1/*	$NetBSD: i915_gem_coherency.c,v 1.2 2021/12/18 23:45:30 riastradh Exp $	*/
2
3/*
4 * SPDX-License-Identifier: MIT
5 *
6 * Copyright �� 2017 Intel Corporation
7 */
8
9#include <sys/cdefs.h>
10__KERNEL_RCSID(0, "$NetBSD: i915_gem_coherency.c,v 1.2 2021/12/18 23:45:30 riastradh Exp $");
11
12#include <linux/prime_numbers.h>
13
14#include "gt/intel_engine_pm.h"
15#include "gt/intel_gt.h"
16#include "gt/intel_gt_pm.h"
17#include "gt/intel_ring.h"
18
19#include "i915_selftest.h"
20#include "selftests/i915_random.h"
21
22struct context {
23	struct drm_i915_gem_object *obj;
24	struct intel_engine_cs *engine;
25};
26
27static int cpu_set(struct context *ctx, unsigned long offset, u32 v)
28{
29	unsigned int needs_clflush;
30	struct page *page;
31	void *map;
32	u32 *cpu;
33	int err;
34
35	err = i915_gem_object_prepare_write(ctx->obj, &needs_clflush);
36	if (err)
37		return err;
38
39	page = i915_gem_object_get_page(ctx->obj, offset >> PAGE_SHIFT);
40	map = kmap_atomic(page);
41	cpu = map + offset_in_page(offset);
42
43	if (needs_clflush & CLFLUSH_BEFORE)
44		drm_clflush_virt_range(cpu, sizeof(*cpu));
45
46	*cpu = v;
47
48	if (needs_clflush & CLFLUSH_AFTER)
49		drm_clflush_virt_range(cpu, sizeof(*cpu));
50
51	kunmap_atomic(map);
52	i915_gem_object_finish_access(ctx->obj);
53
54	return 0;
55}
56
57static int cpu_get(struct context *ctx, unsigned long offset, u32 *v)
58{
59	unsigned int needs_clflush;
60	struct page *page;
61	void *map;
62	u32 *cpu;
63	int err;
64
65	err = i915_gem_object_prepare_read(ctx->obj, &needs_clflush);
66	if (err)
67		return err;
68
69	page = i915_gem_object_get_page(ctx->obj, offset >> PAGE_SHIFT);
70	map = kmap_atomic(page);
71	cpu = map + offset_in_page(offset);
72
73	if (needs_clflush & CLFLUSH_BEFORE)
74		drm_clflush_virt_range(cpu, sizeof(*cpu));
75
76	*v = *cpu;
77
78	kunmap_atomic(map);
79	i915_gem_object_finish_access(ctx->obj);
80
81	return 0;
82}
83
84static int gtt_set(struct context *ctx, unsigned long offset, u32 v)
85{
86	struct i915_vma *vma;
87	u32 __iomem *map;
88	int err = 0;
89
90	i915_gem_object_lock(ctx->obj);
91	err = i915_gem_object_set_to_gtt_domain(ctx->obj, true);
92	i915_gem_object_unlock(ctx->obj);
93	if (err)
94		return err;
95
96	vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, PIN_MAPPABLE);
97	if (IS_ERR(vma))
98		return PTR_ERR(vma);
99
100	intel_gt_pm_get(vma->vm->gt);
101
102	map = i915_vma_pin_iomap(vma);
103	i915_vma_unpin(vma);
104	if (IS_ERR(map)) {
105		err = PTR_ERR(map);
106		goto out_rpm;
107	}
108
109	iowrite32(v, &map[offset / sizeof(*map)]);
110	i915_vma_unpin_iomap(vma);
111
112out_rpm:
113	intel_gt_pm_put(vma->vm->gt);
114	return err;
115}
116
117static int gtt_get(struct context *ctx, unsigned long offset, u32 *v)
118{
119	struct i915_vma *vma;
120	u32 __iomem *map;
121	int err = 0;
122
123	i915_gem_object_lock(ctx->obj);
124	err = i915_gem_object_set_to_gtt_domain(ctx->obj, false);
125	i915_gem_object_unlock(ctx->obj);
126	if (err)
127		return err;
128
129	vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, PIN_MAPPABLE);
130	if (IS_ERR(vma))
131		return PTR_ERR(vma);
132
133	intel_gt_pm_get(vma->vm->gt);
134
135	map = i915_vma_pin_iomap(vma);
136	i915_vma_unpin(vma);
137	if (IS_ERR(map)) {
138		err = PTR_ERR(map);
139		goto out_rpm;
140	}
141
142	*v = ioread32(&map[offset / sizeof(*map)]);
143	i915_vma_unpin_iomap(vma);
144
145out_rpm:
146	intel_gt_pm_put(vma->vm->gt);
147	return err;
148}
149
150static int wc_set(struct context *ctx, unsigned long offset, u32 v)
151{
152	u32 *map;
153	int err;
154
155	i915_gem_object_lock(ctx->obj);
156	err = i915_gem_object_set_to_wc_domain(ctx->obj, true);
157	i915_gem_object_unlock(ctx->obj);
158	if (err)
159		return err;
160
161	map = i915_gem_object_pin_map(ctx->obj, I915_MAP_WC);
162	if (IS_ERR(map))
163		return PTR_ERR(map);
164
165	map[offset / sizeof(*map)] = v;
166	i915_gem_object_unpin_map(ctx->obj);
167
168	return 0;
169}
170
171static int wc_get(struct context *ctx, unsigned long offset, u32 *v)
172{
