Message ID | 1446053858.8018.406.camel@redhat.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On 10/28/2015 06:37 PM, Alex Williamson wrote: > On Wed, 2015-10-28 at 18:10 +0100, Eric Auger wrote: >> Hi Alex, >> On 10/28/2015 05:27 PM, Alex Williamson wrote: >>> On Wed, 2015-10-28 at 13:12 +0000, Eric Auger wrote: >>>> Current vfio_pgsize_bitmap code hides the supported IOMMU page >>>> sizes smaller than PAGE_SIZE. As a result, in case the IOMMU >>>> does not support PAGE_SIZE page, the alignment check on map/unmap >>>> is done with larger page sizes, if any. This can fail although >>>> mapping could be done with pages smaller than PAGE_SIZE. >>>> >>>> vfio_pgsize_bitmap is modified to expose the IOMMU page sizes, >>>> supported by all domains, even those smaller than PAGE_SIZE. The >>>> alignment check on map is performed against PAGE_SIZE if the minimum >>>> IOMMU size is less than PAGE_SIZE or against the min page size greater >>>> than PAGE_SIZE. >>>> >>>> Signed-off-by: Eric Auger <eric.auger@linaro.org> >>>> >>>> --- >>>> >>>> This was tested on AMD Seattle with 64kB page host. ARM MMU 401 >>>> currently expose 4kB, 2MB and 1GB page support. With a 64kB page host, >>>> the map/unmap check is done against 2MB. Some alignment check fail >>>> so VFIO_IOMMU_MAP_DMA fail while we could map using 4kB IOMMU page >>>> size. >>>> --- >>>> drivers/vfio/vfio_iommu_type1.c | 25 +++++++++++-------------- >>>> 1 file changed, 11 insertions(+), 14 deletions(-) >>>> >>>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c >>>> index 57d8c37..13fb974 100644 >>>> --- a/drivers/vfio/vfio_iommu_type1.c >>>> +++ b/drivers/vfio/vfio_iommu_type1.c >>>> @@ -403,7 +403,7 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) >>>> static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) >>>> { >>>> struct vfio_domain *domain; >>>> - unsigned long bitmap = PAGE_MASK; >>>> + unsigned long bitmap = ULONG_MAX; >>> >>> Isn't this and removing the WARN_ON()s the only real change in this >>> patch? The rest looks like conversion to use IS_ALIGNED and the >>> following test, that I don't really understand... >> Yes basically you're right. > > > Ok, so with hopefully correcting my understand of what this does, isn't > this effectively the same: > > diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c > index 57d8c37..7db4f5a 100644 > --- a/drivers/vfio/vfio_iommu_type1.c > +++ b/drivers/vfio/vfio_iommu_type1.c > @@ -403,13 +403,19 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, stru > static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) > { > struct vfio_domain *domain; > - unsigned long bitmap = PAGE_MASK; > + unsigned long bitmap = ULONG_MAX; > > mutex_lock(&iommu->lock); > list_for_each_entry(domain, &iommu->domain_list, next) > bitmap &= domain->domain->ops->pgsize_bitmap; > mutex_unlock(&iommu->lock); > > + /* Some comment about how the IOMMU API splits requests */ > + if (bitmap & ~PAGE_MASK) { > + bitmap &= PAGE_MASK; > + bitmap |= PAGE_SIZE; > + } > + > return bitmap; > } Yes, to me it is indeed the same > > This would also expose to the user that we're accepting PAGE_SIZE, which > we weren't before, so it was not quite right to just let them do it > anyway. I don't think we even need to get rid of the WARN_ONs, do we? > Thanks, The end-user might be afraid of those latter. Personally I would get rid of them but that's definitively up to you. Just let me know and I will respin. Best Regards Eric > > Alex > >>> >>>> >>>> mutex_lock(&iommu->lock); >>>> list_for_each_entry(domain, &iommu->domain_list, next) >>>> @@ -416,20 +416,18 @@ static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) >>>> static int vfio_dma_do_unmap(struct vfio_iommu *iommu, >>>> struct vfio_iommu_type1_dma_unmap *unmap) >>>> { >>>> - uint64_t mask; >>>> struct vfio_dma *dma; >>>> size_t unmapped = 0; >>>> int ret = 0; >>>> + unsigned int min_pagesz = __ffs(vfio_pgsize_bitmap(iommu)); >>>> + unsigned int requested_alignment = (min_pagesz < PAGE_SIZE) ? >>>> + PAGE_SIZE : min_pagesz; >>> >>> This one. If we're going to support sub-PAGE_SIZE mappings, why do we >>> care to cap alignment at PAGE_SIZE? >> My intent in this patch isn't to allow the user-space to map/unmap >> sub-PAGE_SIZE buffers. The new test makes sure the mapped area is bigger >> or equal than a host page whatever the supported page sizes. >> >> I