Improve the tmpfs large folio read performance
Subsystems (2)
Baolin Wang
reviewerv2 cover Improve the tmpfs large folio read performance
The tmpfs has already supported the PMD-sized large folios, but the tmpfs read operation still performs copying at the PAGE SIZE granularity, which is not perfect. This patch changes to copy data at the folio granularity, which can improve the read performance.
Use 'fio bs=64k' to read a 1G tmpfs file populated with 2M THPs, and I can see about 20% performance improvement, and no regression with bs=4k. I also did some functional test with the xfstests suite, and I did not find any regressions with the following xfstests config.
FSTYP=tmpfs
export TEST_DIR=/mnt/tempfs_mnt
export TEST_DEV=/mnt/tempfs_mnt
export SCRATCH_MNT=/mnt/scratchdir
export SCRATCH_DEV=/mnt/scratchdir
Changes from v1:
- Move index calculation to the appropriate place, per Kefeng.
- Fallback to page copy if large folio has poisoned subpages, suggested by Matthew and Yang.
Baolin Wang (2):
mm: shmem: update iocb->ki_pos directly to simplify tmpfs read logic
mm: shmem: improve the tmpfs large folio read performance
mm/shmem.c | 65 +++++++++++++++++++++++++++--------------------------- 1 file changed, 33 insertions(+), 32 deletions(-)
Signature · 1 line
2.39.3
Baolin Wang
reviewerPATCH v2 1/2 mm: shmem: update iocb->ki_pos directly to simplify tmpfs read logic
Use iocb->ki_pos to check if the read bytes exceeds the file size and to calculate the bytes to be read can help simplify the code logic. Meanwhile, this is also a preparation for improving tmpfs large folios read performace in the following patch.
--- mm/shmem.c | 35 +++++++++++------------------------ 1 file changed, 11 insertions(+), 24 deletions(-)
mm/shmem.c +11 -24
@@ -3106,27 +3106,19 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)unsigned long offset;int error =0 ;ssize_t retval =0 ; -loff_t *ppos = &iocb->ki_pos; - index = *ppos >>PAGE_SHIFT ; - offset = *ppos & ~PAGE_MASK ; + offset = iocb->ki_pos & ~PAGE_MASK ;for (;;) {struct folio *folio =NULL ;struct page *page =NULL ; -pgoff_t end_index;unsigned long nr, ret; -loff_t i_size =i_size_read (inode); +loff_t end_offset, i_size =i_size_read (inode); - end_index = i_size >>PAGE_SHIFT ; -if (index > end_index) +if (unlikely (iocb->ki_pos >= i_size))break ; -if (index == end_index) { - nr = i_size & ~PAGE_MASK ; -if (nr <= offset) -break ; - } + index = iocb->ki_pos >>PAGE_SHIFT ; error =shmem_get_folio (inode, index,0 , &folio,SGP_READ );if (error) {if (error == -EINVAL )
@@ -3148,18 +3140,14 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)* We must evaluate after, since reads (unlike writes) * are called without i_rwsem protection against truncate */ - nr =PAGE_SIZE ; i_size =i_size_read (inode); - end_index = i_size >>PAGE_SHIFT ; -if (index == end_index) { - nr = i_size & ~PAGE_MASK ; -if (nr <= offset) { -if (folio) -folio_put (folio); -break ; - } +if (unlikely (iocb->ki_pos >= i_size)) { +if (folio) +folio_put (folio); +break ; } - nr -= offset; + end_offset =min_t (loff_t, i_size, iocb->ki_pos + to->count); + nr =min_t (loff_t, end_offset - iocb->ki_pos,PAGE_SIZE - offset);if (folio) {/*
@@ -3199,8 +3187,8 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to) retval += ret; offset += ret; - index += offset >>PAGE_SHIFT ; offset &= ~PAGE_MASK ; + iocb->ki_pos += ret;if (!iov_iter_count (to))break ;
@@ -3211,7 +3199,6 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)cond_resched (); } - *ppos = ((loff_t ) index <<PAGE_SHIFT ) + offset;file_accessed (file);return retval ? retval : error; }
Signature · 1 line
2.39.3
Baolin Wang
reviewerPATCH v2 2/2 mm: shmem: improve the tmpfs large folio read performance
In reply to Baolin WangThe tmpfs has already supported the PMD-sized large folios, but the tmpfs read operation still performs copying at the PAGE SIZE granularity, which is unreasonable. This patch changes to copy data at the folio granularity, which can improve the read performance, as well as changing to use folio related functions.
Moreoever, if a large folio has a subpage that is hwpoisoned, it will still fallback to page granularity copying.
