[RFC,v8,3/5] mm: mmap: zap pages with read mmap_sem in munmap

Commit Message

Yang Shi Aug. 15, 2018, 6:49 p.m. UTC

When running some mmap/munmap scalability tests with large memory (i.e.
> 300GB), the below hung task issue may happen occasionally.

INFO: task ps:14018 blocked for more than 120 seconds.
       Tainted: G            E 4.9.79-009.ali3000.alios7.x86_64 #1
 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
 ps              D    0 14018      1 0x00000004
  ffff885582f84000 ffff885e8682f000 ffff880972943000 ffff885ebf499bc0
  ffff8828ee120000 ffffc900349bfca8 ffffffff817154d0 0000000000000040
  00ffffff812f872a ffff885ebf499bc0 024000d000948300 ffff880972943000
 Call Trace:
  [<ffffffff817154d0>] ? __schedule+0x250/0x730
  [<ffffffff817159e6>] schedule+0x36/0x80
  [<ffffffff81718560>] rwsem_down_read_failed+0xf0/0x150
  [<ffffffff81390a28>] call_rwsem_down_read_failed+0x18/0x30
  [<ffffffff81717db0>] down_read+0x20/0x40
  [<ffffffff812b9439>] proc_pid_cmdline_read+0xd9/0x4e0
  [<ffffffff81253c95>] ? do_filp_open+0xa5/0x100
  [<ffffffff81241d87>] __vfs_read+0x37/0x150
  [<ffffffff812f824b>] ? security_file_permission+0x9b/0xc0
  [<ffffffff81242266>] vfs_read+0x96/0x130
  [<ffffffff812437b5>] SyS_read+0x55/0xc0
  [<ffffffff8171a6da>] entry_SYSCALL_64_fastpath+0x1a/0xc5

It is because munmap holds mmap_sem exclusively from very beginning to
all the way down to the end, and doesn't release it in the middle. When
unmapping large mapping, it may take long time (take ~18 seconds to
unmap 320GB mapping with every single page mapped on an idle machine).

Zapping pages is the most time consuming part, according to the
suggestion from Michal Hocko [1], zapping pages can be done with holding
read mmap_sem, like what MADV_DONTNEED does. Then re-acquire write
mmap_sem to cleanup vmas.

But, some part may need write mmap_sem, for example, vma splitting. So,
the design is as follows:
        acquire write mmap_sem
        lookup vmas (find and split vmas)
        deal with special mappings
        detach vmas
        downgrade_write

        zap pages
        free page tables
        release mmap_sem

The vm events with read mmap_sem may come in during page zapping, but
since vmas have been detached before, they, i.e. page fault, gup, etc,
will not be able to find valid vma, then just return SIGSEGV or -EFAULT
as expected.

If the vma has VM_HUGETLB | VM_PFNMAP or uprobe, they are considered as
special mappings. They will be handled by falling back to regular
do_munmap() with exclusive mmap_sem held in this patch since they may
update vm flags.
But, with the "detach vmas first" approach, the vmas have been detached
when vm flags are updated, so it sounds safe to update vm flags with
read mmap_sem for this specific case. So, VM_HUGETLB and VM_PFNMAP will
be handled by using the optimized path in the following separate patches
for bisectable sake. However, uprobes mappings will keep using regular
do_munmap() since unmapping uprobe areas may need update mm flags. It
might be not safe with just holding read mmap_sem even though affected
vmas have been detached.

With the "detach vmas first" approach we don't have to re-acquire
mmap_sem again to clean up vmas to avoid race window which might get the
address space changed since downgrade_write() doesn't release the lock
to lead regression, which simply downgrades to read lock.

And, since the lock acquire/release cost is managed to the minimum and
almost as same as before, the optimization could be extended to any size
of mapping without incurring significant penalty to small mappings.

For the time being, just do this in munmap syscall path. Other
vm_munmap() or do_munmap() call sites (i.e mmap, mremap, etc) remain
intact due to some implementation difficulties since they acquire write
mmap_sem from very beginning and hold it until the end, do_munmap()
might be called in the middle. But, the optimized do_munmap would like
to be called without mmap_sem held so that we can do the optimization.
So, if we want to do the similar optimization for mmap/mremap path, I'm
afraid we would have to redesign them. mremap might be called on very
large area depending on the usecases, the optimization to it will be
considered in the future.

With the patches, exclusive mmap_sem hold time when munmap a 80GB
address space on a machine with 32 cores of E5-2680 @ 2.70GHz dropped to
us level from second.

munmap_test-15002 [008]   594.380138: funcgraph_entry: |
vm_munmap_zap_rlock() {
munmap_test-15002 [008]   594.380146: funcgraph_entry:      !2485684 us
|    unmap_region();
munmap_test-15002 [008]   596.865836: funcgraph_exit:       !2485692 us
|  }

Here the excution time of unmap_region() is used to evaluate the time of
holding read mmap_sem, then the remaining time is used with holding
exclusive lock.

