xfs: buffer log item optimisations
cover xfs: buffer log item optimisations
Hi folks,
A couple of optimisations and a bug fix that I don't think we could trigger.
The bug fix was that we weren't passing the segment offset into the buffer log item sizing calculation, so we weren't calculating when we spanned discontiguous pages in the buffers correctly. I don't think this ever mattered, because all buffers larger than a single page are vmapped (so a contiguous virtual address range) and the only direct mapped buffers we have are inode cluster buffers and they never span discontiguous extents. Hence while the code is clearly wrong, we never actually trigger the situation where it results in an incorrect calculation. However, changing the way we calculate the size of the dirty regions is difficult if we don't do this calc the same way as the formatting code, so fix it.
The first optimisation is simply a mechanism to reduce the amount of allocation and freeing overhead on the buffer item shadow buffer as we increase the amount of the buffer that is dirty as we relog it.
The last optimisation is (finally) addressing the overhead of bitmap based dirty tracking of the buffer log item. We walk a bit at a time, calling xfs_buf_offset() at least twice for each bit in both the size and the formatting code to see if the region crosses a discontiguity in the buffer address space. This is expensive. The log recovery code uses contiguous bit range detection to do the same thing, so I've updated the logging code to operate on contiguous bit ranges and only fall back to bit-by-bit checking in the rare case that a contiguous dirty range spans an address space discontiguity and hence has to be split into multiple regions to copy it into the log.
This enables big performance improvements when using large directory block sizes.
Cheers,
Dave.
PATCH 1/3 xfs: reduce buffer log item shadow allocations
From: Dave Chinner <[email protected]>
When we modify btrees repeatedly, we regularly increase the size of the logged region by a single chunk at a time (per transaction commit). This results in the CIL formatting code having to reallocate the log vector buffer every time the buffer dirty region grows. Hence over a typical 4kB btree buffer, we might grow the log vector 4096/128 = 32x over a short period where we repeatedly add or remove records to/from the buffer over a series of running transaction. This means we are doing 32 memory allocations and frees over this time during a performance critical path in the journal.
The amount of space tracked in the CIL for the object is calculated during the ->iop_format() call for the buffer log item, but the buffer memory allocated for it is calculated by the ->iop_size() call. The size callout determines the size of the buffer, the format call determines the space used in the buffer.
Hence we can oversize the buffer space required in the size calculation without impacting the amount of space used and accounted to the CIL for the changes being logged. This allows us to reduce the number of allocations by rounding up the buffer size to allow for future growth. This can safe a substantial amount of CPU time in this path:
- 46.52% 2.02% [kernel] [k] xfs_log_commit_cil
- 44.49% xfs_log_commit_cil
- 30.78% _raw_spin_lock
- 30.75% do_raw_spin_lock
30.27% __pv_queued_spin_lock_slowpath
(oh, ouch!)
....
- 1.05% kmem_alloc_large
- 1.02% kmem_alloc
0.94% __kmalloc
This overhead here us what this patch is aimed at. After:
- 0.76% kmem_alloc_large
- 0.75% kmem_alloc
0.70% __kmalloc
--- fs/xfs/xfs_buf_item.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)
fs/xfs/xfs_buf_item.c +11 -2
@@ -142,6 +142,7 @@ xfs_buf_item_size( {struct xfs_buf_log_item *bip =BUF_ITEM (lip);int i; +int bytes;ASSERT (atomic_read (&bip->bli_refcount) >0 );if (bip->bli_flags &XFS_BLI_STALE ) {
@@ -173,7 +174,7 @@ xfs_buf_item_size( }/* -* the vector count is based on the number of buffer vectors we have +* The vector count is based on the number of buffer vectors we have * dirty bits in. This will only be greater than one when we have a * compound buffer with more than one segment dirty. Hence for compound * buffers we need to track which segment the dirty bits correspond to,
@@ -181,10 +182,18 @@ xfs_buf_item_size(* count for the extra buf log format structure that will need to be * written. */ + bytes =0 ;for (i =0 ; i < bip->bli_format_count; i++) {xfs_buf_item_size_segment (bip, &bip->bli_formats[i], - nvecs, nbytes); + nvecs, &bytes); } + +/* +* Round up the buffer size required to minimise the number of memory +* allocations that need to be done as this item grows when relogged by +* repeated modifications. +*/ + *nbytes =round_up (bytes,512 );trace_xfs_buf_item_size (bip); }
Signature · 1 line
2.28.0
PATCH 2/3 xfs: xfs_buf_item_size_segment() needs to pass segment offset
In reply to Dave ChinnerFrom: Dave Chinner <[email protected]>
Otherwise it doesn't correctly calculate the number of vectors in a logged buffer that has a contiguous map that gets split into multiple regions because the range spans discontigous memory.
Probably never been hit in practice - we don't log contiguous ranges on unmapped buffers (inode clusters).
--- fs/xfs/xfs_buf_item.c | 38 +++++++++++++++++++------------------- 1 file changed, 19 insertions(+), 19 deletions(-)
fs/xfs/xfs_buf_item.c +19 -19
@@ -55,6 +55,18 @@ xfs_buf_log_format_size( (blfp->blf_map_size *sizeof (blfp->blf_data_map[0 ])); } +static inline bool +xfs_buf_item_straddle ( +struct xfs_buf *bp, +uint offset, +int next_bit, +int last_bit) +{ +return xfs_buf_offset (bp, offset + (next_bit <<XFS_BLF_SHIFT )) != + (xfs_buf_offset (bp, offset + (last_bit <<XFS_BLF_SHIFT )) + +XFS_BLF_CHUNK ); +} +/* * Return the number of log iovecs and space needed to log the given buf log * item segment.
