Skip to contentSkip to conversation

ext4 extent status tree LRU locking

DISCUSSION Applied Dave Hansen 4 participants 22 messages Jun 11, 2013 – Jun 18, 2013 ext4 file system

lore
Share & tools
lore

Dave Hansen

author
#0 lore raw

I've got a test case which I intended to use to stress the VM a bit. It fills memory up with page cache a couple of times. It essentially runs 30 or so cp's in parallel.

98% of my CPU is system time, and 96% of _that_ is being spent on the spinlock in ext4_es_lru_add(). I think the LRU list head and its lock end up being *REALLY* hot cachelines and are *the* bottleneck on this test. Note that this is _before_ we go in to reclaim and actually start calling in to the shrinker. There is zero memory pressure in this test.

I'm not sure the benefits of having a proper in-order LRU during reclaim outweigh such a drastic downside for the common case.

Any thoughts?

#1 lore raw

Hi Dave

Thanks for reporting this.

Quoted 3 lines · Dave Hansen

I've got a test case which I intended to use to stress the VM a bit. It fills memory up with page cache a couple of times. It essentially runs 30 or so cp's in parallel.

Could you please share your test case with me? I am glad to look at it and think about how to improve LRU locking.

Quoted 9 lines

98% of my CPU is system time, and 96% of _that_ is being spent on the spinlock in ext4_es_lru_add(). I think the LRU list head and its lock end up being *REALLY* hot cachelines and are *the* bottleneck on this test. Note that this is _before_ we go in to reclaim and actually start calling in to the shrinker. There is zero memory pressure in this test.

I'm not sure the benefits of having a proper in-order LRU during reclaim outweigh such a drastic downside for the common case.

A proper in-order LRU can help us to reclaim some memory from extent status tree when we are under heavy memory pressure. When shrinker tries to reclaim extents from these trees, some extents of files that are accessed infrequnetly will be reclaimed because we hope that frequently accessed files' extents can be kept in memory as much as possible. That is why we need a proper in-order LRU list.

Regards,
- Zheng

author
#2 lore raw
Quoted 7 lines · Zheng Liu
Quoted 3 lines · Dave Hansen

I've got a test case which I intended to use to stress the VM a bit. It fills memory up with page cache a couple of times. It essentially runs 30 or so cp's in parallel.

Could you please share your test case with me? I am glad to look at it and think about how to improve LRU locking.

I'll look in to giving you the actual test case. But I'm not sure of the licensing on it.

Essentially, the test creates an (small (~256MB) ext4 fs on a loopback-mounted ramfs device. It then goes and creates 160 64GB sparse files (one per cpu) and then cp's them all to /dev/null.

Quoted 9 lines
Quoted 8 lines

98% of my CPU is system time, and 96% of _that_ is being spent on the spinlock in ext4_es_lru_add(). I think the LRU list head and its lock end up being *REALLY* hot cachelines and are *the* bottleneck on this test. Note that this is _before_ we go in to reclaim and actually start calling in to the shrinker. There is zero memory pressure in this test.

I'm not sure the benefits of having a proper in-order LRU during reclaim outweigh such a drastic downside for the common case.

Quoted 6 lines

A proper in-order LRU can help us to reclaim some memory from extent status tree when we are under heavy memory pressure. When shrinker tries to reclaim extents from these trees, some extents of files that are accessed infrequnetly will be reclaimed because we hope that frequently accessed files' extents can be kept in memory as much as possible. That is why we need a proper in-order LRU list.

Does it need to be _strictly_ in order, though? In other words, do you truly need a *global*, perfectly in-order LRU?

You could make per-cpu LRUs, and batch movement on and off the global LRU once the local ones get to be a certain size. Or, you could keep them cpu-local *until* the shrinker is called, when the shrinker could go drain all the percpu ones.

Or, you could tag each extent in memory with its last-used time. You write an algorithm to go and walk the tree and attempt to _generally_ free the oldest objects out of a limited window.

#3 lore raw
Quoted 11 lines · Dave Hansen
Quoted 7 lines · Zheng Liu
Quoted 3 lines · Dave Hansen

I've got a test case which I intended to use to stress the VM a bit. It fills memory up with page cache a couple of times. It essentially runs 30 or so cp's in parallel.

Could you please share your test case with me? I am glad to look at it and think about how to improve LRU locking.

I'll look in to giving you the actual test case. But I'm not sure of the licensing on it.

That would be great if you could share it.

Quoted 4 lines

Essentially, the test creates an (small (~256MB) ext4 fs on a loopback-mounted ramfs device. It then goes and creates 160 64GB sparse files (one per cpu) and then cp's them all to /dev/null.

Thanks for letting me know.

Quoted 20 lines
Quoted 10 lines
Quoted 9 lines

98% of my CPU is system time, and 96% of _that_ is being spent on the spinlock in ext4_es_lru_add(). I think the LRU list head and its lock end up being *REALLY* hot cachelines and are *the* bottleneck on this test. Note that this is _before_ we go in to reclaim and actually start calling in to the shrinker. There is zero memory pressure in this test.

I'm not sure the benefits of having a proper in-order LRU during reclaim outweigh such a drastic downside for the common case.