[1] https://lwn.net/Articles/753269/

Suggested-by: Michal Hocko <[e-mail]>
Suggested-by: Kirill A. Shutemov <[e-mail]>
Cc: Matthew Wilcox <[e-mail]>
Cc: Laurent Dufour <[e-mail]>
Cc: Vlastimil Babka <[e-mail]>
Cc: Andrew Morton <[e-mail]>
Signed-off-by: Yang Shi <[e-mail]>
---
 mm/mmap.c | 97 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 95 insertions(+), 2 deletions(-)

Comments

Matthew Wilcox (Oracle) Aug. 15, 2018, 7:16 p.m. UTC | #1

On Thu, Aug 16, 2018 at 02:49:48AM +0800, Yang Shi wrote:
> +static int do_munmap_zap_rlock(struct mm_struct *mm, unsigned long start,
> +			       size_t len, struct list_head *uf)
> +{
> +	unsigned long end;
> +	struct vm_area_struct *start_vma, *prev, *vma;
> +	int ret = 0;
> +
> +	if (!addr_ok(start, len))
> +		return -EINVAL;
> +
> +	len = PAGE_ALIGN(len);
> +
> +	end = start + len;
> +
> +	/*
> +	 * Need write mmap_sem to split vmas and detach vmas
> +	 * splitting vma up-front to save PITA to clean if it is failed
> +	 */
> +	if (down_write_killable(&mm->mmap_sem))
> +		return -EINTR;
> +
> +	start_vma = munmap_lookup_vma(mm, start, end);
> +	if (!start_vma)
> +		goto out;
> +	if (IS_ERR(start_vma)) {
> +		ret = PTR_ERR(start_vma);
> +		goto out;
> +	}
> +
> +	prev = start_vma->vm_prev;
> +
> +	if (unlikely(uf)) {
> +		ret = userfaultfd_unmap_prep(start_vma, start, end, uf);
> +		if (ret)
> +			goto out;
> +	}
> +
> +	/*
> +	 * Unmapping vmas, which have:
> +	 *   VM_HUGETLB or
> +	 *   VM_PFNMAP or
> +	 *   uprobes
> +	 * need get done with write mmap_sem held since they may update
> +	 * vm_flags. Deal with such mappings with regular do_munmap() call.
> +	 */
> +	for (vma = start_vma; vma && vma->vm_start < end; vma = vma->vm_next) {
> +		if ((vma->vm_file &&
> +		    has_uprobes(vma, vma->vm_start, vma->vm_end)) ||
> +		    (vma->vm_flags & (VM_HUGETLB | VM_PFNMAP)))
> +			goto regular_path;

but ... that's going to redo all the work you already did!  Why not just this:

(not even compiled, and I can see a good opportunity for combining the
VM_LOCKED loop with the has_uprobes loop)

diff --git a/mm/mmap.c b/mm/mmap.c
index de699523c0b7..8d121db36efc 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2803,6 +2803,8 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 {
 	unsigned long end;
 	struct vm_area_struct *vma, *prev, *last;
+	int res = 0;
+	bool downgrade = false;
 
 	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
 		return -EINVAL;
@@ -2811,17 +2813,20 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	if (len == 0)
 		return -EINVAL;
 
+	if (down_write_killable(&mm->mmap_sem))
+		return -EINTR;
+
 	/* Find the first overlapping VMA */
 	vma = find_vma(mm, start);
 	if (!vma)
-		return 0;
+		goto unlock;
 	prev = vma->vm_prev;
-	/* we have  start < vma->vm_end  */
+	/* we have start < vma->vm_end  */
 
 	/* if it doesn't overlap, we have nothing.. */
 	end = start + len;
 	if (vma->vm_start >= end)
-		return 0;
+		goto unlock;
 
 	/*
 	 * If we need to split any vma, do it now to save pain later.
@@ -2831,28 +2836,27 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	 * places tmp vma above, and higher split_vma places tmp vma below.
 	 */
 	if (start > vma->vm_start) {
-		int error;
-
 		/*
 		 * Make sure that map_count on return from munmap() will
 		 * not exceed its limit; but let map_count go just above
 		 * its limit temporarily, to help free resources as expected.
 		 */
+		res = -ENOMEM
 		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
-			return -ENOMEM;
+			goto unlock;
 
-		error = __split_vma(mm, vma, start, 0);
-		if (error)
-			return error;
+		res = __split_vma(mm, vma, start, 0);
+		if (res)
+			goto unlock;
 		prev = vma;
 	}
 
 	/* Does it split the last one? */
 	last = find_vma(mm, end);
 	if (last && end > last->vm_start) {
-		int error = __split_vma(mm, last, end, 1);
-		if (error)
-			return error;
+		res = __split_vma(mm, last, end, 1);
+		if (res)
+			goto unlock;
 	}
 	vma = prev ? prev->vm_next : mm->mmap;
 
@@ -2866,9 +2870,19 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 		 * split, despite we could. This is unlikely enough
 		 * failure that it's not worth optimizing it for.
 		 */
-		int error = userfaultfd_unmap_prep(vma, start, end, uf);
-		if (error)
-			return error;
+		result = userfaultfd_unmap_prep(vma, start, end, uf);
+		if (result)
+			goto unlock;
+	}
+
+	downgrade = true;
+
+	for (vma = start_vma; vma && vma->vm_start < end; vma = vma->vm_next) {
+		if (vma->vm_file &&
+				has_uprobes(vma, vma->vm_start, vma->vm_end)) {
+			downgrade = false;
+			break;
+		}
 	}
 
 	/*
@@ -2885,6 +2899,9 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 		}
 	}
 
+	if (downgrade)
+		downgrade_write(&mm->mmap_sem);
+
 	/*
 	 * Remove the vma's, and unmap the actual pages
 	 */
@@ -2896,7 +2913,14 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	/* Fix up all other VM information */
 	remove_vma_list(mm, vma);
 
-	return 0;
+	res = 0;
+unlock:
+	if (downgrade) {
+		up_read(&mm->mmap_sem);
+	} else {
+		up_write(&mm->mmap_sem);
+	}
+	return res;
 }
 
 int vm_munmap(unsigned long start, size_t len)
@@ -2905,11 +2929,7 @@ int vm_munmap(unsigned long start, size_t len)
 	struct mm_struct *mm = current->mm;
 	LIST_HEAD(uf);
 
-	if (down_write_killable(&mm->mmap_sem))
-		return -EINTR;
-
 	ret = do_munmap(mm, start, len, &uf);
-	up_write(&mm->mmap_sem);
 	userfaultfd_unmap_complete(mm, &uf);
 	return ret;
 }

Matthew Wilcox (Oracle) Aug. 15, 2018, 9:09 p.m. UTC | #2

On Wed, Aug 15, 2018 at 12:16:06PM -0700, Matthew Wilcox wrote:
> (not even compiled, and I can see a good opportunity for combining the
> VM_LOCKED loop with the has_uprobes loop)

I was rushing to get that sent earlier.  Here it is tidied up to
actually compile.

