btrfs: allocate page arrays more efficiently
v3 cover btrfs: allocate page arrays more efficiently
In several places, btrfs allocates an array of pages, one at a time. In addition to duplicating code, the mm subsystem provides a helper to allocate multiple pages at once into an array which is suited for our usecase. In the fast path, the batching can result in better allocation decisions and less locking. This changeset first adjusts the users to call a common array-of-pages allocation function, then adjusts that common function to use the batch page allocator.
v3:
- fixed up the other style comments from kdave's v1 comments
v2:
- moved new helper to extent_io.[ch]; fixed title format
- https://lore.kernel.org/linux-btrfs/[email protected]/
v1:
- https://lore.kernel.org/linux-btrfs/[email protected]/
Sweet Tea Dorminy (2):
btrfs: factor out allocating an array of pages
btrfs: allocate page arrays using bulk page allocator
fs/btrfs/check-integrity.c | 8 ++--- fs/btrfs/compression.c | 39 +++++++++------------ fs/btrfs/extent_io.c | 71 +++++++++++++++++++++++++++++--------- fs/btrfs/extent_io.h | 2 ++ fs/btrfs/inode.c | 10 +++--- fs/btrfs/raid56.c | 30 +++------------- 6 files changed, 85 insertions(+), 75 deletions(-)
Signature · 1 line
2.35.1
PATCH v3 1/2 btrfs: factor out allocating an array of pages
Several functions currently populate an array of page pointers one allocated page at a time; factor out the common code so as to allow improvements to all of the sites at once.
---
Changes in v3:
- expanded 'int r' to 'int ret', reflected renaming throughout
- made sure there was space after variable declarations
- fixed up kerneldoc to be more btrfs's style
Changes in v2:
- moved the new btrfs_alloc_page_array() function to extent_io.[ch]
--- fs/btrfs/check-integrity.c | 8 ++--- fs/btrfs/compression.c | 39 ++++++++++------------- fs/btrfs/extent_io.c | 65 ++++++++++++++++++++++++++++---------- fs/btrfs/extent_io.h | 2 ++ fs/btrfs/inode.c | 10 +++--- fs/btrfs/raid56.c | 30 +++--------------- 6 files changed, 79 insertions(+), 75 deletions(-)
fs/btrfs/check-integrity.c +3 -5
@@ -1553,11 +1553,9 @@ static int btrfsic_read_block(struct btrfsic_state *state,return -ENOMEM ; block_ctx->datav = block_ctx->mem_to_free; block_ctx->pagev = (struct page **)(block_ctx->datav + num_pages); -for (i =0 ; i < num_pages; i++) { - block_ctx->pagev[i] =alloc_page (GFP_NOFS ); -if (!block_ctx->pagev[i]) -return -1 ; - } + ret =btrfs_alloc_page_array (num_pages, block_ctx->pagev); +if (ret) +return ret; dev_bytenr = block_ctx->dev_bytenr;for (i =0 ; i < num_pages;) {
fs/btrfs/compression.c +17 -22
@@ -801,8 +801,6 @@ blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,struct extent_map_tree *em_tree;struct compressed_bio *cb;unsigned int compressed_len; -unsigned int nr_pages; -unsigned int pg_index;struct bio *comp_bio =NULL ;const u64 disk_bytenr = bio->bi_iter.bi_sector <<SECTOR_SHIFT ;u64 cur_disk_byte = disk_bytenr;
@@ -812,7 +810,8 @@ blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio,u64 em_start;struct extent_map *em;blk_status_t ret; -int faili =0 ; +int ret2; +int i;u8 *sums; em_tree = &BTRFS_I (inode)->extent_tree;
@@ -855,25 +854,20 @@ blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, cb->compress_type =extent_compress_type (bio_flags); cb->orig_bio = bio; - nr_pages =DIV_ROUND_UP (compressed_len,PAGE_SIZE ); - cb->compressed_pages =kcalloc (nr_pages,sizeof (struct page *), -GFP_NOFS ); + cb->nr_pages =DIV_ROUND_UP (compressed_len,PAGE_SIZE ); + cb->compressed_pages =kcalloc (cb->nr_pages,sizeof (struct page *), +GFP_NOFS );if (!cb->compressed_pages) { ret =BLK_STS_RESOURCE ; - goto fail1; + goto fail; } -for (pg_index =0 ; pg_index < nr_pages; pg_index++) { - cb->compressed_pages[pg_index] =alloc_page (GFP_NOFS ); -if (!cb->compressed_pages[pg_index]) { - faili = pg_index -1 ; - ret =BLK_STS_RESOURCE ; - goto fail2; - } + ret2 =btrfs_alloc_page_array (cb->nr_pages, cb->compressed_pages); +if (ret2) { + ret =BLK_STS_RESOURCE ; + goto fail; } - faili = nr_pages -1 ; - cb->nr_pages = nr_pages; - +add_ra_bio_pages (inode, em_start + em_len, cb);/* include any pages we added in add_ra-bio_pages */