Quoted 6 lines

A proper in-order LRU can help us to reclaim some memory from extent status tree when we are under heavy memory pressure. When shrinker tries to reclaim extents from these trees, some extents of files that are accessed infrequnetly will be reclaimed because we hope that frequently accessed files' extents can be kept in memory as much as possible. That is why we need a proper in-order LRU list.

Does it need to be _strictly_ in order, though? In other words, do you truly need a *global*, perfectly in-order LRU?

Quoted 8 lines

You could make per-cpu LRUs, and batch movement on and off the global LRU once the local ones get to be a certain size. Or, you could keep them cpu-local *until* the shrinker is called, when the shrinker could go drain all the percpu ones.

Or, you could tag each extent in memory with its last-used time. You write an algorithm to go and walk the tree and attempt to _generally_ free the oldest objects out of a limited window.

Thanks for your suggestions. I will try these solutions, and look at which one is best for us.

Regards,
- Zheng

#4 lore raw

FWIW, I'm also seeing this contention to a much smaller extent on much more "normal" hardware. I've got a 6-core (as opposed to 80) desktop machine where I'm doing the same test on a real disk (instead of a loopback-mounted ramfs file).

I'm seeing _about_ 10% CPU time being used on just spinning on the same ext4 LRU lock.

maintainer
#5 lore raw
In reply to Dave Hansen
Quoted 8 lines · Dave Hansen

You could make per-cpu LRUs, and batch movement on and off the global LRU once the local ones get to be a certain size. Or, you could keep them cpu-local *until* the shrinker is called, when the shrinker could go drain all the percpu ones.

Or, you could tag each extent in memory with its last-used time. You write an algorithm to go and walk the tree and attempt to _generally_ free the oldest objects out of a limited window.

Another approach might be to tag each inode with the last time an ext4_map_blocks() was called on the inode, and keep an unsorted list of inodes which has one or more entries in the extent cache.

When we decide to discard entries from the extent cache, we should drop all of the entries for the inode --- and then when we read in part of the extent tree leaf block, we should create entries in the extent cache for all of the extents found in the extent leaf block.

- Ted

#6 lore raw

Hi Ted,

Quoted 10 lines · Theodore Ts'o
Quoted 8 lines · Dave Hansen

You could make per-cpu LRUs, and batch movement on and off the global LRU once the local ones get to be a certain size. Or, you could keep them cpu-local *until* the shrinker is called, when the shrinker could go drain all the percpu ones.

Or, you could tag each extent in memory with its last-used time. You write an algorithm to go and walk the tree and attempt to _generally_ free the oldest objects out of a limited window.

Quoted 8 lines

Another approach might be to tag each inode with the last time an ext4_map_blocks() was called on the inode, and keep an unsorted list of inodes which has one or more entries in the extent cache.

When we decide to discard entries from the extent cache, we should drop all of the entries for the inode --- and then when we read in part of the extent tree leaf block, we should create entries in the extent cache for all of the extents found in the extent leaf block.

Thanks for your suggestion. But, sorry, I couldn't get your point here. As you suggested above, we can tag each inode with the last access time when ext4_map_blocks() is called. Then we will get an unsorted list of inodes with some extent entries. When we tries to reclaim some entries from extent cache, we can call list_sort() to get an sorted list of inodes and try to reclaim some entries according to the last access time. My question is why do we need to drop all of the entries from all inodes.

Af far as I understand, we can do the following improvement. We tag the last access time for each inode. So that would avoid to move the inode in lru list frequently. When we try to reclaim some entries, we call list_sort() to get a sorted lru list, and drop some entries from lru list. Please correct me if I misunderstood.

For creating all extent leaf blocks, that is another improvement. I will try to add it after I solve current problem that Dave reported.

Thanks,
- Zheng

maintainer
#7 lore raw
Quoted 8 lines · Zheng Liu

Thanks for your suggestion. But, sorry, I couldn't get your point here. As you suggested above, we can tag each inode with the last access time when ext4_map_blocks() is called. Then we will get an unsorted list of inodes with some extent entries. When we tries to reclaim some entries from extent cache, we can call list_sort() to get an sorted list of inodes and try to reclaim some entries according to the last access time. My question is why do we need to drop all of the entries from all inodes.

Sorry, what I meant to say is to drop all of the entries from a particular inode. That is, instead of making the decision to drop one or two extents from inode A because they haven't been used in a long time, or one or two extents from inode B because they haven't been used in a long time, we make the decision on an inode by inode basis.

Quoted 5 lines

Af far as I understand, we can do the following improvement. We tag the last access time for each inode. So that would avoid to move the inode in lru list frequently. When we try to reclaim some entries, we call list_sort() to get a sorted lru list, and drop some entries from lru list. Please correct me if I misunderstood.

Yes, that's what I meant; although note that depending on how many items we need to drop, it might not be necesary to do a full sort of the list --- we could just look for the inode with the oldest time, which would be O(n) rather than O(n log n).

If we are under enough memory pressure that the shrinker is getting called a lot, maybe it would be worthwhile keep a list of the oldest N inodes, which only gets updated every so many seconds. (Since we can always validate whether or not an inode has been touched since it had been put on the oldest N inodes very easily.)

