Recently, we see reports [1] of a warning that triggers due to
move_page_tables() doing a downward and overlapping move on a
mutually-aligned offset within a PMD. By mutual alignment, I
mean the source and destination addresses of the mremap are at
the same offset within a PMD.
This mutual alignment along with the fact that the move is downward is
sufficient to cause a warning related to having an allocated PMD that
does not have PTEs in it.
This warning will only trigger when there is mutual alignment in the
move operation. A solution, as suggested by Linus Torvalds [2], is to
initiate the copy process at the PMD level whenever such alignment is
present. Implementing this approach will not only prevent the warning
from being triggered, but it will also optimize the operation as this
method should enhance the speed of the copy process whenever there's a
possibility to start copying at the PMD level.
Some more points:
a. The optimization can be done only when both the source and
destination of the mremap do not have anything mapped below it up to a
PMD boundary. I add support to detect that.
b. #a is not a problem for the call to move_page_tables() from exec.c as
nothing is expected to be mapped below the source. However, for
non-overlapping mutually aligned moves as triggered by mremap(2), I
added support for checking such cases.
c. I currently only optimize for PMD moves, in the future I/we can build
on this work and do PUD moves as well if there is a need for this. But I
want to take it one step at a time.
d. We need to be careful about mremap of ranges within the VMA itself.
For this purpose, I added checks to determine if the address to align
is not the beginning of the VMA which that address corresponds to.
[1] https://lore.kernel.org/all/ZB2GTBD%[email protected]/
[2] https://lore.kernel.org/all/CAHk-=whd7msp8reJPfeGNyt0LiySMT0egExx3TVZSX3Ok6X=9g@mail.gmail.com/
Suggested-by: Linus Torvalds <[email protected]>
Signed-off-by: Joel Fernandes (Google) <[email protected]>
---
mm/mremap.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 61 insertions(+)
diff --git a/mm/mremap.c b/mm/mremap.c
index 411a85682b58..bf355e4d6bd4 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -478,6 +478,51 @@ static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
return moved;
}
+/*
+ * A helper to check if a previous mapping exists. Required for
+ * move_page_tables() and realign_addr() to determine if a previous mapping
+ * exists before we can do realignment optimizations.
+ */
+static bool can_align_down(struct vm_area_struct *vma, unsigned long addr_to_align,
+ unsigned long mask)
+{
+ unsigned long addr_masked = addr_to_align & mask;
+ struct vm_area_struct *prev = NULL, *cur = NULL;
+
+ /*
+ * If @addr_to_align of either source or destination is not the beginning
+ * of the corresponding VMA, we can't align down or we will destroy part
+ * of the current mapping.
+ */
+ if (vma->vm_start != addr_to_align)
+ return false;
+
+ /*
+ * Find the VMA before @vma to see if it subsumes the masked address.
+ * The mmap write lock is held here so the lookup is safe.
+ */
+ cur = find_vma_prev(vma->vm_mm, vma->vm_start, &prev);
+ if (WARN_ON_ONCE(cur != vma))
+ return false;
+
+ return !prev || prev->vm_end <= addr_masked;
+}
+
+/* Opportunistically realign to specified boundary for faster copy. */
+static void realign_addr(unsigned long *old_addr, struct vm_area_struct *old_vma,
+ unsigned long *new_addr, struct vm_area_struct *new_vma,
+ unsigned long mask)
+{
+ bool mutually_aligned = (*old_addr & ~mask) == (*new_addr & ~mask);
+
+ if ((*old_addr & ~mask) && mutually_aligned
+ && can_align_down(old_vma, *old_addr, mask)
+ && can_align_down(new_vma, *new_addr, mask)) {
+ *old_addr = *old_addr & mask;
+ *new_addr = *new_addr & mask;
+ }
+}
+
unsigned long move_page_tables(struct vm_area_struct *vma,
unsigned long old_addr, struct vm_area_struct *new_vma,
unsigned long new_addr, unsigned long len,
@@ -493,6 +538,15 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
old_end = old_addr + len;
+ /*
+ * If possible, realign addresses to PMD boundary for faster copy.
+ * Don't align for intra-VMA moves as we may destroy existing mappings.
+ */
+ if ((vma != new_vma)
+ && (len >= PMD_SIZE - (old_addr & ~PMD_MASK))) {
+ realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK);
+ }
+
if (is_vm_hugetlb_page(vma))
return move_hugetlb_page_tables(vma, new_vma, old_addr,
new_addr, len);
@@ -565,6 +619,13 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
mmu_notifier_invalidate_range_end(&range);
+ /*
+ * Prevent negative return values when {old,new}_addr was realigned
+ * but we broke out of the above loop for the first PMD itself.
+ */
+ if (len + old_addr < old_end)
+ return 0;
+
return len + old_addr - old_end; /* how much done */
}
--
2.41.0.rc2.161.g9c6817b8e7-goog
On Wed, May 31, 2023 at 10:08:01PM +0000, Joel Fernandes (Google) wrote:
> Recently, we see reports [1] of a warning that triggers due to
> move_page_tables() doing a downward and overlapping move on a
> mutually-aligned offset within a PMD. By mutual alignment, I
> mean the source and destination addresses of the mremap are at
> the same offset within a PMD.
>
> This mutual alignment along with the fact that the move is downward is
> sufficient to cause a warning related to having an allocated PMD that
> does not have PTEs in it.
>
> This warning will only trigger when there is mutual alignment in the
> move operation. A solution, as suggested by Linus Torvalds [2], is to
> initiate the copy process at the PMD level whenever such alignment is
> present. Implementing this approach will not only prevent the warning
> from being triggered, but it will also optimize the operation as this
> method should enhance the speed of the copy process whenever there's a
> possibility to start copying at the PMD level.
>
> Some more points:
> a. The optimization can be done only when both the source and
> destination of the mremap do not have anything mapped below it up to a
> PMD boundary. I add support to detect that.
>
> b. #a is not a problem for the call to move_page_tables() from exec.c as
> nothing is expected to be mapped below the source. However, for
> non-overlapping mutually aligned moves as triggered by mremap(2), I
> added support for checking such cases.
>
> c. I currently only optimize for PMD moves, in the future I/we can build
> on this work and do PUD moves as well if there is a need for this. But I
> want to take it one step at a time.
>
> d. We need to be careful about mremap of ranges within the VMA itself.
> For this purpose, I added checks to determine if the address to align
> is not the beginning of the VMA which that address corresponds to.
>
> [1] https://lore.kernel.org/all/ZB2GTBD%[email protected]/
> [2] https://lore.kernel.org/all/CAHk-=whd7msp8reJPfeGNyt0LiySMT0egExx3TVZSX3Ok6X=9g@mail.gmail.com/
>
> Suggested-by: Linus Torvalds <[email protected]>
> Signed-off-by: Joel Fernandes (Google) <[email protected]>
> ---
> mm/mremap.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 61 insertions(+)
>
> diff --git a/mm/mremap.c b/mm/mremap.c
> index 411a85682b58..bf355e4d6bd4 100644
> --- a/mm/mremap.c
> +++ b/mm/mremap.c
> @@ -478,6 +478,51 @@ static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
> return moved;
> }
>
> +/*
> + * A helper to check if a previous mapping exists. Required for
> + * move_page_tables() and realign_addr() to determine if a previous mapping
> + * exists before we can do realignment optimizations.
> + */
> +static bool can_align_down(struct vm_area_struct *vma, unsigned long addr_to_align,
> + unsigned long mask)
> +{
> + unsigned long addr_masked = addr_to_align & mask;
> + struct vm_area_struct *prev = NULL, *cur = NULL;
> +
> + /*
> + * If @addr_to_align of either source or destination is not the beginning
> + * of the corresponding VMA, we can't align down or we will destroy part
> + * of the current mapping.
> + */
> + if (vma->vm_start != addr_to_align)
> + return false;
See below, I think we can eliminate this check.
> +
> + /*
> + * Find the VMA before @vma to see if it subsumes the masked address.
> + * The mmap write lock is held here so the lookup is safe.
> + */
> + cur = find_vma_prev(vma->vm_mm, vma->vm_start, &prev);
> + if (WARN_ON_ONCE(cur != vma))
> + return false;
> +
> + return !prev || prev->vm_end <= addr_masked;
This is a bit clunky, and I don't think we need the WARN_ON_ONCE() check if
we're under the mmap_lock.
How about something like:-
return find_vma_intersection(vma->mm, addr_masked, vma->vm_start) == NULL;
Which explicitly asserts that the range in [addr_masked, vma->vm_start) is
empty.
But actually, we should be able to go further and replace the previous
check with:-
return find_vma_intersection(vma->mm, addr_masked, addr_to_align) == NULL;
Which will fail if addr_to_align is offset within the VMA.
> +}
> +
> +/* Opportunistically realign to specified boundary for faster copy. */
> +static void realign_addr(unsigned long *old_addr, struct vm_area_struct *old_vma,
Something of a nit, but this isn't _always_ realigning the address, so perhaps
something like maybe_realign_addr() or try_realign_addr() is better?
