xref: /linux/drivers/gpu/drm/xe/xe_ggtt.c (revision be239684b18e1cdcafcf8c7face4a2f562c745ad)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include "xe_ggtt.h"
7 
8 #include <linux/sizes.h>
9 
10 #include <drm/drm_managed.h>
11 #include <drm/i915_drm.h>
12 
13 #include "regs/xe_gt_regs.h"
14 #include "regs/xe_regs.h"
15 #include "xe_assert.h"
16 #include "xe_bo.h"
17 #include "xe_device.h"
18 #include "xe_gt.h"
19 #include "xe_gt_printk.h"
20 #include "xe_gt_tlb_invalidation.h"
21 #include "xe_map.h"
22 #include "xe_mmio.h"
23 #include "xe_sriov.h"
24 #include "xe_wopcm.h"
25 
26 #define XELPG_GGTT_PTE_PAT0	BIT_ULL(52)
27 #define XELPG_GGTT_PTE_PAT1	BIT_ULL(53)
28 
29 /* GuC addresses above GUC_GGTT_TOP also don't map through the GTT */
30 #define GUC_GGTT_TOP	0xFEE00000
31 
32 static u64 xelp_ggtt_pte_encode_bo(struct xe_bo *bo, u64 bo_offset,
33 				   u16 pat_index)
34 {
35 	u64 pte;
36 
37 	pte = xe_bo_addr(bo, bo_offset, XE_PAGE_SIZE);
38 	pte |= XE_PAGE_PRESENT;
39 
40 	if (xe_bo_is_vram(bo) || xe_bo_is_stolen_devmem(bo))
41 		pte |= XE_GGTT_PTE_DM;
42 
43 	return pte;
44 }
45 
46 static u64 xelpg_ggtt_pte_encode_bo(struct xe_bo *bo, u64 bo_offset,
47 				    u16 pat_index)
48 {
49 	struct xe_device *xe = xe_bo_device(bo);
50 	u64 pte;
51 
52 	pte = xelp_ggtt_pte_encode_bo(bo, bo_offset, pat_index);
53 
54 	xe_assert(xe, pat_index <= 3);
55 
56 	if (pat_index & BIT(0))
57 		pte |= XELPG_GGTT_PTE_PAT0;
58 
59 	if (pat_index & BIT(1))
60 		pte |= XELPG_GGTT_PTE_PAT1;
61 
62 	return pte;
63 }
64 
65 static unsigned int probe_gsm_size(struct pci_dev *pdev)
66 {
67 	u16 gmch_ctl, ggms;
68 
69 	pci_read_config_word(pdev, SNB_GMCH_CTRL, &gmch_ctl);
70 	ggms = (gmch_ctl >> BDW_GMCH_GGMS_SHIFT) & BDW_GMCH_GGMS_MASK;
71 	return ggms ? SZ_1M << ggms : 0;
72 }
73 
74 void xe_ggtt_set_pte(struct xe_ggtt *ggtt, u64 addr, u64 pte)
75 {
76 	xe_tile_assert(ggtt->tile, !(addr & XE_PTE_MASK));
77 	xe_tile_assert(ggtt->tile, addr < ggtt->size);
78 
79 	writeq(pte, &ggtt->gsm[addr >> XE_PTE_SHIFT]);
80 }
81 
82 static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size)
83 {
84 	u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB];
85 	u64 end = start + size - 1;
86 	u64 scratch_pte;
87 
88 	xe_tile_assert(ggtt->tile, start < end);
89 
90 	if (ggtt->scratch)
91 		scratch_pte = ggtt->pt_ops->pte_encode_bo(ggtt->scratch, 0,
92 							  pat_index);
93 	else
94 		scratch_pte = 0;
95 
96 	while (start < end) {
97 		xe_ggtt_set_pte(ggtt, start, scratch_pte);
98 		start += XE_PAGE_SIZE;
99 	}
100 }
101 
102 static void ggtt_fini_early(struct drm_device *drm, void *arg)
103 {
104 	struct xe_ggtt *ggtt = arg;
105 
106 	mutex_destroy(&ggtt->lock);
107 	drm_mm_takedown(&ggtt->mm);
108 }
109 
110 static void ggtt_fini(struct drm_device *drm, void *arg)
111 {
112 	struct xe_ggtt *ggtt = arg;
113 
114 	ggtt->scratch = NULL;
115 }
116 
117 static void primelockdep(struct xe_ggtt *ggtt)
118 {
119 	if (!IS_ENABLED(CONFIG_LOCKDEP))
120 		return;
121 
122 	fs_reclaim_acquire(GFP_KERNEL);
123 	might_lock(&ggtt->lock);
124 	fs_reclaim_release(GFP_KERNEL);
125 }
126 
127 static const struct xe_ggtt_pt_ops xelp_pt_ops = {
128 	.pte_encode_bo = xelp_ggtt_pte_encode_bo,
129 };
130 
131 static const struct xe_ggtt_pt_ops xelpg_pt_ops = {
132 	.pte_encode_bo = xelpg_ggtt_pte_encode_bo,
133 };
134 
135 /*
136  * Early GGTT initialization, which allows to create new mappings usable by the
137  * GuC.