@@ -949,14 +943,15 @@ blk_status_t btrfs_submit_compressed_read(struct inode *inode, struct bio *bio, }return BLK_STS_OK ; -fail2: -while (faili >=0 ) { -__free_page (cb->compressed_pages[faili]); - faili--; +fail: +if (cb->compressed_pages) { +for (i =0 ; i < cb->nr_pages; i++) { +if (cb->compressed_pages[i]) +__free_page (cb->compressed_pages[i]); + } }kfree (cb->compressed_pages); -fail1:kfree (cb); out:free_extent_map (em);
fs/btrfs/extent_io.c +49 -16
117 lines not on this page · raw message
fs/btrfs/extent_io.h +2 -0
@@ -277,6 +277,8 @@ void extent_range_redirty_for_io(struct inode *inode, u64 start, u64 end);void extent_clear_unlock_delalloc (struct btrfs_inode *inode,u64 start,u64 end,struct page *locked_page,u32 bits_to_clear,unsigned long page_ops); + +int btrfs_alloc_page_array (unsigned long nr_pages,struct page **page_array);struct bio *btrfs_bio_alloc (unsigned int nr_iovecs);struct bio *btrfs_bio_clone (struct bio *bio);struct bio *btrfs_bio_clone_partial (struct bio *orig,u64 offset,u64 size);
fs/btrfs/inode.c +4 -6
@@ -10427,13 +10427,11 @@ static ssize_t btrfs_encoded_read_regular(struct kiocb *iocb, pages =kcalloc (nr_pages,sizeof (struct page *),GFP_NOFS );if (!pages)return -ENOMEM ; -for (i =0 ; i < nr_pages; i++) { - pages[i] =alloc_page (GFP_NOFS ); -if (!pages[i]) { - ret = -ENOMEM ; - goto out; + ret =btrfs_alloc_page_array (nr_pages, pages); +if (ret) { + ret = -ENOMEM ; + goto out; } - } ret =btrfs_encoded_read_regular_fill_pages (inode, start, disk_bytenr, disk_io_size, pages);
fs/btrfs/raid56.c +4 -26
@@ -1026,37 +1026,15 @@ static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info,/* allocate pages for all the stripes in the bio, including parity */ static int alloc_rbio_pages (struct btrfs_raid_bio *rbio) { -int i; -struct page *page; - -for (i =0 ; i < rbio->nr_pages; i++) { -if (rbio->stripe_pages[i]) -continue ; - page =alloc_page (GFP_NOFS ); -if (!page) -return -ENOMEM ; - rbio->stripe_pages[i] = page; - } -return 0 ; +return btrfs_alloc_page_array (rbio->nr_pages, rbio->stripe_pages); }/* only allocate pages for p/q stripes */ static int alloc_rbio_parity_pages (struct btrfs_raid_bio *rbio) { -int i; -struct page *page; - - i =rbio_stripe_page_index (rbio, rbio->nr_data,0 ); - -for (; i < rbio->nr_pages; i++) { -if (rbio->stripe_pages[i]) -continue ; - page =alloc_page (GFP_NOFS ); -if (!page) -return -ENOMEM ; - rbio->stripe_pages[i] = page; - } -return 0 ; +int data_pages =rbio_stripe_page_index (rbio, rbio->nr_data,0 ); +return btrfs_alloc_page_array (rbio->nr_pages - data_pages, + rbio->stripe_pages + data_pages); }/*
Signature · 1 line
2.35.1
Re: [PATCH v3 1/2] btrfs: factor out allocating an array of pages
fs/btrfs/extent_io.c static void end_bio_extent_readpage(struct bio *bio) · line 3145
Quoted 11 lines · Sweet Tea Dorminy
fs/btrfs/extent_io.c +11
-0+/** +* Populate every slot in a provided array with pages. +* +* @nr_pages: the number of pages to request +* @page_array: the array to fill with pages; any existing non-null entries in +* the array will be skipped +* +* Return: 0 if all pages were able to be allocated; -ENOMEM otherwise, and the +* caller is responsible for freeing all non-null page pointers in the array. +*/ +int btrfs_alloc_page_array (unsigned long nr_pages,struct page **page_array)
I've switched nr_pages to 'unsigned int', that's what all callers use and I don't see a reason for long.
