Recently, we have observed some janky issues caused by unpleasantly long
contention on mmap_lock which is held by smaps_rollup when probing large
processes. To address the problem, we let smaps_rollup detect if anyone
wants to acquire mmap_lock for write attempts. If yes, just release the
lock temporarily to ease the contention.
smaps_rollup is a procfs interface which allows users to summarize the
process's memory usage without the overhead of seq_* calls. Android uses
it to sample the memory usage of various processes to balance its memory
pool sizes. If no one wants to take the lock for write requests, smaps_rollup
with this patch will behave like the original one.
Although there are on-going mmap_lock optimizations like range-based locks,
the lock applied to smaps_rollup would be the coarse one, which is hard to
avoid the occurrence of aforementioned issues. So the detection and temporary
release for write attempts on mmap_lock in smaps_rollup is still necessary.
Chinwen Chang (2):
mmap locking API: add mmap_lock_is_contended()
mm: proc: smaps_rollup: do not stall write attempts on mmap_lock
fs/proc/task_mmu.c | 21 +++++++++++++++++++++
include/linux/mmap_lock.h | 5 +++++
2 files changed, 26 insertions(+)
smaps_rollup will try to grab mmap_lock and go through the whole vma
list until it finishes the iterating. When encountering large processes,
the mmap_lock will be held for a longer time, which may block other
write requests like mmap and munmap from progressing smoothly.
There are upcoming mmap_lock optimizations like range-based locks, but
the lock applied to smaps_rollup would be the coarse type, which doesn't
avoid the occurrence of unpleasant contention.
To solve aforementioned issue, we add a check which detects whether
anyone wants to grab mmap_lock for write attempts.
Signed-off-by: Chinwen Chang <[email protected]>
---
fs/proc/task_mmu.c | 21 +++++++++++++++++++++
1 file changed, 21 insertions(+)
diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index dbda449..4b51f25 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -856,6 +856,27 @@ static int show_smaps_rollup(struct seq_file *m, void *v)
for (vma = priv->mm->mmap; vma; vma = vma->vm_next) {
smap_gather_stats(vma, &mss);
last_vma_end = vma->vm_end;
+
+ /*
+ * Release mmap_lock temporarily if someone wants to
+ * access it for write request.
+ */
+ if (mmap_lock_is_contended(mm)) {
+ mmap_read_unlock(mm);
+ ret = mmap_read_lock_killable(mm);
+ if (ret) {
+ release_task_mempolicy(priv);
+ goto out_put_mm;
+ }
+
+ /* Check whether current vma is available */
+ vma = find_vma(mm, last_vma_end - 1);
+ if (vma && vma->vm_start < last_vma_end)
+ continue;
+
+ /* Current vma is not available, just break */
+ break;
+ }
}
show_vma_header_prefix(m, priv->mm->mmap->vm_start,
--
1.9.1
Add new API to query if someone wants to acquire mmap_lock
for write attempts.
Using this instead of rwsem_is_contended makes it more tolerant
of future changes to the lock type.
Signed-off-by: Chinwen Chang <[email protected]>
---
include/linux/mmap_lock.h | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/include/linux/mmap_lock.h b/include/linux/mmap_lock.h
index 0707671..18e7eae 100644
--- a/include/linux/mmap_lock.h
+++ b/include/linux/mmap_lock.h
@@ -87,4 +87,9 @@ static inline void mmap_assert_write_locked(struct mm_struct *mm)
VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_lock), mm);
}
+static inline int mmap_lock_is_contended(struct mm_struct *mm)
+{
+ return rwsem_is_contended(&mm->mmap_lock);
+}
+
#endif /* _LINUX_MMAP_LOCK_H */
--
1.9.1
On 11/08/2020 05:42, Chinwen Chang wrote:
> smaps_rollup will try to grab mmap_lock and go through the whole vma
> list until it finishes the iterating. When encountering large processes,
> the mmap_lock will be held for a longer time, which may block other
> write requests like mmap and munmap from progressing smoothly.
>
> There are upcoming mmap_lock optimizations like range-based locks, but
> the lock applied to smaps_rollup would be the coarse type, which doesn't
> avoid the occurrence of unpleasant contention.
>
> To solve aforementioned issue, we add a check which detects whether
> anyone wants to grab mmap_lock for write attempts.
>
> Signed-off-by: Chinwen Chang <[email protected]>
> ---
> fs/proc/task_mmu.c | 21 +++++++++++++++++++++
> 1 file changed, 21 insertions(+)
>
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index dbda449..4b51f25 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -856,6 +856,27 @@ static int show_smaps_rollup(struct seq_file *m, void *v)
> for (vma = priv->mm->mmap; vma; vma = vma->vm_next) {
> smap_gather_stats(vma, &mss);
> last_vma_end = vma->vm_end;
> +
> + /*
> + * Release mmap_lock temporarily if someone wants to
> + * access it for write request.
> + */
> + if (mmap_lock_is_contended(mm)) {
> + mmap_read_unlock(mm);
> + ret = mmap_read_lock_killable(mm);
> + if (ret) {
> + release_task_mempolicy(priv);
> + goto out_put_mm;
> + }
> +
> + /* Check whether current vma is available */
> + vma = find_vma(mm, last_vma_end - 1);
> + if (vma && vma->vm_start < last_vma_end)
I may be wrong, but this looks like it could return incorrect results.
