Messages in this thread | | | Date | Mon, 12 May 2008 09:10:02 +1000 | From | David Chinner <> | Subject | Re: 2.6.26-rc1: possible circular locking dependency with xfs filesystem |
| |
On Sun, May 11, 2008 at 09:18:07AM +0530, Kamalesh Babulal wrote: > Kamalesh Babulal wrote: > > Adding the cc to kernel-list, Ingo Molnar and Peter Zijlstra > > > > Alexander Beregalov wrote: > >> [ INFO: possible circular locking dependency detected ] > >> 2.6.26-rc1-00279-g28a4acb #13 > >> ------------------------------------------------------- > >> nfsd/3087 is trying to acquire lock: > >> (iprune_mutex){--..}, at: [<c016f947>] shrink_icache_memory+0x38/0x19b > >> > >> but task is already holding lock: > >> (&(&ip->i_iolock)->mr_lock){----}, at: [<c0210b83>] xfs_ilock+0xa2/0xd6 > >> > >> which lock already depends on the new lock. > >> > >> > >> the existing dependency chain (in reverse order) is: > >> > >> -> #1 (&(&ip->i_iolock)->mr_lock){----}: > >> [<c01352e6>] __lock_acquire+0xa0c/0xbc6 > >> [<c013550a>] lock_acquire+0x6a/0x86 > >> [<c012c39a>] down_write_nested+0x33/0x6a > >> [<c0210b5c>] xfs_ilock+0x7b/0xd6 > >> [<c0210cd5>] xfs_ireclaim+0x1d/0x59 > >> [<c022edfe>] xfs_finish_reclaim+0x173/0x195 > >> [<c0230fa3>] xfs_reclaim+0xb3/0x138 > >> [<c023b4cb>] xfs_fs_clear_inode+0x55/0x8e > >> [<c016f60b>] clear_inode+0x83/0xd2 > >> [<c016f88a>] dispose_list+0x3c/0xc1 > >> [<c016fa82>] shrink_icache_memory+0x173/0x19b > >> [<c014a68d>] shrink_slab+0xda/0x14e > >> [<c014a8e5>] try_to_free_pages+0x1e4/0x2a2 > >> [<c0146997>] __alloc_pages_internal+0x23a/0x39d > >> [<c0146b11>] __alloc_pages+0xa/0xc > >> [<c01483b2>] __do_page_cache_readahead+0xaa/0x16a > >> [<c01484bc>] force_page_cache_readahead+0x4a/0x74 > >> [<c014c9b0>] sys_madvise+0x308/0x400 > >> [<c0102b25>] sysenter_past_esp+0x6a/0xb1 > >> [<ffffffff>] 0xffffffff > >> > >> -> #0 (iprune_mutex){--..}: > >> [<c0135203>] __lock_acquire+0x929/0xbc6 > >> [<c013550a>] lock_acquire+0x6a/0x86 > >> [<c0356a6f>] mutex_lock_nested+0xb4/0x226 > >> [<c016f947>] shrink_icache_memory+0x38/0x19b > >> [<c014a68d>] shrink_slab+0xda/0x14e > >> [<c014a8e5>] try_to_free_pages+0x1e4/0x2a2 > >> [<c0146997>] __alloc_pages_internal+0x23a/0x39d > >> [<c0146b11>] __alloc_pages+0xa/0xc > >> [<c01483b2>] __do_page_cache_readahead+0xaa/0x16a > >> [<c014866c>] ondemand_readahead+0x119/0x127 > >> [<c01486cc>] page_cache_async_readahead+0x52/0x5d > >> [<c0178e46>] generic_file_splice_read+0x290/0x4a8 > >> [<c0239f06>] xfs_splice_read+0x4b/0x78 > >> [<c0237713>] xfs_file_splice_read+0x24/0x29 > >> [<c0178182>] do_splice_to+0x45/0x63 > >> [<c01783f6>] splice_direct_to_actor+0xab/0x150 > >> [<c01ce8e1>] nfsd_vfs_read+0x1ed/0x2d0 > >> [<c01ced50>] nfsd_read+0x82/0x99 > >> [<c01d42bc>] nfsd3_proc_read+0xdf/0x12a > >> [<c01cb40b>] nfsd_dispatch+0xcf/0x19e > >> [<c033f484>] svc_process+0x3b3/0x68b > >> [<c01cb939>] nfsd+0x168/0x26b > >> [<c0103747>] kernel_thread_helper+0x7/0x10 > >> [<ffffffff>] 0xffffffff
Oh, yeah, that. Direct inode reclaim through memory pressure.
