| Message ID | 20190129165839.4127-2-jglisse@redhat.com (mailing list archive) |
|---|---|
| State | RFC |
| Headers | show |
| Series | Use HMM for ODP | expand |
On 1/29/2019 6:58 PM, jglisse@redhat.com wrote: > Convert ODP to use HMM so that we can build on common infrastructure > for different class of devices that want to mirror a process address > space into a device. There is no functional changes. Thanks for sending this patch. I think in general it is a good idea to use a common infrastructure for ODP. I have a couple of questions below. > -static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn, > - const struct mmu_notifier_range *range) > -{ > - struct ib_ucontext_per_mm *per_mm = > - container_of(mn, struct ib_ucontext_per_mm, mn); > - > - if (unlikely(!per_mm->active)) > - return; > - > - rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start, > - range->end, > - invalidate_range_end_trampoline, true, NULL); > up_read(&per_mm->umem_rwsem); > + return ret; > } Previously the code held the umem_rwsem between range_start and range_end calls. I guess that was in order to guarantee that no device page faults take reference to the pages being invalidated while the invalidation is ongoing. I assume this is now handled by hmm instead, correct? > + > +static uint64_t odp_hmm_flags[HMM_PFN_FLAG_MAX] = { > + ODP_READ_BIT, /* HMM_PFN_VALID */ > + ODP_WRITE_BIT, /* HMM_PFN_WRITE */ > + ODP_DEVICE_BIT, /* HMM_PFN_DEVICE_PRIVATE */ It seems that the mlx5_ib code in this patch currently ignores the ODP_DEVICE_BIT (e.g., in umem_dma_to_mtt). Is that okay? Or is it handled implicitly by the HMM_PFN_SPECIAL case? > @@ -327,9 +287,10 @@ void put_per_mm(struct ib_umem_odp *umem_odp) > up_write(&per_mm->umem_rwsem); > > WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root)); > - mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm); > + hmm_mirror_unregister(&per_mm->mirror); > put_pid(per_mm->tgid); > - mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm); > + > + kfree(per_mm); > } Previously the per_mm struct was released through call srcu, but now it is released immediately. Is it safe? I saw that hmm_mirror_unregister calls mmu_notifier_unregister_no_release, so I don't understand what prevents concurrently running invalidations from accessing the released per_mm struct. > @@ -578,11 +578,27 @@ static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr, > > next_mr: > size = min_t(size_t, bcnt, ib_umem_end(&odp->umem) - io_virt); > - > page_shift = mr->umem->page_shift; > page_mask = ~(BIT(page_shift) - 1); > + off = (io_virt & (~page_mask)); > + size += (io_virt & (~page_mask)); > + io_virt = io_virt & page_mask; > + off += (size & (~page_mask)); > + size = ALIGN(size, 1UL << page_shift); > + > + if (io_virt < ib_umem_start(&odp->umem)) > + return -EINVAL; > + > start_idx = (io_virt - (mr->mmkey.iova & page_mask)) >> page_shift; > > + if (odp_mr->per_mm == NULL || odp_mr->per_mm->mm == NULL) > + return -ENOENT; > + > + ret = hmm_range_register(&range, odp_mr->per_mm->mm, > + io_virt, io_virt + size, page_shift); > + if (ret) > + return ret; > + > if (prefetch && !downgrade && !mr->umem->writable) { > /* prefetch with write-access must > * be supported by the MR Isn't there a mistake in the calculation of the variable size? Itis first set to the size of the page fault range, but then you add the virtual address, so I guess it is actually the range end. Then you pass io_virt + size to hmm_range_register. Doesn't it double the size of the range > -void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt, > - u64 bound) > +void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, > + u64 virt, u64 bound) > { > + struct device *device = umem_odp->umem.context->device->dma_device; > struct ib_umem *umem = &umem_odp->umem; > - int idx; > - u64 addr; > - struct ib_device *dev = umem->context->device; > + unsigned long idx, page_mask; > + struct hmm_range range; > + long ret; > + > + if (!umem->npages) > + return; > + > + bound = ALIGN(bound, 1UL << umem->page_shift); > + page_mask = ~(BIT(umem->page_shift) - 1); > + virt &= page_mask; > > virt = max_t(u64, virt, ib_umem_start(umem)); > bound = min_t(u64, bound, ib_umem_end(umem)); > - /* Note that during the run of this function, the > - * notifiers_count of the MR is > 0, preventing any racing > - * faults from completion. We might be racing with other > - * invalidations, so we must make sure we free each page only > - * once. */ > + > + idx = ((unsigned long)virt - ib_umem_start(umem)) >> PAGE_SHIFT; > + > + range.page_shift = umem->page_shift; > + range.pfns = &umem_odp->pfns[idx]; > + range.pfn_shift = ODP_FLAGS_BITS; > + range.values = odp_hmm_values; > + range.flags = odp_hmm_flags; > + range.start = virt; > + range.end = bound; > + > mutex_lock(&umem_odp->umem_mutex); > - for (addr = virt; addr < bound; addr += BIT(umem->page_shift)) { > - idx = (addr - ib_umem_start(umem)) >> umem->page_shift; > - if (umem_odp->page_list[idx]) { > - struct page *page = umem_odp->page_list[idx]; > - dma_addr_t dma = umem_odp->dma_list[idx]; > - dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK; > - > - WARN_ON(!dma_addr); > - > - ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE, > - DMA_BIDIRECTIONAL); > - if (dma & ODP_WRITE_ALLOWED_BIT) { > - struct page *head_page = compound_head(page); > - /* > - * set_page_dirty prefers being called with > - * the page lock. However, MMU notifiers are > - * called sometimes with and sometimes without > - * the lock. We rely on the umem_mutex instead > - * to prevent other mmu notifiers from > - * continuing and allowing the page mapping to > - * be removed. > - */ > - set_page_dirty(head_page); > - } > - /* on demand pinning support */ > - if (!umem->context->invalidate_range) > - put_page(page); > - umem_odp->page_list[idx] = NULL; > - umem_odp->dma_list[idx] = 0; > - umem->npages--; > - } > - } > + ret = hmm_range_dma_unmap(&range, NULL, device, > + &umem_odp->dma_list[idx], true); > + if (ret > 0) > + umem->npages -= ret; Can hmm_range_dma_unmap fail? If it does, we do we simply leak the DMA mappings? > mutex_unlock(&umem_odp->umem_mutex); > } Regards, Haggai
On Wed, Feb 06, 2019 at 08:44:26AM +0000, Haggai Eran wrote: > On 1/29/2019 6:58 PM, jglisse@redhat.com wrote: > > Convert ODP to use HMM so that we can build on common infrastructure > > for different class of devices that want to mirror a process address > > space into a device. There is no functional changes. > > Thanks for sending this patch. I think in general it is a good idea to > use a common infrastructure for ODP. > > I have a couple of questions below. > > > -static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn, > > - const struct mmu_notifier_range *range) > > -{ > > - struct ib_ucontext_per_mm *per_mm = > > - container_of(mn, struct ib_ucontext_per_mm, mn); > > - > > - if (unlikely(!per_mm->active)) > > - return; > > - > > - rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start, > > - range->end, > > - invalidate_range_end_trampoline, true, NULL); > > up_read(&per_mm->umem_rwsem); > > + return ret; > > } > Previously the code held the umem_rwsem between range_start and > range_end calls. I guess that was in order to guarantee that no device > page faults take reference to the pages being invalidated while the > invalidation is ongoing. I assume this is now handled by hmm instead, > correct? It is a mix of HMM and driver in