This is probably debatable, as the comment already explains it is opportunistic
:)
> + unsigned long *new_addr, struct vm_area_struct *new_vma,
> + unsigned long mask)
> +{
> + bool mutually_aligned = (*old_addr & ~mask) == (*new_addr & ~mask);
> +
> + if ((*old_addr & ~mask) && mutually_aligned
I may be misunderstanding something here, but doesn't the first condition
here disallow for offset into PMD == 0? Why?
> + && can_align_down(old_vma, *old_addr, mask)
> + && can_align_down(new_vma, *new_addr, mask)) {
> + *old_addr = *old_addr & mask;
> + *new_addr = *new_addr & mask;
> + }
> +}
> +
> unsigned long move_page_tables(struct vm_area_struct *vma,
> unsigned long old_addr, struct vm_area_struct *new_vma,
> unsigned long new_addr, unsigned long len,
> @@ -493,6 +538,15 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
>
> old_end = old_addr + len;
>
> + /*
> + * If possible, realign addresses to PMD boundary for faster copy.
> + * Don't align for intra-VMA moves as we may destroy existing mappings.
> + */
> + if ((vma != new_vma)
Nit but these parens aren't needed. Also if we're deferring the decision as
to whether we realign to this function, why are we doing this check here
and not here?
It feels like it'd be neater to keep all the conditions (including the
length one) together in one place.
> + && (len >= PMD_SIZE - (old_addr & ~PMD_MASK))) {
You don't mention this condition in the above comment (if we have this
altogether as part of the realign function could comment separately there)
- so we only go ahead and do this optimisation if the length of the remap
is such that the entire of old_addr -> end of its PMD (and thus the same
for new_addr) is copied?
I may be missing something/being naive here, but can't we just do a similar
check to the one done for space _below_ the VMA to see if [end, (end of
PMD)) is equally empty?
> + realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK);
> + }
> +
> if (is_vm_hugetlb_page(vma))
> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> new_addr, len);
> @@ -565,6 +619,13 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
>
> mmu_notifier_invalidate_range_end(&range);
>
> + /*
> + * Prevent negative return values when {old,new}_addr was realigned
> + * but we broke out of the above loop for the first PMD itself.
> + */
> + if (len + old_addr < old_end)
> + return 0;
> +
I find this a little iffy, I mean I see that if you align [old,new]_addr to
PMD, then from then on in you're relying on the fact that the loop is just
going from old_addr (now aligned) -> old_end and thus has the correct
length.
Can't we just fix this issue by correcting len? If you take my review above
which checks len in [maybe_]realign_addr(), you could take that as a
pointer and equally update that.
Then you can drop this check.
Also I am concerned in the hugetlb case -> len is passed to
move_hugetlb_page_tables() which is now strictly incorrect, I wonder if
this could cause an issue?
Correcting len seems the neat way of addressing this.
> return len + old_addr - old_end; /* how much done */
> }
>
> --
> 2.41.0.rc2.161.g9c6817b8e7-goog
>
Hi Lorenzo,
Thanks for the review! I replied below:
On 6/17/23 18:49, Lorenzo Stoakes wrote:
> On Wed, May 31, 2023 at 10:08:01PM +0000, Joel Fernandes (Google) wrote:
>> Recently, we see reports [1] of a warning that triggers due to
>> move_page_tables() doing a downward and overlapping move on a
>> mutually-aligned offset within a PMD. By mutual alignment, I
>> mean the source and destination addresses of the mremap are at
>> the same offset within a PMD.
>>
>> This mutual alignment along with the fact that the move is downward is
>> sufficient to cause a warning related to having an allocated PMD that
>> does not have PTEs in it.
>>
>> This warning will only trigger when there is mutual alignment in the
>> move operation. A solution, as suggested by Linus Torvalds [2], is to
>> initiate the copy process at the PMD level whenever such alignment is
>> present. Implementing this approach will not only prevent the warning
>> from being triggered, but it will also optimize the operation as this
>> method should enhance the speed of the copy process whenever there's a
[...]
>> Suggested-by: Linus Torvalds <[email protected]>
>> Signed-off-by: Joel Fernandes (Google) <[email protected]>
>> ---
>> mm/mremap.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>> 1 file changed, 61 insertions(+)
>>
>> diff --git a/mm/mremap.c b/mm/mremap.c
>> index 411a85682b58..bf355e4d6bd4 100644
>> --- a/mm/mremap.c
>> +++ b/mm/mremap.c
>> @@ -478,6 +478,51 @@ static bool move_pgt_entry(enum pgt_entry entry, struct
>> return moved;
>> }
>>
>> +/*
>> + * A helper to check if a previous mapping exists. Required for
>> + * move_page_tables() and realign_addr() to determine if a previous mapping
>> + * exists before we can do realignment optimizations.
>> + */
>> +static bool can_align_down(struct vm_area_struct *vma, unsigned long addr_to_align,
>> + unsigned long mask)
>> +{
>> + unsigned long addr_masked = addr_to_align & mask;
>> + struct vm_area_struct *prev = NULL, *cur = NULL;
>> +
>> + /*
>> + * If @addr_to_align of either source or destination is not the beginning
>> + * of the corresponding VMA, we can't align down or we will destroy part
>> + * of the current mapping.
>> + */
>> + if (vma->vm_start != addr_to_align)
>> + return false;
>
> See below, I think we can eliminate this check.
>
>> +
>> + /*
>> + * Find the VMA before @vma to see if it subsumes the masked address.
>> + * The mmap write lock is held here so the lookup is safe.
>> + */
>> + cur = find_vma_prev(vma->vm_mm, vma->vm_start, &prev);
>> + if (WARN_ON_ONCE(cur != vma))
>> + return false;
>> +
>> + return !prev || prev->vm_end <= addr_masked;
>
> This is a bit clunky, and I don't think we need the WARN_ON_ONCE() check if
> we're under the mmap_lock.
>
> How about something like:-
>
> return find_vma_intersection(vma->mm, addr_masked, vma->vm_start) == NULL;
>
> Which explicitly asserts that the range in [addr_masked, vma->vm_start) is
> empty.
>
> But actually, we should be able to go further and replace the previous
> check with:-
>
> return find_vma_intersection(vma->mm, addr_masked, addr_to_align) == NULL;
>
> Which will fail if addr_to_align is offset within the VMA.
Your suggestion would mean that we do a full VMA search starting from the root. That would
not be a nice thing if say we've 1000s of VMAs?
Actually Liam told me to use find_vma_prev() because given a VMA, the maple tree would not
have to work that hard for the common case to find the previous VMA. Per conversing with
him, there is a chance we may have to go one step above in the tree if we hit the edge of
a node, but that's not supposed to be the common case. In previous code, the previous VMA
could just be obtained using the "previous VMA" pointer, however that pointer has been
remove since the maple tree changes and given a VMA, going to the previous one using the
maple tree is just as fast (as I'm told).
Considering this, I would keep the code as-is and perhaps you/we could consider the
replacement with another API in a subsequent patch as it does the job for this patch.
>> + unsigned long *new_addr, struct vm_area_struct *new_vma,
>> + unsigned long mask)
>> +{
>> + bool mutually_aligned = (*old_addr & ~mask) == (*new_addr & ~mask);
>> +
>> + if ((*old_addr & ~mask) && mutually_aligned
>
> I may be misunderstanding something here, but doesn't the first condition
> here disallow for offset into PMD == 0? Why?
Because in such a situation, the alignment is already done and there's nothing to align.
The patch wants to align down to the PMD and we would not want to waste CPU cycles if
there's nothing to do.
>> + && can_align_down(old_vma, *old_addr, mask)
>> + && can_align_down(new_vma, *new_addr, mask)) {
>> + *old_addr = *old_addr & mask;
>> + *new_addr = *new_addr & mask;
>> + }
>> +}
>> +
>> unsigned long move_page_tables(struct vm_area_struct *vma,
>> unsigned long old_addr, struct vm_area_struct *new_vma,
>> unsigned long new_addr, unsigned long len,
>> @@ -493,6 +538,15 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
>>
>> old_end = old_addr + len;
>>
>> + /*
>> + * If possible, realign addresses to PMD boundary for faster copy.
>> + * Don't align for intra-VMA moves as we may destroy existing mappings.
>> + */
>> + if ((vma != new_vma)
>
> Nit but these parens aren't needed.
Sure, I can drop the parens.
> Also if we're deferring the decision as
> to whether we realign to this function, why are we doing this check here
> and not here?
Hmm, well the function name is realign_addr() so I kept some of the initial checks outside
of it where we should "obviously" not realign. I could do what you're suggesting and
change it to try_realign_addr() or something. And move those checks in there. That would
be a bit better.
> It feels like it'd be neater to keep all the conditions (including the
> length one) together in one place.
>
>
>> + && (len >= PMD_SIZE - (old_addr & ~PMD_MASK))) {
Well, yeah maybe. I'll look into it, thanks.
> You don't mention this condition in the above comment (if we have this
> altogether as part of the realign function could comment separately there)
Ok, sounds good -- I will add a comment with some of the explanation above.
