drm/i915/gen8: page directories rework allocation
authorMichel Thierry <michel.thierry@intel.com>
Wed, 8 Apr 2015 11:13:27 +0000 (12:13 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 10 Apr 2015 06:56:09 +0000 (08:56 +0200)
Start using gen8_for_each_pdpe macro to allocate the page directories.

Similar to PTs, while setting up a page directory, make all entries of
the  pd point to the scratch pd before mapping (and make all its entries
point to the scratch page); this is to be safe in case of out of bound
access or  proactive prefetch. Systems without LLC require an explicit
flush.

v2: Rebased after s/free_pt_*/unmap_and_free_pt/ change.
v3: Rebased after teardown va range logic was removed.
v4: Keep setting up all page directories for systems with less than 4GB
of memory.
v5: Initialize PDs. (Mika)
v6: Initialize also the extra PDs from systems with less than 4GB of
memory. (Mika)

Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Michel Thierry <michel.thierry@intel.com> (v2+)
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem_gtt.c

index c2045bd45e5e626fddc40132d275c44570e47865..e5770189ebff3bc5a9ef2f2a4bac22dbaa9005e7 100644 (file)
@@ -606,6 +606,36 @@ static void gen8_ppgtt_insert_entries(struct i915_address_space *vm,
        }
 }
 
+static void __gen8_do_map_pt(gen8_pde_t * const pde,
+                            struct i915_page_table *pt,
+                            struct drm_device *dev)
+{
+       gen8_pde_t entry =
+               gen8_pde_encode(dev, pt->daddr, I915_CACHE_LLC);
+       *pde = entry;
+}
+
+static void gen8_initialize_pd(struct i915_address_space *vm,
+                              struct i915_page_directory *pd)
+{
+       struct i915_hw_ppgtt *ppgtt =
+                       container_of(vm, struct i915_hw_ppgtt, base);
+       gen8_pde_t *page_directory;
+       struct i915_page_table *pt;
+       int i;
+
+       page_directory = kmap_atomic(pd->page);
+       pt = ppgtt->scratch_pt;
+       for (i = 0; i < I915_PDES; i++)
+               /* Map the PDE to the page table */
+               __gen8_do_map_pt(page_directory + i, pt, vm->dev);
+
+       if (!HAS_LLC(vm->dev))
+               drm_clflush_virt_range(page_directory, PAGE_SIZE);
+
+       kunmap_atomic(page_directory);
+}
+
 static void gen8_free_page_tables(struct i915_page_directory *pd, struct drm_device *dev)
 {
        int i;
@@ -633,6 +663,8 @@ static void gen8_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
                gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev);
                unmap_and_free_pd(ppgtt->pdp.page_directory[i]);
        }
+
+       unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
 }
 
 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
@@ -663,25 +695,55 @@ unwind_out:
        return -ENOMEM;
 }
 
-static int gen8_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt,
-                                               const int max_pdp)
+static int gen8_ppgtt_alloc_page_directories(struct i915_page_directory_pointer *pdp,
+                                    uint64_t start,
+                                    uint64_t length)
 {
-       int i;
+       struct i915_hw_ppgtt *ppgtt =
+               container_of(pdp, struct i915_hw_ppgtt, pdp);
+       struct i915_page_directory *unused;
+       uint64_t temp;
+       uint32_t pdpe;
+
+       /* FIXME: PPGTT container_of won't work for 64b */
+       WARN_ON((start + length) > 0x800000000ULL);
+
+       gen8_for_each_pdpe(unused, pdp, start, length, temp, pdpe) {
+               WARN_ON(unused);
+               pdp->page_directory[pdpe] = alloc_pd_single();
+               if (IS_ERR(ppgtt->pdp.page_directory[pdpe]))
+                       goto unwind_out;
+
+               gen8_initialize_pd(&ppgtt->base,
+                                  ppgtt->pdp.page_directory[pdpe]);
+               ppgtt->num_pd_pages++;
+       }
 
-       for (i = 0; i < max_pdp; i++) {
-               ppgtt->pdp.page_directory[i] = alloc_pd_single();
-               if (IS_ERR(ppgtt->pdp.page_directory[i]))
+       /* XXX: Still alloc all page directories in systems with less than
+        * 4GB of memory. This won't be needed after a subsequent patch.
+        */
+       while (ppgtt->num_pd_pages < GEN8_LEGACY_PDPES) {
+               ppgtt->pdp.page_directory[ppgtt->num_pd_pages] = alloc_pd_single();
+               if (IS_ERR(ppgtt->pdp.page_directory[ppgtt->num_pd_pages]))
                        goto unwind_out;
+
+               gen8_initialize_pd(&ppgtt->base,
+                                  ppgtt->pdp.page_directory[ppgtt->num_pd_pages]);
+               pdpe++;
+               ppgtt->num_pd_pages++;
        }
 
-       ppgtt->num_pd_pages = max_pdp;
        BUG_ON(ppgtt->num_pd_pages > GEN8_LEGACY_PDPES);
 
        return 0;
 
 unwind_out:
-       while (i--)
-               unmap_and_free_pd(ppgtt->pdp.page_directory[i]);
+       while (pdpe--) {
+               unmap_and_free_pd(ppgtt->pdp.page_directory[pdpe]);
+               ppgtt->num_pd_pages--;
+       }
+
+       WARN_ON(ppgtt->num_pd_pages);
 
        return -ENOMEM;
 }
@@ -691,7 +753,8 @@ static int gen8_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt,
 {
        int ret;
 
-       ret = gen8_ppgtt_allocate_page_directories(ppgtt, max_pdp);
+       ret = gen8_ppgtt_alloc_page_directories(&ppgtt->pdp, ppgtt->base.start,
+                                       ppgtt->base.total);
        if (ret)
                return ret;
 
@@ -769,6 +832,17 @@ static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
        if (size % (1<<30))
                DRM_INFO("Pages will be wasted unless GTT size (%llu) is divisible by 1GB\n", size);
 
+       ppgtt->base.start = 0;
+       /* This is the area that we advertise as usable for the caller */
+       ppgtt->base.total = max_pdp * I915_PDES * GEN8_PTES * PAGE_SIZE;
+       WARN_ON(ppgtt->base.total == 0);
+
+       ppgtt->scratch_pt = alloc_pt_single(ppgtt->base.dev);
+       if (IS_ERR(ppgtt->scratch_pt))
+               return PTR_ERR(ppgtt->scratch_pt);
+
+       gen8_initialize_pt(&ppgtt->base, ppgtt->scratch_pt);
+
        /* 1. Do all our allocations for page directories and page tables.
         * We allocate more than was asked so that we can point the unused parts
         * to valid entries that point to scratch page. Dynamic page tables
@@ -794,7 +868,7 @@ static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
        }
 
        /*
-        * 3. Map all the page directory entires to point to the page tables
+        * 3. Map all the page directory entries to point to the page tables
         * we've allocated.
         *
         * For now, the PPGTT helper functions all require that the PDEs are
@@ -820,10 +894,6 @@ static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt, uint64_t size)
        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;
-
-       /* This is the area that we advertise as usable for the caller */
-       ppgtt->base.total = max_pdp * I915_PDES * GEN8_PTES * PAGE_SIZE;
 
        /* Set all ptes to a valid scratch page. Also above requested space */
        ppgtt->base.clear_range(&ppgtt->base, 0,