173	u32 *map;
174	int err;
175
176	i915_gem_object_lock(ctx->obj);
177	err = i915_gem_object_set_to_wc_domain(ctx->obj, false);
178	i915_gem_object_unlock(ctx->obj);
179	if (err)
180		return err;
181
182	map = i915_gem_object_pin_map(ctx->obj, I915_MAP_WC);
183	if (IS_ERR(map))
184		return PTR_ERR(map);
185
186	*v = map[offset / sizeof(*map)];
187	i915_gem_object_unpin_map(ctx->obj);
188
189	return 0;
190}
191
192static int gpu_set(struct context *ctx, unsigned long offset, u32 v)
193{
194	struct i915_request *rq;
195	struct i915_vma *vma;
196	u32 *cs;
197	int err;
198
199	i915_gem_object_lock(ctx->obj);
200	err = i915_gem_object_set_to_gtt_domain(ctx->obj, true);
201	i915_gem_object_unlock(ctx->obj);
202	if (err)
203		return err;
204
205	vma = i915_gem_object_ggtt_pin(ctx->obj, NULL, 0, 0, 0);
206	if (IS_ERR(vma))
207		return PTR_ERR(vma);
208
209	rq = intel_engine_create_kernel_request(ctx->engine);
210	if (IS_ERR(rq)) {
211		i915_vma_unpin(vma);
212		return PTR_ERR(rq);
213	}
214
215	cs = intel_ring_begin(rq, 4);
216	if (IS_ERR(cs)) {
217		i915_request_add(rq);
218		i915_vma_unpin(vma);
219		return PTR_ERR(cs);
220	}
221
222	if (INTEL_GEN(ctx->engine->i915) >= 8) {
223		*cs++ = MI_STORE_DWORD_IMM_GEN4 | 1 << 22;
224		*cs++ = lower_32_bits(i915_ggtt_offset(vma) + offset);
225		*cs++ = upper_32_bits(i915_ggtt_offset(vma) + offset);
226		*cs++ = v;
227	} else if (INTEL_GEN(ctx->engine->i915) >= 4) {
228		*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
229		*cs++ = 0;
230		*cs++ = i915_ggtt_offset(vma) + offset;
231		*cs++ = v;
232	} else {
233		*cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
234		*cs++ = i915_ggtt_offset(vma) + offset;
235		*cs++ = v;
236		*cs++ = MI_NOOP;
237	}
238	intel_ring_advance(rq, cs);
239
240	i915_vma_lock(vma);
241	err = i915_request_await_object(rq, vma->obj, true);
242	if (err == 0)
243		err = i915_vma_move_to_active(vma, rq, EXEC_OBJECT_WRITE);
244	i915_vma_unlock(vma);
245	i915_vma_unpin(vma);
246
247	i915_request_add(rq);
248
249	return err;
250}
251
252static bool always_valid(struct context *ctx)
253{
254	return true;
255}
256
257static bool needs_fence_registers(struct context *ctx)
258{
259	struct intel_gt *gt = ctx->engine->gt;
260
261	if (intel_gt_is_wedged(gt))
262		return false;
263
264	return gt->ggtt->num_fences;
265}
266
267static bool needs_mi_store_dword(struct context *ctx)
268{
269	if (intel_gt_is_wedged(ctx->engine->gt))
270		return false;
271
272	return intel_engine_can_store_dword(ctx->engine);
273}
274
275static const struct igt_coherency_mode {
276	const char *name;
277	int (*set)(struct context *ctx, unsigned long offset, u32 v);
278	int (*get)(struct context *ctx, unsigned long offset, u32 *v);
279	bool (*valid)(struct context *ctx);
280} igt_coherency_mode[] = {
281	{ "cpu", cpu_set, cpu_get, always_valid },
282	{ "gtt", gtt_set, gtt_get, needs_fence_registers },
283	{ "wc", wc_set, wc_get, always_valid },
284	{ "gpu", gpu_set, NULL, needs_mi_store_dword },
285	{ },
286};
287
288static struct intel_engine_cs *
289random_engine(struct drm_i915_private *i915, struct rnd_state *prng)
290{
291	struct intel_engine_cs *engine;
292	unsigned int count;
293
294	count = 0;
295	for_each_uabi_engine(engine, i915)
296		count++;
297
298	count = i915_prandom_u32_max_state(count, prng);
299	for_each_uabi_engine(engine, i915)
300		if (count-- == 0)
301			return engine;
302
303	return NULL;
304}
305
306static int igt_gem_coherency(void *arg)
307{
308	const unsigned int ncachelines = PAGE_SIZE/64;
309	struct drm_i915_private *i915 = arg;
310	const struct igt_coherency_mode *read, *write, *over;
311	unsigned long count, n;
312	u32 *offsets, *values;
313	I915_RND_STATE(prng);
314	struct context ctx;
315	int err = 0;
316
317	/*
318	 * We repeatedly write, overwrite and read from a sequence of
319	 * cachelines in order to try and detect incoherency (unflushed writes
320	 * from either the CPU or GPU). Each setter/getter uses our cache
321	 * domain API which should prevent incoherency.
