On Thu Nov 12, 2026 at 2:19 PM JST, Alexandre Courbot wrote: > On Thu Nov 12, 2026 at 10:43 AM JST, Eliot Courtney wrote: >> On Wed Nov 11, 2026 at 11:49 PM JST, Alexandre Courbot wrote: >>> On Wed Nov 11, 2026 at 10:37 AM JST, Eliot Courtney wrote: >>>> In the msgq v2 protocol, which the r000 firmware uses, the queue >>>> pointers are kept in u32 registers and increase monotonically until they >>>> wrap at 2^32. >>>> >>>> `MSGQ_NUM_PAGES` is currently 63 and does not divide 2^32. This means >>>> that on msgq v2, the slot can jump backwards when the counters wrap. >>>> For example: >>>> >>>> (2^32 - 2) % 63 == 2 >>>> (2^32 - 1) % 63 == 3 >>>> 0 % 63 == 0 (slot 0 instead of slot 4) >>>> >>>> The writer can then overwrite slots before the reader has read them. >>>> >>>> Increase `MSGQ_NUM_PAGES` to 64 and check at compile time that >>>> `MSGQ_NUM_PAGES` divides 2^32. >>> >>> Have we tested that this was also working on msgq v0? >> >> Yerp I tested boot on turing, ampere, ada, blackwell. >> >>> >>>> >>>> Signed-off-by: Eliot Courtney <[email protected]> >>>> --- >>>> drivers/gpu/nova-core/gsp/cmdq.rs | 10 +++++++++- >>>> 1 file changed, 9 insertions(+), 1 deletion(-) >>>> >>>> diff --git a/drivers/gpu/nova-core/gsp/cmdq.rs >>>> b/drivers/gpu/nova-core/gsp/cmdq.rs >>>> index 33074bb53e8c..240ab42db557 100644 >>>> --- a/drivers/gpu/nova-core/gsp/cmdq.rs >>>> +++ b/drivers/gpu/nova-core/gsp/cmdq.rs >>>> @@ -84,7 +84,15 @@ trait CommandElement { >>>> } >>>> >>>> /// Number of GSP pages making the [`Msgq`]. >>>> -pub(crate) const MSGQ_NUM_PAGES: u32 = 0x3f; >>>> +pub(crate) const MSGQ_NUM_PAGES: u32 = 64; >>>> + >>>> +// From version 2 of the msgq protocol, the queue pointers increase >>>> +// monotonically in `u32` counters until they wrap. If `MSGQ_NUM_PAGES` >>>> does not >>>> +// divide `1 << 32`, then the slot, `counter % MSGQ_NUM_PAGES`, can jump >>>> +// backwards when wrapping, which could overwrite slots that still need >>>> to be >>>> +// read. So, require that the last counter value, `u32::MAX`, maps to the >>>> last >>>> +// slot. >>> >>> That's a pretty convoluted way to say the number of slots must be a >>> power of 2 representable as a u32. :) >>> >>>> +static_assert!(u32::MAX % MSGQ_NUM_PAGES == MSGQ_NUM_PAGES - 1); >>> >>> Same here, can't we use `MSGQ_NUM_PAGES.is_power_of_two()`? >> >> I actually had this originally but I think that this form expresses the >> condition more directly. The condition/issue is what happens when you >> wrap the counter - it should go to the next slot (which necessarily >> needs to be zero since a wrapped counter is 0). So you want >> MSGQ_NUM_PAGES to divide 2^32. So if you write >> MSGQ_NUM_PAGES.is_power_of_two() you are implicitly relying on all >> divisors being a power of two, which is an extra logical step. >> >> But in the end I don't really mind, so lmk wdyt. > > The current condition also involves an extra logical step. What you > ideally would have tested is > > (1 << 32).is_multiple_of(MSGQ_NUM_PAGES) > > ... but as `u32::MAX + 1` is not representable using a u32, the one you > chose is the closest one. It does takes some extra labor to understand > what it intends to convey though. > > Trying to put myself into the shoes of someone looking at this code for > the first time, this comment and the condition look a bit hard to chew. > How about using a u64 to test the condition we actually want: > > // The queue pointers are wrapping `u32` counters, and the slot is > // `counter % MSGQ_NUM_PAGES`. For slots to remain contiguous when a counter > // wraps, `MSGQ_NUM_PAGES` must divide `1 << 32`. > static_assert!((1u64 << u32::BITS).is_multiple_of(cv!(MSGQ_NUM_PAGES))); > > (I have removed the mention the the msgq v2 detail - it is useful in the > commit log, but becomes unneeded information in the code as it > eventually will only support v2 anyway) > > WDYT?
Yeah I originally also tried this but I didn't find as nice a form (w.r.t. the conversions to u64) as you did, so I will go with your version. Thanks!