Note the diffstat:

 mmap.c |   71 ++++++++++++++++++++++++++++++++++++++---------------------------
 1 file changed, 42 insertions(+), 29 deletions(-)

I think that's a pretty small extra price to pay for having this improved
scalability.

diff --git a/mm/mmap.c b/mm/mmap.c
index de699523c0b7..b77bb3908f8c 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2802,7 +2802,9 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	      struct list_head *uf)
 {
 	unsigned long end;
-	struct vm_area_struct *vma, *prev, *last;
+	struct vm_area_struct *vma, *prev, *last, *tmp;
+	int res = 0;
+	bool downgrade = false;
 
 	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
 		return -EINVAL;
@@ -2811,17 +2813,20 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	if (len == 0)
 		return -EINVAL;
 
+	if (down_write_killable(&mm->mmap_sem))
+		return -EINTR;
+
 	/* Find the first overlapping VMA */
 	vma = find_vma(mm, start);
 	if (!vma)
-		return 0;
+		goto unlock;
 	prev = vma->vm_prev;
-	/* we have  start < vma->vm_end  */
+	/* we have start < vma->vm_end  */
 
 	/* if it doesn't overlap, we have nothing.. */
 	end = start + len;
 	if (vma->vm_start >= end)
-		return 0;
+		goto unlock;
 
 	/*
 	 * If we need to split any vma, do it now to save pain later.
@@ -2831,28 +2836,27 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	 * places tmp vma above, and higher split_vma places tmp vma below.
 	 */
 	if (start > vma->vm_start) {
-		int error;
-
 		/*
 		 * Make sure that map_count on return from munmap() will
 		 * not exceed its limit; but let map_count go just above
 		 * its limit temporarily, to help free resources as expected.
 		 */
+		res = -ENOMEM;
 		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
-			return -ENOMEM;
+			goto unlock;
 
-		error = __split_vma(mm, vma, start, 0);
-		if (error)
-			return error;
+		res = __split_vma(mm, vma, start, 0);
+		if (res)
+			goto unlock;
 		prev = vma;
 	}
 
 	/* Does it split the last one? */
 	last = find_vma(mm, end);
 	if (last && end > last->vm_start) {
-		int error = __split_vma(mm, last, end, 1);
-		if (error)
-			return error;
+		res = __split_vma(mm, last, end, 1);
+		if (res)
+			goto unlock;
 	}
 	vma = prev ? prev->vm_next : mm->mmap;
 
@@ -2866,25 +2870,31 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 		 * split, despite we could. This is unlikely enough
 		 * failure that it's not worth optimizing it for.
 		 */
-		int error = userfaultfd_unmap_prep(vma, start, end, uf);
-		if (error)
-			return error;
+		res = userfaultfd_unmap_prep(vma, start, end, uf);
+		if (res)
+			goto unlock;
 	}
 
 	/*
 	 * unlock any mlock()ed ranges before detaching vmas
+	 * and check to see if there's any reason we might have to hold
+	 * the mmap_sem write-locked while unmapping regions.
 	 */
-	if (mm->locked_vm) {
-		struct vm_area_struct *tmp = vma;
-		while (tmp && tmp->vm_start < end) {
-			if (tmp->vm_flags & VM_LOCKED) {
-				mm->locked_vm -= vma_pages(tmp);
-				munlock_vma_pages_all(tmp);
-			}
-			tmp = tmp->vm_next;
+	downgrade = true;
+
+	for (tmp = vma; tmp && tmp->vm_start < end; tmp = tmp->vm_next) {
+		if (tmp->vm_flags & VM_LOCKED) {
+			mm->locked_vm -= vma_pages(tmp);
+			munlock_vma_pages_all(tmp);
 		}
+		if (tmp->vm_file &&
+				has_uprobes(tmp, tmp->vm_start, tmp->vm_end))
+			downgrade = false;
 	}
 
+	if (downgrade)
+		downgrade_write(&mm->mmap_sem);
+
 	/*
 	 * Remove the vma's, and unmap the actual pages
 	 */
@@ -2896,7 +2906,14 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	/* Fix up all other VM information */
 	remove_vma_list(mm, vma);
 
-	return 0;
+	res = 0;
+unlock:
+	if (downgrade) {
+		up_read(&mm->mmap_sem);
+	} else {
+		up_write(&mm->mmap_sem);
+	}
+	return res;
 }
 
 int vm_munmap(unsigned long start, size_t len)
@@ -2905,11 +2922,7 @@ int vm_munmap(unsigned long start, size_t len)
 	struct mm_struct *mm = current->mm;
 	LIST_HEAD(uf);
 
-	if (down_write_killable(&mm->mmap_sem))
-		return -EINTR;
-
 	ret = do_munmap(mm, start, len, &uf);
-	up_write(&mm->mmap_sem);
 	userfaultfd_unmap_complete(mm, &uf);
 	return ret;
 }

Yang Shi Aug. 15, 2018, 9:54 p.m. UTC | #3

On 8/15/18 2:09 PM, Matthew Wilcox wrote:
> On Wed, Aug 15, 2018 at 12:16:06PM -0700, Matthew Wilcox wrote:
>> (not even compiled, and I can see a good opportunity for combining the
>> VM_LOCKED loop with the has_uprobes loop)
> I was rushing to get that sent earlier.  Here it is tidied up to
> actually compile.

