| Message ID | 1453226353-61481-2-git-send-email-bfoster@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue 19-01-16 12:59:12, Brian Foster wrote: > From: Dave Chinner <dchinner@redhat.com> > > wait_sb_inodes() currently does a walk of all inodes in the > filesystem to find dirty one to wait on during sync. This is highly > inefficient and wastes a lot of CPU when there are lots of clean > cached inodes that we don't need to wait on. > > To avoid this "all inode" walk, we need to track inodes that are > currently under writeback that we need to wait for. We do this by > adding inodes to a writeback list on the sb when the mapping is > first tagged as having pages under writeback. wait_sb_inodes() can > then walk this list of "inodes under IO" and wait specifically just > for the inodes that the current sync(2) needs to wait for. > > Define a couple helpers to add/remove an inode from the writeback > list and call them when the overall mapping is tagged for or cleared > from writeback. Update wait_sb_inodes() to walk only the inodes > under writeback due to the sync. The patch looks good. Just one comment: This grows struct inode by two longs. Such a growth should be justified by measuring the improvements. So can you measure some numbers showing how much the patch helped? I think it would be interesting to see: a) How much sync(2) speed has improved if there's not much to wait for. b) See whether parallel heavy stat(2) load which is rotating lots of inodes in inode cache sees some improvement when it doesn't have to contend with sync(2) on s_inode_list_lock. I believe Dave Chinner had some loads where the contention on s_inode_list_lock due to sync and rotation of inodes was pretty heavy. Thanks. Honza > Signed-off-by: Dave Chinner <dchinner@redhat.com> > Signed-off-by: Josef Bacik <jbacik@fb.com> > Signed-off-by: Brian Foster <bfoster@redhat.com> > --- > fs/fs-writeback.c | 106 +++++++++++++++++++++++++++++++++++----------- > fs/inode.c | 2 + > fs/super.c | 2 + > include/linux/fs.h | 4 ++ > include/linux/writeback.h | 3 ++ > mm/page-writeback.c | 18 ++++++++ > 6 files changed, 110 insertions(+), 25 deletions(-) > > diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c > index 6915c95..63b878b 100644 > --- a/fs/fs-writeback.c > +++ b/fs/fs-writeback.c > @@ -943,6 +943,37 @@ void inode_io_list_del(struct inode *inode) > } > > /* > + * mark an inode as under writeback on the sb > + */ > +void sb_mark_inode_writeback(struct inode *inode) > +{ > + struct super_block *sb = inode->i_sb; > + unsigned long flags; > + > + if (list_empty(&inode->i_wb_list)) { > + spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); > + if (list_empty(&inode->i_wb_list)) > + list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); > + spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); > + } > +} > + > +/* > + * clear an inode as under writeback on the sb > + */ > +void sb_clear_inode_writeback(struct inode *inode) > +{ > + struct super_block *sb = inode->i_sb; > + unsigned long flags; > + > + if (!list_empty(&inode->i_wb_list)) { > + spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); > + list_del_init(&inode->i_wb_list); > + spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); > + } > +} > + > +/* > * Redirty an inode: set its when-it-was dirtied timestamp and move it to the > * furthest end of its superblock's dirty-inode list. > * > @@ -2106,7 +2137,7 @@ EXPORT_SYMBOL(__mark_inode_dirty); > */ > static void wait_sb_inodes(struct super_block *sb) > { > - struct inode *inode, *old_inode = NULL; > + LIST_HEAD(sync_list); > > /* > * We need to be protected against the filesystem going from > @@ -2115,38 +2146,60 @@ static void wait_sb_inodes(struct super_block *sb) > WARN_ON(!rwsem_is_locked(&sb->s_umount)); > > mutex_lock(&sb->s_sync_lock); > - spin_lock(&sb->s_inode_list_lock); > > /* > - * Data integrity sync. Must wait for all pages under writeback, > - * because there may have been pages dirtied before our sync > - * call, but which had writeout started before we write it out. > - * In which case, the inode may not be on the dirty list, but > - * we still have to wait for that writeout. > + * Splice the writeback list onto a temporary list to avoid waiting on > + * inodes that have started writeback after this point. > + * > + * Use rcu_read_lock() to keep the inodes around until we have a > + * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as > + * the local list because inodes can be dropped from either by writeback > + * completion. > + */ > + rcu_read_lock(); > + spin_lock_irq(&sb->s_inode_wblist_lock); > + list_splice_init(&sb->s_inodes_wb, &sync_list); > + > + /* > + * Data integrity sync. Must wait for all pages under writeback, because > + * there may have been pages dirtied before our sync call, but which had > + * writeout started before we write it out. In which case, the inode > + * may not be on the dirty list, but we still have to wait for that > + * writeout. > */ > - list_for_each_entry(inode, &sb->s_inodes, i_sb_list) { > + while (!list_empty(&sync_list)) { > + struct inode *inode = list_first_entry(&sync_list, struct inode, > + i_wb_list); > struct address_space *mapping = inode->i_mapping; > > + /* > + * Move each inode back to the wb list before we drop the lock > + * to preserve consistency between i_wb_list and the mapping > + * writeback tag. Writeback completion is responsible to remove > + * the inode from either list once the writeback tag is cleared. > + */ > + list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); > + > + /* > + * The mapping can appear untagged while still on-list since we > + * do not have the mapping lock. Skip it here, wb completion > + * will remove it. > + */ > + if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) > + continue; > + > + spin_unlock_irq(&sb->s_inode_wblist_lock); > + > spin_lock(&inode->i_lock); > - if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) || > - (mapping->nrpages == 0)) { > + if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { > spin_unlock(&inode->i_lock); > + > + spin_lock_irq(&sb->s_inode_wblist_lock); > continue; > } > __iget(inode); > spin_unlock(&inode->i_lock); > - spin_unlock(&sb->s_inode_list_lock); > - > - /* > - * We hold a reference to 'inode' so it couldn't have been > - * removed from s_inodes list while we dropped the > - * s_inode_list_lock. We cannot iput the inode now as we can > - * be holding the last reference and we cannot iput it under > - * s_inode_list_lock. So we keep the reference and iput it > - * later. > - */ > - iput(old_inode); > - old_inode = inode; > + rcu_read_unlock(); > > /* > * We keep the error status of individual mapping so that > @@ -2157,10 +2210,13 @@ static void wait_sb_inodes(struct super_block *sb) > > cond_resched(); > > - spin_lock(&sb->s_inode_list_lock); > + iput(inode); > + > + rcu_read_lock(); > + spin_lock_irq(&sb->s_inode_wblist_lock); > } > - spin_unlock(&sb->s_inode_list_lock); > - iput(old_inode); > + spin_unlock_irq(&sb->s_inode_wblist_lock); > + rcu_read_unlock(); > mutex_unlock(&sb->s_sync_lock); > } > > diff --git a/fs/inode.c b/fs/inode.c > index e491e54..f5a7eb9 100644 > --- a/fs/inode.c > +++ b/fs/inode.c > @@ -358,6 +358,7 @@ void inode_init_once(struct inode *inode) > INIT_HLIST_NODE(&inode->i_hash); > INIT_LIST_HEAD(&inode->i_devices); > INIT_LIST_HEAD(&inode->i_io_list); > + INIT_LIST_HEAD(&inode->i_wb_list); > INIT_LIST_HEAD(&inode->i_lru); > address_space_init_once(&inode->i_data); > i_size_ordered_init(inode); > @@ -500,6 +501,7 @@ void clear_inode(struct inode *inode) > BUG_ON(!list_empty(&inode->i_data.private_list)); > BUG_ON(!(inode->i_state & I_FREEING)); > BUG_ON(inode->i_state & I_CLEAR); > + BUG_ON(!list_empty(&inode->i_wb_list)); > /* don't need i_lock here, no concurrent mods to i_state */ > inode->i_state = I_FREEING | I_CLEAR; > } > diff --git a/fs/super.c b/fs/super.c > index 1182af8..60dd44a 100644 > --- a/fs/super.c > +++ b/fs/super.c > @@ -206,6 +206,8 @@ static struct super_block *alloc_super(struct file_system_type *type, int flags) > mutex_init(&s->s_sync_lock); > INIT_LIST_HEAD(&s->s_inodes); > spin_lock_init(&s->s_inode_list_lock); > + INIT_LIST_HEAD(&s->s_inodes_wb); > + spin_lock_init(&s->s_inode_wblist_lock); > > if (list_lru_init_memcg(&s->s_dentry_lru)) > goto fail; > diff --git a/include/linux/fs.h b/include/linux/fs.h > index eb73d74..ac2797d 100644 > --- a/include/linux/fs.h > +++ b/include/linux/fs.h > @@ -651,6 +651,7 @@ struct inode { > #endif > struct list_head i_lru; /* inode LRU list */ > struct list_head i_sb_list; > + struct list_head i_wb_list; /* backing dev writeback list */ > union { > struct hlist_head i_dentry; > struct rcu_head i_rcu; > @@ -1377,6 +1378,9 @@ struct super_block { > /* s_inode_list_lock protects s_inodes */ > spinlock_t s_inode_list_lock ____cacheline_aligned_in_smp; > struct list_head s_inodes; /* all inodes */ > + > + spinlock_t s_inode_wblist_lock; > + struct list_head s_inodes_wb; /* writeback inodes */ > }; > > extern struct timespec current_fs_time(struct super_block *sb); > diff --git a/include/linux/writeback.h b/include/linux/writeback.h > index b333c94..90a380c 100644 > --- a/include/linux/writeback.h > +++ b/include/linux/writeback.h > @@ -379,4 +379,7 @@ void tag_pages_for_writeback(struct address_space *mapping, > > void account_page_redirty(struct page *page); > > +void sb_mark_inode_writeback(struct inode *inode); > +void sb_clear_inode_writeback(struct inode *inode); > + > #endif /* WRITEBACK_H */ > diff --git a/mm/page-writeback.c b/mm/page-writeback.c > index 6fe7d15..a8b718137 100644 > --- a/mm/page-writeback.c > +++ b/mm/page-writeback.c > @@ -2745,6 +2745,11 @@ int test_clear_page_writeback(struct page *page) > __wb_writeout_inc(wb); > } > } > + > + if (mapping->host && !mapping_tagged(mapping, > + PAGECACHE_TAG_WRITEBACK)) > + sb_clear_inode_writeback(mapping->host); > + > spin_unlock_irqrestore(&mapping->tree_lock, flags); > } else { > ret = TestClearPageWriteback(page); > @@ -2773,11 +2778,24 @@ int __test_set_page_writeback(struct page *page, bool keep_write) > spin_lock_irqsave(&mapping->tree_lock, flags); > ret = TestSetPageWriteback(page); > if (!ret) { > + bool on_wblist; > + > + on_wblist = mapping_tagged(mapping, > + PAGECACHE_TAG_WRITEBACK); > + > radix_tree_tag_set(&mapping->page_tree, > page_index(page), > PAGECACHE_TAG_WRITEBACK); > if (bdi_cap_account_writeback(bdi)) > __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK); > + > + /* > + * We can come through here when swapping anonymous > + * pages, so we don't necessarily have an inode to track > + * for sync. > + */ > + if (mapping->host && !on_wblist) > + sb_mark_inode_writeback(mapping->host); > } > if (!PageDirty(page)) > radix_tree_tag_clear(&mapping->page_tree, > -- > 2.4.3 >
