xref: /linux/drivers/gpu/drm/i915/gt/gen6_ppgtt.c (revision fd7d598270724cc787982ea48bbe17ad383a8b7f)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include <linux/log2.h>
7 
8 #include "gem/i915_gem_internal.h"
9 
10 #include "gen6_ppgtt.h"
11 #include "i915_scatterlist.h"
12 #include "i915_trace.h"
13 #include "i915_vgpu.h"
14 #include "intel_gt_regs.h"
15 #include "intel_engine_regs.h"
16 #include "intel_gt.h"
17 
18 /* Write pde (index) from the page directory @pd to the page table @pt */
19 static void gen6_write_pde(const struct gen6_ppgtt *ppgtt,
20 			   const unsigned int pde,
21 			   const struct i915_page_table *pt)
22 {
23 	dma_addr_t addr = pt ? px_dma(pt) : px_dma(ppgtt->base.vm.scratch[1]);
24 
25 	/* Caller needs to make sure the write completes if necessary */
26 	iowrite32(GEN6_PDE_ADDR_ENCODE(addr) | GEN6_PDE_VALID,
27 		  ppgtt->pd_addr + pde);
28 }
29 
30 void gen7_ppgtt_enable(struct intel_gt *gt)
31 {
32 	struct drm_i915_private *i915 = gt->i915;
33 	struct intel_uncore *uncore = gt->uncore;
34 	u32 ecochk;
35 
36 	intel_uncore_rmw(uncore, GAC_ECO_BITS, 0, ECOBITS_PPGTT_CACHE64B);
37 
38 	ecochk = intel_uncore_read(uncore, GAM_ECOCHK);
39 	if (IS_HASWELL(i915)) {
40 		ecochk |= ECOCHK_PPGTT_WB_HSW;
41 	} else {
42 		ecochk |= ECOCHK_PPGTT_LLC_IVB;
43 		ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
44 	}
45 	intel_uncore_write(uncore, GAM_ECOCHK, ecochk);
46 }
47 
48 void gen6_ppgtt_enable(struct intel_gt *gt)
49 {
50 	struct intel_uncore *uncore = gt->uncore;
51 
52 	intel_uncore_rmw(uncore,
53 			 GAC_ECO_BITS,
54 			 0,
55 			 ECOBITS_SNB_BIT | ECOBITS_PPGTT_CACHE64B);
56 
57 	intel_uncore_rmw(uncore,
58 			 GAB_CTL,
59 			 0,
60 			 GAB_CTL_CONT_AFTER_PAGEFAULT);
61 
62 	intel_uncore_rmw(uncore,
63 			 GAM_ECOCHK,
64 			 0,
65 			 ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
66 
67 	if (HAS_PPGTT(uncore->i915)) /* may be disabled for VT-d */
68 		intel_uncore_write(uncore,
69 				   GFX_MODE,
70 				   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
71 }
72 
73 /* PPGTT support for Sandybdrige/Gen6 and later */
74 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
75 				   u64 start, u64 length)
76 {
77 	struct gen6_ppgtt * const ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
78 	const unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
79 	const gen6_pte_t scratch_pte = vm->scratch[0]->encode;
80 	unsigned int pde = first_entry / GEN6_PTES;
81 	unsigned int pte = first_entry % GEN6_PTES;
82 	unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
83 
84 	while (num_entries) {
85 		struct i915_page_table * const pt =
86 			i915_pt_entry(ppgtt->base.pd, pde++);
87 		const unsigned int count = min(num_entries, GEN6_PTES - pte);
88 		gen6_pte_t *vaddr;
89 
90 		num_entries -= count;
91 
92 		GEM_BUG_ON(count > atomic_read(&pt->used));
93 		if (!atomic_sub_return(count, &pt->used))
94 			ppgtt->scan_for_unused_pt = true;
95 
96 		/*
97 		 * Note that the hw doesn't support removing PDE on the fly
98 		 * (they are cached inside the context with no means to
99 		 * invalidate the cache), so we can only reset the PTE
100 		 * entries back to scratch.