Thanks for the example. Yes, I believe the code still can be compacted 
to save some lines. However, the cover letter and the commit log of this 
patch has elaborated the discussion in the earlier reviews about why we 
do it in this way.

Or you just mean I don't have to call do_munmap() for the special 
mappings with the "downgrade" flag to save some cycles since do_munmap() 
will redo something which have been done?

Thanks,
Yang

>
> Note the diffstat:
>
>   mmap.c |   71 ++++++++++++++++++++++++++++++++++++++---------------------------
>   1 file changed, 42 insertions(+), 29 deletions(-)
>
> I think that's a pretty small extra price to pay for having this improved
> scalability.
>
> diff --git a/mm/mmap.c b/mm/mmap.c
> index de699523c0b7..b77bb3908f8c 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -2802,7 +2802,9 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   	      struct list_head *uf)
>   {
>   	unsigned long end;
> -	struct vm_area_struct *vma, *prev, *last;
> +	struct vm_area_struct *vma, *prev, *last, *tmp;
> +	int res = 0;
> +	bool downgrade = false;
>   
>   	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
>   		return -EINVAL;
> @@ -2811,17 +2813,20 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   	if (len == 0)
>   		return -EINVAL;
>   
> +	if (down_write_killable(&mm->mmap_sem))
> +		return -EINTR;
> +
>   	/* Find the first overlapping VMA */
>   	vma = find_vma(mm, start);
>   	if (!vma)
> -		return 0;
> +		goto unlock;
>   	prev = vma->vm_prev;
> -	/* we have  start < vma->vm_end  */
> +	/* we have start < vma->vm_end  */
>   
>   	/* if it doesn't overlap, we have nothing.. */
>   	end = start + len;
>   	if (vma->vm_start >= end)
> -		return 0;
> +		goto unlock;
>   
>   	/*
>   	 * If we need to split any vma, do it now to save pain later.
> @@ -2831,28 +2836,27 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   	 * places tmp vma above, and higher split_vma places tmp vma below.
>   	 */
>   	if (start > vma->vm_start) {
> -		int error;
> -
>   		/*
>   		 * Make sure that map_count on return from munmap() will
>   		 * not exceed its limit; but let map_count go just above
>   		 * its limit temporarily, to help free resources as expected.
>   		 */
> +		res = -ENOMEM;
>   		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
> -			return -ENOMEM;
> +			goto unlock;
>   
> -		error = __split_vma(mm, vma, start, 0);
> -		if (error)
> -			return error;
> +		res = __split_vma(mm, vma, start, 0);
> +		if (res)
> +			goto unlock;
>   		prev = vma;
>   	}
>   
>   	/* Does it split the last one? */
>   	last = find_vma(mm, end);
>   	if (last && end > last->vm_start) {
> -		int error = __split_vma(mm, last, end, 1);
> -		if (error)
> -			return error;
> +		res = __split_vma(mm, last, end, 1);
> +		if (res)
> +			goto unlock;
>   	}
>   	vma = prev ? prev->vm_next : mm->mmap;
>   
> @@ -2866,25 +2870,31 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   		 * split, despite we could. This is unlikely enough
>   		 * failure that it's not worth optimizing it for.
>   		 */
> -		int error = userfaultfd_unmap_prep(vma, start, end, uf);
> -		if (error)
> -			return error;
> +		res = userfaultfd_unmap_prep(vma, start, end, uf);
> +		if (res)
> +			goto unlock;
>   	}
>   
>   	/*
>   	 * unlock any mlock()ed ranges before detaching vmas
> +	 * and check to see if there's any reason we might have to hold
> +	 * the mmap_sem write-locked while unmapping regions.
>   	 */
> -	if (mm->locked_vm) {
> -		struct vm_area_struct *tmp = vma;
> -		while (tmp && tmp->vm_start < end) {
> -			if (tmp->vm_flags & VM_LOCKED) {
> -				mm->locked_vm -= vma_pages(tmp);
> -				munlock_vma_pages_all(tmp);
> -			}
> -			tmp = tmp->vm_next;
> +	downgrade = true;
> +
> +	for (tmp = vma; tmp && tmp->vm_start < end; tmp = tmp->vm_next) {
> +		if (tmp->vm_flags & VM_LOCKED) {
> +			mm->locked_vm -= vma_pages(tmp);
> +			munlock_vma_pages_all(tmp);
>   		}
> +		if (tmp->vm_file &&
> +				has_uprobes(tmp, tmp->vm_start, tmp->vm_end))
> +			downgrade = false;
>   	}
>   
> +	if (downgrade)
> +		downgrade_write(&mm->mmap_sem);
> +
>   	/*
>   	 * Remove the vma's, and unmap the actual pages
>   	 */
> @@ -2896,7 +2906,14 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   	/* Fix up all other VM information */
>   	remove_vma_list(mm, vma);
>   
> -	return 0;
> +	res = 0;
> +unlock:
> +	if (downgrade) {
> +		up_read(&mm->mmap_sem);
> +	} else {
> +		up_write(&mm->mmap_sem);
> +	}
> +	return res;
>   }
>   
>   int vm_munmap(unsigned long start, size_t len)
> @@ -2905,11 +2922,7 @@ int vm_munmap(unsigned long start, size_t len)
>   	struct mm_struct *mm = current->mm;
>   	LIST_HEAD(uf);
>   
> -	if (down_write_killable(&mm->mmap_sem))
> -		return -EINTR;
> -
>   	ret = do_munmap(mm, start, len, &uf);
> -	up_write(&mm->mmap_sem);
>   	userfaultfd_unmap_complete(mm, &uf);
>   	return ret;
>   }

Matthew Wilcox (Oracle) Aug. 16, 2018, 2:46 a.m. UTC | #4

On Wed, Aug 15, 2018 at 02:54:13PM -0700, Yang Shi wrote:
> 
> 
> On 8/15/18 2:09 PM, Matthew Wilcox wrote:
> > On Wed, Aug 15, 2018 at 12:16:06PM -0700, Matthew Wilcox wrote:
> > > (not even compiled, and I can see a good opportunity for combining the
> > > VM_LOCKED loop with the has_uprobes loop)
> > I was rushing to get that sent earlier.  Here it is tidied up to
> > actually compile.
> 
> Thanks for the example. Yes, I believe the code still can be compacted to
> save some lines. However, the cover letter and the commit log of this patch
> has elaborated the discussion in the earlier reviews about why we do it in
> this way.

