Re: [PATCH v2] memory: prevent dma-reentracy issues

[Stefan Hajnoczi]

On Thu, Oct 20, 2022 at 06:11:06PM -0400, Alexander Bulekov wrote:
> On 220712 1034, Stefan Hajnoczi wrote:
> > On Tue, Jun 21, 2022 at 11:53:06AM -0400, Alexander Bulekov wrote:
> > > On 220621 1630, Peter Maydell wrote:
> > > > On Thu, 9 Jun 2022 at 14:59, Alexander Bulekov <alxndr@bu.edu> wrote:
> > > > > diff --git a/include/hw/pci/pci.h b/include/hw/pci/pci.h
> > > > > index 44dacfa224..ab1ad0f7a8 100644
> > > > > --- a/include/hw/pci/pci.h
> > > > > +++ b/include/hw/pci/pci.h
> > > > > @@ -834,8 +834,17 @@ static inline MemTxResult pci_dma_rw(PCIDevice 
> > > > > *dev, dma_addr_t addr,
> > > > >                                       void *buf, dma_addr_t len,
> > > > >                                       DMADirection dir, MemTxAttrs 
> > > > > attrs)
> > > > >  {
> > > > > -    return dma_memory_rw(pci_get_address_space(dev), addr, buf, len,
> > > > > -                         dir, attrs);
> > > > > +    bool prior_engaged_state;
> > > > > +    MemTxResult result;
> > > > > +
> > > > > +    prior_engaged_state = dev->qdev.engaged_in_io;
> > > > > +
> > > > > +    dev->qdev.engaged_in_io = true;
> > > > > +    result = dma_memory_rw(pci_get_address_space(dev), addr, buf, 
> > > > > len,
> > > > > +                           dir, attrs);
> > > > > +    dev->qdev.engaged_in_io = prior_engaged_state;
> > > > > +
> > > > > +    return result;
> > > > 
> > > > Why do we need to do something in this pci-specific function ?
> > > > I was expecting this to only need changes at the generic-to-all-devices
> > > > level.
> > > 
> > > Both of these handle the BH->DMA->MMIO case. Unlike MMIO, I don't think
> > > there is any neat way to set the engaged_in_io flag as we enter a BH. So
> > > instead, we try to set it when a device initiates DMA.
> > > 
> > > The pci function lets us do that since we get a glimpse of the dev/qdev
> > > (unlike the dma_memory_...  functions).
> > ...
> > > > > @@ -302,6 +310,10 @@ static MemTxResult dma_buf_rw(void *buf, 
> > > > > dma_addr_t len, dma_addr_t *residual,
> > > > >          xresidual -= xfer;
> > > > >      }
> > > > >
> > > > > +    if (dev) {
> > > > > +        dev->engaged_in_io = prior_engaged_state;
> > > > > +    }
> > > > 
> > > > Not all DMA goes through dma_buf_rw() -- why does it need changes?
> > > 
> > > This one has the same goal, but accesses the qdev through sg, instead of
> > > PCI.
> > 
> > Should dma_*() APIs take a reentrancy guard argument so that all DMA
> > accesses are systematically covered?
> > 
> >   /* Define this in the memory API */
> >   typedef struct {
> >       bool engaged_in_io;
> >   } MemReentrancyGuard;
> > 
> >   /* Embed MemReentrancyGuard in DeviceState */
> >   ...
> > 
> >   /* Require it in dma_*() APIs */
> >   static inline MemTxResult dma_memory_rw(AddressSpace *as, dma_addr_t addr,
> >                                           void *buf, dma_addr_t len,
> >                                           DMADirection dir, MemTxAttrs 
> > attrs,
> >                                       MemReentrancyGuard *guard);
> > 
> >   /* Call dma_*() APIs like this... */
> >   static inline MemTxResult pci_dma_rw(PCIDevice *dev, dma_addr_t addr,
> >                                        void *buf, dma_addr_t len,
> >                                        DMADirection dir, MemTxAttrs attrs)
> >   {
> >       return dma_memory_rw(pci_get_address_space(dev), addr, buf, len,
> >                            dir, attrs, &dev->qdev.reentrancy_guard);
> >   }
> > 
> 
> Taking a stab at this. Here is the list of DMA APIs that appear to need
> changes:
> dma_memory_valid (1 usage)
> dma_memory_rw (~5 uses)
> dma_memory_read (~92 uses)
> dma_memory_write (~71 uses)
> dma_memory_set (~4 uses)
> dma_memory_map (~18 uses)
> dma_memory_unmap (~21 uses)
> {ld,st}_{le,be}_{uw,l,q}_dma (~10 uses)
> ldub_dma (does not appear to be used anywhere)
> stb_dma (1 usage)
> dma_buf_read (~18 uses)
> dma_buf_write (~7 uses)
> 
> These appear to be internal to the DMA API and probably don't need to be
> changed:
> dma_memory_read_relaxed (does not appear to be used anywhere)
> dma_memory_write_relaxed (does not appear to be used anywhere)
> dma_memory_rw_relaxed 
> 
> I don't think the sglist APIs need to be changed since we can get
> DeviceState from the QEMUSGList.
> 
> Does this look more-or-less right?

That's along the lines of what I would expect. Interesting that
map/unmap is also on the list; it makes sense when considering bounce
buffers.

Stefan

signature.asc
Description: PGP signature