Use 'fio bs=64k' to read a 1G tmpfs file populated with 2M THPs, and I can see about 20% performance improvement, and no regression with bs=4k. Before the patch:
READ: bw=10.0GiB/s
After the patch:
READ: bw=12.0GiB/s
--- mm/shmem.c | 34 ++++++++++++++++++++++++---------- 1 file changed, 24 insertions(+), 10 deletions(-)
mm/shmem.c +24 -10
@@ -3107,13 +3107,13 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)int error =0 ;ssize_t retval =0 ; - offset = iocb->ki_pos & ~PAGE_MASK ; -for (;;) {struct folio *folio =NULL ;struct page *page =NULL ;unsigned long nr, ret;loff_t end_offset, i_size =i_size_read (inode); +bool fallback_page_copy = false; +size_t fsize;if (unlikely (iocb->ki_pos >= i_size))break ;
@@ -3134,6 +3134,10 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to) error = -EIO ;break ; } + +if (folio_test_large (folio) && +folio_test_has_hwpoisoned (folio)) + fallback_page_copy = true; }/*
@@ -3147,7 +3151,12 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)break ; } end_offset =min_t (loff_t, i_size, iocb->ki_pos + to->count); - nr =min_t (loff_t, end_offset - iocb->ki_pos,PAGE_SIZE - offset); +if (folio &&likely (!fallback_page_copy)) + fsize =folio_size (folio); +else + fsize =PAGE_SIZE ; + offset = iocb->ki_pos & (fsize -1 ); + nr =min_t (loff_t, end_offset - iocb->ki_pos, fsize - offset);if (folio) {/*
@@ -3155,10 +3164,15 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)* virtual addresses, take care about potential aliasing * before reading the page on the kernel side. */ -if (mapping_writably_mapped (mapping)) -flush_dcache_page (page); +if (mapping_writably_mapped (mapping)) { +if (likely (!fallback_page_copy)) +flush_dcache_folio (folio); +else +flush_dcache_page (page); + } +/* -* Mark the page accessed if we read the beginning. +* Mark the folio accessed if we read the beginning. */ if (!offset)folio_mark_accessed (folio);
@@ -3166,9 +3180,11 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)* Ok, we have the page, and it's up-to-date, so * now we can copy it to user space... */ - ret =copy_page_to_iter (page, offset, nr, to); +if (likely (!fallback_page_copy)) + ret =copy_folio_to_iter (folio, offset, nr, to); +else + ret =copy_page_to_iter (page, offset, nr, to);folio_put (folio); - }else if (user_backed_iter (to)) {/* * Copy to user tends to be so well optimized, but
@@ -3186,8 +3202,6 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to) } retval += ret; - offset += ret; - offset &= ~PAGE_MASK ; iocb->ki_pos += ret;if (!iov_iter_count (to))
Signature · 1 line
2.39.3
Re: [PATCH v2 1/2] mm: shmem: update iocb->ki_pos directly to simplify tmpfs read logic
In reply to Baolin WangQuoted 4 lines · Baolin Wang
Use
iocb->ki_posto check if the read bytes exceeds the file size and to calculate the bytes to be read can help simplify the code logic. Meanwhile, this is also a preparation for improving tmpfs large folios read performace
s/performace/performance
Quoted 3 lines
in the following patch.
Signed-off-by Baolin Wang <[email protected]>
The patch looks good to me. Reviewed-by: Yang Shi <[email protected]>
Quoted 86 lines
--- mm/shmem.c | 35 +++++++++++------------------------ 1 file changed, 11 insertions(+), 24 deletions(-)mm/shmem.c +11
-25@@ -3106,27 +3106,19 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)unsigned long offset;int error =0 ;ssize_t retval =0 ; -loff_t *ppos = &iocb->ki_pos; - index = *ppos >>PAGE_SHIFT ; - offset = *ppos & ~PAGE_MASK ; + offset = iocb->ki_pos & ~PAGE_MASK ;for (;;) {struct folio *folio =NULL ;struct page *page =NULL ; -pgoff_t end_index;unsigned long nr, ret; -loff_t i_size =i_size_read (inode); +loff_t end_offset, i_size =i_size_read (inode); - end_index = i_size >>PAGE_SHIFT ; -if (index > end_index) +if (unlikely (iocb->ki_pos >= i_size))break ; -if (index == end_index) { - nr = i_size & ~PAGE_MASK ; -if (nr <= offset) -break ; - } + index = iocb->ki_pos >>PAGE_SHIFT ; error =shmem_get_folio (inode, index,0 , &folio,SGP_READ );if (error) {if (error == -EINVAL )@@ -3148,18 +3140,14 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)* We must evaluate after, since reads (unlike writes) * are called without i_rwsem protection