| Message ID | 1466797811-5873-3-git-send-email-rpeterso@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi, I think the idea looks good. A couple of comments below though... On 24/06/16 20:50, Bob Peterson wrote: > This patch adds a new prune_icache_sb function for the VFS slab > shrinker to call. Trying to directly free the inodes from memory > might deadlock because it evicts inodes, which calls into DLM > to acquire the glock. The DLM, in turn, may block on a pending > fence operation, which may already be blocked on memory allocation > that caused the slab shrinker to be called in the first place. > > Signed-off-by: Bob Peterson <rpeterso@redhat.com> > --- > fs/gfs2/incore.h | 2 ++ > fs/gfs2/ops_fstype.c | 1 + > fs/gfs2/quota.c | 25 +++++++++++++++++++++++++ > fs/gfs2/super.c | 13 +++++++++++++ > 4 files changed, 41 insertions(+) > > diff --git a/fs/gfs2/incore.h b/fs/gfs2/incore.h > index a6a3389..a367459 100644 > --- a/fs/gfs2/incore.h > +++ b/fs/gfs2/incore.h > @@ -757,6 +757,8 @@ struct gfs2_sbd { > > struct task_struct *sd_logd_process; > struct task_struct *sd_quotad_process; > + int sd_iprune; /* inodes to prune */ > + spinlock_t sd_shrinkspin; > > /* Quota stuff */ > > diff --git a/fs/gfs2/ops_fstype.c b/fs/gfs2/ops_fstype.c > index 4546360..65a69be 100644 > --- a/fs/gfs2/ops_fstype.c > +++ b/fs/gfs2/ops_fstype.c > @@ -95,6 +95,7 @@ static struct gfs2_sbd *init_sbd(struct super_block *sb) > spin_lock_init(&sdp->sd_jindex_spin); > mutex_init(&sdp->sd_jindex_mutex); > init_completion(&sdp->sd_journal_ready); > + spin_lock_init(&sdp->sd_shrinkspin); > > INIT_LIST_HEAD(&sdp->sd_quota_list); > mutex_init(&sdp->sd_quota_mutex); > diff --git a/fs/gfs2/quota.c b/fs/gfs2/quota.c > index ce7d69a..5810a2c 100644 > --- a/fs/gfs2/quota.c > +++ b/fs/gfs2/quota.c > @@ -1528,14 +1528,39 @@ void gfs2_wake_up_statfs(struct gfs2_sbd *sdp) { > int gfs2_quotad(void *data) > { > struct gfs2_sbd *sdp = data; > + struct super_block *sb = sdp->sd_vfs; > struct gfs2_tune *tune = &sdp->sd_tune; > unsigned long statfs_timeo = 0; > unsigned long quotad_timeo = 0; > unsigned long t = 0; > DEFINE_WAIT(wait); > int empty; > + int rc; > + struct shrink_control sc = {.gfp_mask = GFP_KERNEL, }; > > while (!kthread_should_stop()) { > + /* TODO: Deal with shrinking of dcache */ > + /* Prune any inode cache intended by the shrinker. */ > + spin_lock(&sdp->sd_shrinkspin); > + if (sdp->sd_iprune > 0) { > + sc.nr_to_scan = sdp->sd_iprune; > + if (sc.nr_to_scan > 1024) > + sc.nr_to_scan = 1024; > + sdp->sd_iprune -= sc.nr_to_scan; > + spin_unlock(&sdp->sd_shrinkspin); > + rc = prune_icache_sb(sb, &sc); > + if (rc < 0) { > + spin_lock(&sdp->sd_shrinkspin); > + sdp->sd_iprune = 0; > + spin_unlock(&sdp->sd_shrinkspin); > + } > + if (sdp->sd_iprune) { > + cond_resched(); > + continue; > + } > + } else { > + spin_unlock(&sdp->sd_shrinkspin); > + } > > /* Update the master statfs file */ > if (sdp->sd_statfs_force_sync) { > diff --git a/fs/gfs2/super.c b/fs/gfs2/super.c > index 9b2ff353..75e8a85 100644 > --- a/fs/gfs2/super.c > +++ b/fs/gfs2/super.c > @@ -1667,6 +1667,18 @@ static void gfs2_destroy_inode(struct inode *inode) > call_rcu(&inode->i_rcu, gfs2_i_callback); > } > > +static long gfs2_prune_icache_sb(struct super_block *sb, > + struct shrink_control *sc) > +{ > + struct gfs2_sbd *sdp; > + > + sdp = sb->s_fs_info; > + spin_lock(&sdp->sd_shrinkspin); > + sdp->sd_iprune = sc->nr_to_scan + 1; > + spin_unlock(&sdp->sd_shrinkspin); > + return 0; > +} This doesn't wake up the thread that will do the reclaim, so that there may be a significant delay between the request to shrink and the actual shrink. Also, using quotad is not a good plan, since it might itself block waiting for memory. This