Re: [PATCH v2 0/2] PCI/DOE: Speed up mailbox polling

[Bjorn Helgaas <helgaas@kernel.org>]

[+cc DOE and MSI/MSI-X folks]

On Tue, Nov 03, 2026 at 04:30:59PM +0000, Michal Clapinski wrote:
> The DOE code waits for mailbox responses by sleeping
> PCI_DOE_POLL_INTERVAL (7 ms at HZ=1000, 4 ms at HZ=250) between status
> reads. Every DOE-capable endpoint goes through several such exchanges
> during enumeration: an abort when the mailbox is created and one
> discovery exchange per supported feature. Each pays at least one
> interval, even though devices typically respond within about a hundred
> microseconds.
> 
> Patch 1 fixes PCI_DOE_POLL_INTERVAL being 0 when HZ < 128, which makes
> the polling loops spin for up to a second instead of sleeping.
> 
> Patch 2 replaces the fixed sleep with an exponential backoff that
> busy-waits briefly while high resolution timers are not yet enabled, as
> is the case during enumeration at boot. Time spent in pci_device_add()
> for an NVIDIA B200 GPU drops from 45.9 ms to 6.4 ms.
> 
> Changes in v2:
> - New patch 1, fixing busy-polling when HZ < 128.
> - Patch 2: share the fast poll with pci_doe_abort() by moving it into
>   pci_doe_wait().
> - Patch 2: replace the fixed ~2 ms fast-polling window with an
>   exponential backoff that busy-waits for its first ~255 us, but only
>   until high resolution timers are enabled.
> 
> Michal Clapinski (2):
>   PCI/DOE: Don't busy-poll the mailbox when HZ < 128
>   PCI/DOE: Poll briefly for a response before sleeping
> 
>  drivers/pci/doe.c | 49 ++++++++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 44 insertions(+), 5 deletions(-)

v1 at https://lore.kernel.org/linux-pci/20261012114213.1757207-1-mclapinski@google.com/

I wish we had a way to use interrupts instead of polling, but it
doesn't seem like we have a way to share MSI/MSI-X vectors between
PCI core code like this and the eventual driver.  Added Thomas in case
he has any brilliant ideas about that.