* Re: [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390
2015-11-19 8:22 ` [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390 Christian Borntraeger
@ 2015-11-19 9:31 ` Vlastimil Babka
2015-11-19 9:43 ` Dr. David Alan Gilbert
2015-11-19 13:10 ` Dr. David Alan Gilbert
2 siblings, 0 replies; 5+ messages in thread
From: Vlastimil Babka @ 2015-11-19 9:31 UTC (permalink / raw)
To: Christian Borntraeger, Jason J. Herne, linux-s390
Cc: aarcange, linux-man, Juan Quintela, linux-api, qemu-devel,
Dr. David Alan Gilbert, linux-mm, akpm
On 11/19/2015 09:22 AM, Christian Borntraeger wrote:
> On 11/18/2015 02:31 PM, Vlastimil Babka wrote:
>> [CC += linux-api@vger.kernel.org]
>> Anyway, I agree that it doesn't make sense to fail madvise when the given flag
>> is already set. On the other hand, I don't think the userspace app should fail
>> just because of madvise failing? It should in general be an advice that the
>> kernel is also strictly speaking free to ignore as it shouldn't affect
>> correctnes, just performance. Yeah, there are exceptions today like
>> MADV_DONTNEED, but that shouldn't apply to hugepages?
>> So I think Qemu needs fixing too.
>
> yes, I agree. David, Juan. I think The postcopy code should not fail if the madvise.
> Can you fix that?
>
> Also what happens if the kernel is build
>> without CONFIG_TRANSPARENT_HUGEPAGE? Then madvise also returns EINVAL,
>
> Does it? To me it looks more like we would trigger a kernel bug.
>
> mm/madvise.c:
> case MADV_HUGEPAGE:
> case MADV_NOHUGEPAGE:
> error = hugepage_madvise(vma, &new_flags, behavior); <-----
> if (error)
> goto out;
> break;
> }
>
>
> include/linux/huge_mm.h:
> static inline int hugepage_madvise(struct vm_area_struct *vma,
> unsigned long *vm_flags, int advice)
> {
> BUG();
> return 0;
> }
>
> If we just remove the BUG() statement the code would actually be correct
> in simply ignoring an MADVISE it cannot handle. If you agree, I can
> spin a patch.
Yeah this looks suspicious at first, but the code is not reachable
thanks to madvise_behavior_valid() returning false:
...
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
case MADV_HUGEPAGE:
case MADV_NOHUGEPAGE:
#endif
case MADV_DONTDUMP:
case MADV_DODUMP:
return true;
default:
return false;
I think the BUG() is pointless (KSM doesn't use it) but not wrong. I
wouldn't object to removal.
>
>
>> how does Qemu handle that?
>
> The normal qemu startup ignores the return value of the madvise. Only the
> recent post migration changes want to disable huge pages for userfaultd.
> And this code checks the return value. And yes, we should change that
> in QEMU.
Great, thanks :)
^ permalink raw reply [flat|nested] 5+ messages in thread* Re: [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390
2015-11-19 8:22 ` [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390 Christian Borntraeger
2015-11-19 9:31 ` Vlastimil Babka
@ 2015-11-19 9:43 ` Dr. David Alan Gilbert
2015-11-19 13:10 ` Dr. David Alan Gilbert
2 siblings, 0 replies; 5+ messages in thread
From: Dr. David Alan Gilbert @ 2015-11-19 9:43 UTC (permalink / raw)
To: Christian Borntraeger
Cc: aarcange, linux-s390, linux-man, linux-api, Juan Quintela,
qemu-devel, linux-mm, Jason J. Herne, akpm, Vlastimil Babka
* Christian Borntraeger (borntraeger@de.ibm.com) wrote:
> On 11/18/2015 02:31 PM, Vlastimil Babka wrote:
> > [CC += linux-api@vger.kernel.org]
> >
> > Since this is a kernel-user-space API change, please CC linux-api@. The kernel
> > source file Documentation/SubmitChecklist notes that all Linux kernel patches
> > that change userspace interfaces should be CCed to linux-api@vger.kernel.org, so
> > that the various parties who are interested in API changes are informed. For
> > further information, see https://www.kernel.org/doc/man-pages/linux-api-ml.html
> >
> > On 11/12/2015 04:18 PM, Jason J. Herne wrote:
> >> MADV_NOHUGEPAGE processing is too restrictive. kvm already disables
> >> hugepage but hugepage_madvise() takes the error path when we ask to turn
> >> on the MADV_NOHUGEPAGE bit and the bit is already on. This causes Qemu's
> >> new postcopy migration feature to fail on s390 because its first action is
> >> to madvise the guest address space as NOHUGEPAGE. This patch modifies the
> >> code so that the operation succeeds without error now.
