From: "Michael S. Tsirkin" <mst@redhat.com>
To: Peter Maydell <peter.maydell@linaro.org>
Cc: "Gavin Shan" <gshan@redhat.com>, "Peter Xu" <peterx@redhat.com>,
"Pavel Hrdina" <phrdina@redhat.com>,
"Daniel P. Berrangé" <berrange@redhat.com>,
qemu-devel@nongnu.org, qemu-arm@nongnu.org, jugraham@redhat.com,
shan.gavin@gmail.com, "Alex Williamson" <alex@shazbot.org>,
"David Hildenbrand" <david@kernel.org>
Subject: Re: [PATCH RFCv1] virtio: Inherit max bounce buffer size from bus parent if possible
Date: Thu, 11 Jun 2026 11:25:01 -0400 [thread overview]
Message-ID: <20260611112248-mutt-send-email-mst@kernel.org> (raw)
In-Reply-To: <20260611110156-mutt-send-email-mst@kernel.org>
On Thu, Jun 11, 2026 at 11:05:35AM -0400, Michael S. Tsirkin wrote:
> On Thu, Jun 11, 2026 at 03:55:12PM +0100, Peter Maydell wrote:
> > On Thu, 11 Jun 2026 at 15:10, Michael S. Tsirkin <mst@redhat.com> wrote:
> > >
> > > On Thu, Jun 11, 2026 at 01:48:51PM +0100, Peter Maydell wrote:
> > > > On Thu, 11 Jun 2026 at 13:34, Gavin Shan <gshan@redhat.com> wrote:
> > > > >
> > > > > Let me try to summarize what I understood. As VFIO is concerned, there
> > > > > are multiple memory regions for one particular PCI BAR, and they're stacked
> > > > > up. The memory regions for PCI BAR#4 of the GH100 card looks as below.
> > > > >
> > > > > (qemu) info mtree
> > > > > :
> > > > > address-space: pci_bridge_pci_mem
> > > > > 0000000000000000-ffffffffffffffff (prio 0, container): pci_bridge_pci
> > > > > 0000042000000000-0000043fffffffff (prio 1, i/o): 0009:01:00.0 base BAR 4 <---- (1) VFIOBAR::mr
> > > > > 0000042000000000-0000043fffffffff (prio 0, i/o): 0009:01:00.0 BAR 4 <---- (2) VFIOBAR::VFIORegion::mem
> > > > > 0000042000000000-000004379fffffff (prio 0, ramd): 0009:01:00.0 BAR 4 mmaps[0] <---- (3) VFIOBAR::VFIORegion::VFIOMap::mem
> > > > >
> > > > > (1) Its MemoryRegionOps is NULL. No data accesses are routed to this region
> > > > > (2) The data accesses routed to this region is handled by pread() and pwrite()
> > > > > (3) The data accesses routed to this region is handled by memcpy() before
> > > > > commit 4a2e242bbb.
> > > > >
> > > > > There are identified PCI devices who have quirks, see vfio_bar_quirk_setup().
> > > > > Accesses to part of the PCI BAR have to be emulated by the extra IO regions,
> > > > > something like below for rtl8168 PCI device, where two extra IO regions are
> > > > > stacked up for the quirks.
> > > > >
> > > > > address-space: pci_bridge_pci_mem
> > > > > 0000000000000000-ffffffffffffffff (prio 0, container): pci_bridge_pci
> > > > > 0000042000000000-0000043fffffffff (prio 1, i/o): 0009:01:00.0 base BAR 4 <---- (1) VFIOBAR::mr
> > > > > 0000042000000000-0000043fffffffff (prio 0, i/o): 0009:01:00.0 BAR 4 <---- (2) VFIOBAR::VFIORegion::mem
> > > > > 0000042000000000-000004379fffffff (prio 0, ramd): 0009:01:00.0 BAR 4 mmaps[0] <---- (3) VFIOBAR::VFIORegion::VFIOMap::mem
> > > > > 0000042000000010-0000042000000014 (prio 1, i/o): 0009:01:00.0 BAR 4 quirk[0] <---- (4) quirk[0]
> > > > > 0000042000000018-000004200000001c (prio 1, i/o): 0009:01:00.0 BAR 4 quirk[1] <---- (5) quirk[1]
> > > > >
> > > > > Access on 0000042000000010-0000042000000014 should be routed to region (4) quirk[0]
> > > > > and access on 0000042000000018-000004200000001c should be routed to region (5) quirk[1].