Quoted 2 lines

For creating all extent leaf blocks, that is another improvement. I will try to add it after I solve current problem that Dave reported.

Yes, it's a separable issue, but it's related to the same idea. We should drop items in the cache which can be easily instaniated at the time same time (i.e., drop all of the items related to a particular extent leaf block, or all of the items related to a particular inode).

There are of course, other hueristics we could use, such as only searching inodes which have open file descriptors. It's not clear the benefit is worth the cost of determining which inodes fit a particular criteria or not, though.

One other possibility while we're brainstorming is using hints from posix_fadvise or madvise to drop items from the extent cache. The cautions with this idea are (a) not that many applications provide those hints, and (b) just because one process declares that it doesn't anticipate needing to access a particular file, doesn't mean that it is true for other users of the same file. A related classical problem is the backup program which reads every single file in the file system once. There's no particular good reason to have the backup program's activities affect the caches in the system, whether it be the page cache or the extent cache. The trick is determining whether or not a particular program is behaving like a backup program or not, either via a hueristic or via an explicit hint/declaration from the program.

Cheers,

- Ted

#8 lore raw
Quoted 10 lines · Theodore Ts'o
Quoted 8 lines · Zheng Liu

Thanks for your suggestion. But, sorry, I couldn't get your point here. As you suggested above, we can tag each inode with the last access time when ext4_map_blocks() is called. Then we will get an unsorted list of inodes with some extent entries. When we tries to reclaim some entries from extent cache, we can call list_sort() to get an sorted list of inodes and try to reclaim some entries according to the last access time. My question is why do we need to drop all of the entries from all inodes.

Quoted 5 lines

Sorry, what I meant to say is to drop all of the entries from a particular inode. That is, instead of making the decision to drop one or two extents from inode A because they haven't been used in a long time, or one or two extents from inode B because they haven't been used in a long time, we make the decision on an inode by inode basis.

Sorry, but it seems that we have implemented this algorithm that tries to reclaim all extents except delayed extent from a particular inode.

Let me describe it, please. When we access extent status tree of a particular inode, this inode will be moved into the tail of LRU list (sbi->s_es_lru). When we try to reclaim some extents, we will traverse this list, and reclaim all extents except delayed extent from this inode until the number of reclaimed objects is equal to 'nr_to_scan' or all extents has been discarded.

BTW, I have a patch to avoid traverse all entries in LRU list to speed up this process.

Quoted 8 lines
Quoted 6 lines

Af far as I understand, we can do the following improvement. We tag the last access time for each inode. So that would avoid to move the inode in lru list frequently. When we try to reclaim some entries, we call list_sort() to get a sorted lru list, and drop some entries from lru list. Please correct me if I misunderstood.

Yes, that's what I meant;

I have a new patch to implement it. But as I measure it using 'perf top', there is no any improvement. I see that the spin lock burns about 1.3% CPU time in my desktop with 2 Intel CPUs. Maybe a x86 server with 8 CPUs could highlight this improvement.

Quoted 10 lines

although note that depending on how many
items we need to drop, it might not be necesary to do a full sort of the list --- we could just look for the inode with the oldest time, which would be O(n) rather than O(n log n).

If we are under enough memory pressure that the shrinker is getting called a lot, maybe it would be worthwhile keep a list of the oldest N inodes, which only gets updated every so many seconds. (Since we can always validate whether or not an inode has been touched since it had been put on the oldest N inodes very easily.)

Yes, I will give this a try.

Quoted 9 lines
Quoted 3 lines

For creating all extent leaf blocks, that is another improvement. I will try to add it after I solve current problem that Dave reported.

Yes, it's a separable issue, but it's related to the same idea. We should drop items in the cache which can be easily instaniated at the time same time (i.e., drop all of the items related to a particular extent leaf block, or all of the items related to a particular inode).

Quoted 4 lines

There are of course, other hueristics we could use, such as only searching inodes which have open file descriptors. It's not clear the benefit is worth the cost of determining which inodes fit a particular criteria or not, though.

Yes, but as you said, we need to measure the benefit from it.

Quoted 13 lines

One other possibility while we're brainstorming is using hints from posix_fadvise or madvise to drop items from the extent cache. The cautions with this idea are (a) not that many applications provide those hints, and (b) just because one process declares that it doesn't anticipate needing to access a particular file, doesn't mean that it is true for other users of the same file. A related classical problem is the backup program which reads every single file in the file system once. There's no particular good reason to have the backup program's activities affect the caches in the system, whether it be the page cache or the extent cache. The trick is determining whether or not a particular program is behaving like a backup program or not, either via a hueristic or via an explicit hint/declaration from the program.

To be honest, it is difficulty for us to add a new flag in fadvise/madvise because other file systems might not provide this hint, and app A using this hint could affect app B as you described. But I agree with you that we could provide a interface to let application manage extent cache.

Regards,
- Zheng

#9 lore raw
In reply to Dave Hansen
Quoted 4 lines · Dave Hansen

I've got a test case which I intended to use to stress the VM a bit. It fills memory up with page cache a couple of times. It essentially runs 30 or so cp's in parallel.

