xref: /openbmc/linux/drivers/gpu/drm/i915/gt/gen6_ppgtt.c (revision 2326f3cd)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2020 Intel Corporation
4  */
5 
6 #include <linux/log2.h>
7 
8 #include "gen6_ppgtt.h"
9 #include "i915_scatterlist.h"
10 #include "i915_trace.h"
11 #include "i915_vgpu.h"
12 #include "intel_gt.h"
13 
14 /* Write pde (index) from the page directory @pd to the page table @pt */
15 static void gen6_write_pde(const struct gen6_ppgtt *ppgtt,
16 			   const unsigned int pde,
17 			   const struct i915_page_table *pt)
18 {
19 	dma_addr_t addr = pt ? px_dma(pt) : px_dma(ppgtt->base.vm.scratch[1]);
20 
21 	/* Caller needs to make sure the write completes if necessary */
22 	iowrite32(GEN6_PDE_ADDR_ENCODE(addr) | GEN6_PDE_VALID,
23 		  ppgtt->pd_addr + pde);
24 }
25 
26 void gen7_ppgtt_enable(struct intel_gt *gt)
27 {
28 	struct drm_i915_private *i915 = gt->i915;
29 	struct intel_uncore *uncore = gt->uncore;
30 	u32 ecochk;
31 
32 	intel_uncore_rmw(uncore, GAC_ECO_BITS, 0, ECOBITS_PPGTT_CACHE64B);
33 
34 	ecochk = intel_uncore_read(uncore, GAM_ECOCHK);
35 	if (IS_HASWELL(i915)) {
36 		ecochk |= ECOCHK_PPGTT_WB_HSW;
37 	} else {
38 		ecochk |= ECOCHK_PPGTT_LLC_IVB;
39 		ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
40 	}
41 	intel_uncore_write(uncore, GAM_ECOCHK, ecochk);
42 }
43 
44 void gen6_ppgtt_enable(struct intel_gt *gt)
45 {
46 	struct intel_uncore *uncore = gt->uncore;
47 
48 	intel_uncore_rmw(uncore,
49 			 GAC_ECO_BITS,
50 			 0,
51 			 ECOBITS_SNB_BIT | ECOBITS_PPGTT_CACHE64B);
52 
53 	intel_uncore_rmw(uncore,
54 			 GAB_CTL,
55 			 0,
56 			 GAB_CTL_CONT_AFTER_PAGEFAULT);
57 
58 	intel_uncore_rmw(uncore,
59 			 GAM_ECOCHK,
60 			 0,
61 			 ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
62 
63 	if (HAS_PPGTT(uncore->i915)) /* may be disabled for VT-d */
64 		intel_uncore_write(uncore,
65 				   GFX_MODE,
66 				   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
67 }
68 
69 /* PPGTT support for Sandybdrige/Gen6 and later */
70 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
71 				   u64 start, u64 length)
72 {
73 	struct gen6_ppgtt * const ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
74 	const unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
75 	const gen6_pte_t scratch_pte = vm->scratch[0]->encode;
76 	unsigned int pde = first_entry / GEN6_PTES;
77 	unsigned int pte = first_entry % GEN6_PTES;
78 	unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
79 
80 	while (num_entries) {
81 		struct i915_page_table * const pt =
82 			i915_pt_entry(ppgtt->base.pd, pde++);
83 		const unsigned int count = min(num_entries, GEN6_PTES - pte);
84 		gen6_pte_t *vaddr;
85 
86 		num_entries -= count;
87 
88 		GEM_BUG_ON(count > atomic_read(&pt->used));
89 		if (!atomic_sub_return(count, &pt->used))
90 			ppgtt->scan_for_unused_pt = true;
91 
92 		/*
93 		 * Note that the hw doesn't support removing PDE on the fly
94 		 * (they are cached inside the context with no means to
95 		 * invalidate the cache), so we can only reset the PTE
96 		 * entries back to scratch.
