[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH v7 10/20] hw/arm/smmuv3: Implement translate cal

From: Peter Maydell
Subject: Re: [Qemu-devel] [PATCH v7 10/20] hw/arm/smmuv3: Implement translate callback
Date: Tue, 6 Feb 2018 12:43:06 +0000

On 6 February 2018 at 12:19, Auger Eric <address@hidden> wrote:
> On 09/10/17 19:45, Peter Maydell wrote:
>> Incidentally, the spec requires us to perform memory accesses as
>> at least 64-bit single-copy atomic (see 3.21.3) -- does this do that?
>> (This gets important with SMP when the guest on another CPU might
>> be updating the STE or page table entry at the same time as we're
>> reading it...)
> Among all your comments on v7, here is the one I am the least
> comfortable with. I was not able to figure out whether
> dma_memory_read(), which I use now for all the descriptor accesses
> guarantees this 64b single copy atomicity.

It doesn't -- it winds up in flatview_read_continue(), which will
do a memcpy() from guest ram into the buffer, and since it's
using an arbitrary passed-in length value the chances are high
it won't end up using an atomic access on the host.

This is a nasty issue which we haven't figured out at all yet;
see also this thread:

For the moment my suggestion would be that when you need
to do guest memory accesses that have atomicity requirements you:
 (a) use an accessor function which specifically loads a
  quantity of the correct size, rather than one which takes
  an arbitrary start-and-length
  (when we add APIs which do have atomicity guarantees they'll
  be "load 64 bit word" etc, so using our existing APIs of
  that form should avoid needing to restructure this code later)
 (b) add a TODO comment noting the required atomicity

-- PMM

reply via email to

[Prev in Thread] Current Thread [Next in Thread]