1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include "gem/i915_gem_domain.h" 7 #include "gem/i915_gem_internal.h" 8 #include "gt/gen8_ppgtt.h" 9 10 #include "i915_drv.h" 11 #include "intel_display_types.h" 12 #include "intel_dpt.h" 13 #include "intel_fb.h" 14 15 struct i915_dpt { 16 struct i915_address_space vm; 17 18 struct drm_i915_gem_object *obj; 19 struct i915_vma *vma; 20 void __iomem *iomem; 21 }; 22 23 #define i915_is_dpt(vm) ((vm)->is_dpt) 24 25 static inline struct i915_dpt * 26 i915_vm_to_dpt(struct i915_address_space *vm) 27 { 28 BUILD_BUG_ON(offsetof(struct i915_dpt, vm)); 29 GEM_BUG_ON(!i915_is_dpt(vm)); 30 return container_of(vm, struct i915_dpt, vm); 31 } 32 33 #define dpt_total_entries(dpt) ((dpt)->vm.total >> PAGE_SHIFT) 34 35 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte) 36 { 37 writeq(pte, addr); 38 } 39 40 static void dpt_insert_page(struct i915_address_space *vm, 41 dma_addr_t addr, 42 u64 offset, 43 enum i915_cache_level level, 44 u32 flags) 45 { 46 struct i915_dpt *dpt = i915_vm_to_dpt(vm); 47 gen8_pte_t __iomem *base = dpt->iomem; 48 49 gen8_set_pte(base + offset / I915_GTT_PAGE_SIZE, 50 vm->pte_encode(addr, level, flags)); 51 } 52 53 static void dpt_insert_entries(struct i915_address_space *vm, 54 struct i915_vma_resource *vma_res, 55 enum i915_cache_level level, 56 u32 flags) 57 { 58 struct i915_dpt *dpt = i915_vm_to_dpt(vm); 59 gen8_pte_t __iomem *base = dpt->iomem; 60 const gen8_pte_t pte_encode = vm->pte_encode(0, level, flags); 61 struct sgt_iter sgt_iter; 62 dma_addr_t addr; 63 int i; 64 65 /* 66 * Note that we ignore PTE_READ_ONLY here. The caller must be careful 67 * not to allow the user to override access to a read only page. 68 */ 69 70 i = vma_res->start / I915_GTT_PAGE_SIZE; 71 for_each_sgt_daddr(addr, sgt_iter, vma_res->bi.pages) 72 gen8_set_pte(&base[i++], pte_encode | addr); 73 } 74 75 static void dpt_clear_range(struct i915_address_space *vm, 76 u64 start, u64 length) 77 { 78 } 79 80 static void dpt_bind_vma(struct i915_address_space *vm, 81 struct i915_vm_pt_stash *stash, 82 struct i915_vma_resource *vma_res, 83 enum i915_cache_level cache_level, 84 u32 flags) 85 { 86 u32 pte_flags; 87 88 if (vma_res->bound_flags) 89 return; 90 91 /* Applicable to VLV (gen8+ do not support RO in the GGTT) */ 92 pte_flags = 0; 93 if (vm->has_read_only && vma_res->bi.readonly) 94 pte_flags |= PTE_READ_ONLY; 95 if (vma_res->bi.lmem) 96 pte_flags |= PTE_LM; 97 98 vm->insert_entries(vm, vma_res, cache_level, pte_flags); 99 100 vma_res->page_sizes_gtt = I915_GTT_PAGE_SIZE; 101 102 /* 103 * Without aliasing PPGTT there's no difference between 104 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally 105 * upgrade to both bound if we bind either to avoid double-binding. 