97 		 */
98 
99 		vaddr = px_vaddr(pt);
100 		memset32(vaddr + pte, scratch_pte, count);
101 
102 		pte = 0;
103 	}
104 }
105 
106 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
107 				      struct i915_vma *vma,
108 				      enum i915_cache_level cache_level,
109 				      u32 flags)
110 {
111 	struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
112 	struct i915_page_directory * const pd = ppgtt->pd;
113 	unsigned int first_entry = vma->node.start / I915_GTT_PAGE_SIZE;
114 	unsigned int act_pt = first_entry / GEN6_PTES;
115 	unsigned int act_pte = first_entry % GEN6_PTES;
116 	const u32 pte_encode = vm->pte_encode(0, cache_level, flags);
117 	struct sgt_dma iter = sgt_dma(vma);
118 	gen6_pte_t *vaddr;
119 
120 	GEM_BUG_ON(!pd->entry[act_pt]);
121 
122 	vaddr = px_vaddr(i915_pt_entry(pd, act_pt));
123 	do {
124 		GEM_BUG_ON(sg_dma_len(iter.sg) < I915_GTT_PAGE_SIZE);
125 		vaddr[act_pte] = pte_encode | GEN6_PTE_ADDR_ENCODE(iter.dma);
126 
127 		iter.dma += I915_GTT_PAGE_SIZE;
128 		if (iter.dma == iter.max) {
129 			iter.sg = __sg_next(iter.sg);
130 			if (!iter.sg || sg_dma_len(iter.sg) == 0)
131 				break;
132 
133 			iter.dma = sg_dma_address(iter.sg);
134 			iter.max = iter.dma + sg_dma_len(iter.sg);
135 		}
136 
137 		if (++act_pte == GEN6_PTES) {
138 			vaddr = px_vaddr(i915_pt_entry(pd, ++act_pt));
139 			act_pte = 0;
140 		}
141 	} while (1);
142 
143 	vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
144 }
145 
146 static void gen6_flush_pd(struct gen6_ppgtt *ppgtt, u64 start, u64 end)
147 {
148 	struct i915_page_directory * const pd = ppgtt->base.pd;
149 	struct i915_page_table *pt;
150 	unsigned int pde;
151 
152 	start = round_down(start, SZ_64K);
153 	end = round_up(end, SZ_64K) - start;
154 
155 	mutex_lock(&ppgtt->flush);
156 
157 	gen6_for_each_pde(pt, pd, start, end, pde)
158 		gen6_write_pde(ppgtt, pde, pt);
159 
160 	mb();
161 	ioread32(ppgtt->pd_addr + pde - 1);
162 	gen6_ggtt_invalidate(ppgtt->base.vm.gt->ggtt);
163 	mb();
164 
165 	mutex_unlock(&ppgtt->flush);
166 }
167 
168 static void gen6_alloc_va_range(struct i915_address_space *vm,
169 				struct i915_vm_pt_stash *stash,
170 				u64 start, u64 length)
171 {
172 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
173 	struct i915_page_directory * const pd = ppgtt->base.pd;
174 	struct i915_page_table *pt;
175 	bool flush = false;
176 	u64 from = start;
177 	unsigned int pde;
178 
179 	spin_lock(&pd->lock);
180 	gen6_for_each_pde(pt, pd, start, length, pde) {
181 		const unsigned int count = gen6_pte_count(start, length);
182 
183 		if (!pt) {
184 			spin_unlock(&pd->lock);
185 
186 			pt = stash->pt[0];
187 			__i915_gem_object_pin_pages(pt->base);
188 			i915_gem_object_make_unshrinkable(pt->base);
189 
190 			fill32_px(pt, vm->scratch[0]->encode);
191 
192 			spin_lock(&pd->lock);
193 			if (!pd->entry[pde]) {
194 				stash->pt[0] = pt->stash;
195 				atomic_set(&pt->used, 0);
196 				pd->entry[pde] = pt;
197 			} else {
198 				pt = pd->entry[pde];
199 			}
200 
201 			flush = true;
202 		}
203 
204 		atomic_add(count, &pt->used);
205 	}
206 	spin_unlock(&pd->lock);
207 
208 	if (flush && i915_vma_is_bound(ppgtt->vma, I915_VMA_GLOBAL_BIND)) {
209 		intel_wakeref_t wakeref;
210 
211 		with_intel_runtime_pm(&vm->i915->runtime_pm, wakeref)
212 			gen6_flush_pd(ppgtt, from, start);
213 	}
214 }
215 
216 static int gen6_ppgtt_init_scratch(struct gen6_ppgtt *ppgtt)
217 {
218 	struct i915_address_space * const vm = &ppgtt->base.vm;
219 	int ret;
220 
221 	ret = setup_scratch_page(vm);
222 	if (ret)
223 		return ret;
224 
225 	vm->scratch[0]->encode =
226 		vm->pte_encode(px_dma(vm->scratch[0]),
227 			       I915_CACHE_NONE, PTE_READ_ONLY);
228 
229 	vm->scratch[1] = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
230 	if (IS_ERR(vm->scratch[1])) {
231 		ret = PTR_ERR(vm->scratch[1]);