noticed that chunk construction, pinning and other many things are >> based on PAGE_SIZE and far be it from me to change that code! I want to >> keep that minimal granularity for all those computation. >> >> However on iommu side, I would like to rely on the fact the iommu driver >> is clever enough to choose the right page size and even to choose a size >> that is smaller than PAGE_SIZE if this latter is not supported. >>> >>>> - mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; >>>> - >>>> - if (unmap->iova & mask) >>>> + if (!IS_ALIGNED(unmap->iova, requested_alignment)) >>>> return -EINVAL; >>>> - if (!unmap->size || unmap->size & mask) >>>> + if (!unmap->size || !IS_ALIGNED(unmap->size, requested_alignment)) >>>> return -EINVAL; >>>> >>>> - WARN_ON(mask & PAGE_MASK); >>>> - >>>> mutex_lock(&iommu->lock); >>>> >>>> /* >>>> @@ -553,25 +551,24 @@ static int vfio_dma_do_map(struct vfio_iommu *iommu, >>>> size_t size = map->size; >>>> long npage; >>>> int ret = 0, prot = 0; >>>> - uint64_t mask; >>>> struct vfio_dma *dma; >>>> unsigned long pfn; >>>> + unsigned int min_pagesz = __ffs(vfio_pgsize_bitmap(iommu)); >>>> + unsigned int requested_alignment = (min_pagesz < PAGE_SIZE) ? >>>> + PAGE_SIZE : min_pagesz; >>>> >>>> /* Verify that none of our __u64 fields overflow */ >>>> if (map->size != size || map->vaddr != vaddr || map->iova != iova) >>>> return -EINVAL; >>>> >>>> - mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; >>>> - >>>> - WARN_ON(mask & PAGE_MASK); >>>> - >>>> /* READ/WRITE from device perspective */ >>>> if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) >>>> prot |= IOMMU_WRITE; >>>> if (map->flags & VFIO_DMA_MAP_FLAG_READ) >>>> prot |= IOMMU_READ; >>>> >>>> - if (!prot || !size || (size | iova | vaddr) & mask) >>>> + if (!prot || !size || >>>> + !IS_ALIGNED(size | iova | vaddr, requested_alignment)) >>>> return -EINVAL; >>>> >>>> /* Don't allow IOVA or virtual address wrap */ >>> >>> This is mostly ignoring the problems with sub-PAGE_SIZE mappings. For >>> instance, we can only pin on PAGE_SIZE and therefore we only do >>> accounting on PAGE_SIZE, so if the user does 4K mappings across your 64K >>> page, that page gets pinned and accounted 16 times. Are we going to >>> tell users that their locked memory limit needs to be 16x now? The rest >>> of the code would need an audit as well to see what other sub-page bugs >>> might be hiding. Thanks, >> So if the user is not allowed to map sub-PAGE_SIZE buffers, accounting >> still is based on PAGE_SIZE while iommu mapping can be based on >> sub-PAGE_SIZE pages. I am misunderstanding something? >> >> Best Regards >> >> Eric >>> >>> Alex >>> >>> >>> >> > > >
On Wed, Oct 28, 2015 at 06:48:41PM +0100, Eric Auger wrote: > On 10/28/2015 06:37 PM, Alex Williamson wrote: > > Ok, so with hopefully correcting my understand of what this does, isn't > > this effectively the same: > > > > diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c > > index 57d8c37..7db4f5a 100644 > > --- a/drivers/vfio/vfio_iommu_type1.c > > +++ b/drivers/vfio/vfio_iommu_type1.c > > @@ -403,13 +403,19 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, stru > > static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) > > { > > struct vfio_domain *domain; > > - unsigned long bitmap = PAGE_MASK; > > + unsigned long bitmap = ULONG_MAX; > > > > mutex_lock(&iommu->lock); > > list_for_each_entry(domain, &iommu->domain_list, next) > > bitmap &= domain->domain->ops->pgsize_bitmap; > > mutex_unlock(&iommu->lock); > > > > + /* Some comment about how the IOMMU API splits requests */ > > + if (bitmap & ~PAGE_MASK) { > > + bitmap &= PAGE_MASK; > > + bitmap |= PAGE_SIZE; > > + } > > + > > return bitmap; > > } > Yes, to me it is indeed the same > > > > This would also expose to the user that we're accepting PAGE_SIZE, which > > we weren't before, so it was not quite right to just let them do it > > anyway. I don't think we even need to get rid of the WARN_ONs, do we? > > Thanks, > > The end-user might be afraid of those latter. Personally I would get rid > of them but that's definitively up to you. I think Alex's point is that the WARN_ON's won't trigger with this patch, because he clears those lower bits in the bitmap. Will
