| Message ID | 20200819120542.3780727-1-yukuai3@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [RFC,V3] iomap: add support to track dirty state of sub pages | expand |
On Wed, Aug 19, 2020 at 08:05:42PM +0800, Yu Kuai wrote: ... > +static void > +iomap_iop_set_range_dirty(struct page *page, unsigned int off, > + unsigned int len) > +{ > + struct iomap_page *iop = to_iomap_page(page); > + struct inode *inode = page->mapping->host; > + unsigned int first = DIRTY_BITS(off >> inode->i_blkbits); > + unsigned int last = DIRTY_BITS((off + len - 1) >> inode->i_blkbits); > + unsigned long flags; > + unsigned int i; > + > + spin_lock_irqsave(&iop->state_lock, flags); > + for (i = first; i <= last; i++) > + set_bit(i, iop->state); > + > + if (last >= first) > + iomap_set_page_dirty(page); set_page_dirty() in the atomic context? > + > + spin_unlock_irqrestore(&iop->state_lock, flags); > +} > + > +static void > +iomap_set_range_dirty(struct page *page, unsigned int off, > + unsigned int len) > +{ > + if (PageError(page)) > + return; > + > + if (page_has_private(page)) > + iomap_iop_set_range_dirty(page, off, len); I vaguely remembered iomap doesn't always set up PagePrivate. @@ -705,7 +770,7 @@ __iomap_write_end(struct inode *inode, loff_t pos, unsigned len, if (unlikely(copied < len && !PageUptodate(page))) return 0; iomap_set_range_uptodate(page, offset_in_page(pos), len); - iomap_set_page_dirty(page); + iomap_set_range_dirty(page, offset_in_page(pos), len); return copied; } so here could be suspectable, but I might be wrong here since I just take a quick look. Thanks, Gao Xiang
On 2020/8/19 20:56, Gao Xiang wrote: > On Wed, Aug 19, 2020 at 08:05:42PM +0800, Yu Kuai wrote: > > ... > >> +static void >> +iomap_iop_set_range_dirty(struct page *page, unsigned int off, >> + unsigned int len) >> +{ >> + struct iomap_page *iop = to_iomap_page(page); >> + struct inode *inode = page->mapping->host; >> + unsigned int first = DIRTY_BITS(off >> inode->i_blkbits); >> + unsigned int last = DIRTY_BITS((off + len - 1) >> inode->i_blkbits); >> + unsigned long flags; >> + unsigned int i; >> + >> + spin_lock_irqsave(&iop->state_lock, flags); >> + for (i = first; i <= last; i++) >> + set_bit(i, iop->state); >> + >> + if (last >= first) >> + iomap_set_page_dirty(page); > > set_page_dirty() in the atomic context? > Hi, You'are right, this shouldn't be inside spin_lock. >> + >> + spin_unlock_irqrestore(&iop->state_lock, flags); >> +} >> + >> +static void >> +iomap_set_range_dirty(struct page *page, unsigned int off, >> + unsigned int len) >> +{ >> + if (PageError(page)) >> + return; >> + >> + if (page_has_private(page)) >> + iomap_iop_set_range_dirty(page, off, len); > > > I vaguely remembered iomap doesn't always set up PagePrivate. > If so, maybe I should move iomap_set_page_dirty() to ioamp_set_range_dirty(). Thanks, Yu Kuai > > @@ -705,7 +770,7 @@ __iomap_write_end(struct inode *inode, loff_t pos, unsigned len, > if (unlikely(copied < len && !PageUptodate(page))) > return 0; > iomap_set_range_uptodate(page, offset_in_page(pos), len); > - iomap_set_page_dirty(page); > + iomap_set_range_dirty(page, offset_in_page(pos), len); > return copied; > } > > so here could be suspectable, but I might be wrong here since > I just take a quick look. > > Thanks, > Gao Xiang > > > . >
On Thu, Aug 20, 2020 at 09:38:53AM +0800, yukuai (C) wrote: > > > +iomap_set_range_dirty(struct page *page, unsigned int off, > > > + unsigned int len) > > > +{ > > > + if (PageError(page)) > > > + return; > > > + > > > + if (page_has_private(page)) > > > + iomap_iop_set_range_dirty(page, off, len); > > > > > > I vaguely remembered iomap doesn't always set up PagePrivate. > > > > If so, maybe I should move iomap_set_page_dirty() to > ioamp_set_range_dirty(). iomap sets PagePrivate if a iomap_page structure is allocated. Right now that means for all pages on a file system with a block size smaller than the page size, although I hope we reduce that scope a little.
> diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c > index bcfc288dba3f..b6a7457d8581 100644 > --- a/fs/iomap/buffered-io.c > +++ b/fs/iomap/buffered-io.c > @@ -21,15 +21,20 @@ > > #include "../internal.h" > > +#define DIRTY_BITS(x) ((x) + PAGE_SIZE / SECTOR_SIZE) > /* Nit: please keep an empty line between a definition and a comment. > + * The first half bits are used to track sub-page uptodate status, > + * the second half bits are for dirty status. > + */ > + DECLARE_BITMAP(state, PAGE_SIZE * 2 / SECTOR_SIZE); > }; > > static inline struct iomap_page *to_iomap_page(struct page *page) > @@ -52,8 +57,8 @@ iomap_page_create(struct inode *inode, struct page *page) > iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL); > atomic_set(&iop->read_count, 0); > atomic_set(&iop->write_count, 0); > - spin_lock_init(&iop->uptodate_lock); > - bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE); > + spin_lock_init(&iop->state_lock); > + bitmap_zero(iop->state, PAGE_SIZE * 2 / SECTOR_SIZE); Maybe add a #define IOMAP_STATE_ARRAY_SIZE (PAGE_SIZE * 2 / SECTOR_SIZE) and use? That isn't much shorter, but a little easier to read at least. > + spin_lock_irqsave(&iop->state_lock, flags); > + for (i = first; i <= last; i++) > + set_bit(i, iop->state); I think Matthew had some patches to move the these days pointlessly atomic bitops to use the bitmap_* routines. It might make sense to start out that way for new code as well. > + > + if (last >= first) > + iomap_set_page_dirty(page); > + > + spin_unlock_irqrestore(&iop->state_lock, flags); As aready pointed out, this probably needs to move out of the lock. > +static void > +iomap_set_range_dirty(struct page *page, unsigned int off, > + unsigned int len) > +{ > + if (PageError(page)) > + return; > + > + if (page_has_private(page)) > + iomap_iop_set_range_dirty(page, off, len); I'd be tempted to merge this function and iomap_iop_set_range_dirty, and just return early if there is an error or no iomap_page structure, relying on the fact that to_iomap_page returns NULL for that case. > +static void > +iomap_iop_clear_range_dirty(struct page *page, unsigned int off, > + unsigned int len) > +{ > + struct iomap_page *iop = to_iomap_page(page); > + struct inode *inode = page->mapping->host; > + unsigned int first = DIRTY_BITS(off >> inode->i_blkbits); > + unsigned int last = DIRTY_BITS((off + len - 1) >> inode->i_blkbits); > + unsigned long flags; > + unsigned int i; > + > + spin_lock_irqsave(&iop->state_lock, flags); > + for (i = first; i <= last; i++) > + clear_bit(i, iop->state); should probably use bitmap_clear(). > +static void > +iomap_clear_range_dirty(struct page *page, unsigned int off, > + unsigned int len) > +{ > + if (PageError(page)) > + return; > + > + if (page_has_private(page)) > + iomap_iop_clear_range_dirty(page, off, len); > +} Same comment about merging the two functions as above.
On Fri, Aug 21, 2020 at 07:12:34AM +0100, Christoph Hellwig wrote: > iomap sets PagePrivate if a iomap_page structure is allocated. Right > now that means for all pages on a file system with a block size smaller > than the page size, although I hope we reduce that scope a little. I was thinking about that. Is there a problem where we initially allocate the page with a contiguous extent larger than the page, then later need to write the page to a pair of extents? If we're doing an unshare operation, then we know our src and dest iomaps and can allocate the iop then. But if we readahead, we don't necessarily know our eventual dest. So the conditions for skipping allocating an iop are tricky to be sure we'll never need it.