138  * Mappings are not usable by the HW engines, as it doesn't have scratch /
139  * initial clear done to it yet. That will happen in the regular, non-early
140  * GGTT init.
141  */
142 int xe_ggtt_init_early(struct xe_ggtt *ggtt)
143 {
144 	struct xe_device *xe = tile_to_xe(ggtt->tile);
145 	struct pci_dev *pdev = to_pci_dev(xe->drm.dev);
146 	unsigned int gsm_size;
147 
148 	if (IS_SRIOV_VF(xe))
149 		gsm_size = SZ_8M; /* GGTT is expected to be 4GiB */
150 	else
151 		gsm_size = probe_gsm_size(pdev);
152 
153 	if (gsm_size == 0) {
154 		drm_err(&xe->drm, "Hardware reported no preallocated GSM\n");
155 		return -ENOMEM;
156 	}
157 
158 	ggtt->gsm = ggtt->tile->mmio.regs + SZ_8M;
159 	ggtt->size = (gsm_size / 8) * (u64) XE_PAGE_SIZE;
160 
161 	if (IS_DGFX(xe) && xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K)
162 		ggtt->flags |= XE_GGTT_FLAGS_64K;
163 
164 	/*
165 	 * 8B per entry, each points to a 4KB page.
166 	 *
167 	 * The GuC address space is limited on both ends of the GGTT, because
168 	 * the GuC shim HW redirects accesses to those addresses to other HW
169 	 * areas instead of going through the GGTT. On the bottom end, the GuC
170 	 * can't access offsets below the WOPCM size, while on the top side the
171 	 * limit is fixed at GUC_GGTT_TOP. To keep things simple, instead of
172 	 * checking each object to see if they are accessed by GuC or not, we
173 	 * just exclude those areas from the allocator. Additionally, to
174 	 * simplify the driver load, we use the maximum WOPCM size in this logic
175 	 * instead of the programmed one, so we don't need to wait until the
176 	 * actual size to be programmed is determined (which requires FW fetch)
177 	 * before initializing the GGTT. These simplifications might waste space
178 	 * in the GGTT (about 20-25 MBs depending on the platform) but we can
179 	 * live with this.
180 	 *
181 	 * Another benifit of this is the GuC bootrom can't access anything
182 	 * below the WOPCM max size so anything the bootom needs to access (e.g.
183 	 * a RSA key) needs to be placed in the GGTT above the WOPCM max size.
184 	 * Starting the GGTT allocations above the WOPCM max give us the correct
185 	 * placement for free.