> > > > > However, accesses to 0000042000000000-0000042000000020 are routed to region (3) before
> > > > > commit 4a2e242bbb and the data transfer is done by memcpy(), bypassing region (4) and
> > > > > (5). It's not the expected behavior and why memcpy() isn't expected on device rtl8168's
> > > > > PCI BAR due to the quirks, answering your question.
> > > > >
> > > > > With commit 4a2e242bbb applied, the accesses will be routed to the correct region.
> > > >
> > > > The way I read 4a2e242bbb's commit message, it isn't about things being routed
> > > > to the wrong region. It's about the handling of areas which aren't in the small
> > > > quirk regions but which are in the same 4K page as them. These have to
> > > > be handled
> > > > via the memory subsystem's "subpage" mechanism. This does route
> > > > everything to the
> > > > correct region, but if the region (3) is marked as "direct access is OK" then
> > > > QEMU assumes that any kind of direct access is OK, i.e. this behaves
> > > > like true RAM.
> > > > It then does a memcpy access to a BAR that's really a bank of device registers,
> > > > and this goes wrong.
> > > >
> > > > > Back to our case (GH100 card), there are no quirks for the PCI BAR (0009:01:00.0 BAR 4)
> > > > > so it's fine mark the RAM DEVICE region as directly accessible. We perhaps needn't host
> > > > > to export the capability (VFIO_REGION_INFO_CAP_DIRECT_ACCESS) suggested by you. It's
> > > > > safe to mark any PCI BARs as directly accessible if they have no quirks attached. All
> > > > > the devices except those listed in vfio_bar_quirk_setup() are capable of this.
> > > >
> > > > I still feel like there are different kinds of PCI BAR here ("this BAR is
> > > > true RAM and can be accessed arbitrarily" vs "this BAR is full of registers
> > > > and can't be handled that way") and the vfio code in QEMU needs to set up
> > > > the memory regions differently for the two cases. For your example I think
> > > > it would be fine to have direct-access even if there were some kind of
> > > > quirk memory region, because for the parts of the BAR that aren't covered
> > > > by a quirk overlay the underlying BAR still allows "entirely like RAM,
> > > > any alignment and size is OK" accesses.
> >
> > > Yea, and I feel this is the main part:
> > >
> > > The assumption here is that accesses initiated by the VM are
> > > driven by a device specific driver, which knows the device
> > > capabilities.
> > >
> > >
> > > Frankly I don't get why a big hammer of disabling direct access
> > > was taken, when all we apparently need to do is to make sure
> > > small aligned accesses through BAR stay aligned and same size.
> >
> > If you say an MR is OK for direct access then in my view you are
> > saying "any access of any kind is OK", because you're permitting
> > the pointer to be directly returned as a host pointer from
> > address_space_map() (at which point the caller might do anything
> > with that memory). That is not OK for every BAR, only for ones where
> > it's really RAM.
> >
>
> What is "OK"? If the BAR is RAM and I write into it, I will overwrite
> data guest stored there. Is that "OK"?
>
>
>
>
> > > I guess it felt safe - a vfio specific change, and emulating device
> > > accesses was assumed to be slow path, anyway.
> > >
> > > Except it no longer is with people wanting to do direct io
> > > into device BARs.
> >
> > > Isn't it basically the below?
> > > At least I checked asm and it produces the correct code.
> > > And then the whole pile of hacks can be reverted?