On Fri, Aug 21, 2020 at 02:36:57PM +0100, Matthew Wilcox wrote: > On Fri, Aug 21, 2020 at 07:12:34AM +0100, Christoph Hellwig wrote: > > iomap sets PagePrivate if a iomap_page structure is allocated. Right > > now that means for all pages on a file system with a block size smaller > > than the page size, although I hope we reduce that scope a little. > > I was thinking about that. Is there a problem where we initially allocate > the page with a contiguous extent larger than the page, then later need > to write the page to a pair of extents? > > If we're doing an unshare operation, then we know our src and dest iomaps > and can allocate the iop then. But if we readahead, we don't necessarily > know our eventual dest. So the conditions for skipping allocating an > iop are tricky to be sure we'll never need it. So with the current codebase (that is without your THP work that I need to re-review) the decision should be pretty easy: - check if block size >= PAGE, and if yes don't allocate - check if the extent fully covers the page, and if yes don't allocate Now with THP we'd just need to check the thp size instead of the page above and be fine, or do I miss something?
On Sat, Aug 22, 2020 at 07:03:45AM +0100, Christoph Hellwig wrote: > On Fri, Aug 21, 2020 at 02:36:57PM +0100, Matthew Wilcox wrote: > > On Fri, Aug 21, 2020 at 07:12:34AM +0100, Christoph Hellwig wrote: > > > iomap sets PagePrivate if a iomap_page structure is allocated. Right > > > now that means for all pages on a file system with a block size smaller > > > than the page size, although I hope we reduce that scope a little. > > > > I was thinking about that. Is there a problem where we initially allocate > > the page with a contiguous extent larger than the page, then later need > > to write the page to a pair of extents? > > > > If we're doing an unshare operation, then we know our src and dest iomaps > > and can allocate the iop then. But if we readahead, we don't necessarily > > know our eventual dest. So the conditions for skipping allocating an > > iop are tricky to be sure we'll never need it. > > So with the current codebase (that is without your THP work that I need > to re-review) the decision should be pretty easy: > > - check if block size >= PAGE, and if yes don't allocate > - check if the extent fully covers the page, and if yes don't allocate > > Now with THP we'd just need to check the thp size instead of the page > above and be fine, or do I miss something? The case I was worrying about: fill a filesystem so that free space is very fragmented readahead into a hole hole is large, don't allocate an iop writeback the page don't have an iop, can't track the write count I'd be fine with choosing to allocate an iop later (and indeed I do that as part of the THP work). But does this scenario make you think of any other corner cases?
On Sat, Aug 22, 2020 at 03:24:14PM +0100, Matthew Wilcox wrote: > The case I was worrying about: > > fill a filesystem so that free space is very fragmented > readahead into a hole > hole is large, don't allocate an iop > writeback the page > don't have an iop, can't track the write count > > I'd be fine with choosing to allocate an iop later (and indeed I do that > as part of the THP work). But does this scenario make you think of any > other corner cases? Can't think of a corner case. And as said last time this comes up I think trying to allocate the iop as late and lazy as possible is probably a good thing. I just went for the dumb way because it was simpler and already a huge improvement over buffer heads.