106 */ 107 vma_res->bound_flags = I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND; 108 } 109 110 static void dpt_unbind_vma(struct i915_address_space *vm, 111 struct i915_vma_resource *vma_res) 112 { 113 vm->clear_range(vm, vma_res->start, vma_res->vma_size); 114 } 115 116 static void dpt_cleanup(struct i915_address_space *vm) 117 { 118 struct i915_dpt *dpt = i915_vm_to_dpt(vm); 119 120 i915_gem_object_put(dpt->obj); 121 } 122 123 struct i915_vma *intel_dpt_pin(struct i915_address_space *vm) 124 { 125 struct drm_i915_private *i915 = vm->i915; 126 struct i915_dpt *dpt = i915_vm_to_dpt(vm); 127 intel_wakeref_t wakeref; 128 struct i915_vma *vma; 129 void __iomem *iomem; 130 struct i915_gem_ww_ctx ww; 131 u64 pin_flags = 0; 132 int err; 133 134 if (i915_gem_object_is_stolen(dpt->obj)) 135 pin_flags |= PIN_MAPPABLE; 136 137 wakeref = intel_runtime_pm_get(&i915->runtime_pm); 138 atomic_inc(&i915->gpu_error.pending_fb_pin); 139 140 for_i915_gem_ww(&ww, err, true) { 141 err = i915_gem_object_lock(dpt->obj, &ww); 142 if (err) 143 continue; 144 145 vma = i915_gem_object_ggtt_pin_ww(dpt->obj, &ww, NULL, 0, 4096, 146 pin_flags); 147 if (IS_ERR(vma)) { 148 err = PTR_ERR(vma); 149 continue; 150 } 151 152 iomem = i915_vma_pin_iomap(vma); 153 i915_vma_unpin(vma); 154 155 if (IS_ERR(iomem)) { 156 err = PTR_ERR(iomem); 157 continue; 158 } 159 160 dpt->vma = vma; 161 dpt->iomem = iomem; 162 163 i915_vma_get(vma); 164 } 165 166 atomic_dec(&i915->gpu_error.pending_fb_pin); 167 intel_runtime_pm_put(&i915->runtime_pm, wakeref); 168 169 return err ? ERR_PTR(err) : vma; 170 } 171 172 void intel_dpt_unpin(struct i915_address_space *vm) 173 { 174 struct i915_dpt *dpt = i915_vm_to_dpt(vm); 175 176 i915_vma_unpin_iomap(dpt->vma); 177 i915_vma_put(dpt->vma); 178 } 179 180 /** 181 * intel_dpt_resume - restore the memory mapping for all DPT FBs during system resume 182 * @i915: device instance 183 * 184 * Restore the memory mapping during system resume for all framebuffers which 185 * are mapped to HW via a GGTT->DPT page table. The content of these page 186 * tables are not stored in the hibernation image during S4 and S3RST->S4 187 * transitions, so here we reprogram the PTE entries in those tables. 188 * 189 * This function must be called after the mappings in GGTT have been restored calling 190 * i915_ggtt_resume(). 191 */ 192 void intel_dpt_resume(struct drm_i915_private *i915) 193 { 194 struct drm_framebuffer *drm_fb; 195 196 if (!HAS_DISPLAY(i915)) 197 return; 198 199 mutex_lock(&i915->drm.mode_config.fb_lock); 200 drm_for_each_fb(drm_fb, &i915->drm) { 201 struct intel_framebuffer *fb = to_intel_framebuffer(drm_fb); 202 203 if (fb->dpt_vm) 204 i915_ggtt_resume_vm(fb->dpt_vm); 205 } 206 mutex_unlock(&i915->drm.mode_config.fb_lock); 207 } 208 209 /** 210 * intel_dpt_suspend - suspend the memory mapping for all DPT FBs during system suspend 211 * @i915: device instance 212 * 213 * Suspend the memory mapping during system suspend for all framebuffers which 214 * are mapped to HW via a GGTT->DPT page table. 