> - so we only go ahead and do this optimisation if the length of the remap
> is such that the entire of old_addr -> end of its PMD (and thus the same
> for new_addr) is copied?
Yes, correct. And in the future that could also be optimized (if say there is no
subsequent mapping, so we can copy the tail PMD as well, however one step at a time and
all that.)
> I may be missing something/being naive here, but can't we just do a similar
> check to the one done for space _below_ the VMA to see if [end, (end of
> PMD)) is equally empty?
We can, but the warning that was triggering does not really need that to be silenced. I am
happy to do that in a later patch if needed, or you can. ;-) But I'd like to keep the risk
low since this was itself hard enough to get right.
>> + realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK);
>> + }
>> +
>> if (is_vm_hugetlb_page(vma))
>> return move_hugetlb_page_tables(vma, new_vma, old_addr,
>> new_addr, len);
>> @@ -565,6 +619,13 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
>>
>> mmu_notifier_invalidate_range_end(&range);
>>
>> + /*
>> + * Prevent negative return values when {old,new}_addr was realigned
>> + * but we broke out of the above loop for the first PMD itself.
>> + */
>> + if (len + old_addr < old_end)
>> + return 0;
>> +
>
> I find this a little iffy, I mean I see that if you align [old,new]_addr to
> PMD, then from then on in you're relying on the fact that the loop is just
> going from old_addr (now aligned) -> old_end and thus has the correct
> length.
>
> Can't we just fix this issue by correcting len? If you take my review above
> which checks len in [maybe_]realign_addr(), you could take that as a
> pointer and equally update that.
>
> Then you can drop this check.
The drawback of adjusting len is it changes what move_page_tables() users were previously
expecting.
I think we should look at the return value of move_page_tables() as well, not just len
independently.
len is what the user requested.
"len + old_addr - old_end" is how much was actually copied and is the return value.
If everything was copied, old_addr == old_end and len is unchanged.
The users of move_page_tables(), like move_vma() should not care whether we copied a full
PMD or not. In fact telling them anything like may cause problems with the interpretation
of the return value I think.
They asked us to copy len, did we copy it? hell yeah.
Note that after the first loop iteration's PMD copy, old_addr is now at the PMD boundary
and the functionality of this function is not changed with this patch. We end up doing a
PMD-copy just like we used to without this patch. So this patch does not really change
anything from before.
The following are the cases:
1. If we realign and copy, yes we copied a PMD, but really it was to satisfy the requested
length. In this situation, "len + old_addr - old_end" is accurate and just like before.
We copied whatever the user requested. Yes we copied a little more, but who cares? We
copied into a mapping that does not exist anyway. It may be absurd for us to return a len
that is greater than the requested len IMO.
2. If there are no errors (example first PMD copy did not fail), "len + old_addr -
old_end" is identical to what it was without this patch -- as it should be. That's true
whether we realigned or not.
3. If we realigned and the first PMD copy failed (unlikely error) -- that's where there's
a problem. We would end up returning a negative value. That's what Linus found and
suggested to correct. Because (old_addr - old_end) will be greater than len in such a
situation, however unlikely.
>> return len + old_addr - old_end; /* how much done */
>> }
> Also I am concerned in the hugetlb case -> len is passed to
> move_hugetlb_page_tables() which is now strictly incorrect, I wonder if
> this could cause an issue?
>
> Correcting len seems the neat way of addressing this.
That's a good point. I am wondering if we can just change that from:
if (is_vm_hugetlb_page(vma))
return move_hugetlb_page_tables(vma, new_vma, old_addr,
new_addr, len);
to:
if (is_vm_hugetlb_page(vma))
return move_hugetlb_page_tables(vma, new_vma, old_addr,
new_addr, old_addr - new_addr);
Or, another option is to turn it off for hugetlb by just moving:
if (len >= PMD_SIZE - (old_addr & ~PMD_MASK))
realign_addr(...);
to after:
if (is_vm_hugetlb_page(vma))
return move_hugetlb_page_tables(...);
thanks,
- Joel
On Mon, Jun 19, 2023 at 11:55:08AM -0400, Joel Fernandes wrote:
> Hi Lorenzo,
> Thanks for the review! I replied below:
>
> On 6/17/23 18:49, Lorenzo Stoakes wrote:
> > On Wed, May 31, 2023 at 10:08:01PM +0000, Joel Fernandes (Google) wrote:
> >> Recently, we see reports [1] of a warning that triggers due to
> >> move_page_tables() doing a downward and overlapping move on a
> >> mutually-aligned offset within a PMD. By mutual alignment, I
> >> mean the source and destination addresses of the mremap are at
> >> the same offset within a PMD.
> >>
> >> This mutual alignment along with the fact that the move is downward is
> >> sufficient to cause a warning related to having an allocated PMD that
> >> does not have PTEs in it.
> >>
> >> This warning will only trigger when there is mutual alignment in the
> >> move operation. A solution, as suggested by Linus Torvalds [2], is to
> >> initiate the copy process at the PMD level whenever such alignment is
> >> present. Implementing this approach will not only prevent the warning
> >> from being triggered, but it will also optimize the operation as this
> >> method should enhance the speed of the copy process whenever there's a
>
> [...]
>
> >> Suggested-by: Linus Torvalds <[email protected]>
> >> Signed-off-by: Joel Fernandes (Google) <[email protected]>
> >> ---
> >> mm/mremap.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++
> >> 1 file changed, 61 insertions(+)
> >>
> >> diff --git a/mm/mremap.c b/mm/mremap.c
> >> index 411a85682b58..bf355e4d6bd4 100644
> >> --- a/mm/mremap.c
> >> +++ b/mm/mremap.c
> >> @@ -478,6 +478,51 @@ static bool move_pgt_entry(enum pgt_entry entry, struct
> >> return moved;
> >> }
> >>
> >> +/*
> >> + * A helper to check if a previous mapping exists. Required for
> >> + * move_page_tables() and realign_addr() to determine if a previous mapping
> >> + * exists before we can do realignment optimizations.
> >> + */
> >> +static bool can_align_down(struct vm_area_struct *vma, unsigned long addr_to_align,
> >> + unsigned long mask)
> >> +{
> >> + unsigned long addr_masked = addr_to_align & mask;
> >> + struct vm_area_struct *prev = NULL, *cur = NULL;
> >> +
> >> + /*
> >> + * If @addr_to_align of either source or destination is not the beginning
> >> + * of the corresponding VMA, we can't align down or we will destroy part
> >> + * of the current mapping.
> >> + */
> >> + if (vma->vm_start != addr_to_align)
> >> + return false;
> >
> > See below, I think we can eliminate this check.
> >
> >> +
> >> + /*
> >> + * Find the VMA before @vma to see if it subsumes the masked address.
> >> + * The mmap write lock is held here so the lookup is safe.
> >> + */
> >> + cur = find_vma_prev(vma->vm_mm, vma->vm_start, &prev);
> >> + if (WARN_ON_ONCE(cur != vma))
> >> + return false;
> >> +
> >> + return !prev || prev->vm_end <= addr_masked;
> >
> > This is a bit clunky, and I don't think we need the WARN_ON_ONCE() check if
> > we're under the mmap_lock.
> >
> > How about something like:-
> >
> > return find_vma_intersection(vma->mm, addr_masked, vma->vm_start) == NULL;
> >
> > Which explicitly asserts that the range in [addr_masked, vma->vm_start) is
> > empty.
> >
> > But actually, we should be able to go further and replace the previous
> > check with:-
> >
> > return find_vma_intersection(vma->mm, addr_masked, addr_to_align) == NULL;
> >
> > Which will fail if addr_to_align is offset within the VMA.
>
> Your suggestion would mean that we do a full VMA search starting from the
> root. That would not be a nice thing if say we've 1000s of VMAs?
>
> Actually Liam told me to use find_vma_prev() because given a VMA, the maple
> tree would not have to work that hard for the common case to find the
> previous VMA. Per conversing with him, there is a chance we may have to go
> one step above in the tree if we hit the edge of a node, but that's not
> supposed to be the common case. In previous code, the previous VMA could
> just be obtained using the "previous VMA" pointer, however that pointer has
> been remove since the maple tree changes and given a VMA, going to the
> previous one using the maple tree is just as fast (as I'm told).
As far as I can tell, find_vma_prev() already does a walk? I mean this is
equivalent to find_vma() only retrieving the previous VMA right? I defer to
Liam, but I'm not sure this would be that much more involved? Perhaps he
can comment.
An alternative is to create an iterator and use vma_prev(). I find it
extremely clunky that we search for a VMA we already possess (and it's
previous one) while not needing the the former.
I'm not hugely familiar with the maple tree (perhaps Liam can comment) but
I suspect that'd be more performant if that's the concern. Either way I
would be surprised if this is the correct approach.
>
> Considering this, I would keep the code as-is and perhaps you/we could
> consider the replacement with another API in a subsequent patch as it does
> the job for this patch.
