Re: PCI memory sync question (kvm,dpdk,e1000,packet stalled)

[Stefan Hajnoczi <stefanha@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2635 bytes --]

On Wed, Nov 20, 2019 at 08:36:32PM +0300, ASM wrote:
> I trying solve the problem, with packets stopping (e1000,tap,kvm).
> My studies led to the following:
> 1. From flatview_write_continue() I see, what e1000 writes the number
> "7" to the STAT register.
> 2. The driver from target OS reads STAT register with number "7" and
> writes to the register the number "0".
> 3. From flatview_write_continue() (I make edits):
>             memcpy(ptr, buf, l);
>             new1=ptr[0xc];
>             usleep(100);
>             new2=ptr[0xc];
>             invalidate_and_set_dirty(mr, addr1, l);
>             new3=ptr[0xc];
> printf("Old: %i, new1, %i, new2: %i, new3: %i\n", old,new1,new2,new3);
> 
> I see what memory in first printf is "7", but after usleep() is "0".
> Do I understand correctly that this should not be? Or RCU lock
> suggests the ability to the multiple writers?
> 
> The problem is that qemu(e1000) writes the number 7, after which
> target(dpdk driver) reads 7, on the basis of this it writes the number
> 0, but as a result (extremely rarely), the value STATUS still remains
> 7. Therefore, packet processing is interrupted. This behavior is
> observed only on kvm (it is not observed on tcg).
> 
> Please help with advice or ideas.

Hi Leonid,
Could you be seeing weird behavior with KVM due to MMIO write
coalescing?

  static void e1000_mmio_setup(E1000State *d)
  {
      int i;
      const uint32_t excluded_regs[] = {
          E1000_MDIC, E1000_ICR, E1000_ICS, E1000_IMS,
          E1000_IMC, E1000_TCTL, E1000_TDT, PNPMMIO_SIZE
      };

      memory_region_init_io(&d->mmio, OBJECT(d), &e1000_mmio_ops, d,
                            "e1000-mmio", PNPMMIO_SIZE);
      memory_region_add_coalescing(&d->mmio, 0, excluded_regs[0]);
      for (i = 0; excluded_regs[i] != PNPMMIO_SIZE; i++)
          memory_region_add_coalescing(&d->mmio, excluded_regs[i] + 4,
                                       excluded_regs[i+1] - excluded_regs[i] - 4);
      memory_region_init_io(&d->io, OBJECT(d), &e1000_io_ops, d, "e1000-io", IOPORT_SIZE);
  }

MMIO write coalescing means that QEMU doesn't see the register writes
immediately.  Instead kvm.ko records them into a ring buffer and QEMU
processes the ring when the next ioctl(KVM_RUN) exit occurs.

See Linux Documentation/virt/kvm/api.txt "4.116
KVM_(UN)REGISTER_COALESCED_MMIO" for more details.

I don't really understand your printf debugging explanation.  It would
help to see the DPDK code and the exact printf() output.

Also, is DPDK accessing the e1000 device from more than 1 vCPU?

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]