> >>
> >> Signed-off-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
> >> Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
> >
> > Looks like the manpage should be fine, as it wasn't very specific wrt these
> > madvise flags. The only thing that potentially applies is:
> >
> > "EINVAL advice is not a valid."
> >
> > which itself looks like it needs fixing. Valid what, value? As in completely
> > unknown flags, or flags not valid for the given vma?
> >
> > Anyway, I agree that it doesn't make sense to fail madvise when the given flag
> > is already set. On the other hand, I don't think the userspace app should fail
> > just because of madvise failing? It should in general be an advice that the
> > kernel is also strictly speaking free to ignore as it shouldn't affect
> > correctnes, just performance. Yeah, there are exceptions today like
> > MADV_DONTNEED, but that shouldn't apply to hugepages?
> > So I think Qemu needs fixing too.
>
> yes, I agree. David, Juan. I think The postcopy code should not fail if the madvise.
> Can you fix that?
Yes, I can change that.
> Also what happens if the kernel is build
> > without CONFIG_TRANSPARENT_HUGEPAGE? Then madvise also returns EINVAL,
>
> Does it? To me it looks more like we would trigger a kernel bug.
Yes, it does return EINVAL; it's a shame - there's no way to distinguish between
a kernel that doesn't have hugepage and a screwup in the addresses passed;
everything gets squashed into EINVAL.
Dave
> mm/madvise.c:
> case MADV_HUGEPAGE:
> case MADV_NOHUGEPAGE:
> error = hugepage_madvise(vma, &new_flags, behavior); <-----
> if (error)
> goto out;
> break;
> }
>
>
> include/linux/huge_mm.h:
> static inline int hugepage_madvise(struct vm_area_struct *vma,
> unsigned long *vm_flags, int advice)
> {
> BUG();
> return 0;
> }
>
> If we just remove the BUG() statement the code would actually be correct
> in simply ignoring an MADVISE it cannot handle. If you agree, I can
> spin a patch.
>
>
>
>
> > how does Qemu handle that?
>
> The normal qemu startup ignores the return value of the madvise. Only the
> recent post migration changes want to disable huge pages for userfaultd.
> And this code checks the return value. And yes, we should change that
> in QEMU.
>
>
>
>
> >
> >> ---
> >> mm/huge_memory.c | 4 ++--
> >> 1 file changed, 2 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> >> index c29ddeb..62fe06b 100644
> >> --- a/mm/huge_memory.c
> >> +++ b/mm/huge_memory.c
> >> @@ -2009,7 +2009,7 @@ int hugepage_madvise(struct vm_area_struct *vma,
> >> /*
> >> * Be somewhat over-protective like KSM for now!
> >> */
> >> - if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
> >> + if (*vm_flags & VM_NO_THP)
> >> return -EINVAL;
> >> *vm_flags &= ~VM_NOHUGEPAGE;
> >> *vm_flags |= VM_HUGEPAGE;
> >> @@ -2025,7 +2025,7 @@ int hugepage_madvise(struct vm_area_struct *vma,
> >> /*
> >> * Be somewhat over-protective like KSM for now!