101 		 */
102 
103 		vaddr = px_vaddr(pt);
104 		memset32(vaddr + pte, scratch_pte, count);
105 
106 		pte = 0;
107 	}
108 }
109 
110 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
111 				      struct i915_vma_resource *vma_res,
112 				      unsigned int pat_index,
113 				      u32 flags)
114 {
115 	struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
116 	struct i915_page_directory * const pd = ppgtt->pd;
117 	unsigned int first_entry = vma_res->start / I915_GTT_PAGE_SIZE;
118 	unsigned int act_pt = first_entry / GEN6_PTES;
119 	unsigned int act_pte = first_entry % GEN6_PTES;
120 	const u32 pte_encode = vm->pte_encode(0, pat_index, flags);
121 	struct sgt_dma iter = sgt_dma(vma_res);
122 	gen6_pte_t *vaddr;
123 
124 	GEM_BUG_ON(!pd->entry[act_pt]);
125 
126 	vaddr = px_vaddr(i915_pt_entry(pd, act_pt));
127 	do {
128 		GEM_BUG_ON(sg_dma_len(iter.sg) < I915_GTT_PAGE_SIZE);
129 		vaddr[act_pte] = pte_encode | GEN6_PTE_ADDR_ENCODE(iter.dma);
130 
131 		iter.dma += I915_GTT_PAGE_SIZE;
132 		if (iter.dma == iter.max) {
133 			iter.sg = __sg_next(iter.sg);
134 			if (!iter.sg || sg_dma_len(iter.sg) == 0)
135 				break;
136 
137 			iter.dma = sg_dma_address(iter.sg);
138 			iter.max = iter.dma + sg_dma_len(iter.sg);
139 		}
140 
141 		if (++act_pte == GEN6_PTES) {
142 			vaddr = px_vaddr(i915_pt_entry(pd, ++act_pt));
143 			act_pte = 0;
144 		}
145 	} while (1);
146 
147 	vma_res->page_sizes_gtt = I915_GTT_PAGE_SIZE;
148 }
149 
150 static void gen6_flush_pd(struct gen6_ppgtt *ppgtt, u64 start, u64 end)
151 {
152 	struct i915_page_directory * const pd = ppgtt->base.pd;
153 	struct i915_page_table *pt;
154 	unsigned int pde;
155 
156 	start = round_down(start, SZ_64K);
157 	end = round_up(end, SZ_64K) - start;
158 
159 	mutex_lock(&ppgtt->flush);
160 
161 	gen6_for_each_pde(pt, pd, start, end, pde)
162 		gen6_write_pde(ppgtt, pde, pt);
163 
164 	mb();
165 	ioread32(ppgtt->pd_addr + pde - 1);
166 	gen6_ggtt_invalidate(ppgtt->base.vm.gt->ggtt);
167 	mb();
168 
169 	mutex_unlock(&ppgtt->flush);
170 }
171 
172 static void gen6_alloc_va_range(struct i915_address_space *vm,
173 				struct i915_vm_pt_stash *stash,
174 				u64 start, u64 length)
175 {
176 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
177 	struct i915_page_directory * const pd = ppgtt->base.pd;
178 	struct i915_page_table *pt;
179 	bool flush = false;
180 	u64 from = start;
181 	unsigned int pde;
182 
183 	spin_lock(&pd->lock);
184 	gen6_for_each_pde(pt, pd, start, length, pde) {
185 		const unsigned int count = gen6_pte_count(start, length);
186 
187 		if (!pt) {
188 			spin_unlock(&pd->lock);
189 
190 			pt = stash->pt[0];
191 			__i915_gem_object_pin_pages(pt->base);
192 
193 			fill32_px(pt, vm->scratch[0]->encode);
194 
195 			spin_lock(&pd->lock);
196 			if (!pd->entry[pde]) {
197 				stash->pt[0] = pt->stash;
198 				atomic_set(&pt->used, 0);
199 				pd->entry[pde] = pt;
200 			} else {
201 				pt = pd->entry[pde];
202 			}
203 
204 			flush = true;
205 		}
206 
207 		atomic_add(count, &pt->used);
208 	}
209 	spin_unlock(&pd->lock);
210 
211 	if (flush && i915_vma_is_bound(ppgtt->vma, I915_VMA_GLOBAL_BIND)) {
212 		intel_wakeref_t wakeref;
213 
214 		with_intel_runtime_pm(&vm->i915->runtime_pm, wakeref)
215 			gen6_flush_pd(ppgtt, from, start);
216 	}
217 }
218 
219 static int gen6_ppgtt_init_scratch(struct gen6_ppgtt *ppgtt)
220 {
221 	struct i915_address_space * const vm = &ppgtt->base.vm;
222 	int ret;
223 
224 	ret = setup_scratch_page(vm);
225 	if (ret)
226 		return ret;
227 
228 	vm->scratch[0]->encode =