> > >
> > >
> > > diff --git a/system/physmem.c b/system/physmem.c
> > > index 7bcbf87573..aab4390d40 100644
> > > --- a/system/physmem.c
> > > +++ b/system/physmem.c
> > > @@ -3272,7 +3272,29 @@ static MemTxResult flatview_write_continue_step(MemTxAttrs attrs,
> > > uint8_t *ram_ptr = qemu_ram_ptr_length(mr->ram_block, mr_addr, l,
> > > false, true);
> > >
> > > - memmove(ram_ptr, buf, *l);
> > > + switch (*l) {
> > > + case 1:
> > > + __builtin_memmove(ram_ptr, buf, 1);
> > > + break;
> > > + case 2:
> > > + __builtin_memmove(ram_ptr, buf, 2);
> > > + break;
> > > + case 4:
> > > + __builtin_memmove(ram_ptr, buf, 4);
> > > + break;
> > > + case 8:
> > > + __builtin_memmove(ram_ptr, buf, 8);
> >
> > Nothing says that __builtin_memmove() is required to do only a single
> > access of the right size.
> > memmove() can do all kinds of things.
> > For instance on aarch64 if you get into the actual glibc memmove then
> > a memmove of 1 byte will actually store the same byte multiple times.
> > Or an architecture might decide that it's more efficient to do an
> > 8-byte write with two 4-byte writes even if there is an 8-byte access
> > instruction.
> >
> > If and where we need to provide guarantees about "this access will
> > really definitely only do this size access and it won't break it
> > apart or anything like that" then we need to either be using compiler
> > atomics or else inline asm.
> >
> > (I think there are places where we do need to be more careful about
> > what we do with accesses to real RAM, for where we're emulating a
> > device write to RAM that's updating a data structure shared with the
> > guest, and things like writing multiple times can cause problems.
> > https://lore.kernel.org/qemu-devel/CAFEAcA8dwHV8F48kb-013rxkG9kKcZhym9_qarKmoeUfeh0YWw@mail.gmail.com/
> > is an unrelated example of that, which I haven't done detailed
> > analysis of yet.)
> >
> > thanks
> > -- PMM
>
> If it does not work, then QEMU is broken:
>
>
> /*
> * Any compiler worth its salt will turn these memcpy into native unaligned
> * operations. Thus we don't need to play games with packed attributes, or
> * inline byte-by-byte stores.
> * Some compilation environments (eg some fortify-source implementations)
> * may intercept memcpy() in a way that defeats the compiler optimization,
> * though, so we use __builtin_memcpy() to give ourselves the best chance
> * of good performance.
> */
>
> static inline int lduw_he_p(const void *ptr)
> {
> uint16_t r;
> __builtin_memcpy(&r, ptr, sizeof(r));
> return r;
> }
>
> static inline int ldsw_he_p(const void *ptr)
> {
> int16_t r;
> __builtin_memcpy(&r, ptr, sizeof(r));
> return r;
> }
>
> static inline void stw_he_p(void *ptr, uint16_t v)
> {
> __builtin_memcpy(ptr, &v, sizeof(v));
> }
>
>
In fact, it's not even complete, mmapped version does memmove,
which only happens to be same as memcpy because the compiler
is not insane.
In short, it's a pile of hacks, so if Alex can confirm his usecase
still works with above flatview tweak, I think we can revert the
whole thing.
>
> Somehow things have worked for years without atomics,
> but if you want to fix that, be my guest.