should be done by a thread on its own to avoid any deadlock possibility here. There also appears to be a limit of 1024 to scan per run of quotad, which means it would take a very long time to push out any significant number, and it does seem a bit arbitrary - was there a reason for selecting that number? It would probably be better to simply yield every now and then if there are a lot of items to process, Steve. > + > const struct super_operations gfs2_super_ops = { > .alloc_inode = gfs2_alloc_inode, > .destroy_inode = gfs2_destroy_inode, > @@ -1681,5 +1693,6 @@ const struct super_operations gfs2_super_ops = { > .remount_fs = gfs2_remount_fs, > .drop_inode = gfs2_drop_inode, > .show_options = gfs2_show_options, > + .prune_icache_sb = gfs2_prune_icache_sb, > }; > -- 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 Fri, Jun 24, 2016 at 02:50:11PM -0500, Bob Peterson wrote: > This patch adds a new prune_icache_sb function for the VFS slab > shrinker to call. Trying to directly free the inodes from memory > might deadlock because it evicts inodes, which calls into DLM > to acquire the glock. The DLM, in turn, may block on a pending > fence operation, which may already be blocked on memory allocation > that caused the slab shrinker to be called in the first place. > > Signed-off-by: Bob Peterson <rpeterso@redhat.com> All sorts of problems with this. > --- > fs/gfs2/incore.h | 2 ++ > fs/gfs2/ops_fstype.c | 1 + > fs/gfs2/quota.c | 25 +++++++++++++++++++++++++ > fs/gfs2/super.c | 13 +++++++++++++ > 4 files changed, 41 insertions(+) > > diff --git a/fs/gfs2/incore.h b/fs/gfs2/incore.h > index a6a3389..a367459 100644 > --- a/fs/gfs2/incore.h > +++ b/fs/gfs2/incore.h > @@ -757,6 +757,8 @@ struct gfs2_sbd { > > struct task_struct *sd_logd_process; > struct task_struct *sd_quotad_process; > + int sd_iprune; /* inodes to prune */ > + spinlock_t sd_shrinkspin; > > /* Quota stuff */ > > diff --git a/fs/gfs2/ops_fstype.c b/fs/gfs2/ops_fstype.c > index 4546360..65a69be 100644 > --- a/fs/gfs2/ops_fstype.c > +++ b/fs/gfs2/ops_fstype.c > @@ -95,6 +95,7 @@ static struct gfs2_sbd *init_sbd(struct super_block *sb) > spin_lock_init(&sdp->sd_jindex_spin); > mutex_init(&sdp->sd_jindex_mutex); > init_completion(&sdp->sd_journal_ready); > + spin_lock_init(&sdp->sd_shrinkspin); > > INIT_LIST_HEAD(&sdp->sd_quota_list); > mutex_init(&sdp->sd_quota_mutex); > diff --git a/fs/gfs2/quota.c b/fs/gfs2/quota.c > index ce7d69a..5810a2c 100644 > --- a/fs/gfs2/quota.c > +++ b/fs/gfs2/quota.c > @@ -1528,14 +1528,39 @@ void gfs2_wake_up_statfs(struct gfs2_sbd *sdp) { > int gfs2_quotad(void *data) > { > struct gfs2_sbd *sdp = data; > + struct super_block *sb = sdp->sd_vfs; > struct gfs2_tune *tune = &sdp->sd_tune; > unsigned long statfs_timeo = 0; > unsigned long quotad_timeo = 0; > unsigned long t = 0; > DEFINE_WAIT(wait); > int empty; > + int rc; > + struct shrink_control sc = {.gfp_mask = GFP_KERNEL, }; > Has not set PF_MEMALLOC context here, so memory reclaim operations that require allocation to complete won't get access to memory reserves and hence are likely to deadlock. > while (!kthread_should_stop()) { > + /* TODO: Deal with shrinking of dcache */ I don't think that's ever going to be allowed. > + /* Prune any inode cache intended by the shrinker. */ > + spin_lock(&sdp->sd_shrinkspin); > + if (sdp->sd_iprune > 0) { > + sc.nr_to_scan = sdp->sd_iprune; > + if (sc.nr_to_scan > 1024) > + sc.nr_to_scan = 1024; Magic number warning. Where's that come from? > + sdp->sd_iprune -= sc.nr_to_scan; > + spin_unlock(&sdp->sd_shrinkspin); > + rc = prune_icache_sb(sb, &sc); > + if (rc < 0) { > + spin_lock(&sdp->sd_shrinkspin); > + sdp->sd_iprune = 0; > + spin_unlock(&sdp->sd_shrinkspin); > + } > + if (sdp->sd_iprune) { > + cond_resched(); > + continue; > + } > + } else { > + spin_unlock(&sdp->sd_shrinkspin); > + } Ok, so this only scans 1024 inodes per loop. AFAICT, that's only once every quotad/statfs timeout, which appears to be 60/30 jiffies by default. What this means is that memory pressure is not going to be able to evict gfs2 inodes at the rate required by memory pressure. If we have a million cached inodes, and we can only reclaim 1000 every ~50ms, then we've got a severe memory imblance issue. in that same time, we could be pulling thousands on inodes into the cache.... > > /* Update the master statfs file */ > if (sdp->sd_statfs_force_sync) { > diff --git a/fs/gfs2/super.c b/fs/gfs2/super.c > index 9b2ff353..75e8a85 100644 > --- a/fs/gfs2/super.c > +++ b/fs/gfs2/super.c > @@ -1667,6 +1667,18 @@ static void gfs2_destroy_inode(struct inode *inode) > call_rcu(&inode->i_rcu, gfs2_i_callback); > } > > +static long gfs2_prune_icache_sb(struct super_block *sb, > + struct shrink_control *sc) > +{ > + struct gfs2_sbd *sdp; > + > + sdp = sb->s_fs_info; > + spin_lock(&sdp->sd_shrinkspin); > + sdp->sd_iprune = sc->nr_to_scan + 1; > + spin_unlock(&sdp->sd_shrinkspin); > + return 0; And so this is even more problematic. sc->nr_to_scan is set to to either the shrinker batch size or SHRINK_BATCH (128), and the return value of the shrinker is the number of objects freed. When the shrinker calculates the number of objects that need to be freed from the slab cache, it breaks it up into batch size chunks, and it calls iteratively until it's scanned the number of objects it calculated. This is how the shrinkers convey memory pressure - the more calls made, the higher the memory pressure, the more the cache should be scanned for reclaim. The code above, if called iteratively, continually overwrites the prune count, so it is only ever going to have a maximum of the shrinker batch size. IOWs, if the kernel wants 100,000 inodes to be scaned for reclaim, it will make 100 calls into this function, but this will only pass one call on to the background thread, which then runs some time in the next 50-500ms to do that reclaim. Hence reclaim is simply not going to be effective under heavy inode cache memory pressure. And then there's an even bigger problem - the superblock shrinker has set the flags: s->s_shrink.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE; This new shrinker is not aware of either node based reclaim or memcg based reclaim, and hence this will break more than system balance - it will break containers or anything else that uses memcgs to limit memory usage of an application. It will probably also trash the gfs2 inode cache system wide as prune_icache_sb() is given no directions for NUMA or memcg based reclaim. [ Actually, now that I look at the code and remind myself how the NUMA aware reclaim works, it's worse than I thought. Passing a sc->nid = 0 to prune_icache_sb() means it will only attempt to shrink the inode cache on node 0. This means the gfs2 inode cache on nodes 1..N can never be reclaimed by memory pressure. IOWs, it simply does not work properly.... ] Finally, returning 0 immediately from the shrinker without doing anything essentially prevents the superblock shrinker from throttling inode allocation to the rate at which inodes can be reclaimed. Because reclaim will now skip all attempts to free inodes, the memory pressure feedback loop has been broken. e.g. if you have a pure inode cache workload (e.g. doing lots of concurrent finds) that is generating all the memory pressure, there's now nothing to throttle memory allocation to the rate that memory can be reclaimed. This will cause premature ENOMEM and OOM situations to occur, as