pagefault_mr() mlx5/odp.c mutex_lock(&odp->umem_mutex); if (hmm_vma_range_done(range)) { ... This is what serialize programming the hw and any concurrent CPU page table invalidation. This is also one of the thing i want to improve long term as mlx5_ib_update_xlt() can do memory allocation and i would like to avoid that ie make mlx5_ib_update_xlt() and its sub-functions as small and to the points as possible so that they could only fail if the hardware is in bad state not because of memory allocation issues. > > > + > > +static uint64_t odp_hmm_flags[HMM_PFN_FLAG_MAX] = { > > + ODP_READ_BIT, /* HMM_PFN_VALID */ > > + ODP_WRITE_BIT, /* HMM_PFN_WRITE */ > > + ODP_DEVICE_BIT, /* HMM_PFN_DEVICE_PRIVATE */ > It seems that the mlx5_ib code in this patch currently ignores the > ODP_DEVICE_BIT (e.g., in umem_dma_to_mtt). Is that okay? Or is it > handled implicitly by the HMM_PFN_SPECIAL case? This is because HMM except a bit for device memory as same API is use for GPU which have device memory. I can add a comment explaining that it is not use for ODP but there just to comply with HMM API. > > > @@ -327,9 +287,10 @@ void put_per_mm(struct ib_umem_odp *umem_odp) > > up_write(&per_mm->umem_rwsem); > > > > WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root)); > > - mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm); > > + hmm_mirror_unregister(&per_mm->mirror); > > put_pid(per_mm->tgid); > > - mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm); > > + > > + kfree(per_mm); > > } > Previously the per_mm struct was released through call srcu, but now it > is released immediately. Is it safe? I saw that hmm_mirror_unregister > calls mmu_notifier_unregister_no_release, so I don't understand what > prevents concurrently running invalidations from accessing the released > per_mm struct. Yes it is safe, the hmm struct has its own refcount and mirror holds a reference on it, the mm struct itself has a reference on the mm struct. So no structure can vanish before the other. However once release call- back happens you can no longer fault anything it will -EFAULT if you try to (not to mention that by then all the vma have been tear down). So even if some kernel thread race with destruction it will not be able to fault anything or use mirror struct in any meaning full way. Note that in a regular tear down the ODP put_per_mm() will happen before the release callback as iirc file including device file get close before the mm is teardown. But in anycase it would work no matter what the order is. > > > @@ -578,11 +578,27 @@ static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr, > > > > next_mr: > > size = min_t(size_t, bcnt, ib_umem_end(&odp->umem) - io_virt); > > - > > page_shift = mr->umem->page_shift; > > page_mask = ~(BIT(page_shift) - 1); > > + off = (io_virt & (~page_mask)); > > + size += (io_virt & (~page_mask)); > > + io_virt = io_virt & page_mask; > > + off += (size & (~page_mask)); > > + size = ALIGN(size, 1UL << page_shift); > > + > > + if (io_virt < ib_umem_start(&odp->umem)) > > + return -EINVAL; > > + > > start_idx = (io_virt - (mr->mmkey.iova & page_mask)) >> page_shift; > > > > + if (odp_mr->per_mm == NULL || odp_mr->per_mm->mm == NULL) > > + return -ENOENT; > > + > > + ret = hmm_range_register(&range, odp_mr->per_mm->mm, > > + io_virt, io_virt + size, page_shift); > > + if (ret) > > + return ret; > > + > > if (prefetch && !downgrade && !mr->umem->writable) { > > /* prefetch with write-access must > > * be supported by the MR > Isn't there a mistake in the calculation of the variable size? Itis > first set to the size of the page fault range, but then you add the > virtual address, so I guess it is actually the range end. Then you pass > io_virt + size to hmm_range_register. Doesn't it double the size of the > range No i think it is correct, bcnt is the byte count we are ask to fault, we align that on the maximum size the current mr covers (min_t above) then we align with the page size so that fault address is page align. hmm_range_register() takes start address and end address which is the start address + size. off is the offset ie the number of extra byte we are faulting to align start on page size. If there is a bug this might be: off += (size & (~page_mask)); I need to review before and after to double check that again. > > > -void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt, > > - u64 bound) > > +void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, > > + u64 virt, u64 bound) > > { > > + struct device *device = umem_odp->umem.context->device->dma_device; > > struct ib_umem *umem = &umem_odp->umem; > > - int idx; > > - u64 addr; > > - struct ib_device *dev = umem->context->device; > > + unsigned long idx, page_mask; > > + struct hmm_range range; > > + long ret; > > + > > + if (!umem->npages) > > + return; > > + > > + bound = ALIGN(bound, 1UL << umem->page_shift); > > + page_mask = ~(BIT(umem->page_shift) - 1); > > + virt &= page_mask; > > > > virt = max_t(u64, virt, ib_umem_start(umem)); > > bound = min_t(u64, bound, ib_umem_end(umem)); > > - /* Note that during the run of this function, the > > - * notifiers_count of the MR is > 0, preventing any racing > > - * faults from completion. We might be racing with other > > - * invalidations, so we must make sure we free each page only > > - * once. */ > > + > > + idx = ((unsigned long)virt - ib_umem_start(umem)) >> PAGE_SHIFT; > > + > > + range.page_shift = umem->page_shift; > > + range.pfns = &umem_odp->pfns[idx]; > > + range.pfn_shift = ODP_FLAGS_BITS; > > + range.values = odp_hmm_values; > > + range.flags = odp_hmm_flags; > > + range.start = virt; > > + range.end = bound; > > + > > mutex_lock(&umem_odp->umem_mutex); > > - for (addr = virt; addr < bound; addr += BIT(umem->page_shift)) { > > - idx = (addr - ib_umem_start(umem)) >> umem->page_shift; > > - if (umem_odp->page_list[idx]) { > > - struct page *page = umem_odp->page_list[idx]; > > - dma_addr_t dma = umem_odp->dma_list[idx]; > > - dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK; > > - > > - WARN_ON(!dma_addr); > > - > > - ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE, > > - DMA_BIDIRECTIONAL); > > - if (dma & ODP_WRITE_ALLOWED_BIT) { > > - struct page *head_page = compound_head(page); > > - /* > > - * set_page_dirty prefers being called with > > - * the page lock. However, MMU notifiers are > > - * called sometimes with and sometimes without > > - * the lock. We rely on the umem_mutex instead > > - * to prevent other mmu notifiers from > > - * continuing and allowing the page mapping to > > - * be removed. > > - */ > > - set_page_dirty(head_page); > > - } > > - /* on demand pinning support */ > > - if (!umem->context->invalidate_range) > > - put_page(page); > > - umem_odp->page_list[idx] = NULL; > > - umem_odp->dma_list[idx] = 0; > > - umem->npages--; > > - } > > - } > > + ret = hmm_range_dma_unmap(&range, NULL, device, > > + &umem_odp->dma_list[idx], true); > > + if (ret > 0) > > + umem->npages -= ret; > Can hmm_range_dma_unmap fail? If it does, we do we simply leak the DMA > mappings? It can only fails if you provide bogus range structure (like end address before start address). This is just a safety next really. It also returns the number of page that have been unmap just like hmm_range_dma_map() returns the number of page that have been map. So you can keep a count of the number of pages map with umem->npages Cheers, Jérôme
On 2/12/2019 6:11 PM, Jerome Glisse wrote: > On Wed, Feb 06, 2019 at 08:44:26AM +0000, Haggai Eran wrote: >> On 1/29/2019 6:58 PM, jglisse@redhat.com wrote: >> > Convert ODP to use HMM so that we can build on common infrastructure >> > for different class of devices that want to mirror a process address >> > space into a device. There is no functional changes. >> >> Thanks for sending this patch. I think in general it is a good idea to >> use a common infrastructure for ODP. >> >> I have a couple of questions below. >> >>> -static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn, >>> - const struct mmu_notifier_range *range) >>> -{ >>> - struct ib_ucontext_per_mm *per_mm = >>> - container_of(mn, struct ib_ucontext_per_mm, mn); >>> - >>> - if (unlikely(!per_mm->active)) >>> - return; >>> - >>> - rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start, >>> - range->end, >>> - invalidate_range_end_trampoline, true, NULL); >>> up_read(&per_mm->umem_rwsem); >>> + return ret; >>> } >> Previously the code held the umem_rwsem between range_start and >> range_end calls. I guess that was in order to guarantee that no device >> page faults take reference to the pages being invalidated while the >> invalidation is ongoing. I assume this is now handled by hmm instead, >> correct? > > It is a mix of HMM and driver in pagefault_mr() mlx5/odp.c > mutex_lock(&odp->umem_mutex); > if (hmm_vma_range_done(range)) { > ... > > This is what serialize programming the hw and any concurrent CPU page > table invalidation. This is also one of the thing i want to improve > long term as mlx5_ib_update_xlt() can do memory allocation and i would > like to avoid that ie make mlx5_ib_update_xlt() and its sub-functions > as small and to the points as possible so that they could only fail if > the hardware is in bad state not because of memory allocation issues. I wonder if it would be possible to make use of the memory that is already allocated (ib_umem_odp->dma_list) for that purpose. This would probably mean that this area will need to be formatted according to the device hardware requirements (e.g., big endian), and then you can instruct the device to DMA the updated translations directly from their. > > >> >>> + >>> +static uint64_t odp_hmm_flags[HMM_PFN_FLAG_MAX] = { >>> + ODP_READ_BIT, /* HMM_PFN_VALID */ >>> + ODP_WRITE_BIT, /* HMM_PFN_WRITE */ >>> + ODP_DEVICE_BIT, /* HMM_PFN_DEVICE_PRIVATE */ >> It seems that the mlx5_ib code in this patch currently ignores the >> ODP_DEVICE_BIT (e.g., in umem_dma_to_mtt). Is that okay? Or is it >> handled implicitly by the HMM_PFN_SPECIAL case? > > This is because HMM except a bit for device memory as same API is > use for GPU which have device memory. I can add a comment explaining > that it is not use for ODP but there just to comply with HMM API. > >> >>> @@ -327,9 +287,10 @@ void put_per_mm(struct ib_umem_odp *umem_odp) >>> up_write(&per_mm->umem_rwsem); >>> >>> WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root)); >>> - mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm); >>> + hmm_mirror_unregister(&per_mm->mirror); >>> put_pid(per_mm->tgid); >>> - mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm); >>> + >>> + kfree(per_mm); >>> } >> Previously the per_mm struct was released through call srcu, but now it >> is released immediately. Is it safe? I saw that hmm_mirror_unregister >> calls mmu_notifier_unregister_no_release, so I don't understand what >> prevents concurrently running invalidations from accessing the released >> per_mm struct. > > Yes it is safe, the hmm struct has its own refcount and mirror holds a > reference on it, the mm struct itself has a reference on the mm struct. > So no structure can vanish before the other. However once release call- > back happens you can no longer fault anything it will -EFAULT if you > try to (not to mention that by then all the vma have been tear down). > So even if some kernel thread race with destruction it will not be able > to fault anything or use mirror struct in any meaning full way. > > Note that in a regular tear down the ODP put_per_mm() will happen before > the release callback as iirc file including device file get close before > the mm is teardown. But in anycase it would work no matter what the order > is. I see. I was worried about concurrent invalidations and ibv_dereg_mr, but I understand hmm protects against that internally through the mirrors_sem semaphore. > >> >>> @@ -578,11 +578,27 @@ static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr, >>> >>> next_mr: >>> size = min_t(size_t, bcnt, ib_umem_end(&odp->umem) - io_virt); >>> - >>> page_shift = mr->umem->page_shift; >>> page_mask = ~(BIT(page_shift) - 1); >>> + off = (io_virt & (~page_mask)); >>> + size += (io_virt & (~page_mask)); >>> + io_virt = io_virt & page_mask; >>> + off += (size & (~page_mask)); >>> + size = ALIGN(size, 1UL << page_shift); >>> + >>> + if (io_virt < ib_umem_start(&odp->umem)) >>> + return -EINVAL; >>> + >>> start_idx = (io_virt - (mr->mmkey.iova & page_mask)) >> page_shift; >>> >>> + if (odp_mr->per_mm == NULL || odp_mr->per_mm->mm == NULL) >>> + return -ENOENT; >>> + >>> + ret = hmm_range_register(&range, odp_mr->per_mm->mm, >>> + io_virt, io_virt + size, page_shift); >>> + if (ret) >>> + return ret; >>> + >>> if (prefetch && !downgrade && !mr->umem->writable) { >>> /* prefetch with write-access must >>> * be supported by the MR >> Isn't there a mistake in the calculation of the variable size? Itis >> first set to the size of the page fault range, but then you add the >> virtual address, so I guess it is actually the range end. Then you pass >> io_virt + size to hmm_range_register. Doesn't it double the size of the >> range > > No i think it is correct, bcnt is the byte count we are ask to fault, > we align that on the maximum size the current mr covers (min_t above) > then we align with the page size so that fault address is page align. There are three lines that update size above, not two: >>> size = min_t(size_t, bcnt, ib_umem_end(&odp->umem) - io_virt); >>> + size += (io_virt & (~page_mask)); >>> + size = ALIGN(size, 1UL << page_shift); As you said, the first one aligns to the end of the MR, the third one aligns to page size, but the middle one seems to add the start. > hmm_range_register() takes start address and end address which is the > start address + size. > > off is the offset ie the number of extra byte we are faulting to align > start on page size. If there is a bug this might be: > off += (size & (~page_mask)); off is used only to calculate the number of bytes mapped in order to proceed with the next MR in the page fault with an updated byte count. I think it should be off = start & ~page_mask; meaning it shouldn't also add in (end & ~page_mask). Regards, Haggai
