Re: [PATCH] x86/msi: dynamically map pages for MSI-X tables

["Roger Pau Monné" <roger.pau@citrix.com>]

On Tue, May 02, 2023 at 12:18:06PM +0200, Jan Beulich wrote:
> On 02.05.2023 12:10, Roger Pau Monné wrote:
> > On Wed, Apr 26, 2023 at 02:55:20PM +0000, Ruben Hakobyan wrote:
> >> Xen reserves a constant number of pages that can be used for mapping
> >> MSI-X tables. This limit is defined by FIX_MSIX_MAX_PAGES in fixmap.h.
> >>
> >> Reserving a fixed number of pages could result in an -ENOMEM if a
> >> device requests a new page when the fixmap limit is exhausted and will
> >> necessitate manually adjusting the limit before compilation.
> >>
> >> To avoid the issues with the current fixmap implementation, we modify
> >> the MSI-X page mapping logic to instead dynamically map new pages when
> >> they are needed by making use of ioremap().
> > 
> > I wonder if Arm plans to reuse this code, and whether then arm32 would
> > better keep the fixmap implementation to avoid exhausting virtual
> > address space in that case.
> 
> I think this would then need to be something that 32-bit architectures
> do specially. Right now aiui PCI (and hence MSI-X) work on Arm targets
> only Arm64.
> 
> > This also have the side effect of ioremap() now possibly allocating a
> > page in order to fill the page table for the newly allocated VA.
> 
> Indeed, but I think the (vague) plan to switch to ioremap() has been
> around for a pretty long time (perhaps forever since 32-bit support
> was purged).

Yup, I'm not saying the above should block the patch, but might be
worth mentioning in the commit message.

Roger.