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 "gem/i915_gem_lmem.h"
9 #include "gt/gen8_ppgtt.h"
10 
11 #include "i915_drv.h"
12 #include "i915_reg.h"
13 #include "intel_de.h"
14 #include "intel_display_types.h"
15 #include "intel_dpt.h"
16 #include "intel_fb.h"
17 
18 struct i915_dpt {
19 	struct i915_address_space vm;
20 
21 	struct drm_i915_gem_object *obj;
22 	struct i915_vma *vma;
23 	void __iomem *iomem;
24 };
25 
26 #define i915_is_dpt(vm) ((vm)->is_dpt)
27 
28 static inline struct i915_dpt *
29 i915_vm_to_dpt(struct i915_address_space *vm)
30 {
31 	BUILD_BUG_ON(offsetof(struct i915_dpt, vm));
32 	GEM_BUG_ON(!i915_is_dpt(vm));
33 	return container_of(vm, struct i915_dpt, vm);
34 }
35 
36 #define dpt_total_entries(dpt) ((dpt)->vm.total >> PAGE_SHIFT)
37 
38 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
39 {
40 	writeq(pte, addr);
41 }
42 
43 static void dpt_insert_page(struct i915_address_space *vm,
44 			    dma_addr_t addr,
45 			    u64 offset,
46 			    enum i915_cache_level level,
47 			    u32 flags)
48 {
49 	struct i915_dpt *dpt = i915_vm_to_dpt(vm);
50 	gen8_pte_t __iomem *base = dpt->iomem;
51 
52 	gen8_set_pte(base + offset / I915_GTT_PAGE_SIZE,
53 		     vm->pte_encode(addr, level, flags));
54 }
55 
56 static void dpt_insert_entries(struct i915_address_space *vm,
57 			       struct i915_vma_resource *vma_res,
58 			       enum i915_cache_level level,
59 			       u32 flags)
60 {
61 	struct i915_dpt *dpt = i915_vm_to_dpt(vm);
62 	gen8_pte_t __iomem *base = dpt->iomem;
63 	const gen8_pte_t pte_encode = vm->pte_encode(0, level, flags);
64 	struct sgt_iter sgt_iter;
65 	dma_addr_t addr;
66 	int i;
67 
68 	/*
69 	 * Note that we ignore PTE_READ_ONLY here. The caller must be careful
70 	 * not to allow the user to override access to a read only page.
71 	 */
72 
73 	i = vma_res->start / I915_GTT_PAGE_SIZE;
74 	for_each_sgt_daddr(addr, sgt_iter, vma_res->bi.pages)
75 		gen8_set_pte(&base[i++], pte_encode | addr);
76 }
77 
78 static void dpt_clear_range(struct i915_address_space *vm,
79 			    u64 start, u64 length)
80 {
81 }
82 
83 static void dpt_bind_vma(struct i915_address_space *vm,
84 			 struct i915_vm_pt_stash *stash,
85 			 struct i915_vma_resource *vma_res,
86 			 enum i915_cache_level cache_level,
87 			 u32 flags)
88 {
89 	u32 pte_flags;
90 
91 	if (vma_res->bound_flags)
92 		return;
93 
94 	/* Applicable to VLV (gen8+ do not support RO in the GGTT) */
95 	pte_flags = 0;
96 	if (vm->has_read_only && vma_res->bi.readonly)
97 		pte_flags |= PTE_READ_ONLY;
98 	if (vma_res->bi.lmem)
99 		pte_flags |= PTE_LM;
100 
101 	vm->insert_entries(vm, vma_res, cache_level, pte_flags);
102 
103 	vma_res->page_sizes_gtt = I915_GTT_PAGE_SIZE;
104 
105 	/*
106 	 * Without aliasing PPGTT there's no difference between
107 	 * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
108 	 * upgrade to both bound if we bind either to avoid double-binding.