See above. I don't think this kind of comment is helpful in code
review. Your disagreement above suffices, I've responded to it and of
course if there is no other way this is fine.
But I'd be surprised, and re-looking up a VMA we already have is just
horrid. It's not really a nitpick, it's a code quality issue in my view.
In any case, let's please try to avoid 'if you are bothered, write a follow
up patch' style responses. If you disagree with something just say so, it's
fine! :)
>
> >> + unsigned long *new_addr, struct vm_area_struct *new_vma,
> >> + unsigned long mask)
> >> +{
> >> + bool mutually_aligned = (*old_addr & ~mask) == (*new_addr & ~mask);
> >> +
> >> + if ((*old_addr & ~mask) && mutually_aligned
> >
> > I may be misunderstanding something here, but doesn't the first condition
> > here disallow for offset into PMD == 0? Why?
>
> Because in such a situation, the alignment is already done and there's
> nothing to align. The patch wants to align down to the PMD and we would not
> want to waste CPU cycles if there's nothing to do.
OK, makes sense. It'd be useful to have a comment to this effect.
>
> >> + && can_align_down(old_vma, *old_addr, mask)
> >> + && can_align_down(new_vma, *new_addr, mask)) {
> >> + *old_addr = *old_addr & mask;
> >> + *new_addr = *new_addr & mask;
> >> + }
> >> +}
> >> +
> >> unsigned long move_page_tables(struct vm_area_struct *vma,
> >> unsigned long old_addr, struct vm_area_struct *new_vma,
> >> unsigned long new_addr, unsigned long len,
> >> @@ -493,6 +538,15 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
> >>
> >> old_end = old_addr + len;
> >>
> >> + /*
> >> + * If possible, realign addresses to PMD boundary for faster copy.
> >> + * Don't align for intra-VMA moves as we may destroy existing mappings.
> >> + */
> >> + if ((vma != new_vma)
> >
> > Nit but these parens aren't needed.
>
> Sure, I can drop the parens.
Thanks.
>
> > Also if we're deferring the decision as
> > to whether we realign to this function, why are we doing this check here
> > and not here?
>
> Hmm, well the function name is realign_addr() so I kept some of the initial
> checks outside of it where we should "obviously" not realign. I could do
> what you're suggesting and change it to try_realign_addr() or something. And
> move those checks in there. That would be a bit better.
Thanks.
>
> > It feels like it'd be neater to keep all the conditions (including the
> > length one) together in one place.
> >
> >
> >> + && (len >= PMD_SIZE - (old_addr & ~PMD_MASK))) {
>
> Well, yeah maybe. I'll look into it, thanks.
I mean it's not a huge big deal, but reading your code having a bunch of
conditions in two different places is a little hard to parse and jarring.
>
> > You don't mention this condition in the above comment (if we have this
> > altogether as part of the realign function could comment separately there)
>
> Ok, sounds good -- I will add a comment with some of the explanation above.
>
> > - so we only go ahead and do this optimisation if the length of the remap
> > is such that the entire of old_addr -> end of its PMD (and thus the same
> > for new_addr) is copied?
>
> Yes, correct. And in the future that could also be optimized (if say there
> is no subsequent mapping, so we can copy the tail PMD as well, however one
> step at a time and all that.)
>
OK cool makes sense.
> > I may be missing something/being naive here, but can't we just do a similar
> > check to the one done for space _below_ the VMA to see if [end, (end of
> > PMD)) is equally empty?
>
> We can, but the warning that was triggering does not really need that to be
> silenced. I am happy to do that in a later patch if needed, or you can. ;-)
> But I'd like to keep the risk low since this was itself hard enough to get
> right.
(see above about 'later patch' comments...)
Sure, this is not a big deal.
>
> >> + realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK);
> >> + }
> >> +
> >> if (is_vm_hugetlb_page(vma))
> >> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> >> new_addr, len);
> >> @@ -565,6 +619,13 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
> >>
> >> mmu_notifier_invalidate_range_end(&range);
> >>
> >> + /*
> >> + * Prevent negative return values when {old,new}_addr was realigned
> >> + * but we broke out of the above loop for the first PMD itself.
> >> + */
> >> + if (len + old_addr < old_end)
> >> + return 0;
> >> +
> >
> > I find this a little iffy, I mean I see that if you align [old,new]_addr to
> > PMD, then from then on in you're relying on the fact that the loop is just
> > going from old_addr (now aligned) -> old_end and thus has the correct
> > length.
> >
> > Can't we just fix this issue by correcting len? If you take my review above
> > which checks len in [maybe_]realign_addr(), you could take that as a
> > pointer and equally update that.
> >
> > Then you can drop this check.
>
> The drawback of adjusting len is it changes what move_page_tables() users
> were previously expecting.
>
> I think we should look at the return value of move_page_tables() as well,
> not just len independently.
>
> len is what the user requested.
>
> "len + old_addr - old_end" is how much was actually copied and is the return value.
>
> If everything was copied, old_addr == old_end and len is unchanged.
Ah yeah I see, sorry I missed the fact we're returning a value, that does
complicate things...
If we retain the hugetlb logic, then we could work around the issue with
that instance of len by storing the 'actual length' of the range in
a new var actual_len and passing that.
If we choose to instead just not do this for hugetlb (I wonder if the
hugetlb handling code actually does the equivalent of this since surely
these pages have to be handled a PMD at a time?) then we can drop the whole
actual_len idea [see below on response to hugetlb thing].
>
> The users of move_page_tables(), like move_vma() should not care whether we
> copied a full PMD or not. In fact telling them anything like may cause
> problems with the interpretation of the return value I think.
>
> They asked us to copy len, did we copy it? hell yeah.
>
> Note that after the first loop iteration's PMD copy, old_addr is now at the
> PMD boundary and the functionality of this function is not changed with this
> patch. We end up doing a PMD-copy just like we used to without this patch.
> So this patch does not really change anything from before.
>
> The following are the cases:
>
> 1. If we realign and copy, yes we copied a PMD, but really it was to satisfy
> the requested length. In this situation, "len + old_addr - old_end" is
> accurate and just like before. We copied whatever the user requested. Yes we
> copied a little more, but who cares? We copied into a mapping that does not
> exist anyway. It may be absurd for us to return a len that is greater than
> the requested len IMO.
>
> 2. If there are no errors (example first PMD copy did not fail), "len +
> old_addr - old_end" is identical to what it was without this patch -- as it
> should be. That's true whether we realigned or not.
>
> 3. If we realigned and the first PMD copy failed (unlikely error) -- that's
> where there's a problem. We would end up returning a negative value. That's
> what Linus found and suggested to correct. Because (old_addr - old_end) will
> be greater than len in such a situation, however unlikely.
>
Right. Yeah that is thorny, sorry I did miss the degree of the complexity
with that... ugh ye gods. Probably then this has to be retained.
I was thinking we could use min(actual_len + old_addr - old_end, len), but
then we'd over-report what was 'copied' (actually not copied) because
that'd include the address range spanned by the empty PTE entries up to the
start of the PMD entry.
> >> return len + old_addr - old_end; /* how much done */
> >> }
> > Also I am concerned in the hugetlb case -> len is passed to
> > move_hugetlb_page_tables() which is now strictly incorrect, I wonder if
> > this could cause an issue?
> >
> > Correcting len seems the neat way of addressing this.
>
> That's a good point. I am wondering if we can just change that from:
>
> if (is_vm_hugetlb_page(vma))
> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> new_addr, len);
>
> to:
> if (is_vm_hugetlb_page(vma))
> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> new_addr, old_addr - new_addr);
>
> Or, another option is to turn it off for hugetlb by just moving:
>
> if (len >= PMD_SIZE - (old_addr & ~PMD_MASK))
> realign_addr(...);
>
> to after:
>
> if (is_vm_hugetlb_page(vma))
> return move_hugetlb_page_tables(...);
>
> thanks,
I think the actual_len solution should sort this right? If not maybe better
to be conservative and disable for the hugetlb case (I'm not sure if this
would help given you'd need to be PMD aligned anyway right?), so not to
hold up the series.
If we do decide not to include hugetlb (the endless 'special case' for so
much code...) in this then we can drop the actual_len idea altogether.
(Yes I realise it's ironic I'm suggesting deferring to a later patch here
but there you go ;)
>
> - Joel
>
Hi Lorenzo,
On 6/20/23 07:02, Lorenzo Stoakes wrote:
> On Mon, Jun 19, 2023 at 11:55:08AM -0400, Joel Fernandes wrote:
>> Hi Lorenzo,
>> Thanks for the review! I replied below:
>>
>> On 6/17/23 18:49, Lorenzo Stoakes wrote:
>>> On Wed, May 31, 2023 at 10:08:01PM +0000, Joel Fernandes (Google) wrote:
>>>> Recently, we see reports [1] of a warning that triggers due to
>>>> move_page_tables() doing a downward and overlapping move on a
>>>> mutually-aligned offset within a PMD. By mutual alignment, I
>>>> mean the source and destination addresses of the mremap are at
>>>> the same offset within a PMD.