Effectively memory reclaim inverts locking order w.r.t. iprune_mutex when it recurses into the filesystem. False positive - can never cause a deadlock on XFS. Can't be solved from the XFS side of things without effectively turning off lockdep checking for xfs inode locking.
The fix is needed to lockdep via iprune_mutex annotations here....
> May 9 02:16:46 nomad64 kernel: [42951853.992965] the existing dependency chain (in reverse order) is: > May 9 02:16:46 nomad64 kernel: [42951853.992967] > May 9 02:16:46 nomad64 kernel: [42951853.992968] -> #1 (&(&ip->i_iolock)->mr_lock){----}: > May 9 02:16:46 nomad64 kernel: [42951853.992974] [<ffffffff80261d72>] __lock_acquire+0xf92/0x1080 > May 9 02:16:46 nomad64 kernel: [42951853.992989] [<ffffffff80261f02>] lock_acquire+0xa2/0xd0 > May 9 02:16:46 nomad64 kernel: [42951853.993002] [<ffffffff80255556>] down_write_nested+0x46/0x80 > May 9 02:16:46 nomad64 kernel: [42951853.993018] [<ffffffff80387fb9>] xfs_ilock+0x99/0xa0 > May 9 02:16:46 nomad64 kernel: [42951853.993034] [<ffffffff803a5117>] xfs_free_eofblocks+0x1c7/0x250 > May 9 02:16:46 nomad64 kernel: [42951853.993049] [<ffffffff803a8a26>] xfs_release+0x186/0x1d0 > May 9 02:16:46 nomad64 kernel: [42951853.993062] [<ffffffff803aeeb0>] xfs_file_release+0x10/0x20 > May 9 02:16:46 nomad64 kernel: [42951853.993076] [<ffffffff802a01cc>] __fput+0xcc/0x1c0 > May 9 02:16:46 nomad64 kernel: [42951853.993091] [<ffffffff802a05e6>] fput+0x16/0x20 > May 9 02:16:46 nomad64 kernel: [42951853.993105] [<ffffffff8028865a>] remove_vma+0x4a/0x80 > May 9 02:16:46 nomad64 kernel: [42951853.993120] [<ffffffff802894e1>] do_munmap+0x281/0x2e0 > May 9 02:16:46 nomad64 kernel: [42951853.993134] [<ffffffff8028958b>] sys_munmap+0x4b/0x70 > May 9 02:16:46 nomad64 kernel: [42951853.993148] [<ffffffff8020b62b>] system_call_after_swapgs+0x7b/0x80 > May 9 02:16:46 nomad64 kernel: [42951853.993161] [<ffffffffffffffff>] 0xffffffffffffffff
hmmmm. Sounds like:
fd = open() addr = mmap(fd) close(fd) ..... munmap(addr);
But yes, XFS takes locks in ->release which means.....