186 	 */
187 	if (ggtt->size > GUC_GGTT_TOP)
188 		ggtt->size = GUC_GGTT_TOP;
189 
190 	if (GRAPHICS_VERx100(xe) >= 1270)
191 		ggtt->pt_ops = &xelpg_pt_ops;
192 	else
193 		ggtt->pt_ops = &xelp_pt_ops;
194 
195 	drm_mm_init(&ggtt->mm, xe_wopcm_size(xe),
196 		    ggtt->size - xe_wopcm_size(xe));
197 	mutex_init(&ggtt->lock);
198 	primelockdep(ggtt);
199 
200 	return drmm_add_action_or_reset(&xe->drm, ggtt_fini_early, ggtt);
201 }
202 
203 static void xe_ggtt_initial_clear(struct xe_ggtt *ggtt)
204 {
205 	struct drm_mm_node *hole;
206 	u64 start, end;
207 
208 	/* Display may have allocated inside ggtt, so be careful with clearing here */
209 	xe_device_mem_access_get(tile_to_xe(ggtt->tile));
210 	mutex_lock(&ggtt->lock);
211 	drm_mm_for_each_hole(hole, &ggtt->mm, start, end)
212 		xe_ggtt_clear(ggtt, start, end - start);
213 
214 	xe_ggtt_invalidate(ggtt);
215 	mutex_unlock(&ggtt->lock);
216 	xe_device_mem_access_put(tile_to_xe(ggtt->tile));
217 }
218 
219 int xe_ggtt_init(struct xe_ggtt *ggtt)
220 {
221 	struct xe_device *xe = tile_to_xe(ggtt->tile);
222 	unsigned int flags;
223 	int err;
224 
225 	/*
226 	 * So we don't need to worry about 64K GGTT layout when dealing with
227 	 * scratch entires, rather keep the scratch page in system memory on
228 	 * platforms where 64K pages are needed for VRAM.
229 	 */
230 	flags = XE_BO_CREATE_PINNED_BIT;
231 	if (ggtt->flags & XE_GGTT_FLAGS_64K)
232 		flags |= XE_BO_CREATE_SYSTEM_BIT;
233 	else
234 		flags |= XE_BO_CREATE_VRAM_IF_DGFX(ggtt->tile);
235 
236 	ggtt->scratch = xe_managed_bo_create_pin_map(xe, ggtt->tile, XE_PAGE_SIZE, flags);
237 	if (IS_ERR(ggtt->scratch)) {
238 		err = PTR_ERR(ggtt->scratch);
239 		goto err;
240 	}
241 
242 	xe_map_memset(xe, &ggtt->scratch->vmap, 0, 0, ggtt->scratch->size);
243 
244 	xe_ggtt_initial_clear(ggtt);
245 
246 	return drmm_add_action_or_reset(&xe->drm, ggtt_fini, ggtt);
247 err:
248 	ggtt->scratch = NULL;
249 	return err;
250 }
251 
252 #define GUC_TLB_INV_CR				XE_REG(0xcee8)
253 #define   GUC_TLB_INV_CR_INVALIDATE		REG_BIT(0)
254 #define PVC_GUC_TLB_INV_DESC0			XE_REG(0xcf7c)
255 #define   PVC_GUC_TLB_INV_DESC0_VALID		REG_BIT(0)
256 #define PVC_GUC_TLB_INV_DESC1			XE_REG(0xcf80)
257 #define   PVC_GUC_TLB_INV_DESC1_INVALIDATE	REG_BIT(6)
258 
259 static void ggtt_invalidate_gt_tlb(struct xe_gt *gt)
260 {
261 	if (!gt)
262 		return;
263 
264 	/*
265 	 * Invalidation can happen when there's no in-flight work keeping the
266 	 * GT awake.  We need to explicitly grab forcewake to ensure the GT
267 	 * and GuC are accessible.
268 	 */
269 	xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
270 
271 	/* TODO: vfunc for GuC vs. non-GuC */
272 
273 	if (gt->uc.guc.submission_state.enabled) {
274 		int seqno;
275 
276 		seqno = xe_gt_tlb_invalidation_guc(gt);
277 		xe_gt_assert(gt, seqno > 0);
278 		if (seqno > 0)
279 			xe_gt_tlb_invalidation_wait(gt, seqno);
280 	} else if (xe_device_uc_enabled(gt_to_xe(gt))) {
281 		struct xe_device *xe = gt_to_xe(gt);
282 
283 		if (xe->info.platform == XE_PVC || GRAPHICS_VER(xe) >= 20) {
284 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC1,