229 		vm->pte_encode(px_dma(vm->scratch[0]),
230 			       i915_gem_get_pat_index(vm->i915,
231 						      I915_CACHE_NONE),
232 			       PTE_READ_ONLY);
233 
234 	vm->scratch[1] = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
235 	if (IS_ERR(vm->scratch[1])) {
236 		ret = PTR_ERR(vm->scratch[1]);
237 		goto err_scratch0;
238 	}
239 
240 	ret = map_pt_dma(vm, vm->scratch[1]);
241 	if (ret)
242 		goto err_scratch1;
243 
244 	fill32_px(vm->scratch[1], vm->scratch[0]->encode);
245 
246 	return 0;
247 
248 err_scratch1:
249 	i915_gem_object_put(vm->scratch[1]);
250 err_scratch0:
251 	i915_gem_object_put(vm->scratch[0]);
252 	vm->scratch[0] = NULL;
253 	return ret;
254 }
255 
256 static void gen6_ppgtt_free_pd(struct gen6_ppgtt *ppgtt)
257 {
258 	struct i915_page_directory * const pd = ppgtt->base.pd;
259 	struct i915_page_table *pt;
260 	u32 pde;
261 
262 	gen6_for_all_pdes(pt, pd, pde)
263 		if (pt)
264 			free_pt(&ppgtt->base.vm, pt);
265 }
266 
267 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
268 {
269 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
270 
271 	gen6_ppgtt_free_pd(ppgtt);
272 	free_scratch(vm);
273 
274 	if (ppgtt->base.pd)
275 		free_pd(&ppgtt->base.vm, ppgtt->base.pd);
276 
277 	mutex_destroy(&ppgtt->flush);
278 }
279 
280 static void pd_vma_bind(struct i915_address_space *vm,
281 			struct i915_vm_pt_stash *stash,
282 			struct i915_vma_resource *vma_res,
283 			unsigned int pat_index,
284 			u32 unused)
285 {
286 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
287 	struct gen6_ppgtt *ppgtt = vma_res->private;
288 	u32 ggtt_offset = vma_res->start / I915_GTT_PAGE_SIZE;
289 
290 	ppgtt->pp_dir = ggtt_offset * sizeof(gen6_pte_t) << 10;
291 	ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm + ggtt_offset;
292 
293 	gen6_flush_pd(ppgtt, 0, ppgtt->base.vm.total);
294 }
295 
296 static void pd_vma_unbind(struct i915_address_space *vm,
297 			  struct i915_vma_resource *vma_res)
298 {
299 	struct gen6_ppgtt *ppgtt = vma_res->private;
300 	struct i915_page_directory * const pd = ppgtt->base.pd;
301 	struct i915_page_table *pt;
302 	unsigned int pde;
303 
304 	if (!ppgtt->scan_for_unused_pt)
305 		return;
306 
307 	/* Free all no longer used page tables */
308 	gen6_for_all_pdes(pt, ppgtt->base.pd, pde) {
309 		if (!pt || atomic_read(&pt->used))
310 			continue;
311 
312 		free_pt(&ppgtt->base.vm, pt);
313 		pd->entry[pde] = NULL;
314 	}
315 
316 	ppgtt->scan_for_unused_pt = false;
317 }
318 
319 static const struct i915_vma_ops pd_vma_ops = {
320 	.bind_vma = pd_vma_bind,
321 	.unbind_vma = pd_vma_unbind,
322 };
323 
324 int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww)
325 {
326 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
327 	int err;
328 
329 	GEM_BUG_ON(!kref_read(&ppgtt->base.vm.ref));
330 
331 	/*
332 	 * Workaround the limited maximum vma->pin_count and the aliasing_ppgtt
333 	 * which will be pinned into every active context.
334 	 * (When vma->pin_count becomes atomic, I expect we will naturally
335 	 * need a larger, unpacked, type and kill this redundancy.)
336 	 */
337 	if (atomic_add_unless(&ppgtt->pin_count, 1, 0))
338 		return 0;
339 
340 	/* grab the ppgtt resv to pin the object */
341 	err = i915_vm_lock_objects(&ppgtt->base.vm, ww);
342 	if (err)
343 		return err;
344 
345 	/*
346 	 * PPGTT PDEs reside in the GGTT and consists of 512 entries. The
347 	 * allocator works in address space sizes, so it's multiplied by page
348 	 * size. We allocate at the top of the GTT to avoid fragmentation.