You mean the other callers which need to hold mmap_sem write-locked for
longer?  I hadn't really considered those; how about this?

 mmap.c |   47 +++++++++++++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 18 deletions(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index de699523c0b7..06dc31d1da8c 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2798,11 +2798,11 @@ int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
  * work.  This now handles partial unmappings.
  * Jeremy Fitzhardinge <[e-mail]>
  */
-int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
-	      struct list_head *uf)
+static int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
+	      struct list_head *uf, bool downgrade)
 {
 	unsigned long end;
-	struct vm_area_struct *vma, *prev, *last;
+	struct vm_area_struct *vma, *prev, *last, *tmp;
 
 	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
 		return -EINVAL;
@@ -2816,7 +2816,7 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	if (!vma)
 		return 0;
 	prev = vma->vm_prev;
-	/* we have  start < vma->vm_end  */
+	/* we have start < vma->vm_end  */
 
 	/* if it doesn't overlap, we have nothing.. */
 	end = start + len;
@@ -2873,18 +2873,22 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 
 	/*
 	 * unlock any mlock()ed ranges before detaching vmas
+	 * and check to see if there's any reason we might have to hold
+	 * the mmap_sem write-locked while unmapping regions.
 	 */
-	if (mm->locked_vm) {
-		struct vm_area_struct *tmp = vma;
-		while (tmp && tmp->vm_start < end) {
-			if (tmp->vm_flags & VM_LOCKED) {
-				mm->locked_vm -= vma_pages(tmp);
-				munlock_vma_pages_all(tmp);
-			}
-			tmp = tmp->vm_next;
+	for (tmp = vma; tmp && tmp->vm_start < end; tmp = tmp->vm_next) {
+		if (tmp->vm_flags & VM_LOCKED) {
+			mm->locked_vm -= vma_pages(tmp);
+			munlock_vma_pages_all(tmp);
 		}
+		if (tmp->vm_file &&
+				has_uprobes(tmp, tmp->vm_start, tmp->vm_end))
+			downgrade = false;
 	}
 
+	if (downgrade)
+		downgrade_write(&mm->mmap_sem);
+
 	/*
 	 * Remove the vma's, and unmap the actual pages
 	 */
@@ -2896,7 +2900,13 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	/* Fix up all other VM information */
 	remove_vma_list(mm, vma);
 
-	return 0;
+	return downgrade ? 1 : 0;
+}
+
+int do_unmap(struct mm_struct *mm, unsigned long start, size_t len,
+		struct list_head *uf)
+{
+	return __do_munmap(mm, start, len, uf, false);
 }
 
 int vm_munmap(unsigned long start, size_t len)
@@ -2905,11 +2915,12 @@ int vm_munmap(unsigned long start, size_t len)
 	struct mm_struct *mm = current->mm;
 	LIST_HEAD(uf);
 
-	if (down_write_killable(&mm->mmap_sem))
-		return -EINTR;
-
-	ret = do_munmap(mm, start, len, &uf);
-	up_write(&mm->mmap_sem);
+	down_write(&mm->mmap_sem);
+	ret = __do_munmap(mm, start, len, &uf, true);
+	if (ret == 1)
+		up_read(&mm->mmap_sem);
+	else
+		up_write(&mm->mmap_sem);
 	userfaultfd_unmap_complete(mm, &uf);
 	return ret;
 }

Yang Shi Aug. 16, 2018, 6:11 a.m. UTC | #5

On 8/15/18 7:46 PM, Matthew Wilcox wrote:
> On Wed, Aug 15, 2018 at 02:54:13PM -0700, Yang Shi wrote:
>>
>> On 8/15/18 2:09 PM, Matthew Wilcox wrote:
>>> On Wed, Aug 15, 2018 at 12:16:06PM -0700, Matthew Wilcox wrote:
>>>> (not even compiled, and I can see a good opportunity for combining the
>>>> VM_LOCKED loop with the has_uprobes loop)
>>> I was rushing to get that sent earlier.  Here it is tidied up to
>>> actually compile.
>> Thanks for the example. Yes, I believe the code still can be compacted to
>> save some lines. However, the cover letter and the commit log of this patch
>> has elaborated the discussion in the earlier reviews about why we do it in
>> this way.
> You mean the other callers which need to hold mmap_sem write-locked for
> longer?  I hadn't really considered those; how about this?

Thanks. Yes, this is the other potential implementation. My rationale 
about a separate function for the optimized path is I would prefer 
optimize this step by step by starting with some relatively simple way, 
then add enhancement on top of it.

And, I would prefer keep the current implementation of do_munmap since 
it is called somewhere else and it might be called by the optimized path 
for some reason until we are confident enough that the optimization 
doesn't have regression.

This sounds like separate function vs an extra parameter. We do save 
some lines with extra parameter instead of a separate function.