On Wed, Jan 20, 2016 at 02:26:26PM +0100, Jan Kara wrote: > On Tue 19-01-16 12:59:12, Brian Foster wrote: > > From: Dave Chinner <dchinner@redhat.com> > > > > wait_sb_inodes() currently does a walk of all inodes in the > > filesystem to find dirty one to wait on during sync. This is highly > > inefficient and wastes a lot of CPU when there are lots of clean > > cached inodes that we don't need to wait on. > > > > To avoid this "all inode" walk, we need to track inodes that are > > currently under writeback that we need to wait for. We do this by > > adding inodes to a writeback list on the sb when the mapping is > > first tagged as having pages under writeback. wait_sb_inodes() can > > then walk this list of "inodes under IO" and wait specifically just > > for the inodes that the current sync(2) needs to wait for. > > > > Define a couple helpers to add/remove an inode from the writeback > > list and call them when the overall mapping is tagged for or cleared > > from writeback. Update wait_sb_inodes() to walk only the inodes > > under writeback due to the sync. > > The patch looks good. Just one comment: This grows struct inode by two > longs. Such a growth should be justified by measuring the improvements. So > can you measure some numbers showing how much the patch helped? I think it > would be interesting to see: > > a) How much sync(2) speed has improved if there's not much to wait for. Depends on the size of the inode cache when sync is run. If it's empty it's not noticable. When you have tens of millions of cached, clean inodes the inode list traversal can takes tens of seconds. This is the sort of problem Josef reported that FB were having... > b) See whether parallel heavy stat(2) load which is rotating lots of inodes > in inode cache sees some improvement when it doesn't have to contend with > sync(2) on s_inode_list_lock. I believe Dave Chinner had some loads where > the contention on s_inode_list_lock due to sync and rotation of inodes was > pretty heavy. Just my usual fsmark workloads - they have parallel find and parallel ls -lR traversals over the created fileset. Even just running sync during creation (because there are millions of cached inodes, and ~250,000 inodes being instiated and reclaimed every second) causes lock contention problems.... Cheers, Dave.
On Thu, Jan 21, 2016 at 07:11:59AM +1100, Dave Chinner wrote: > On Wed, Jan 20, 2016 at 02:26:26PM +0100, Jan Kara wrote: > > On Tue 19-01-16 12:59:12, Brian Foster wrote: > > > From: Dave Chinner <dchinner@redhat.com> > > > > > > wait_sb_inodes() currently does a walk of all inodes in the > > > filesystem to find dirty one to wait on during sync. This is highly > > > inefficient and wastes a lot of CPU when there are lots of clean > > > cached inodes that we don't need to wait on. > > > > > > To avoid this "all inode" walk, we need to track inodes that are > > > currently under writeback that we need to wait for. We do this by > > > adding inodes to a writeback list on the sb when the mapping is > > > first tagged as having pages under writeback. wait_sb_inodes() can > > > then walk this list of "inodes under IO" and wait specifically just > > > for the inodes that the current sync(2) needs to wait for. > > > > > > Define a couple helpers to add/remove an inode from the writeback > > > list and call them when the overall mapping is tagged for or cleared > > > from writeback. Update wait_sb_inodes() to walk only the inodes > > > under writeback due to the sync. > > Hi Jan, Dave, > > The patch looks good. Just one comment: This grows struct inode by two > > longs. Such a growth should be justified by measuring the improvements. So > > can you measure some numbers showing how much the patch helped? I think it > > would be interesting to see: > > Thanks.. indeed, I had run some simple tests that demonstrate the effectiveness of the change. I reran them recently against the latest version. Some results are appended to this mail. Note that I don't have anything at the moment that demonstrates a notable improvement with rcu over the original spin lock approach. I can play with that a bit more, but that's not really the crux of the patch. > > a) How much sync(2) speed has improved if there's not much to wait for. > > Depends on the size of the inode cache when sync is run. If it's > empty it's not noticable. When you have tens of millions of cached, > clean inodes the inode list traversal can takes tens of seconds. > This is the sort of problem Josef reported that FB were having... > FWIW, Ceph has indicated this is a pain point for them as well. The results at [0] below show the difference in sync time with a largely populated inode cache before and after this patch. > > b) See whether parallel heavy stat(2) load which is rotating lots of inodes > > in inode cache sees some improvement when it doesn't have to contend with > > sync(2) on s_inode_list_lock. I believe Dave Chinner had some loads where > > the contention on s_inode_list_lock due to sync and rotation of inodes was > > pretty heavy. > > Just my usual fsmark workloads - they have parallel find and > parallel ls -lR traversals over the created fileset. Even just > running sync during creation (because there are millions of cached > inodes, and ~250,000 inodes being instiated and reclaimed every > second) causes lock contention problems.... > I ran a similar parallel (16x) fs_mark workload using '-S 4,' which incorporates a sync() per pass. Without this patch, this demonstrates a slow degradation as the inode cache grows. Results at [1]. Brian > Cheers, > > Dave. > -- > Dave Chinner > david@fromorbit.com 16xcpu, 32GB RAM x86-64 server Storage is LVM volumes on hw raid0. [0] -- sync test w/ ~10m clean inode cache - 10TB pre-populated XFS fs, cache populated via parallel find/stat workload --- 4.4.0+ # cat /proc/slabinfo | grep xfs xfs_dqtrx 0 0 528 62 8 : tunables 0 0 0 : slabdata 0 0 0 xfs_dquot 0 0 656 49 8 : tunables 0 0 0 : slabdata 0 0 0 xfs_buf 496293 496893 640 51 8 : tunables 0 0 0 : slabdata 9743 9743 0 xfs_icr 0 0 144 56 2 : tunables 0 0 0 : slabdata 0 0 0 xfs_inode 10528071 10529150 1728 18 8 : tunables 0 0 0 : slabdata 584999 584999 0 xfs_efd_item 0 0 400 40 4 : tunables 0 0 0 : slabdata 0 0 0 xfs_da_state 544 544 480 34 4 : tunables 0 0 0 : slabdata 16 