>>>>
>>>> This mutual alignment along with the fact that the move is downward is
>>>> sufficient to cause a warning related to having an allocated PMD that
>>>> does not have PTEs in it.
>>>>
>>>> This warning will only trigger when there is mutual alignment in the
>>>> move operation. A solution, as suggested by Linus Torvalds [2], is to
>>>> initiate the copy process at the PMD level whenever such alignment is
>>>> present. Implementing this approach will not only prevent the warning
>>>> from being triggered, but it will also optimize the operation as this
>>>> method should enhance the speed of the copy process whenever there's a
>>
>> [...]
>>
>>>> Suggested-by: Linus Torvalds <[email protected]>
>>>> Signed-off-by: Joel Fernandes (Google) <[email protected]>
>>>> ---
>>>> mm/mremap.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>>>> 1 file changed, 61 insertions(+)
>>>>
>>>> diff --git a/mm/mremap.c b/mm/mremap.c
>>>> index 411a85682b58..bf355e4d6bd4 100644
>>>> --- a/mm/mremap.c
>>>> +++ b/mm/mremap.c
>>>> @@ -478,6 +478,51 @@ static bool move_pgt_entry(enum pgt_entry entry, struct
>>>> return moved;
>>>> }
>>>>
>>>> +/*
>>>> + * A helper to check if a previous mapping exists. Required for
>>>> + * move_page_tables() and realign_addr() to determine if a previous mapping
>>>> + * exists before we can do realignment optimizations.
>>>> + */
>>>> +static bool can_align_down(struct vm_area_struct *vma, unsigned long addr_to_align,
>>>> + unsigned long mask)
>>>> +{
>>>> + unsigned long addr_masked = addr_to_align & mask;
>>>> + struct vm_area_struct *prev = NULL, *cur = NULL;
>>>> +
>>>> + /*
>>>> + * If @addr_to_align of either source or destination is not the beginning
>>>> + * of the corresponding VMA, we can't align down or we will destroy part
>>>> + * of the current mapping.
>>>> + */
>>>> + if (vma->vm_start != addr_to_align)
>>>> + return false;
>>>
>>> See below, I think we can eliminate this check.
>>>
>>>> +
>>>> + /*
>>>> + * Find the VMA before @vma to see if it subsumes the masked address.
>>>> + * The mmap write lock is held here so the lookup is safe.
>>>> + */
>>>> + cur = find_vma_prev(vma->vm_mm, vma->vm_start, &prev);
>>>> + if (WARN_ON_ONCE(cur != vma))
>>>> + return false;
>>>> +
>>>> + return !prev || prev->vm_end <= addr_masked;
>>>
>>> This is a bit clunky, and I don't think we need the WARN_ON_ONCE() check if
>>> we're under the mmap_lock.
>>>
>>> How about something like:-
>>>
>>> return find_vma_intersection(vma->mm, addr_masked, vma->vm_start) == NULL;
>>>
>>> Which explicitly asserts that the range in [addr_masked, vma->vm_start) is
>>> empty.
>>>
>>> But actually, we should be able to go further and replace the previous
>>> check with:-
>>>
>>> return find_vma_intersection(vma->mm, addr_masked, addr_to_align) == NULL;
>>>
>>> Which will fail if addr_to_align is offset within the VMA.
>>
>> Your suggestion would mean that we do a full VMA search starting from the
>> root. That would not be a nice thing if say we've 1000s of VMAs?
>>
>> Actually Liam told me to use find_vma_prev() because given a VMA, the maple
>> tree would not have to work that hard for the common case to find the
>> previous VMA. Per conversing with him, there is a chance we may have to go
>> one step above in the tree if we hit the edge of a node, but that's not
>> supposed to be the common case. In previous code, the previous VMA could
>> just be obtained using the "previous VMA" pointer, however that pointer has
>> been remove since the maple tree changes and given a VMA, going to the
>> previous one using the maple tree is just as fast (as I'm told).
>
> As far as I can tell, find_vma_prev() already does a walk? I mean this is
> equivalent to find_vma() only retrieving the previous VMA right? I defer to
> Liam, but I'm not sure this would be that much more involved? Perhaps he
> can comment.
>
> An alternative is to create an iterator and use vma_prev(). I find it
> extremely clunky that we search for a VMA we already possess (and it's
> previous one) while not needing the the former.
>
> I'm not hugely familiar with the maple tree (perhaps Liam can comment) but
> I suspect that'd be more performant if that's the concern. Either way I
> would be surprised if this is the correct approach.
I see your point. I am not sure myself, the maple tree functions for both APIs
are indeed similar. We already have looked up the VMA being aligned down. If
there is a way to get the previous VMA quickly, given an existing VMA, I can
incorporate that change.
Ideally, if I had access to the ma_state used for lookup of the VMA being
aligned down, I could perhaps reuse that somehow. But when I checked, that
seemed a lot more invasive to pass that state down to these align functions.
But there is a merit to your suggestion itself in the sense it cuts down a few
more lines of code.
>> Considering this, I would keep the code as-is and perhaps you/we could
>> consider the replacement with another API in a subsequent patch as it does
>> the job for this patch.
>
> See above. I don't think this kind of comment is helpful in code
> review. Your disagreement above suffices, I've responded to it and of
> course if there is no other way this is fine.
>
> But I'd be surprised, and re-looking up a VMA we already have is just
> horrid. It's not really a nitpick, it's a code quality issue in my view.
>
> In any case, let's please try to avoid 'if you are bothered, write a follow
> up patch' style responses. If you disagree with something just say so, it's
> fine! :)
I wasn't disagreeing :) Just saying that the find_vma_prev() suggested in a
previous conversation with Liam fixes the issue (and has been tested a lot in
this series, on my side) so I was hoping to stick to that and we could iterate
more on that in the future.
However, after taking a deeper look at the maple tree, I'd like to give the
find_vma_intersection() option at least a try (with appropriate attribution to you).
Apologies if the response style in my previous email came across badly. That
wasn't my intent and I will try to improve myself.
[..]
>>>> + realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK);
>>>> + }
>>>> +
>>>> if (is_vm_hugetlb_page(vma))
>>>> return move_hugetlb_page_tables(vma, new_vma, old_addr,
>>>> new_addr, len);
>>>> @@ -565,6 +619,13 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
>>>>
>>>> mmu_notifier_invalidate_range_end(&range);
>>>>
>>>> + /*
>>>> + * Prevent negative return values when {old,new}_addr was realigned
>>>> + * but we broke out of the above loop for the first PMD itself.
>>>> + */
>>>> + if (len + old_addr < old_end)
>>>> + return 0;
>>>> +
>>>
>>> I find this a little iffy, I mean I see that if you align [old,new]_addr to
>>> PMD, then from then on in you're relying on the fact that the loop is just
>>> going from old_addr (now aligned) -> old_end and thus has the correct
>>> length.
>>>
>>> Can't we just fix this issue by correcting len? If you take my review above
>>> which checks len in [maybe_]realign_addr(), you could take that as a
>>> pointer and equally update that.
>>>
>>> Then you can drop this check.
>>
>> The drawback of adjusting len is it changes what move_page_tables() users
>> were previously expecting.
>>
>> I think we should look at the return value of move_page_tables() as well,
>> not just len independently.
>>
>> len is what the user requested.
>>
>> "len + old_addr - old_end" is how much was actually copied and is the return value.
>>
>> If everything was copied, old_addr == old_end and len is unchanged.
>
> Ah yeah I see, sorry I missed the fact we're returning a value, that does
> complicate things...
>
> If we retain the hugetlb logic, then we could work around the issue with
> that instance of len by storing the 'actual length' of the range in
> a new var actual_len and passing that.
>
> If we choose to instead just not do this for hugetlb (I wonder if the
> hugetlb handling code actually does the equivalent of this since surely
> these pages have to be handled a PMD at a time?) then we can drop the whole
> actual_len idea [see below on response to hugetlb thing].
Thanks. Yes, you are right. We should already b good with hugetlb handling as
it does appear that hugetlb_move_page_tables() does copy by huge_page_size(h),
so the old_addr should already be PMD-aligned for it to be able to do that.
[..]
>>>> return len + old_addr - old_end; /* how much done */
>>>> }
>>> Also I am concerned in the hugetlb case -> len is passed to
>>> move_hugetlb_page_tables() which is now strictly incorrect, I wonder if
>>> this could cause an issue?
>>>
>>> Correcting len seems the neat way of addressing this.
>>
>> That's a good point. I am wondering if we can just change that from:
>>
>> if (is_vm_hugetlb_page(vma))
>> return move_hugetlb_page_tables(vma, new_vma, old_addr,
>> new_addr, len);
>>
>> to:
>> if (is_vm_hugetlb_page(vma))
>> return move_hugetlb_page_tables(vma, new_vma, old_addr,
>> new_addr, old_addr - new_addr);
>>
>> Or, another option is to turn it off for hugetlb by just moving:
>>
>> if (len >= PMD_SIZE - (old_addr & ~PMD_MASK))
>> realign_addr(...);
>>
>> to after:
>>
>> if (is_vm_hugetlb_page(vma))
>> return move_hugetlb_page_tables(...);
>>
>> thanks,
>
> I think the actual_len solution should sort this right? If not maybe better
> to be conservative and disable for the hugetlb case (I'm not sure if this
> would help given you'd need to be PMD aligned anyway right?), so not to
> hold up the series.