Thanks,
Yang

>
>   mmap.c |   47 +++++++++++++++++++++++++++++------------------
>   1 file changed, 29 insertions(+), 18 deletions(-)
>
> diff --git a/mm/mmap.c b/mm/mmap.c
> index de699523c0b7..06dc31d1da8c 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -2798,11 +2798,11 @@ int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
>    * work.  This now handles partial unmappings.
>    * Jeremy Fitzhardinge <[e-mail]>
>    */
> -int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
> -	      struct list_head *uf)
> +static int __do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
> +	      struct list_head *uf, bool downgrade)
>   {
>   	unsigned long end;
> -	struct vm_area_struct *vma, *prev, *last;
> +	struct vm_area_struct *vma, *prev, *last, *tmp;
>   
>   	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
>   		return -EINVAL;
> @@ -2816,7 +2816,7 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   	if (!vma)
>   		return 0;
>   	prev = vma->vm_prev;
> -	/* we have  start < vma->vm_end  */
> +	/* we have start < vma->vm_end  */
>   
>   	/* if it doesn't overlap, we have nothing.. */
>   	end = start + len;
> @@ -2873,18 +2873,22 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   
>   	/*
>   	 * unlock any mlock()ed ranges before detaching vmas
> +	 * and check to see if there's any reason we might have to hold
> +	 * the mmap_sem write-locked while unmapping regions.
>   	 */
> -	if (mm->locked_vm) {
> -		struct vm_area_struct *tmp = vma;
> -		while (tmp && tmp->vm_start < end) {
> -			if (tmp->vm_flags & VM_LOCKED) {
> -				mm->locked_vm -= vma_pages(tmp);
> -				munlock_vma_pages_all(tmp);
> -			}
> -			tmp = tmp->vm_next;
> +	for (tmp = vma; tmp && tmp->vm_start < end; tmp = tmp->vm_next) {
> +		if (tmp->vm_flags & VM_LOCKED) {
> +			mm->locked_vm -= vma_pages(tmp);
> +			munlock_vma_pages_all(tmp);
>   		}
> +		if (tmp->vm_file &&
> +				has_uprobes(tmp, tmp->vm_start, tmp->vm_end))
> +			downgrade = false;
>   	}
>   
> +	if (downgrade)
> +		downgrade_write(&mm->mmap_sem);
> +
>   	/*
>   	 * Remove the vma's, and unmap the actual pages
>   	 */
> @@ -2896,7 +2900,13 @@ int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
>   	/* Fix up all other VM information */
>   	remove_vma_list(mm, vma);
>   
> -	return 0;
> +	return downgrade ? 1 : 0;
> +}
> +
> +int do_unmap(struct mm_struct *mm, unsigned long start, size_t len,
> +		struct list_head *uf)
> +{
> +	return __do_munmap(mm, start, len, uf, false);
>   }
>   
>   int vm_munmap(unsigned long start, size_t len)
> @@ -2905,11 +2915,12 @@ int vm_munmap(unsigned long start, size_t len)
>   	struct mm_struct *mm = current->mm;
>   	LIST_HEAD(uf);
>   
> -	if (down_write_killable(&mm->mmap_sem))
> -		return -EINTR;
> -
> -	ret = do_munmap(mm, start, len, &uf);
> -	up_write(&mm->mmap_sem);
> +	down_write(&mm->mmap_sem);
> +	ret = __do_munmap(mm, start, len, &uf, true);
> +	if (ret == 1)
> +		up_read(&mm->mmap_sem);
> +	else
> +		up_write(&mm->mmap_sem);
>   	userfaultfd_unmap_complete(mm, &uf);
>   	return ret;
>   }

Vlastimil Babka Aug. 22, 2018, 11:11 a.m. UTC | #6

On 08/15/2018 08:49 PM, Yang Shi wrote:

> +	start_vma = munmap_lookup_vma(mm, start, end);
> +	if (!start_vma)
> +		goto out;
> +	if (IS_ERR(start_vma)) {
> +		ret = PTR_ERR(start_vma);
> +		goto out;
> +	}
> +
> +	prev = start_vma->vm_prev;
> +
> +	if (unlikely(uf)) {
> +		ret = userfaultfd_unmap_prep(start_vma, start, end, uf);
> +		if (ret)
> +			goto out;
> +	}
> +

You sure it's ok to redo this in case of goto regular path? The
preparations have some side-effects... I would rather move this after
the regular path check?

Vlastimil Babka Aug. 22, 2018, 11:19 a.m. UTC | #7

On 08/15/2018 08:49 PM, Yang Shi wrote:
> +	downgrade_write(&mm->mmap_sem);
> +
> +	/* Zap mappings with read mmap_sem */
> +	unmap_region(mm, start_vma, prev, start, end);
> +
> +	arch_unmap(mm, start_vma, start, end);

Hmm, did you check that all architectures' arch_unmap() is safe with
read mmap_sem instead of write mmap_sem? E.g. x86 does
mpx_notify_unmap() there where I would be far from sure at first glance...

> +	remove_vma_list(mm, start_vma);
> +	up_read(&mm->mmap_sem);

Yang Shi Aug. 22, 2018, 7:20 p.m. UTC | #8

On 8/22/18 4:11 AM, Vlastimil Babka wrote:
> On 08/15/2018 08:49 PM, Yang Shi wrote:
>
>> +	start_vma = munmap_lookup_vma(mm, start, end);
>> +	if (!start_vma)
>> +		goto out;
>> +	if (IS_ERR(start_vma)) {
>> +		ret = PTR_ERR(start_vma);
>> +		goto out;
>> +	}
>> +
>> +	prev = start_vma->vm_prev;
>> +
>> +	if (unlikely(uf)) {
>> +		ret = userfaultfd_unmap_prep(start_vma, start, end, uf);
>> +		if (ret)
>> +			goto out;
>> +	}
>> +
> You sure it's ok to redo this in case of goto regular path? The
> preparations have some side-effects... I would rather move this after
> the regular path check?