@@ -67,6 +79,7 @@ STATIC voidxfs_buf_item_size_segment ( struct xfs_buf_log_item *bip, struct xfs_buf_log_format *blfp, + uint offset, int *nvecs, int *nbytes) {
@@ -101,12 +114,8 @@ xfs_buf_item_size_segment(*/ if (next_bit ==-1 ) {break ; - }else if (next_bit != last_bit +1 ) { - last_bit = next_bit; - (*nvecs)++; - }else if (xfs_buf_offset (bp, next_bit *XFS_BLF_CHUNK ) != - (xfs_buf_offset (bp, last_bit *XFS_BLF_CHUNK ) + -XFS_BLF_CHUNK )) { + }else if (next_bit != last_bit +1 || +xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { last_bit = next_bit; (*nvecs)++; }else {
@@ -141,8 +150,10 @@ xfs_buf_item_size( int *nbytes) {struct xfs_buf_log_item *bip =BUF_ITEM (lip); +struct xfs_buf *bp = bip->bli_buf;int i;int bytes; +uint offset =0 ;ASSERT (atomic_read (&bip->bli_refcount) >0 );if (bip->bli_flags &XFS_BLI_STALE ) {
@@ -184,8 +195,9 @@ xfs_buf_item_size(*/ bytes =0 ;for (i =0 ; i < bip->bli_format_count; i++) { -xfs_buf_item_size_segment (bip, &bip->bli_formats[i], +xfs_buf_item_size_segment (bip, &bip->bli_formats[i], offset, nvecs, &bytes); + offset +=BBTOB (bp->b_maps[i].bm_len); }/*
@@ -212,18 +224,6 @@ xfs_buf_item_copy_iovec( nbits * XFS_BLF_CHUNK); } -static inline bool -xfs_buf_item_straddle ( -struct xfs_buf *bp, -uint offset, -int next_bit, -int last_bit) -{ -return xfs_buf_offset (bp, offset + (next_bit <<XFS_BLF_SHIFT )) != - (xfs_buf_offset (bp, offset + (last_bit <<XFS_BLF_SHIFT )) + -XFS_BLF_CHUNK ); -} - static void xfs_buf_item_format_segment( struct xfs_buf_log_item *bip,
Signature · 1 line
2.28.0
PATCH 3/3 xfs: optimise xfs_buf_item_size/format for contiguous regions
In reply to Dave ChinnerFrom: Dave Chinner <[email protected]>
We process the buf_log_item bitmap one set bit at a time with xfs_next_bit() so we can detect if a region crosses a memcpy discontinuity in the buffer data address. This has massive overhead on large buffers (e.g. 64k directory blocks) because we do a lot of unnecessary checks and xfs_buf_offset() calls.
For example, 16-way concurrent create workload on debug kernel running CPU bound has this at the top of the profile at ~120k create/s on 64kb directory block size:
20.66% [kernel] [k] xfs_dir3_leaf_check_int 7.10% [kernel] [k] memcpy 6.22% [kernel] [k] xfs_next_bit 3.55% [kernel] [k] xfs_buf_offset 3.53% [kernel] [k] xfs_buf_item_format 3.34% [kernel] [k] __pv_queued_spin_lock_slowpath 3.04% [kernel] [k] do_raw_spin_lock 2.84% [kernel] [k] xfs_buf_item_size_segment.isra.0 2.31% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock 1.36% [kernel] [k] xfs_log_commit_cil
(debug checks hurt large blocks)
The only buffers with discontinuities in the data address are unmapped buffers, and they are only used for inode cluster buffers and only for logging unlinked pointers. IOWs, it is -rare- that we even need to detect a discontinuity in the buffer item formatting code.
Optimise all this by using xfs_contig_bits() to find the size of the contiguous regions, then test for a discontiunity inside it. If we find one, do the slow "bit at a time" method we do now. If we don't, then just copy the entire contiguous range in one go.
Profile now looks like:
25.26% [kernel] [k] xfs_dir3_leaf_check_int 9.25% [kernel] [k] memcpy 5.01% [kernel] [k] __pv_queued_spin_lock_slowpath 2.84% [kernel] [k] do_raw_spin_lock 2.22% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock 1.88% [kernel] [k] xfs_buf_find 1.53% [kernel] [k] memmove 1.47% [kernel] [k] xfs_log_commit_cil
....
0.34% [kernel] [k] xfs_buf_item_format
....
0.21% [kernel] [k] xfs_buf_offset
....
0.16% [kernel] [k] xfs_contig_bits
....
0.13% [kernel] [k] xfs_buf_item_size_segment.isra.0
So the bit scanning over for the dirty region tracking for the buffer log items is basically gone. Debug overhead hurts even more now...