16 0 xfs_btree_cur 0 0 208 39 2 : tunables 0 0 0 : slabdata 0 0 0 # time sync real 0m7.322s user 0m0.000s sys 0m7.314s # time sync real 0m7.299s user 0m0.000s sys 0m7.296s --- 4.4.0+ w/ sync patch # cat /proc/slabinfo | grep xfs xfs_dqtrx 0 0 528 62 8 : tunables 0 0 0 : slabdata 0 0 0 xfs_dquot 0 0 656 49 8 : tunables 0 0 0 : slabdata 0 0 0 xfs_buf 428214 428514 640 51 8 : tunables 0 0 0 : slabdata 8719 8719 0 xfs_icr 0 0 144 56 2 : tunables 0 0 0 : slabdata 0 0 0 xfs_inode 11054375 11054438 1728 18 8 : tunables 0 0 0 : slabdata 721323 721323 0 xfs_efd_item 0 0 400 40 4 : tunables 0 0 0 : slabdata 0 0 0 xfs_da_state 544 544 480 34 4 : tunables 0 0 0 : slabdata 16 16 0 xfs_btree_cur 0 0 208 39 2 : tunables 0 0 0 : slabdata 0 0 0 # time sync real 0m0.040s user 0m0.001s sys 0m0.003s # time sync real 0m0.002s user 0m0.001s sys 0m0.002s [1] -- fs_mark -D 1000 -S4 -n 1000 -d /mnt/0 ... -d /mnt/15 -L 32 - 1TB XFS fs --- 4.4.0+ FSUse% Count Size Files/sec App Overhead 2 16000 51200 3313.3 822514 2 32000 51200 3353.6 310268 2 48000 51200 3475.2 289941 2 64000 51200 3104.6 289993 2 80000 51200 2944.9 292124 2 96000 51200 3010.4 288042 3 112000 51200 2756.4 289761 3 128000 51200 2753.2 288096 3 144000 51200 2474.4 290797 3 160000 51200 2657.9 290898 3 176000 51200 2498.0 288247 3 192000 51200 2415.5 287329 3 208000 51200 2336.1 291113 3 224000 51200 2352.9 290103 3 240000 51200 2309.6 289580 3 256000 51200 2344.3 289828 3 272000 51200 2293.0 291282 3 288000 51200 2295.5 286538 4 304000 51200 2119.0 288906 4 320000 51200 2059.6 293605 4 336000 51200 2129.1 289825 4 352000 51200 1929.8 288186 4 368000 51200 1987.5 294596 4 384000 51200 1929.1 293528 4 400000 51200 1934.8 288138 4 416000 51200 1823.6 292318 4 432000 51200 1838.7 290890 4 448000 51200 1797.5 288816 4 464000 51200 1823.2 287190 4 480000 51200 1738.7 295745 4 496000 51200 1716.4 293821 5 512000 51200 1726.7 290445 --- 4.4.0+ w/ sync patch FSUse% Count Size Files/sec App Overhead 2 16000 51200 3409.7 999579 2 32000 51200 3481.3 286877 2 48000 51200 3447.3 282743 2 64000 51200 3522.3 283400 2 80000 51200 3427.0 286360 2 96000 51200 3360.2 307219 3 112000 51200 3377.7 286625 3 128000 51200 3363.7 285929 3 144000 51200 3345.7 283138 3 160000 51200 3384.9 291081 3 176000 51200 3084.1 285265 3 192000 51200 3388.4 291439 3 208000 51200 3242.8 286332 3 224000 51200 3337.9 285006 3 240000 51200 3442.8 292109 3 256000 51200 3230.3 283432 3 272000 51200 3358.3 286996 3 288000 51200 3309.0 288058 4 304000 51200 3293.4 284309 4 320000 51200 3221.4 284476 4 336000 51200 3241.5 283968 4 352000 51200 3228.3 284354 4 368000 51200 3255.7 286072 4 384000 51200 3094.6 290240 4 400000 51200 3385.6 288158 4 416000 51200 3265.2 284387 4 432000 51200 3315.2 289656 4 448000 51200 3275.1 284562 4 464000 51200 3238.4 294976 4 480000 51200 3060.0 290088 4 496000 51200 3359.5 286949 5 512000 51200 3156.2 288126 -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu 21-01-16 10:22:57, Brian Foster wrote: > On Thu, Jan 21, 2016 at 07:11:59AM +1100, Dave Chinner wrote: > > On Wed, Jan 20, 2016 at 02:26:26PM +0100, Jan Kara wrote: > > > On Tue 19-01-16 12:59:12, Brian Foster wrote: > > > > From: Dave Chinner <dchinner@redhat.com> > > > > > > > > wait_sb_inodes() currently does a walk of all inodes in the > > > > filesystem to find dirty one to wait on during sync. This is highly > > > > inefficient and wastes a lot of CPU when there are lots of clean > > > > cached inodes that we don't need to wait on. > > > > > > > > To avoid this "all inode" walk, we need to track inodes that are > > > > currently under writeback that we need to wait for. We do this by > > > > adding inodes to a writeback list on the sb when the mapping is > > > > first tagged as having pages under writeback. wait_sb_inodes() can > > > > then walk this list of "inodes under IO" and wait specifically just > > > > for the inodes that the current sync(2) needs to wait for. > > > > > > > > Define a couple helpers to add/remove an inode from the writeback > > > > list and call them when the overall mapping is tagged for or cleared > > > > from writeback. Update wait_sb_inodes() to walk only the inodes > > > > under writeback due to the sync. > > > > > Hi Jan, Dave, > > > > The patch looks good. Just one comment: This grows struct inode by two > > > longs. Such a growth should be justified by measuring the improvements. So > > > can you measure some numbers showing how much the patch helped? I think it > > > would be interesting to see: > > > > > Thanks.. indeed, I had run some simple tests that demonstrate the > effectiveness of the change. I reran them recently against the latest > version. Some results are appended to this mail. > > Note that I don't have anything at the moment that demonstrates a > notable improvement with rcu over the original spin lock approach. I can > play with that a bit more, but that's not really the crux of the patch. > > > > a) How much sync(2) speed has improved if there's not much to wait for. > > > > Depends on the size of the inode cache when sync is run. If it's > > empty it's not noticable. When you have tens of millions of cached, > > clean inodes the inode list traversal can takes tens of seconds. > > This is the sort of problem Josef reported that FB were having... > > > > FWIW, Ceph has indicated this is a pain point for them as well. The > results at [0] below show the difference in sync time with a largely > populated inode cache before and after this patch. > > > > b) See whether parallel heavy stat(2) load which is rotating lots of inodes > > > in inode cache sees some improvement when it doesn't have to contend with > > > sync(2) on s_inode_list_lock. I believe Dave Chinner