For example if we start reading with the following VMAs:
+------+------+-----------+
| VMA1 | VMA2 | VMA3 |
+------+------+-----------+
| | | |
4k 8k 16k 400k
Then after reading VMA2 we drop the lock due to contention. So:
last_vma_end = 16k
Then if VMA2 is freed while the lock is dropped, so we have:
+------+ +-----------+
| VMA1 | | VMA3 |
+------+ +-----------+
| | | |
4k 8k 16k 400k
find_vma(mm, 16k-1) will then return VMA3 and the condition vm_start <
last_vma_end will be false.
> + continue;
> +
> + /* Current vma is not available, just break */
> + break;
Which means we break out here and report an incomplete output (the
numbers will be much smaller than reality).
Would it be better to have a loop like:
for (vma = priv->mm->mmap; vma;) {
smap_gather_stats(vma, &mss);
last_vma_end = vma->vm_end;
if (contended) {
/* drop/acquire lock */
vma = find_vma(mm, last_vma_end - 1);
if (!vma)
break;
if (vma->vm_start >= last_vma_end)
continue;
}
vma = vma->vm_next;
}
that way if the VMA is removed while the lock is dropped the loop can
just continue from the next VMA.
Or perhaps I missed something obvious? I haven't actually tested
anything above.
Steve
> + }
> }
>
> show_vma_header_prefix(m, priv->mm->mmap->vm_start,
>
On Wed, 2020-08-12 at 09:39 +0100, Steven Price wrote:
> On 11/08/2020 05:42, Chinwen Chang wrote:
> > smaps_rollup will try to grab mmap_lock and go through the whole vma
> > list until it finishes the iterating. When encountering large processes,
> > the mmap_lock will be held for a longer time, which may block other
> > write requests like mmap and munmap from progressing smoothly.
> >
> > There are upcoming mmap_lock optimizations like range-based locks, but
> > the lock applied to smaps_rollup would be the coarse type, which doesn't
> > avoid the occurrence of unpleasant contention.
> >
> > To solve aforementioned issue, we add a check which detects whether
> > anyone wants to grab mmap_lock for write attempts.
> >
> > Signed-off-by: Chinwen Chang <[email protected]>
> > ---
> > fs/proc/task_mmu.c | 21 +++++++++++++++++++++
> > 1 file changed, 21 insertions(+)
> >
> > diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> > index dbda449..4b51f25 100644
> > --- a/fs/proc/task_mmu.c
> > +++ b/fs/proc/task_mmu.c
> > @@ -856,6 +856,27 @@ static int show_smaps_rollup(struct seq_file *m, void *v)
> > for (vma = priv->mm->mmap; vma; vma = vma->vm_next) {
> > smap_gather_stats(vma, &mss);
> > last_vma_end = vma->vm_end;
> > +
> > + /*
> > + * Release mmap_lock temporarily if someone wants to
> > + * access it for write request.
> > + */
> > + if (mmap_lock_is_contended(mm)) {
> > + mmap_read_unlock(mm);
> > + ret = mmap_read_lock_killable(mm);
> > + if (ret) {
> > + release_task_mempolicy(priv);
> > + goto out_put_mm;
> > + }
> > +
> > + /* Check whether current vma is available */
> > + vma = find_vma(mm, last_vma_end - 1);
> > + if (vma && vma->vm_start < last_vma_end)
>
> I may be wrong, but this looks like it could return incorrect results.
> For example if we start reading with the following VMAs:
>
> +------+------+-----------+
> | VMA1 | VMA2 | VMA3 |
> +------+------+-----------+
> | | | |
> 4k 8k 16k 400k
>
> Then after reading VMA2 we drop the lock due to contention. So:
>
> last_vma_end = 16k
>
> Then if VMA2 is freed while the lock is dropped, so we have:
>
> +------+ +-----------+
> | VMA1 | | VMA3 |
> +------+ +-----------+
> | | | |
> 4k 8k 16k 400k
>
> find_vma(mm, 16k-1) will then return VMA3 and the condition vm_start <
> last_vma_end will be false.
>
Hi Steve,
Thank you for reviewing this patch.
You are correct. If the contention is detected and the current vma(here
is VMA2) is freed while the lock is dropped, it will report an
incomplete result.
> > + continue;
> > +
> > + /* Current vma is not available, just break */
> > + break;
>
> Which means we break out here and report an incomplete output (the
> numbers will be much smaller than reality).
>
> Would it be better to have a loop like:
>
> for (vma = priv->mm->mmap; vma;) {
> smap_gather_stats(vma, &mss);
> last_vma_end = vma->vm_end;
>
> if (contended) {
> /* drop/acquire lock */
>
> vma = find_vma(mm, last_vma_end - 1);
> if (!vma)
> break;
> if (vma->vm_start >= last_vma_end)
> continue;
> }
> vma = vma->vm_next;
> }
>
> that way if the VMA is removed while the lock is dropped the loop can
> just continue from the next VMA.
>
Thanks a lot for your great suggestion.
> Or perhaps I missed something obvious? I haven't actually tested
> anything above.
>
> Steve
I will prepare new patch series for further reviews.
Thank you.
Chinwen
>
> > + }
> > }
> >
> > show_vma_header_prefix(m, priv->mm->mmap->vm_start,
> >
>