Quoted 15 lines
fs/btrfs/extent_io.c +15
-0+{ +int i; + +for (i =0 ; i < nr_pages; i++) { +struct page *page; + +if (page_array[i]) +continue ; + page =alloc_page (GFP_NOFS ); +if (!page) +return -ENOMEM ; + page_array[i] = page; + } +return 0 ; +}
PATCH v3 2/2 btrfs: allocate page arrays using bulk page allocator
In reply to Sweet Tea DorminyWhile calling alloc_page() in a loop is an effective way to populate an array of pages, the kernel provides a method to allocate pages in bulk. alloc_pages_bulk_array() populates the NULL slots in a page array, trying to grab more than one page at a time.
Unfortunately, it doesn't guarantee allocating all slots in the array, but it's easy to call it in a loop and return an error if no progress occurs. Similar code can be found in xfs/xfs_buf.c:xfs_buf_alloc_pages().
---
Changes in v3:
- Added a newline after variable declaration
Changes in v2:
- Moved from ctree.c to extent_io.c
--- fs/btrfs/extent_io.c | 24 +++++++++++++++--------- 1 file changed, 15 insertions(+), 9 deletions(-)
fs/btrfs/extent_io.c +15 -9
@@ -3144,19 +3144,25 @@ static void end_bio_extent_readpage(struct bio *bio)*/ int btrfs_alloc_page_array (unsigned long nr_pages,struct page **page_array) { -int i; +long allocated =0 ; + +for (;;) { +long last = allocated; -for (i =0 ; i < nr_pages; i++) { -struct page *page; + allocated =alloc_pages_bulk_array (GFP_NOFS , nr_pages, + page_array); +if (allocated == nr_pages) +return 0 ; -if (page_array[i]) +if (allocated != last)continue ; - page =alloc_page (GFP_NOFS ); -if (!page) -return -ENOMEM ; - page_array[i] = page; +/* +* During this iteration, no page could be allocated, even +* though alloc_pages_bulk_array() falls back to alloc_page() +* if it could not bulk-allocate. So we must be out of memory. +*/ +return -ENOMEM ; } -return 0 ; }/*
Signature · 1 line
2.35.1
Re: [PATCH v3 2/2] btrfs: allocate page arrays using bulk page allocator
Quoted 45 lines · Sweet Tea Dorminy
While calling
alloc_page()in a loop is an effective way to populate an array of pages, the kernel provides a method to allocate pages in bulk.alloc_pages_bulk_array()populates the NULL slots in a page array, trying to grab more than one page at a time.Unfortunately, it doesn't guarantee allocating all slots in the array, but it's easy to call it in a loop and return an error if no progress occurs. Similar code can be found in xfs/xfs_buf.c:xfs_buf_alloc_pages().
Signed-off-by Sweet Tea Dorminy <[email protected]>Reviewed-by Nikolay Borisov <[email protected]>---
Changes in v3:
- Added a newline after variable declaration
Changes in v2:
- Moved fromctree.ctoextent_io.c--- fs/btrfs/extent_io.c | 24 +++++++++++++++--------- 1 file changed, 15 insertions(+), 9 deletions(-)fs/btrfs/extent_io.c +9
-5@@ -3144,19 +3144,25 @@ static void end_bio_extent_readpage(struct bio *bio)*/ int btrfs_alloc_page_array (unsigned long nr_pages,struct page **page_array) { -int i; +long allocated =0 ; + +for (;;) { +long last = allocated; -for (i =0 ; i < nr_pages; i++) { -struct page *page; + allocated =alloc_pages_bulk_array (GFP_NOFS , nr_pages, + page_array); +if (allocated == nr_pages) +return 0 ; -if (page_array[i]) +if (allocated != last)continue ; - page =alloc_page (GFP_NOFS );
fs/btrfs/extent_io.c static void end_bio_extent_readpage(struct bio *bio) · line 3165
Quoted 10 lines
fs/btrfs/extent_io.c +6
-3-if (!page) -return -ENOMEM ; - page_array[i] = page; +/* +* During this iteration, no page could be allocated, even +* though alloc_pages_bulk_array() falls back to alloc_page() +* if it could not bulk-allocate. So we must be out of memory. +*/ +return -ENOMEM ; }
I find the way the loop is structured a bit cumbersome so I'd suggest to rewrite it as:
int btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array)
{
unsigned int allocated;
for (allocated = 0; allocated < nr_pages;) {
unsigned int last = allocated;
allocated = alloc_pages_bulk_array(GFP_NOFS, nr_pages, page_array);
/*
* During this iteration, no page could be allocated, even
* though alloc_pages_bulk_array() falls back to alloc_page()
* if it could not bulk-allocate. So we must be out of memory.
*/
if (allocated == last)
return -ENOMEM;
}
return 0;
}Also in the xfs code there's memalloc_retry_wait() which is supposed to be called when repeated memory allocation is retried. What was the reason you removed it?