had some loads where > > > the contention on s_inode_list_lock due to sync and rotation of inodes was > > > pretty heavy. > > > > Just my usual fsmark workloads - they have parallel find and > > parallel ls -lR traversals over the created fileset. Even just > > running sync during creation (because there are millions of cached > > inodes, and ~250,000 inodes being instiated and reclaimed every > > second) causes lock contention problems.... > > > > I ran a similar parallel (16x) fs_mark workload using '-S 4,' which > incorporates a sync() per pass. Without this patch, this demonstrates a > slow degradation as the inode cache grows. Results at [1]. Thanks for the results. I think it would be good if you incorporated them in the changelog since other people will likely be asking similar questions when seeing the inode is growing. Other than that feel free to add: Reviewed-by: Jan Kara <jack@suse.cz> Honza > 16xcpu, 32GB RAM x86-64 server > Storage is LVM volumes on hw raid0. > > [0] -- sync test w/ ~10m clean inode cache > - 10TB pre-populated XFS fs, cache populated via parallel find/stat > workload > > --- 4.4.0+ > > # cat /proc/slabinfo | grep xfs > xfs_dqtrx 0 0 528 62 8 : tunables 0 0 0 : slabdata 0 0 0 > xfs_dquot 0 0 656 49 8 : tunables 0 0 0 : slabdata 0 0 0 > xfs_buf 496293 496893 640 51 8 : tunables 0 0 0 : slabdata 9743 9743 0 > xfs_icr 0 0 144 56 2 : tunables 0 0 0 : slabdata 0 0 0 > xfs_inode 10528071 10529150 1728 18 8 : tunables 0 0 0 : slabdata 584999 584999 0 > xfs_efd_item 0 0 400 40 4 : tunables 0 0 0 : slabdata 0 0 0 > xfs_da_state 544 544 480 34 4 : tunables 0 0 0 : slabdata 16 16 0 > xfs_btree_cur 0 0 208 39 2 : tunables 0 0 0 : slabdata 0 0 0 > > # time sync > > real 0m7.322s > user 0m0.000s > sys 0m7.314s > # time sync > > real 0m7.299s > user 0m0.000s > sys 0m7.296s > > --- 4.4.0+ w/ sync patch > > # cat /proc/slabinfo | grep xfs > xfs_dqtrx 0 0 528 62 8 : tunables 0 0 0 : slabdata 0 0 0 > xfs_dquot 0 0 656 49 8 : tunables 0 0 0 : slabdata 0 0 0 > xfs_buf 428214 428514 640 51 8 : tunables 0 0 0 : slabdata 8719 8719 0 > xfs_icr 0 0 144 56 2 : tunables 0 0 0 : slabdata 0 0 0 > xfs_inode 11054375 11054438 1728 18 8 : tunables 0 0 0 : slabdata 721323 721323 0 > xfs_efd_item 0 0 400 40 4 : tunables 0 0 0 : slabdata 0 0 0 > xfs_da_state 544 544 480 34 4 : tunables 0 0 0 : slabdata 16 16 0 > xfs_btree_cur 0 0 208 39 2 : tunables 0 0 0 : slabdata 0 0 0 > > # time sync > > real 0m0.040s > user 0m0.001s > sys 0m0.003s > # time sync > > real 0m0.002s > user 0m0.001s > sys 0m0.002s > > [1] -- fs_mark -D 1000 -S4 -n 1000 -d /mnt/0 ... -d /mnt/15 -L 32 > - 1TB XFS fs > > --- 4.4.0+ > > FSUse% Count Size Files/sec App Overhead > 2 16000 51200 3313.3 822514 > 2 32000 51200 3353.6 310268 > 2 48000 51200 3475.2 289941 > 2 64000 51200 3104.6 289993 > 2 80000 51200 2944.9 292124 > 2 96000 51200 3010.4 288042 > 3 112000 51200 2756.4 289761 > 3 128000 51200 2753.2 288096 > 3 144000 51200 2474.4 290797 > 3 160000 51200 2657.9 290898 > 3 176000 51200 2498.0 288247 > 3 192000 51200 2415.5 287329 > 3 208000 51200 2336.1 291113 > 3 224000 51200 2352.9 290103 > 3 240000 51200 2309.6 289580 > 3 256000 51200 2344.3 289828 > 3 272000 51200 2293.0 291282 > 3 288000 51200 2295.5 286538 > 4 304000 51200 2119.0 288906 > 4 320000 51200 2059.6 293605 > 4 336000 51200 2129.1 289825 > 4 352000 51200 1929.8 288186 > 4 368000 51200 1987.5 294596 > 4 384000 51200 1929.1 293528 > 4 400000 51200 1934.8 288138 > 4 416000 51200 1823.6 292318 > 4 432000 51200 1838.7 290890 > 4 448000 51200 1797.5 288816 > 4 464000 51200 1823.2 287190 > 4 480000 51200 1738.7 295745 > 4 496000 51200 1716.4 293821 > 5 512000 51200 1726.7 290445 > > --- 4.4.0+ w/ sync patch > > FSUse% Count Size Files/sec App Overhead > 2 16000 51200 3409.7 999579 > 2 32000 51200 3481.3 286877 > 2 48000 51200 3447.3 282743 > 2 64000 51200 3522.3 283400 > 2 80000 51200 3427.0 286360 > 2 96000 51200 3360.2 307219 > 3 112000 51200 3377.7 286625 > 3 128000 51200 3363.7 285929 > 3 144000 51200 3345.7 283138 > 3 160000 51200 3384.9 291081 > 3 176000 51200 3084.1 285265 > 3 192000 51200 3388.4 291439 > 3 208000 51200 3242.8 286332 > 3 224000 51200 3337.9 285006 > 3 240000 51200 3442.8 292109 > 3 256000 51200 3230.3 283432 > 3 272000 51200 3358.3 286996 > 3 288000 51200 3309.0 288058 > 4 304000 51200 3293.4 284309 > 4 320000 51200 3221.4 284476 > 4 336000 51200 3241.5 283968 > 4 352000 51200 3228.3 284354 > 4 368000 51200 3255.7 286072 > 4 384000 51200 3094.6 290240 > 4 400000 51200 3385.6 288158 > 4 416000 51200 3265.2 284387 > 4 432000 51200 3315.2 289656 > 4 448000 51200 3275.1 284562 > 4 464000 51200 3238.4 294976 > 4 480000 51200 3060.0 290088 > 4 496000 51200 3359.5 286949 > 5 512000 51200 3156.2 288126 >
On Thu, Jan 21, 2016 at 05:34:11PM +0100, Jan Kara wrote: > On Thu 21-01-16 10:22:57, Brian Foster wrote: > > On Thu, Jan 21, 2016 at 07:11:59AM +1100, Dave Chinner wrote: > > > On Wed, Jan 20, 2016 at 02:26:26PM +0100, Jan Kara wrote: > > > > On Tue 19-01-16 12:59:12, Brian Foster wrote: > > > > > From: Dave Chinner <dchinner@redhat.com> > > > > > ... > > > > > > > > Hi Jan, Dave, > > ... > > > > a) How much sync(2) speed has improved if there's not much to wait for. > > > > > > Depends on the size of the inode cache when sync is run. If it's > > > empty it's not noticable. When you have tens of millions of cached, > > > clean inodes the inode list traversal can takes tens of seconds. > > > This is the sort of problem Josef reported that FB were having... > > > > > > > FWIW, Ceph has indicated this is a pain point for them as well. The > > results at [0] below show the difference in sync time with a largely > > populated inode cache before and after this