Re: [PATCH 7/7] xen/evtchn: read producer index only once

["Jürgen Groß" <jgross@suse.com>]

[-- Attachment #1.1.1: Type: text/plain, Size: 2808 bytes --]

On 08.02.21 11:51, Jan Beulich wrote:
> On 08.02.2021 11:41, Jürgen Groß wrote:
>> On 08.02.21 10:48, Jan Beulich wrote:
>>> On 06.02.2021 11:49, Juergen Gross wrote:
>>>> In evtchn_read() use READ_ONCE() for reading the producer index in
>>>> order to avoid the compiler generating multiple accesses.
>>>>
>>>> Signed-off-by: Juergen Gross <jgross@suse.com>
>>>> ---
>>>>    drivers/xen/evtchn.c | 2 +-
>>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/xen/evtchn.c b/drivers/xen/evtchn.c
>>>> index 421382c73d88..f6b199b597bf 100644
>>>> --- a/drivers/xen/evtchn.c
>>>> +++ b/drivers/xen/evtchn.c
>>>> @@ -211,7 +211,7 @@ static ssize_t evtchn_read(struct file *file, char __user *buf,
>>>>    			goto unlock_out;
>>>>    
>>>>    		c = u->ring_cons;
>>>> -		p = u->ring_prod;
>>>> +		p = READ_ONCE(u->ring_prod);
>>>>    		if (c != p)
>>>>    			break;
>>>
>>> Why only here and not also in
>>>
>>> 		rc = wait_event_interruptible(u->evtchn_wait,
>>> 					      u->ring_cons != u->ring_prod);
>>>
>>> or in evtchn_poll()? I understand it's not needed when
>>> ring_prod_lock is held, but that's not the case in the two
>>> afaics named places. Plus isn't the same then true for
>>> ring_cons and ring_cons_mutex, i.e. aren't the two named
>>> places plus evtchn_interrupt() also in need of READ_ONCE()
>>> for ring_cons?
>>
>> The problem solved here is the further processing using "p" multiple
>> times. p must not be silently replaced with u->ring_prod by the
>> compiler, so I probably should reword the commit message to say:
>>
>> ... in order to not allow the compiler to refetch p.
> 
> I still wouldn't understand the change (and the lack of
> further changes) then: The first further use of p is
> outside the loop, alongside one of c. IOW why would c
> then not need treating the same as p?

Its value wouldn't change, as ring_cons is being modified only at
the bottom of this function, and nowhere else (apart from the reset
case, but this can't run concurrently due to ring_cons_mutex).

> I also still don't see the difference between latching a
> value into a local variable vs a "freestanding" access -
> neither are guaranteed to result in exactly one memory
> access afaict.

READ_ONCE() is using a pointer to volatile, so any refetching by
the compiler would be a bug.

> And of course there's also our beloved topic of access
> tearing here: READ_ONCE() also excludes that (at least as
> per its intentions aiui).

Yes, but I don't see an urgent need to fix that, as there would
be thousands of accesses in the kernel needing a fix. A compiler
tearing a naturally aligned access into multiple memory accesses
would be rejected as buggy from the kernel community IMO.


Juergen

[-- Attachment #1.1.2: OpenPGP_0xB0DE9DD628BF132F.asc --]
[-- Type: application/pgp-keys, Size: 3135 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]