This preparation sets vma->vm_userfaultfd_ctx.ctx for each vmas. But, 
before doing this, it calls has_unmap_ctx() to check if the ctx has been 
set or not. If it has been set, it just skip the vma. It sounds ok, right?

Yang Shi Aug. 22, 2018, 8:45 p.m. UTC | #9

On 8/22/18 4:19 AM, Vlastimil Babka wrote:
> On 08/15/2018 08:49 PM, Yang Shi wrote:
>> +	downgrade_write(&mm->mmap_sem);
>> +
>> +	/* Zap mappings with read mmap_sem */
>> +	unmap_region(mm, start_vma, prev, start, end);
>> +
>> +	arch_unmap(mm, start_vma, start, end);
> Hmm, did you check that all architectures' arch_unmap() is safe with
> read mmap_sem instead of write mmap_sem? E.g. x86 does
> mpx_notify_unmap() there where I would be far from sure at first glance...

Yes, I'm also not quite sure if it is 100% safe or not. I was trying to 
move this before downgrade_write, however, I'm not sure if it is ok or 
not too, so I keep the calling sequence.

For architectures, just x86 and ppc really do something. PPC just uses 
it for vdso unmap which should just happen during process exit, so it 
sounds safe.

For x86, mpx_notify_unmap() looks finally zap the VM_MPX vmas in bound 
table range with zap_page_range() and doesn't update vm flags, so it 
sounds ok to me since vmas have been detached, nobody can find those 
vmas. But, I'm not familiar with the details of mpx, maybe Kirill could 
help to confirm this?

Thanks,
Yang

>
>> +	remove_vma_list(mm, start_vma);
>> +	up_read(&mm->mmap_sem);

Kirill A. Shutemov Aug. 22, 2018, 9:10 p.m. UTC | #10

On Wed, Aug 22, 2018 at 01:45:44PM -0700, Yang Shi wrote:
> 
> 
> On 8/22/18 4:19 AM, Vlastimil Babka wrote:
> > On 08/15/2018 08:49 PM, Yang Shi wrote:
> > > +	downgrade_write(&mm->mmap_sem);
> > > +
> > > +	/* Zap mappings with read mmap_sem */
> > > +	unmap_region(mm, start_vma, prev, start, end);
> > > +
> > > +	arch_unmap(mm, start_vma, start, end);
> > Hmm, did you check that all architectures' arch_unmap() is safe with
> > read mmap_sem instead of write mmap_sem? E.g. x86 does
> > mpx_notify_unmap() there where I would be far from sure at first glance...
> 
> Yes, I'm also not quite sure if it is 100% safe or not. I was trying to move
> this before downgrade_write, however, I'm not sure if it is ok or not too,
> so I keep the calling sequence.
> 
> For architectures, just x86 and ppc really do something. PPC just uses it
> for vdso unmap which should just happen during process exit, so it sounds
> safe.
> 
> For x86, mpx_notify_unmap() looks finally zap the VM_MPX vmas in bound table
> range with zap_page_range() and doesn't update vm flags, so it sounds ok to
> me since vmas have been detached, nobody can find those vmas. But, I'm not
> familiar with the details of mpx, maybe Kirill could help to confirm this?

I don't see anything obviously dependent on down_write() in
mpx_notify_unmap(), but Dave should know better.

Dave Hansen Aug. 22, 2018, 9:42 p.m. UTC | #11

On 08/22/2018 02:10 PM, Kirill A. Shutemov wrote:
>> For x86, mpx_notify_unmap() looks finally zap the VM_MPX vmas in bound table
>> range with zap_page_range() and doesn't update vm flags, so it sounds ok to
>> me since vmas have been detached, nobody can find those vmas. But, I'm not
>> familiar with the details of mpx, maybe Kirill could help to confirm this?
> I don't see anything obviously dependent on down_write() in
> mpx_notify_unmap(), but Dave should know better.

We need mmap_sem for write in mpx_notify_unmap().

Its job is to clean up bounds tables, but bounds tables are dynamically
allocated and destroyed by the kernel.  When we destroy a table, we also
destroy the VMA for the bounds table *itself*, separate from the VMA
being unmapped.

But, this code is very likely to go away soon.  If it's causing a
problem for you, let me know and I'll see if I can get to removing it
faster.

Yang Shi Aug. 22, 2018, 9:56 p.m. UTC | #12

On 8/22/18 2:42 PM, Dave Hansen wrote:
> On 08/22/2018 02:10 PM, Kirill A. Shutemov wrote:
>>> For x86, mpx_notify_unmap() looks finally zap the VM_MPX vmas in bound table
>>> range with zap_page_range() and doesn't update vm flags, so it sounds ok to
>>> me since vmas have been detached, nobody can find those vmas. But, I'm not
>>> familiar with the details of mpx, maybe Kirill could help to confirm this?
>> I don't see anything obviously dependent on down_write() in
>> mpx_notify_unmap(), but Dave should know better.
> We need mmap_sem for write in mpx_notify_unmap().
>
> Its job is to clean up bounds tables, but bounds tables are dynamically
> allocated and destroyed by the kernel.  When we destroy a table, we also
> destroy the VMA for the bounds table *itself*, separate from the VMA
> being unmapped.

Thanks for confirming this. I didn't realize there is VMA for bounds 
table itself.

>
> But, this code is very likely to go away soon.  If it's causing a
> problem for you, let me know and I'll see if I can get to removing it
> faster.

Does it depends on unmap_region()? Or IOW, does it has to be called 
after unmap_region()? Now the calling sequence is:

detach vmas
unmap_region()
mpx_notify_unmap()

I'm wondering if it is safe to move it up before unmap_region() like:

detach vmas
mpx_notify_unmap()
unmap_region()

With this change we also can do our optimization to do unmap_region() 
with read mmap_sem. Otherwise it does cause problem.