memory reclaim will blow through without reclaiming any inodes and hence freeing no memory. Shrinker contexts are complex and the VFS shrinkers are much more complex than most people understand. I strongly recommend that filesystems leave inode cache reclaim to the superblock shrinker for the above reasons - it's far too easy for people to get badly wrong. If there are specific limitations on how inodes can be freed, then move the parts of inode *freeing* that cause problems to a different context via the ->evict/destroy callouts and trigger that external context processing on demand. That external context can just do bulk "if it is on the list then free it" processing, because the reclaim policy has already been executed to place that inode on the reclaim list. This is essentially what XFS does, but it also uses the ->nr_cached_objects/->free_cached_objects() callouts in the superblock shrinker to provide the reclaim rate feedback mechanism required to throttle incoming memory allocations. Cheers, Dave.
Hi, On 28/06/16 03:08, Dave Chinner wrote: > On Fri, Jun 24, 2016 at 02:50:11PM -0500, Bob Peterson wrote: >> This patch adds a new prune_icache_sb function for the VFS slab >> shrinker to call. Trying to directly free the inodes from memory >> might deadlock because it evicts inodes, which calls into DLM >> to acquire the glock. The DLM, in turn, may block on a pending >> fence operation, which may already be blocked on memory allocation >> that caused the slab shrinker to be called in the first place. >> >> Signed-off-by: Bob Peterson <rpeterso@redhat.com> > All sorts of problems with this. > >> --- >> fs/gfs2/incore.h | 2 ++ >> fs/gfs2/ops_fstype.c | 1 + >> fs/gfs2/quota.c | 25 +++++++++++++++++++++++++ >> fs/gfs2/super.c | 13 +++++++++++++ >> 4 files changed, 41 insertions(+) >> >> diff --git a/fs/gfs2/incore.h b/fs/gfs2/incore.h >> index a6a3389..a367459 100644 >> --- a/fs/gfs2/incore.h >> +++ b/fs/gfs2/incore.h >> @@ -757,6 +757,8 @@ struct gfs2_sbd { >> >> struct task_struct *sd_logd_process; >> struct task_struct *sd_quotad_process; >> + int sd_iprune; /* inodes to prune */ >> + spinlock_t sd_shrinkspin; >> >> /* Quota stuff */ >> >> diff --git a/fs/gfs2/ops_fstype.c b/fs/gfs2/ops_fstype.c >> index 4546360..65a69be 100644 >> --- a/fs/gfs2/ops_fstype.c >> +++ b/fs/gfs2/ops_fstype.c >> @@ -95,6 +95,7 @@ static struct gfs2_sbd *init_sbd(struct super_block *sb) >> spin_lock_init(&sdp->sd_jindex_spin); >> mutex_init(&sdp->sd_jindex_mutex); >> init_completion(&sdp->sd_journal_ready); >> + spin_lock_init(&sdp->sd_shrinkspin); >> >> INIT_LIST_HEAD(&sdp->sd_quota_list); >> mutex_init(&sdp->sd_quota_mutex); >> diff --git a/fs/gfs2/quota.c b/fs/gfs2/quota.c >> index ce7d69a..5810a2c 100644 >> --- a/fs/gfs2/quota.c >> +++ b/fs/gfs2/quota.c >> @@ -1528,14 +1528,39 @@ void gfs2_wake_up_statfs(struct gfs2_sbd *sdp) { >> int gfs2_quotad(void *data) >> { >> struct gfs2_sbd *sdp = data; >> + struct super_block *sb = sdp->sd_vfs; >> struct gfs2_tune *tune = &sdp->sd_tune; >> unsigned long statfs_timeo = 0; >> unsigned long quotad_timeo = 0; >> unsigned long t = 0; >> DEFINE_WAIT(wait); >> int empty; >> + int rc; >> + struct shrink_control sc = {.gfp_mask = GFP_KERNEL, }; >> > Has not set PF_MEMALLOC context here, so memory reclaim operations that > require allocation to complete won't get access to memory reserves and > hence are likely to deadlock. > >> while (!kthread_should_stop()) { >> + /* TODO: Deal with shrinking of dcache */ > I don't think that's ever going to be allowed. > >> + /* Prune any inode cache intended by the shrinker. */ >> + spin_lock(&sdp->sd_shrinkspin); >> + if (sdp->sd_iprune > 0) { >> + sc.nr_to_scan = sdp->sd_iprune; >> + if (sc.nr_to_scan > 1024) >> + sc.nr_to_scan = 1024; > Magic number warning. Where's that come from? > >> + sdp->sd_iprune -= sc.nr_to_scan; >> + spin_unlock(&sdp->sd_shrinkspin); >> + rc = prune_icache_sb(sb, &sc); >> + if (rc < 0) { >> + spin_lock(&sdp->sd_shrinkspin); >> + sdp->sd_iprune = 0; >> + spin_unlock(&sdp->sd_shrinkspin); >> + } >> + if (sdp->sd_iprune) { >> + cond_resched(); >> + continue; >> + } >> + } else { >> + spin_unlock(&sdp->sd_shrinkspin); >> + } > Ok, so this only scans 1024 inodes per loop. AFAICT, that's only > once every quotad/statfs timeout, which appears to be 60/30 jiffies > by default. > > What this means is that memory pressure is not going to be able to > evict gfs2 inodes at the rate required by memory pressure. If we > have a million cached inodes, and we can only reclaim 1000 every > ~50ms, then we've got a severe memory imblance issue. in that same > time, we could be pulling thousands on inodes into the cache.... > >> >> /* Update the master statfs file */ >> if (sdp->sd_statfs_force_sync) { >> diff --git a/fs/gfs2/super.c b/fs/gfs2/super.c >> index 9b2ff353..75e8a85 100644 >> --- a/fs/gfs2/super.c >> +++ b/fs/gfs2/super.c >> @@ -1667,6 +1667,18 @@ static void gfs2_destroy_inode(struct inode *inode) >> call_rcu(&inode->i_rcu, gfs2_i_callback); >> } >> >> +static long gfs2_prune_icache_sb(struct super_block *sb, >> + struct shrink_control *sc) >> +{ >> + struct gfs2_sbd *sdp; >> + >> + sdp = sb->s_fs_info; >> + spin_lock(&sdp->sd_shrinkspin); >> + sdp->sd_iprune = sc->nr_to_scan + 1; >> + spin_unlock(&sdp->sd_shrinkspin); >> + return 0; > And so this is even more problematic. sc->nr_to_scan is set to to > either the shrinker batch size or SHRINK_BATCH (128), and the return > value of the shrinker is the number of objects freed. When the > shrinker calculates the number of objects that need to be freed from > the slab cache, it breaks it up into batch size chunks, and it calls > iteratively until it's scanned the number of objects it calculated. > This is how the shrinkers convey memory pressure - the more calls > made, the higher the memory pressure, the more the cache should be > scanned for reclaim. > > The code above, if called iteratively, continually overwrites the > prune count, so it is only ever going to have a maximum of the > shrinker batch size. IOWs, if the kernel wants 100,000 inodes to be > scaned for reclaim, it will make 100 calls into this function, but > this will only pass one call on to the background thread, which then > runs some time in the next 50-500ms to do that reclaim. Hence > reclaim is simply not going to be effective under heavy inode cache > memory pressure. > > And then there's an even bigger problem - the superblock shrinker > has set the flags: > > s->s_shrink.flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE; > > This new shrinker is not aware of either node based reclaim or memcg > based reclaim, and hence this will break more than system balance - > it will break containers or anything else that uses memcgs to limit > memory usage of an application. It will probably also trash the gfs2 > inode cache system wide as prune_icache_sb() is given no directions > for NUMA or memcg based reclaim. > > [ Actually, now that I look at the code and remind myself how the > NUMA aware reclaim works, it's worse than I thought. Passing a > sc->nid = 0 to prune_icache_sb() means it will only attempt to > shrink the inode cache on node 0. This means the gfs2 inode cache on > nodes 1..N can never be reclaimed by memory pressure. IOWs, it > simply does not work properly.... ] > > Finally, returning 0 immediately from the shrinker without doing > anything essentially prevents the superblock shrinker from > throttling