Perf comparison
dir block creates unlink
size (kb) time rate time
Original 4 4m08s 220k 5m13s
Original 64 7m21s 115k 13m25s
Patched 4 3m59s 230k 5m03s
Patched 64 6m23s 143k 12m33s
--- fs/xfs/xfs_buf_item.c | 102 +++++++++++++++++++++++++++++++++++------- 1 file changed, 87 insertions(+), 15 deletions(-)
fs/xfs/xfs_buf_item.c +87 -15
@@ -59,12 +59,18 @@ static inline boolxfs_buf_item_straddle ( struct xfs_buf *bp, uint offset, - int next_bit, - int last_bit) + int first_bit, + int nbits) { - return xfs_buf_offset(bp, offset + (next_bit << XFS_BLF_SHIFT)) != - (xfs_buf_offset(bp, offset + (last_bit << XFS_BLF_SHIFT)) + -XFS_BLF_CHUNK ); + void *first, *last; + + first =xfs_buf_offset (bp, offset + (first_bit <<XFS_BLF_SHIFT )); + last =xfs_buf_offset (bp, + offset + ((first_bit + nbits) <<XFS_BLF_SHIFT )); + +if (last - first != nbits *XFS_BLF_CHUNK ) +return true; +return false; }/*
@@ -84,20 +90,51 @@ xfs_buf_item_size_segment( int *nbytes) {struct xfs_buf *bp = bip->bli_buf; +int first_bit; +int nbits;int next_bit;int last_bit; - last_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size,0 ); -if (last_bit ==-1 ) + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size,0 ); +if (first_bit ==-1 )return ; -/* -* initial count for a dirty buffer is 2 vectors - the format structure -* and the first dirty region. -*/ - *nvecs +=2 ; - *nbytes +=xfs_buf_log_format_size (blfp) +XFS_BLF_CHUNK ; + (*nvecs)++; + *nbytes +=xfs_buf_log_format_size (blfp); + +do { + nbits =xfs_contig_bits (blfp->blf_data_map, + blfp->blf_map_size, first_bit); +ASSERT (nbits >0 ); + +/* +* Straddling a page is rare because we don't log contiguous +* chunks of unmapped buffers anywhere. +*/ +if (nbits >1 && +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) + goto slow_scan; + + (*nvecs)++; + *nbytes += nbits *XFS_BLF_CHUNK ; + +/* +* This takes the bit number to start looking from and +* returns the next set bit from there. It returns -1 +* if there are no more bits set or the start bit is +* beyond the end of the bitmap. +*/ + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size, + (uint )first_bit + nbits +1 ); + }while (first_bit !=-1 ); +return ; + +slow_scan: +/* Count the first bit we jumped out of the above loop from */ + (*nvecs)++; + *nbytes +=XFS_BLF_CHUNK ; + last_bit = first_bit;while (last_bit !=-1 ) {/* * This takes the bit number to start looking from and
@@ -115,11 +152,14 @@ xfs_buf_item_size_segment(if (next_bit ==-1 ) {break ; }else if (next_bit != last_bit +1 || -xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) { last_bit = next_bit; + first_bit = next_bit; (*nvecs)++; + nbits =1 ; }else { last_bit++; + nbits++; } *nbytes +=XFS_BLF_CHUNK ; }
@@ -276,6 +316,38 @@ xfs_buf_item_format_segment(/* * Fill in an iovec for each set of contiguous chunks. */ +do { +ASSERT (first_bit >=0 ); + nbits =xfs_contig_bits (blfp->blf_data_map, + blfp->blf_map_size, first_bit); +ASSERT (nbits >0 ); + +/* +* Straddling a page is rare because we don't log contiguous +* chunks of unmapped buffers anywhere. +*/ +if (nbits >1 && +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) + goto slow_scan; + +xfs_buf_item_copy_iovec (lv, vecp, bp, offset, + first_bit, nbits); + blfp->blf_size++; + +/* +* This takes the bit number to start looking from and +* returns the next set bit from there. It returns -1 +* if there are no more bits set or the start bit is +* beyond the end of the bitmap. +*/ + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size, + (uint )first_bit + nbits +1 ); + }while (first_bit !=-1 ); + +return ; + +slow_scan: +ASSERT (bp->b_addr ==NULL ); last_bit = first_bit; nbits =1 ;for (;;) {
@@ -300,7 +372,7 @@ xfs_buf_item_format_segment( blfp->blf_size++;break ; }else if (next_bit != last_bit +1 || -xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) {xfs_buf_item_copy_iovec (lv, vecp, bp, offset, first_bit, nbits); blfp->blf_size++;
Signature · 1 line
2.28.0
Re: [PATCH 1/3] xfs: reduce buffer log item shadow allocations
In reply to Dave ChinnerQuoted 37 lines · Dave Chinner
From: Dave Chinner <[email protected]>
When we modify btrees repeatedly, we regularly increase the size of the logged region by a single chunk at a time (per transaction commit). This results in the CIL formatting code having to reallocate the log vector buffer every time the buffer dirty region grows. Hence over a typical 4kB btree buffer, we might grow the log vector 4096/128 = 32x over a short period where we repeatedly add or remove records to/from the buffer over a series of running transaction. This means we are doing 32 memory allocations and frees over this time during a performance critical path in the journal.
The amount of space tracked in the CIL for the object is calculated during the ->iop_format() call for the buffer log item, but the buffer memory allocated for it is calculated by the ->iop_size() call. The size callout determines the size of the buffer, the format call determines the space used in the buffer.
Hence we can oversize the buffer space required in the size calculation without impacting the amount of space used and accounted to the CIL for the changes being logged. This allows us to reduce the number of allocations by rounding up the buffer size to allow for future growth. This can safe a substantial amount of CPU time in this path:
- 46.52% 2.02% [kernel] [k]
xfs_log_commit_cil
- 44.49%xfs_log_commit_cil
- 30.78%_raw_spin_lock
- 30.75%do_raw_spin_lock
30.27%__pv_queued_spin_lock_slowpath(oh, ouch!)
....
- 1.05%kmem_alloc_large
- 1.02%kmem_alloc
0.94%__kmalloc
Quoted 8 lines
This overhead here us what this patch is aimed at. After:
- 0.76%
kmem_alloc_large
- 0.75%kmem_alloc
0.70%__kmallocSigned-off-by Dave Chinner <[email protected]>Reviewed-by Chandan Babu R <[email protected]>
Any particular reason for 512? It looks like you simply picked an arbitrary power of 2, but was there a particular target in mind? i.e. we never need to realloc for the usual 4k filesystem?