109 	 */
110 	vma_res->bound_flags = I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
111 }
112 
113 static void dpt_unbind_vma(struct i915_address_space *vm,
114 			   struct i915_vma_resource *vma_res)
115 {
116 	vm->clear_range(vm, vma_res->start, vma_res->vma_size);
117 }
118 
119 static void dpt_cleanup(struct i915_address_space *vm)
120 {
121 	struct i915_dpt *dpt = i915_vm_to_dpt(vm);
122 
123 	i915_gem_object_put(dpt->obj);
124 }
125 
126 struct i915_vma *intel_dpt_pin(struct i915_address_space *vm)
127 {
128 	struct drm_i915_private *i915 = vm->i915;
129 	struct i915_dpt *dpt = i915_vm_to_dpt(vm);
130 	intel_wakeref_t wakeref;
131 	struct i915_vma *vma;
132 	void __iomem *iomem;
133 	struct i915_gem_ww_ctx ww;
134 	u64 pin_flags = 0;
135 	int err;
136 
137 	if (i915_gem_object_is_stolen(dpt->obj))
138 		pin_flags |= PIN_MAPPABLE;
139 
140 	wakeref = intel_runtime_pm_get(&i915->runtime_pm);
141 	atomic_inc(&i915->gpu_error.pending_fb_pin);
142 
143 	for_i915_gem_ww(&ww, err, true) {
144 		err = i915_gem_object_lock(dpt->obj, &ww);
145 		if (err)
146 			continue;
147 
148 		vma = i915_gem_object_ggtt_pin_ww(dpt->obj, &ww, NULL, 0, 4096,
149 						  pin_flags);
150 		if (IS_ERR(vma)) {
151 			err = PTR_ERR(vma);
152 			continue;
153 		}
154 
155 		iomem = i915_vma_pin_iomap(vma);
156 		i915_vma_unpin(vma);
157 
158 		if (IS_ERR(iomem)) {
159 			err = PTR_ERR(iomem);
160 			continue;
161 		}
162 
163 		dpt->vma = vma;
164 		dpt->iomem = iomem;
165 
166 		i915_vma_get(vma);
167 	}
168 
169 	atomic_dec(&i915->gpu_error.pending_fb_pin);
170 	intel_runtime_pm_put(&i915->runtime_pm, wakeref);
171 
172 	return err ? ERR_PTR(err) : vma;
173 }
174 
175 void intel_dpt_unpin(struct i915_address_space *vm)
176 {
177 	struct i915_dpt *dpt = i915_vm_to_dpt(vm);
178 
179 	i915_vma_unpin_iomap(dpt->vma);
180 	i915_vma_put(dpt->vma);
181 }
182 
183 /**
184  * intel_dpt_resume - restore the memory mapping for all DPT FBs during system resume
185  * @i915: device instance
186  *
187  * Restore the memory mapping during system resume for all framebuffers which
188  * are mapped to HW via a GGTT->DPT page table. The content of these page
189  * tables are not stored in the hibernation image during S4 and S3RST->S4
190  * transitions, so here we reprogram the PTE entries in those tables.
191  *
192  * This function must be called after the mappings in GGTT have been restored calling
193  * i915_ggtt_resume().
194  */
195 void intel_dpt_resume(struct drm_i915_private *i915)
196 {
197 	struct drm_framebuffer *drm_fb;
198 
199 	if (!HAS_DISPLAY(i915))
200 		return;
201 
202 	mutex_lock(&i915->drm.mode_config.fb_lock);
203 	drm_for_each_fb(drm_fb, &i915->drm) {
204 		struct intel_framebuffer *fb = to_intel_framebuffer(drm_fb);
205 
206 		if (fb->dpt_vm)
207 			i915_ggtt_resume_vm(fb->dpt_vm);
208 	}
209 	mutex_unlock(&i915->drm.mode_config.fb_lock);
210 }
211 
212 /**
213  * intel_dpt_suspend - suspend the memory mapping for all DPT FBs during system suspend
214  * @i915: device instance
215  *
216  * Suspend the memory mapping during system suspend for all framebuffers which
217  * are mapped to HW via a GGTT->DPT page table.
218  *
219  * This function must be called before the mappings in GGTT are suspended calling
220  * i915_ggtt_suspend().
