Re: [PATCH v2 2/2] PCI/DOE: Poll briefly for a response before sleeping

[sashiko-bot@kernel.org]

> doe_statemachine_work() and pci_doe_abort() wait for the mailbox by
> sleeping PCI_DOE_POLL_INTERVAL (PCI_DOE_TIMEOUT / 128) between status
> reads, which is 7 ms at HZ=1000 and 4 ms at HZ=250. Mailbox operations
> are served from the device's local mailbox logic and complete much
> faster: on NVIDIA B200 GPUs, an abort completes in ~70 us and a
> discovery exchange in ~110 us, including the status reads polling for
> them. Every DOE-capable endpoint pays the sleep once for the abort
> issued during mailbox creation and once per supported feature during
> enumeration.
> 
> Make pci_doe_wait(), which both paths use, back off exponentially:
> start with a 1 us delay and double it on every poll until it reaches
> PCI_DOE_POLL_INTERVAL, then fall back to the existing jiffies-based
> wait.
> 
> Enumeration at boot runs before high resolution timers are enabled, so
> any sleep lasts until the next tick, which at HZ=250 is no shorter than
> PCI_DOE_POLL_INTERVAL. Until they are enabled, busy-wait instead of
> sleeping for delays of up to 128 us. Operations that complete within
> ~255 us of accumulated delay, like the ones above, then finish without
> sleeping, while slower ones waste at most that much CPU time before
> sleeping anyway. Once high resolution timers are enabled, short sleeps
> are accurate and no busy-wait beyond what fsleep() does is needed.
> 
> Time spent in pci_device_add() for a B200 GPU drops from 45.9 ms to
> 6.4 ms.
> 
> Signed-off-by: Michal Clapinski <mclapinski@google.com>

Sashiko has reviewed this patch and found no issues.

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261103163101.1696520-1-mclapinski@google.com?part=2