inode allocation to the rate at which inodes can be > reclaimed. Because reclaim will now skip all attempts to free > inodes, the memory pressure feedback loop has been broken. e.g. if > you have a pure inode cache workload (e.g. doing lots of > concurrent finds) that is generating all the memory pressure, > there's now nothing to throttle memory allocation to the rate that > memory can be reclaimed. This will cause premature ENOMEM and OOM > situations to occur, as memory reclaim will blow through without > reclaiming any inodes and hence freeing no memory. > > Shrinker contexts are complex and the VFS shrinkers are much more > complex than most people understand. I strongly recommend that > filesystems leave inode cache reclaim to the superblock shrinker for > the above reasons - it's far too easy for people to get badly wrong. > > If there are specific limitations on how inodes can be freed, then > move the parts of inode *freeing* that cause problems to a different > context via the ->evict/destroy callouts and trigger that external > context processing on demand. That external context can just do bulk > "if it is on the list then free it" processing, because the reclaim > policy has already been executed to place that inode on the reclaim > list. That causes problems, unfortunately. As soon as the inode is taken out of the hash, then a new inode can be created by another process doing a lookup (for whatever reason - NFS fh lookup, scan for unlinked but still allocated inodes) and we'd land up with two in memory inodes referring to the same on disk inode at the same time. We rely on the VFS to provide mutual exclusion there. We did look at leaving the inode hashed and making the evict async, but that didn't work either - the exact issue escapes me, but Bob's memory is probably better than mine on this issue. The issue is that there is a dependency loop involving the glocks required for inode deallocation, so we have: 1. Event occurs which requires recovery somewhere in the cluster, which memory is short 2. Userspace gets blocked waiting for memory (on inode shrinker) due to requirement for a glock, to complete eviction 3. The glock is blocked on the DLM 4. The DLM is blocked until recovery is complete -> deadlock So we have various option on where to break this loop, the big question being where is the best place to do it, as the "obvious" ones seem to have various disadvantages to them, Steve. > > This is essentially what XFS does, but it also uses the > ->nr_cached_objects/->free_cached_objects() callouts in the > superblock shrinker to provide the reclaim rate feedback mechanism > required to throttle incoming memory allocations. > > Cheers, > > Dave. -- 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 Tue, Jun 28, 2016 at 10:13:32AM +0100, Steven Whitehouse wrote: > Hi, > > On 28/06/16 03:08, Dave Chinner wrote: > >On Fri, Jun 24, 2016 at 02:50:11PM -0500, Bob Peterson wrote: > >>This patch adds a new prune_icache_sb function for the VFS slab > >>shrinker to call. Trying to directly free the inodes from memory > >>might deadlock because it evicts inodes, which calls into DLM > >>to acquire the glock. The DLM, in turn, may block on a pending > >>fence operation, which may already be blocked on memory allocation > >>that caused the slab shrinker to be called in the first place. > >> > >>Signed-off-by: Bob Peterson <rpeterso@redhat.com> > >All sorts of problems with this. ..... > >Shrinker contexts are complex and the VFS shrinkers are much more > >complex than most people understand. I strongly recommend that > >filesystems leave inode cache reclaim to the superblock shrinker for > >the above reasons - it's far too easy for people to get badly