On Thu 10-11-22 11:27:01, Jeremi Piotrowski wrote: > On Thu, Nov 10, 2022 at 04:26:37PM +0100, Jan Kara wrote: > > On Thu 10-11-22 04:57:58, Jeremi Piotrowski wrote: > > > On Wed, Oct 26, 2022 at 12:18:54PM +0200, Jan Kara wrote: > > > > On Mon 24-10-22 18:32:51, Thilo Fromm wrote: > > > > > Hello Honza, > > > > > > > > > > > Yeah, I was pondering about this for some time but still I have no clue who > > > > > > could be holding the buffer lock (which blocks the task holding the > > > > > > transaction open) or how this could related to the commit you have > > > > > > identified. I have two things to try: > > > > > > > > > > > > 1) Can you please check whether the deadlock reproduces also with 6.0 > > > > > > kernel? The thing is that xattr handling code in ext4 has there some > > > > > > additional changes, commit 307af6c8793 ("mbcache: automatically delete > > > > > > entries from cache on freeing") in particular. > > > > > > > > > > This would be complex; we currently do not integrate 6.0 with Flatcar and > > > > > would need to spend quite some effort ingesting it first (mostly, make sure > > > > > the new kernel does not break something unrelated). Flatcar is an > > > > > image-based distro, so kernel updates imply full distro updates. > > > > > > > > OK, understood. > > > > > > > > > > 2) I have created a debug patch (against 5.15.x stable kernel). Can you > > > > > > please reproduce the failure with it and post the output of "echo w > > > > > > > /proc/sysrq-trigger" and also the output the debug patch will put into the > > > > > > kernel log? It will dump the information about buffer lock owner if we > cannot get the lock for more than 32 seconds. > > > > > > > > > > This would be more straightforward - I can reach out to one of our users > > > > > suffering from the issue; they can reliably reproduce it and don't shy away > > > > > from patching their kernel. Where can I find the patch? > > > > > > > > Ha, my bad. I forgot to attach it. Here it is. > > > > > > > > > > Unfortunately this patch produced no output, but I have been able to repro so I > > > understand why: except for the hung tasks, we have 1+ tasks busy-looping through > > > the following code in ext4_xattr_block_set(): > > > > > > inserted: > > > if (!IS_LAST_ENTRY(s->first)) { > > > new_bh = ext4_xattr_block_cache_find(inode, header(s->base), > > > &ce); > > > if (new_bh) { > > > /* We found an identical block in the cache. */ > > > if (new_bh == bs->bh) > > > ea_bdebug(new_bh, "keeping"); > > > else { > > > u32 ref; > > > > > > WARN_ON_ONCE(dquot_initialize_needed(inode)); > > > > > > /* The old block is released after updating > > > the inode. */ > > > error = dquot_alloc_block(inode, > > > EXT4_C2B(EXT4_SB(sb), 1)); > > > if (error) > > > goto cleanup; > > > BUFFER_TRACE(new_bh, "get_write_access"); > > > error = ext4_journal_get_write_access( > > > handle, sb, new_bh, > > > EXT4_JTR_NONE); > > > if (error) > > > goto cleanup_dquot; > > > lock_buffer(new_bh); > > > /* > > > * We have to be careful about races with > > > * adding references to xattr block. Once we > > > * hold buffer lock xattr block's state is > > > * stable so we can check the additional > > > * reference fits. > > > */ > > > ref = le32_to_cpu(BHDR(new_bh)->h_refcount) + 1; > > > if (ref > EXT4_XATTR_REFCOUNT_MAX) { > > > /* > > > * Undo everything and check mbcache > > > * again. > > > */ > > > unlock_buffer(new_bh); > > > dquot_free_block(inode, > > > EXT4_C2B(EXT4_SB(sb), > > > 1)); > > > brelse(new_bh); > > > mb_cache_entry_put(ea_block_cache, ce); > > > ce = NULL; > > > new_bh = NULL; > > > goto inserted; > > > } > > > > > > The tasks keep taking the 'goto inserted' branch, and never finish. I've been > > > able to repro with kernel v6.0.7 as well. > > > > Interesting! That makes is much clearer (and also makes my debug patch > > unnecessary). So clearly the e_reusable variable in the mb_cache_entry got > > out of sync with the number of references really in the xattr block - in > > particular the block likely has h_refcount >= EXT4_XATTR_REFCOUNT_MAX but > > e_reusable is set to true. Now I can see how e_reusable can stay at false due > > to a race when refcount is actually smaller but I don't see how it could > > stay at true when refcount is big enough - that part seems to be locked > > properly. If you can reproduce reasonably easily, can you try reproducing > > with attached patch? Thanks! > > > > Sure, with that patch I'm getting the following output, reusable is false on > most items until we hit something with reusable true and then that loops > indefinitely: Thanks. So that is what I've suspected. I'm still not 100% clear on how this inconsistency can happen although I have a suspicion - does attached patch fix the problem for you? Also is it possible to share the reproducer or it needs some special infrastructure? Honza -- Jan Kara SUSE Labs, CR