285 					PVC_GUC_TLB_INV_DESC1_INVALIDATE);
286 			xe_mmio_write32(gt, PVC_GUC_TLB_INV_DESC0,
287 					PVC_GUC_TLB_INV_DESC0_VALID);
288 		} else
289 			xe_mmio_write32(gt, GUC_TLB_INV_CR,
290 					GUC_TLB_INV_CR_INVALIDATE);
291 	}
292 
293 	xe_force_wake_put(gt_to_fw(gt), XE_FW_GT);
294 }
295 
296 void xe_ggtt_invalidate(struct xe_ggtt *ggtt)
297 {
298 	/* Each GT in a tile has its own TLB to cache GGTT lookups */
299 	ggtt_invalidate_gt_tlb(ggtt->tile->primary_gt);
300 	ggtt_invalidate_gt_tlb(ggtt->tile->media_gt);
301 }
302 
303 void xe_ggtt_printk(struct xe_ggtt *ggtt, const char *prefix)
304 {
305 	u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB];
306 	u64 addr, scratch_pte;
307 
308 	scratch_pte = ggtt->pt_ops->pte_encode_bo(ggtt->scratch, 0, pat_index);
309 
310 	printk("%sGlobal GTT:", prefix);
311 	for (addr = 0; addr < ggtt->size; addr += XE_PAGE_SIZE) {
312 		unsigned int i = addr / XE_PAGE_SIZE;
313 
314 		xe_tile_assert(ggtt->tile, addr <= U32_MAX);
315 		if (ggtt->gsm[i] == scratch_pte)
316 			continue;
317 
318 		printk("%s    ggtt[0x%08x] = 0x%016llx",
319 		       prefix, (u32)addr, ggtt->gsm[i]);
320 	}
321 }
322 
323 static void xe_ggtt_dump_node(struct xe_ggtt *ggtt,
324 			      const struct drm_mm_node *node, const char *description)
325 {
326 	char buf[10];
327 
328 	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
329 		string_get_size(node->size, 1, STRING_UNITS_2, buf, sizeof(buf));
330 		xe_gt_dbg(ggtt->tile->primary_gt, "GGTT %#llx-%#llx (%s) %s\n",
331 			  node->start, node->start + node->size, buf, description);
332 	}
333 }
334 
335 /**
336  * xe_ggtt_balloon - prevent allocation of specified GGTT addresses
337  * @ggtt: the &xe_ggtt where we want to make reservation
338  * @start: the starting GGTT address of the reserved region
339  * @end: then end GGTT address of the reserved region
340  * @node: the &drm_mm_node to hold reserved GGTT node
341  *
342  * Use xe_ggtt_deballoon() to release a reserved GGTT node.
343  *
344  * Return: 0 on success or a negative error code on failure.
345  */
346 int xe_ggtt_balloon(struct xe_ggtt *ggtt, u64 start, u64 end, struct drm_mm_node *node)
347 {
348 	int err;
349 
350 	xe_tile_assert(ggtt->tile, start < end);
351 	xe_tile_assert(ggtt->tile, IS_ALIGNED(start, XE_PAGE_SIZE));
352 	xe_tile_assert(ggtt->tile, IS_ALIGNED(end, XE_PAGE_SIZE));
353 	xe_tile_assert(ggtt->tile, !drm_mm_node_allocated(node));
354 
355 	node->color = 0;
356 	node->start = start;
357 	node->size = end - start;
358 
359 	mutex_lock(&ggtt->lock);
360 	err = drm_mm_reserve_node(&ggtt->mm, node);
361 	mutex_unlock(&ggtt->lock);
362 
363 	if (xe_gt_WARN(ggtt->tile->primary_gt, err,
364 		       "Failed to balloon GGTT %#llx-%#llx (%pe)\n",
365 		       node->start, node->start + node->size, ERR_PTR(err)))
366 		return err;
367 
368 	xe_ggtt_dump_node(ggtt, node, "balloon");
369 	return 0;
370 }
371 
372 /**
373  * xe_ggtt_deballoon - release a reserved GGTT region
374  * @ggtt: the &xe_ggtt where reserved node belongs
375  * @node: the &drm_mm_node with reserved GGTT region
376  *
377  * See xe_ggtt_balloon() for details.
