]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/i915/selftest_migrate: Check CCS meta data clear
authorRamalingam C <ramalingam.c@intel.com>
Tue, 5 Apr 2022 15:08:38 +0000 (20:38 +0530)
committerRamalingam C <ramalingam.c@intel.com>
Thu, 14 Apr 2022 07:50:29 +0000 (13:20 +0530)
Extend the live migrate selftest, to verify the ccs surface clearing
during the Flat-CCS capable lmem obj clear.

v2:
  Look at right places for ccs data [Thomas]

Signed-off-by: Ramalingam C <ramalingam.c@intel.com>
Reviewed-by: Thomas Hellstrom <thomas.hellstrom@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220405150840.29351-8-ramalingam.c@intel.com
drivers/gpu/drm/i915/gt/selftest_migrate.c

index 1c2417ebeca0f1678662a6e5536d5c248d997ca2..2b0c8799994900312a744ef94f4ed1f642a1f962 100644 (file)
@@ -132,6 +132,124 @@ err_free_src:
        return err;
 }
 
+static int intel_context_copy_ccs(struct intel_context *ce,
+                                 const struct i915_deps *deps,
+                                 struct scatterlist *sg,
+                                 enum i915_cache_level cache_level,
+                                 bool write_to_ccs,
+                                 struct i915_request **out)
+{
+       u8 src_access = write_to_ccs ? DIRECT_ACCESS : INDIRECT_ACCESS;
+       u8 dst_access = write_to_ccs ? INDIRECT_ACCESS : DIRECT_ACCESS;
+       struct sgt_dma it = sg_sgt(sg);
+       struct i915_request *rq;
+       u32 offset;
+       int err;
+
+       GEM_BUG_ON(ce->vm != ce->engine->gt->migrate.context->vm);
+       *out = NULL;
+
+       GEM_BUG_ON(ce->ring->size < SZ_64K);
+
+       offset = 0;
+       if (HAS_64K_PAGES(ce->engine->i915))
+               offset = CHUNK_SZ;
+
+       do {
+               int len;
+
+               rq = i915_request_create(ce);
+               if (IS_ERR(rq)) {
+                       err = PTR_ERR(rq);
+                       goto out_ce;
+               }
+
+               if (deps) {
+                       err = i915_request_await_deps(rq, deps);
+                       if (err)
+                               goto out_rq;
+
+                       if (rq->engine->emit_init_breadcrumb) {
+                               err = rq->engine->emit_init_breadcrumb(rq);
+                               if (err)
+                                       goto out_rq;
+                       }
+
+                       deps = NULL;
+               }
+
+               /* The PTE updates + clear must not be interrupted. */
+               err = emit_no_arbitration(rq);
+               if (err)
+                       goto out_rq;
+
+               len = emit_pte(rq, &it, cache_level, true, offset, CHUNK_SZ);
+               if (len <= 0) {
+                       err = len;
+                       goto out_rq;
+               }
+
+               err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
+               if (err)
+                       goto out_rq;
+
+               err = emit_copy_ccs(rq, offset, dst_access,
+                                   offset, src_access, len);
+               if (err)
+                       goto out_rq;
+
+               err = rq->engine->emit_flush(rq, EMIT_INVALIDATE);
+
+               /* Arbitration is re-enabled between requests. */
+out_rq:
+               if (*out)
+                       i915_request_put(*out);
+               *out = i915_request_get(rq);
+               i915_request_add(rq);
+               if (err || !it.sg || !sg_dma_len(it.sg))
+                       break;
+
+               cond_resched();
+       } while (1);
+
+out_ce:
+       return err;
+}
+
+static int
+intel_migrate_ccs_copy(struct intel_migrate *m,
+                      struct i915_gem_ww_ctx *ww,
+                      const struct i915_deps *deps,
+                      struct scatterlist *sg,
+                      enum i915_cache_level cache_level,
+                      bool write_to_ccs,
+                      struct i915_request **out)
+{
+       struct intel_context *ce;
+       int err;
+
+       *out = NULL;
+       if (!m->context)
+               return -ENODEV;
+
+       ce = intel_migrate_create_context(m);
+       if (IS_ERR(ce))
+               ce = intel_context_get(m->context);
+       GEM_BUG_ON(IS_ERR(ce));
+
+       err = intel_context_pin_ww(ce, ww);
+       if (err)
+               goto out;
+
+       err = intel_context_copy_ccs(ce, deps, sg, cache_level,
+                                    write_to_ccs, out);
+
+       intel_context_unpin(ce);
+out:
+       intel_context_put(ce);
+       return err;
+}
+
 static int clear(struct intel_migrate *migrate,
                 int (*fn)(struct intel_migrate *migrate,
                           struct i915_gem_ww_ctx *ww,
@@ -144,7 +262,8 @@ static int clear(struct intel_migrate *migrate,
        struct drm_i915_gem_object *obj;
        struct i915_request *rq;
        struct i915_gem_ww_ctx ww;
-       u32 *vaddr;
+       u32 *vaddr, val = 0;
+       bool ccs_cap = false;
        int err = 0;
        int i;
 
@@ -155,7 +274,12 @@ static int clear(struct intel_migrate *migrate,
        /* Consider the rounded up memory too */
        sz = obj->base.size;
 
+       if (HAS_FLAT_CCS(i915) && i915_gem_object_is_lmem(obj))
+               ccs_cap = true;
+
        for_i915_gem_ww(&ww, err, true) {
+               int ccs_bytes, ccs_bytes_per_chunk;
+
                err = i915_gem_object_lock(obj, &ww);
                if (err)
                        continue;
@@ -170,44 +294,114 @@ static int clear(struct intel_migrate *migrate,
                        vaddr[i] = ~i;
                i915_gem_object_flush_map(obj);
 
