-2497 [000] 616145.763148: wbc_writeback_start: dev 253:16 wbc=9d20 towrt=1024 skip=0 sb=1 mode=1 kupd=0 bgrd=0 reclm=0 cyclic=0 more=0 older=0x0 start=0x0 end=0x7fffffffffffffff > flush-253:16-2497 [000] 616145.763160: wbc_writeback_written: dev 253:16 wbc=9d20 towrt=1024 skip=0 sb=1 mode=1 kupd=0 bgrd=0 reclm=0 cyclic=0 more=0 older=0x0 start=0x0 end=0x7fffffffffffffff > flush-253:16-2497 [000] 616145.763161: writeback_clear: work=ffff88002e783e58, refs=1 > ^^^^^ > > The flush thread does not appear to be cycling through 1024 pages at > a time as the wbc structure says it should - it appears to be doing > all the writeback. Indeed, it is almost always blocked here: > > task PC stack pid father > flush-253:16 D 00000000ffffffff 0 2511 2 0x00000000 > ffff880038409690 0000000000000046 ffff880038409610 00000000001d42c0 > ffff880038409fd8 00000000001d42c0 ffff880038409fd8 ffff88003840c340 > 00000000001d42c0 00000000001d42c0 ffff880038409fd8 00000000001d42c0 > Call Trace: > [<ffffffff81790197>] io_schedule+0x47/0x70 > [<ffffffff8141b637>] get_request_wait+0xc7/0x190 > [<ffffffff8109d880>] ? autoremove_wake_function+0x0/0x40 > [<ffffffff81414817>] ? elv_merge+0x47/0x220 > [<ffffffff8141bce3>] __make_request+0x93/0x480 > [<ffffffff8141a359>] generic_make_request+0x1f9/0x510 > [<ffffffff810b41bd>] ? trace_hardirqs_on+0xd/0x10 > [<ffffffff8117e462>] ? bvec_alloc_bs+0x62/0x110 > [<ffffffff8141a6ca>] submit_bio+0x5a/0xd0 > [<ffffffff8134f874>] xfs_submit_ioend_bio+0x74/0xa0 > [<ffffffff8134fbb1>] xfs_submit_ioend+0xb1/0x110 > [<ffffffff81350e34>] xfs_page_state_convert+0x3a4/0x730 > [<ffffffff810b416d>] ? trace_hardirqs_on_caller+0x14d/0x190 > [<ffffffff8135137c>] xfs_vm_writepage+0x8c/0x160 > [<ffffffff81112cfa>] __writepage+0x1a/0x50 > [<ffffffff81113b17>] write_cache_pages+0x1f7/0x400 > [<ffffffff81112ce0>] ? __writepage+0x0/0x50 > [<ffffffff81113d47>] generic_writepages+0x27/0x30 > [<ffffffff8134f28d>] xfs_vm_writepages+0x5d/0x80 > [<ffffffff81113d74>] do_writepages+0x24/0x40 > [<ffffffff811722f7>] writeback_single_inode+0xe7/0x3b0 > [<ffffffff81172d65>] writeback_inodes_wb+0x2e5/0x550 > [<ffffffff811247fb>] ? ftrace_raw_event_id_wbc_class+0x16b/0x190 > [<ffffffff811730c2>] wb_writeback+0xf2/0x2d0 > [<ffffffff811243aa>] ? ftrace_raw_event_writeback_exec+0xea/0xf0 > [<ffffffff811734c8>] wb_do_writeback+0x108/0x240 > [<ffffffff811733f0>] ? wb_do_writeback+0x30/0x240 > [<ffffffff8117365b>] bdi_writeback_task+0x5b/0x180 > [<ffffffff81125ac0>] ? bdi_start_fn+0x0/0x100 > [<ffffffff81125b46>] bdi_start_fn+0x86/0x100 > [<ffffffff81125ac0>] ? bdi_start_fn+0x0/0x100 > [<ffffffff8109d396>] kthread+0x96/0xa0 > [<ffffffff81036e24>] kernel_thread_helper+0x4/0x10 > [<ffffffff817934d0>] ? restore_args+0x0/0x30 > [<ffffffff8109d300>] ? kthread+0x0/0xa0 > [<ffffffff81036e20>] ? kernel_thread_helper+0x0/0x10 > > Waiting on block device congestion. > > Because I have this in wb_writeback(): > > 756 trace_wbc_writeback_start(&wbc); > 757 writeback_inodes_wb(wb, &wbc); > 758 trace_wbc_writeback_written(&wbc); > > I know that we are stuck in a single iteration of > writeback_inodes_wb(). This also implies that we are stuck in a > single do_writepages() call. > > Indeed, looking at write_cache_pages(): > > 838 long nr_to_write = wbc->nr_to_write; > ... > 920 ret = (*writepage)(page, wbc, data); > ... > 940 if (nr_to_write > 0) { > 941 nr_to_write--; > 942 if (nr_to_write == 0 && > 943 wbc->sync_mode == WB_SYNC_NONE) { > 944 /* > 945 * We stop writing back only if we are > 946 * not doing integrity sync. In