Re: [PATCH] softmmu: Support concurrent bounce buffers

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Sep 10, 2024 at 11:36:08PM +0200, Mattias Nissler wrote:
> On Tue, Sep 10, 2024 at 6:40 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Tue, Sep 10, 2024 at 06:10:50PM +0200, Mattias Nissler wrote:
> > > On Tue, Sep 10, 2024 at 5:44 PM Peter Maydell <peter.maydell@linaro.org> wrote:
> > > >
> > > > On Tue, 10 Sept 2024 at 15:53, Michael S. Tsirkin <mst@redhat.com> wrote:
> > > > >
> > > > > On Mon, Aug 19, 2024 at 06:54:54AM -0700, Mattias Nissler wrote:
> > > > > > When DMA memory can't be directly accessed, as is the case when
> > > > > > running the device model in a separate process without shareable DMA
> > > > > > file descriptors, bounce buffering is used.
> > > > > >
> > > > > > It is not uncommon for device models to request mapping of several DMA
> > > > > > regions at the same time. Examples include:
> > > > > >  * net devices, e.g. when transmitting a packet that is split across
> > > > > >    several TX descriptors (observed with igb)
> > > > > >  * USB host controllers, when handling a packet with multiple data TRBs
> > > > > >    (observed with xhci)
> > > > > >
> > > > > > Previously, qemu only provided a single bounce buffer per AddressSpace
> > > > > > and would fail DMA map requests while the buffer was already in use. In
> > > > > > turn, this would cause DMA failures that ultimately manifest as hardware
> > > > > > errors from the guest perspective.
> > > > > >
> > > > > > This change allocates DMA bounce buffers dynamically instead of
> > > > > > supporting only a single buffer. Thus, multiple DMA mappings work
> > > > > > correctly also when RAM can't be mmap()-ed.
> > > > > >
> > > > > > The total bounce buffer allocation size is limited individually for each
> > > > > > AddressSpace. The default limit is 4096 bytes, matching the previous
> > > > > > maximum buffer size. A new x-max-bounce-buffer-size parameter is
> > > > > > provided to configure the limit for PCI devices.
> > > > > >
> > > > > > Signed-off-by: Mattias Nissler <mnissler@rivosinc.com>
> > > > > > Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
> > > > > > Acked-by: Peter Xu <peterx@redhat.com>
> > > > > > ---
> > > > > > This patch is split out from my "Support message-based DMA in vfio-user server"
> > > > > > series. With the series having been partially applied, I'm splitting this one
> > > > > > out as the only remaining patch to system emulation code in the hope to
> > > > > > simplify getting it landed. The code has previously been reviewed by Stefan
> > > > > > Hajnoczi and Peter Xu. This latest version includes changes to switch the
> > > > > > bounce buffer size bookkeeping to `size_t` as requested and LGTM'd by Phil in
> > > > > > v9.
> > > > > > ---
> > > > > >  hw/pci/pci.c                |  8 ++++
> > > > > >  include/exec/memory.h       | 14 +++----
> > > > > >  include/hw/pci/pci_device.h |  3 ++
> > > > > >  system/memory.c             |  5 ++-
> > > > > >  system/physmem.c            | 82 ++++++++++++++++++++++++++-----------
> > > > > >  5 files changed, 76 insertions(+), 36 deletions(-)
> > > > > >
> > > > > > diff --git a/hw/pci/pci.c b/hw/pci/pci.c
> > > > > > index fab86d0567..d2caf3ee8b 100644
> > > > > > --- a/hw/pci/pci.c
> > > > > > +++ b/hw/pci/pci.c
> > > > > > @@ -85,6 +85,8 @@ static Property pci_props[] = {
> > > > > >                      QEMU_PCIE_ERR_UNC_MASK_BITNR, true),
> > > > > >      DEFINE_PROP_BIT("x-pcie-ari-nextfn-1", PCIDevice, cap_present,
> > > > > >                      QEMU_PCIE_ARI_NEXTFN_1_BITNR, false),
> > > > > > +    DEFINE_PROP_SIZE32("x-max-bounce-buffer-size", PCIDevice,
> > > > > > +                     max_bounce_buffer_size, DEFAULT_MAX_BOUNCE_BUFFER_SIZE),
> > > > > >      DEFINE_PROP_END_OF_LIST()
> > > > > >  };
> > > > > >
> > > > >
> > > > > I'm a bit puzzled by now there being two fields named
> > > > > max_bounce_buffer_size, one directly controllable by
> > > > > a property.
> > >
> > > One is one the pci device, the other is on the address space. The
> > > former can be set via a command line parameter, and that value is used
> > > to initialize the field on the address space, which is then consulted
> > > when allocating bounce buffers.
> > >
> > > I'm not sure which aspect of this is unclear and/or deserves
> > > additional commenting - let me know and I'll be happy to send a patch.
> >
> > I'd document what does each field do.
> 
> I have just sent a patch to expand the comments, let's discuss details there.
> 
> >
> > > > >
> > > > > Pls add code comments explaining how they are related.
> > > > >
> > > > >
> > > > > Also, what is the point of adding a property without
> > > > > making it part of an API? No one will be able to rely on
> > > > > it working.
> 
> All I needed was a practical way to allow the bounce buffer size limit
> to be adjusted in the somewhat exotic situations where we're making
> DMA requests to indirect memory regions (in my case it is a qemu
> vfio-user server accessed by a client that can't or doesn't want to
> provide direct memory-mapped access to its RAM). There was some
> discussion about the nature of the parameter when I first proposed the
> patch, see https://lore.kernel.org/qemu-devel/20230823092905.2259418-2-mnissler@rivosinc.com/
> - an x-prefixed experimental command-line parameter was suggested
> there as a practical way to allow this without qemu committing to
> supporting this forever. For the unlikely case that this parameter
> proves popular, it can still be added to a stable API (or
> alternatively we could discuss whether a large-enough limit is
> feasible after all, or even consider DMA API changes to obviate the
> need for bounce buffering).


Yes but how happy will you be if we rename the parameter in the
future? All your scripts will break.

> > > >
> > > > Note that this patch is already upstream as commit 637b0aa13.
> > > >
> > > > thanks
> > > > -- PMM
> >
> > Maybe you can answer this?
> >