Quoted 8 lines

98% of my CPU is system time, and 96% of _that_ is being spent on the spinlock in ext4_es_lru_add(). I think the LRU list head and its lock end up being *REALLY* hot cachelines and are *the* bottleneck on this test. Note that this is _before_ we go in to reclaim and actually start calling in to the shrinker. There is zero memory pressure in this test.

I'm not sure the benefits of having a proper in-order LRU during reclaim outweigh such a drastic downside for the common case.

Hi Dave,

I paste a patch that tries to fix this problem that you reported. I do the following test in my sand box, and it seeems that this patch could fix the problem. This patch is not very mature. But I hope to be sure that my direction is right. So I really appreciate if you could confirm that this patch can fix the problem in your test case.

I try to reproduce the test as you described. I creaete a ext4 file system that is a loopback device in a x86_64 server with 16 CPUs, 24G memory. Then I create 160 64G sparse files, and copy them to /dev/null. I use 'perf record/report' to observe the overhead, and the results are as below:

unpatched kernel
----------------
65.32% cp [kernel.kallsyms] [k] _raw_spin_lock

                         |
                         --- _raw_spin_lock
                            |          
                            |--99.30%-- ext4_es_lru_add
                            |          ext4_es_lookup_extent
                            |          ext4_map_blocks
                            |          _ext4_get_block
                            |          ext4_get_block
                            |          do_mpage_readpage

The result shows that ext4_es_lru_add() burns about 65% CPU time.

patched kernel
--------------
2.26% cp [kernel.kallsyms] [k] _raw_spin_lock

                         |
                         --- _raw_spin_lock
                            |          
                            |--96.36%-- rmqueue_bulk.clone.0
                            |          get_page_from_freelist
                            |          |          
                            |          |--60.13%-- __alloc_pages_nodemask

After applied the patch, ext4_es_lru_add() is never a bottleneck.

I am not sure whether I have reproduced your test. So that would be great if you can confirm that this patch can fix the problem. The patch bases against the master branch of ext4 tree.

Any comment or suggestion is welcome.

Thanks,
- Zheng

Subject: [PATCH] ext4: improve extent cache shrink mechanism to avoid to burn CPU time

From: Zheng Liu <[email protected]>

Now we maintain an proper in-order LRU list in ext4 to reclaim entries from extent status tree when we are under heavy memory pressure. For keeping this order, a spin lock is used to protect this list. But this lock burns a lot of CPU time.

This commit tries to fix this problem. Now a new member called i_touch_when is added into ext4_inode_info to record the last access time for an inode. Meanwhile we never need to keep a proper in-order LRU list. So this can avoid to burns some CPU time. When we try to reclaim some entries from extent status tree, we use list_sort() to get a proper in-order list. Then we traverse this list to discard some entries.

In this commit, we break the loop if s_extent_cache_cnt == 0 because that means that all extents in extent status tree have been reclaimed.

Reported-by Dave Hansen <[email protected]>
Signed-off-by Zheng Liu <[email protected]>
---
 fs/ext4/ext4.h           |    1 +
 fs/ext4/extents_status.c |   43 ++++++++++++++++++++++++++++++++++---------
 fs/ext4/inode.c          |    4 ++++
 fs/ext4/super.c          |    1 +
 4 files changed, 40 insertions(+), 9 deletions(-)
fs/ext4/ext4.h +1 -0
index 019db3c..1c31f27 100644
@@ -874,6 +874,7 @@ struct ext4_inode_info {
 	rwlock_t i_es_lock;
 	struct list_head i_es_lru;
 	unsigned int i_es_lru_nr;	/* protected by i_es_lock */
+	unsigned long i_touch_when;	/* jiffies of last accessing */
 
 	/* ialloc */
 	ext4_group_t	i_last_alloc_group;
fs/ext4/extents_status.c +34 -9

94 lines not on this page · raw message

fs/ext4/inode.c +4 -0
index 38f03dc..1477406 100644
@@ -571,6 +571,8 @@ int ext4_map_blocks(handle_t *handle, struct inode *inode,
 		  "logical block %lu\n", inode->i_ino, flags, map->m_len,
 		  (unsigned long) map->m_lblk);
 
+	ext4_es_lru_add(inode);
+
 	/* Lookup extent status tree firstly */
 	if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
 		if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {

@@ -1888,6 +1890,8 @@ static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
 		  "logical block %lu\n", inode->i_ino, map->m_len,
 		  (unsigned long) map->m_lblk);
 
+	ext4_es_lru_add(inode);
+
 	/* Lookup extent status tree firstly */
 	if (ext4_es_lookup_extent(inode, iblock, &es)) {
fs/ext4/super.c +1 -0
index a9c1438..b365695 100644
@@ -849,6 +849,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
 	rwlock_init(&ei->i_es_lock);
 	INIT_LIST_HEAD(&ei->i_es_lru);
 	ei->i_es_lru_nr = 0;
+	ei->i_touch_when = 0;
 	ei->i_reserved_data_blocks = 0;
 	ei->i_reserved_meta_blocks = 0;
 	ei->i_allocated_meta_blocks = 0;
Signature · 1 line
1.7.9.7
maintainer
#10 lore raw
Quoted 7 lines · Zheng Liu

This commit tries to fix this problem. Now a new member called i_touch_when is added into ext4_inode_info to record the last access time for an inode. Meanwhile we never need to keep a proper in-order LRU list. So this can avoid to burns some CPU time. When we try to reclaim some entries from extent status tree, we use list_sort() to get a proper in-order list. Then we traverse this list to discard some entries.

I think this approach is very sound. I have one reservation, however. If there are a large number of inodes on this list, list_sort() could burn a lot of CPU time, especially if the shrinker is called a very large number of times in a short time period (i.e., when the system is thrashing, or when one or more memory containers are under a large amount of memory pressure).

I have a suggestion for how to address this: Keep a timestamp of when the list last has been sorted in struct ext4_super_info. When iterating over the list, looking for a candidate inode, if inode's i_touch_when is greater than the last time sorted, skip the inode. If there are no inodes left in the list, or more than some threshold inodes have been skipped, only then resort the list (and update the timestamp in the ext4_super_info structure).

What do you think?

- Ted

#11 lore raw
Quoted 24 lines · Theodore Ts'o
Quoted 7 lines · Zheng Liu

This commit tries to fix this problem. Now a new member called i_touch_when is added into ext4_inode_info to record the last access time for an inode. Meanwhile we never need to keep a proper in-order LRU list. So this can avoid to burns some CPU time. When we try to reclaim some entries from extent status tree, we use list_sort() to get a proper in-order list. Then we traverse this list to discard some entries.

I think this approach is very sound. I have one reservation, however. If there are a large number of inodes on this list, list_sort() could burn a lot of CPU time, especially if the shrinker is called a very large number of times in a short time period (i.e., when the system is thrashing, or when one or more memory containers are under a large amount of memory pressure).

I have a suggestion for how to address this: Keep a timestamp of when the list last has been sorted in struct ext4_super_info. When iterating over the list, looking for a candidate inode, if inode's i_touch_when is greater than the last time sorted, skip the inode. If there are no inodes left in the list, or more than some threshold inodes have been skipped, only then resort the list (and update the timestamp in the ext4_super_info structure).

Quoted 1 line

What do you think?

Thanks for your suggestion. I fully agree with you. What I concern is how to define this threshold. A fixed value is very simple but too inflexible. If this value is too big, this lru list could never be sorted. So my proposal is that we set this value accroding to the number of used inodes. For example, if the number of used inodes is 10,000, this threshold could be set to 10 (10,000 x 0.1%). Meanwhile, we need to provide a sysfs interface to let user set this value. Does it make sense?

Thanks,
- Zheng

maintainer
#12 lore raw
Quoted 9 lines · Zheng Liu
Quoted 8 lines

I have a suggestion for how to address this: Keep a timestamp of when the list last has been sorted in struct ext4_super_info. When iterating over the list, looking for a candidate inode, if inode's i_touch_when is greater than the last time sorted, skip the inode. If there are no inodes left in the list, or more than some threshold inodes have been skipped, only then resort the list (and update the timestamp in the ext4_super_info structure).

Quoted 4 lines

Thanks for your suggestion. I fully agree with you. What I concern is how to define this threshold. A fixed value is very simple but too inflexible. If this value is too big, this lru list could never be sorted.

In my proposal, if the list is never sorted before, we would definitely sort the list the first time there is a request to shrink the number of extent caches. So the list would always be sorted at least once.

If an inode gets accessed after the time when the list has been sorted, then the last_touch time will be greater than the last_sorted time, right? And our goal is to use the least recently used inode, so as long as we know they have been used more recently than the time the list was sorted, and there are still inodes on the list where the last_touch time is older than the last_sort time, we don't have a problem. So we will evict all of the non-delalloc extent cache items from that inode, and remove it from the list.

The original reason for having the threshold at all is so that if we skip a huge number of inodes on the list, this would take a long time. But thinking about this some more, we don't even need a threshold. What we do instead is if the the last_touch time is newer than the last_sort time, we just move the inode to the end of the list. That means the end of the list will be unsorted, but that's OK, because the oldest inodes are still at the beginning of the list. Once the head of the list is newer than the last_sorted time, then we know we need to resort the entire list.

Does that make sense to you?

- Ted

#13 lore raw
Quoted 39 lines · Theodore Ts'o
Quoted 9 lines · Zheng Liu
Quoted 8 lines

I have a suggestion for how to address this: Keep a timestamp of when the list last has been sorted in struct ext4_super_info. When iterating over the list, looking for a candidate inode, if inode's i_touch_when is greater than the last time sorted, skip the inode. If there are no inodes left in the list, or more than some threshold inodes have been skipped, only then resort the list (and update the timestamp in the ext4_super_info structure).

Quoted 4 lines

Thanks for your suggestion. I fully agree with you. What I concern is how to define this threshold. A fixed value is very simple but too inflexible. If this value is too big, this lru list could never be sorted.

In my proposal, if the list is never sorted before, we would definitely sort the list the first time there is a request to shrink the number of extent caches. So the list would always be sorted at least once.

If an inode gets accessed after the time when the list has been sorted, then the last_touch time will be greater than the last_sorted time, right? And our goal is to use the least recently used inode, so as long as we know they have been used more recently than the time the list was sorted, and there are still inodes on the list where the last_touch time is older than the last_sort time, we don't have a problem. So we will evict all of the non-delalloc extent cache items from that inode, and remove it from the list.

The original reason for having the threshold at all is so that if we skip a huge number of inodes on the list, this would take a long time. But thinking about this some more, we don't even need a threshold. What we do instead is if the the last_touch time is newer than the last_sort time, we just move the inode to the end of the list. That means the end of the list will be unsorted, but that's OK, because the oldest inodes are still at the beginning of the list. Once the head of the list is newer than the last_sorted time, then we know we need to resort the entire list.

Quoted 1 line

Does that make sense to you?

Hi Ted and Dave,

Sorry for the late. It makes sense to me. Here is the second version of patch. Could you please review it?

Dave, that would be great if you could do your testing again to confirm this patch is useful.

Any feedback is welcome.

Thanks,
- Zheng

Subject: [PATCH v2] ext4: improve extent cache shrink mechanism to avoid to burn CPU time

From: Zheng Liu <[email protected]>

Now we maintain an proper in-order LRU list in ext4 to reclaim entries from extent status tree when we are under heavy memory pressure. For keeping this order, a spin lock is used to protect this list. But this lock burns a lot of CPU time. We can use the following steps to trigger it.

% cd /dev/shm
% dd if=/dev/zero of=ext4-img bs=1M count=2k % mkfs.ext4 ext4-img
% mount -t ext4 -o loop ext4-img /mnt
% cd /mnt
% for ((i=0;i<160;i++)); do truncate -s 64g $i; done % for ((i=0;i<160;i++)); do cp $i /dev/null &; done % perf record -a -g
% perf report

This commit tries to fix this problem. Now a new member called i_touch_when is added into ext4_inode_info to record the last access time for an inode. Meanwhile we never need to keep a proper in-order LRU list. So this can avoid to burns some CPU time. When we try to reclaim some entries from extent status tree, we use list_sort() to get a proper in-order list. Then we traverse this list to discard some entries. In ext4_sb_info, we use s_es_last_sorted to record the last time of sorting this list. When we traverse the list, we skip the inode that is newer than this time, and move this inode to the tail of LRU list. When the head of the list is newer than s_es_last_sorted, we will sort the LRU list again.

In this commit, we break the loop if s_extent_cache_cnt == 0 because that means that all extents in extent status tree have been reclaimed.

Meanwhile in this commit, ext4_es_{un}register_shrinker()'s prototype is changed to save a local variable in these functions.

Reported-by Dave Hansen <[email protected]>
Cc "Theodore Ts'o" <[email protected]>
Signed-off-by Zheng Liu <[email protected]>

---
changelog:
v2:
* avoid to sort the LRU list
* change ext4_es_{un}register_shrinker()'s prototype to save a local variable

 fs/ext4/ext4.h           |    2 ++
 fs/ext4/extents_status.c |   77 ++++++++++++++++++++++++++++++++++------------
 fs/ext4/extents_status.h |    5 +--
 fs/ext4/inode.c          |    4 +++
 fs/ext4/super.c          |    7 +++--
 5 files changed, 70 insertions(+), 25 deletions(-)
fs/ext4/ext4.h +2 -0
index 019db3c..f30ff8c 100644
@@ -874,6 +874,7 @@ struct ext4_inode_info {
 	rwlock_t i_es_lock;
 	struct list_head i_es_lru;
 	unsigned int i_es_lru_nr;	/* protected by i_es_lock */
+	unsigned long i_touch_when;	/* jiffies of last accessing */
 
 	/* ialloc */
 	ext4_group_t	i_last_alloc_group;

@@ -1302,6 +1303,7 @@ struct ext4_sb_info {
 	/* Reclaim extents from extent status tree */
 	struct shrinker s_es_shrinker;
 	struct list_head s_es_lru;
+	unsigned long s_es_last_sorted;
 	struct percpu_counter s_extent_cache_cnt;
 	spinlock_t s_es_lru_lock ____cacheline_aligned_in_smp;
 };
fs/ext4/extents_status.c +57 -20

158 lines not on this page · raw message

fs/ext4/extents_status.h +3 -2
index f740eb03..e936730 100644
@@ -39,6 +39,7 @@
 				 EXTENT_STATUS_DELAYED | \
 				 EXTENT_STATUS_HOLE)
 
+struct ext4_sb_info;
 struct ext4_extent;
 
 struct extent_status {

@@ -119,8 +120,8 @@ static inline void ext4_es_store_status(struct extent_status *es,
 	es->es_pblk = block;
 }
 
-extern void ext4_es_register_shrinker(struct super_block *sb);
-extern void ext4_es_unregister_shrinker(struct super_block *sb);
+extern void ext4_es_register_shrinker(struct ext4_sb_info *sbi);
+extern void ext4_es_unregister_shrinker(struct ext4_sb_info *sbi);
 extern void ext4_es_lru_add(struct inode *inode);
 extern void ext4_es_lru_del(struct inode *inode);
fs/ext4/inode.c +4 -0
index 38f03dc..1477406 100644
@@ -571,6 +571,8 @@ int ext4_map_blocks(handle_t *handle, struct inode *inode,
 		  "logical block %lu\n", inode->i_ino, flags, map->m_len,
 		  (unsigned long) map->m_lblk);
 
+	ext4_es_lru_add(inode);
+
 	/* Lookup extent status tree firstly */
 	if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
 		if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {

@@ -1888,6 +1890,8 @@ static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
 		  "logical block %lu\n", inode->i_ino, map->m_len,
 		  (unsigned long) map->m_lblk);
 
+	ext4_es_lru_add(inode);
+
 	/* Lookup extent status tree firstly */
 	if (ext4_es_lookup_extent(inode, iblock, &es)) {
fs/ext4/super.c +4 -3
index a9c1438..a7054e7 100644
@@ -760,7 +760,7 @@ static void ext4_put_super(struct super_block *sb)
 			ext4_abort(sb, "Couldn't clean up the journal");
 	}
 
-	ext4_es_unregister_shrinker(sb);
+	ext4_es_unregister_shrinker(sbi);
 	del_timer(&sbi->s_err_report);
 	ext4_release_system_zone(sb);
 	ext4_mb_release(sb);

@@ -849,6 +849,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
 	rwlock_init(&ei->i_es_lock);
 	INIT_LIST_HEAD(&ei->i_es_lru);
 	ei->i_es_lru_nr = 0;
+	ei->i_touch_when = 0;
 	ei->i_reserved_data_blocks = 0;
 	ei->i_reserved_meta_blocks = 0;
 	ei->i_allocated_meta_blocks = 0;

@@ -3766,7 +3767,7 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent)
 	sbi->s_err_report.data = (unsigned long) sb;
 
 	/* Register extent status tree shrinker */
-	ext4_es_register_shrinker(sb);
+	ext4_es_register_shrinker(sbi);
 
 	err = percpu_counter_init(&sbi->s_freeclusters_counter,
 			ext4_count_free_clusters(sb));

@@ -4084,7 +4085,7 @@ failed_mount_wq:
 		sbi->s_journal = NULL;
 	}
 failed_mount3:
-	ext4_es_unregister_shrinker(sb);
+	ext4_es_unregister_shrinker(sbi);
 	del_timer(&sbi->s_err_report);
 	if (sbi->s_flex_groups)
 		ext4_kvfree(sbi->s_flex_groups);
Signature · 1 line
1.7.9.7
author
#14 lore raw
Quoted 2 lines · Zheng Liu

Dave, that would be great if you could do your testing again to confirm this patch is useful.

I was able to apply this to Ted's

https://git.kernel.org/cgit/linux/kernel/git/tytso/ext4.git/

"ext4/dev" tree with a bit of fuzz. What did you generate this against, btw?

It does seem to be functioning OK for me, and passes the other tests that saw so much spinlock contention previously. You can add my Tested-by if you like.

#15 lore raw
Quoted 10 lines · Dave Hansen
Quoted 2 lines · Zheng Liu

Dave, that would be great if you could do your testing again to confirm this patch is useful.

I was able to apply this to Ted's

https://git.kernel.org/cgit/linux/kernel/git/tytso/ext4.git/

"ext4/dev" tree with a bit of fuzz. What did you generate this against, btw?

Ah, sorry, I forgot to mention that this patch bases against ext4/master branch. Now ext4/dev branch has some regression when I run xfstests. So I don't base against this branch to generate my patch.

Ted, I notice that now in ext4 tree we have 'dev', 'dev-with-revert', and 'dev2' branches. Which one is the best to generate a new patch for the next merge window?

Quoted 4 lines

It does seem to be functioning OK for me, and passes the other tests that saw so much spinlock contention previously. You can add my Tested-by if you like.

Thanks for your testing.

Regards,
- Zheng

maintainer
#16 lore raw
Quoted 2 lines · Zheng Liu

Ah, sorry, I forgot to mention that this patch bases against ext4/master branch. Now ext4/dev branch has some regression when I run xfstests.

What regressions are you seeing?

Quoted 3 lines

Ted, I notice that now in ext4 tree we have 'dev', 'dev-with-revert', and 'dev2' branches. Which one is the best to generate a new patch for the next merge window?

Either the dev branch or the master branch.

The dev-with-revert and dev2 were branches that I had created when investigating a potential regression with the invalidage page range patch set. I've since determined that it's a timing issue and it's not a new regression --- we've had xfstests failures with test generic/300 for a while now.

- Ted

#17 lore raw
Quoted 5 lines · Theodore Ts'o
Quoted 2 lines · Zheng Liu

Ah, sorry, I forgot to mention that this patch bases against ext4/master branch. Now ext4/dev branch has some regression when I run xfstests.

What regressions are you seeing?

generic/300.

When I try to test my patch, I know that there has a report that invalidate page range patch set causes a regression, and I am not sure whether invalidate page range patch set causes it or not. So I decide to generate my patch against ext4/master. So, don't worry. :-)

BTW, I will run xfstests this week. If I meet any regression, I will let you know.