patch. > > > > > > b) See whether parallel heavy stat(2) load which is rotating lots of inodes > > > > in inode cache sees some improvement when it doesn't have to contend with > > > > sync(2) on s_inode_list_lock. I believe Dave Chinner had some loads where > > > > the contention on s_inode_list_lock due to sync and rotation of inodes was > > > > pretty heavy. > > > > > > Just my usual fsmark workloads - they have parallel find and > > > parallel ls -lR traversals over the created fileset. Even just > > > running sync during creation (because there are millions of cached > > > inodes, and ~250,000 inodes being instiated and reclaimed every > > > second) causes lock contention problems.... > > > > > > > I ran a similar parallel (16x) fs_mark workload using '-S 4,' which > > incorporates a sync() per pass. Without this patch, this demonstrates a > > slow degradation as the inode cache grows. Results at [1]. > > Thanks for the results. I think it would be good if you incorporated them > in the changelog since other people will likely be asking similar > questions when seeing the inode is growing. Other than that feel free to > add: > > Reviewed-by: Jan Kara <jack@suse.cz> > No problem, thanks! Sure, I don't want to dump the raw stuff into the commit log description to avoid making it too long, but I can reference the core sync time impact. I've appended the following for now: "With this change, filesystem sync times are significantly reduced for fs' with largely populated inode caches and otherwise no other work to do. For example, on a 16xcpu 2GHz x86-64 server, 10TB XFS filesystem with a ~10m entry inode cache, sync times are reduced from ~7.3s to less than 0.1s when the filesystem is fully clean." I'll repost in a day or so if I don't receive any other feedback. Brian > Honza > > 16xcpu, 32GB RAM x86-64 server > > Storage is LVM volumes on hw raid0. > > > > [0] -- sync test w/ ~10m clean inode cache > > - 10TB pre-populated XFS fs, cache populated via parallel find/stat > > workload > > > > --- 4.4.0+ > > > > # cat /proc/slabinfo | grep xfs > > xfs_dqtrx 0 0 528 62 8 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_dquot 0 0 656 49 8 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_buf 496293 496893 640 51 8 : tunables 0 0 0 : slabdata 9743 9743 0 > > xfs_icr 0 0 144 56 2 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_inode 10528071 10529150 1728 18 8 : tunables 0 0 0 : slabdata 584999 584999 0 > > xfs_efd_item 0 0 400 40 4 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_da_state 544 544 480 34 4 : tunables 0 0 0 : slabdata 16 16 0 > > xfs_btree_cur 0 0 208 39 2 : tunables 0 0 0 : slabdata 0 0 0 > > > > # time sync > > > > real 0m7.322s > > user 0m0.000s > > sys 0m7.314s > > # time sync > > > > real 0m7.299s > > user 0m0.000s > > sys 0m7.296s > > > > --- 4.4.0+ w/ sync patch > > > > # cat /proc/slabinfo | grep xfs > > xfs_dqtrx 0 0 528 62 8 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_dquot 0 0 656 49 8 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_buf 428214 428514 640 51 8 : tunables 0 0 0 : slabdata 8719 8719 0 > > xfs_icr 0 0 144 56 2 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_inode 11054375 11054438 1728 18 8 : tunables 0 0 0 : slabdata 721323 721323 0 > > xfs_efd_item 0 0 400 40 4 : tunables 0 0 0 : slabdata 0 0 0 > > xfs_da_state 544 544 480 34 4 : tunables 0 0 0 : slabdata 16 16 0 > > xfs_btree_cur 0 0 208 39 2 : tunables 0 0 0 : slabdata 0 0 0 > > > > # time sync > > > > real 0m0.040s > > user 0m0.001s > > sys 0m0.003s > > # time sync > > > > real 0m0.002s > > user 0m0.001s > > sys 0m0.002s > > > > [1] -- fs_mark -D 1000 -S4 -n 1000 -d /mnt/0 ... -d /mnt/15 -L 32 > > - 1TB XFS fs > > > > --- 4.4.0+ > > > > FSUse% Count Size Files/sec App Overhead > > 2 16000 51200 3313.3 822514 > > 2 32000 51200 3353.6 310268 > > 2 48000 51200 3475.2 289941 > > 2 64000 51200 3104.6 289993 > > 2 80000 51200 2944.9 292124 > > 2 96000 51200 3010.4 288042 > > 3 112000 51200 2756.4 289761 > > 3 128000 51200 2753.2 288096 > > 3 144000 51200 2474.4 290797 > > 3 160000 51200 2657.9 290898 > > 3 176000 51200 2498.0 288247 > > 3 192000 51200 2415.5 287329 > > 3 208000 51200 2336.1 291113 > > 3 224000 51200 2352.9 290103 > > 3 240000 51200 2309.6 289580 > > 3 256000 51200 2344.3 289828 > > 3 272000 51200 2293.0 291282 > > 3 288000 51200 2295.5 286538 > > 4 304000 51200 2119.0 288906 > > 4 320000 51200 2059.6 293605 > > 4 336000 51200 2129.1 289825 > > 4 352000 51200 1929.8 288186 > > 4 368000 51200 1987.5 294596 > > 4 384000 51200 1929.1 293528 > > 4 400000 51200 1934.8 288138 > > 4 416000 51200 1823.6 292318 > > 4 432000 51200 1838.7 290890 > > 4 448000 51200 1797.5 288816 > > 4 464000 51200 1823.2 287190 > > 4 480000 51200 1738.7 295745 > > 4 496000 51200 1716.4 293821 > > 5 512000 51200 1726.7 290445 > > > > --- 4.4.0+ w/ sync patch > > > > FSUse% Count Size Files/sec App Overhead > > 2 16000 51200 3409.7 999579 > > 2 32000 51200 3481.3 286877 > > 2 48000 51200 3447.3 282743 > > 2 64000 51200 3522.3 283400 > > 2 80000 51200 3427.0 286360 > > 2 96000 51200 3360.2 307219 > > 3 112000 51200 3377.7 286625 > > 3 128000 51200 3363.7 285929 > > 3 144000 51200 3345.7 283138 > > 3 160000 51200 3384.9 291081 > > 3 176000 51200 3084.1 285265 > > 3 192000 51200 3388.4 291439 > > 3 208000 51200 3242.8 286332 > > 3 224000 51200 3337.9 285006 > > 3 240000 51200 3442.8 292109 > > 3 256000 51200 3230.3 283432 > > 3 272000 51200 3358.3 286996 > > 3 288000 51200 3309.0 288058 > > 4 304000 51200 3293.4 284309 > > 4 320000 51200 3221.4 284476 > > 4 336000 51200 3241.5 283968 > > 4 352000 51200 3228.3 284354 > > 4 368000 51200 3255.7 286072 > > 4 384000 51200 3094.6 290240 > > 4 400000 51200 3385.6 288158 > > 4 416000 51200 3265.2 284387 > > 4 432000 51200 3315.2 289656 > > 4 448000 51200 3275.1 284562 > > 4 464000 51200 3238.4 294976 > > 4 480000 51200 3060.0 290088 > > 4 496000 51200 3359.5 286949 > > 5 512000 51200 3156.2 288126 > > > -- > Jan Kara <jack@suse.com> > SUSE Labs, CR -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 01/21/2016 12:13 PM, Brian Foster wrote: > On Thu, Jan 21, 2016 at 05:34:11PM +0100, Jan Kara wrote: >> On Thu 21-01-16 10:22:57, Brian Foster wrote: >>> On Thu, Jan 21, 2016 at 07:11:59AM +1100, Dave Chinner wrote: >>>> On Wed, Jan 20, 2016 at 02:26:26PM +0100, Jan Kara wrote: >>>>> On Tue 19-01-16 12:59:12, Brian Foster wrote: >>>>>> From: Dave Chinner <dchinner@redhat.com> >>>>>> > ... >>>>> >>> >>> Hi Jan, Dave, >>> > ... >>>>> a) How much sync(2) speed has improved if there's not much to wait for. >>>> >>>> Depends on the size of the inode cache when sync is run. If it's >>>> empty it's not noticable. When you have tens of millions of cached, >>>> clean inodes the inode list traversal can takes tens of seconds. >>>> This is the sort of problem Josef reported that FB were having... >>>> >>> >>> FWIW, Ceph has indicated this is a pain point for them as well. The >>> results at [0] below show the difference in sync time with a largely >>> populated inode cache before and after this patch. >>> >>>>> b) See whether parallel heavy stat(2) load which is rotating lots of inodes >>>>> in inode cache sees some improvement when it doesn't have to contend with >>>>> sync(2) on s_inode_list_lock. I believe Dave Chinner had some loads where >>>>> the contention on s_inode_list_lock due to sync and rotation of inodes was >>>>> pretty heavy. >>>> >>>> Just my usual fsmark workloads - they have parallel find and >>>> parallel ls -lR traversals over the created fileset. Even just >>>> running sync during creation (because there are millions of cached >>>> inodes, and ~250,000 inodes being instiated and reclaimed every >>>> second) causes lock contention problems.... >>>> >>> >>> I ran a similar parallel (16x) fs_mark workload using '-S 4,' which >>> incorporates a sync() per pass. Without this patch, this demonstrates a >>> slow degradation as the inode cache grows. Results at [1]. >> >> Thanks for the results. I think it would be good if you incorporated them >> in the changelog since other people will likely be asking similar >> questions when seeing the inode is growing. Other than that feel free to >> add: >> >> Reviewed-by: Jan Kara <jack@suse.cz> >> > > No problem, thanks! Sure, I don't want to dump the raw stuff into the > commit log description to avoid making it too long, but I can reference > the core sync time impact. I've appended the following for now: > > "With this change, filesystem sync times are significantly reduced for > fs' with largely populated inode caches and otherwise no other work to > do. For example, on a 16xcpu 2GHz x86-64 server, 10TB XFS filesystem > with a ~10m entry inode cache, sync times are reduced from ~7.3s to less > than 0.1s when the filesystem is fully clean." > > I'll repost in a day or so if I don't receive any other feedback. > Sorry I dropped the ball on this guys, thanks for picking it up Brian! I think that changelog is acceptable. Thanks, Josef -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c index 6915c95..63b878b 100644 --- a/fs/fs-writeback.c +++ b/fs/fs-writeback.c @@ -943,6 +943,37 @@ void inode_io_list_del(struct inode *inode) } /* + * mark an inode as under writeback on the sb + */ +void sb_mark_inode_writeback(struct inode *inode) +{ + struct super_block *sb = inode->i_sb; + unsigned long flags; + + if (list_empty(&inode->i_wb_list)) { + spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); + if (list_empty(&inode->i_wb_list)) + list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); + spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); + } +} + +/* + * clear an inode as under writeback on the sb + */ +void sb_clear_inode_writeback(struct inode *inode) +{ + struct super_block *sb = inode->i_sb; + unsigned long flags; + + if (!list_empty(&inode->i_wb_list)) { + spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); + list_del_init(&inode->i_wb_list); + spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); + } +} + +/* * Redirty an inode: set its when-it-was dirtied timestamp and move it to the * furthest end of its superblock's dirty-inode list. * @@ -2106,7 +2137,7 @@ EXPORT_SYMBOL(__mark_inode_dirty); */ static void wait_sb_inodes(struct super_block *sb) { - struct inode *inode, *old_inode = NULL; + LIST_HEAD(sync_list); /* * We need to be protected against the filesystem going from @@ -2115,38 +2146,60 @@ static void wait_sb_inodes(struct super_block *sb) WARN_ON(!rwsem_is_locked(&sb->s_umount)); mutex_lock(&sb->s_sync_lock); - spin_lock(&sb->s_inode_list_lock); /* - * Data integrity sync. Must wait for all pages under writeback, - * because there may have been pages dirtied before our sync - * call, but which had writeout started before we write it out. - * In which case, the inode may not be on the dirty list, but - * we still have to wait for that writeout. + * Splice the writeback list onto a temporary list to avoid waiting on + * inodes that have started writeback after this point. + * + * Use rcu_read_lock() to keep the inodes around until we have a + * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as + * the local list because inodes can be dropped from either by writeback + * completion. + */ + rcu_read_lock(); + spin_lock_irq(&sb->s_inode_wblist_lock); + list_splice_init(&sb->s_inodes_wb, &sync_list); + + /* + * Data integrity sync. Must wait for all pages under writeback, because + * there may have been pages dirtied before our sync call, but which had + * writeout started before we write it out. In which case, the inode + * may not be on the dirty list, but we still have to wait for that + * writeout. */ - list_for_each_entry(inode, &sb->s_inodes, i_sb_list) { + while (!list_empty(&sync_list)) { + struct inode *inode = list_first_entry(&sync_list, struct inode, + i_wb_list); struct address_space *mapping = inode->i_mapping; + /* + * Move each inode back to the wb list before we drop the lock + * to preserve consistency between i_wb_list and the mapping + * writeback tag. Writeback completion is responsible to remove + * the inode from either list once the writeback tag is cleared. + */ + list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); + + /* + * The mapping can appear untagged while still on-list since we + * do not have the mapping lock. Skip it here, wb completion + * will remove it. + */ + if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) + continue; + + spin_unlock_irq(&sb->s_inode_wblist_lock); + spin_lock(&inode->i_lock); - if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) || - (mapping->nrpages == 0)) { + if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { spin_unlock(&inode->i_lock); + + spin_lock_irq(&sb->s_inode_wblist_lock); continue; } __iget(inode); spin_unlock(&inode->i_lock); - spin_unlock(&sb->s_inode_list_lock); - - /* - * We hold a reference to 'inode' so it couldn't have been - * removed from s_inodes list while we dropped the - * s_inode_list_lock. We cannot iput the inode now as we can - * be holding the last reference and we cannot iput it under - * s_inode_list_lock. So we keep the reference and iput it - * later. - */ - iput(old_inode); - old_inode = inode; + rcu_read_unlock(); /* * We keep the error status of individual mapping so that @@ -2157,10 +2210,13 @@ static void wait_sb_inodes(struct super_block *sb) cond_resched(); - spin_lock(&sb->s_inode_list_lock); + iput(inode); + + rcu_read_lock(); + spin_lock_irq(&sb->s_inode_wblist_lock); } - spin_unlock(&sb->s_inode_list_lock); - iput(old_inode); + spin_unlock_irq(&sb->s_inode_wblist_lock); + rcu_read_unlock(); mutex_unlock(&sb->s_sync_lock); } diff --git a/fs/inode.c b/fs/inode.c index e491e54..f5a7eb9 100644 --- a/fs/inode.c +++ b/fs/inode.c @@ -358,6 +358,7 @@ void inode_init_once(struct inode *inode) INIT_HLIST_NODE(&inode->i_hash); INIT_LIST_HEAD(&inode->i_devices); INIT_LIST_HEAD(&inode->i_io_list); + INIT_LIST_HEAD(&inode->i_wb_list); INIT_LIST_HEAD(&inode->i_lru); address_space_init_once(&inode->i_data); i_size_ordered_init(inode); @@ -500,6 +501,7 @@ void clear_inode(struct inode *inode) BUG_ON(!list_empty(&inode->i_data.private_list)); BUG_ON(!(inode->i_state & I_FREEING)); BUG_ON(inode->i_state & I_CLEAR); + BUG_ON(!list_empty(&inode->i_wb_list)); /* don't need i_lock here, no concurrent mods to i_state */ inode->i_state = I_FREEING | I_CLEAR; } diff --git a/fs/super.c b/fs/super.c index 1182af8..60dd44a 100644 --- a/fs/super.c +++ b/fs/super.c @@ -206,6 +206,8 @@ static struct super_block *alloc_super(struct file_system_type *type, int flags) mutex_init(&s->s_sync_lock); INIT_LIST_HEAD(&s->s_inodes); spin_lock_init(&s->s_inode_list_lock); + INIT_LIST_HEAD(&s->s_inodes_wb); + spin_lock_init(&s->s_inode_wblist_lock); if (list_lru_init_memcg(&s->s_dentry_lru)) goto fail; diff --git a/include/linux/fs.h b/include/linux/fs.h index eb73d74..ac2797d 100644 --- a/include/linux/fs.h +++ b/include/linux/fs.h @@ -651,6 +651,7 @@ struct inode { #endif struct list_head i_lru; /* inode LRU list */ struct list_head i_sb_list; + struct list_head i_wb_list; /* backing dev writeback list */ union { struct hlist_head i_dentry; struct rcu_head i_rcu; @@ -1377,6 +1378,9 @@ struct super_block { /* s_inode_list_lock protects s_inodes */ spinlock_t s_inode_list_lock ____cacheline_aligned_in_smp; struct list_head s_inodes; /* all inodes */ + + spinlock_t s_inode_wblist_lock; + struct list_head s_inodes_wb; /* writeback inodes */ }; extern struct timespec current_fs_time(struct super_block *sb); diff --git a/include/linux/writeback.h b/include/linux/writeback.h index b333c94..90a380c 100644 --- a/include/linux/writeback.h +++ b/include/linux/writeback.h @@ -379,4 +379,7 @@ void tag_pages_for_writeback(struct address_space *mapping, void account_page_redirty(struct page *page); +void sb_mark_inode_writeback(struct inode *inode); +void sb_clear_inode_writeback(struct inode *inode); + #endif /* WRITEBACK_H */ diff --git a/mm/page-writeback.c b/mm/page-writeback.c index 6fe7d15..a8b718137 100644 --- a/mm/page-writeback.c +++ b/mm/page-writeback.c @@ -2745,6 +2745,11 @@ int test_clear_page_writeback(struct page *page) __wb_writeout_inc(wb); } } + + if (mapping->host && !mapping_tagged(mapping, + PAGECACHE_TAG_WRITEBACK)) + sb_clear_inode_writeback(mapping->host); + spin_unlock_irqrestore(&mapping->tree_lock, flags); } else { ret = TestClearPageWriteback(page); @@ -2773,11 +2778,24 @@ int __test_set_page_writeback(struct page *page, bool keep_write) spin_lock_irqsave(&mapping->tree_lock, flags); ret = TestSetPageWriteback(page); if (!ret) { + bool on_wblist; + + on_wblist = mapping_tagged(mapping, + PAGECACHE_TAG_WRITEBACK); + radix_tree_tag_set(&mapping->page_tree, page_index(page), PAGECACHE_TAG_WRITEBACK); if (bdi_cap_account_writeback(bdi)) __inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK); + + /* + * We can come through here when swapping anonymous + * pages, so we don't necessarily have an inode to track + * for sync. + */ + if (mapping->host && !on_wblist) + sb_mark_inode_writeback(mapping->host); } if (!PageDirty(page)) radix_tree_tag_clear(&mapping->page_tree,