[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20251007095129.4b0e88a8@fedora>
Date: Tue, 7 Oct 2025 09:51:29 +0200
From: Boris Brezillon <boris.brezillon@...labora.com>
To: Adrián Larumbe <adrian.larumbe@...labora.com>
Cc: linux-kernel@...r.kernel.org, dri-devel@...ts.freedesktop.org, Steven
Price <steven.price@....com>, kernel@...labora.com, Rob Herring
<robh@...nel.org>, Maarten Lankhorst <maarten.lankhorst@...ux.intel.com>,
Maxime Ripard <mripard@...nel.org>, Thomas Zimmermann
<tzimmermann@...e.de>, David Airlie <airlied@...il.com>, Simona Vetter
<simona@...ll.ch>
Subject: Re: [PATCH v4 05/10] drm/panfrost: Handle page mapping failure
On Tue, 7 Oct 2025 02:04:00 +0100
Adrián Larumbe <adrian.larumbe@...labora.com> wrote:
> Hi Boris,
>
> On 01.10.2025 12:58, Boris Brezillon wrote:
> > On Wed, 1 Oct 2025 03:20:26 +0100
> > Adrián Larumbe <adrian.larumbe@...labora.com> wrote:
> >
> > > When mapping the pages of a BO, either a heap type at page fault time or
> > > else a non-heap BO at object creation time, if the ARM page table mapping
> > > function fails, we unmap what had been mapped so far and bail out.
> > >
> > > Signed-off-by: Adrián Larumbe <adrian.larumbe@...labora.com>
> > > ---
> > > drivers/gpu/drm/panfrost/panfrost_mmu.c | 49 ++++++++++++++++++++++---
> > > 1 file changed, 44 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/panfrost/panfrost_mmu.c b/drivers/gpu/drm/panfrost/panfrost_mmu.c
> > > index cf272b167feb..fb17c32855a5 100644
> > > --- a/drivers/gpu/drm/panfrost/panfrost_mmu.c
> > > +++ b/drivers/gpu/drm/panfrost/panfrost_mmu.c
> > > @@ -393,13 +393,32 @@ static void panfrost_mmu_flush_range(struct panfrost_device *pfdev,
> > > pm_runtime_put_autosuspend(pfdev->base.dev);
> > > }
> > >
> > > +static void mmu_unmap_range(struct panfrost_mmu *mmu, u64 iova, size_t len)
> > > +{
> > > + struct io_pgtable_ops *ops = mmu->pgtbl_ops;
> > > + size_t pgsize, unmapped_len = 0;
> > > + size_t unmapped_page, pgcount;
> > > +
> > > + while (unmapped_len < len) {
> > > + pgsize = get_pgsize(iova, len - unmapped_len, &pgcount);
> > > +
> > > + unmapped_page = ops->unmap_pages(ops, iova, pgsize, pgcount, NULL);
> > > + WARN_ON(unmapped_page != pgsize * pgcount);
> > > +
> > > + iova += pgsize * pgcount;
> > > + unmapped_len += pgsize * pgcount;
> > > + }
> > > +}
> > > +
> > > static int mmu_map_sg(struct panfrost_device *pfdev, struct panfrost_mmu *mmu,
> > > u64 iova, int prot, struct sg_table *sgt)
> > > {
> > > unsigned int count;
> > > struct scatterlist *sgl;
> > > struct io_pgtable_ops *ops = mmu->pgtbl_ops;
> > > + size_t total_mapped = 0;
> > > u64 start_iova = iova;
> > > + int ret;
> > >
> > > for_each_sgtable_dma_sg(sgt, sgl, count) {
> > > unsigned long paddr = sg_dma_address(sgl);
> > > @@ -413,10 +432,14 @@ static int mmu_map_sg(struct panfrost_device *pfdev, struct panfrost_mmu *mmu,
> > > size_t pgcount, mapped = 0;
> > > size_t pgsize = get_pgsize(iova | paddr, len, &pgcount);
> > >
> > > - ops->map_pages(ops, iova, paddr, pgsize, pgcount, prot,
> > > + ret = ops->map_pages(ops, iova, paddr, pgsize, pgcount, prot,
> > > GFP_KERNEL, &mapped);
> > > + if (ret)
> > > + goto err_unmap_pages;
> > > +
> > > /* Don't get stuck if things have gone wrong */
> > > mapped = max(mapped, pgsize);
> > > + total_mapped += mapped;
> > > iova += mapped;
> > > paddr += mapped;
> > > len -= mapped;
> > > @@ -426,6 +449,10 @@ static int mmu_map_sg(struct panfrost_device *pfdev, struct panfrost_mmu *mmu,
> > > panfrost_mmu_flush_range(pfdev, mmu, start_iova, iova - start_iova);
> > >
> > > return 0;