349 	 */
350 	if (!atomic_read(&ppgtt->pin_count)) {
351 		err = i915_ggtt_pin(ppgtt->vma, ww, GEN6_PD_ALIGN, PIN_HIGH);
352 
353 		GEM_BUG_ON(ppgtt->vma->fence);
354 		clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(ppgtt->vma));
355 	}
356 	if (!err)
357 		atomic_inc(&ppgtt->pin_count);
358 
359 	return err;
360 }
361 
362 static int pd_dummy_obj_get_pages(struct drm_i915_gem_object *obj)
363 {
364 	obj->mm.pages = ZERO_SIZE_PTR;
365 	return 0;
366 }
367 
368 static void pd_dummy_obj_put_pages(struct drm_i915_gem_object *obj,
369 				   struct sg_table *pages)
370 {
371 }
372 
373 static const struct drm_i915_gem_object_ops pd_dummy_obj_ops = {
374 	.name = "pd_dummy_obj",
375 	.get_pages = pd_dummy_obj_get_pages,
376 	.put_pages = pd_dummy_obj_put_pages,
377 };
378 
379 static struct i915_page_directory *
380 gen6_alloc_top_pd(struct gen6_ppgtt *ppgtt)
381 {
382 	struct i915_ggtt * const ggtt = ppgtt->base.vm.gt->ggtt;
383 	struct i915_page_directory *pd;
384 	int err;
385 
386 	pd = __alloc_pd(I915_PDES);
387 	if (unlikely(!pd))
388 		return ERR_PTR(-ENOMEM);
389 
390 	pd->pt.base = __i915_gem_object_create_internal(ppgtt->base.vm.gt->i915,
391 							&pd_dummy_obj_ops,
392 							I915_PDES * SZ_4K);
393 	if (IS_ERR(pd->pt.base)) {
394 		err = PTR_ERR(pd->pt.base);
395 		pd->pt.base = NULL;
396 		goto err_pd;
397 	}
398 
399 	pd->pt.base->base.resv = i915_vm_resv_get(&ppgtt->base.vm);
400 	pd->pt.base->shares_resv_from = &ppgtt->base.vm;
401 
402 	ppgtt->vma = i915_vma_instance(pd->pt.base, &ggtt->vm, NULL);
403 	if (IS_ERR(ppgtt->vma)) {
404 		err = PTR_ERR(ppgtt->vma);
405 		ppgtt->vma = NULL;
406 		goto err_pd;
407 	}
408 
409 	/* The dummy object we create is special, override ops.. */
410 	ppgtt->vma->ops = &pd_vma_ops;
411 	ppgtt->vma->private = ppgtt;
412 	return pd;
413 
414 err_pd:
415 	free_pd(&ppgtt->base.vm, pd);
416 	return ERR_PTR(err);
417 }
418 
419 void gen6_ppgtt_unpin(struct i915_ppgtt *base)
420 {
421 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
422 
423 	GEM_BUG_ON(!atomic_read(&ppgtt->pin_count));
424 	if (atomic_dec_and_test(&ppgtt->pin_count))
425 		i915_vma_unpin(ppgtt->vma);
426 }
427 
428 struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt)
429 {
430 	struct i915_ggtt * const ggtt = gt->ggtt;
431 	struct gen6_ppgtt *ppgtt;
432 	int err;
433 
434 	ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
435 	if (!ppgtt)
436 		return ERR_PTR(-ENOMEM);
437 
438 	mutex_init(&ppgtt->flush);
439 
440 	ppgtt_init(&ppgtt->base, gt, 0);
441 	ppgtt->base.vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen6_pte_t));
442 	ppgtt->base.vm.top = 1;
443 
444 	ppgtt->base.vm.bind_async_flags = I915_VMA_LOCAL_BIND;
445 	ppgtt->base.vm.allocate_va_range = gen6_alloc_va_range;
446 	ppgtt->base.vm.clear_range = gen6_ppgtt_clear_range;
447 	ppgtt->base.vm.insert_entries = gen6_ppgtt_insert_entries;
448 	ppgtt->base.vm.cleanup = gen6_ppgtt_cleanup;
449 
450 	ppgtt->base.vm.alloc_pt_dma = alloc_pt_dma;
451 	ppgtt->base.vm.alloc_scratch_dma = alloc_pt_dma;
452 	ppgtt->base.vm.pte_encode = ggtt->vm.pte_encode;
453 
454 	err = gen6_ppgtt_init_scratch(ppgtt);
455 	if (err)
456 		goto err_put;
457 
458 	ppgtt->base.pd = gen6_alloc_top_pd(ppgtt);
459 	if (IS_ERR(ppgtt->base.pd)) {
460 		err = PTR_ERR(ppgtt->base.pd);
461 		goto err_put;
462 	}
463 
464 	return &ppgtt->base;
465 
466 err_put:
467 	i915_vm_put(&ppgtt->base.vm);
468 	return ERR_PTR(err);
469 }
470