On Tue, Feb 12, 2019 at 11:11:24AM -0500, Jerome Glisse wrote: > This is what serialize programming the hw and any concurrent CPU page > table invalidation. This is also one of the thing i want to improve > long term as mlx5_ib_update_xlt() can do memory allocation and i would > like to avoid that ie make mlx5_ib_update_xlt() and its sub-functions > as small and to the points as possible so that they could only fail if > the hardware is in bad state not because of memory allocation issues. How can the translation table memory consumption be dynamic (ie use tables sized huge pages until the OS breaks into 4k pages) if the tables are pre-allocated? > > > > > + > > > +static uint64_t odp_hmm_flags[HMM_PFN_FLAG_MAX] = { > > > + ODP_READ_BIT, /* HMM_PFN_VALID */ > > > + ODP_WRITE_BIT, /* HMM_PFN_WRITE */ > > > + ODP_DEVICE_BIT, /* HMM_PFN_DEVICE_PRIVATE */ > > It seems that the mlx5_ib code in this patch currently ignores the > > ODP_DEVICE_BIT (e.g., in umem_dma_to_mtt). Is that okay? Or is it > > handled implicitly by the HMM_PFN_SPECIAL case? > > This is because HMM except a bit for device memory as same API is > use for GPU which have device memory. I can add a comment explaining > that it is not use for ODP but there just to comply with HMM API. > > > > > > @@ -327,9 +287,10 @@ void put_per_mm(struct ib_umem_odp *umem_odp) > > > up_write(&per_mm->umem_rwsem); > > > > > > WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root)); > > > - mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm); > > > + hmm_mirror_unregister(&per_mm->mirror); > > > put_pid(per_mm->tgid); > > > - mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm); > > > + > > > + kfree(per_mm); > > > } > > Previously the per_mm struct was released through call srcu, but now it > > is released immediately. Is it safe? I saw that hmm_mirror_unregister > > calls mmu_notifier_unregister_no_release, so I don't understand what > > prevents concurrently running invalidations from accessing the released > > per_mm struct. > > Yes it is safe, the hmm struct has its own refcount and mirror holds a > reference on it, the mm struct itself has a reference on the mm > struct. The issue here is that that hmm_mirror_unregister() must be a strong fence that guarentees no callback is running or will run after return. mmu_notifier_unregister did not provide that. I think I saw locking in hmm that was doing this.. Jason
On Wed, Feb 20, 2019 at 10:20:27PM +0000, Jason Gunthorpe wrote: > On Tue, Feb 12, 2019 at 11:11:24AM -0500, Jerome Glisse wrote: > > This is what serialize programming the hw and any concurrent CPU page > > table invalidation. This is also one of the thing i want to improve > > long term as mlx5_ib_update_xlt() can do memory allocation and i would > > like to avoid that ie make mlx5_ib_update_xlt() and its sub-functions > > as small and to the points as possible so that they could only fail if > > the hardware is in bad state not because of memory allocation issues. > > How can the translation table memory consumption be dynamic (ie use > tables sized huge pages until the OS breaks into 4k pages) if the > tables are pre-allocated? The idea is to have HMM handle DMA mapping so from the DMA mapping page table (wether you have an IOMMU or not) you can build the device page table (leveraging contiguous DMA address into huge dma page for the hardware). This can happen before calling mlx5_ib_update_xlt(). Then mlx5_ib_update_xlt() would only need to program the hardware. > > > > > > > + > > > > +static uint64_t odp_hmm_flags[HMM_PFN_FLAG_MAX] = { > > > > + ODP_READ_BIT, /* HMM_PFN_VALID */ > > > > + ODP_WRITE_BIT, /* HMM_PFN_WRITE */ > > > > + ODP_DEVICE_BIT, /* HMM_PFN_DEVICE_PRIVATE */ > > > It seems that the mlx5_ib code in this patch currently ignores the > > > ODP_DEVICE_BIT (e.g., in umem_dma_to_mtt). Is that okay? Or is it > > > handled implicitly by the HMM_PFN_SPECIAL case? > > > > This is because HMM except a bit for device memory as same API is > > use for GPU which have device memory. I can add a comment explaining > > that it is not use for ODP but there just to comply with HMM API. > > > > > > > > > @@ -327,9 +287,10 @@ void put_per_mm(struct ib_umem_odp *umem_odp) > > > > up_write(&per_mm->umem_rwsem); > > > > > > > > WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root)); > > > > - mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm); > > > > + hmm_mirror_unregister(&per_mm->mirror); > > > > put_pid(per_mm->tgid); > > > > - mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm); > > > > + > > > > + kfree(per_mm); > > > > } > > > Previously the per_mm struct was released through call srcu, but now it > > > is released immediately. Is it safe? I saw that hmm_mirror_unregister > > > calls mmu_notifier_unregister_no_release, so I don't understand what > > > prevents concurrently running invalidations from accessing the released > > > per_mm struct. > > > > Yes it is safe, the hmm struct has its own refcount and mirror holds a > > reference on it, the mm struct itself has a reference on the mm > > struct. > > The issue here is that that hmm_mirror_unregister() must be a strong > fence that guarentees no callback is running or will run after > return. mmu_notifier_unregister did not provide that. > > I think I saw locking in hmm that was doing this.. So pattern is: hmm_mirror_register(mirror); // Safe for driver to call within HMM with mirror no matter what hmm_mirror_unregister(mirror) // Driver must no stop calling within HMM, it would be a use after // free scenario Cheers, Jérôme
On Wed, Feb 20, 2019 at 05:29:24PM -0500, Jerome Glisse wrote: > > > > > > Yes it is safe, the hmm struct has its own refcount and mirror holds a > > > reference on it, the mm struct itself has a reference on the mm > > > struct. > > > > The issue here is that that hmm_mirror_unregister() must be a strong > > fence that guarentees no callback is running or will run after > > return. mmu_notifier_unregister did not provide that. > > > > I think I saw locking in hmm that was doing this.. > > So pattern is: > hmm_mirror_register(mirror); > > // Safe for driver to call within HMM with mirror no matter what > > hmm_mirror_unregister(mirror) > > // Driver must no stop calling within HMM, it would be a use after > // free scenario This statement is the opposite direction I want to know that HMM doesn't allow any driver callbacks to be running after unregister - because I am going to kfree mirror and other memory touched by the driver callbacks. Jason
On Thu, Feb 21, 2019 at 03:59:37PM -0700, Jason Gunthorpe wrote: > On Wed, Feb 20, 2019 at 05:29:24PM -0500, Jerome Glisse wrote: > > > > > > > > Yes it is safe, the hmm struct has its own refcount and mirror holds a > > > > reference on it, the mm struct itself has a reference on the mm > > > > struct. > > > > > > The issue here is that that hmm_mirror_unregister() must be a strong > > > fence that guarentees no callback is running or will run after > > > return. mmu_notifier_unregister did not provide that. > > > > > > I think I saw locking in hmm that was doing this.. > > > > So pattern is: > > hmm_mirror_register(mirror); > > > > // Safe for driver to call within HMM with mirror no matter what > > > > hmm_mirror_unregister(mirror) > > > > // Driver must no stop calling within HMM, it would be a use after > > // free scenario > > This statement is the opposite direction > > I want to know that HMM doesn't allow any driver callbacks to be > running after unregister - because I am going to kfree mirror and > other memory touched by the driver callbacks. Sorry i miss-understood your question. Yes after hmm_mirror_unregister() you will no longer get a callback from HMM ie it is safe for you to free any data structure associated with HMM. Cheers, Jérôme