--D
Quoted 48 lines
--- fs/xfs/xfs_buf_item.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)fs/xfs/xfs_buf_item.c +11
-2@@ -142,6 +142,7 @@ xfs_buf_item_size( {struct xfs_buf_log_item *bip =BUF_ITEM (lip);int i; +int bytes;ASSERT (atomic_read (&bip->bli_refcount) >0 );if (bip->bli_flags &XFS_BLI_STALE ) {@@ -173,7 +174,7 @@ xfs_buf_item_size( }/* -* the vector count is based on the number of buffer vectors we have +* The vector count is based on the number of buffer vectors we have * dirty bits in. This will only be greater than one when we have a * compound buffer with more than one segment dirty. Hence for compound * buffers we need to track which segment the dirty bits correspond to, @@ -181,10 +182,18 @@ xfs_buf_item_size(* count for the extra buf log format structure that will need to be * written. */ + bytes =0 ;for (i =0 ; i < bip->bli_format_count; i++) {xfs_buf_item_size_segment (bip, &bip->bli_formats[i], - nvecs, nbytes); + nvecs, &bytes); } + +/* +* Round up the buffer size required to minimise the number of memory +* allocations that need to be done as this item grows when relogged by +* repeated modifications. +*/ + *nbytes =round_up (bytes,512 );trace_xfs_buf_item_size (bip); }Signature · 1 line
2.28.0
Re: [PATCH 2/3] xfs: xfs_buf_item_size_segment() needs to pass segment offset
In reply to Dave ChinnerQuoted 10 lines · Dave Chinner
From: Dave Chinner <[email protected]>
Otherwise it doesn't correctly calculate the number of vectors in a logged buffer that has a contiguous map that gets split into multiple regions because the range spans discontigous memory.
Probably never been hit in practice - we don't log contiguous ranges on unmapped buffers (inode clusters).
Signed-off-by Dave Chinner <[email protected]>
Subtle.
--D
Quoted 94 lines
--- fs/xfs/xfs_buf_item.c | 38 +++++++++++++++++++------------------- 1 file changed, 19 insertions(+), 19 deletions(-)fs/xfs/xfs_buf_item.c +19
-19@@ -55,6 +55,18 @@ xfs_buf_log_format_size( (blfp->blf_map_size *sizeof (blfp->blf_data_map[0 ])); } +static inline bool +xfs_buf_item_straddle ( +struct xfs_buf *bp, +uint offset, +int next_bit, +int last_bit) +{ +return xfs_buf_offset (bp, offset + (next_bit <<XFS_BLF_SHIFT )) != + (xfs_buf_offset (bp, offset + (last_bit <<XFS_BLF_SHIFT )) + +XFS_BLF_CHUNK ); +} +/* * Return the number of log iovecs and space needed to log the given buf log * item segment. @@ -67,6 +79,7 @@ STATIC voidxfs_buf_item_size_segment ( struct xfs_buf_log_item *bip, struct xfs_buf_log_format *blfp, + uint offset, int *nvecs, int *nbytes) {@@ -101,12 +114,8 @@ xfs_buf_item_size_segment(*/ if (next_bit ==-1 ) {break ; - }else if (next_bit != last_bit +1 ) { - last_bit = next_bit; - (*nvecs)++; - }else if (xfs_buf_offset (bp, next_bit *XFS_BLF_CHUNK ) != - (xfs_buf_offset (bp, last_bit *XFS_BLF_CHUNK ) + -XFS_BLF_CHUNK )) { + }else if (next_bit != last_bit +1 || +xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { last_bit = next_bit; (*nvecs)++; }else {@@ -141,8 +150,10 @@ xfs_buf_item_size( int *nbytes) {struct xfs_buf_log_item *bip =BUF_ITEM (lip); +struct xfs_buf *bp = bip->bli_buf;int i;int bytes; +uint offset =0 ;ASSERT (atomic_read (&bip->bli_refcount) >0 );if (bip->bli_flags &XFS_BLI_STALE ) {@@ -184,8 +195,9 @@ xfs_buf_item_size(*/ bytes =0 ;for (i =0 ; i < bip->bli_format_count; i++) { -xfs_buf_item_size_segment (bip, &bip->bli_formats[i], +xfs_buf_item_size_segment (bip, &bip->bli_formats[i], offset, nvecs, &bytes); + offset +=BBTOB (bp->b_maps[i].bm_len); }/* @@ -212,18 +224,6 @@ xfs_buf_item_copy_iovec( nbits * XFS_BLF_CHUNK); } -static inline bool -xfs_buf_item_straddle ( -struct xfs_buf *bp, -uint offset, -int next_bit, -int last_bit) -{ -return xfs_buf_offset (bp, offset + (next_bit <<XFS_BLF_SHIFT )) != - (xfs_buf_offset (bp, offset + (last_bit <<XFS_BLF_SHIFT )) + -XFS_BLF_CHUNK ); -} - static void xfs_buf_item_format_segment( struct xfs_buf_log_item *bip,Signature · 1 line
2.28.0
Re: [PATCH 3/3] xfs: optimise xfs_buf_item_size/format for contiguous regions
In reply to Dave ChinnerQuoted 64 lines · Dave Chinner
From: Dave Chinner <[email protected]>
We process the
buf_log_itembitmap one set bit at a time withxfs_next_bit()so we can detect if a region crosses a memcpy discontinuity in the buffer data address. This has massive overhead on large buffers (e.g. 64k directory blocks) because we do a lot of unnecessary checks andxfs_buf_offset()calls.For example, 16-way concurrent create workload on debug kernel running CPU bound has this at the top of the profile at ~120k create/s on 64kb directory block size:
20.66% [kernel] [k] xfs_dir3_leaf_check_int 7.10% [kernel] [k] memcpy 6.22% [kernel] [k] xfs_next_bit 3.55% [kernel] [k] xfs_buf_offset 3.53% [kernel] [k] xfs_buf_item_format 3.34% [kernel] [k] __pv_queued_spin_lock_slowpath 3.04% [kernel] [k] do_raw_spin_lock 2.84% [kernel] [k] xfs_buf_item_size_segment.isra.0 2.31% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock 1.36% [kernel] [k] xfs_log_commit_cil(debug checks hurt large blocks)
The only buffers with discontinuities in the data address are unmapped buffers, and they are only used for inode cluster buffers and only for logging unlinked pointers. IOWs, it is -rare- that we even need to detect a discontinuity in the buffer item formatting code.