case of > 947 * integrity sync we have to keep going > 948 * because someone may be concurrently > 949 * dirtying pages, and we might have > 950 * synced a lot of newly appeared dirty > 951 * pages, but have not synced all of the > 952 * old dirty pages. > 953 */ > 954 done = 1; > 955 break; > 956 } > 957 } > ... > 973 if (!wbc->no_nrwrite_index_update) { > 974 if (wbc->range_cyclic || (range_whole && nr_to_write > 0)) > 975 mapping->writeback_index = done_index; > 976 wbc->nr_to_write = nr_to_write; > 977 } > > It even hides this fact from the higher layers by rewriting > wbc->nr_to_write with what it thinks it did, not what really > happened. So, where did this come from? > > <git blame> > > commit 89e12190 - fix bug in nr_to_write introduced by dcf6a79d > commit dcf6a79d - fix bug in nr_to_write introduced by 05fe478d > commit 05fe478d - data integrity write fix: ignore nr_to_write for > WB_SYNC_ALL writes. > "This change does indeed make the possibility of > long stalls la[r]ger, and that's not a good thing, > but lying about data integrity is even worse." > > IOWs, the observed sync behaviour is as intended - if you keep > dirtying the file, sync will keep cleaning it because it defaults to > being safe. I'd say "not a bug" then. I agree it's not ideal, but > until Jan's inode sync sweep code is accepted I don't think there's > much that can be done about it. > > However, what this is doing to XFS writeback is really, really nasty > - it's effectively causing single page allocation and IO submission > instead of doing it in much, much larger chunks. > > Adding a wbc trace to xfs_vm_writepage(), I see: > > flush-253:16-2586 [000] 620402.417931: wbc_writepage: dev 253:16 wbc=3d20 towrt=1024 skip=0 sb=0 mode=0 kupd=0 bgrd=1 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > flush-253:16-2586 [000] 620402.442765: wbc_writepage: dev 253:16 wbc=3d20 towrt=0 skip=0 sb=0 mode=0 kupd=0 bgrd=1 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > flush-253:16-2586 [000] 620402.442899: wbc_writepage: dev 253:16 wbc=3d20 towrt=-1 skip=0 sb=0 mode=0 kupd=0 bgrd=1 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > flush-253:16-2586 [000] 620402.442910: wbc_writepage: dev 253:16 wbc=3d20 towrt=-5 skip=0 sb=0 mode=0 kupd=0 bgrd=1 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > flush-253:16-2586 [000] 620402.442918: wbc_writepage: dev 253:16 wbc=3d20 towrt=-21 skip=0 sb=0 mode=0 kupd=0 bgrd=1 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > flush-253:16-2586 [000] 620402.442927: wbc_writepage: dev 253:16 wbc=3d20 towrt=-85 skip=0 sb=0 mode=0 kupd=0 bgrd=1 reclm=0 cyclic=1 more=0 older=0x0 start=0x0 end=0x0 > > Which shows why XFS is having problems. Basically, if a filesystem > writes more than one page in ->writepage and updates > wbc->nr_to_write to indicate this, write_cache_pages completely > ignores it. IOWs, write_cache_pages() wants to call ->writepage() > nr_to_write times, not write nr_to_write pages. And by sending a > negative number down to ->writepage, XFs is writing a single page > and then returning, completely defeating all the page clustering > optimisations XFS has in ->writepage.... > > I'll post some patches for the tracing and the XFS fixes soon, but i > don't have anything for sync except understanding, though... > > Cheers, > > Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx > -- > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html > Please read the FAQ at http://www.tux.org/lkml/ > -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