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c index bcfc288dba3f..b6a7457d8581 100644 --- a/fs/iomap/buffered-io.c +++ b/fs/iomap/buffered-io.c @@ -21,15 +21,20 @@ #include "../internal.h" +#define DIRTY_BITS(x) ((x) + PAGE_SIZE / SECTOR_SIZE) /* * Structure allocated for each page when block size < PAGE_SIZE to track - * sub-page uptodate status and I/O completions. + * sub-page status and I/O completions. */ struct iomap_page { atomic_t read_count; atomic_t write_count; - spinlock_t uptodate_lock; - DECLARE_BITMAP(uptodate, PAGE_SIZE / 512); + spinlock_t state_lock; + /* + * The first half bits are used to track sub-page uptodate status, + * the second half bits are for dirty status. + */ + DECLARE_BITMAP(state, PAGE_SIZE * 2 / SECTOR_SIZE); }; static inline struct iomap_page *to_iomap_page(struct page *page) @@ -52,8 +57,8 @@ iomap_page_create(struct inode *inode, struct page *page) iop = kmalloc(sizeof(*iop), GFP_NOFS | __GFP_NOFAIL); atomic_set(&iop->read_count, 0); atomic_set(&iop->write_count, 0); - spin_lock_init(&iop->uptodate_lock); - bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE); + spin_lock_init(&iop->state_lock); + bitmap_zero(iop->state, PAGE_SIZE * 2 / SECTOR_SIZE); /* * migrate_page_move_mapping() assumes that pages with private data have @@ -101,7 +106,7 @@ iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop, /* move forward for each leading block marked uptodate */ for (i = first; i <= last; i++) { - if (!test_bit(i, iop->uptodate)) + if (!test_bit(i, iop->state)) break; *pos += block_size; poff += block_size; @@ -111,7 +116,7 @@ iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop, /* truncate len if we find any trailing uptodate block(s) */ for ( ; i <= last; i++) { - if (test_bit(i, iop->uptodate)) { + if (test_bit(i, iop->state)) { plen -= (last - i + 1) * block_size; last = i - 1; break; @@ -135,6 +140,66 @@ iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop, *lenp = plen; } +static void +iomap_iop_set_range_dirty(struct page *page, unsigned int off, + unsigned int len) +{ + struct iomap_page *iop = to_iomap_page(page); + struct inode *inode = page->mapping->host; + unsigned int first = DIRTY_BITS(off >> inode->i_blkbits); + unsigned int last = DIRTY_BITS((off + len - 1) >> inode->i_blkbits); + unsigned long flags; + unsigned int i; + + spin_lock_irqsave(&iop->state_lock, flags); + for (i = first; i <= last; i++) + set_bit(i, iop->state); + + if (last >= first) + iomap_set_page_dirty(page); + + spin_unlock_irqrestore(&iop->state_lock, flags); +} + +static void +iomap_set_range_dirty(struct page *page, unsigned int off, + unsigned int len) +{ + if (PageError(page)) + return; + + if (page_has_private(page)) + iomap_iop_set_range_dirty(page, off, len); +} + +static void +iomap_iop_clear_range_dirty(struct page *page, unsigned int off, + unsigned int len) +{ + struct iomap_page *iop = to_iomap_page(page); + struct inode *inode = page->mapping->host; + unsigned int first = DIRTY_BITS(off >> inode->i_blkbits); + unsigned int last = DIRTY_BITS((off + len - 1) >> inode->i_blkbits); + unsigned long flags; + unsigned int i; + + spin_lock_irqsave(&iop->state_lock, flags); + for (i = first; i <= last; i++) + clear_bit(i, iop->state); + spin_unlock_irqrestore(&iop->state_lock, flags); +} + +static void +iomap_clear_range_dirty(struct