Optimise all this by using
xfs_contig_bits()to find the size of the contiguous regions, then test for a discontiunity inside it. If we find one, do the slow "bit at a time" method we do now. If we don't, then just copy the entire contiguous range in one go.Profile now looks like:
25.26% [kernel] [k] xfs_dir3_leaf_check_int 9.25% [kernel] [k] memcpy 5.01% [kernel] [k] __pv_queued_spin_lock_slowpath 2.84% [kernel] [k] do_raw_spin_lock 2.22% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock 1.88% [kernel] [k] xfs_buf_find 1.53% [kernel] [k] memmove 1.47% [kernel] [k] xfs_log_commit_cil....
0.34% [kernel] [k]xfs_buf_item_format
....
0.21% [kernel] [k]xfs_buf_offset
....
0.16% [kernel] [k]xfs_contig_bits
....
0.13% [kernel] [k] xfs_buf_item_size_segment.isra.0So the bit scanning over for the dirty region tracking for the buffer log items is basically gone. Debug overhead hurts even more now...
Perf comparison
dir block creates unlink
size (kb) time rate time
Quoted 7 lines
Original 4 4m08s 220k 5m13s
Original 64 7m21s 115k 13m25s
Patched 4 3m59s 230k 5m03s
Patched 64 6m23s 143k 12m33sSigned-off-by Dave Chinner <[email protected]>
Seems straightforward enough...
--D
Quoted 159 lines
--- fs/xfs/xfs_buf_item.c | 102 +++++++++++++++++++++++++++++++++++------- 1 file changed, 87 insertions(+), 15 deletions(-)fs/xfs/xfs_buf_item.c +87
-15@@ -59,12 +59,18 @@ static inline boolxfs_buf_item_straddle ( struct xfs_buf *bp, uint offset, - int next_bit, - int last_bit) + int first_bit, + int nbits) { - return xfs_buf_offset(bp, offset + (next_bit << XFS_BLF_SHIFT)) != - (xfs_buf_offset(bp, offset + (last_bit << XFS_BLF_SHIFT)) + -XFS_BLF_CHUNK ); + void *first, *last; + + first =xfs_buf_offset (bp, offset + (first_bit <<XFS_BLF_SHIFT )); + last =xfs_buf_offset (bp, + offset + ((first_bit + nbits) <<XFS_BLF_SHIFT )); + +if (last - first != nbits *XFS_BLF_CHUNK ) +return true; +return false; }/* @@ -84,20 +90,51 @@ xfs_buf_item_size_segment( int *nbytes) {struct xfs_buf *bp = bip->bli_buf; +int first_bit; +int nbits;int next_bit;int last_bit; - last_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size,0 ); -if (last_bit ==-1 ) + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size,0 ); +if (first_bit ==-1 )return ; -/* -* initial count for a dirty buffer is 2 vectors - the format structure -* and the first dirty region. -*/ - *nvecs +=2 ; - *nbytes +=xfs_buf_log_format_size (blfp) +XFS_BLF_CHUNK ; + (*nvecs)++; + *nbytes +=xfs_buf_log_format_size (blfp); + +do { + nbits =xfs_contig_bits (blfp->blf_data_map, + blfp->blf_map_size, first_bit); +ASSERT (nbits >0 ); + +/* +* Straddling a page is rare because we don't log contiguous +* chunks of unmapped buffers anywhere. +*/ +if (nbits >1 && +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) + goto slow_scan; + + (*nvecs)++; + *nbytes += nbits *XFS_BLF_CHUNK ; + +/* +* This takes the bit number to start looking from and +* returns the next set bit from there. It returns -1 +* if there are no more bits set or the start bit is +* beyond the end of the bitmap. +*/ + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size, + (uint )first_bit + nbits +1 ); + }while (first_bit !=-1 ); +return ; + +slow_scan: +/* Count the first bit we jumped out of the above loop from */ + (*nvecs)++; + *nbytes +=XFS_BLF_CHUNK ; + last_bit = first_bit;while (last_bit !=-1 ) {/* * This takes the bit number to start looking from and @@ -115,11 +152,14 @@ xfs_buf_item_size_segment(if (next_bit ==-1 ) {break ; }else if (next_bit != last_bit +1 || -xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) { last_bit = next_bit; + first_bit = next_bit; (*nvecs)++; + nbits =1 ; }else { last_bit++; + nbits++; } *nbytes +=XFS_BLF_CHUNK ; }@@ -276,6 +316,38 @@ xfs_buf_item_format_segment(/* * Fill in an iovec for each set of contiguous chunks. */ +do { +ASSERT (first_bit >=0 ); + nbits =xfs_contig_bits (blfp->blf_data_map, + blfp->blf_map_size, first_bit); +ASSERT (nbits >0 ); + +/* +* Straddling a page is rare because we don't log contiguous +* chunks of unmapped buffers anywhere. +*/ +if (nbits >1 && +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) + goto slow_scan; + +xfs_buf_item_copy_iovec (lv, vecp, bp, offset, + first_bit, nbits); + blfp->blf_size++; + +/* +* This takes the bit number to start looking from and +* returns the next set bit from there. It returns -1 +* if there are no more bits set or the start bit is +* beyond the end of the bitmap. +*/ + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size, + (uint )first_bit + nbits +1 ); + }while (first_bit !=-1 ); + +return ; + +slow_scan: +ASSERT (bp->b_addr ==NULL ); last_bit = first_bit; nbits =1 ;for (;;) {@@ -300,7 +372,7 @@ xfs_buf_item_format_segment( blfp->blf_size++;break ; }else if (next_bit != last_bit +1 || -xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) {xfs_buf_item_copy_iovec (lv, vecp, bp, offset, first_bit, nbits); blfp->blf_size++;Signature · 1 line
2.28.0
Re: [PATCH 1/3] xfs: reduce buffer log item shadow allocations
In reply to Darrick J. WongQuoted 44 lines · Darrick J. Wong
Quoted 43 lines · Dave Chinner
From: Dave Chinner <[email protected]>
When we modify btrees repeatedly, we regularly increase the size of the logged region by a single chunk at a time (per transaction commit). This results in the CIL formatting code having to reallocate the log vector buffer every time the buffer dirty region grows. Hence over a typical 4kB btree buffer, we might grow the log vector 4096/128 = 32x over a short period where we repeatedly add or remove records to/from the buffer over a series of running transaction. This means we are doing 32 memory allocations and frees over this time during a performance critical path in the journal.