>
> If we do decide not to include hugetlb (the endless 'special case' for so
> much code...) in this then we can drop the actual_len idea altogether.
>
> (Yes I realise it's ironic I'm suggesting deferring to a later patch here
> but there you go ;)
;-). Considering our discussion above that hugetlb mremap addresses should
always starts at a PMD boundary, maybe I can just add a warning to the if() like
so to detect any potential?
if (is_vm_hugetlb_page(vma)) {
WARN_ON_ONCE(old_addr - old_end != len);
return move_hugetlb_page_tables(vma, new_vma, old_addr,
new_addr, len);
}
Thank you so much and I learnt a lot from you and others in -mm community.
- Joel
On 6/20/23 17:16, Joel Fernandes wrote:
>
> Considering our discussion above that hugetlb mremap addresses should always
> starts at a PMD boundary, maybe I can just add a warning to the if() like so to
> detect any potential?
>
> if (is_vm_hugetlb_page(vma)) {
> WARN_ON_ONCE(old_addr - old_end != len);
Oops, I meant WARN_ON_ONCE(old_end - old_addr != len);
to make sure we did not mess up hugetlb mremaps.
thanks,
- Joel
On Tue, Jun 20, 2023 at 05:16:57PM -0400, Joel Fernandes wrote:
> Hi Lorenzo,
>
> > As far as I can tell, find_vma_prev() already does a walk? I mean this is
> > equivalent to find_vma() only retrieving the previous VMA right? I defer to
> > Liam, but I'm not sure this would be that much more involved? Perhaps he
> > can comment.
> >
> > An alternative is to create an iterator and use vma_prev(). I find it
> > extremely clunky that we search for a VMA we already possess (and it's
> > previous one) while not needing the the former.
> >
> > I'm not hugely familiar with the maple tree (perhaps Liam can comment) but
> > I suspect that'd be more performant if that's the concern. Either way I
> > would be surprised if this is the correct approach.
>
> I see your point. I am not sure myself, the maple tree functions for both
> APIs are indeed similar. We already have looked up the VMA being aligned
> down. If there is a way to get the previous VMA quickly, given an existing
> VMA, I can incorporate that change.
>
> Ideally, if I had access to the ma_state used for lookup of the VMA being
> aligned down, I could perhaps reuse that somehow. But when I checked, that
> seemed a lot more invasive to pass that state down to these align functions.
>
> But there is a merit to your suggestion itself in the sense it cuts down a
> few more lines of code.
Yeah it's thorny, the maple tree seems to add a separation between the vmi
and vma that didn't exist previously, and you'd have to thread through the
mas or vmi that was used in the first instance or end up having to rewalk
the tree anyway.
I have spesnt too much time staring at v6 code where it was trivial :)
I think given we have to walk the tree either way, we may as well do the
find_vma_intersection() [happy to stand corrected by Liam if this turns out
not to be less efficient but I don't think it is].
>
> > > Considering this, I would keep the code as-is and perhaps you/we could
> > > consider the replacement with another API in a subsequent patch as it does
> > > the job for this patch.
> >
> > See above. I don't think this kind of comment is helpful in code
> > review. Your disagreement above suffices, I've responded to it and of
> > course if there is no other way this is fine.
> >
> > But I'd be surprised, and re-looking up a VMA we already have is just
> > horrid. It's not really a nitpick, it's a code quality issue in my view.
> >
> > In any case, let's please try to avoid 'if you are bothered, write a follow
> > up patch' style responses. If you disagree with something just say so, it's
> > fine! :)
>
> I wasn't disagreeing :) Just saying that the find_vma_prev() suggested in a
> previous conversation with Liam fixes the issue (and has been tested a lot
> in this series, on my side) so I was hoping to stick to that and we could
> iterate more on that in the future.
>
> However, after taking a deeper look at the maple tree, I'd like to give the
> find_vma_intersection() option at least a try (with appropriate attribution
> to you).
Cheers appreciate it, sorry to be a pain and nitpicky but I think if that
works as well it could be quite a nice solution.
>
> Apologies if the response style in my previous email came across badly. That
> wasn't my intent and I will try to improve myself.
No worries, text is a sucky medium and likely I misinterpreted you in any
case. We're having a productive discussion which is what matters! :)
>
> [..]
>
> > > > > + realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK);
> > > > > + }
> > > > > +
> > > > > if (is_vm_hugetlb_page(vma))
> > > > > return move_hugetlb_page_tables(vma, new_vma, old_addr,
> > > > > new_addr, len);
> > > > > @@ -565,6 +619,13 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
> > > > >
> > > > > mmu_notifier_invalidate_range_end(&range);
> > > > >
> > > > > + /*
> > > > > + * Prevent negative return values when {old,new}_addr was realigned
> > > > > + * but we broke out of the above loop for the first PMD itself.
> > > > > + */
> > > > > + if (len + old_addr < old_end)
> > > > > + return 0;
> > > > > +
> > > >
> > > > I find this a little iffy, I mean I see that if you align [old,new]_addr to
> > > > PMD, then from then on in you're relying on the fact that the loop is just
> > > > going from old_addr (now aligned) -> old_end and thus has the correct
> > > > length.
> > > >
> > > > Can't we just fix this issue by correcting len? If you take my review above
> > > > which checks len in [maybe_]realign_addr(), you could take that as a
> > > > pointer and equally update that.
> > > >
> > > > Then you can drop this check.
> > >
> > > The drawback of adjusting len is it changes what move_page_tables() users
> > > were previously expecting.
> > >
> > > I think we should look at the return value of move_page_tables() as well,
> > > not just len independently.
> > >
> > > len is what the user requested.
> > >
> > > "len + old_addr - old_end" is how much was actually copied and is the return value.
> > >
> > > If everything was copied, old_addr == old_end and len is unchanged.
> >
> > Ah yeah I see, sorry I missed the fact we're returning a value, that does
> > complicate things...
> >
> > If we retain the hugetlb logic, then we could work around the issue with
> > that instance of len by storing the 'actual length' of the range in
> > a new var actual_len and passing that.
> >
> > If we choose to instead just not do this for hugetlb (I wonder if the
> > hugetlb handling code actually does the equivalent of this since surely
> > these pages have to be handled a PMD at a time?) then we can drop the whole
> > actual_len idea [see below on response to hugetlb thing].
>
> Thanks. Yes, you are right. We should already b good with hugetlb handling
> as it does appear that hugetlb_move_page_tables() does copy by
> huge_page_size(h), so the old_addr should already be PMD-aligned for it to
> be able to do that.
Cool, in which case we can drop the actual_len idea as this is really the
only place we'd need it.
>
> [..]
>
> > > > > return len + old_addr - old_end; /* how much done */
> > > > > }
> > > > Also I am concerned in the hugetlb case -> len is passed to
> > > > move_hugetlb_page_tables() which is now strictly incorrect, I wonder if
> > > > this could cause an issue?
> > > >
> > > > Correcting len seems the neat way of addressing this.
> > >
> > > That's a good point. I am wondering if we can just change that from:
> > >
> > > if (is_vm_hugetlb_page(vma))
> > > return move_hugetlb_page_tables(vma, new_vma, old_addr,
> > > new_addr, len);
> > >
> > > to:
> > > if (is_vm_hugetlb_page(vma))
> > > return move_hugetlb_page_tables(vma, new_vma, old_addr,
> > > new_addr, old_addr - new_addr);
> > >
> > > Or, another option is to turn it off for hugetlb by just moving:
> > >
> > > if (len >= PMD_SIZE - (old_addr & ~PMD_MASK))
> > > realign_addr(...);
> > >
> > > to after:
> > >
> > > if (is_vm_hugetlb_page(vma))
> > > return move_hugetlb_page_tables(...);
> > >
> > > thanks,
> >
> > I think the actual_len solution should sort this right? If not maybe better
> > to be conservative and disable for the hugetlb case (I'm not sure if this
> > would help given you'd need to be PMD aligned anyway right?), so not to
> > hold up the series.
> >
> > If we do decide not to include hugetlb (the endless 'special case' for so
> > much code...) in this then we can drop the actual_len idea altogether.
> >
> > (Yes I realise it's ironic I'm suggesting deferring to a later patch here
> > but there you go ;)
>
> ;-). Considering our discussion above that hugetlb mremap addresses should
> always starts at a PMD boundary, maybe I can just add a warning to the if()
> like so to detect any potential?
>
> if (is_vm_hugetlb_page(vma)) {
> WARN_ON_ONCE(old_addr - old_end != len);
> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> new_addr, len);
> }
>
Yeah looks good [ack your follow up that it should be old_end - old_addr],
maybe adding a comment explaining why this is a problem here too.
