diff options
Diffstat (limited to 'drivers/infiniband/core/umem_odp.c')
| -rw-r--r-- | drivers/infiniband/core/umem_odp.c | 943 |
1 files changed, 306 insertions, 637 deletions
diff --git a/drivers/infiniband/core/umem_odp.c b/drivers/infiniband/core/umem_odp.c index a4ec43093cb3..572a91a62a7b 100644 --- a/drivers/infiniband/core/umem_odp.c +++ b/drivers/infiniband/core/umem_odp.c @@ -39,421 +39,241 @@ #include <linux/export.h> #include <linux/vmalloc.h> #include <linux/hugetlb.h> -#include <linux/interval_tree_generic.h> +#include <linux/interval_tree.h> +#include <linux/hmm.h> +#include <linux/hmm-dma.h> +#include <linux/pagemap.h> -#include <rdma/ib_verbs.h> -#include <rdma/ib_umem.h> #include <rdma/ib_umem_odp.h> -/* - * The ib_umem list keeps track of memory regions for which the HW - * device request to receive notification when the related memory - * mapping is changed. - * - * ib_umem_lock protects the list. - */ - -static u64 node_start(struct umem_odp_node *n) -{ - struct ib_umem_odp *umem_odp = - container_of(n, struct ib_umem_odp, interval_tree); - - return ib_umem_start(&umem_odp->umem); -} +#include "uverbs.h" -/* Note that the representation of the intervals in the interval tree - * considers the ending point as contained in the interval, while the - * function ib_umem_end returns the first address which is not contained - * in the umem. - */ -static u64 node_last(struct umem_odp_node *n) +static void ib_init_umem_implicit_odp(struct ib_umem_odp *umem_odp) { - struct ib_umem_odp *umem_odp = - container_of(n, struct ib_umem_odp, interval_tree); - - return ib_umem_end(&umem_odp->umem) - 1; -} - -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); + umem_odp->is_implicit_odp = 1; + umem_odp->umem.is_odp = 1; + mutex_init(&umem_odp->umem_mutex); } -static void ib_umem_notifier_end_account(struct ib_umem_odp *umem_odp) +static int ib_init_umem_odp(struct ib_umem_odp *umem_odp, + const struct mmu_interval_notifier_ops *ops) { - 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); -} + struct ib_device *dev = umem_odp->umem.ibdev; + size_t page_size = 1UL << umem_odp->page_shift; + struct hmm_dma_map *map; + unsigned long start; + unsigned long end; + size_t nr_entries; + int ret = 0; -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; + umem_odp->umem.is_odp = 1; + mutex_init(&umem_odp->umem_mutex); + start = ALIGN_DOWN(umem_odp->umem.address, page_size); + if (check_add_overflow(umem_odp->umem.address, + (unsigned long)umem_odp->umem.length, &end)) + return -EOVERFLOW; + end = ALIGN(end, page_size); + if (unlikely(end < page_size)) + return -EOVERFLOW; /* - * Increase the number of notifiers running, to - * prevent any further fault handling on this MR. + * The mmu notifier can be called within reclaim contexts and takes the + * umem_mutex. This is rare to trigger in testing, teach lockdep about + * it. */ - 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) -{ - struct ib_ucontext_per_mm *per_mm = - container_of(mn, struct ib_ucontext_per_mm, mn); - - down_read(&per_mm->umem_rwsem); - if (per_mm->active) - rbt_ib_umem_for_each_in_range( - &per_mm->umem_tree, 0, ULLONG_MAX, - ib_umem_notifier_release_trampoline, true, NULL); - up_read(&per_mm->umem_rwsem); -} - -static int invalidate_range_start_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) -{ - struct ib_ucontext_per_mm *per_mm = - container_of(mn, struct ib_ucontext_per_mm, mn); - - if (range->blockable) - down_read(&per_mm->umem_rwsem); - else if (!down_read_trylock(&per_mm->umem_rwsem)) - return -EAGAIN; - - if (!per_mm->active) { - up_read(&per_mm->umem_rwsem); - /* - * At this point active is permanently set and visible to this - * CPU without a lock, that fact is relied on to skip the unlock - * in range_end. - */ - return 0; + if (IS_ENABLED(CONFIG_LOCKDEP)) { + fs_reclaim_acquire(GFP_KERNEL); + mutex_lock(&umem_odp->umem_mutex); + mutex_unlock(&umem_odp->umem_mutex); + fs_reclaim_release(GFP_KERNEL); } - return rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, range->start, - range->end, - invalidate_range_start_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; + nr_entries = (end - start) >> PAGE_SHIFT; + if (!