drm/amdgpu: support userptr cross VMAs case with HMM
authorPhilip Yang <Philip.Yang@amd.com>
Mon, 4 Mar 2019 19:41:03 +0000 (14:41 -0500)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 19 Mar 2019 20:36:58 +0000 (15:36 -0500)
userptr may cross two VMAs if the forked child process (not call exec
after fork) malloc buffer, then free it, and then malloc larger size
buf, kerenl will create new VMA adjacent to old VMA which was cloned
from parent process, some pages of userptr are in the first VMA, the
rest pages are in the second VMA.

HMM expects range only have one VMA, loop over all VMAs in the address
range, create multiple ranges to handle this case. See
is_mergeable_anon_vma in mm/mmap.c for details.

Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c

index c1240bf243ba69ab8749ad97fbe62c9deefac944..c14198737dcdf852bf9afda605df98ba7c0ff6c0 100644 (file)
@@ -711,7 +711,8 @@ struct amdgpu_ttm_tt {
        struct task_struct      *usertask;
        uint32_t                userflags;
 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
-       struct hmm_range        range;
+       struct hmm_range        *ranges;
+       int                     nr_ranges;
 #endif
 };
 
@@ -723,62 +724,108 @@ struct amdgpu_ttm_tt {
  * once afterwards to stop HMM tracking
  */
 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
+
+/* Support Userptr pages cross max 16 vmas */
+#define MAX_NR_VMAS    (16)
+
 int amdgpu_ttm_tt_get_user_pages(struct ttm_tt *ttm, struct page **pages)
 {
        struct amdgpu_ttm_tt *gtt = (void *)ttm;
        struct mm_struct *mm = gtt->usertask->mm;
-       unsigned long end = gtt->userptr + ttm->num_pages * PAGE_SIZE;
-       struct hmm_range *range = &gtt->range;
-       int r = 0, i;
+       unsigned long start = gtt->userptr;
+       unsigned long end = start + ttm->num_pages * PAGE_SIZE;
+       struct hmm_range *ranges;
+       struct vm_area_struct *vma = NULL, *vmas[MAX_NR_VMAS];
+       uint64_t *pfns, f;
+       int r = 0, i, nr_pages;
 
        if (!mm) /* Happens during process shutdown */
                return -ESRCH;
 
-       amdgpu_hmm_init_range(range);
-
        down_read(&mm->mmap_sem);
 
-       range->vma = find_vma(mm, gtt->userptr);
-       if (!range_in_vma(range->vma, gtt->userptr, end))
-               r = -EFAULT;
-       else if ((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
-               range->vma->vm_file)
+       /* user pages may cross multiple VMAs */
+       gtt->nr_ranges = 0;
+       do {
+               unsigned long vm_start;
+
+               if (gtt->nr_ranges >= MAX_NR_VMAS) {
+                       DRM_ERROR("Too many VMAs in userptr range\n");
+                       r = -EFAULT;
+                       goto out;
+               }
+
+               vm_start = vma ? vma->vm_end : start;
+               vma = find_vma(mm, vm_start);
+               if (unlikely(!vma || vm_start < vma->vm_start)) {
+                       r = -EFAULT;
+                       goto out;
+               }
+               vmas[gtt->nr_ranges++] = vma;
+       } while (end > vma->vm_end);
+
+       DRM_DEBUG_DRIVER("0x%lx nr_ranges %d pages 0x%lx\n",
+               start, gtt->nr_ranges, ttm->num_pages);
+
+       if (unlikely((gtt->userflags & AMDGPU_GEM_USERPTR_ANONONLY) &&
+               vmas[0]->vm_file)) {
                r = -EPERM;
-       if (r)
                goto out;
+       }
 
-       range->pfns = kvmalloc_array(ttm->num_pages, sizeof(uint64_t),
-                                    GFP_KERNEL);
-       if (range->pfns == NULL) {
+       ranges = kvmalloc_array(gtt->nr_ranges, sizeof(*ranges), GFP_KERNEL);
+       if (unlikely(!ranges)) {
                r = -ENOMEM;
                goto out;
        }
-       range->start = gtt->userptr;
-       range->end = end;
 