>
> Thank you so much and I learnt a lot from you and others in -mm community.
No worries, thanks very much for this patch series, this is a nice fixup
for a quite stupid failure we were experiencing before with a neat
solution. Your hard work is appreciated!
>
> - Joel
>
* Joel Fernandes <[email protected]> [230619 11:55]:
> Hi Lorenzo,
> Thanks for the review! I replied below:
>
> On 6/17/23 18:49, Lorenzo Stoakes wrote:
> > On Wed, May 31, 2023 at 10:08:01PM +0000, Joel Fernandes (Google) wrote:
> >> Recently, we see reports [1] of a warning that triggers due to
> >> move_page_tables() doing a downward and overlapping move on a
> >> mutually-aligned offset within a PMD. By mutual alignment, I
> >> mean the source and destination addresses of the mremap are at
> >> the same offset within a PMD.
> >>
> >> This mutual alignment along with the fact that the move is downward is
> >> sufficient to cause a warning related to having an allocated PMD that
> >> does not have PTEs in it.
> >>
> >> This warning will only trigger when there is mutual alignment in the
> >> move operation. A solution, as suggested by Linus Torvalds [2], is to
> >> initiate the copy process at the PMD level whenever such alignment is
> >> present. Implementing this approach will not only prevent the warning
> >> from being triggered, but it will also optimize the operation as this
> >> method should enhance the speed of the copy process whenever there's a
>
> [...]
>
> >> Suggested-by: Linus Torvalds <[email protected]>
> >> Signed-off-by: Joel Fernandes (Google) <[email protected]>
> >> ---
> >> mm/mremap.c | 61 +++++++++++++++++++++++++++++++++++++++++++++++++++++
> >> 1 file changed, 61 insertions(+)
> >>
> >> diff --git a/mm/mremap.c b/mm/mremap.c
> >> index 411a85682b58..bf355e4d6bd4 100644
> >> --- a/mm/mremap.c
> >> +++ b/mm/mremap.c
> >> @@ -478,6 +478,51 @@ static bool move_pgt_entry(enum pgt_entry entry, struct
> >> return moved;
> >> }
> >>
> >> +/*
> >> + * A helper to check if a previous mapping exists. Required for
> >> + * move_page_tables() and realign_addr() to determine if a previous mapping
> >> + * exists before we can do realignment optimizations.
> >> + */
> >> +static bool can_align_down(struct vm_area_struct *vma, unsigned long addr_to_align,
> >> + unsigned long mask)
> >> +{
> >> + unsigned long addr_masked = addr_to_align & mask;
> >> + struct vm_area_struct *prev = NULL, *cur = NULL;
> >> +
> >> + /*
> >> + * If @addr_to_align of either source or destination is not the beginning
> >> + * of the corresponding VMA, we can't align down or we will destroy part
> >> + * of the current mapping.
> >> + */
> >> + if (vma->vm_start != addr_to_align)
> >> + return false;
> >
> > See below, I think we can eliminate this check.
> >
> >> +
> >> + /*
> >> + * Find the VMA before @vma to see if it subsumes the masked address.
> >> + * The mmap write lock is held here so the lookup is safe.
> >> + */
> >> + cur = find_vma_prev(vma->vm_mm, vma->vm_start, &prev);
> >> + if (WARN_ON_ONCE(cur != vma))
> >> + return false;
> >> +
> >> + return !prev || prev->vm_end <= addr_masked;
> >
> > This is a bit clunky, and I don't think we need the WARN_ON_ONCE() check if
> > we're under the mmap_lock.
> >
> > How about something like:-
> >
> > return find_vma_intersection(vma->mm, addr_masked, vma->vm_start) == NULL;
> >
> > Which explicitly asserts that the range in [addr_masked, vma->vm_start) is
> > empty.
> >
> > But actually, we should be able to go further and replace the previous
> > check with:-
> >
> > return find_vma_intersection(vma->mm, addr_masked, addr_to_align) == NULL;
> >
> > Which will fail if addr_to_align is offset within the VMA.
>
> Your suggestion would mean that we do a full VMA search starting from the
> root. That would not be a nice thing if say we've 1000s of VMAs?
>
> Actually Liam told me to use find_vma_prev() because given a VMA, the maple
> tree would not have to work that hard for the common case to find the
> previous VMA. Per conversing with him, there is a chance we may have to go
> one step above in the tree if we hit the edge of a node, but that's not
> supposed to be the common case. In previous code, the previous VMA could
> just be obtained using the "previous VMA" pointer, however that pointer has
> been remove since the maple tree changes and given a VMA, going to the
> previous one using the maple tree is just as fast (as I'm told).
I think there's been a bit of a miscommunication on that..
If you have already found the VMA and are using the maple state, then
it's very little effort to get the next/prev. Leaf nodes can hold 16
entries/NULL ranges, so the chances to go to the next/prev is usually in
the cpu cache already.. if you go up a level in the tree, then you will
have 10 nodes each with 16 entries each, etc, etc.. So the chances of
being on an edge node and having to walk up multiple levels to get to
the prev/next becomes rather rare.. and if you've just walked down, the
nodes on the way up will still be cached.
Here, you're not using the maple state but searching for an address
using find_vma_prev(), but internally, that function does use a maple
state to get you the previous. So you are looking up the VMA from the
root, but the prev will very likely be in the CPU cache.
Assuming the worst case tree (each VMA has a gap next to it, not really
going to happen as they tend to be grouped together), then we are
looking at a 4 level tree to get to 8,000 VMAs. 5 levels gets you a
minimum 80,000. I've never seen a tree of height 6 in the wild, but you
can fit 1.6M to 800K in one.
I think the code is fine, but I wanted to clarify what we discussed.
>
> Considering this, I would keep the code as-is and perhaps you/we could
> consider the replacement with another API in a subsequent patch as it does
> the job for this patch.
>
> >> + unsigned long *new_addr, struct vm_area_struct *new_vma,
> >> + unsigned long mask)
> >> +{
> >> + bool mutually_aligned = (*old_addr & ~mask) == (*new_addr & ~mask);
> >> +
> >> + if ((*old_addr & ~mask) && mutually_aligned
> >
> > I may be misunderstanding something here, but doesn't the first condition
> > here disallow for offset into PMD == 0? Why?
>
> Because in such a situation, the alignment is already done and there's
> nothing to align. The patch wants to align down to the PMD and we would not
> want to waste CPU cycles if there's nothing to do.
>
> >> + && can_align_down(old_vma, *old_addr, mask)
> >> + && can_align_down(new_vma, *new_addr, mask)) {
> >> + *old_addr = *old_addr & mask;
> >> + *new_addr = *new_addr & mask;
> >> + }
> >> +}
> >> +
> >> unsigned long move_page_tables(struct vm_area_struct *vma,
> >> unsigned long old_addr, struct vm_area_struct *new_vma,
> >> unsigned long new_addr, unsigned long len,
> >> @@ -493,6 +538,15 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
> >>
> >> old_end = old_addr + len;
> >>
> >> + /*
> >> + * If possible, realign addresses to PMD boundary for faster copy.
> >> + * Don't align for intra-VMA moves as we may destroy existing mappings.
> >> + */
> >> + if ((vma != new_vma)
> >
> > Nit but these parens aren't needed.
>
> Sure, I can drop the parens.
>
> > Also if we're deferring the decision as
> > to whether we realign to this function, why are we doing this check here
> > and not here?
>
> Hmm, well the function name is realign_addr() so I kept some of the initial
> checks outside of it where we should "obviously" not realign. I could do
> what you're suggesting and change it to try_realign_addr() or something. And
> move those checks in there. That would be a bit better.
>
> > It feels like it'd be neater to keep all the conditions (including the
> > length one) together in one place.
> >
> >
> >> + && (len >= PMD_SIZE - (old_addr & ~PMD_MASK))) {
>
> Well, yeah maybe. I'll look into it, thanks.
>
> > You don't mention this condition in the above comment (if we have this
> > altogether as part of the realign function could comment separately there)
>
> Ok, sounds good -- I will add a comment with some of the explanation above.
>
> > - so we only go ahead and do this optimisation if the length of the remap
> > is such that the entire of old_addr -> end of its PMD (and thus the same
> > for new_addr) is copied?
>
> Yes, correct. And in the future that could also be optimized (if say there
> is no subsequent mapping, so we can copy the tail PMD as well, however one
> step at a time and all that.)
>
> > I may be missing something/being naive here, but can't we just do a similar
> > check to the one done for space _below_ the VMA to see if [end, (end of
> > PMD)) is equally empty?
>
> We can, but the warning that was triggering does not really need that to be
> silenced. I am happy to do that in a later patch if needed, or you can. ;-)
> But I'd like to keep the risk low since this was itself hard enough to get
> right.
>
> >> + realign_addr(&old_addr, vma, &new_addr, new_vma, PMD_MASK);
> >> + }
> >> +
> >> if (is_vm_hugetlb_page(vma))
> >> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> >> new_addr, len);
> >> @@ -565,6 +619,13 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
> >>
> >> mmu_notifier_invalidate_range_end(&range);
> >>
> >> + /*
> >> + * Prevent negative return values when {old,new}_addr was realigned
> >> + * but we broke out of the above loop for the first PMD itself.