The amount of space tracked in the CIL for the object is calculated during the ->iop_format() call for the buffer log item, but the buffer memory allocated for it is calculated by the ->iop_size() call. The size callout determines the size of the buffer, the format call determines the space used in the buffer.
Hence we can oversize the buffer space required in the size calculation without impacting the amount of space used and accounted to the CIL for the changes being logged. This allows us to reduce the number of allocations by rounding up the buffer size to allow for future growth. This can safe a substantial amount of CPU time in this path:
- 46.52% 2.02% [kernel] [k]
xfs_log_commit_cil
- 44.49%xfs_log_commit_cil
- 30.78%_raw_spin_lock
- 30.75%do_raw_spin_lock
30.27%__pv_queued_spin_lock_slowpath(oh, ouch!)
....
- 1.05%kmem_alloc_large
- 1.02%kmem_alloc
0.94%__kmallocThis overhead here us what this patch is aimed at. After:
- 0.76%
kmem_alloc_large
- 0.75%kmem_alloc
0.70%__kmalloc
Quoted 6 lines
Quoted 2 lines
Signed-off-by Dave Chinner <[email protected]>Reviewed-by Chandan Babu R <[email protected]>Any particular reason for 512? It looks like you simply picked an arbitrary power of 2, but was there a particular target in mind? i.e. we never need to realloc for the usual 4k filesystem?
It is based on the bitmap chunk size being 128 bytes and that random directory entry updates almost never require more than 3-4 128 byte regions to be logged in the directory block.
The other observation is for per-ag btrees. When we are inserting into a new btree block, we'll pack it from the front. Hence the first few records land in the first 128 bytes so we log only 128 bytes, the next 8-16 records land in the second region so now we log 256 bytes. And so on. If we are doing random updates, it will only allocate every 4 random 128 byte regions that are dirtied instead of every single one.
Any larger than this and I noticed an increase in memory footprint in my scalability workloads. Any less than this and I didn't really see any significant benefit to CPU usage.
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]
The whole series looks good to me:
Re: [PATCH 1/3] xfs: reduce buffer log item shadow allocations
In reply to Dave ChinnerQuoted 39 lines · Dave Chinner
From: Dave Chinner <[email protected]>
When we modify btrees repeatedly, we regularly increase the size of the logged region by a single chunk at a time (per transaction commit). This results in the CIL formatting code having to reallocate the log vector buffer every time the buffer dirty region grows. Hence over a typical 4kB btree buffer, we might grow the log vector 4096/128 = 32x over a short period where we repeatedly add or remove records to/from the buffer over a series of running transaction. This means we are doing 32 memory allocations and frees over this time during a performance critical path in the journal.
The amount of space tracked in the CIL for the object is calculated during the ->iop_format() call for the buffer log item, but the buffer memory allocated for it is calculated by the ->iop_size() call. The size callout determines the size of the buffer, the format call determines the space used in the buffer.
Hence we can oversize the buffer space required in the size calculation without impacting the amount of space used and accounted to the CIL for the changes being logged. This allows us to reduce the number of allocations by rounding up the buffer size to allow for future growth. This can safe a substantial amount of CPU time in this path:
- 46.52% 2.02% [kernel] [k]
xfs_log_commit_cil
- 44.49%xfs_log_commit_cil
- 30.78%_raw_spin_lock
- 30.75%do_raw_spin_lock
30.27%__pv_queued_spin_lock_slowpath(oh, ouch!)
....