232 		goto err_scratch0;
233 	}
234 
235 	ret = map_pt_dma(vm, vm->scratch[1]);
236 	if (ret)
237 		goto err_scratch1;
238 
239 	fill32_px(vm->scratch[1], vm->scratch[0]->encode);
240 
241 	return 0;
242 
243 err_scratch1:
244 	i915_gem_object_put(vm->scratch[1]);
245 err_scratch0:
246 	i915_gem_object_put(vm->scratch[0]);
247 	return ret;
248 }
249 
250 static void gen6_ppgtt_free_pd(struct gen6_ppgtt *ppgtt)
251 {
252 	struct i915_page_directory * const pd = ppgtt->base.pd;
253 	struct i915_page_table *pt;
254 	u32 pde;
255 
256 	gen6_for_all_pdes(pt, pd, pde)
257 		if (pt)
258 			free_pt(&ppgtt->base.vm, pt);
259 }
260 
261 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
262 {
263 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
264 
265 	__i915_vma_put(ppgtt->vma);
266 
267 	gen6_ppgtt_free_pd(ppgtt);
268 	free_scratch(vm);
269 
270 	mutex_destroy(&ppgtt->flush);
271 	mutex_destroy(&ppgtt->pin_mutex);
272 
273 	free_pd(&ppgtt->base.vm, ppgtt->base.pd);
274 }
275 
276 static int pd_vma_set_pages(struct i915_vma *vma)
277 {
278 	vma->pages = ERR_PTR(-ENODEV);
279 	return 0;
280 }
281 
282 static void pd_vma_clear_pages(struct i915_vma *vma)
283 {
284 	GEM_BUG_ON(!vma->pages);
285 
286 	vma->pages = NULL;
287 }
288 
289 static void pd_vma_bind(struct i915_address_space *vm,
290 			struct i915_vm_pt_stash *stash,
291 			struct i915_vma *vma,
292 			enum i915_cache_level cache_level,
293 			u32 unused)
294 {
295 	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
296 	struct gen6_ppgtt *ppgtt = vma->private;
297 	u32 ggtt_offset = i915_ggtt_offset(vma) / I915_GTT_PAGE_SIZE;
298 
299 	ppgtt->pp_dir = ggtt_offset * sizeof(gen6_pte_t) << 10;
300 	ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm + ggtt_offset;
301 
302 	gen6_flush_pd(ppgtt, 0, ppgtt->base.vm.total);
303 }
304 
305 static void pd_vma_unbind(struct i915_address_space *vm, struct i915_vma *vma)
306 {
307 	struct gen6_ppgtt *ppgtt = vma->private;
308 	struct i915_page_directory * const pd = ppgtt->base.pd;
309 	struct i915_page_table *pt;
310 	unsigned int pde;
311 
312 	if (!ppgtt->scan_for_unused_pt)
313 		return;
314 
315 	/* Free all no longer used page tables */
316 	gen6_for_all_pdes(pt, ppgtt->base.pd, pde) {
317 		if (!pt || atomic_read(&pt->used))
318 			continue;
319 
320 		free_pt(&ppgtt->base.vm, pt);
321 		pd->entry[pde] = NULL;
322 	}
323 
324 	ppgtt->scan_for_unused_pt = false;
325 }
326 
327 static const struct i915_vma_ops pd_vma_ops = {
328 	.set_pages = pd_vma_set_pages,
329 	.clear_pages = pd_vma_clear_pages,
330 	.bind_vma = pd_vma_bind,
331 	.unbind_vma = pd_vma_unbind,
332 };
333 
334 static struct i915_vma *pd_vma_create(struct gen6_ppgtt *ppgtt, int size)
335 {
336 	struct i915_ggtt *ggtt = ppgtt->base.vm.gt->ggtt;
337 	struct i915_vma *vma;
338 
339 	GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
340 	GEM_BUG_ON(size > ggtt->vm.total);
341 
342 	vma = i915_vma_alloc();
343 	if (!vma)
344 		return ERR_PTR(-ENOMEM);
345 
346 	i915_active_init(&vma->active, NULL, NULL, 0);
347 
348 	kref_init(&vma->ref);
349 	mutex_init(&vma->pages_mutex);
350 	vma->vm = i915_vm_get(&ggtt->vm);
351 	vma->ops = &pd_vma_ops;
352 	vma->private = ppgtt;
353 
354 	vma->size = size;
355 	vma->fence_size = size;
356 	atomic_set(&vma->flags, I915_VMA_GGTT);
357 	vma->ggtt_view.type = I915_GGTT_VIEW_ROTATED; /* prevent fencing */
358 
359 	INIT_LIST_HEAD(&vma->obj_link);
360 	INIT_LIST_HEAD(&vma->closed_link);
361 
362 	return vma;
363 }
364 
365 int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww)
366 {
367 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
368 	int err;
369 
370 	GEM_BUG_ON(!atomic_read(&ppgtt->base.vm.open));
371 
372 	/*
373 	 * Workaround the limited maximum vma->pin_count and the aliasing_ppgtt
374 	 * which will be pinned into every active context.
375 	 * (When vma->pin_count becomes atomic, I expect we will naturally
376 	 * need a larger, unpacked, type and kill this redundancy.)
377 	 */
378 	if (atomic_add_unless(&ppgtt->pin_count, 1, 0))
379 		return 0;
380 
381 	if (mutex_lock_interruptible(&ppgtt->pin_mutex))
382 		return -EINTR;
383 
384 	/*
385 	 * PPGTT PDEs reside in the GGTT and consists of 512 entries. The
386 	 * allocator works in address space sizes, so it's multiplied by page
387 	 * size. We allocate at the top of the GTT to avoid fragmentation.
388 	 */
389 	err = 0;
390 	if (!atomic_read(&ppgtt->pin_count))
391 		err = i915_ggtt_pin(ppgtt->vma, ww, GEN6_PD_ALIGN, PIN_HIGH);
392 	if (!err)
393 		atomic_inc(&ppgtt->pin_count);
394 	mutex_unlock(&ppgtt->pin_mutex);
395 
396 	return err;
397 }
398 
399 void gen6_ppgtt_unpin(struct i915_ppgtt *base)
400 {
401 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
402 
403 	GEM_BUG_ON(!atomic_read(&ppgtt->pin_count));
404 	if (atomic_dec_and_test(&ppgtt->pin_count))
405 		i915_vma_unpin(ppgtt->vma);
406 }
407 
408 void gen6_ppgtt_unpin_all(struct i915_ppgtt *base)
409 {
410 	struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base);
411 
412 	if (!atomic_read(&ppgtt->pin_count))
413 		return;
414 
415 	i915_vma_unpin(ppgtt->vma);
416 	atomic_set(&ppgtt->pin_count, 0);
417 }
418 
419 struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt)
420 {
421 	struct i915_ggtt * const ggtt = gt->ggtt;
422 	struct gen6_ppgtt *ppgtt;
423 	int err;
424 
425 	ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
426 	if (!ppgtt)
427 		return ERR_PTR(-ENOMEM);
428 
429 	mutex_init(&ppgtt->flush);
430 	mutex_init(&ppgtt->pin_mutex);
431 
432 	ppgtt_init(&ppgtt->base, gt);
433 	ppgtt->base.vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen6_pte_t));
434 	ppgtt->base.vm.top = 1;
435 
436 	ppgtt->base.vm.bind_async_flags = I915_VMA_LOCAL_BIND;
437 	ppgtt->base.vm.allocate_va_range = gen6_alloc_va_range;
438 	ppgtt->base.vm.clear_range = gen6_ppgtt_clear_range;
439 	ppgtt->base.vm.insert_entries = gen6_ppgtt_insert_entries;
440 	ppgtt->base.vm.cleanup = gen6_ppgtt_cleanup;
441 
442 	ppgtt->base.vm.alloc_pt_dma = alloc_pt_dma;
443 	ppgtt->base.vm.pte_encode = ggtt->vm.pte_encode;
444 
445 	ppgtt->base.pd = __alloc_pd(I915_PDES);
446 	if (!ppgtt->base.pd) {
447 		err = -ENOMEM;
448 		goto err_free;
449 	}
450 
451 	err = gen6_ppgtt_init_scratch(ppgtt);
452 	if (err)
453 		goto err_pd;
454 
455 	ppgtt->vma = pd_vma_create(ppgtt, GEN6_PD_SIZE);
456 	if (IS_ERR(ppgtt->vma)) {
457 		err = PTR_ERR(ppgtt->vma);
458 		goto err_scratch;
459 	}
460 
461 	return &ppgtt->base;
462 
463 err_scratch:
464 	free_scratch(&ppgtt->base.vm);
465 err_pd:
466 	free_pd(&ppgtt->base.vm, ppgtt->base.pd);
467 err_free:
468 	mutex_destroy(&ppgtt->pin_mutex);
469 	kfree(ppgtt);
470 	return ERR_PTR(err);
471 }
472