Re: [PATCH v3 2/2] btrfs: allocate page arrays using bulk page allocator
Quoted 71 lines · David Sterba
Quoted 45 lines · Sweet Tea Dorminy
While calling
alloc_page()in a loop is an effective way to populate an array of pages, the kernel provides a method to allocate pages in bulk.alloc_pages_bulk_array()populates the NULL slots in a page array, trying to grab more than one page at a time.Unfortunately, it doesn't guarantee allocating all slots in the array, but it's easy to call it in a loop and return an error if no progress occurs. Similar code can be found in xfs/xfs_buf.c:xfs_buf_alloc_pages().
Signed-off-by Sweet Tea Dorminy <[email protected]>Reviewed-by Nikolay Borisov <[email protected]>---
Changes in v3:
- Added a newline after variable declaration
Changes in v2:
- Moved fromctree.ctoextent_io.c--- fs/btrfs/extent_io.c | 24 +++++++++++++++--------- 1 file changed, 15 insertions(+), 9 deletions(-)fs/btrfs/extent_io.c +9
-5@@ -3144,19 +3144,25 @@ static void end_bio_extent_readpage(struct bio *bio)*/ int btrfs_alloc_page_array (unsigned long nr_pages,struct page **page_array) { -int i; +long allocated =0 ; + +for (;;) { +long last = allocated; -for (i =0 ; i < nr_pages; i++) { -struct page *page; + allocated =alloc_pages_bulk_array (GFP_NOFS , nr_pages, + page_array); +if (allocated == nr_pages) +return 0 ; -if (page_array[i]) +if (allocated != last)continue ; - page =alloc_page (GFP_NOFS );fs/btrfs/extent_io.c static void end_bio_extent_readpage(struct bio *bio) · line 3165
Quoted 10 lines
fs/btrfs/extent_io.c +6
-3-if (!page) -return -ENOMEM ; - page_array[i] = page; +/* +* During this iteration, no page could be allocated, even +* though alloc_pages_bulk_array() falls back to alloc_page() +* if it could not bulk-allocate. So we must be out of memory. +*/ +return -ENOMEM ; }I find the way the loop is structured a bit cumbersome so I'd suggest to rewrite it as:
int
btrfs_alloc_page_array(unsigned int nr_pages, struct page **page_array){ unsigned int allocated; for (allocated = 0; allocated < nr_pages;) { unsigned int last = allocated; allocated = alloc_pages_bulk_array(GFP_NOFS, nr_pages, page_array); /* * During this iteration, no page could be allocated, even
Quoted 8 lines
* though alloc_pages_bulk_array() falls back to alloc_page()
* if it could not bulk-allocate. So we must be out of memory.
*/
if (allocated == last)
return -ENOMEM;
}
return 0;
}Sounds good, I'll amend it that way.
Quoted 4 lines
Also in the xfs code there's
memalloc_retry_wait()which is supposed to be called when repeated memory allocation is retried. What was the reason you removed it?
Trying to keep the behavior as close as possible to the existing behavior.
The current behavior of each alloc_page loop is to fail if alloc_page() fails; in the worst case, alloc_pages_bulk_array() calls alloc_page() after trying to get a batch, so I figured the worst case is still basically a loop calling alloc_page() and failing if it ever fails.
Reading up on it, though, arguably the memalloc_retry_wait() should already be in all the callsites, so maybe I should insert a patch in the middle that just adds the memalloc_retry_wait() into btrfs_alloc_page_array()? Since it's an orthogonal fixup to either the refactoring or the conversion to alloc_pages_bulk_array()?
Thanks!
Sweet Tea
Re: [PATCH v3 2/2] btrfs: allocate page arrays using bulk page allocator
Quoted 5 lines · Sweet Tea Dorminy
Quoted 3 lines
Also in the xfs code there's
memalloc_retry_wait()which is supposed to be called when repeated memory allocation is retried. What was the reason you removed it?Trying to keep the behavior as close as possible to the existing behavior.
I see, makes sense.
Quoted 10 lines
The current behavior of each
alloc_pageloop is to fail ifalloc_page()fails; in the worst case,alloc_pages_bulk_array()callsalloc_page()after trying to get a batch, so I figured the worst case is still basically a loop callingalloc_page()and failing if it ever fails.Reading up on it, though, arguably the
memalloc_retry_wait()should already be in all the callsites, so maybe I should insert a patch in the middle that just adds thememalloc_retry_wait()intobtrfs_alloc_page_array()? Since it's an orthogonal fixup to either the refactoring or the conversion toalloc_pages_bulk_array()?
Yeah a separate patch with the reasonig about the potential effects is better. The v3 is now in misc-next with the suggested loop refactoring, so please send the memalloc_retry_wait() update on top of that. Thanks.