Quoted 5 lines
Quoted 4 lines

Ted, I notice that now in ext4 tree we have 'dev', 'dev-with-revert', and 'dev2' branches. Which one is the best to generate a new patch for the next merge window?

Quoted 7 lines

Either the dev branch or the master branch.

The dev-with-revert and dev2 were branches that I had created when investigating a potential regression with the invalidage page range patch set. I've since determined that it's a timing issue and it's not a new regression --- we've had xfstests failures with test generic/300 for a while now.

Thanks for pointing it out.

- Zheng

maintainer Applied
#18 lore raw
In reply to Zheng Liu
Quoted 35 lines

Subject: [PATCH v2] ext4: improve extent cache shrink mechanism to avoid to burn CPU time

From: Zheng Liu <[email protected]>

Now we maintain an proper in-order LRU list in ext4 to reclaim entries from extent status tree when we are under heavy memory pressure. For keeping this order, a spin lock is used to protect this list. But this lock burns a lot of CPU time. We can use the following steps to trigger it.

% cd /dev/shm
% dd if=/dev/zero of=ext4-img bs=1M count=2k % mkfs.ext4 ext4-img
% mount -t ext4 -o loop ext4-img /mnt
% cd /mnt
% for ((i=0;i<160;i++)); do truncate -s 64g $i; done % for ((i=0;i<160;i++)); do cp $i /dev/null &; done % perf record -a -g
% perf report

This commit tries to fix this problem. Now a new member called i_touch_when is added into ext4_inode_info to record the last access time for an inode. Meanwhile we never need to keep a proper in-order LRU list. So this can avoid to burns some CPU time. When we try to reclaim some entries from extent status tree, we use list_sort() to get a proper in-order list. Then we traverse this list to discard some entries. In ext4_sb_info, we use s_es_last_sorted to record the last time of sorting this list. When we traverse the list, we skip the inode that is newer than this time, and move this inode to the tail of LRU list. When the head of the list is newer than s_es_last_sorted, we will sort the LRU list again.

In this commit, we break the loop if s_extent_cache_cnt == 0 because that means that all extents in extent status tree have been reclaimed.

Quoted 6 lines

Meanwhile in this commit, ext4_es_{un}register_shrinker()'s prototype is changed to save a local variable in these functions.

Reported-by Dave Hansen <[email protected]>
Cc "Theodore Ts'o" <[email protected]>
Signed-off-by Zheng Liu <[email protected]>

Applied, thanks.

- Ted

author
#19 lore raw
In reply to Zheng Liu

fs/ext4/extents_status.c static int ext4_es_shrink(struct shrinker *shrink, struct shrink_control *sc) · line 914

Quoted 5 lines · Zheng Liu
fs/ext4/extents_status.c +1 -2
-	INIT_LIST_HEAD(&scanned);
-
 	spin_lock(&sbi->s_es_lru_lock);
+	list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
 	list_for_each_safe(cur, tmp, &sbi->s_es_lru) {

How long can this list get? I have the feeling this might get a bit painful, especially on a NUMA machine.

But, it definitely eliminates the spinlock contention that I was seeing.
The top ext4 function in my profiles is way down under 1% of CPU time
now. Thanks for the quick response, and please let me know if you need any further testing.

#20 lore raw
Quoted 9 lines · Dave Hansen

fs/ext4/extents_status.c static int ext4_es_shrink(struct shrinker *shrink, struct shrink_control *sc) · line 914

Quoted 5 lines · Zheng Liu
fs/ext4/extents_status.c +1 -2
-	INIT_LIST_HEAD(&scanned);
-
 	spin_lock(&sbi->s_es_lru_lock);
+	list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
 	list_for_each_safe(cur, tmp, &sbi->s_es_lru) {

How long can this list get? I have the feeling this might get a bit painful, especially on a NUMA machine.

I guess that you worry about the time of sorting a lru list, right? Ted and I also worry about this, especially when shrinker is called() very frequently. Ted has presented a solution. I will give it a try.

Quoted 5 lines

But, it definitely eliminates the spinlock contention that I was seeing.
The top ext4 function in my profiles is way down under 1% of CPU time
now. Thanks for the quick response, and please let me know if you need any further testing.

Thanks for your help. As I said above, I will try to improve this patch later. My server's numa has been turned off in BIOS and I haven't a privilege to turn it on. So that would be great if you could help me to do some further testing.

Thanks in advance,
- Zheng

author
#21 lore raw
Quoted 11 lines · Zheng Liu
Quoted 9 lines · Dave Hansen
Quoted 9 lines · Zheng Liu
Quoted 5 lines

fs/ext4/extents_status.c static int ext4_es_shrink(struct shrinker *shrink, struct shrink_control *sc) · line 914

Quoted 5 lines
fs/ext4/extents_status.c +1 -2
-	INIT_LIST_HEAD(&scanned);
-
 	spin_lock(&sbi->s_es_lru_lock);
+	list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
 	list_for_each_safe(cur, tmp, &sbi->s_es_lru) {

How long can this list get? I have the feeling this might get a bit painful, especially on a NUMA machine.

I guess that you worry about the time of sorting a lru list, right?

Yeah, just worried about the amount of time it takes to examine (and possibly write to) every item on the list.