322	 */
323
324	offsets = kmalloc_array(ncachelines, 2*sizeof(u32), GFP_KERNEL);
325	if (!offsets)
326		return -ENOMEM;
327	for (count = 0; count < ncachelines; count++)
328		offsets[count] = count * 64 + 4 * (count % 16);
329
330	values = offsets + ncachelines;
331
332	ctx.engine = random_engine(i915, &prng);
333	if (!ctx.engine) {
334		err = -ENODEV;
335		goto out_free;
336	}
337	pr_info("%s: using %s\n", __func__, ctx.engine->name);
338	intel_engine_pm_get(ctx.engine);
339
340	for (over = igt_coherency_mode; over->name; over++) {
341		if (!over->set)
342			continue;
343
344		if (!over->valid(&ctx))
345			continue;
346
347		for (write = igt_coherency_mode; write->name; write++) {
348			if (!write->set)
349				continue;
350
351			if (!write->valid(&ctx))
352				continue;
353
354			for (read = igt_coherency_mode; read->name; read++) {
355				if (!read->get)
356					continue;
357
358				if (!read->valid(&ctx))
359					continue;
360
361				for_each_prime_number_from(count, 1, ncachelines) {
362					ctx.obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
363					if (IS_ERR(ctx.obj)) {
364						err = PTR_ERR(ctx.obj);
365						goto out_pm;
366					}
367
368					i915_random_reorder(offsets, ncachelines, &prng);
369					for (n = 0; n < count; n++)
370						values[n] = prandom_u32_state(&prng);
371
372					for (n = 0; n < count; n++) {
373						err = over->set(&ctx, offsets[n], ~values[n]);
374						if (err) {
375							pr_err("Failed to set stale value[%ld/%ld] in object using %s, err=%d\n",
376							       n, count, over->name, err);
377							goto put_object;
378						}
379					}
380
381					for (n = 0; n < count; n++) {
382						err = write->set(&ctx, offsets[n], values[n]);
383						if (err) {
384							pr_err("Failed to set value[%ld/%ld] in object using %s, err=%d\n",
385							       n, count, write->name, err);
386							goto put_object;
387						}
388					}
389
390					for (n = 0; n < count; n++) {
391						u32 found;
392
393						err = read->get(&ctx, offsets[n], &found);
394						if (err) {
395							pr_err("Failed to get value[%ld/%ld] in object using %s, err=%d\n",
396							       n, count, read->name, err);
397							goto put_object;
398						}
399
400						if (found != values[n]) {
401							pr_err("Value[%ld/%ld] mismatch, (overwrite with %s) wrote [%s] %x read [%s] %x (inverse %x), at offset %x\n",
402							       n, count, over->name,
403							       write->name, values[n],
404							       read->name, found,
405							       ~values[n], offsets[n]);
406							err = -EINVAL;
407							goto put_object;
408						}
409					}
410
411					i915_gem_object_put(ctx.obj);
412				}
413			}
414		}
415	}
416out_pm:
417	intel_engine_pm_put(ctx.engine);
418out_free:
419	kfree(offsets);
420	return err;
421
422put_object:
423	i915_gem_object_put(ctx.obj);
424	goto out_pm;
425}
426
427int i915_gem_coherency_live_selftests(struct drm_i915_private *i915)
428{
429	static const struct i915_subtest tests[] = {
430		SUBTEST(igt_gem_coherency),
431	};
432
433	return i915_subtests(tests, i915);
434}
435