> > > +
> > > +err_unmap_pages:
> > > + mmu_unmap_range(mmu, start_iova, total_mapped);
> > > + return ret;
> > > }
> > >
> > > int panfrost_mmu_map(struct panfrost_gem_mapping *mapping)
> > > @@ -436,6 +463,7 @@ int panfrost_mmu_map(struct panfrost_gem_mapping *mapping)
> > > struct panfrost_device *pfdev = to_panfrost_device(obj->dev);
> > > struct sg_table *sgt;
> > > int prot = IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE;
> > > + int ret;
> > >
> > > if (WARN_ON(mapping->active))
> > > return 0;
> > > @@ -447,11 +475,18 @@ int panfrost_mmu_map(struct panfrost_gem_mapping *mapping)
> > > if (WARN_ON(IS_ERR(sgt)))
> > > return PTR_ERR(sgt);
> > >
> > > - mmu_map_sg(pfdev, mapping->mmu, mapping->mmnode.start << PAGE_SHIFT,
> > > - prot, sgt);
> > > + ret = mmu_map_sg(pfdev, mapping->mmu, mapping->mmnode.start << PAGE_SHIFT,
> > > + prot, sgt);
> > > + if (ret)
> > > + goto err_put_pages;
> > > +
> > > mapping->active = true;
> > >
> > > return 0;
> > > +
> > > +err_put_pages:
> > > + drm_gem_shmem_put_pages_locked(shmem);
> > > + return ret;
> > > }
> > >
> > > void panfrost_mmu_unmap(struct panfrost_gem_mapping *mapping)
> > > @@ -635,8 +670,10 @@ static int panfrost_mmu_map_fault_addr(struct panfrost_device *pfdev, int as,
> > > if (ret)
> > > goto err_map;
> > >
> > > - mmu_map_sg(pfdev, bomapping->mmu, addr,
> > > - IOMMU_WRITE | IOMMU_READ | IOMMU_CACHE | IOMMU_NOEXEC, sgt);
> > > + ret = mmu_map_sg(pfdev, bomapping->mmu, addr,
> > > + IOMMU_WRITE | IOMMU_READ | IOMMU_CACHE | IOMMU_NOEXEC, sgt);
> > > + if (ret)
> > > + goto err_mmu_map_sg;
> > >
> > > bomapping->active = true;
> > > bo->heap_rss_size += SZ_2M;
> > > @@ -650,6 +687,8 @@ static int panfrost_mmu_map_fault_addr(struct panfrost_device *pfdev, int as,
> > >
> > > return 0;
> > >
> > > +err_mmu_map_sg:
> > > + dma_unmap_sgtable(pfdev->base.dev, sgt, DMA_BIDIRECTIONAL, 0);
> >
> > You also need to clear the sgts[]/pages[] entries you added earlier,
> > otherwise the next time you have a fault it will bail-out before
> > attempting an mmu_map_sg().
> >
> > IIRC, Dmitry had a similar fix in his shmem-shrinker series.
>
> Went over the mailing list and I think the commit you had in mind was 1fc9af813b25 ("drm/panfrost: Fix the error path in panfrost_mmu_map_fault_addr()")
>
> I suspect there's a problem with the present code. If shmem_read_mapping_page() fails for let's say, page_offset+5, then when the interrupt is triggered
> again, because the page array had already been allocated and pages[page_offset] populated in the first try, then it would bail out immediately even though
> most pages haven't been retrieved yet.
>
> On the other hand, depopulating the array for the IRQ to be triggered again seems wasteful. Because for any virtual address, a fault will map all the
> pages within its 2MiB boundaries, maybe we could change
>
> if (pages[page_offset]) {
> /* Pages are already mapped, bail out. */
> goto out;
> }
>
> to 'pages[page_offset+NUM_FAULT_PAGES-1]'
Or, we simply don't check the
pages[page_offset..page_offset+NUM_FAULT_PAGES-1] range and we let
the following loop walk over all entries and fill the missing ones, if
any.
>
> And then, in the event that mmu_map_sg() fails:
>
> err_mmu_map_sg:
> pages[page_offset+NUM_FAULT_PAGES-1] = NULL;
> dma_unmap_sgtable(pfdev->base.dev, sgt, DMA_BIDIRECTIONAL, 0);
>
> So that it'll only fetch the very last page instead of all of them in case of a page fault reattempt.
>
> > > err_map:
> > > sg_free_table(sgt);
> > > err_unlock:
>
>
> Adrian Larumbe
Powered by blists - more mailing lists