>
>
> --
> MST
next prev parent reply other threads:[~2026-06-11 15:25 UTC|newest]
Thread overview: 67+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-06-08 0:18 [PATCH RFCv1] virtio: Inherit max bounce buffer size from bus parent if possible Gavin Shan
2026-06-08 8:55 ` Daniel P. Berrangé
2026-06-08 11:11 ` Gavin Shan
2026-06-08 11:38 ` Daniel P. Berrangé
2026-06-09 2:08 ` Gavin Shan
2026-06-09 16:25 ` Peter Xu
2026-06-10 0:32 ` Gavin Shan
2026-06-10 9:54 ` Pavel Hrdina
2026-06-10 10:55 ` Gavin Shan
2026-06-10 12:12 ` Michael S. Tsirkin
2026-06-10 12:19 ` Gavin Shan
2026-06-10 12:27 ` Michael S. Tsirkin
2026-06-10 13:00 ` Gavin Shan
2026-06-10 13:54 ` Gavin Shan
2026-06-10 14:06 ` Michael S. Tsirkin
2026-06-10 15:36 ` Peter Xu
2026-06-10 16:11 ` Peter Maydell
2026-06-10 16:19 ` Michael S. Tsirkin
2026-06-10 19:10 ` Peter Xu
2026-06-10 21:03 ` Michael S. Tsirkin
2026-06-10 21:27 ` Peter Xu
2026-06-10 21:44 ` Michael S. Tsirkin
2026-06-10 16:18 ` Michael S. Tsirkin
2026-06-11 4:33 ` Gavin Shan
2026-06-11 5:31 ` Michael S. Tsirkin
2026-06-11 6:28 ` Gavin Shan
2026-06-11 6:34 ` Michael S. Tsirkin
2026-06-11 12:33 ` Gavin Shan
2026-06-11 12:48 ` Peter Maydell
2026-06-11 14:10 ` Michael S. Tsirkin
2026-06-11 14:55 ` Peter Maydell
2026-06-11 15:05 ` Michael S. Tsirkin
2026-06-11 15:25 ` Michael S. Tsirkin [this message]
2026-06-11 15:29 ` Peter Maydell
2026-06-11 15:57 ` Michael S. Tsirkin
2026-06-11 16:16 ` Peter Maydell
2026-06-11 16:42 ` Michael S. Tsirkin
2026-06-11 16:53 ` Peter Maydell
2026-06-11 17:02 ` Michael S. Tsirkin
2026-06-11 18:20 ` Peter Xu
2026-06-11 20:52 ` Michael S. Tsirkin
2026-06-11 21:20 ` Peter Xu
2026-06-11 21:59 ` Michael S. Tsirkin
2026-06-12 19:44 ` Alex Williamson
2026-06-12 21:24 ` Peter Xu
2026-06-14 16:10 ` Michael S. Tsirkin
2026-06-11 21:15 ` Michael S. Tsirkin
2026-06-11 16:13 ` Michael S. Tsirkin
2026-06-12 4:25 ` Gavin Shan
2026-06-12 8:43 ` Michael S. Tsirkin
2026-06-12 10:25 ` Gavin Shan
2026-06-11 6:51 ` Michael S. Tsirkin
2026-06-10 12:23 ` Pavel Hrdina
2026-06-10 14:04 ` Gavin Shan
2026-06-10 14:08 ` Michael S. Tsirkin
2026-06-10 9:49 ` Michael S. Tsirkin
2026-06-10 18:30 ` Stefan Hajnoczi
2026-06-10 21:00 ` Michael S. Tsirkin
2026-06-11 14:20 ` Stefan Hajnoczi
2026-06-11 14:45 ` Michael S. Tsirkin
2026-06-11 15:04 ` Peter Maydell
2026-06-11 15:09 ` Michael S. Tsirkin
2026-06-11 18:37 ` Stefan Hajnoczi
2026-06-11 20:54 ` Michael S. Tsirkin
2026-06-15 15:54 ` Stefan Hajnoczi
2026-06-15 19:59 ` Michael S. Tsirkin
2026-06-11 1:19 ` Gavin Shan
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20260611112248-mutt-send-email-mst@kernel.org \
--to=mst@redhat.com \
--cc=alex@shazbot.org \
--cc=berrange@redhat.com \
--cc=david@kernel.org \
--cc=gshan@redhat.com \
--cc=jugraham@redhat.com \
--cc=peter.maydell@linaro.org \
--cc=peterx@redhat.com \
--cc=phrdina@redhat.com \
--cc=qemu-arm@nongnu.org \
--cc=qemu-devel@nongnu.org \
--cc=shan.gavin@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox