]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
drm/i915/bdw: Reorganize PPGTT init
authorBen Widawsky <benjamin.widawsky@intel.com>
Thu, 20 Feb 2014 06:05:42 +0000 (22:05 -0800)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 4 Mar 2014 14:54:31 +0000 (15:54 +0100)
Create 3 clear stages in PPGTT init. This will help with upcoming
changes be more readable. The 3 stages are, allocation, dma mapping, and
writing the P[DT]Es

One nice benefit to the patches is that it makes 2 very clear error
points, allocation, and mapping, and avoids having to do any handling
after writing PTEs (something which was likely buggy before). This
simplified error handling I suspect will be helpful when we move to
deferred/dynamic page table allocation and mapping.

The patches also attempts to break up some of the steps into more
logical reviewable chunks, particularly when we free.

v2: Don't call cleanup on the error path since that takes down the
drm_mm and list entry, which aren't setup at this point.

v3: Fixes addressing Imre's comments from:
<1392821989.19792.13.camel@intelbox>

Don't do dynamic allocation for the page table DMA addresses. I can't
remember why I did it in the first place. This addresses one of Imre's
other issues.

Fix error path leak of page tables.

v4: Fix the fix of the error path leak. Original fix still leaked page
tables. (Imre)

Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem_gtt.c

index 49e79fbee45d28a27f6bb759a63456316abd834f..2bc07fb447bdb81b14ee95fdda71302a5d37a954 100644 (file)
@@ -332,6 +332,7 @@ static void gen8_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
 
 static void gen8_ppgtt_unmap_pages(struct i915_hw_ppgtt *ppgtt)
 {
+       struct pci_dev *hwdev = ppgtt->base.dev->pdev;
        int i, j;
 
        for (i = 0; i < ppgtt->num_pd_pages; i++) {
@@ -340,18 +341,14 @@ static void gen8_ppgtt_unmap_pages(struct i915_hw_ppgtt *ppgtt)
                if (!ppgtt->pd_dma_addr[i])
                        continue;
 
-               pci_unmap_page(ppgtt->base.dev->pdev,
-                              ppgtt->pd_dma_addr[i],
-                              PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+               pci_unmap_page(hwdev, ppgtt->pd_dma_addr[i], PAGE_SIZE,
+                              PCI_DMA_BIDIRECTIONAL);
 
                for (j = 0; j < GEN8_PDES_PER_PAGE; j++) {
                        dma_addr_t addr = ppgtt->gen8_pt_dma_addr[i][j];
                        if (addr)
-                               pci_unmap_page(ppgtt->base.dev->pdev,
-                                      addr,
-                                      PAGE_SIZE,
-                                      PCI_DMA_BIDIRECTIONAL);
-
+                               pci_unmap_page(hwdev, addr, PAGE_SIZE,
+                                              PCI_DMA_BIDIRECTIONAL);
                }
        }
 }
@@ -369,27 +366,27 @@ static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
 }
 
 /**
- * GEN8 legacy ppgtt programming is accomplished through 4 PDP registers with a
- * net effect resembling a 2-level page table in normal x86 terms. Each PDP
- * represents 1GB of memory
- * 4 * 512 * 512 * 4096 = 4GB legacy 32b address space.
+ * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
+ * with a net effect resembling a 2-level page table in normal x86 terms. Each
+ * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
+ * space.
  *
+ * FIXME: split allocation into smaller pieces. For now we only ever do this
+ * once, but with full PPGTT, the multiple contiguous allocations will be bad.
  * TODO: Do something with the size parameter
- **/
+ */
 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
 {
        struct page *pt_pages;
-       int i, j, ret = -ENOMEM;
        const int max_pdp = DIV_ROUND_UP(size, 1 << 30);
        const int num_pt_pages = GEN8_PDES_PER_PAGE * max_pdp;
+       struct pci_dev *hwdev = ppgtt->base.dev->pdev;
+       int i, j, ret;
 
        if (size % (1<<30))
                DRM_INFO("Pages will be wasted unless GTT size (%llu) is divisible by 1GB\n", size);
 
-       /* FIXME: split allocation into smaller pieces. For now we only ever do
-        * this once, but with full PPGTT, the multiple contiguous allocations
-        * will be bad.
-        */
+       /* 1. Do all our allocations for page directories and page tables */
        ppgtt->pd_pages = alloc_pages(GFP_KERNEL, get_order(max_pdp << PAGE_SHIFT));
        if (!ppgtt->pd_pages)
                return -ENOMEM;
@@ -404,52 +401,56 @@ static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
        ppgtt->num_pd_pages = 1 << get_order(max_pdp << PAGE_SHIFT);
        ppgtt->num_pt_pages = 1 << get_order(num_pt_pages << PAGE_SHIFT);
        ppgtt->num_pd_entries = max_pdp * GEN8_PDES_PER_PAGE;
-       ppgtt->enable = gen8_ppgtt_enable;
-       ppgtt->switch_mm = gen8_mm_switch;
-       ppgtt->base.clear_range = gen8_ppgtt_clear_range;
-       ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
-       ppgtt->base.cleanup = gen8_ppgtt_cleanup;
-       ppgtt->base.start = 0;
-       ppgtt->base.total = ppgtt->num_pt_pages * GEN8_PTES_PER_PAGE * PAGE_SIZE;
-
        BUG_ON(ppgtt->num_pd_pages > GEN8_LEGACY_PDPS);
 
+       for (i = 0; i < max_pdp; i++) {
+               ppgtt->gen8_pt_dma_addr[i] = kcalloc(GEN8_PDES_PER_PAGE,
+                                                    sizeof(dma_addr_t),
+                                                    GFP_KERNEL);
+               if (!ppgtt->gen8_pt_dma_addr[i]) {
+                       ret = -ENOMEM;
+                       goto bail;
+               }
+       }
+
        /*
-        * - Create a mapping for the page directories.
-        * - For each page directory:
-        *      allocate space for page table mappings.
-        *      map each page table
+        * 2. Create all the DMA mappings for the page directories and page
+        * tables
         */
        for (i = 0; i < max_pdp; i++) {
-               dma_addr_t temp;
-               temp = pci_map_page(ppgtt->base.dev->pdev,
-                                   &ppgtt->pd_pages[i], 0,
-                                   PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
-               if (pci_dma_mapping_error(ppgtt->base.dev->pdev, temp))
-                       goto err_out;
+               dma_addr_t pd_addr, pt_addr;
 
-               ppgtt->pd_dma_addr[i] = temp;
+               /* Get the page directory mappings */
+               pd_addr = pci_map_page(hwdev, &ppgtt->pd_pages[i], 0,
+                                      PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+               ret = pci_dma_mapping_error(ppgtt->base.dev->pdev, pd_addr);
+               if (ret)
+                       goto bail;
 
-               ppgtt->gen8_pt_dma_addr[i] = kmalloc(sizeof(dma_addr_t) * GEN8_PDES_PER_PAGE, GFP_KERNEL);
-               if (!ppgtt->gen8_pt_dma_addr[i])
-                       goto err_out;
+               ppgtt->pd_dma_addr[i] = pd_addr;
 
+               /* And the page table mappings per page directory */
                for (j = 0; j < GEN8_PDES_PER_PAGE; j++) {
                        struct page *p = &pt_pages[i * GEN8_PDES_PER_PAGE + j];
-                       temp = pci_map_page(ppgtt->base.dev->pdev,
-                                           p, 0, PAGE_SIZE,
-                                           PCI_DMA_BIDIRECTIONAL);
 
-                       if (pci_dma_mapping_error(ppgtt->base.dev->pdev, temp))
-                               goto err_out;
+                       pt_addr = pci_map_page(hwdev, p, 0, PAGE_SIZE,
+                                              PCI_DMA_BIDIRECTIONAL);
+                       ret = pci_dma_mapping_error(hwdev, pt_addr);
+                       if (ret)
+                               goto bail;
 
-                       ppgtt->gen8_pt_dma_addr[i][j] = temp;
+                       ppgtt->gen8_pt_dma_addr[i][j] = pt_addr;
                }
        }
 
-       /* For now, the PPGTT helper functions all require that the PDEs are
+       /*
+        * 3. Map all the page directory entires to point to the page tables
+        * we've allocated.
+        *
+        * For now, the PPGTT helper functions all require that the PDEs are
         * plugged in correctly. So we do that now/here. For aliasing PPGTT, we
-        * will never need to touch the PDEs again */
+        * will never need to touch the PDEs again.
+        */
        for (i = 0; i < max_pdp; i++) {
                gen8_ppgtt_pde_t *pd_vaddr;
                pd_vaddr = kmap_atomic(&ppgtt->pd_pages[i]);
@@ -461,6 +462,14 @@ static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
                kunmap_atomic(pd_vaddr);
        }
 
+       ppgtt->enable = gen8_ppgtt_enable;
+       ppgtt->switch_mm = gen8_mm_switch;
+       ppgtt->base.clear_range = gen8_ppgtt_clear_range;
+       ppgtt->base.insert_entries = gen8_ppgtt_insert_entries;
+       ppgtt->base.cleanup = gen8_ppgtt_cleanup;
+       ppgtt->base.start = 0;
+       ppgtt->base.total = ppgtt->num_pt_pages * GEN8_PTES_PER_PAGE * PAGE_SIZE;
+
        ppgtt->base.clear_range(&ppgtt->base, 0,
                                ppgtt->num_pd_entries * GEN8_PTES_PER_PAGE,
                                true);
@@ -473,8 +482,9 @@ static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
                         size % (1<<30));
        return 0;
 
-err_out:
-       ppgtt->base.cleanup(&ppgtt->base);
+bail:
+       gen8_ppgtt_unmap_pages(ppgtt);
+       gen8_ppgtt_free(ppgtt);
        return ret;
 }