On 10/28/2015 06:55 PM, Will Deacon wrote: > On Wed, Oct 28, 2015 at 06:48:41PM +0100, Eric Auger wrote: >> On 10/28/2015 06:37 PM, Alex Williamson wrote: >>> Ok, so with hopefully correcting my understand of what this does, isn't >>> this effectively the same: >>> >>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c >>> index 57d8c37..7db4f5a 100644 >>> --- a/drivers/vfio/vfio_iommu_type1.c >>> +++ b/drivers/vfio/vfio_iommu_type1.c >>> @@ -403,13 +403,19 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, stru >>> static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) >>> { >>> struct vfio_domain *domain; >>> - unsigned long bitmap = PAGE_MASK; >>> + unsigned long bitmap = ULONG_MAX; >>> >>> mutex_lock(&iommu->lock); >>> list_for_each_entry(domain, &iommu->domain_list, next) >>> bitmap &= domain->domain->ops->pgsize_bitmap; >>> mutex_unlock(&iommu->lock); >>> >>> + /* Some comment about how the IOMMU API splits requests */ >>> + if (bitmap & ~PAGE_MASK) { >>> + bitmap &= PAGE_MASK; >>> + bitmap |= PAGE_SIZE; >>> + } >>> + >>> return bitmap; >>> } >> Yes, to me it is indeed the same >>> >>> This would also expose to the user that we're accepting PAGE_SIZE, which >>> we weren't before, so it was not quite right to just let them do it >>> anyway. I don't think we even need to get rid of the WARN_ONs, do we? >>> Thanks, >> >> The end-user might be afraid of those latter. Personally I would get rid >> of them but that's definitively up to you. > > I think Alex's point is that the WARN_ON's won't trigger with this patch, > because he clears those lower bits in the bitmap. ah yes sure! Thanks Eric > > Will >
On Wed, 2015-10-28 at 19:00 +0100, Eric Auger wrote: > On 10/28/2015 06:55 PM, Will Deacon wrote: > > On Wed, Oct 28, 2015 at 06:48:41PM +0100, Eric Auger wrote: > >> On 10/28/2015 06:37 PM, Alex Williamson wrote: > >>> Ok, so with hopefully correcting my understand of what this does, isn't > >>> this effectively the same: > >>> > >>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c > >>> index 57d8c37..7db4f5a 100644 > >>> --- a/drivers/vfio/vfio_iommu_type1.c > >>> +++ b/drivers/vfio/vfio_iommu_type1.c > >>> @@ -403,13 +403,19 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, stru > >>> static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) > >>> { > >>> struct vfio_domain *domain; > >>> - unsigned long bitmap = PAGE_MASK; > >>> + unsigned long bitmap = ULONG_MAX; > >>> > >>> mutex_lock(&iommu->lock); > >>> list_for_each_entry(domain, &iommu->domain_list, next) > >>> bitmap &= domain->domain->ops->pgsize_bitmap; > >>> mutex_unlock(&iommu->lock); > >>> > >>> + /* Some comment about how the IOMMU API splits requests */ > >>> + if (bitmap & ~PAGE_MASK) { > >>> + bitmap &= PAGE_MASK; > >>> + bitmap |= PAGE_SIZE; > >>> + } > >>> + > >>> return bitmap; > >>> } > >> Yes, to me it is indeed the same > >>> > >>> This would also expose to the user that we're accepting PAGE_SIZE, which > >>> we weren't before, so it was not quite right to just let them do it > >>> anyway. I don't think we even need to get rid of the WARN_ONs, do we? > >>> Thanks, > >> > >> The end-user might be afraid of those latter. Personally I would get rid > >> of them but that's definitively up to you. > > > > I think Alex's point is that the WARN_ON's won't trigger with this patch, > > because he clears those lower bits in the bitmap. > ah yes sure! The WARN_ON triggers when the IOMMU mask is greater than PAGE_SIZE, which means we can't operate on the IOMMU with PAGE_SIZE granularity, which we do in a couple places. So I think the WARN_ON is actually valid for the code and won't trigger for you now that the IOMMU mask is always at least ~PAGE_MASK if we can use the IOMMU at anything less than PAGE_SIZE granularity. Thanks, Alex
diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c index 57d8c37..7db4f5a 100644 --- a/drivers/vfio/vfio_iommu_type1.c +++ b/drivers/vfio/vfio_iommu_type1.c @@ -403,13 +403,19 @@ static void vfio_remove_dma(struct vfio_iommu *iommu, stru static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) { struct vfio_domain *domain; - unsigned long bitmap = PAGE_MASK; + unsigned long bitmap = ULONG_MAX; mutex_lock(&iommu->lock); list_for_each_entry(domain, &iommu->domain_list, next) bitmap &= domain->domain->ops->pgsize_bitmap; mutex_unlock(&iommu->lock); + /* Some comment about how the IOMMU API splits requests */ + if (bitmap & ~PAGE_MASK) { + bitmap &= PAGE_MASK; + bitmap |= PAGE_SIZE; + } + return bitmap; }