against truncate */ - nr =PAGE_SIZE ; i_size =i_size_read (inode); - end_index = i_size >>PAGE_SHIFT ; -if (index == end_index) { - nr = i_size & ~PAGE_MASK ; -if (nr <= offset) { -if (folio) -folio_put (folio); -break ; - } +if (unlikely (iocb->ki_pos >= i_size)) { +if (folio) +folio_put (folio); +break ; } - nr -= offset; + end_offset =min_t (loff_t, i_size, iocb->ki_pos + to->count); + nr =min_t (loff_t, end_offset - iocb->ki_pos,PAGE_SIZE - offset);if (folio) {/* @@ -3199,8 +3187,8 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to) retval += ret; offset += ret; - index += offset >>PAGE_SHIFT ; offset &= ~PAGE_MASK ; + iocb->ki_pos += ret;if (!iov_iter_count (to))break ;@@ -3211,7 +3199,6 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)cond_resched (); } - *ppos = ((loff_t ) index <<PAGE_SHIFT ) + offset;file_accessed (file);return retval ? retval : error; } --2.39.3
Re: [PATCH v2 2/2] mm: shmem: improve the tmpfs large folio read performance
In reply to Baolin WangQuoted 9 lines · Baolin Wang
The tmpfs has already supported the PMD-sized large folios, but the tmpfs read operation still performs copying at the PAGE SIZE granularity, which is unreasonable. This patch changes to copy data at the folio granularity, which can improve the read performance, as well as changing to use folio related functions.
Moreoever, if a large folio has a subpage that is hwpoisoned, it will still fallback to page granularity copying.
s/Moreoever/Moreover
Quoted 10 lines
Use 'fio bs=64k' to read a 1G tmpfs file populated with 2M THPs, and I can see about 20% performance improvement, and no regression with bs=4k. Before the patch:
READ: bw=10.0GiB/sAfter the patch:
READ: bw=12.0GiB/sSigned-off-by Baolin Wang <[email protected]>
The patch looks fine to me. Reviewed-by: Yang Shi <[email protected]>
Quoted 94 lines
--- mm/shmem.c | 34 ++++++++++++++++++++++++---------- 1 file changed, 24 insertions(+), 10 deletions(-)mm/shmem.c +24
-11@@ -3107,13 +3107,13 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)int error =0 ;ssize_t retval =0 ; - offset = iocb->ki_pos & ~PAGE_MASK ; -for (;;) {struct folio *folio =NULL ;struct page *page =NULL ;unsigned long nr, ret;loff_t end_offset, i_size =i_size_read (inode); +bool fallback_page_copy = false; +size_t fsize;if (unlikely (iocb->ki_pos >= i_size))break ;@@ -3134,6 +3134,10 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to) error = -EIO ;break ; } + +if (folio_test_large (folio) && +folio_test_has_hwpoisoned (folio)) + fallback_page_copy = true; }/* @@ -3147,7 +3151,12 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)break ; } end_offset =min_t (loff_t, i_size, iocb->ki_pos + to->count); - nr =min_t (loff_t, end_offset - iocb->ki_pos,PAGE_SIZE - offset); +if (folio &&likely (!fallback_page_copy)) + fsize =folio_size (folio); +else + fsize =PAGE_SIZE ; + offset = iocb->ki_pos & (fsize -1 ); + nr =min_t (loff_t, end_offset - iocb->ki_pos, fsize - offset);if (folio) {/* @@ -3155,10 +3164,15 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)* virtual addresses, take care about potential aliasing * before reading the page on the kernel side. */ -if (mapping_writably_mapped (mapping)) -flush_dcache_page (page); +if (mapping_writably_mapped (mapping)) { +if (likely (!fallback_page_copy)) +flush_dcache_folio (folio); +else +flush_dcache_page (page); + } +/* -* Mark the page accessed if we read the beginning. +* Mark the folio accessed if we read the beginning. */ if (!offset)folio_mark_accessed (folio);@@ -3166,9 +3180,11 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to)* Ok, we have the page, and it's up-to-date, so * now we can copy it to user space... */ - ret =copy_page_to_iter (page, offset, nr, to); +if (likely (!fallback_page_copy)) + ret =copy_folio_to_iter (folio, offset, nr, to); +else + ret =copy_page_to_iter (page, offset, nr, to);folio_put (folio); - }else if (user_backed_iter (to)) {/* * Copy to user tends to be so well optimized, but @@ -3186,8 +3202,6 @@ static ssize_t shmem_file_read_iter(struct kiocb *iocb, struct iov_iter *to) } retval += ret; - offset += ret; - offset &= ~PAGE_MASK ; iocb->ki_pos += ret;if (!iov_iter_count (to)) --2.39.3