> May 9 02:16:46 nomad64 kernel: [42951853.993293] Call Trace: > May 9 02:16:46 nomad64 kernel: [42951853.993297] [<ffffffff8025f2b3>] print_circular_bug_tail+0x83/0x90 > May 9 02:16:46 nomad64 kernel: [42951853.993302] [<ffffffff80261b90>] __lock_acquire+0xdb0/0x1080 > May 9 02:16:46 nomad64 kernel: [42951853.993306] [<ffffffff80222bbd>] ? do_page_fault+0xdd/0x890 > May 9 02:16:46 nomad64 kernel: [42951853.993310] [<ffffffff80261f02>] lock_acquire+0xa2/0xd0 > May 9 02:16:46 nomad64 kernel: [42951853.993313] [<ffffffff80222bbd>] ? do_page_fault+0xdd/0x890 > May 9 02:16:46 nomad64 kernel: [42951853.993317] [<ffffffff806b887b>] down_read+0x3b/0x70 > May 9 02:16:46 nomad64 kernel: [42951853.993320] [<ffffffff80222bbd>] do_page_fault+0xdd/0x890 > May 9 02:16:46 nomad64 kernel: [42951853.993324] [<ffffffff806ba5dd>] error_exit+0x0/0xa9 > May 9 02:16:46 nomad64 kernel: [42951853.993328] [<ffffffff802739b6>] ? file_read_actor+0x46/0x1b0 > May 9 02:16:46 nomad64 kernel: [42951853.993331] [<ffffffff806ba3d6>] ? _read_unlock_irq+0x36/0x60 > May 9 02:16:46 nomad64 kernel: [42951853.993335] [<ffffffff80275dbc>] ? generic_file_aio_read+0x2cc/0x5d0 > May 9 02:16:46 nomad64 kernel: [42951853.993339] [<ffffffff8025ddb9>] ? get_lock_stats+0x19/0x70 > May 9 02:16:46 nomad64 kernel: [42951853.993343] [<ffffffff803b2769>] ? xfs_read+0x139/0x220 > May 9 02:16:46 nomad64 kernel: [42951853.993347] [<ffffffff803af06d>] ? xfs_file_aio_read+0x4d/0x60 > May 9 02:16:46 nomad64 kernel: [42951853.993350] [<ffffffff8029eeb1>] ? do_sync_read+0xf1/0x130 > May 9 02:16:46 nomad64 kernel: [42951853.993354] [<ffffffff802516e0>] ? autoremove_wake_function+0x0/0x40 > May 9 02:16:46 nomad64 kernel: [42951853.993358] [<ffffffff8026089a>] ? trace_hardirqs_on+0xda/0x170 > May 9 02:16:46 nomad64 kernel: [42951853.993361] [<ffffffff80272e45>] ? __rcu_read_unlock+0xb5/0xc0 > May 9 02:16:46 nomad64 kernel: [42951853.993365] [<ffffffff8026089a>] ? trace_hardirqs_on+0xda/0x170 > May 9 02:16:46 nomad64 kernel: [42951853.993369] [<ffffffff803c4381>] ? security_file_permission+0x11/0x20 > May 9 02:16:46 nomad64 kernel: [42951853.993374] [<ffffffff8029f794>] ? vfs_read+0xc4/0x160 > May 9 02:16:46 nomad64 kernel: [42951853.993377] [<ffffffff8029fc30>] ? sys_read+0x50/0x90 > May 9 02:16:46 nomad64 kernel: [42951853.993380] [<ffffffff8020b62b>] ? system_call_after_swapgs+0x7b/0x80
Oh, joy - a page fault during a read() call triggers lock order inversions on the mmap->sem. I don't think this can deadlock (can't be page faulting in a vma that is being torn down), but it's clear from the last trace that the VM has a mmap->sem inversion problem with ->release vs ->read and page faults...
Basically what we are seeing here in both cases is that the VM is calling inode ->release or ->clear_inode methods with different high level locks held. If the filesystem has to take the same locks in these methods as it does in, say, ->read (like XFS does), then we are guaranteed to get reports like this. AFAICT there's nothing we can do from the filesystem perspective to prevent false positives like this from being reported....
Cheers,
Dave. -- Dave Chinner Principal Engineer SGI Australian Software Group
| |