378  */
379 void xe_ggtt_deballoon(struct xe_ggtt *ggtt, struct drm_mm_node *node)
380 {
381 	if (!drm_mm_node_allocated(node))
382 		return;
383 
384 	xe_ggtt_dump_node(ggtt, node, "deballoon");
385 
386 	mutex_lock(&ggtt->lock);
387 	drm_mm_remove_node(node);
388 	mutex_unlock(&ggtt->lock);
389 }
390 
391 int xe_ggtt_insert_special_node_locked(struct xe_ggtt *ggtt, struct drm_mm_node *node,
392 				       u32 size, u32 align, u32 mm_flags)
393 {
394 	return drm_mm_insert_node_generic(&ggtt->mm, node, size, align, 0,
395 					  mm_flags);
396 }
397 
398 int xe_ggtt_insert_special_node(struct xe_ggtt *ggtt, struct drm_mm_node *node,
399 				u32 size, u32 align)
400 {
401 	int ret;
402 
403 	mutex_lock(&ggtt->lock);
404 	ret = xe_ggtt_insert_special_node_locked(ggtt, node, size,
405 						 align, DRM_MM_INSERT_HIGH);
406 	mutex_unlock(&ggtt->lock);
407 
408 	return ret;
409 }
410 
411 void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
412 {
413 	u16 cache_mode = bo->flags & XE_BO_NEEDS_UC ? XE_CACHE_NONE : XE_CACHE_WB;
414 	u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[cache_mode];
415 	u64 start = bo->ggtt_node.start;
416 	u64 offset, pte;
417 
418 	for (offset = 0; offset < bo->size; offset += XE_PAGE_SIZE) {
419 		pte = ggtt->pt_ops->pte_encode_bo(bo, offset, pat_index);
420 		xe_ggtt_set_pte(ggtt, start + offset, pte);
421 	}
422 
423 	xe_ggtt_invalidate(ggtt);
424 }
425 
426 static int __xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo,
427 				  u64 start, u64 end)
428 {
429 	int err;
430 	u64 alignment = XE_PAGE_SIZE;
431 
432 	if (xe_bo_is_vram(bo) && ggtt->flags & XE_GGTT_FLAGS_64K)
433 		alignment = SZ_64K;
434 
435 	if (XE_WARN_ON(bo->ggtt_node.size)) {
436 		/* Someone's already inserted this BO in the GGTT */
437 		xe_tile_assert(ggtt->tile, bo->ggtt_node.size == bo->size);
438 		return 0;
439 	}
440 
441 	err = xe_bo_validate(bo, NULL, false);
442 	if (err)
443 		return err;
444 
445 	xe_device_mem_access_get(tile_to_xe(ggtt->tile));
446 	mutex_lock(&ggtt->lock);
447 	err = drm_mm_insert_node_in_range(&ggtt->mm, &bo->ggtt_node, bo->size,
448 					  alignment, 0, start, end, 0);
449 	if (!err)
450 		xe_ggtt_map_bo(ggtt, bo);
451 	mutex_unlock(&ggtt->lock);
452 	xe_device_mem_access_put(tile_to_xe(ggtt->tile));
453 
454 	return err;
455 }
456 
457 int xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo,
458 			 u64 start, u64 end)
459 {
460 	return __xe_ggtt_insert_bo_at(ggtt, bo, start, end);
461 }
462 
463 int xe_ggtt_insert_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
464 {
465 	return __xe_ggtt_insert_bo_at(ggtt, bo, 0, U64_MAX);
466 }
467 
468 void xe_ggtt_remove_node(struct xe_ggtt *ggtt, struct drm_mm_node *node)
469 {
470 	xe_device_mem_access_get(tile_to_xe(ggtt->tile));
471 	mutex_lock(&ggtt->lock);
472 
473 	xe_ggtt_clear(ggtt, node->start, node->size);
474 	drm_mm_remove_node(node);
475 	node->size = 0;
476 
477 	xe_ggtt_invalidate(ggtt);
478 
479 	mutex_unlock(&ggtt->lock);
480 	xe_device_mem_access_put(tile_to_xe(ggtt->tile));
481 }
482 
483 void xe_ggtt_remove_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
484 {
485 	if (XE_WARN_ON(!bo->ggtt_node.size))
486 		return;
487 
488 	/* This BO is not currently in the GGTT */
489 	xe_tile_assert(ggtt->tile, bo->ggtt_node.size == bo->size);
490 
491 	xe_ggtt_remove_node(ggtt, &bo->ggtt_node);
492 }
493 
494 int xe_ggtt_dump(struct xe_ggtt *ggtt, struct drm_printer *p)
495 {
496 	int err;
497 
498 	err = mutex_lock_interruptible(&ggtt->lock);
499 	if (err)
500 		return err;
501 
502 	drm_mm_print(&ggtt->mm, p);
503 	mutex_unlock(&ggtt->lock);
504 	return err;
505 }
506