On Fri, Feb 26, 2021 at 10:41 AM Thomas Hellström (Intel) thomas_os@shipmail.org wrote:

On 2/25/21 4:49 PM, Daniel Vetter wrote:

...

On Thu, Feb 25, 2021 at 11:44 AM Daniel Vetter daniel@ffwll.ch wrote:

...

On Thu, Feb 25, 2021 at 11:28:31AM +0100, Christian König wrote:

...

Am 24.02.21 um 10:31 schrieb Daniel Vetter:

...

On Wed, Feb 24, 2021 at 10:16 AM Thomas Hellström (Intel) thomas_os@shipmail.org wrote:

...

On 2/24/21 9:45 AM, Daniel Vetter wrote: > On Wed, Feb 24, 2021 at 8:46 AM Thomas Hellström (Intel) > thomas_os@shipmail.org wrote: >> On 2/23/21 11:59 AM, Daniel Vetter wrote: >>> tldr; DMA buffers aren't normal memory, expecting that you can use >>> them like that (like calling get_user_pages works, or that they're >>> accounting like any other normal memory) cannot be guaranteed. >>> >>> Since some userspace only runs on integrated devices, where all >>> buffers are actually all resident system memory, there's a huge >>> temptation to assume that a struct page is always present and useable >>> like for any more pagecache backed mmap. This has the potential to >>> result in a uapi nightmare. >>> >>> To stop this gap require that DMA buffer mmaps are VM_PFNMAP, which >>> blocks get_user_pages and all the other struct page based >>> infrastructure for everyone. In spirit this is the uapi counterpart to >>> the kernel-internal CONFIG_DMABUF_DEBUG. >>> >>> Motivated by a recent patch which wanted to swich the system dma-buf >>> heap to vm_insert_page instead of vm_insert_pfn. >>> >>> v2: >>> >>> Jason brought up that we also want to guarantee that all ptes have the >>> pte_special flag set, to catch fast get_user_pages (on architectures >>> that support this). Allowing VM_MIXEDMAP (like VM_SPECIAL does) would >>> still allow vm_insert_page, but limiting to VM_PFNMAP will catch that. >>> >>> From auditing the various functions to insert pfn pte entires >>> (vm_insert_pfn_prot, remap_pfn_range and all it's callers like >>> dma_mmap_wc) it looks like VM_PFNMAP is already required anyway, so >>> this should be the correct flag to check for. >>> >> If we require VM_PFNMAP, for ordinary page mappings, we also need to >> disallow COW mappings, since it will not work on architectures that >> don't have CONFIG_ARCH_HAS_PTE_SPECIAL, (see the docs for vm_normal_page()). > Hm I figured everyone just uses MAP_SHARED for buffer objects since > COW really makes absolutely no sense. How would we enforce this? Perhaps returning -EINVAL on is_cow_mapping() at mmap time. Either that or allowing MIXEDMAP.

>> Also worth noting is the comment in ttm_bo_mmap_vma_setup() with >> possible performance implications with x86 + PAT + VM_PFNMAP + normal >> pages. That's a very old comment, though, and might not be valid anymore. > I think that's why ttm has a page cache for these, because it indeed > sucks. The PAT changes on pages are rather expensive. IIRC the page cache was implemented because of the slowness of the caching mode transition itself, more specifically the wbinvd() call + global TLB flush.

Yes, exactly that. The global TLB flush is what really breaks our neck here from a performance perspective.

...

...

> There is still an issue for iomem mappings, because the PAT validation > does a linear walk of the resource tree (lol) for every vm_insert_pfn. > But for i915 at least this is fixed by using the io_mapping > infrastructure, which does the PAT reservation only once when you set > up the mapping area at driver load. Yes, I guess that was the issue that the comment describes, but the issue wasn't there with vm_insert_mixed() + VM_MIXEDMAP.

> Also TTM uses VM_PFNMAP right now for everything, so it can't be a > problem that hurts much :-) Hmm, both 5.11 and drm-tip appears to still use MIXEDMAP?

https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/ttm/ttm_bo_vm...

Uh that's bad, because mixed maps pointing at struct page wont stop gup. At least afaik.

Hui? I'm pretty sure MIXEDMAP stops gup as well. Otherwise we would have already seen tons of problems with the page cache.

On any architecture which has CONFIG_ARCH_HAS_PTE_SPECIAL vm_insert_mixed boils down to vm_insert_pfn wrt gup. And special pte stops gup fast path.

But if you don't have VM_IO or VM_PFNMAP set, then I'm not seeing how you're stopping gup slow path. See check_vma_flags() in mm/gup.c.

Also if you don't have CONFIG_ARCH_HAS_PTE_SPECIAL then I don't think vm_insert_mixed even works on iomem pfns. There's the devmap exception, but we're not devmap. Worse ttm abuses some accidental codepath to smuggle in hugepte support by intentionally not being devmap.

So I'm really not sure this works as we think it should. Maybe good to do a quick test program on amdgpu with a buffer in system memory only and try to do direct io into it. If it works, you have a problem, and a bad one.

That's probably impossible, since a quick git grep shows that pretty much anything reasonable has special ptes: arc, arm, arm64, powerpc, riscv, s390, sh, sparc, x86. I don't think you'll have a platform where you can plug an amdgpu in and actually exercise the bug :-)

Hm. AFAIK _insert_mixed() doesn't set PTE_SPECIAL on system pages, so I don't see what should be stopping gup to those?

If you have an arch with pte special we use insert_pfn(), which afaict will use pte_mkspecial for the !devmap case. And ttm isn't devmap (otherwise our hugepte abuse of devmap hugeptes would go rather wrong).

So I think it stops gup. But I haven't verified at all. Would be good if Christian can check this with some direct io to a buffer in system memory. -Daniel