- 1.05%kmem_alloc_large
- 1.02%kmem_alloc
0.94%__kmallocThis overhead here us what this patch is aimed at. After:
Quoted 7 lines
- 0.76%
kmem_alloc_large
- 0.75%kmem_alloc
0.70%__kmallocSigned-off-by Dave Chinner <[email protected]>Reviewed-by Chandan Babu R <[email protected]>---
Quoted 47 lines
fs/xfs/xfs_buf_item.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)fs/xfs/xfs_buf_item.c +11
-2@@ -142,6 +142,7 @@ xfs_buf_item_size( {struct xfs_buf_log_item *bip =BUF_ITEM (lip);int i; +int bytes;ASSERT (atomic_read (&bip->bli_refcount) >0 );if (bip->bli_flags &XFS_BLI_STALE ) {@@ -173,7 +174,7 @@ xfs_buf_item_size( }/* -* the vector count is based on the number of buffer vectors we have +* The vector count is based on the number of buffer vectors we have * dirty bits in. This will only be greater than one when we have a * compound buffer with more than one segment dirty. Hence for compound * buffers we need to track which segment the dirty bits correspond to, @@ -181,10 +182,18 @@ xfs_buf_item_size(* count for the extra buf log format structure that will need to be * written. */ + bytes =0 ;for (i =0 ; i < bip->bli_format_count; i++) {xfs_buf_item_size_segment (bip, &bip->bli_formats[i], - nvecs, nbytes); + nvecs, &bytes); } + +/* +* Round up the buffer size required to minimise the number of memory +* allocations that need to be done as this item grows when relogged by +* repeated modifications. +*/ + *nbytes =round_up (bytes,512 );trace_xfs_buf_item_size (bip); }Signature · 1 line
2.28.0
Re: [PATCH 3/3] xfs: optimise xfs_buf_item_size/format for contiguous regions
In reply to Dave ChinnerQuoted 79 lines · Dave Chinner
From: Dave Chinner <[email protected]>
We process the
buf_log_itembitmap one set bit at a time withxfs_next_bit()so we can detect if a region crosses a memcpy discontinuity in the buffer data address. This has massive overhead on large buffers (e.g. 64k directory blocks) because we do a lot of unnecessary checks andxfs_buf_offset()calls.For example, 16-way concurrent create workload on debug kernel running CPU bound has this at the top of the profile at ~120k create/s on 64kb directory block size:
20.66% [kernel] [k] xfs_dir3_leaf_check_int 7.10% [kernel] [k] memcpy 6.22% [kernel] [k] xfs_next_bit 3.55% [kernel] [k] xfs_buf_offset 3.53% [kernel] [k] xfs_buf_item_format 3.34% [kernel] [k] __pv_queued_spin_lock_slowpath 3.04% [kernel] [k] do_raw_spin_lock 2.84% [kernel] [k] xfs_buf_item_size_segment.isra.0 2.31% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock 1.36% [kernel] [k] xfs_log_commit_cil(debug checks hurt large blocks)
The only buffers with discontinuities in the data address are unmapped buffers, and they are only used for inode cluster buffers and only for logging unlinked pointers. IOWs, it is -rare- that we even need to detect a discontinuity in the buffer item formatting code.
Optimise all this by using
xfs_contig_bits()to find the size of the contiguous regions, then test for a discontiunity inside it. If we find one, do the slow "bit at a time" method we do now. If we don't, then just copy the entire contiguous range in one go.Profile now looks like:
25.26% [kernel] [k] xfs_dir3_leaf_check_int 9.25% [kernel] [k] memcpy 5.01% [kernel] [k] __pv_queued_spin_lock_slowpath 2.84% [kernel] [k] do_raw_spin_lock 2.22% [kernel] [k] __raw_callee_save___pv_queued_spin_unlock 1.88% [kernel] [k] xfs_buf_find 1.53% [kernel] [k] memmove 1.47% [kernel] [k] xfs_log_commit_cil....
0.34% [kernel] [k]xfs_buf_item_format
....
0.21% [kernel] [k]xfs_buf_offset
....
0.16% [kernel] [k]xfs_contig_bits
....
0.13% [kernel] [k] xfs_buf_item_size_segment.isra.0So the bit scanning over for the dirty region tracking for the buffer log items is basically gone. Debug overhead hurts even more now...
Perf comparison
dir block creates unlink
size (kb) time rate timeOriginal 4 4m08s 220k 5m13s
Original 64 7m21s 115k 13m25s
Patched 4 3m59s 230k 5m03s
Patched 64 6m23s 143k 12m33sSigned-off-by Dave Chinner <[email protected]>--- fs/xfs/xfs_buf_item.c | 102 +++++++++++++++++++++++++++++++++++------- 1 file changed, 87 insertions(+), 15 deletions(-)fs/xfs/xfs_buf_item.c +0
-0
...
Quoted 38 lines
fs/xfs/xfs_buf_item.c +21
-8@@ -84,20 +90,51 @@ xfs_buf_item_size_segment( int *nbytes) {struct xfs_buf *bp = bip->bli_buf; +int first_bit; +int nbits;int next_bit;int last_bit; - last_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size,0 ); -if (last_bit ==-1 ) + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size,0 ); +if (first_bit ==-1 )return ; -/* -* initial count for a dirty buffer is 2 vectors - the format structure -* and the first dirty region. -*/ - *nvecs +=2 ; - *nbytes +=xfs_buf_log_format_size (blfp) +XFS_BLF_CHUNK ; + (*nvecs)++; + *nbytes +=xfs_buf_log_format_size (blfp); + + do { + nbits = xfs_contig_bits(blfp->blf_data_map, + blfp->blf_map_size, first_bit); + ASSERT(nbits > 0); + + /* + * Straddling a page is rare because we don't log contiguous + * chunks of unmapped buffers anywhere. + */ + if (nbits > 1 && + xfs_buf_item_straddle(bp, offset, first_bit, nbits)) + goto slow_scan; + + (*nvecs)++;
fs/xfs/xfs_buf_item.c xfs_buf_item_size_segment( · line 128
Quoted 10 lines
fs/xfs/xfs_buf_item.c +10
-0+ *nbytes += nbits *XFS_BLF_CHUNK ; + +/* +* This takes the bit number to start looking from and +* returns the next set bit from there. It returns -1 +* if there are no more bits set or the start bit is +* beyond the end of the bitmap. +*/ + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size, + (uint )first_bit + nbits +1 );
I think the range tracking logic would be a bit more robust to not +1 here and thus not make assumptions on how the rest of the loop is implemented, but regardless this looks Ok to me:
Quoted 79 lines
fs/xfs/xfs_buf_item.c +45
-2+ }while (first_bit !=-1 ); +return ; + +slow_scan: +/* Count the first bit we jumped out of the above loop from */ + (*nvecs)++; + *nbytes +=XFS_BLF_CHUNK ; + last_bit = first_bit;while (last_bit !=-1 ) {/* * This takes the bit number to start looking from and @@ -115,11 +152,14 @@ xfs_buf_item_size_segment(if (next_bit ==-1 ) {break ; }else if (next_bit != last_bit +1 || -xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) { last_bit = next_bit; + first_bit = next_bit; (*nvecs)++; + nbits =1 ; }else { last_bit++; + nbits++; } *nbytes +=XFS_BLF_CHUNK ; }@@ -276,6 +316,38 @@ xfs_buf_item_format_segment(/* * Fill in an iovec for each set of contiguous chunks. */ +do { +ASSERT (first_bit >=0 ); + nbits =xfs_contig_bits (blfp->blf_data_map, + blfp->blf_map_size, first_bit); +ASSERT (nbits >0 ); + +/* +* Straddling a page is rare because we don't log contiguous +* chunks of unmapped buffers anywhere. +*/ +if (nbits >1 && +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) + goto slow_scan; + +xfs_buf_item_copy_iovec (lv, vecp, bp, offset, + first_bit, nbits); + blfp->blf_size++; + +/* +* This takes the bit number to start looking from and +* returns the next set bit from there. It returns -1 +* if there are no more bits set or the start bit is +* beyond the end of the bitmap. +*/ + first_bit =xfs_next_bit (blfp->blf_data_map, blfp->blf_map_size, + (uint )first_bit + nbits +1 ); + }while (first_bit !=-1 ); + +return ; + +slow_scan: +ASSERT (bp->b_addr ==NULL ); last_bit = first_bit; nbits =1 ;for (;;) {@@ -300,7 +372,7 @@ xfs_buf_item_format_segment( blfp->blf_size++;break ; }else if (next_bit != last_bit +1 || -xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { +xfs_buf_item_straddle (bp, offset, first_bit, nbits)) {xfs_buf_item_copy_iovec (lv, vecp, bp, offset, first_bit, nbits); blfp->blf_size++;Signature · 1 line
2.28.0
Re: [PATCH 2/3] xfs: xfs_buf_item_size_segment() needs to pass segment offset
In reply to Dave ChinnerQuoted 11 lines · Dave Chinner
From: Dave Chinner <[email protected]>
Otherwise it doesn't correctly calculate the number of vectors in a logged buffer that has a contiguous map that gets split into multiple regions because the range spans discontigous memory.
Probably never been hit in practice - we don't log contiguous ranges on unmapped buffers (inode clusters).
Signed-off-by Dave Chinner <[email protected]>---
Quoted 93 lines
fs/xfs/xfs_buf_item.c | 38 +++++++++++++++++++------------------- 1 file changed, 19 insertions(+), 19 deletions(-)fs/xfs/xfs_buf_item.c +19
-19@@ -55,6 +55,18 @@ xfs_buf_log_format_size( (blfp->blf_map_size *sizeof (blfp->blf_data_map[0 ])); } +static inline bool +xfs_buf_item_straddle ( +struct xfs_buf *bp, +uint offset, +int next_bit, +int last_bit) +{ +return xfs_buf_offset (bp, offset + (next_bit <<XFS_BLF_SHIFT )) != + (xfs_buf_offset (bp, offset + (last_bit <<XFS_BLF_SHIFT )) + +XFS_BLF_CHUNK ); +} +/* * Return the number of log iovecs and space needed to log the given buf log * item segment. @@ -67,6 +79,7 @@ STATIC voidxfs_buf_item_size_segment ( struct xfs_buf_log_item *bip, struct xfs_buf_log_format *blfp, + uint offset, int *nvecs, int *nbytes) {@@ -101,12 +114,8 @@ xfs_buf_item_size_segment(*/ if (next_bit ==-1 ) {break ; - }else if (next_bit != last_bit +1 ) { - last_bit = next_bit; - (*nvecs)++; - }else if (xfs_buf_offset (bp, next_bit *XFS_BLF_CHUNK ) != - (xfs_buf_offset (bp, last_bit *XFS_BLF_CHUNK ) + -XFS_BLF_CHUNK )) { + }else if (next_bit != last_bit +1 || +xfs_buf_item_straddle (bp, offset, next_bit, last_bit)) { last_bit = next_bit; (*nvecs)++; }else {@@ -141,8 +150,10 @@ xfs_buf_item_size( int *nbytes) {struct xfs_buf_log_item *bip =BUF_ITEM (lip); +struct xfs_buf *bp = bip->bli_buf;int i;int bytes; +uint offset =0 ;ASSERT (atomic_read (&bip->bli_refcount) >0 );if (bip->bli_flags &XFS_BLI_STALE ) {@@ -184,8 +195,9 @@ xfs_buf_item_size(*/ bytes =0 ;for (i =0 ; i < bip->bli_format_count; i++) { -xfs_buf_item_size_segment (bip, &bip->bli_formats[i], +xfs_buf_item_size_segment (bip, &bip->bli_formats[i], offset, nvecs, &bytes); + offset +=BBTOB (bp->b_maps[i].bm_len); }/* @@ -212,18 +224,6 @@ xfs_buf_item_copy_iovec( nbits * XFS_BLF_CHUNK); } -static inline bool -xfs_buf_item_straddle ( -struct xfs_buf *bp, -uint offset, -int next_bit, -int last_bit) -{ -return xfs_buf_offset (bp, offset + (next_bit <<XFS_BLF_SHIFT )) != - (xfs_buf_offset (bp, offset + (last_bit <<XFS_BLF_SHIFT )) + -XFS_BLF_CHUNK ); -} - static void xfs_buf_item_format_segment( struct xfs_buf_log_item *bip,Signature · 1 line
2.28.0