> >> */
> >> - if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
> >> + if (*vm_flags & VM_NO_THP)
> >> return -EINVAL;
> >> *vm_flags &= ~VM_HUGEPAGE;
> >> *vm_flags |= VM_NOHUGEPAGE;
> >>
> >
>
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
^ permalink raw reply [flat|nested] 5+ messages in thread* Re: [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390
2015-11-19 8:22 ` [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390 Christian Borntraeger
2015-11-19 9:31 ` Vlastimil Babka
2015-11-19 9:43 ` Dr. David Alan Gilbert
@ 2015-11-19 13:10 ` Dr. David Alan Gilbert
2015-11-19 14:30 ` Jason J. Herne
2 siblings, 1 reply; 5+ messages in thread
From: Dr. David Alan Gilbert @ 2015-11-19 13:10 UTC (permalink / raw)
To: Christian Borntraeger
Cc: aarcange, Juan Quintela, qemu-devel, Jason J. Herne, akpm,
Vlastimil Babka
* Christian Borntraeger (borntraeger@de.ibm.com) wrote:
> On 11/18/2015 02:31 PM, Vlastimil Babka wrote:
> > [CC += linux-api@vger.kernel.org]
> >
> > Since this is a kernel-user-space API change, please CC linux-api@. The kernel
> > source file Documentation/SubmitChecklist notes that all Linux kernel patches
> > that change userspace interfaces should be CCed to linux-api@vger.kernel.org, so
> > that the various parties who are interested in API changes are informed. For
> > further information, see https://www.kernel.org/doc/man-pages/linux-api-ml.html
> >
> > On 11/12/2015 04:18 PM, Jason J. Herne wrote:
> >> MADV_NOHUGEPAGE processing is too restrictive. kvm already disables
> >> hugepage but hugepage_madvise() takes the error path when we ask to turn
> >> on the MADV_NOHUGEPAGE bit and the bit is already on. This causes Qemu's
> >> new postcopy migration feature to fail on s390 because its first action is
> >> to madvise the guest address space as NOHUGEPAGE. This patch modifies the
> >> code so that the operation succeeds without error now.
> >>
> >> Signed-off-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
> >> Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
> >
> > Looks like the manpage should be fine, as it wasn't very specific wrt these
> > madvise flags. The only thing that potentially applies is:
> >
> > "EINVAL advice is not a valid."
> >
> > which itself looks like it needs fixing. Valid what, value? As in completely
> > unknown flags, or flags not valid for the given vma?
> >
> > Anyway, I agree that it doesn't make sense to fail madvise when the given flag
> > is already set. On the other hand, I don't think the userspace app should fail
> > just because of madvise failing? It should in general be an advice that the
> > kernel is also strictly speaking free to ignore as it shouldn't affect
> > correctnes, just performance. Yeah, there are exceptions today like
> > MADV_DONTNEED, but that shouldn't apply to hugepages?
> > So I think Qemu needs fixing too.
>
> yes, I agree. David, Juan. I think The postcopy code should not fail if the madvise.
> Can you fix that?
(cutting down cc list a bit)
Can you tell me if the following works for you:
From 545809a18fa768eccdaafe9bd842490c3390b00c Mon Sep 17 00:00:00 2001
From: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
Date: Thu, 19 Nov 2015 12:05:36 +0000
Subject: [PATCH] Assume madvise for (no)hugepage works
madvise() returns EINVAL in the case of many failures, but also
returns it in cases where the host kernel doesn't have THP enabled.
Postcopy only really cares that THP is off before it detects faults,
and turns it back on afterwards; so we're going to have
to assume that if the madvise fails then the host just doesn't do
THP and we can carry on with the postcopy.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
---
migration/postcopy-ram.c | 10 ++--------
1 file changed, 2 insertions(+), 8 deletions(-)
diff --git a/migration/postcopy-ram.c b/migration/postcopy-ram.c
index 22d6b18..3946aa9 100644
--- a/migration/postcopy-ram.c
+++ b/migration/postcopy-ram.c
@@ -241,10 +241,7 @@ static int cleanup_range(const char *block_name, void *host_addr,
* We turned off hugepage for the precopy stage with postcopy enabled
* we can turn it back on now.
