xref: /linux/drivers/gpu/drm/xe/xe_memirq.c (revision 2c1ed907520c50326b8f604907a8478b27881a2e)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2023 Intel Corporation
4  */
5 
6 #include <drm/drm_managed.h>
7 
8 #include "regs/xe_guc_regs.h"
9 #include "regs/xe_irq_regs.h"
10 #include "regs/xe_regs.h"
11 
12 #include "xe_assert.h"
13 #include "xe_bo.h"
14 #include "xe_device.h"
15 #include "xe_device_types.h"
16 #include "xe_gt.h"
17 #include "xe_gt_printk.h"
18 #include "xe_guc.h"
19 #include "xe_hw_engine.h"
20 #include "xe_map.h"
21 #include "xe_memirq.h"
22 
23 #define memirq_assert(m, condition)	xe_tile_assert(memirq_to_tile(m), condition)
24 #define memirq_printk(m, _level, _fmt, ...)			\
25 	drm_##_level(&memirq_to_xe(m)->drm, "MEMIRQ%u: " _fmt,	\
26 		     memirq_to_tile(m)->id, ##__VA_ARGS__)
27 
28 #ifdef CONFIG_DRM_XE_DEBUG_MEMIRQ
29 #define memirq_debug(m, _fmt, ...)	memirq_printk(m, dbg, _fmt, ##__VA_ARGS__)
30 #else
31 #define memirq_debug(...)
32 #endif
33 
34 #define memirq_err(m, _fmt, ...)	memirq_printk(m, err, _fmt, ##__VA_ARGS__)
35 #define memirq_err_ratelimited(m, _fmt, ...)	\
36 	memirq_printk(m, err_ratelimited, _fmt, ##__VA_ARGS__)
37 
memirq_to_tile(struct xe_memirq * memirq)38 static struct xe_tile *memirq_to_tile(struct xe_memirq *memirq)
39 {
40 	return container_of(memirq, struct xe_tile, memirq);
41 }
42 
memirq_to_xe(struct xe_memirq * memirq)43 static struct xe_device *memirq_to_xe(struct xe_memirq *memirq)
44 {
45 	return tile_to_xe(memirq_to_tile(memirq));
46 }
47 
guc_name(struct xe_guc * guc)48 static const char *guc_name(struct xe_guc *guc)
49 {
50 	return xe_gt_is_media_type(guc_to_gt(guc)) ? "media GuC" : "GuC";
51 }
52 
53 /**
54  * DOC: Memory Based Interrupts
55  *
56  * MMIO register based interrupts infrastructure used for non-virtualized mode
57  * or SRIOV-8 (which supports 8 Virtual Functions) does not scale efficiently
58  * to allow delivering interrupts to a large number of Virtual machines or
59  * containers. Memory based interrupt status reporting provides an efficient
60  * and scalable infrastructure.
61  *
62  * For memory based interrupt status reporting hardware sequence is:
63  *  * Engine writes the interrupt event to memory
64  *    (Pointer to memory location is provided by SW. This memory surface must
65  *    be mapped to system memory and must be marked as un-cacheable (UC) on
66  *    Graphics IP Caches)
67  *  * Engine triggers an interrupt to host.
68  */
69 
70 /**
71  * DOC: Memory Based Interrupts Page Layout
72  *
73  * `Memory Based Interrupts`_ requires three different objects, which are
74  * called "page" in the specs, even if they aren't page-sized or aligned.
75  *
76  * To simplify the code we allocate a single page size object and then use
77  * offsets to embedded "pages". The address of those "pages" are then
78  * programmed in the HW via LRI and LRM in the context image.
79  *
80  * - _`Interrupt Status Report Page`: this page contains the interrupt
81  *   status vectors for each unit. Each bit in the interrupt vectors is
82  *   converted to a byte, with the byte being set to 0xFF when an
83  *   interrupt is triggered; interrupt vectors are 16b big so each unit
84  *   gets 16B. One space is reserved for each bit in one of the
85  *   GT_INTR_DWx registers, so this object needs a total of 1024B.