Thanks,
Yang

Dave Hansen Aug. 22, 2018, 10:03 p.m. UTC | #13

On 08/22/2018 02:56 PM, [e-mail] wrote:
> 
> 
> On 8/22/18 2:42 PM, Dave Hansen wrote:
>> On 08/22/2018 02:10 PM, Kirill A. Shutemov wrote:
>>>> For x86, mpx_notify_unmap() looks finally zap the VM_MPX vmas in
>>>> bound table
>>>> range with zap_page_range() and doesn't update vm flags, so it
>>>> sounds ok to
>>>> me since vmas have been detached, nobody can find those vmas. But,
>>>> I'm not
>>>> familiar with the details of mpx, maybe Kirill could help to confirm
>>>> this?
>>> I don't see anything obviously dependent on down_write() in
>>> mpx_notify_unmap(), but Dave should know better.
>> We need mmap_sem for write in mpx_notify_unmap().
>>
>> Its job is to clean up bounds tables, but bounds tables are dynamically
>> allocated and destroyed by the kernel.  When we destroy a table, we also
>> destroy the VMA for the bounds table *itself*, separate from the VMA
>> being unmapped.
...
> Does it depends on unmap_region()? Or IOW, does it has to be called
> after unmap_region()? Now the calling sequence is:
> 
> detach vmas
> unmap_region()
> mpx_notify_unmap()
> 
> I'm wondering if it is safe to move it up before unmap_region() like:
> 
> detach vmas
> mpx_notify_unmap()
> unmap_region()
> 
> With this change we also can do our optimization to do unmap_region()
> with read mmap_sem. Otherwise it does cause problem.

I think changing the ordering is fine.

The MPX bounds table unmapping is entirely driven by the VMAs being
unmapped, so the page table unmapping in unmap_region() should not
affect it.

Patch

diff --git a/mm/mmap.c b/mm/mmap.c
index f05f49b..e92f680 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2768,6 +2768,89 @@  static inline void munlock_vmas(struct vm_area_struct *vma,
 	}
 }
 
+/*
+ * Zap pages with read mmap_sem held
+ *
+ * uf is the list for userfaultfd
+ */
+static int do_munmap_zap_rlock(struct mm_struct *mm, unsigned long start,
+			       size_t len, struct list_head *uf)
+{
+	unsigned long end;
+	struct vm_area_struct *start_vma, *prev, *vma;
+	int ret = 0;
+
+	if (!addr_ok(start, len))
+		return -EINVAL;
+
+	len = PAGE_ALIGN(len);
+
+	end = start + len;
+
+	/*
+	 * Need write mmap_sem to split vmas and detach vmas
+	 * splitting vma up-front to save PITA to clean if it is failed
+	 */
+	if (down_write_killable(&mm->mmap_sem))
+		return -EINTR;
+
+	start_vma = munmap_lookup_vma(mm, start, end);
+	if (!start_vma)
+		goto out;
+	if (IS_ERR(start_vma)) {
+		ret = PTR_ERR(start_vma);
+		goto out;
+	}
+
+	prev = start_vma->vm_prev;
+
+	if (unlikely(uf)) {
+		ret = userfaultfd_unmap_prep(start_vma, start, end, uf);
+		if (ret)
+			goto out;
+	}
+
+	/*
+	 * Unmapping vmas, which have:
+	 *   VM_HUGETLB or
+	 *   VM_PFNMAP or
+	 *   uprobes
+	 * need get done with write mmap_sem held since they may update
+	 * vm_flags. Deal with such mappings with regular do_munmap() call.
+	 */
+	for (vma = start_vma; vma && vma->vm_start < end; vma = vma->vm_next) {
+		if ((vma->vm_file &&
+		    has_uprobes(vma, vma->vm_start, vma->vm_end)) ||
+		    (vma->vm_flags & (VM_HUGETLB | VM_PFNMAP)))
+			goto regular_path;
+	}
+
+	/* Handle mlocked vmas */
+	if (mm->locked_vm)
+		munlock_vmas(start_vma, end);
+
+	/* Detach vmas from rbtree */
+	detach_vmas_to_be_unmapped(mm, start_vma, prev, end);
+
+	downgrade_write(&mm->mmap_sem);
+
+	/* Zap mappings with read mmap_sem */
+	unmap_region(mm, start_vma, prev, start, end);
+
+	arch_unmap(mm, start_vma, start, end);
+	remove_vma_list(mm, start_vma);
+	up_read(&mm->mmap_sem);
+
+	return 0;
+
+regular_path:
+	ret = do_munmap(mm, start, len, uf);
+
+out:
+	up_write(&mm->mmap_sem);
+	return ret;
+}
+
 /* Munmap is split into 2 main parts -- this part which finds
  * what needs doing, and the areas themselves, which do the
  * work.  This now handles partial unmappings.
@@ -2829,6 +2912,17 @@  int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
 	return 0;
 }
 
+static int vm_munmap_zap_rlock(unsigned long start, size_t len)
+{
+	int ret;
+	struct mm_struct *mm = current->mm;
+	LIST_HEAD(uf);
+
+	ret = do_munmap_zap_rlock(mm, start, len, &uf);
+	userfaultfd_unmap_complete(mm, &uf);
+	return ret;
+}
+
 int vm_munmap(unsigned long start, size_t len)
 {
 	int ret;
@@ -2848,10 +2942,9 @@  int vm_munmap(unsigned long start, size_t len)
 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
 {
 	profile_munmap(addr);
-	return vm_munmap(addr, len);
+	return vm_munmap_zap_rlock(addr, len);
 }
 
-
 /*
  * Emulation of deprecated remap_file_pages() syscall.
  */