*/
- if (qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE)) {
- error_report("%s HUGEPAGE: %s", __func__, strerror(errno));
- return -1;
- }
+ qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
/*
* We can also turn off userfault now since we should have all the
@@ -345,10 +342,7 @@ static int nhp_range(const char *block_name, void *host_addr,
* do delete areas of the page, even if THP thinks a hugepage would
* be a good idea, so force hugepages off.
*/
- if (qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE)) {
- error_report("%s: NOHUGEPAGE: %s", __func__, strerror(errno));
- return -1;
- }
+ qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
return 0;
}
--
2.5.0
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK
^ permalink raw reply related [flat|nested] 5+ messages in thread* Re: [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390
2015-11-19 13:10 ` Dr. David Alan Gilbert
@ 2015-11-19 14:30 ` Jason J. Herne
0 siblings, 0 replies; 5+ messages in thread
From: Jason J. Herne @ 2015-11-19 14:30 UTC (permalink / raw)
To: Dr. David Alan Gilbert, Christian Borntraeger
Cc: aarcange, akpm, qemu-devel, Vlastimil Babka, Juan Quintela
On 11/19/2015 08:10 AM, Dr. David Alan Gilbert wrote:
> * Christian Borntraeger (borntraeger@de.ibm.com) wrote:
>> On 11/18/2015 02:31 PM, Vlastimil Babka wrote:
>>> [CC += linux-api@vger.kernel.org]
...
> Can you tell me if the following works for you:
>
>
> From 545809a18fa768eccdaafe9bd842490c3390b00c Mon Sep 17 00:00:00 2001
> From: "Dr. David Alan Gilbert" <dgilbert@redhat.com>
> Date: Thu, 19 Nov 2015 12:05:36 +0000
> Subject: [PATCH] Assume madvise for (no)hugepage works
>
> madvise() returns EINVAL in the case of many failures, but also
> returns it in cases where the host kernel doesn't have THP enabled.
> Postcopy only really cares that THP is off before it detects faults,
> and turns it back on afterwards; so we're going to have
> to assume that if the madvise fails then the host just doesn't do
> THP and we can carry on with the postcopy.
>
> Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
> ---
> migration/postcopy-ram.c | 10 ++--------
> 1 file changed, 2 insertions(+), 8 deletions(-)
>
> diff --git a/migration/postcopy-ram.c b/migration/postcopy-ram.c
> index 22d6b18..3946aa9 100644
> --- a/migration/postcopy-ram.c
> +++ b/migration/postcopy-ram.c
> @@ -241,10 +241,7 @@ static int cleanup_range(const char *block_name, void *host_addr,
> * We turned off hugepage for the precopy stage with postcopy enabled
> * we can turn it back on now.
> */
> - if (qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE)) {
> - error_report("%s HUGEPAGE: %s", __func__, strerror(errno));
> - return -1;
> - }
> + qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
>
> /*
> * We can also turn off userfault now since we should have all the
> @@ -345,10 +342,7 @@ static int nhp_range(const char *block_name, void *host_addr,
> * do delete areas of the page, even if THP thinks a hugepage would
> * be a good idea, so force hugepages off.
> */
> - if (qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE)) {
> - error_report("%s: NOHUGEPAGE: %s", __func__, strerror(errno));
> - return -1;
> - }
> + qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
>
> return 0;
> }
>
I tried this without my madvise kernel patch, and was able to get by the
problem as expected.
We still need the kernel patch set "Allow gmap fault to retry" as
posted to linux-mm to get
userfaultfd support playing nicely with s390 async page faults. But that
is a separate problem
entirely.
Tested-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
--
-- Jason J. Herne (jjherne@linux.vnet.ibm.com)
^ permalink raw reply [flat|nested] 5+ messages in thread