86  *   This object needs to be 4KiB aligned.
87  *
88  * - _`Interrupt Source Report Page`: this is the equivalent of the
89  *   GEN11_GT_INTR_DWx registers, with each bit in those registers being
90  *   mapped to a byte here. The offsets are the same, just bytes instead
91  *   of bits. This object needs to be cacheline aligned.
92  *
93  * - Interrupt Mask: the HW needs a location to fetch the interrupt
94  *   mask vector to be used by the LRM in the context, so we just use
95  *   the next available space in the interrupt page.
96  *
97  * ::
98  *
99  *   0x0000   +===========+  <== Interrupt Status Report Page
100  *            |           |
101  *            |           |     ____ +----+----------------+
102  *            |           |    /     |  0 | USER INTERRUPT |
103  *            +-----------+ __/      |  1 |                |
104  *            |  HWE(n)   | __       |    | CTX SWITCH     |
105  *            +-----------+   \      |    | WAIT SEMAPHORE |
106  *            |           |    \____ | 15 |                |
107  *            |           |          +----+----------------+
108  *            |           |
109  *   0x0400   +===========+  <== Interrupt Source Report Page
110  *            |  HWE(0)   |
111  *            |  HWE(1)   |
112  *            |           |
113  *            |  HWE(x)   |
114  *   0x0440   +===========+  <== Interrupt Enable Mask
115  *            |           |
116  *            |           |
117  *            +-----------+
118  *
119  *
120  * MSI-X use case
121  *
122  * When using MSI-X, hw engines report interrupt status and source to engine
123  * instance 0. For this scenario, in order to differentiate between the
124  * engines, we need to pass different status/source pointers in the LRC.
125  *
126  * The requirements on those pointers are:
127  * - Interrupt status should be 4KiB aligned
128  * - Interrupt source should be 64 bytes aligned
129  *
130  * To accommodate this, we duplicate the memirq page layout above -
131  * allocating a page for each engine instance and pass this page in the LRC.
132  * Note that the same page can be reused for different engine types.
133  * For example, an LRC executing on CCS #x will have pointers to page #x,
134  * and an LRC executing on BCS #x will have the same pointers.
135  *
136  * ::
137  *
138  *   0x0000   +==============================+  <== page for instance 0 (BCS0, CCS0, etc.)
139  *            | Interrupt Status Report Page |
140  *   0x0400   +==============================+
141  *            | Interrupt Source Report Page |
142  *   0x0440   +==============================+
143  *            | Interrupt Enable Mask        |
144  *            +==============================+
145  *            | Not used                     |
146  *   0x1000   +==============================+  <== page for instance 1 (BCS1, CCS1, etc.)
147  *            | Interrupt Status Report Page |
148  *   0x1400   +==============================+
149  *            | Interrupt Source Report Page |
150  *   0x1440   +==============================+
151  *            | Not used                     |
152  *   0x2000   +==============================+  <== page for instance 2 (BCS2, CCS2, etc.)
153  *            | ...                          |
154  *            +==============================+
155  *
156  */
157 
hw_reports_to_instance_zero(struct xe_memirq * memirq)158 static inline bool hw_reports_to_instance_zero(struct xe_memirq *memirq)
159 {
160 	/*
161 	 * When the HW engines are configured to use MSI-X,
162 	 * they report interrupt status and source to the offset of
163 	 * engine instance 0.
164 	 */
165 	return xe_device_has_msix(memirq_to_xe(memirq));
166 }
167 
memirq_alloc_pages(struct xe_memirq * memirq)168 static int memirq_alloc_pages(struct xe_memirq *memirq)
169 {
170 	struct xe_device *xe = memirq_to_xe(memirq);
171 	struct xe_tile *tile = memirq_to_tile(memirq);
172 	size_t bo_size = hw_reports_to_instance_zero(memirq) ?