wrong. > > > >If there are specific limitations on how inodes can be freed, then > >move the parts of inode *freeing* that cause problems to a different > >context via the ->evict/destroy callouts and trigger that external > >context processing on demand. That external context can just do bulk > >"if it is on the list then free it" processing, because the reclaim > >policy has already been executed to place that inode on the reclaim > >list. > That causes problems, unfortunately. As soon as the inode is taken > out of the hash, then a new inode can be created by another process > doing a lookup (for whatever reason - NFS fh lookup, scan for > unlinked but still allocated inodes) and we'd land up with two in > memory inodes referring to the same on disk inode at the same time. Yes, that's a problem XFS has to deal with, too. the lookup needs to search the reclaim list, too, and if it is found there recycle that inode rather than allocate a new one. This generaly involves clearing the reclaim state, calling init_inode_always(inode) on the reclaimed inode and then treating it like a freshly allocated struct inode... > We did look at > leaving the inode hashed and making the evict async, but that didn't > work either - the exact issue escapes me, but Bob's memory is > probably better than mine on this issue. There's no need to play games with hashing if you layer the reclaim list as per above. The VFS state is always consistent, and the only times you need to care about inodes in the "limbo" state of waiting for reclaim is when a new lookup comes in or a superblock state transition such as freeze, remount-ro and unmount... > The issue is that there is a dependency loop involving the glocks > required for inode deallocation, so we have: > > 1. Event occurs which requires recovery somewhere in the cluster, > which memory is short > 2. Userspace gets blocked waiting for memory (on inode shrinker) > due to requirement for a glock, to complete eviction > 3. The glock is blocked on the DLM > 4. The DLM is blocked until recovery is complete -> deadlock > > So we have various option on where to break this loop, the big > question being where is the best place to do it, as the "obvious" > ones seem to have various disadvantages to them, Yes, I understand the problem. Essentially you have a "non-blocking" requirement for reclaiming inodes because the DLM dependency means that reclaim progress cannot be guaranteed. The superblock shrinker has always assumed that unreferenced inodes can be reclaimed at any time when GFP_KERNEL reclaim context is given. This is clearly not true for gfs2 - it is only safe to attempt reclaim on inodes whose glock can be obtained without blocking. There are two ways of dealing with this. The first is as I described above - defer final cleanup/freeing of specific inodes to a separate list and context out of sight of the VFS inode life cycle and hook your inode lookup code into it. This way the reclaim of the inode can be deferred until the glock can be obtained without affecting normal VFS level operations. The second is to enable the shrinker to skip inodes that cannot be immediately reclaimed. The shrinker already does this based on the internal VFS inode state in inode_lru_isolate(). It would seem to me that we could add a callout in this function to allow a non-blocking determination of inode reclaimability by the filesystem. It would truly have to be non-blocking - I think we could only sanely call it under the lru and inode spinlocks because if we can't reclaim it immediately we have to skip the inode. Dropping the locks mean we can't move the inode in the LRU, and traversal must restart. Hence we'd replace a deadlock with a livelock if we can't skip inodes we can't reclaim because we dropped the lru lock. I don't know whether you can grab a glock in a truly non-blocking fashion, which is why I didn't initially suggest it. It seems like a lot more surgery to both the inode reclaim code and the GFS2 code to enable a hook into reclaim in this way... Cheers, Dave.