221  */
222 void intel_dpt_suspend(struct drm_i915_private *i915)
223 {
224 	struct drm_framebuffer *drm_fb;
225 
226 	if (!HAS_DISPLAY(i915))
227 		return;
228 
229 	mutex_lock(&i915->drm.mode_config.fb_lock);
230 
231 	drm_for_each_fb(drm_fb, &i915->drm) {
232 		struct intel_framebuffer *fb = to_intel_framebuffer(drm_fb);
233 
234 		if (fb->dpt_vm)
235 			i915_ggtt_suspend_vm(fb->dpt_vm);
236 	}
237 
238 	mutex_unlock(&i915->drm.mode_config.fb_lock);
239 }
240 
241 struct i915_address_space *
242 intel_dpt_create(struct intel_framebuffer *fb)
243 {
244 	struct drm_gem_object *obj = &intel_fb_obj(&fb->base)->base;
245 	struct drm_i915_private *i915 = to_i915(obj->dev);
246 	struct drm_i915_gem_object *dpt_obj;
247 	struct i915_address_space *vm;
248 	struct i915_dpt *dpt;
249 	size_t size;
250 	int ret;
251 
252 	if (intel_fb_needs_pot_stride_remap(fb))
253 		size = intel_remapped_info_size(&fb->remapped_view.gtt.remapped);
254 	else
255 		size = DIV_ROUND_UP_ULL(obj->size, I915_GTT_PAGE_SIZE);
256 
257 	size = round_up(size * sizeof(gen8_pte_t), I915_GTT_PAGE_SIZE);
258 
259 	dpt_obj = i915_gem_object_create_lmem(i915, size, I915_BO_ALLOC_CONTIGUOUS);
260 	if (IS_ERR(dpt_obj) && i915_ggtt_has_aperture(to_gt(i915)->ggtt))
261 		dpt_obj = i915_gem_object_create_stolen(i915, size);
262 	if (IS_ERR(dpt_obj) && !HAS_LMEM(i915)) {
263 		drm_dbg_kms(&i915->drm, "Allocating dpt from smem\n");
264 		dpt_obj = i915_gem_object_create_internal(i915, size);
265 	}
266 	if (IS_ERR(dpt_obj))
267 		return ERR_CAST(dpt_obj);
268 
269 	ret = i915_gem_object_lock_interruptible(dpt_obj, NULL);
270 	if (!ret) {
271 		ret = i915_gem_object_set_cache_level(dpt_obj, I915_CACHE_NONE);
272 		i915_gem_object_unlock(dpt_obj);
273 	}
274 	if (ret) {
275 		i915_gem_object_put(dpt_obj);
276 		return ERR_PTR(ret);
277 	}
278 
279 	dpt = kzalloc(sizeof(*dpt), GFP_KERNEL);
280 	if (!dpt) {
281 		i915_gem_object_put(dpt_obj);
282 		return ERR_PTR(-ENOMEM);
283 	}
284 
285 	vm = &dpt->vm;
286 
287 	vm->gt = to_gt(i915);
288 	vm->i915 = i915;
289 	vm->dma = i915->drm.dev;
290 	vm->total = (size / sizeof(gen8_pte_t)) * I915_GTT_PAGE_SIZE;
291 	vm->is_dpt = true;
292 
293 	i915_address_space_init(vm, VM_CLASS_DPT);
294 
295 	vm->insert_page = dpt_insert_page;
296 	vm->clear_range = dpt_clear_range;
297 	vm->insert_entries = dpt_insert_entries;
298 	vm->cleanup = dpt_cleanup;
299 
300 	vm->vma_ops.bind_vma    = dpt_bind_vma;
301 	vm->vma_ops.unbind_vma  = dpt_unbind_vma;
302 
303 	vm->pte_encode = gen8_ggtt_pte_encode;
304 
305 	dpt->obj = dpt_obj;
306 	dpt->obj->is_dpt = true;
307 
308 	return &dpt->vm;
309 }
310 
311 void intel_dpt_destroy(struct i915_address_space *vm)
312 {
313 	struct i915_dpt *dpt = i915_vm_to_dpt(vm);
314 
315 	dpt->obj->is_dpt = false;
316 	i915_vm_put(&dpt->vm);
317 }
318 
319 void intel_dpt_configure(struct intel_crtc *crtc)
320 {
321 	struct drm_i915_private *i915 = to_i915(crtc->base.dev);
322 
323 	if (DISPLAY_VER(i915) == 14) {
324 		enum pipe pipe = crtc->pipe;
325 		enum plane_id plane_id;
326 
327 		for_each_plane_id_on_crtc(crtc, plane_id) {
328 			if (plane_id == PLANE_CURSOR)
329 				continue;
330 
331 			intel_de_rmw(i915, PLANE_CHICKEN(pipe, plane_id),
332 				     PLANE_CHICKEN_DISABLE_DPT,
333 				     i915->params.enable_dpt ? 0 : PLANE_CHICKEN_DISABLE_DPT);
334 		}
335 	} else if (DISPLAY_VER(i915) == 13) {
336 		intel_de_rmw(i915, CHICKEN_MISC_2,
337 			     CHICKEN_MISC_DISABLE_DPT,
338 			     i915->params.enable_dpt ? 0 : CHICKEN_MISC_DISABLE_DPT);
339 	}
340 }
341