> >> + */
> >> + if (len + old_addr < old_end)
> >> + return 0;
> >> +
> >
> > I find this a little iffy, I mean I see that if you align [old,new]_addr to
> > PMD, then from then on in you're relying on the fact that the loop is just
> > going from old_addr (now aligned) -> old_end and thus has the correct
> > length.
> >
> > Can't we just fix this issue by correcting len? If you take my review above
> > which checks len in [maybe_]realign_addr(), you could take that as a
> > pointer and equally update that.
> >
> > Then you can drop this check.
>
> The drawback of adjusting len is it changes what move_page_tables() users
> were previously expecting.
>
> I think we should look at the return value of move_page_tables() as well,
> not just len independently.
>
> len is what the user requested.
>
> "len + old_addr - old_end" is how much was actually copied and is the return value.
>
> If everything was copied, old_addr == old_end and len is unchanged.
>
> The users of move_page_tables(), like move_vma() should not care whether we
> copied a full PMD or not. In fact telling them anything like may cause
> problems with the interpretation of the return value I think.
>
> They asked us to copy len, did we copy it? hell yeah.
>
> Note that after the first loop iteration's PMD copy, old_addr is now at the
> PMD boundary and the functionality of this function is not changed with this
> patch. We end up doing a PMD-copy just like we used to without this patch.
> So this patch does not really change anything from before.
>
> The following are the cases:
>
> 1. If we realign and copy, yes we copied a PMD, but really it was to satisfy
> the requested length. In this situation, "len + old_addr - old_end" is
> accurate and just like before. We copied whatever the user requested. Yes we
> copied a little more, but who cares? We copied into a mapping that does not
> exist anyway. It may be absurd for us to return a len that is greater than
> the requested len IMO.
>
> 2. If there are no errors (example first PMD copy did not fail), "len +
> old_addr - old_end" is identical to what it was without this patch -- as it
> should be. That's true whether we realigned or not.
>
> 3. If we realigned and the first PMD copy failed (unlikely error) -- that's
> where there's a problem. We would end up returning a negative value. That's
> what Linus found and suggested to correct. Because (old_addr - old_end) will
> be greater than len in such a situation, however unlikely.
>
> >> return len + old_addr - old_end; /* how much done */
> >> }
> > Also I am concerned in the hugetlb case -> len is passed to
> > move_hugetlb_page_tables() which is now strictly incorrect, I wonder if
> > this could cause an issue?
> >
> > Correcting len seems the neat way of addressing this.
>
> That's a good point. I am wondering if we can just change that from:
>
> if (is_vm_hugetlb_page(vma))
> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> new_addr, len);
>
> to:
> if (is_vm_hugetlb_page(vma))
> return move_hugetlb_page_tables(vma, new_vma, old_addr,
> new_addr, old_addr - new_addr);
>
> Or, another option is to turn it off for hugetlb by just moving:
>
> if (len >= PMD_SIZE - (old_addr & ~PMD_MASK))
> realign_addr(...);
>
> to after:
>
> if (is_vm_hugetlb_page(vma))
> return move_hugetlb_page_tables(...);
>
> thanks,
>
> - Joel
>
On Tue, Jun 27, 2023 at 01:56:09PM -0400, Liam R. Howlett wrote:
[snip]
> > > How about something like:-
> > >
> > > return find_vma_intersection(vma->mm, addr_masked, vma->vm_start) == NULL;
> > >
> > > Which explicitly asserts that the range in [addr_masked, vma->vm_start) is
> > > empty.
> > >
> > > But actually, we should be able to go further and replace the previous
> > > check with:-
> > >
> > > return find_vma_intersection(vma->mm, addr_masked, addr_to_align) == NULL;
> > >
> > > Which will fail if addr_to_align is offset within the VMA.
> >
> > Your suggestion would mean that we do a full VMA search starting from the
> > root. That would not be a nice thing if say we've 1000s of VMAs?
> >
> > Actually Liam told me to use find_vma_prev() because given a VMA, the maple
> > tree would not have to work that hard for the common case to find the
> > previous VMA. Per conversing with him, there is a chance we may have to go
> > one step above in the tree if we hit the edge of a node, but that's not
> > supposed to be the common case. In previous code, the previous VMA could
> > just be obtained using the "previous VMA" pointer, however that pointer has
> > been remove since the maple tree changes and given a VMA, going to the
> > previous one using the maple tree is just as fast (as I'm told).
>
> I think there's been a bit of a miscommunication on that..
>
> If you have already found the VMA and are using the maple state, then
> it's very little effort to get the next/prev. Leaf nodes can hold 16
> entries/NULL ranges, so the chances to go to the next/prev is usually in
> the cpu cache already.. if you go up a level in the tree, then you will
> have 10 nodes each with 16 entries each, etc, etc.. So the chances of
> being on an edge node and having to walk up multiple levels to get to
> the prev/next becomes rather rare.. and if you've just walked down, the
> nodes on the way up will still be cached.
>
> Here, you're not using the maple state but searching for an address
> using find_vma_prev(), but internally, that function does use a maple
> state to get you the previous. So you are looking up the VMA from the
> root, but the prev will very likely be in the CPU cache.
>
> Assuming the worst case tree (each VMA has a gap next to it, not really
> going to happen as they tend to be grouped together), then we are
> looking at a 4 level tree to get to 8,000 VMAs. 5 levels gets you a
> minimum 80,000. I've never seen a tree of height 6 in the wild, but you
> can fit 1.6M to 800K in one.
>
> I think the code is fine, but I wanted to clarify what we discussed.
Would the same apply to find_vma_intersection(), as they equally searches
from the root and allows the code to be made fairly succinct?
I really am not a huge fan of find_vma_prev() searching for a VMA you
already have just to get the previous one... would at lesat like to use
vma_prev() on a newly defined vmi, but if find_vma_intersection() is fine
then can reduce code to this.
[snip]
* Lorenzo Stoakes <[email protected]> [230627 14:02]:
> On Tue, Jun 27, 2023 at 01:56:09PM -0400, Liam R. Howlett wrote:
> [snip]
> > > > How about something like:-
> > > >
> > > > return find_vma_intersection(vma->mm, addr_masked, vma->vm_start) == NULL;
> > > >
> > > > Which explicitly asserts that the range in [addr_masked, vma->vm_start) is
> > > > empty.
> > > >
> > > > But actually, we should be able to go further and replace the previous
> > > > check with:-
> > > >
> > > > return find_vma_intersection(vma->mm, addr_masked, addr_to_align) == NULL;
> > > >
> > > > Which will fail if addr_to_align is offset within the VMA.
> > >
> > > Your suggestion would mean that we do a full VMA search starting from the
> > > root. That would not be a nice thing if say we've 1000s of VMAs?
> > >
> > > Actually Liam told me to use find_vma_prev() because given a VMA, the maple
> > > tree would not have to work that hard for the common case to find the
> > > previous VMA. Per conversing with him, there is a chance we may have to go
> > > one step above in the tree if we hit the edge of a node, but that's not
> > > supposed to be the common case. In previous code, the previous VMA could
> > > just be obtained using the "previous VMA" pointer, however that pointer has
> > > been remove since the maple tree changes and given a VMA, going to the
> > > previous one using the maple tree is just as fast (as I'm told).
> >
> > I think there's been a bit of a miscommunication on that..
> >
> > If you have already found the VMA and are using the maple state, then
> > it's very little effort to get the next/prev. Leaf nodes can hold 16
> > entries/NULL ranges, so the chances to go to the next/prev is usually in
> > the cpu cache already.. if you go up a level in the tree, then you will
> > have 10 nodes each with 16 entries each, etc, etc.. So the chances of
> > being on an edge node and having to walk up multiple levels to get to
> > the prev/next becomes rather rare.. and if you've just walked down, the
> > nodes on the way up will still be cached.
> >
> > Here, you're not using the maple state but searching for an address
> > using find_vma_prev(), but internally, that function does use a maple
> > state to get you the previous. So you are looking up the VMA from the
> > root, but the prev will very likely be in the CPU cache.
> >
> > Assuming the worst case tree (each VMA has a gap next to it, not really
> > going to happen as they tend to be grouped together), then we are
> > looking at a 4 level tree to get to 8,000 VMAs. 5 levels gets you a
> > minimum 80,000. I've never seen a tree of height 6 in the wild, but you
> > can fit 1.6M to 800K in one.
> >
> > I think the code is fine, but I wanted to clarify what we discussed.
>
> Would the same apply to find_vma_intersection(), as they equally searches
> from the root and allows the code to be made fairly succinct?
I think so.
>
> I really am not a huge fan of find_vma_prev() searching for a VMA you
> already have just to get the previous one... would at lesat like to use
> vma_prev() on a newly defined vmi, but if find_vma_intersection() is fine
> then can reduce code to this.
find_vma_intersection() will work as well.
> [snip]