page *page, unsigned int off, + unsigned int len) +{ + if (PageError(page)) + return; + + if (page_has_private(page)) + iomap_iop_clear_range_dirty(page, off, len); +} + static void iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len) { @@ -146,17 +211,17 @@ iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len) unsigned long flags; unsigned int i; - spin_lock_irqsave(&iop->uptodate_lock, flags); + spin_lock_irqsave(&iop->state_lock, flags); for (i = 0; i < PAGE_SIZE / i_blocksize(inode); i++) { if (i >= first && i <= last) - set_bit(i, iop->uptodate); - else if (!test_bit(i, iop->uptodate)) + set_bit(i, iop->state); + else if (!test_bit(i, iop->state)) uptodate = false; } if (uptodate) SetPageUptodate(page); - spin_unlock_irqrestore(&iop->uptodate_lock, flags); + spin_unlock_irqrestore(&iop->state_lock, flags); } static void @@ -466,7 +531,7 @@ iomap_is_partially_uptodate(struct page *page, unsigned long from, if (iop) { for (i = first; i <= last; i++) - if (!test_bit(i, iop->uptodate)) + if (!test_bit(i, iop->state)) return 0; return 1; } @@ -705,7 +770,7 @@ __iomap_write_end(struct inode *inode, loff_t pos, unsigned len, if (unlikely(copied < len && !PageUptodate(page))) return 0; iomap_set_range_uptodate(page, offset_in_page(pos), len); - iomap_set_page_dirty(page); + iomap_set_range_dirty(page, offset_in_page(pos), len); return copied; } @@ -1029,7 +1094,6 @@ iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length, } else { WARN_ON_ONCE(!PageUptodate(page)); iomap_page_create(inode, page); - set_page_dirty(page); } return length; @@ -1039,7 +1103,7 @@ vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops) { struct page *page = vmf->page; struct inode *inode = file_inode(vmf->vma->vm_file); - unsigned long length; + unsigned int length, bytes_in_page; loff_t offset; ssize_t ret; @@ -1048,6 +1112,7 @@ vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops) if (ret < 0) goto out_unlock; length = ret; + bytes_in_page = ret; offset = page_offset(page); while (length > 0) { @@ -1060,6 +1125,7 @@ vm_fault_t iomap_page_mkwrite(struct vm_fault *vmf, const struct iomap_ops *ops) length -= ret; } + iomap_set_range_dirty(page, 0, bytes_in_page); wait_for_stable_page(page); return VM_FAULT_LOCKED; out_unlock: @@ -1386,7 +1452,7 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc, for (i = 0, file_offset = page_offset(page); i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset; i++, file_offset += len) { - if (iop && !test_bit(i, iop->uptodate)) + if (iop && !test_bit(DIRTY_BITS(i), iop->state)) continue; error = wpc->ops->map_blocks(wpc, inode, file_offset); @@ -1435,6 +1501,7 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc, */ set_page_writeback_keepwrite(page); } else { + iomap_clear_range_dirty(page, 0, PAGE_SIZE); clear_page_dirty_for_io(page); set_page_writeback(page); }