-       range->pfns[0] = range->flags[HMM_PFN_VALID];
-       range->pfns[0] |= amdgpu_ttm_tt_is_readonly(ttm) ?
-                               0 : range->flags[HMM_PFN_WRITE];
-       for (i = 1; i < ttm->num_pages; i++)
-               range->pfns[i] = range->pfns[0];
+       pfns = kvmalloc_array(ttm->num_pages, sizeof(*pfns), GFP_KERNEL);
+       if (unlikely(!pfns)) {
+               r = -ENOMEM;
+               goto out_free_ranges;
+       }
+
+       for (i = 0; i < gtt->nr_ranges; i++)
+               amdgpu_hmm_init_range(&ranges[i]);
+
+       f = ranges[0].flags[HMM_PFN_VALID];
+       f |= amdgpu_ttm_tt_is_readonly(ttm) ?
+                               0 : ranges[0].flags[HMM_PFN_WRITE];
+       memset64(pfns, f, ttm->num_pages);
+
+       for (nr_pages = 0, i = 0; i < gtt->nr_ranges; i++) {
+               ranges[i].vma = vmas[i];
+               ranges[i].start = max(start, vmas[i]->vm_start);
+               ranges[i].end = min(end, vmas[i]->vm_end);
+               ranges[i].pfns = pfns + nr_pages;
+               nr_pages += (ranges[i].end - ranges[i].start) / PAGE_SIZE;
+
+               r = hmm_vma_fault(&ranges[i], true);
+               if (unlikely(r))
+                       break;
+       }
+       if (unlikely(r)) {
+               while (i--)
+                       hmm_vma_range_done(&ranges[i]);
 
-       /* This may trigger page table update */
-       r = hmm_vma_fault(range, true);
-       if (r)
                goto out_free_pfns;
+       }
 
        up_read(&mm->mmap_sem);
 
        for (i = 0; i < ttm->num_pages; i++)
-               pages[i] = hmm_pfn_to_page(range, range->pfns[i]);
+               pages[i] = hmm_pfn_to_page(&ranges[0], pfns[i]);
+       gtt->ranges = ranges;
 
        return 0;
 
 out_free_pfns:
-       kvfree(range->pfns);
-       range->pfns = NULL;
+       kvfree(pfns);
+out_free_ranges:
+       kvfree(ranges);
 out:
        up_read(&mm->mmap_sem);
+
        return r;
 }
 
@@ -792,15 +839,23 @@ bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
 {
        struct amdgpu_ttm_tt *gtt = (void *)ttm;
        bool r = false;
+       int i;
 
        if (!gtt || !gtt->userptr)
                return false;
 
-       WARN_ONCE(!gtt->range.pfns, "No user pages to check\n");
-       if (gtt->range.pfns) {
-               r = hmm_vma_range_done(&gtt->range);
-               kvfree(gtt->range.pfns);
-               gtt->range.pfns = NULL;
+       DRM_DEBUG_DRIVER("user_pages_done 0x%llx nr_ranges %d pages 0x%lx\n",
+               gtt->userptr, gtt->nr_ranges, ttm->num_pages);
+
+       WARN_ONCE(!gtt->ranges || !gtt->ranges[0].pfns,
+               "No user pages to check\n");
+
+       if (gtt->ranges) {
+               for (i = 0; i < gtt->nr_ranges; i++)
+                       r |= hmm_vma_range_done(&gtt->ranges[i]);
+               kvfree(gtt->ranges[0].pfns);
+               kvfree(gtt->ranges);
+               gtt->ranges = NULL;
        }
 
        return r;
@@ -884,8 +939,9 @@ static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
        sg_free_table(ttm->sg);
 
 #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
-       if (gtt->range.pfns &&
-           ttm->pages[0] == hmm_pfn_to_page(&gtt->range, gtt->range.pfns[0]))
+       if (gtt->ranges &&
+           ttm->pages[0] == hmm_pfn_to_page(&gtt->ranges[0],
+                                            gtt->ranges[0].pfns[0]))
                WARN_ONCE(1, "Missing get_user_page_done\n");
 #endif
 }