On Tue, 22 Mar 2022 16:53:05 +0000 "Dr. David Alan Gilbert" <dgilb...@redhat.com> wrote:
> * Andrew Deason (adea...@sinenomine.net) wrote: > > On Wed, 16 Mar 2022 10:41:41 +0100 > > David Hildenbrand <da...@redhat.com> wrote: > > > > > On 16.03.22 10:37, Dr. David Alan Gilbert wrote: > > > > * Peter Maydell (peter.mayd...@linaro.org) wrote: > > > >> On Wed, 16 Mar 2022 at 07:53, David Hildenbrand <da...@redhat.com> > > > >> wrote: > > > >>> > > > >>> On 16.03.22 05:04, Andrew Deason wrote: > > > >>>> We have a thin wrapper around madvise, called qemu_madvise, which > > > >>>> provides consistent behavior for the !CONFIG_MADVISE case, and works > > > >>>> around some platform-specific quirks (some platforms only provide > > > >>>> posix_madvise, and some don't offer all 'advise' types). This > > > >>>> specific > > > >>>> caller of madvise has never used it, tracing back to its original > > > >>>> introduction in commit e0b266f01dd2 ("migration_completion: Take > > > >>>> current state"). > > > >>>> > > > >>>> Call qemu_madvise here, to follow the same logic as all of our other > > > >>>> madvise callers. This slightly changes the behavior for > > > >>>> !CONFIG_MADVISE (EINVAL instead of ENOSYS, and a slightly different > > > >>>> error message), but this is now more consistent with other callers > > > >>>> that use qemu_madvise. > > > >>>> > > > >>>> Signed-off-by: Andrew Deason <adea...@sinenomine.net> > > > >>>> --- > > > >>>> Looking at the history of commits that touch this madvise() call, it > > > >>>> doesn't _look_ like there's any reason to be directly calling > > > >>>> madvise vs > > > >>>> qemu_advise (I don't see anything mentioned), but I'm not sure. > > > >>>> > > > >>>> softmmu/physmem.c | 12 ++---------- > > > >>>> 1 file changed, 2 insertions(+), 10 deletions(-) > > > >>>> > > > >>>> diff --git a/softmmu/physmem.c b/softmmu/physmem.c > > > >>>> index 43ae70fbe2..900c692b5e 100644 > > > >>>> --- a/softmmu/physmem.c > > > >>>> +++ b/softmmu/physmem.c > > > >>>> @@ -3584,40 +3584,32 @@ int ram_block_discard_range(RAMBlock *rb, > > > >>>> uint64_t start, size_t length) > > > >>>> rb->idstr, start, length, ret); > > > >>>> goto err; > > > >>>> #endif > > > >>>> } > > > >>>> if (need_madvise) { > > > >>>> /* For normal RAM this causes it to be unmapped, > > > >>>> * for shared memory it causes the local mapping to > > > >>>> disappear > > > >>>> * and to fall back on the file contents (which we just > > > >>>> * fallocate'd away). > > > >>>> */ > > > >>>> -#if defined(CONFIG_MADVISE) > > > >>>> if (qemu_ram_is_shared(rb) && rb->fd < 0) { > > > >>>> - ret = madvise(host_startaddr, length, > > > >>>> QEMU_MADV_REMOVE); > > > >>>> + ret = qemu_madvise(host_startaddr, length, > > > >>>> QEMU_MADV_REMOVE); > > > >>>> } else { > > > >>>> - ret = madvise(host_startaddr, length, > > > >>>> QEMU_MADV_DONTNEED); > > > >>>> + ret = qemu_madvise(host_startaddr, length, > > > >>>> QEMU_MADV_DONTNEED); > > > >>> > > > >>> posix_madvise(QEMU_MADV_DONTNEED) has completely different semantics > > > >>> then madvise() -- it's not a discard that we need here. > > > >>> > > > >>> So ram_block_discard_range() would now succeed in environments (BSD?) > > > >>> where it's supposed to fail. > > > >>> > > > >>> So AFAIKs this isn't sane. > > > >> > > > >> But CONFIG_MADVISE just means "host has madvise()"; it doesn't imply > > > >> "this is a Linux madvise() with MADV_DONTNEED". Solaris madvise() > > > >> doesn't seem to have MADV_DONTNEED at all; a quick look at the > > > >> FreeBSD manpage suggests its madvise MADV_DONTNEED is identical > > > >> to its posix_madvise MADV_DONTNEED. > > > >> > > > >> If we need "specifically Linux MADV_DONTNEED semantics" maybe we > > > >> should define a QEMU_MADV_LINUX_DONTNEED which either (a) does the > > > >> right thing or (b) fails, and use qemu_madvise() regardless. > > > >> > > > >> Certainly the current code is pretty fragile to being changed by > > > >> people who don't understand the undocumented subtlety behind > > > >> the use of a direct madvise() call here. > > > > > > > > Yeh and I'm not sure I can remembe rall the subtleties; there's a big > > > > hairy set of ifdef's in include/qemu/madvise.h that makes > > > > sure we always have the definition of QEMU_MADV_REMOVE/DONTNEED > > > > even on platforms that might not define it themselves. > > > > > > > > But I think this code is used for things with different degrees > > > > of care about the semantics; e.g. 'balloon' just cares that > > > > it frees memory up and doesn't care about the detailed semantics > > > > that much; so it's probably fine with that. > > > > Postcopy is much more touchy, but then it's only going to be > > > > calling this on Linux anyway (because of the userfault dependency). > > > > > > MADV_DONTNEED/MADV_REMOVE only provides discard semantics on Linux IIRC > > > -- and that's what we want to achieve: ram_block_discard_range() > > > > > > So I agree with Peter that we might want to make this more explicit. > > > > I was looking at the comments/history around this code to try to make > > this more explicit/clear, and it seems like the whole function is very > > Linux-specific. All we ever do is: > > > > - fallocate(FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE) > > - madvise(MADV_REMOVE) > > - madvise(MADV_DONTNEED) with Linux semantics > > > > All of those operations are Linux-only, so trying to figure out the > > cross-platform way to model this seems kind of pointless. Is it fine to > > just #ifdef the whole thing to be just for Linux? > > For ballooning we don't really need Linux semantics; we just need it to > use less host memory. Postcopy needs the more careful semantics though. Right, sorry, you mentioned that, but I was thinking it sounded like that applied to the MADV_DONTNEED path, which is only 1 case of the 3. But reading into this a bit, maybe all of these cases are fine on non-Linux for ballooning: fallocate is never called ('rb->fd' is always -1), and QEMU_MADV_REMOVE falls back to QEMU_MADV_DONTNEED, which is fine for ballooning. Am I understanding that correctly? But regardless, it's simpler/more-conservative to make it all Linux-specific. If I just take ram_block_discard_range away from non-Linux (that is, make it always return an error), is that breaking actual functionality, or is it removing a niche code path that non-Linux isn't really using anyway, and it's not worth the time to figure out? I am not familiar with qemu internals, so this is not obvious to me. For context: I'm just trying to get the tree to compile on other platforms (immediate focus is the guest agent on Solaris). The madvise() calls here generate warnings due to platform-specific quirks that qemu_madvise() has some logic to deal with. So my question is whether to adapt these to the cross-platform qemu_advise(), or treat the function as platform-specific code. -- Andrew Deason adea...@sinenomine.net