changes from v2: as suggested by Mathew: - move iomap_set_page_dirty() into iomap_set_range_dirty() - add DIRTY_BITS() - move ioamp_set_rannge_dirty() from iomap_page_mkwrite_actor() to iomap_page_mkwrite() - clear the dirty bits of entire page in iomap_writepage_map changes from v1: - separate set dirty and clear dirty functions - don't test uptodate bit in iomap_writepage_map() - use one bitmap array for uptodate and dirty. commit 9dc55f1389f9 ("iomap: add support for sub-pagesize buffered I/O without buffer heads") replace the per-block structure buffer_head with the per-page structure iomap_page. However, iomap_page can't track the dirty state of sub pages, which will cause performance issue since sub pages will be writeback even if they are not dirty. For example, if block size is 4k and page size is 64k: dd if=/dev/zero of=testfile bs=4k count=16 oflag=sync With buffer_head implementation, the above dd cmd will writeback 4k in each round. However, with iomap_page implementation, the range of writeback in each round is from the start of the page to the end offset we just wrote. Thus add support to track dirty state in iomap_page. I tested this path with: test environment: platform: arm64 kernel: v5.8 pagesize: 64k blocksize: 4k test case: dd if=/dev/zero of=/mnt/testfile bs=1M count=128 fio --ioengine=sync --rw=randwrite --iodepth=64 --name=test --filename=/mnt/testfile --bs=4k --fsync=1 The test result is: a. with patch ``` Jobs: 1 (f=1): [w(1)][100.0%][r=0KiB/s,w=4460KiB/s][r=0,w=1115 IOPS][eta 00m:00s] test: (groupid=0, jobs=1): err= 0: pid=3158: Tue Aug 18 07:38:53 2020 write: IOPS=1087, BW=4350KiB/s (4455kB/s)(128MiB/30129msec) clat (nsec): min=3020, max=22320, avg=4990.47, stdev=1613.56 lat (nsec): min=3180, max=23220, avg=5157.69, stdev=1617.42 clat percentiles (nsec): | 1.00th=[ 3376], 5.00th=[ 3568], 10.00th=[ 3824], 20.00th=[ 4016], | 30.00th=[ 4128], 40.00th=[ 4192], 50.00th=[ 4256], 60.00th=[ 4320], | 70.00th=[ 4512], 80.00th=[ 7392], 90.00th=[ 7840], 95.00th=[ 8032], | 99.00th=[ 8512], 99.50th=[ 8896], 99.90th=[12096], 99.95th=[14144], | 99.99th=[20096] bw ( KiB/s): min= 1504, max= 4496, per=100.00%, avg=4350.65, stdev=392.33, samples=60 iops : min= 376, max= 1124, avg=1087.65, stdev=98.08, samples=60 lat (usec) : 4=17.40%, 10=82.40%, 20=0.19%, 50=0.02% fsync/fdatasync/sync_file_range: sync (usec): min=677, max=24318, avg=903.99, stdev=455.75 sync percentiles (usec): | 1.00th=[ 685], 5.00th=[ 693], 10.00th=[ 701], 20.00th=[ 701], | 30.00th=[ 709], 40.00th=[ 709], 50.00th=[ 717], 60.00th=[ 717], | 70.00th=[ 725], 80.00th=[ 1467], 90.00th=[ 1483], 95.00th=[ 1500], | 99.00th=[ 1532], 99.50th=[ 1762], 99.90th=[ 7767], 99.95th=[ 7832], | 99.99th=[ 8094] cpu : usr=0.33%, sys=2.13%, ctx=98405, majf=0, minf=4 IO depths : 1=200.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0% submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% issued rwts: total=0,32768,0,32767 short=0,0,0,0 dropped=0,0,0,0 latency : target=0, window=0, percentile=100.00%, depth=64 Run status group 0 (all jobs): WRITE: bw=4350KiB/s (4455kB/s), 4350KiB/s-4350KiB/s (4455kB/s-4455kB/s), io=128MiB (134MB), run=30129-30129msec Disk stats (read/write): sda: ios=4/65596, merge=0/5, ticks=3/30579, in_queue=58279, util=99.72% ``` b. without patch ``` Jobs: 1 (f=1): [w(1)][100.0%][r=0KiB/s,w=3003KiB/s][r=0,w=750 IOPS][eta 00m:00s] test: (groupid=0, jobs=1): err= 0: pid=9174: Tue Aug 18 04:17:16 2020 write: IOPS=678, BW=2714KiB/s (2780kB/s)(128MiB/48286msec) clat (nsec): min=3420, max=26240, avg=5898.60, stdev=1824.49 lat (nsec): min=3600, max=26860, avg=6065.21, stdev=1826.90 clat percentiles (nsec): | 1.00th=[ 3792], 5.00th=[ 4128], 10.00th=[ 4320], 20.00th=[ 4512], | 30.00th=[ 4576], 40.00th=[ 4704], 50.00th=[ 4832], 60.00th=[ 