(nr_entries * PAGE_SIZE / page_size)) + return -EINVAL; - 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); -} + map = &umem_odp->map; + if (ib_uses_virt_dma(dev)) { + map->pfn_list = kvcalloc(nr_entries, sizeof(*map->pfn_list), + GFP_KERNEL | __GFP_NOWARN); + if (!map->pfn_list) + ret = -ENOMEM; + } else + ret = hmm_dma_map_alloc(dev->dma_device, map, + (end - start) >> PAGE_SHIFT, + 1 << umem_odp->page_shift); + if (ret) + return ret; + + ret = mmu_interval_notifier_insert(&umem_odp->notifier, + umem_odp->umem.owning_mm, start, + end - start, ops); + if (ret) + goto out_free_map; -static const struct mmu_notifier_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, -}; + return 0; -static void add_umem_to_per_mm(struct ib_umem_odp *umem_odp) -{ - struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm; - struct ib_umem *umem = &umem_odp->umem; - - down_write(&per_mm->umem_rwsem); - if (likely(ib_umem_start(umem) != ib_umem_end(umem))) - rbt_ib_umem_insert(&umem_odp->interval_tree, - &per_mm->umem_tree); - up_write(&per_mm->umem_rwsem); +out_free_map: + if (ib_uses_virt_dma(dev)) + kvfree(map->pfn_list); + else + hmm_dma_map_free(dev->dma_device, map); + return ret; } -static void remove_umem_from_per_mm(struct ib_umem_odp *umem_odp) +/** + * ib_umem_odp_alloc_implicit - Allocate a parent implicit ODP umem + * + * Implicit ODP umems do not have a VA range and do not have any page lists. + * They exist only to hold the per_mm reference to help the driver create + * children umems. + * + * @device: IB device to create UMEM + * @access: ib_reg_mr access flags + */ +struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device, + int access) { - struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm; - struct ib_umem *umem = &umem_odp->umem; - - down_write(&per_mm->umem_rwsem); - 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); -} + struct ib_umem *umem; + struct ib_umem_odp *umem_odp; -static struct ib_ucontext_per_mm *alloc_per_mm(struct ib_ucontext *ctx, - struct mm_struct *mm) -{ - struct ib_ucontext_per_mm *per_mm; - int ret; + if (access & IB_ACCESS_HUGETLB) + return ERR_PTR(-EINVAL); - per_mm = kzalloc(sizeof(*per_mm), GFP_KERNEL); - if (!per_mm) + umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL); + if (!umem_odp) return ERR_PTR(-ENOMEM); - - per_mm->context = ctx; - per_mm->mm = mm; - per_mm->umem_tree = RB_ROOT_CACHED; - init_rwsem(&per_mm->umem_rwsem); - per_mm->active = ctx->invalidate_range; - - rcu_read_lock(); - per_mm->tgid = get_task_pid(current->group_leader, PIDTYPE_PID); - rcu_read_unlock(); - - WARN_ON(mm != current->mm); - - per_mm->mn.ops = &ib_umem_notifiers; - ret = mmu_notifier_register(&per_mm->mn, per_mm->mm); - if (ret) { - dev_err(&ctx->device->dev, - "Failed to register mmu_notifier %d\n", ret); - goto out_pid; - } - - list_add(&per_mm->ucontext_list, &ctx->per_mm_list); - return per_mm; - -out_pid: - put_pid(per_mm->tgid); - kfree(per_mm); - return ERR_PTR(ret); -} - -static int get_per_mm(struct ib_umem_odp *umem_odp) -{ - struct ib_ucontext *ctx = umem_odp->umem.context; - struct ib_ucontext_per_mm *per_mm; - - /* - * Generally speaking we expect only one or two per_mm in this list, - * so no reason to optimize this search today. - */ - mutex_lock(&ctx->per_mm_list_lock); - list_for_each_entry(per_mm, &ctx->per_mm_list, ucontext_list) { - if (per_mm->mm == umem_odp->umem.owning_mm) - goto found; - } - - per_mm = alloc_per_mm(ctx, umem_odp->umem.owning_mm); - if (IS_ERR(per_mm)) { - mutex_unlock(&ctx->per_mm_list_lock); - return PTR_ERR(per_mm); - } - -found: - umem_odp->per_mm = per_mm; - per_mm->odp_mrs_count++; - mutex_unlock(&ctx->per_mm_list_lock); - - return 0; -} - -static void free_per_mm(struct rcu_head *rcu) -{ - kfree(container_of(rcu, struct ib_ucontext_per_mm, rcu)); + umem = &umem_odp->umem; + umem->ibdev = device; + umem->writable = ib_access_writable(access); + umem->owning_mm = current->mm; + umem_odp->page_shift = PAGE_SHIFT; + + umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID); + ib_init_umem_implicit_odp(umem_odp); + return umem_odp; } +EXPORT_SYMBOL(ib_umem_odp_alloc_implicit); -void put_per_mm(struct ib_umem_odp *umem_odp) +/** + * ib_umem_odp_alloc_child - Allocate a child ODP umem under an implicit + * parent ODP umem + * + * @root: The parent umem enclosing the child. This must be allocated using + * ib_alloc_implicit_odp_umem() + * @addr: The starting userspace VA + * @size: The length of the userspace VA + * @ops: MMU interval ops, currently only @invalidate + */ +struct ib_umem_odp * +ib_umem_odp_alloc_child(struct ib_umem_odp *root, unsigned long addr, + size_t size, + const struct mmu_interval_notifier_ops *ops) { - struct ib_ucontext_per_mm *per_mm = umem_odp->per_mm; - struct ib_ucontext *ctx = umem_odp->umem.context; - bool need_free; - - mutex_lock(&ctx->per_mm_list_lock); - umem_odp->per_mm = NULL; - per_mm->odp_mrs_count--; - need_free = per_mm->odp_mrs_count == 0; - if (need_free) - list_del(&per_mm->ucontext_list); - mutex_unlock(&ctx->per_mm_list_lock); - - if (!need_free) - return; - /* - * NOTE! mmu_notifier_unregister() can happen between a start/end - * callback, resulting in an start/end, and thus an unbalanced - * lock. This doesn't really matter to us since we are about to kfree - * the memory that holds the lock, however LOCKDEP doesn't like this. + * Caller must ensure that root cannot be freed during the call to + * ib_alloc_odp_umem. */ - down_write(&per_mm->umem_rwsem); - per_mm->active = false; - 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); - put_pid(per_mm->tgid); - mmu_notifier_call_srcu(&per_mm->rcu, free_per_mm); -} - -struct ib_umem_odp *ib_alloc_odp_umem(struct ib_ucontext_per_mm *per_mm, - unsigned long addr, size_t size) -{ - struct ib_ucontext *ctx = per_mm->context; struct ib_umem_odp *odp_data; struct ib_umem *umem; - int pages = size >> PAGE_SHIFT; int ret; + if (WARN_ON(!root->is_implicit_odp)) + return ERR_PTR(-EINVAL); + odp_data = kzalloc(sizeof(*odp_data), GFP_KERNEL); if (!odp_data) return ERR_PTR(-ENOMEM); umem = &odp_data->umem; - umem->context = ctx; + umem->ibdev = root->umem.ibdev; umem->length = size; umem->address = addr; - umem->page_shift = PAGE_SHIFT; - umem->writable = 1; - umem->is_odp = 1; - 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) { - ret = -ENOMEM; - goto out_odp_data; - } - - odp_data->dma_list = - vzalloc(array_size(pages, sizeof(*odp_data->dma_list))); - if (!odp_data->dma_list) { - ret = -ENOMEM; - goto out_page_list; - } + umem->writable = root->umem.writable; + umem->owning_mm = root->umem.owning_mm; + odp_data->page_shift = PAGE_SHIFT; + odp_data->notifier.ops = ops; /* - * Caller must ensure that the umem_odp that the per_mm came from - * cannot be freed during the call to ib_alloc_odp_umem. + * A mmget must be held when registering a notifier, the owming_mm only + * has a mm_grab at this point. */ - mutex_lock(&ctx->per_mm_list_lock); - per_mm->odp_mrs_count++; - mutex_unlock(&ctx->per_mm_list_lock); - add_umem_to_per_mm(odp_data); + if (!mmget_not_zero(umem->owning_mm)) { + ret = -EFAULT; + goto out_free; + } + odp_data->tgid = get_pid(root->tgid); + ret = ib_init_umem_odp(odp_data, ops); + if (ret) + goto out_tgid; + mmput(umem->owning_mm); return odp_data; -out_page_list: - vfree(odp_data->page_list); -out_odp_data: +out_tgid: + put_pid(odp_data->tgid); + mmput(umem->owning_mm); +out_free: kfree(odp_data); return ERR_PTR(ret); } -EXPORT_SYMBOL(ib_alloc_odp_umem); +EXPORT_SYMBOL(ib_umem_odp_alloc_child); -int ib_umem_odp_get(struct ib_umem_odp *umem_odp, int access) +/** + * ib_umem_odp_get - Create a umem_odp for a userspace va + * + * @device: IB device struct to get UMEM + * @addr: userspace virtual address to start at + * @size: length of region to pin + * @access: IB_ACCESS_xxx flags for memory being pinned + * @ops: MMU interval ops, currently only @invalidate + * + * The driver should use when the access flags indicate ODP memory. It avoids + * pinning, instead, stores the mm for future page fault handling in + * conjunction with MMU notifiers. + */ +struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device, + unsigned long addr, size_t size, int access, + const struct mmu_interval_notifier_ops *ops) { - struct ib_umem *umem = &umem_odp->umem; - /* - * NOTE: This must called in a process context where umem->owning_mm - * == current->mm - */ - struct mm_struct *mm = umem->owning_mm; - int ret_val; - - if (access & IB_ACCESS_HUGETLB) { - struct vm_area_struct *vma; - struct hstate *h; - - down_read(&mm->mmap_sem); - vma = find_vma(mm, ib_umem_start(umem)); - if (!vma || !is_vm_hugetlb_page(vma)) { - up_read(&mm->mmap_sem); - return -EINVAL; - } - h = hstate_vma(vma); - umem->page_shift = huge_page_shift(h); - up_read(&mm->mmap_sem); - umem->hugetlb = 1; - } else { - umem->hugetlb = 0; - } - - 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), - ib_umem_num_pages(umem))); - if (!umem_odp->page_list) - return -ENOMEM; - - umem_odp->dma_list = - vzalloc(array_size(sizeof(*umem_odp->dma_list), - ib_umem_num_pages(umem))); - if (!umem_odp->dma_list) { - ret_val = -ENOMEM; - goto out_page_list; - } - } + struct ib_umem_odp *umem_odp; + int ret; - ret_val = get_per_mm(umem_odp); - if (ret_val) - goto out_dma_list; - add_umem_to_per_mm(umem_odp); + if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND))) + return ERR_PTR(-EINVAL); - return 0; + umem_odp = kzalloc(sizeof(struct ib_umem_odp), GFP_KERNEL); + if (!umem_odp) + return ERR_PTR(-ENOMEM); -out_dma_list: - vfree(umem_odp->dma_list); -out_page_list: - vfree(umem_odp->page_list); - return ret_val; + umem_odp->umem.ibdev = device; + umem_odp->umem.length = size; + umem_odp->umem.address = addr; + umem_odp->umem.writable = ib_access_writable(access); + umem_odp->umem.owning_mm = current->mm; + umem_odp->notifier.ops = ops; + + umem_odp->page_shift = PAGE_SHIFT; +#ifdef CONFIG_HUGETLB_PAGE + if (access & IB_ACCESS_HUGETLB) + umem_odp->page_shift = HPAGE_SHIFT; +#endif + + umem_odp->tgid = get_task_pid(current->group_leader, PIDTYPE_PID); + ret = ib_init_umem_odp(umem_odp, ops); + if (ret) + goto err_put_pid; + return umem_odp; + +err_put_pid: + put_pid(umem_odp->tgid); + kfree(umem_odp); + return ERR_PTR(ret); } +EXPORT_SYMBOL(ib_umem_odp_get); -void ib_umem_odp_release(struct ib_umem_odp *umem_odp) +static void ib_umem_odp_free(struct ib_umem_odp *umem_odp) { - struct ib_umem *umem = &umem_odp->umem; + struct ib_device *dev = umem_odp->umem.ibdev; /* * Ensure that no more pages are mapped in the umem. @@ -461,111 +281,36 @@ void ib_umem_odp_release(struct ib_umem_odp *umem_odp) * It is the driver's responsibility to ensure, before calling us, * that the hardware will not attempt to access the MR any more. */ - ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem), - ib_umem_end(umem)); - - remove_umem_from_per_mm(umem_odp); - put_per_mm(umem_odp); - vfree(umem_odp->dma_list); - vfree(umem_odp->page_list); + mutex_lock(&umem_odp->umem_mutex); + ib_umem_odp_unmap_dma_pages(umem_odp, ib_umem_start(umem_odp), + ib_umem_end(umem_odp)); + mutex_unlock(&umem_odp->umem_mutex); + mmu_interval_notifier_remove(&umem_odp->notifier); + if (ib_uses_virt_dma(dev)) + kvfree(umem_odp->map.pfn_list); + else + hmm_dma_map_free(dev->dma_device, &umem_odp->map); } -/* - * 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. - * - * 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. - */ -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) +void ib_umem_odp_release(struct ib_umem_odp *umem_odp) { - 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; - } - -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 (!umem_odp->is_implicit_odp) + ib_umem_odp_free(umem_odp); - return ret; + put_pid(umem_odp->tgid); + kfree(umem_odp); } +EXPORT_SYMBOL(ib_umem_odp_release); /** - * ib_umem_odp_map_dma_pages - Pin and DMA map userspace memory in an ODP MR. + * ib_umem_odp_map_dma_and_lock - DMA map userspace memory in an ODP MR and lock it. * - * 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. + * Maps the range passed in the argument to DMA addresses. + * Upon success the ODP MR will be locked to let caller complete its device + * page table update. * * 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 @@ -574,227 +319,151 @@ out: * 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 + * @fault: is faulting required for the given range */ -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) +int ib_umem_odp_map_dma_and_lock(struct ib_umem_odp *umem_odp, u64 user_virt, + u64 bcnt, u64 access_mask, bool fault) + __acquires(&umem_odp->umem_mutex) { - 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) - return -EINVAL; - - if (user_virt < ib_umem_start(umem) || - user_virt + bcnt > ib_umem_end(umem)) + int pfn_index, dma_index, ret = 0, start_idx; + unsigned int page_shift, hmm_order, pfn_start_idx; + unsigned long num_pfns, current_seq; + struct hmm_range range = {}; + unsigned long timeout; + + if (user_virt < ib_umem_start(umem_odp) || + user_virt + bcnt > ib_umem_end(umem_odp)) return -EFAULT; - local_page_list = (struct page **)__get_free_page(GFP_KERNEL); - if (!local_page_list) - return -ENOMEM; - - 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. */ + page_shift = umem_odp->page_shift; /* * 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. */ - owning_process = get_pid_task(umem_odp->per_mm->tgid, PIDTYPE_PID); - if (WARN_ON(!mmget_not_zero(umem_odp->umem.owning_mm))) { + owning_process = get_pid_task(umem_odp->tgid, PIDTYPE_PID); + if (!owning_process || !mmget_not_zero(owning_mm)) { ret = -EINVAL; goto out_put_task; } - if (access_mask & ODP_WRITE_ALLOWED_BIT) - flags |= FOLL_WRITE; + range.notifier = &umem_odp->notifier; + range.start = ALIGN_DOWN(user_virt, 1UL << page_shift); + range.end = ALIGN(user_virt + bcnt, 1UL << page_shift); + pfn_start_idx = (range.start - ib_umem_start(umem_odp)) >> PAGE_SHIFT; + num_pfns = (range.end - range.start) >> PAGE_SHIFT; + if (fault) { + range.default_flags = HMM_PFN_REQ_FAULT; - start_idx = (user_virt - ib_umem_start(umem)) >> page_shift; - k = start_idx; + if (access_mask & HMM_PFN_WRITE) + range.default_flags |= HMM_PFN_REQ_WRITE; + } + + range.hmm_pfns = &(umem_odp->map.pfn_list[pfn_start_idx]); + timeout = jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT); + +retry: + current_seq = range.notifier_seq = + mmu_interval_read_begin(&umem_odp->notifier); - 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 *)); + mmap_read_lock(owning_mm); + ret = hmm_range_fault(&range); + mmap_read_unlock(owning_mm); + if (unlikely(ret)) { + if (ret == -EBUSY && !time_after(jiffies, timeout)) + goto retry; + goto out_put_mm; + } + + start_idx = (range.start - ib_umem_start(umem_odp)) >> page_shift; + dma_index = start_idx; + + mutex_lock(&umem_odp->umem_mutex); + if (mmu_interval_read_retry(&umem_odp->notifier, current_seq)) { + mutex_unlock(&umem_odp->umem_mutex); + goto retry; + } + + for (pfn_index = 0; pfn_index < num_pfns; + pfn_index += 1 << (page_shift - PAGE_SHIFT), dma_index++) { - 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). + * Since we asked for hmm_range_fault() to populate + * pages it shouldn't return an error entry on success. */ - 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; - } + WARN_ON(fault && range.hmm_pfns[pfn_index] & HMM_PFN_ERROR); + WARN_ON(fault && !(range.hmm_pfns[pfn_index] & HMM_PFN_VALID)); + if (!(range.hmm_pfns[pfn_index] & HMM_PFN_VALID)) + continue; - 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 (range.hmm_pfns[pfn_index] & HMM_PFN_DMA_MAPPED) + continue; - if (ret < 0) { - /* Release left over pages when handling errors. */ - for (++j; j < npages; ++j) - put_page(local_page_list[j]); + hmm_order = hmm_pfn_to_map_order(range.hmm_pfns[pfn_index]); + /* If a hugepage was detected and ODP wasn't set for, the umem + * page_shift will be used, the opposite case is an error. + */ + if (hmm_order + PAGE_SHIFT < page_shift) { + ret = -EINVAL; + ibdev_dbg(umem_odp->umem.ibdev, + "%s: un-expected hmm_order %u, page_shift %u\n", + __func__, hmm_order, page_shift); break; } } + /* upon success lock should stay on hold for the callee */ + if (!ret) + ret = dma_index - start_idx; + else + mutex_unlock(&umem_odp->umem_mutex); - if (ret >= 0) { - if (npages < 0 && k == start_idx) - ret = npages; - else - ret = k - start_idx; - } - - mmput(owning_mm); +out_put_mm: + mmput_async(owning_mm); out_put_task: if (owning_process) put_task_struct(owning_process); - free_page((unsigned long)local_page_list); return ret; } -EXPORT_SYMBOL(ib_umem_odp_map_dma_pages); +EXPORT_SYMBOL(ib_umem_odp_map_dma_and_lock); void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 virt, u64 bound) { - struct ib_umem *umem = &umem_odp->umem; - int idx; + struct ib_device *dev = umem_odp->umem.ibdev; u64 addr; - struct ib_device *dev = umem->context->device; - - 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. */ - 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--; + + lockdep_assert_held(&umem_odp->umem_mutex); + + virt = max_t(u64, virt, ib_umem_start(umem_odp)); + bound = min_t(u64, bound, ib_umem_end(umem_odp)); + for (addr = virt; addr < bound; addr += BIT(umem_odp->page_shift)) { + u64 offset = addr - ib_umem_start(umem_odp); + size_t idx = offset >> umem_odp->page_shift; + unsigned long pfn = umem_odp->map.pfn_list[idx]; + + if (!hmm_dma_unmap_pfn(dev->dma_device, &umem_odp->map, idx)) + goto clear; + + if (pfn & HMM_PFN_WRITE) { + struct page *page = hmm_pfn_to_page(pfn); + 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); } + umem_odp->npages--; +clear: + umem_odp->map.pfn_list[idx] &= ~HMM_PFN_FLAGS; } - mutex_unlock(&umem_odp->umem_mutex); } EXPORT_SYMBOL(ib_umem_odp_unmap_dma_pages); - -/* @last is not a part of the interval. See comment for function - * node_last. - */ -int rbt_ib_umem_for_each_in_range(struct rb_root_cached *root, - u64 start, u64 last, - umem_call_back cb, - bool blockable, - void *cookie) -{ - int ret_val = 0; - struct umem_odp_node *node, *next; - struct ib_umem_odp *umem; - - if (unlikely(start == last)) - return ret_val; - - for (node = rbt_ib_umem_iter_first(root, start, last - 1); - node; node = next) { - /* TODO move the blockable decision up to the callback */ - if (!blockable) - return -EAGAIN; - next = rbt_ib_umem_iter_next(node, start, last - 1); - umem = container_of(node, struct ib_umem_odp, interval_tree); - ret_val = cb(umem, start, last, cookie) || ret_val; - } - - return ret_val; -} -EXPORT_SYMBOL(rbt_ib_umem_for_each_in_range); - -struct ib_umem_odp *rbt_ib_umem_lookup(struct rb_root_cached *root, - u64 addr, u64 length) -{ - struct umem_odp_node *node; - - node = rbt_ib_umem_iter_first(root, addr, addr + length - 1); - if (node) - return container_of(node, struct ib_umem_odp, interval_tree); - return NULL; - -} -EXPORT_SYMBOL(rbt_ib_umem_lookup); |