-               err = fn(migrate, &ww, obj, sz, &rq);
-               if (!err)
-                       continue;
+               if (ccs_cap && !val) {
+                       /* Write the obj data into ccs surface */
+                       err = intel_migrate_ccs_copy(migrate, &ww, NULL,
+                                                    obj->mm.pages->sgl,
+                                                    obj->cache_level,
+                                                    true, &rq);
+                       if (rq && !err) {
+                               if (i915_request_wait(rq, 0, HZ) < 0) {
+                                       pr_err("%ps timed out, size: %u\n",
+                                              fn, sz);
+                                       err = -ETIME;
+                               }
+                               i915_request_put(rq);
+                               rq = NULL;
+                       }
+                       if (err)
+                               continue;
+               }
 
-               if (err != -EDEADLK && err != -EINTR && err != -ERESTARTSYS)
-                       pr_err("%ps failed, size: %u\n", fn, sz);
-               if (rq) {
-                       i915_request_wait(rq, 0, HZ);
+               err = fn(migrate, &ww, obj, val, &rq);
+               if (rq && !err) {
+                       if (i915_request_wait(rq, 0, HZ) < 0) {
+                               pr_err("%ps timed out, size: %u\n", fn, sz);
+                               err = -ETIME;
+                       }
                        i915_request_put(rq);
+                       rq = NULL;
                }
-               i915_gem_object_unpin_map(obj);
-       }
-       if (err)
-               goto err_out;
+               if (err)
+                       continue;
 
-       if (rq) {
-               if (i915_request_wait(rq, 0, HZ) < 0) {
-                       pr_err("%ps timed out, size: %u\n", fn, sz);
-                       err = -ETIME;
+               i915_gem_object_flush_map(obj);
+
+               /* Verify the set/clear of the obj mem */
+               for (i = 0; !err && i < sz / PAGE_SIZE; i++) {
+                       int x = i * 1024 +
+                               i915_prandom_u32_max_state(1024, prng);
+
+                       if (vaddr[x] != val) {
+                               pr_err("%ps failed, (%u != %u), offset: %zu\n",
+                                      fn, vaddr[x], val,  x * sizeof(u32));
+                               igt_hexdump(vaddr + i * 1024, 4096);
+                               err = -EINVAL;
+                       }
                }
-               i915_request_put(rq);
-       }
+               if (err)
+                       continue;
 
-       for (i = 0; !err && i < sz / PAGE_SIZE; i++) {
-               int x = i * 1024 + i915_prandom_u32_max_state(1024, prng);
+               if (ccs_cap && !val) {
+                       for (i = 0; i < sz / sizeof(u32); i++)
+                               vaddr[i] = ~i;
+                       i915_gem_object_flush_map(obj);
+
+                       err = intel_migrate_ccs_copy(migrate, &ww, NULL,
+                                                    obj->mm.pages->sgl,
+                                                    obj->cache_level,
+                                                    false, &rq);
+                       if (rq && !err) {
+                               if (i915_request_wait(rq, 0, HZ) < 0) {
+                                       pr_err("%ps timed out, size: %u\n",
+                                              fn, sz);
+                                       err = -ETIME;
+                               }
+                               i915_request_put(rq);
+                               rq = NULL;
+                       }
+                       if (err)
+                               continue;
+
+                       ccs_bytes = GET_CCS_BYTES(i915, sz);
+                       ccs_bytes_per_chunk = GET_CCS_BYTES(i915, CHUNK_SZ);
+                       i915_gem_object_flush_map(obj);
+
+                       for (i = 0; !err && i < DIV_ROUND_UP(ccs_bytes, PAGE_SIZE); i++) {
+                               int offset = ((i * PAGE_SIZE)  /
+                                       ccs_bytes_per_chunk) * CHUNK_SZ / sizeof(u32);
+                               int ccs_bytes_left = (ccs_bytes - i * PAGE_SIZE) / sizeof(u32);
+                               int x = i915_prandom_u32_max_state(min_t(int, 1024,
+                                                                        ccs_bytes_left), prng);
+
+                               if (vaddr[offset + x]) {
+                                       pr_err("%ps ccs clearing failed, offset: %ld/%d\n",
+                                              fn, i * PAGE_SIZE + x * sizeof(u32), ccs_bytes);
+                                       igt_hexdump(vaddr + offset,
+                                                   min_t(int, 4096,
+                                                         ccs_bytes_left * sizeof(u32)));
+                                       err = -EINVAL;
+                               }
+                       }
+
+                       if (err)
+                               continue;
+               }
+               i915_gem_object_unpin_map(obj);
+       }
 
-               if (vaddr[x] != sz) {
-                       pr_err("%ps failed, size: %u, offset: %zu\n",
-                              fn, sz, x * sizeof(u32));
-                       igt_hexdump(vaddr + i * 1024, 4096);
-                       err = -EINVAL;
+       if (err) {
+               if (err != -EDEADLK && err != -EINTR && err != -ERESTARTSYS)
+                       pr_err("%ps failed, size: %u\n", fn, sz);
+               if (rq && err != -EINVAL) {
+                       i915_request_wait(rq, 0, HZ);
+                       i915_request_put(rq);
                }
+
+               i915_gem_object_unpin_map(obj);
        }
 
-       i915_gem_object_unpin_map(obj);
-err_out:
        i915_gem_object_put(obj);
-
        return err;
 }