diff --git a/drivers/infiniband/core/umem_odp.c b/drivers/infiniband/core/umem_odp.c index a4ec43093cb3..8afa707f1d9a 100644 --- a/drivers/infiniband/core/umem_odp.c +++ b/drivers/infiniband/core/umem_odp.c @@ -45,6 +45,20 @@ #include <rdma/ib_umem.h> #include <rdma/ib_umem_odp.h> + +static uint64_t odp_hmm_flags[HMM_PFN_FLAG_MAX] = { + ODP_READ_BIT, /* HMM_PFN_VALID */ + ODP_WRITE_BIT, /* HMM_PFN_WRITE */ + ODP_DEVICE_BIT, /* HMM_PFN_DEVICE_PRIVATE */ +}; + +static uint64_t odp_hmm_values[HMM_PFN_VALUE_MAX] = { + -1UL, /* HMM_PFN_ERROR */ + 0UL, /* HMM_PFN_NONE */ + -2UL, /* HMM_PFN_SPECIAL */ +}; + + /* * The ib_umem list keeps track of memory regions for which the HW * device request to receive notification when the related memory @@ -77,57 +91,25 @@ static u64 node_last(struct umem_odp_node *n) INTERVAL_TREE_DEFINE(struct umem_odp_node, rb, u64, __subtree_last, node_start, node_last, static, rbt_ib_umem) -static void ib_umem_notifier_start_account(struct ib_umem_odp *umem_odp) -{ - mutex_lock(&umem_odp->umem_mutex); - if (umem_odp->notifiers_count++ == 0) - /* - * Initialize the completion object for waiting on - * notifiers. Since notifier_count is zero, no one should be - * waiting right now. - */ - reinit_completion(&umem_odp->notifier_completion); - mutex_unlock(&umem_odp->umem_mutex); -} - -static void ib_umem_notifier_end_account(struct ib_umem_odp *umem_odp) -{ - mutex_lock(&umem_odp->umem_mutex); - /* - * This sequence increase will notify the QP page fault that the page - * that is going to be mapped in the spte could have been freed. - */ - ++umem_odp->notifiers_seq; - if (--umem_odp->notifiers_count == 0) - complete_all(&umem_odp->notifier_completion); - mutex_unlock(&umem_odp->umem_mutex); -} - static int ib_umem_notifier_release_trampoline(struct ib_umem_odp *umem_odp, u64 start, u64 end, void *cookie) { struct ib_umem *umem = &umem_odp->umem; - /* - * Increase the number of notifiers running, to - * prevent any further fault handling on this MR. - */ - ib_umem_notifier_start_account(umem_odp); umem_odp->dying = 1; /* Make sure that the fact the umem is dying is out before we release * all pending page faults. */ smp_wmb(); - complete_all(&umem_odp->notifier_completion); umem->context->invalidate_range(umem_odp, ib_umem_start(umem), ib_umem_end(umem)); return 0; } -static void ib_umem_notifier_release(struct mmu_notifier *mn, - struct mm_struct *mm) +static void ib_umem_notifier_release(struct hmm_mirror *mirror) { - struct ib_ucontext_per_mm *per_mm = - container_of(mn, struct ib_ucontext_per_mm, mn); + struct ib_ucontext_per_mm *per_mm; + + per_mm = container_of(mirror, struct ib_ucontext_per_mm, mirror); down_read(&per_mm->umem_rwsem); if (per_mm->active) @@ -135,21 +117,24 @@ static void ib_umem_notifier_release(struct mmu_notifier *mn, &per_mm->umem_tree, 0, ULLONG_MAX, ib_umem_notifier_release_trampoline, true, NULL); up_read(&per_mm->umem_rwsem); + + per_mm->mm = NULL; } -static int invalidate_range_start_trampoline(struct ib_umem_odp *item, - u64 start, u64 end, void *cookie) +static int invalidate_range_trampoline(struct ib_umem_odp *item, + u64 start, u64 end, void *cookie) { - ib_umem_notifier_start_account(item); item->umem.context->invalidate_range(item, start, end); return 0; } -static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn, - const struct mmu_notifier_range *range) +static int ib_sync_cpu_device_pagetables(struct hmm_mirror *mirror, + const struct hmm_update *range) { - struct ib_ucontext_per_mm *per_mm = - container_of(mn, struct ib_ucontext_per_mm, mn); + struct ib_ucontext_per_mm *per_mm; + int ret; + + per_mm = container_of(mirror, struct ib_ucontext_per_mm, mirror); if (range->blockable) down_read(&per_mm->umem_rwsem); @@ -166,38 +151,17 @@ static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn, return 0; } - return rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start, + ret = rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start, range->end, - invalidate_range_start_trampoline, + invalidate_range_trampoline, range->blockable, NULL); -} - -static int invalidate_range_end_trampoline(struct ib_umem_odp *item, u64 start, - u64 end, void *cookie) -{ - ib_umem_notifier_end_account(item); - return 0; -} - -static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn, - const struct mmu_notifier_range *range) -{ - struct ib_ucontext_per_mm *per_mm = - container_of(mn, struct ib_ucontext_per_mm, mn); - - if (unlikely(!per_mm->active)) - return; - - rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start, - range->end, - invalidate_range_end_trampoline, true, NULL); up_read(&per_mm->umem_rwsem); + return ret; } -static const struct mmu_notifier_ops ib_umem_notifiers = { +static const struct hmm_mirror_ops ib_umem_notifiers = { .release = ib_umem_notifier_release, - .invalidate_range_start = ib_umem_notifier_invalidate_range_start, - .invalidate_range_end = ib_umem_notifier_invalidate_range_end, + .sync_cpu_device_pagetables = ib_sync_cpu_device_pagetables, }; static void add_umem_to_per_mm(struct ib_umem_odp *umem_odp) @@ -221,7 +185,6 @@ static void remove_umem_from_per_mm(struct ib_umem_odp *umem_odp) if (likely(ib_umem_start(umem) != ib_umem_end(umem))) rbt_ib_umem_remove(&umem_odp->interval_tree, &per_mm->umem_tree); - complete_all(&umem_odp->notifier_completion); up_write(&per_mm->umem_rwsem); } @@ -248,11 +211,13 @@ static struct ib_ucontext_per_mm *alloc_per_mm(struct ib_ucontext *ctx, WARN_ON(mm != current->mm); - per_mm->mn.ops = &ib_umem_notifiers; - ret = mmu_notifier_register(&per_mm->mn, per_mm->mm); + per_mm->mirror.ops = &ib_umem_notifiers; + down_write(&mm->mmap_sem); + ret = hmm_mirror_register(&per_mm->mirror, per_mm->mm); + up_write(&mm->mmap_sem); if (ret) { dev_err(&ctx->device->dev, - "Failed to register mmu_notifier %d\n", ret); + "Failed to register HMM mirror %d\n", ret); goto out_pid; } @@ -294,11 +259,6 @@ static int get_per_mm(struct ib_umem_odp *umem_odp) return 0; } -static void free_per_mm(struct rcu_head *rcu) -{ - kfree(container_of(rcu, struct ib_ucontext_per_mm, rcu)); -} - void put_per_mm(struct ib_umem_odp *umem_odp) { struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm; @@ -327,9 +287,10 @@ void put_per_mm(struct ib_umem_odp *umem_odp) up_write(&per_mm->umem_rwsem); WARN_ON(!RB_EMPTY_ROOT(&per_mm->umem_tree.rb_root)); - mmu_notifier_unregister_no_release(&per_mm->mn, per_mm->mm); + hmm_mirror_unregister(&per_mm->mirror); put_pid(per_mm->tgid); - mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm); + + kfree(per_mm); } struct ib_umem_odp *ib_alloc_odp_umem(struct ib_ucontext_per_mm *per_mm, @@ -354,11 +315,9 @@ struct ib_umem_odp *ib_alloc_odp_umem(struct ib_ucontext_per_mm *per_mm, odp_data->per_mm = per_mm; mutex_init(&odp_data->umem_mutex); - init_completion(&odp_data->notifier_completion); - odp_data->page_list = - vzalloc(array_size(pages, sizeof(*odp_data->page_list))); - if (!odp_data->page_list) { + odp_data->pfns = vzalloc(array_size(pages, sizeof(*odp_data->pfns))); + if (!odp_data->pfns) { ret = -ENOMEM; goto out_odp_data; } @@ -367,7 +326,7 @@ struct ib_umem_odp *ib_alloc_odp_umem(struct ib_ucontext_per_mm *per_mm, vzalloc(array_size(pages, sizeof(*odp_data->dma_list))); if (!odp_data->dma_list) { ret = -ENOMEM; - goto out_page_list; + goto out_pfns; } /* @@ -381,8 +340,8 @@ struct ib_umem_odp *ib_alloc_odp_umem(struct ib_ucontext_per_mm *per_mm, return odp_data; -out_page_list: - vfree(odp_data->page_list); +out_pfns: + vfree(odp_data->pfns); out_odp_data: kfree(odp_data); return ERR_PTR(ret); @@ -419,13 +378,11 @@ int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access) mutex_init(&umem_odp->umem_mutex); - init_completion(&umem_odp->notifier_completion); - if (ib_umem_num_pages(umem)) { - umem_odp->page_list = - vzalloc(array_size(sizeof(*umem_odp->page_list), + umem_odp->pfns = + vzalloc(array_size(sizeof(*umem_odp->pfns), ib_umem_num_pages(umem))); - if (!umem_odp->page_list) + if (!umem_odp->pfns) return -ENOMEM; umem_odp->dma_list = @@ -433,7 +390,7 @@ int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access) ib_umem_num_pages(umem))); if (!umem_odp->dma_list) { ret_val = -ENOMEM; - goto out_page_list; + goto out_pfns; } } @@ -446,8 +403,8 @@ int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access) out_dma_list: vfree(umem_odp->dma_list); -out_page_list: - vfree(umem_odp->page_list); +out_pfns: + vfree(umem_odp->pfns); return ret_val; } @@ -467,291 +424,113 @@ void ib_umem_odp_release(struct ib_umem_odp *umem_odp) remove_umem_from_per_mm(umem_odp); put_per_mm(umem_odp); vfree(umem_odp->dma_list); - vfree(umem_odp->page_list); + vfree(umem_odp->pfns); } -/* - * Map for DMA and insert a single page into the on-demand paging page tables. - * - * @umem: the umem to insert the page to. - * @page_index: index in the umem to add the page to. - * @page: the page struct to map and add. - * @access_mask: access permissions needed for this page. - * @current_seq: sequence number for synchronization with invalidations. - * the sequence number is taken from - * umem_odp->notifiers_seq. - * - * The function returns -EFAULT if the DMA mapping operation fails. It returns - * -EAGAIN if a concurrent invalidation prevents us from updating the page. +/** + * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR. + * @umem_odp: the umem to map and pin + * @range: range of virtual address to be mapped to the device + * Returns: -EINVAL some invalid arguments, -EAGAIN need to try again, -ENOENT + * if process is being terminated, number of pages mapped otherwise. * - * The page is released via put_page even if the operation failed. For - * on-demand pinning, the page is released whenever it isn't stored in the - * umem. + * Map to device a range of virtual address passed in the argument. The DMA + * addresses are in umem_odp->dma_list and the corresponding page informations + * in umem_odp->pfns. */ -static int ib_umem_odp_map_dma_single_page( - struct ib_umem_odp *umem_odp, - int page_index, - struct page *page, - u64 access_mask, - unsigned long current_seq) +long ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, + struct hmm_range *range) { + struct device *device = umem_odp->umem.context->device->dma_device; + struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm; struct ib_umem *umem = &umem_odp->umem; - struct ib_device *dev = umem->context->device; - dma_addr_t dma_addr; - int stored_page = 0; - int remove_existing_mapping = 0; - int ret = 0; - - /* - * Note: we avoid writing if seq is different from the initial seq, to - * handle case of a racing notifier. This check also allows us to bail - * early if we have a notifier running in parallel with us. - */ - if (ib_umem_mmu_notifier_retry(umem_odp, current_seq)) { - ret = -EAGAIN; - goto out; - } - if (!(umem_odp->dma_list[page_index])) { - dma_addr = ib_dma_map_page(dev, - page, - 0, BIT(umem->page_shift), - DMA_BIDIRECTIONAL); - if (ib_dma_mapping_error(dev, dma_addr)) { - ret = -EFAULT; - goto out; - } - umem_odp->dma_list[page_index] = dma_addr | access_mask; - umem_odp->page_list[page_index] = page; - umem->npages++; - stored_page = 1; - } else if (umem_odp->page_list[page_index] == page) { - umem_odp->dma_list[page_index] |= access_mask; - } else { - pr_err("error: got different pages in IB device and from get_user_pages. IB device page: %p, gup page: %p\n", - umem_odp->page_list[page_index], page); - /* Better remove the mapping now, to prevent any further - * damage. */ - remove_existing_mapping = 1; - } + struct mm_struct *mm = per_mm->mm; + unsigned long idx, npages; + long ret; -out: - /* On Demand Paging - avoid pinning the page */ - if (umem->context->invalidate_range || !stored_page) - put_page(page); - - if (remove_existing_mapping && umem->context->invalidate_range) { - ib_umem_notifier_start_account(umem_odp); - umem->context->invalidate_range( - umem_odp, - ib_umem_start(umem) + (page_index << umem->page_shift), - ib_umem_start(umem) + - ((page_index + 1) << umem->page_shift)); - ib_umem_notifier_end_account(umem_odp); - ret = -EAGAIN; - } + if (mm == NULL) + return -ENOENT; - return ret; -} + /* Only drivers with invalidate support can use this function. */ + if (!umem->context->invalidate_range) + return -EINVAL; -/** - * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR. - * - * Pins the range of pages passed in the argument, and maps them to - * DMA addresses. The DMA addresses of the mapped pages is updated in - * umem_odp->dma_list. - * - * Returns the number of pages mapped in success, negative error code - * for failure. - * An -EAGAIN error code is returned when a concurrent mmu notifier prevents - * the function from completing its task. - * An -ENOENT error code indicates that userspace process is being terminated - * and mm was already destroyed. - * @umem_odp: the umem to map and pin - * @user_virt: the address from which we need to map. - * @bcnt: the minimal number of bytes to pin and map. The mapping might be - * bigger due to alignment, and may also be smaller in case of an error - * pinning or mapping a page. The actual pages mapped is returned in - * the return value. - * @access_mask: bit mask of the requested access permissions for the given - * range. - * @current_seq: the MMU notifiers sequance value for synchronization with - * invalidations. the sequance number is read from - * umem_odp->notifiers_seq before calling this function - */ -int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt, - u64 bcnt, u64 access_mask, - unsigned long current_seq) -{ - struct ib_umem *umem = &umem_odp->umem; - struct task_struct *owning_process = NULL; - struct mm_struct *owning_mm = umem_odp->umem.owning_mm; - struct page **local_page_list = NULL; - u64 page_mask, off; - int j, k, ret = 0, start_idx, npages = 0, page_shift; - unsigned int flags = 0; - phys_addr_t p = 0; - - if (access_mask == 0) + /* Sanity checks. */ + if (range->default_flags == 0) return -EINVAL; - if (user_virt < ib_umem_start(umem) || - user_virt + bcnt > ib_umem_end(umem)) - return -EFAULT; + if (range->start < ib_umem_start(umem) || + range->end > ib_umem_end(umem)) + return -EINVAL; - local_page_list = (struct page **)__get_free_page(GFP_KERNEL); - if (!local_page_list) - return -ENOMEM; + idx = (range->start - ib_umem_start(umem)) >> umem->page_shift; + range->pfns = &umem_odp->pfns[idx]; + range->pfn_shift = ODP_FLAGS_BITS; + range->values = odp_hmm_values; + range->flags = odp_hmm_flags; - page_shift = umem->page_shift; - page_mask = ~(BIT(page_shift) - 1); - off = user_virt & (~page_mask); - user_virt = user_virt & page_mask; - bcnt += off; /* Charge for the first page offset as well. */ + ret = hmm_mirror_mm_down_read(&per_mm->mirror); + if (ret) + return ret; + mutex_lock(&umem_odp->umem_mutex); + ret = hmm_range_dma_map(range, device, + &umem_odp->dma_list[idx], true); + mutex_unlock(&umem_odp->umem_mutex); + npages = ret; /* - * owning_process is allowed to be NULL, this means somehow the mm is - * existing beyond the lifetime of the originating process.. Presumably - * mmget_not_zero will fail in this case. + * The mmap_sem have been drop if hmm_vma_fault_and_dma_map() returned + * with -EAGAIN. In which case we need to retry as -EBUSY but we also + * need to take the mmap_sem again. */ - owning_process = get_pid_task(umem_odp->per_mm->tgid, PIDTYPE_PID); - if (WARN_ON(!mmget_not_zero(umem_odp->umem.owning_mm))) { - ret = -EINVAL; - goto out_put_task; - } - - if (access_mask & ODP_WRITE_ALLOWED_BIT) - flags |= FOLL_WRITE; - - start_idx = (user_virt - ib_umem_start(umem)) >> page_shift; - k = start_idx; - - while (bcnt > 0) { - const size_t gup_num_pages = min_t(size_t, - (bcnt + BIT(page_shift) - 1) >> page_shift, - PAGE_SIZE / sizeof(struct page *)); - - down_read(&owning_mm->mmap_sem); - /* - * Note: this might result in redundent page getting. We can - * avoid this by checking dma_list to be 0 before calling - * get_user_pages. However, this make the code much more - * complex (and doesn't gain us much performance in most use - * cases). - */ - npages = get_user_pages_remote(owning_process, owning_mm, - user_virt, gup_num_pages, - flags, local_page_list, NULL, NULL); - up_read(&owning_mm->mmap_sem); - - if (npages < 0) { - if (npages != -EAGAIN) - pr_warn("fail to get %zu user pages with error %d\n", gup_num_pages, npages); - else - pr_debug("fail to get %zu user pages with error %d\n", gup_num_pages, npages); - break; - } - - bcnt -= min_t(size_t, npages << PAGE_SHIFT, bcnt); - mutex_lock(&umem_odp->umem_mutex); - for (j = 0; j < npages; j++, user_virt += PAGE_SIZE) { - if (user_virt & ~page_mask) { - p += PAGE_SIZE; - if (page_to_phys(local_page_list[j]) != p) { - ret = -EFAULT; - break; - } - put_page(local_page_list[j]); - continue; - } - - ret = ib_umem_odp_map_dma_single_page( - umem_odp, k, local_page_list[j], - access_mask, current_seq); - if (ret < 0) { - if (ret != -EAGAIN) - pr_warn("ib_umem_odp_map_dma_single_page failed with error %d\n", ret); - else - pr_debug("ib_umem_odp_map_dma_single_page failed with error %d\n", ret); - break; - } - - p = page_to_phys(local_page_list[j]); - k++; - } - mutex_unlock(&umem_odp->umem_mutex); + if (ret != -EAGAIN) + hmm_mirror_mm_up_read(&per_mm->mirror); - if (ret < 0) { - /* Release left over pages when handling errors. */ - for (++j; j < npages; ++j) - put_page(local_page_list[j]); - break; - } - } - - if (ret >= 0) { - if (npages < 0 && k == start_idx) - ret = npages; - else - ret = k - start_idx; + if (ret <= 0) { + /* Convert -EBUSY to -EAGAIN and 0 to -EAGAIN */ + ret = ret == -EBUSY ? -EAGAIN : ret; + return ret ? ret : -EAGAIN; } - mmput(owning_mm); -out_put_task: - if (owning_process) - put_task_struct(owning_process); - free_page((unsigned long)local_page_list); - return ret; + umem->npages += npages; + return npages; } EXPORT_SYMBOL(ib_umem_odp_map_dma_pages); -void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt, - u64 bound) +void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, + u64 virt, u64 bound) { + struct device *device = umem_odp->umem.context->device->dma_device; struct ib_umem *umem = &umem_odp->umem; - int idx; - u64 addr; - struct ib_device *dev = umem->context->device; + unsigned long idx, page_mask; + struct hmm_range range; + long ret; + + if (!umem->npages) + return; + + bound = ALIGN(bound, 1UL << umem->page_shift); + page_mask = ~(BIT(umem->page_shift) - 1); + virt &= page_mask; virt = max_t(u64, virt, ib_umem_start(umem)); bound = min_t(u64, bound, ib_umem_end(umem)); - /* Note that during the run of this function, the - * notifiers_count of the MR is > 0, preventing any racing - * faults from completion. We might be racing with other - * invalidations, so we must make sure we free each page only - * once. */ + + idx = ((unsigned long)virt - ib_umem_start(umem)) >> PAGE_SHIFT; + + range.page_shift = umem->page_shift; + range.pfns = &umem_odp->pfns[idx]; + range.pfn_shift = ODP_FLAGS_BITS; + range.values = odp_hmm_values; + range.flags = odp_hmm_flags; + range.start = virt; + range.end = bound; + mutex_lock(&umem_odp->umem_mutex); - for (addr = virt; addr < bound; addr += BIT(umem->page_shift)) { - idx = (addr - ib_umem_start(umem)) >> umem->page_shift; - if (umem_odp->page_list[idx]) { - struct page *page = umem_odp->page_list[idx]; - dma_addr_t dma = umem_odp->dma_list[idx]; - dma_addr_t dma_addr = dma & ODP_DMA_ADDR_MASK; - - WARN_ON(!dma_addr); - - ib_dma_unmap_page(dev, dma_addr, PAGE_SIZE, - DMA_BIDIRECTIONAL); - if (dma & ODP_WRITE_ALLOWED_BIT) { - struct page *head_page = compound_head(page); - /* - * set_page_dirty prefers being called with - * the page lock. However, MMU notifiers are - * called sometimes with and sometimes without - * the lock. We rely on the umem_mutex instead - * to prevent other mmu notifiers from - * continuing and allowing the page mapping to - * be removed. - */ - set_page_dirty(head_page); - } - /* on demand pinning support */ - if (!umem->context->invalidate_range) - put_page(page); - umem_odp->page_list[idx] = NULL; - umem_odp->dma_list[idx] = 0; - umem->npages--; - } - } + ret = hmm_range_dma_unmap(&range, NULL, device, + &umem_odp->dma_list[idx], true); + if (ret > 0) + umem->npages -= ret; mutex_unlock(&umem_odp->umem_mutex); } EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages); diff --git a/drivers/infiniband/hw/mlx5/mem.c b/drivers/infiniband/hw/mlx5/mem.c index 549234988bb4..b9621d4d80b1 100644 --- a/drivers/infiniband/hw/mlx5/mem.c +++ b/drivers/infiniband/hw/mlx5/mem.c @@ -112,16 +112,16 @@ void mlx5_ib_cont_pages(struct ib_umem *umem, u64 addr, } #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING -static u64 umem_dma_to_mtt(dma_addr_t umem_dma) +static u64 umem_dma_to_mtt(struct ib_umem_odp *odp, size_t idx) { - u64 mtt_entry = umem_dma & ODP_DMA_ADDR_MASK; + u64 mtt_entry = odp->dma_list[idx]; - if (umem_dma & ODP_READ_ALLOWED_BIT) + if (odp->pfns[idx] & ODP_READ_BIT) mtt_entry |= MLX5_IB_MTT_READ; - if (umem_dma & ODP_WRITE_ALLOWED_BIT) + if (odp->pfns[idx] & ODP_WRITE_BIT) mtt_entry |= MLX5_IB_MTT_WRITE; - return mtt_entry; + return cpu_to_be64(mtt_entry); } #endif @@ -153,15 +153,13 @@ void __mlx5_ib_populate_pas(struct mlx5_ib_dev *dev, struct ib_umem *umem, int entry; #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING if (umem->is_odp) { + struct ib_umem_odp *odp = to_ib_umem_odp(umem); + WARN_ON(shift != 0); WARN_ON(access_flags != (MLX5_IB_MTT_READ | MLX5_IB_MTT_WRITE)); - for (i = 0; i < num_pages; ++i) { - dma_addr_t pa = - to_ib_umem_odp(umem)->dma_list[offset + i]; - - pas[i] = cpu_to_be64(umem_dma_to_mtt(pa)); - } + for (i = 0; i < num_pages; ++i) + pas[i] = umem_dma_to_mtt(odp, offset + i); return; } #endif diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index fd6ea1f75085..e764855e4044 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -1650,7 +1650,7 @@ static void dereg_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr) /* Wait for all running page-fault handlers to finish. */ synchronize_srcu(&dev->mr_srcu); /* Destroy all page mappings */ - if (umem_odp->page_list) + if (umem_odp->pfns) mlx5_ib_invalidate_range(umem_odp, ib_umem_start(umem), ib_umem_end(umem)); else diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index 01e0f6200631..9d551b044017 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -257,8 +257,7 @@ void mlx5_ib_invalidate_range(struct ib_umem_odp *umem_odp, unsigned long start, * estimate the cost of another UMR vs. the cost of bigger * UMR. */ - if (umem_odp->dma_list[idx] & - (ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) { + if (umem_odp->pfns[idx] & ODP_READ_BIT) { if (!in_block) { blk_start_idx = idx; in_block = 1; @@ -555,17 +554,18 @@ static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt, u32 *bytes_mapped, u32 flags) { - int npages = 0, current_seq, page_shift, ret, np; - bool implicit = false; struct ib_umem_odp *odp_mr = to_ib_umem_odp(mr->umem); bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE; bool prefetch = flags & MLX5_PF_FLAGS_PREFETCH; - u64 access_mask = ODP_READ_ALLOWED_BIT; + unsigned long npages = 0, page_shift, np, off; u64 start_idx, page_mask; struct ib_umem_odp *odp; + struct hmm_range range; + bool implicit = false; size_t size; + long ret; - if (!odp_mr->page_list) { + if (!odp_mr->pfns) { odp = implicit_mr_get_data(mr, io_virt, bcnt); if (IS_ERR(odp)) @@ -578,11 +578,27 @@ static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr, next_mr: size = min_t(size_t, bcnt, ib_umem_end(&odp->umem) - io_virt); - page_shift = mr->umem->page_shift; page_mask = ~(BIT(page_shift) - 1); + off = (io_virt & (~page_mask)); + size += (io_virt & (~page_mask)); + io_virt = io_virt & page_mask; + off += (size & (~page_mask)); + size = ALIGN(size, 1UL << page_shift); + + if (io_virt < ib_umem_start(&odp->umem)) + return -EINVAL; + start_idx = (io_virt - (mr->mmkey.iova & page_mask)) >> page_shift; + if (odp_mr->per_mm == NULL || odp_mr->per_mm->mm == NULL) + return -ENOENT; + + ret = hmm_range_register(&range, odp_mr->per_mm->mm, + io_virt, io_virt + size, page_shift); + if (ret) + return ret; + if (prefetch && !downgrade && !mr->umem->writable) { /* prefetch with write-access must * be supported by the MR @@ -591,55 +607,48 @@ static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr, goto out; } + range.default_flags = ODP_READ_BIT; if (mr->umem->writable && !downgrade) - access_mask |= ODP_WRITE_ALLOWED_BIT; - - current_seq = READ_ONCE(odp->notifiers_seq); - /* - * Ensure the sequence number is valid for some time before we call - * gup. - */ - smp_rmb(); - - ret = ib_umem_odp_map_dma_pages(to_ib_umem_odp(mr->umem), io_virt, size, - access_mask, current_seq); + range.default_flags |= ODP_WRITE_BIT; + ret = ib_umem_odp_map_dma_pages(to_ib_umem_odp(mr->umem), &range); if (ret < 0) - goto out; + goto again; np = ret; mutex_lock(&odp->umem_mutex); - if (!ib_umem_mmu_notifier_retry(to_ib_umem_odp(mr->umem), - current_seq)) { + if (hmm_range_valid(&range)) { /* * No need to check whether the MTTs really belong to - * this MR, since ib_umem_odp_map_dma_pages already + * this MR, since ib_umem_odp_map_dma_pages() already * checks this. */ ret = mlx5_ib_update_xlt(mr, start_idx, np, page_shift, MLX5_IB_UPD_XLT_ATOMIC); - } else { + } else ret = -EAGAIN; - } mutex_unlock(&odp->umem_mutex); if (ret < 0) { - if (ret != -EAGAIN) + if (ret != -EAGAIN) { mlx5_ib_err(dev, "Failed to update mkey page tables\n"); - goto out; + goto out; + } + goto again; } if (bytes_mapped) { - u32 new_mappings = (np << page_shift) - - (io_virt - round_down(io_virt, 1 << page_shift)); + long new_mappings = (np << page_shift) - off; + new_mappings = new_mappings < 0 ? 0 : new_mappings; *bytes_mapped += min_t(u32, new_mappings, size); } npages += np << (page_shift - PAGE_SHIFT); + hmm_range_unregister(&range); bcnt -= size; - if (unlikely(bcnt)) { + if (unlikely(bcnt > 0)) { struct ib_umem_odp *next; io_virt += size; @@ -656,24 +665,18 @@ static int pagefault_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr, return npages; -out: - if (ret == -EAGAIN) { - if (implicit || !odp->dying) { - unsigned long timeout = - msecs_to_jiffies(MMU_NOTIFIER_TIMEOUT); - - if (!wait_for_completion_timeout( - &odp->notifier_completion, - timeout)) { - mlx5_ib_warn(dev, "timeout waiting for mmu notifier. seq %d against %d. notifiers_count=%d\n", - current_seq, odp->notifiers_seq, odp->notifiers_count); - } - } else { - /* The MR is being killed, kill the QP as well. */ - ret = -EFAULT; - } - } +again: + if (ret != -EAGAIN) + goto out; + /* Check if the MR is being killed, kill the QP as well. */ + if (!implicit || odp->dying) + ret = -EFAULT; + else if (!hmm_range_wait_until_valid(&range, MMU_NOTIFIER_TIMEOUT)) + mlx5_ib_warn(dev, "timeout waiting for mmu notifier.\n"); + +out: + hmm_range_unregister(&range); return ret; } diff --git a/include/rdma/ib_umem_odp.h b/include/rdma/ib_umem_odp.h index 0b1446fe2fab..1d61a616b3b6 100644 --- a/include/rdma/ib_umem_odp.h +++ b/include/rdma/ib_umem_odp.h @@ -36,6 +36,7 @@ #include <rdma/ib_umem.h> #include <rdma/ib_verbs.h> #include <linux/interval_tree.h> +#include <linux/hmm.h> struct umem_odp_node { u64 __subtree_last; @@ -47,11 +48,11 @@ struct ib_umem_odp { struct ib_ucontext_per_mm *per_mm; /* - * An array of the pages included in the on-demand paging umem. - * Indices of pages that are currently not mapped into the device will - * contain NULL. + * An array of the pages included in the on-demand paging umem. Indices + * of pages that are currently not mapped into the device will contain + * 0. */ - struct page **page_list; + uint64_t *pfns; /* * An array of the same size as page_list, with DMA addresses mapped * for pages the pages in page_list. The lower two bits designate @@ -67,13 +68,9 @@ struct ib_umem_odp { struct mutex umem_mutex; void *private; /* for the HW driver to use. */ - int notifiers_seq; - int notifiers_count; - /* Tree tracking */ struct umem_odp_node interval_tree; - struct completion notifier_completion; int dying; struct work_struct work; }; @@ -95,11 +92,10 @@ struct ib_ucontext_per_mm { /* Protects umem_tree */ struct rw_semaphore umem_rwsem; - struct mmu_notifier mn; + struct hmm_mirror mirror; unsigned int odp_mrs_count; struct list_head ucontext_list; - struct rcu_head rcu; }; int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access); @@ -107,22 +103,13 @@ struct ib_umem_odp *ib_alloc_odp_umem(struct ib_ucontext_per_mm *per_mm, unsigned long addr, size_t size); void ib_umem_odp_release(struct ib_umem_odp *umem_odp); -/* - * The lower 2 bits of the DMA address signal the R/W permissions for - * the entry. To upgrade the permissions, provide the appropriate - * bitmask to the map_dma_pages function. - * - * Be aware that upgrading a mapped address might result in change of - * the DMA address for the page. - */ -#define ODP_READ_ALLOWED_BIT (1<<0ULL) -#define ODP_WRITE_ALLOWED_BIT (1<<1ULL) - -#define ODP_DMA_ADDR_MASK (~(ODP_READ_ALLOWED_BIT | ODP_WRITE_ALLOWED_BIT)) +#define ODP_READ_BIT (1<<0ULL) +#define ODP_WRITE_BIT (1<<1ULL) +#define ODP_DEVICE_BIT (1<<2ULL) +#define ODP_FLAGS_BITS 3 -int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 start_offset, - u64 bcnt, u64 access_mask, - unsigned long current_seq); +long ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, + struct hmm_range *range); void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 start_offset, u64 bound); @@ -145,23 +132,6 @@ int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root, struct ib_umem_odp *rbt_ib_umem_lookup(struct rb_root_cached *root, u64 addr, u64 length); -static inline int ib_umem_mmu_notifier_retry(struct ib_umem_odp *umem_odp, - unsigned long mmu_seq) -{ - /* - * This code is strongly based on the KVM code from - * mmu_notifier_retry. Should be called with - * the relevant locks taken (umem_odp->umem_mutex - * and the ucontext umem_mutex semaphore locked for read). - */ - - if (unlikely(umem_odp->notifiers_count)) - return 1; - if (umem_odp->notifiers_seq != mmu_seq) - return 1; - return 0; -} - #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */ static inline int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access)