drm/i915: Avoid allocating a page array for the gpu coredump
authorThomas Hellström <thomas.hellstrom@linux.intel.com>
Mon, 8 Nov 2021 17:45:44 +0000 (18:45 +0100)
committerThomas Hellström <thomas.hellstrom@linux.intel.com>
Fri, 26 Nov 2021 07:26:08 +0000 (08:26 +0100)
The gpu coredump typically takes place in a dma_fence signalling
critical path, and hence can't use GFP_KERNEL allocations, as that
means we might hit deadlocks under memory pressure. However
changing to __GFP_KSWAPD_RECLAIM which will be done in an upcoming
patch will instead mean a lower chance of the allocation succeeding.
In particular large contigous allocations like the coredump page
vector.
Remove the page vector in favor of a linked list of single pages.
Use the page lru list head as the list link, as the page owner is
allowed to do that.

Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Ramalingam C <ramalingam.c@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211108174547.979714-2-thomas.hellstrom@linux.intel.com
drivers/gpu/drm/i915/i915_gpu_error.c
drivers/gpu/drm/i915/i915_gpu_error.h

index 2a2d764..14de642 100644 (file)
@@ -275,16 +275,16 @@ static bool compress_start(struct i915_vma_compress *c)
 static void *compress_next_page(struct i915_vma_compress *c,
                                struct i915_vma_coredump *dst)
 {
-       void *page;
+       void *page_addr;
+       struct page *page;
 
-       if (dst->page_count >= dst->num_pages)
-               return ERR_PTR(-ENOSPC);
-
-       page = pool_alloc(&c->pool, ALLOW_FAIL);
-       if (!page)
+       page_addr = pool_alloc(&c->pool, ALLOW_FAIL);
+       if (!page_addr)
                return ERR_PTR(-ENOMEM);
 
-       return dst->pages[dst->page_count++] = page;
+       page = virt_to_page(page_addr);
+       list_add_tail(&page->lru, &dst->page_list);
+       return page_addr;
 }
 
 static int compress_page(struct i915_vma_compress *c,
@@ -397,7 +397,7 @@ static int compress_page(struct i915_vma_compress *c,
 
        if (!(wc && i915_memcpy_from_wc(ptr, src, PAGE_SIZE)))
                memcpy(ptr, src, PAGE_SIZE);
-       dst->pages[dst->page_count++] = ptr;
+       list_add_tail(&virt_to_page(ptr)->lru, &dst->page_list);
        cond_resched();
 
        return 0;
@@ -614,7 +614,7 @@ static void print_error_vma(struct drm_i915_error_state_buf *m,
                            const struct i915_vma_coredump *vma)
 {
        char out[ASCII85_BUFSZ];
-       int page;
+       struct page *page;
 
        if (!vma)
                return;
@@ -628,16 +628,17 @@ static void print_error_vma(struct drm_i915_error_state_buf *m,
                err_printf(m, "gtt_page_sizes = 0x%08x\n", vma->gtt_page_sizes);
 
        err_compression_marker(m);
-       for (page = 0; page < vma->page_count; page++) {
+       list_for_each_entry(page, &vma->page_list, lru) {
                int i, len;
+               const u32 *addr = page_address(page);
 
                len = PAGE_SIZE;
-               if (page == vma->page_count - 1)
+               if (page == list_last_entry(&vma->page_list, typeof(*page), lru))
                        len -= vma->unused;
                len = ascii85_encode_len(len);
 
                for (i = 0; i < len; i++)
-                       err_puts(m, ascii85_encode(vma->pages[page][i], out));
+                       err_puts(m, ascii85_encode(addr[i], out));
        }
        err_puts(m, "\n");
 }
@@ -946,10 +947,12 @@ static void i915_vma_coredump_free(struct i915_vma_coredump *vma)
 {
        while (vma) {
                struct i915_vma_coredump *next = vma->next;
-               int page;
+               struct page *page, *n;
 
-               for (page = 0; page < vma->page_count; page++)
-                       free_page((unsigned long)vma->pages[page]);
+               list_for_each_entry_safe(page, n, &vma->page_list, lru) {
+                       list_del_init(&page->lru);
+                       __free_page(page);
+               }
 
                kfree(vma);
                vma = next;
@@ -1016,7 +1019,6 @@ i915_vma_coredump_create(const struct intel_gt *gt,
        struct i915_ggtt *ggtt = gt->ggtt;
        const u64 slot = ggtt->error_capture.start;
        struct i915_vma_coredump *dst;
-       unsigned long num_pages;
        struct sgt_iter iter;
        int ret;
 
@@ -1025,9 +1027,7 @@ i915_vma_coredump_create(const struct intel_gt *gt,
        if (!vma || !vma->pages || !compress)
                return NULL;
 
-       num_pages = min_t(u64, vma->size, vma->obj->base.size) >> PAGE_SHIFT;
-       num_pages = DIV_ROUND_UP(10 * num_pages, 8); /* worstcase zlib growth */
-       dst = kmalloc(sizeof(*dst) + num_pages * sizeof(u32 *), ALLOW_FAIL);
+       dst = kmalloc(sizeof(*dst), ALLOW_FAIL);
        if (!dst)
                return NULL;
 
@@ -1036,14 +1036,13 @@ i915_vma_coredump_create(const struct intel_gt *gt,
                return NULL;
        }
 
+       INIT_LIST_HEAD(&dst->page_list);
        strcpy(dst->name, name);
        dst->next = NULL;
 
        dst->gtt_offset = vma->node.start;
        dst->gtt_size = vma->node.size;
        dst->gtt_page_sizes = vma->page_sizes.gtt;
-       dst->num_pages = num_pages;
-       dst->page_count = 0;
        dst->unused = 0;
 
        ret = -EINVAL;
@@ -1106,8 +1105,13 @@ i915_vma_coredump_create(const struct intel_gt *gt,
        }
 
        if (ret || compress_flush(compress, dst)) {
-               while (dst->page_count--)
-                       pool_free(&compress->pool, dst->pages[dst->page_count]);
+               struct page *page, *n;
+
+               list_for_each_entry_safe_reverse(page, n, &dst->page_list, lru) {
+                       list_del_init(&page->lru);
+                       pool_free(&compress->pool, page_address(page));
+               }
+
                kfree(dst);
                dst = NULL;
        }
index b98d8cd..5aedf51 100644 (file)
@@ -39,10 +39,8 @@ struct i915_vma_coredump {
        u64 gtt_size;
        u32 gtt_page_sizes;
 
-       int num_pages;
-       int page_count;
        int unused;
-       u32 *pages[];
+       struct list_head page_list;
 };
 
 struct i915_request_coredump {