4960], | 70.00th=[ 7968], 80.00th=[ 8256], 90.00th=[ 8512], 95.00th=[ 8768], | 99.00th=[ 9152], 99.50th=[ 9408], 99.90th=[11840], 99.95th=[13376], | 99.99th=[18560] bw ( KiB/s): min= 1016, max= 3128, per=99.92%, avg=2711.92, stdev=357.89, samples=96 iops : min= 254, max= 782, avg=677.98, stdev=89.47, samples=96 lat (usec) : 4=3.14%, 10=96.66%, 20=0.20%, 50=0.01% fsync/fdatasync/sync_file_range: sync (usec): min=814, max=24221, avg=1456.82, stdev=543.48 sync percentiles (usec): | 1.00th=[ 988], 5.00th=[ 996], 10.00th=[ 996], 20.00th=[ 1012], | 30.00th=[ 1029], 40.00th=[ 1221], 50.00th=[ 1270], 60.00th=[ 1287], | 70.00th=[ 1795], 80.00th=[ 1844], 90.00th=[ 2245], 95.00th=[ 2278], | 99.00th=[ 2442], 99.50th=[ 2737], 99.90th=[ 5407], 99.95th=[ 5538], | 99.99th=[ 5735] cpu : usr=0.19%, sys=1.54%, ctx=98412, majf=0, minf=4 IO depths : 1=200.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0% submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% issued rwts: total=0,32768,0,32767 short=0,0,0,0 dropped=0,0,0,0 latency : target=0, window=0, percentile=100.00%, depth=64 Run status group 0 (all jobs): WRITE: bw=2714KiB/s (2780kB/s), 2714KiB/s-2714KiB/s (2780kB/s-2780kB/s), io=128MiB (134MB), run=48286-48286msec Disk stats (read/write): sda: ios=4/65344, merge=0/5, ticks=2/48198, in_queue=88938, util=99.83% ``` c. ext4 ``` Jobs: 1 (f=1): [w(1)][100.0%][r=0KiB/s,w=3919KiB/s][r=0,w=979 IOPS][eta 00m:00s] test: (groupid=0, jobs=1): err= 0: pid=8682: Tue Aug 18 04:15:43 2020 write: IOPS=960, BW=3840KiB/s (3932kB/s)(128MiB/34133msec) clat (usec): min=4, max=349, avg= 8.92, stdev= 2.94 lat (usec): min=4, max=349, avg= 9.06, stdev= 2.94 clat percentiles (nsec): | 1.00th=[ 6112], 5.00th=[ 6624], 10.00th=[ 6880], 20.00th=[ 7200], | 30.00th=[ 7456], 40.00th=[ 7712], 50.00th=[ 8032], 60.00th=[ 8384], | 70.00th=[ 9024], 80.00th=[11712], 90.00th=[12608], 95.00th=[13120], | 99.00th=[14272], 99.50th=[14656], 99.90th=[17536], 99.95th=[20352], | 99.99th=[33536] bw ( KiB/s): min= 1344, max= 3992, per=100.00%, avg=3839.88, stdev=314.69, samples=68 iops : min= 336, max= 998, avg=959.97, stdev=78.67, samples=68 lat (usec) : 10=74.64%, 20=25.31%, 50=0.05%, 100=0.01%, 500=0.01% fsync/fdatasync/sync_file_range: sync (usec): min=666, max=25174, avg=1021.69, stdev=871.62 sync percentiles (usec): | 1.00th=[ 685], 5.00th=[ 693], 10.00th=[ 701], 20.00th=[ 701], | 30.00th=[ 709], 40.00th=[ 717], 50.00th=[ 717], 60.00th=[ 725], | 70.00th=[ 734], 80.00th=[ 1500], 90.00th=[ 1516], 95.00th=[ 1532], | 99.00th=[ 6128], 99.50th=[ 6128], 99.90th=[ 7832], 99.95th=[ 8225], | 99.99th=[ 9634] cpu : usr=0.32%, sys=2.87%, ctx=90254, majf=0, minf=4 IO depths : 1=200.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0% submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% issued rwts: total=0,32768,0,32767 short=0,0,0,0 dropped=0,0,0,0 latency : target=0, window=0, percentile=100.00%, depth=64 Run status group 0 (all jobs): WRITE: bw=3840KiB/s (3932kB/s), 3840KiB/s-3840KiB/s (3932kB/s-3932kB/s), io=128MiB (134MB), run=34133-34133msec Disk stats (read/write): sda: ios=0/75055, merge=0/8822, ticks=0/40565, in_queue=68469, util=99.80% ``` Signed-off-by: Yu Kuai <yukuai3@huawei.com> --- fs/iomap/buffered-io.c | 99 +++++++++++++++++++++++++++++++++++------- 1 file changed, 83 insertions(+), 16 deletions(-)