[PATCH v3 00/10] IOMMU memory observability

[Pasha Tatashin]

On Thu, Jan 4, 2024 at 12:04 PM Michal Koutný <mkoutny at suse.com> wrote:
>
> On Thu, Jan 04, 2024 at 11:29:43AM -0500, Pasha Tatashin <pasha.tatashin at soleen.com> wrote:
> > Thank you for taking a look at this. The two patches [1] [2] which add
> > GFP_KERNEL_ACCOUNT were sent separate from this series at request of
> > reviewers:
>
> Ah, I didn't catch that.
>
> Though, I mean the patch 02/10 calls iommu_alloc_pages() with GFP_KERNEL
> (and not a passed gfp from iommu_map).
> Then patch 09/10 accounts all iommu_alloc_pages() under NR_IOMMU_PAGES.
>
> I think there is a difference between what's shown NR_IOMMU_PAGES and
> what will have __GFP_ACCOUNT because of that.
>
> I.e. is it the intention that this difference is not subject to
> limiting?

Yes, we will have a difference between GFP_ACCOUNT and what
NR_IOMMU_PAGES shows. GFP_ACCOUNT is set only where it makes sense to
charge to user processes, i.e. IOMMU Page Tables, but there more IOMMU
shared data that should not really be charged to a specific process.
The charged and uncharged data will be visible via /proc/vmstat
nr_iommu_pages field.

Pasha

>
> (Note: I'm not familiar with iommu code and moreover I'm only looking at
> the two patch sets, not the complete code applied. So you may correct my
> reasoning.)
>
>
> Thanks,
> Michal