Re: [Qemu-devel] [PATCH] mm: Loosen MADV_NOHUGEPAGE to enable Qemu postcopy on s390

[Dr. David Alan Gilbert]

* Christian Borntraeger (address@hidden) wrote:
> On 11/18/2015 02:31 PM, Vlastimil Babka wrote:
> > [CC += address@hidden
> > 
> > Since this is a kernel-user-space API change, please CC address@hidden The 
> > kernel
> > source file Documentation/SubmitChecklist notes that all Linux kernel 
> > patches
> > that change userspace interfaces should be CCed to address@hidden, 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 <address@hidden>
> >> Reviewed-by: Andrea Arcangeli <address@hidden>
> > 
> > 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" <address@hidden>
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 <address@hidden>
---
 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 / address@hidden / Manchester, UK