diff --git a/fs/gfs2/incore.h b/fs/gfs2/incore.h index a6a3389..a367459 100644 --- a/fs/gfs2/incore.h +++ b/fs/gfs2/incore.h @@ -757,6 +757,8 @@ struct gfs2_sbd { struct task_struct *sd_logd_process; struct task_struct *sd_quotad_process; + int sd_iprune; /* inodes to prune */ + spinlock_t sd_shrinkspin; /* Quota stuff */ diff --git a/fs/gfs2/ops_fstype.c b/fs/gfs2/ops_fstype.c index 4546360..65a69be 100644 --- a/fs/gfs2/ops_fstype.c +++ b/fs/gfs2/ops_fstype.c @@ -95,6 +95,7 @@ static struct gfs2_sbd *init_sbd(struct super_block *sb) spin_lock_init(&sdp->sd_jindex_spin); mutex_init(&sdp->sd_jindex_mutex); init_completion(&sdp->sd_journal_ready); + spin_lock_init(&sdp->sd_shrinkspin); INIT_LIST_HEAD(&sdp->sd_quota_list); mutex_init(&sdp->sd_quota_mutex); diff --git a/fs/gfs2/quota.c b/fs/gfs2/quota.c index ce7d69a..5810a2c 100644 --- a/fs/gfs2/quota.c +++ b/fs/gfs2/quota.c @@ -1528,14 +1528,39 @@ void gfs2_wake_up_statfs(struct gfs2_sbd *sdp) { int gfs2_quotad(void *data) { struct gfs2_sbd *sdp = data; + struct super_block *sb = sdp->sd_vfs; struct gfs2_tune *tune = &sdp->sd_tune; unsigned long statfs_timeo = 0; unsigned long quotad_timeo = 0; unsigned long t = 0; DEFINE_WAIT(wait); int empty; + int rc; + struct shrink_control sc = {.gfp_mask = GFP_KERNEL, }; while (!kthread_should_stop()) { + /* TODO: Deal with shrinking of dcache */ + /* Prune any inode cache intended by the shrinker. */ + spin_lock(&sdp->sd_shrinkspin); + if (sdp->sd_iprune > 0) { + sc.nr_to_scan = sdp->sd_iprune; + if (sc.nr_to_scan > 1024) + sc.nr_to_scan = 1024; + sdp->sd_iprune -= sc.nr_to_scan; + spin_unlock(&sdp->sd_shrinkspin); + rc = prune_icache_sb(sb, &sc); + if (rc < 0) { + spin_lock(&sdp->sd_shrinkspin); + sdp->sd_iprune = 0; + spin_unlock(&sdp->sd_shrinkspin); + } + if (sdp->sd_iprune) { + cond_resched(); + continue; + } + } else { + spin_unlock(&sdp->sd_shrinkspin); + } /* Update the master statfs file */ if (sdp->sd_statfs_force_sync) { diff --git a/fs/gfs2/super.c b/fs/gfs2/super.c index 9b2ff353..75e8a85 100644 --- a/fs/gfs2/super.c +++ b/fs/gfs2/super.c @@ -1667,6 +1667,18 @@ static void gfs2_destroy_inode(struct inode *inode) call_rcu(&inode->i_rcu, gfs2_i_callback); } +static long gfs2_prune_icache_sb(struct super_block *sb, + struct shrink_control *sc) +{ + struct gfs2_sbd *sdp; + + sdp = sb->s_fs_info; + spin_lock(&sdp->sd_shrinkspin); + sdp->sd_iprune = sc->nr_to_scan + 1; + spin_unlock(&sdp->sd_shrinkspin); + return 0; +} + const struct super_operations gfs2_super_ops = { .alloc_inode = gfs2_alloc_inode, .destroy_inode = gfs2_destroy_inode, @@ -1681,5 +1693,6 @@ const struct super_operations gfs2_super_ops = { .remount_fs = gfs2_remount_fs, .drop_inode = gfs2_drop_inode, .show_options = gfs2_show_options, + .prune_icache_sb = gfs2_prune_icache_sb, };
This patch adds a new prune_icache_sb function for the VFS slab shrinker to call. Trying to directly free the inodes from memory might deadlock because it evicts inodes, which calls into DLM to acquire the glock. The DLM, in turn, may block on a pending fence operation, which may already be blocked on memory allocation that caused the slab shrinker to be called in the first place. Signed-off-by: Bob Peterson <rpeterso@redhat.com> --- fs/gfs2/incore.h | 2 ++ fs/gfs2/ops_fstype.c | 1 + fs/gfs2/quota.c | 25 +++++++++++++++++++++++++ fs/gfs2/super.c | 13 +++++++++++++ 4 files changed, 41 insertions(+)