215 * 216 * This function must be called before the mappings in GGTT are suspended calling 217 * i915_ggtt_suspend(). 218 */ 219 void intel_dpt_suspend(struct drm_i915_private *i915) 220 { 221 struct drm_framebuffer *drm_fb; 222 223 if (!HAS_DISPLAY(i915)) 224 return; 225 226 mutex_lock(&i915->drm.mode_config.fb_lock); 227 228 drm_for_each_fb(drm_fb, &i915->drm) { 229 struct intel_framebuffer *fb = to_intel_framebuffer(drm_fb); 230 231 if (fb->dpt_vm) 232 i915_ggtt_suspend_vm(fb->dpt_vm); 233 } 234 235 mutex_unlock(&i915->drm.mode_config.fb_lock); 236 } 237 238 struct i915_address_space * 239 intel_dpt_create(struct intel_framebuffer *fb) 240 { 241 struct drm_gem_object *obj = &intel_fb_obj(&fb->base)->base; 242 struct drm_i915_private *i915 = to_i915(obj->dev); 243 struct drm_i915_gem_object *dpt_obj; 244 struct i915_address_space *vm; 245 struct i915_dpt *dpt; 246 size_t size; 247 int ret; 248 249 if (intel_fb_needs_pot_stride_remap(fb)) 250 size = intel_remapped_info_size(&fb->remapped_view.gtt.remapped); 251 else 252 size = DIV_ROUND_UP_ULL(obj->size, I915_GTT_PAGE_SIZE); 253 254 size = round_up(size * sizeof(gen8_pte_t), I915_GTT_PAGE_SIZE); 255 256 dpt_obj = i915_gem_object_create_lmem(i915, size, I915_BO_ALLOC_CONTIGUOUS); 257 if (IS_ERR(dpt_obj) && i915_ggtt_has_aperture(to_gt(i915)->ggtt)) 258 dpt_obj = i915_gem_object_create_stolen(i915, size); 259 if (IS_ERR(dpt_obj) && !HAS_LMEM(i915)) { 260 drm_dbg_kms(&i915->drm, "Allocating dpt from smem\n"); 261 dpt_obj = i915_gem_object_create_internal(i915, size); 262 } 263 if (IS_ERR(dpt_obj)) 264 return ERR_CAST(dpt_obj); 265 266 ret = i915_gem_object_lock_interruptible(dpt_obj, NULL); 267 if (!ret) { 268 ret = i915_gem_object_set_cache_level(dpt_obj, I915_CACHE_NONE); 269 i915_gem_object_unlock(dpt_obj); 270 } 271 if (ret) { 272 i915_gem_object_put(dpt_obj); 273 return ERR_PTR(ret); 274 } 275 276 dpt = kzalloc(sizeof(*dpt), GFP_KERNEL); 277 if (!dpt) { 278 i915_gem_object_put(dpt_obj); 279 return ERR_PTR(-ENOMEM); 280 } 281 282 vm = &dpt->vm; 283 284 vm->gt = to_gt(i915); 285 vm->i915 = i915; 286 vm->dma = i915->drm.dev; 287 vm->total = (size / sizeof(gen8_pte_t)) * I915_GTT_PAGE_SIZE; 288 vm->is_dpt = true; 289 290 i915_address_space_init(vm, VM_CLASS_DPT); 291 292 vm->insert_page = dpt_insert_page; 293 vm->clear_range = dpt_clear_range; 294 vm->insert_entries = dpt_insert_entries; 295 vm->cleanup = dpt_cleanup; 296 297 vm->vma_ops.bind_vma = dpt_bind_vma; 298 vm->vma_ops.unbind_vma = dpt_unbind_vma; 299 300 vm->pte_encode = gen8_ggtt_pte_encode; 301 302 dpt->obj = dpt_obj; 303 304 return &dpt->vm; 305 } 306 307 void intel_dpt_destroy(struct i915_address_space *vm) 308 { 309 struct i915_dpt *dpt = i915_vm_to_dpt(vm); 310 311 i915_vm_put(&dpt->vm); 312 } 313