> On Apr 19, 2023, at 4:55 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > > On Fri, Apr 14, 2023 at 03:58:35PM -0700, Wengang Wang wrote: >> At IO time, make sure an EFI contains only one extent. Transaction rolling in >> xfs_defer_finish() would commit the busy blocks for previous EFIs. By that we >> avoid holding busy extents (for previously extents in the same EFI) in current >> transaction when allocating blocks for AGFL where we could be otherwise stuck >> waiting the busy extents held by current transaction to be flushed (thus a >> deadlock). >> >> The log changes >> 1) before change: >> >> 358 rbbn 13 rec_lsn: 1,12 Oper 5: tid: ee327fd2 len: 48 flags: None >> 359 EFI nextents:2 id:ffff9fef708ba940 >> 360 EFI id=ffff9fef708ba940 (0x21, 7) >> 361 EFI id=ffff9fef708ba940 (0x18, 8) >> 362 ----------------------------------------------------------------- >> 363 rbbn 13 rec_lsn: 1,12 Oper 6: tid: ee327fd2 len: 48 flags: None >> 364 EFD nextents:2 id:ffff9fef708ba940 >> 365 EFD id=ffff9fef708ba940 (0x21, 7) >> 366 EFD id=ffff9fef708ba940 (0x18, 8) >> >> 2) after change: >> >> 830 rbbn 31 rec_lsn: 1,30 Oper 5: tid: 319f015f len: 32 flags: None >> 831 EFI nextents:1 id:ffff9fef708b9b80 >> 832 EFI id=ffff9fef708b9b80 (0x21, 7) >> 833 ----------------------------------------------------------------- >> 834 rbbn 31 rec_lsn: 1,30 Oper 6: tid: 319f015f len: 32 flags: None >> 835 EFI nextents:1 id:ffff9fef708b9d38 >> 836 EFI id=ffff9fef708b9d38 (0x18, 8) >> 837 ----------------------------------------------------------------- >> 838 rbbn 31 rec_lsn: 1,30 Oper 7: tid: 319f015f len: 32 flags: None >> 839 EFD nextents:1 id:ffff9fef708b9b80 >> 840 EFD id=ffff9fef708b9b80 (0x21, 7) >> 841 ----------------------------------------------------------------- >> 842 rbbn 31 rec_lsn: 1,30 Oper 8: tid: 319f015f len: 32 flags: None >> 843 EFD nextents:1 id:ffff9fef708b9d38 >> 844 EFD id=ffff9fef708b9d38 (0x18, 8) >> >> Signed-off-by: Wengang Wang <wen.gang.wang@xxxxxxxxxx> >> --- >> fs/xfs/xfs_extfree_item.h | 9 ++++++++- >> 1 file changed, 8 insertions(+), 1 deletion(-) >> >> diff --git a/fs/xfs/xfs_extfree_item.h b/fs/xfs/xfs_extfree_item.h >> index da6a5afa607c..ae84d77eaf30 100644 >> --- a/fs/xfs/xfs_extfree_item.h >> +++ b/fs/xfs/xfs_extfree_item.h >> @@ -13,8 +13,15 @@ struct kmem_cache; >> >> /* >> * Max number of extents in fast allocation path. >> + * >> + * At IO time, make sure an EFI contains only one extent. Transaction rolling >> + * in xfs_defer_finish() would commit the busy blocks for previous EFIs. By >> + * that we avoid holding busy extents (for previously extents in the same EFI) >> + * in current transaction when allocating blocks for AGFL where we could be >> + * otherwise stuck waiting the busy extents held by current transaction to be >> + * flushed (thus a deadlock). >> */ >> -#define XFS_EFI_MAX_FAST_EXTENTS 16 >> +#define XFS_EFI_MAX_FAST_EXTENTS 1 > > IIRC, this doesn't have anything to do with the number of extents an > EFI can hold. All it does is control how the memory for the EFI > allocated. > Yes, it ensures that one EFI contains at most one extent. And because each deferred intent goes with one transaction roll, it would solve the AGFL allocation deadlock (because no busy extents held by the process when it is doing the AGFL allocation). And yes, this would requires more log space if the EFI/EFD pair doesn’t appear in same checkpoint. > Oh, at some point it got overloaded code to define the max items in > a defer ops work item. Ok, I now see why you changed this, but I > don't think this is right way to solve the problem. We can handle > processing multiple extents per EFI just fine, we just need to > update the EFD and roll the transaction on each extent we process, > yes? > I am not quite sure what does “update the EFD” mean. My original concern is that (without your updated EFD), the extents in original EFI can be partially done before a crush. And during the recovery, the already done extents would also be replayed and hit error (because the in-place metadata could be flushed since the transaction is rolled.). Now consider your “update the EFD”, you meant the following? EFI: ID: THISISID1 extent1 extent2 free extent extent1 EFD: ID: THISISID1 extent1 free extent extent2 another EFD: ID: THISISID1 (same ID as above) extent2 Do we currently support that? ( I am thinking NO). thanks, wengang > In hindsight, the use of XFS_EFI_MAX_FAST_EXTENTS to limit intent > size is pretty awful. I also see the same pattern copied in every > other intent. > > Darrick, if the deferops code has been limiting the number of > extents in a work item to this value, when will we ever see an > intent with more extents that .max_items in it? And if that is the > case, then why wouldn't we consider an intent with more extents than > we support in log recovery a corruption event? > > Cheers, > > Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx
