iomap: don't lose folio dropbehind state for overwrites
Subsystems (3)
PATCH 1/1 iomap: don't lose folio dropbehind state for overwrites
DONTCACHE I/O must have the completion punted to a workqueue, just like what is done for unwritten extents, as the completion needs task context to perform the invalidation of the folio(s). However, if writeback is started off filemap_fdatawrite_range() off generic_sync() and it's an overwrite, then the DONTCACHE marking gets lost as iomap_add_to_ioend() don't look at the folio being added and no further state is passed down to help it know that this is a dropbehind/DONTCACHE write.
Check if the folio being added is marked as dropbehind, and set IOMAP_IOEND_DONTCACHE if that is the case. Then XFS can factor this into the decision making of completion context in xfs_submit_ioend(). Additionally include this ioend flag in the NOMERGE flags, to avoid mixing it with unrelated IO.
This fixes extra page cache being instantiated when the write performed is an overwrite, rather than newly instantiated blocks.
---
Found this one while testing the unrelated issue of invalidation being a bit broken before 6.15 release. We need this to ensure that overwrites also prune correctly, just like unwritten extents currently do.
fs/iomap/buffered-io.c +2 -0
@@ -1691,6 +1691,8 @@ static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc, ioend_flags |=IOMAP_IOEND_UNWRITTEN ;if (wpc->iomap.flags &IOMAP_F_SHARED ) ioend_flags |=IOMAP_IOEND_SHARED ; +if (folio_test_dropbehind (folio)) + ioend_flags |=IOMAP_IOEND_DONTCACHE ;if (pos == wpc->iomap.offset && (wpc->iomap.flags &IOMAP_F_BOUNDARY )) ioend_flags |=IOMAP_IOEND_BOUNDARY ;
fs/xfs/xfs_aops.c +4 -2
@@ -436,6 +436,9 @@ xfs_map_blocks(return 0 ; } +#define IOEND_WQ_FLAGS (IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_SHARED | \ + IOMAP_IOEND_DONTCACHE) + static int xfs_submit_ioend( struct iomap_writepage_ctx *wpc,
@@ -460,8 +463,7 @@ xfs_submit_ioend(memalloc_nofs_restore (nofs_flag);/* send ioends that might require a transaction to the completion wq */ -if (xfs_ioend_is_append (ioend) || - (ioend->io_flags & (IOMAP_IOEND_UNWRITTEN |IOMAP_IOEND_SHARED ))) +if (xfs_ioend_is_append (ioend) || ioend->io_flags &IOEND_WQ_FLAGS ) ioend->io_bio.bi_end_io = xfs_end_bio;if (status)
include/linux/iomap.h +4 -1
@@ -377,13 +377,16 @@ sector_t iomap_bmap(struct address_space *mapping, sector_t bno,#define IOMAP_IOEND_BOUNDARY (1U << 2)/* is direct I/O */ #define IOMAP_IOEND_DIRECT (1U << 3) +/* is DONTCACHE I/O */ +#define IOMAP_IOEND_DONTCACHE (1U << 4)/* * Flags that if set on either ioend prevent the merge of two ioends. * (IOMAP_IOEND_BOUNDARY also prevents merges, but only one-way) */ #define IOMAP_IOEND_NOMERGE_FLAGS \ - (IOMAP_IOEND_SHARED | IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_DIRECT) + (IOMAP_IOEND_SHARED | IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_DIRECT | \ + IOMAP_IOEND_DONTCACHE)/* * Structure for writeback I/O completions.
Signature · 1 line
Jens Axboe
Quoted 17 lines · Jens Axboe
DONTCACHE I/O must have the completion punted to a workqueue, just like what is done for unwritten extents, as the completion needs task context to perform the invalidation of the folio(s). However, if writeback is started off
filemap_fdatawrite_range()offgeneric_sync()and it's an overwrite, then the DONTCACHE marking gets lost asiomap_add_to_ioend()don't look at the folio being added and no further state is passed down to help it know that this is a dropbehind/DONTCACHE write.Check if the folio being added is marked as dropbehind, and set
IOMAP_IOEND_DONTCACHEif that is the case. Then XFS can factor this into the decision making of completion context inxfs_submit_ioend(). Additionally include this ioend flag in the NOMERGE flags, to avoid mixing it with unrelated IO.This fixes extra page cache being instantiated when the write performed is an overwrite, rather than newly instantiated blocks.
Quoted 8 lines
Fixes b2cd5ae693a3 ("iomap: make buffered writes work with RWF_DONTCACHE")Signed-off-by Jens Axboe <[email protected]>---
Found this one while testing the unrelated issue of invalidation being a bit broken before 6.15 release. We need this to ensure that overwrites also prune correctly, just like unwritten extents currently do.
I wondered about the stack traces showing DONTCACHE writeback completion being handled from irq context[*] when I read the -fsdevel thread about broken DONTCACHE functionality yesterday.
[*] second trace in the failure reported in this comment:
https://lore.kernel.org/linux-fsdevel/[email protected]/
Quoted 28 lines
fs/iomap/buffered-io.c +2
-0@@ -1691,6 +1691,8 @@ static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc, ioend_flags |=IOMAP_IOEND_UNWRITTEN ;if (wpc->iomap.flags &IOMAP_F_SHARED ) ioend_flags |=IOMAP_IOEND_SHARED ; +if (folio_test_dropbehind (folio)) + ioend_flags |=IOMAP_IOEND_DONTCACHE ;if (pos == wpc->iomap.offset && (wpc->iomap.flags &IOMAP_F_BOUNDARY )) ioend_flags |=IOMAP_IOEND_BOUNDARY ;fs/xfs/xfs_aops.c +3
-0@@ -436,6 +436,9 @@ xfs_map_blocks(return 0 ; } +#define IOEND_WQ_FLAGS (IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_SHARED | \ + IOMAP_IOEND_DONTCACHE) + static int xfs_submit_ioend( struct iomap_writepage_ctx *wpc,
fs/xfs/xfs_aops.c xfs_submit_ioend( · line 469
Quoted 9 lines
fs/xfs/xfs_aops.c +1
-2@@ -460,8 +463,7 @@ xfs_submit_ioend(memalloc_nofs_restore (nofs_flag);/* send ioends that might require a transaction to the completion wq */ -if (xfs_ioend_is_append (ioend) || - (ioend->io_flags & (IOMAP_IOEND_UNWRITTEN |IOMAP_IOEND_SHARED ))) +if (xfs_ioend_is_append (ioend) || ioend->io_flags &IOEND_WQ_FLAGS ) ioend->io_bio.bi_end_io = xfs_end_bio;if (status)
IMO, this would be cleaner as a helper so that individual cases can be commented correctly, as page cache invalidation does not actually require a transaction...
Something like:
static bool
xfs_ioend_needs_wq_completion(
struct xfs_ioend *ioend)
{
/* Changing inode size requires a transaction. */
if (xfs_ioend_is_append(ioend))
return true;
/* Extent manipulation requires a transaction. */
if (ioend->io_flags & (IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_SHARED))
return true;
/* Page cache invalidation cannot be done in irq context. */
if (ioend->io_flags & IOMAP_IOEND_DONTCACHE)
return true;
return false;
}Otherwise seems fine.
-Dave.
Signature · 2 lines
Dave Chinner [email protected]
Quoted 27 lines · Dave Chinner
Quoted 17 lines · Jens Axboe
DONTCACHE I/O must have the completion punted to a workqueue, just like what is done for unwritten extents, as the completion needs task context to perform the invalidation of the folio(s). However, if writeback is started off
filemap_fdatawrite_range()offgeneric_sync()and it's an overwrite, then the DONTCACHE marking gets lost asiomap_add_to_ioend()don't look at the folio being added and no further state is passed down to help it know that this is a dropbehind/DONTCACHE write.Check if the folio being added is marked as dropbehind, and set
IOMAP_IOEND_DONTCACHEif that is the case. Then XFS can factor this into the decision making of completion context inxfs_submit_ioend(). Additionally include this ioend flag in the NOMERGE flags, to avoid mixing it with unrelated IO.This fixes extra page cache being instantiated when the write performed is an overwrite, rather than newly instantiated blocks.
Quoted 8 lines
Fixes b2cd5ae693a3 ("iomap: make buffered writes work with RWF_DONTCACHE")Signed-off-by Jens Axboe <[email protected]>---
Found this one while testing the unrelated issue of invalidation being a bit broken before 6.15 release. We need this to ensure that overwrites also prune correctly, just like unwritten extents currently do.
Quoted 7 lines
I wondered about the stack traces showing DONTCACHE writeback completion being handled from irq context[*] when I read the -fsdevel thread about broken DONTCACHE functionality yesterday.
[*] second trace in the failure reported in this comment:
Indeed, though that could've been a "normal" write and not a DONTCACHE one. But with the bug being fixed by this one, both would've gone that path...
Quoted 63 lines
Quoted 28 lines
fs/iomap/buffered-io.c +2
-0@@ -1691,6 +1691,8 @@ static int iomap_add_to_ioend(struct iomap_writepage_ctx *wpc, ioend_flags |=IOMAP_IOEND_UNWRITTEN ;if (wpc->iomap.flags &IOMAP_F_SHARED ) ioend_flags |=IOMAP_IOEND_SHARED ; +if (folio_test_dropbehind (folio)) + ioend_flags |=IOMAP_IOEND_DONTCACHE ;if (pos == wpc->iomap.offset && (wpc->iomap.flags &IOMAP_F_BOUNDARY )) ioend_flags |=IOMAP_IOEND_BOUNDARY ;fs/xfs/xfs_aops.c +3
-0@@ -436,6 +436,9 @@ xfs_map_blocks(return 0 ; } +#define IOEND_WQ_FLAGS (IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_SHARED | \ + IOMAP_IOEND_DONTCACHE) + static int xfs_submit_ioend( struct iomap_writepage_ctx *wpc,fs/xfs/xfs_aops.c xfs_submit_ioend( · line 469
Quoted 9 lines
fs/xfs/xfs_aops.c +1
-2@@ -460,8 +463,7 @@ xfs_submit_ioend(memalloc_nofs_restore (nofs_flag);/* send ioends that might require a transaction to the completion wq */ -if (xfs_ioend_is_append (ioend) || - (ioend->io_flags & (IOMAP_IOEND_UNWRITTEN |IOMAP_IOEND_SHARED ))) +if (xfs_ioend_is_append (ioend) || ioend->io_flags &IOEND_WQ_FLAGS ) ioend->io_bio.bi_end_io = xfs_end_bio;if (status)IMO, this would be cleaner as a helper so that individual cases can be commented correctly, as page cache invalidation does not actually require a transaction...
Something like:
static bool
xfs_ioend_needs_wq_completion(
structxfs_ioend*ioend){ /* Changing inode size requires a transaction. */ if (xfs_ioend_is_append(ioend)) return true; /* Extent manipulation requires a transaction. */ if (ioend->io_flags & (IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_SHARED)) return true; /* Page cache invalidation cannot be done in irq context. */ if (ioend->io_flags & IOMAP_IOEND_DONTCACHE) return true; return false; }
Quoted 1 line
Otherwise seems fine.
Yeah I like that, gets rid of the need to add the mask as well. I'll spin a v2 and add the helper.
Signature · 1 line
Jens Axboe
Quoted 17 lines · Jens Axboe
DONTCACHE I/O must have the completion punted to a workqueue, just like what is done for unwritten extents, as the completion needs task context to perform the invalidation of the folio(s). However, if writeback is started off
filemap_fdatawrite_range()offgeneric_sync()and it's an overwrite, then the DONTCACHE marking gets lost asiomap_add_to_ioend()don't look at the folio being added and no further state is passed down to help it know that this is a dropbehind/DONTCACHE write.Check if the folio being added is marked as dropbehind, and set
IOMAP_IOEND_DONTCACHEif that is the case. Then XFS can factor this into the decision making of completion context inxfs_submit_ioend(). Additionally include this ioend flag in the NOMERGE flags, to avoid mixing it with unrelated IO.This fixes extra page cache being instantiated when the write performed is an overwrite, rather than newly instantiated blocks.
Quoted 8 lines
Fixes b2cd5ae693a3 ("iomap: make buffered writes work with RWF_DONTCACHE")Signed-off-by Jens Axboe <[email protected]>---
Found this one while testing the unrelated issue of invalidation being a bit broken before 6.15 release. We need this to ensure that overwrites also prune correctly, just like unwritten extents currently do.
I guess I did report this to you a while ago when I was adding support for uncahed buffered-io to xfs_io. But I never heard back from you :(
https://lore.kernel.org/all/[email protected]/
No worries, good that we finally have this fixed.
-ritesh
Quoted 28 lines · Ritesh Harjani (IBM)
Quoted 17 lines · Jens Axboe
DONTCACHE I/O must have the completion punted to a workqueue, just like what is done for unwritten extents, as the completion needs task context to perform the invalidation of the folio(s). However, if writeback is started off
filemap_fdatawrite_range()offgeneric_sync()and it's an overwrite, then the DONTCACHE marking gets lost asiomap_add_to_ioend()don't look at the folio being added and no further state is passed down to help it know that this is a dropbehind/DONTCACHE write.Check if the folio being added is marked as dropbehind, and set
IOMAP_IOEND_DONTCACHEif that is the case. Then XFS can factor this into the decision making of completion context inxfs_submit_ioend(). Additionally include this ioend flag in the NOMERGE flags, to avoid mixing it with unrelated IO.This fixes extra page cache being instantiated when the write performed is an overwrite, rather than newly instantiated blocks.
Quoted 8 lines
Fixes b2cd5ae693a3 ("iomap: make buffered writes work with RWF_DONTCACHE")Signed-off-by Jens Axboe <[email protected]>---
Found this one while testing the unrelated issue of invalidation being a bit broken before 6.15 release. We need this to ensure that overwrites also prune correctly, just like unwritten extents currently do.
Quoted 6 lines
I guess I did report this to you a while ago when I was adding support for uncahed buffered-io to
xfs_io. But I never heard back from you :(https://lore.kernel.org/all/[email protected]/
No worries, good that we finally have this fixed.
Sorry not sure how I missed that, to be fair there were a lot of emails in various threads on this topic back at that time. Not trying to make excuses... Yes, at least the "would otherwise have been !task context" issue is resolved.
Signature · 1 line
Jens Axboe