173 		XE_HW_ENGINE_MAX_INSTANCE * SZ_4K : SZ_4K;
174 	struct xe_bo *bo;
175 	int err;
176 
177 	BUILD_BUG_ON(!IS_ALIGNED(XE_MEMIRQ_SOURCE_OFFSET(0), SZ_64));
178 	BUILD_BUG_ON(!IS_ALIGNED(XE_MEMIRQ_STATUS_OFFSET(0), SZ_4K));
179 
180 	bo = xe_managed_bo_create_pin_map(xe, tile, bo_size,
181 					  XE_BO_FLAG_SYSTEM |
182 					  XE_BO_FLAG_GGTT |
183 					  XE_BO_FLAG_GGTT_INVALIDATE |
184 					  XE_BO_FLAG_NEEDS_UC |
185 					  XE_BO_FLAG_NEEDS_CPU_ACCESS);
186 	if (IS_ERR(bo)) {
187 		err = PTR_ERR(bo);
188 		goto out;
189 	}
190 
191 	memirq_assert(memirq, !xe_bo_is_vram(bo));
192 	memirq_assert(memirq, !memirq->bo);
193 
194 	iosys_map_memset(&bo->vmap, 0, 0, bo_size);
195 
196 	memirq->bo = bo;
197 	memirq->source = IOSYS_MAP_INIT_OFFSET(&bo->vmap, XE_MEMIRQ_SOURCE_OFFSET(0));
198 	memirq->status = IOSYS_MAP_INIT_OFFSET(&bo->vmap, XE_MEMIRQ_STATUS_OFFSET(0));
199 	memirq->mask = IOSYS_MAP_INIT_OFFSET(&bo->vmap, XE_MEMIRQ_ENABLE_OFFSET);
200 
201 	memirq_assert(memirq, !memirq->source.is_iomem);
202 	memirq_assert(memirq, !memirq->status.is_iomem);
203 	memirq_assert(memirq, !memirq->mask.is_iomem);
204 
205 	memirq_debug(memirq, "page offsets: bo %#x bo_size %zu source %#x status %#x\n",
206 		     xe_bo_ggtt_addr(bo), bo_size, XE_MEMIRQ_SOURCE_OFFSET(0),
207 		     XE_MEMIRQ_STATUS_OFFSET(0));
208 
209 	return 0;
210 
211 out:
212 	memirq_err(memirq, "Failed to allocate memirq page (%pe)\n", ERR_PTR(err));
213 	return err;
214 }
215 
memirq_set_enable(struct xe_memirq * memirq,bool enable)216 static void memirq_set_enable(struct xe_memirq *memirq, bool enable)
217 {
218 	iosys_map_wr(&memirq->mask, 0, u32, enable ? GENMASK(15, 0) : 0);
219 
220 	memirq->enabled = enable;
221 }
222 
223 /**
224  * xe_memirq_init - Initialize data used by `Memory Based Interrupts`_.
225  * @memirq: the &xe_memirq to initialize
226  *
227  * Allocate `Interrupt Source Report Page`_ and `Interrupt Status Report Page`_
228  * used by `Memory Based Interrupts`_.
229  *
230  * These allocations are managed and will be implicitly released on unload.
231  *
232  * If this function fails then the driver won't be able to operate correctly.
233  * If `Memory Based Interrupts`_ are not used this function will return 0.
234  *
235  * Return: 0 on success or a negative error code on failure.
236  */
xe_memirq_init(struct xe_memirq * memirq)237 int xe_memirq_init(struct xe_memirq *memirq)
238 {
239 	struct xe_device *xe = memirq_to_xe(memirq);
240 	int err;
241 
242 	if (!xe_device_uses_memirq(xe))
243 		return 0;
244 
245 	err = memirq_alloc_pages(memirq);
246 	if (unlikely(err))
247 		return err;
248 
249 	/* we need to start with all irqs enabled */
250 	memirq_set_enable(memirq, true);
251 
252 	return 0;
253 }
254 
__memirq_source_page(struct xe_memirq * memirq,u16 instance)255 static u32 __memirq_source_page(struct xe_memirq *memirq, u16 instance)
256 {
257 	memirq_assert(memirq, instance <= XE_HW_ENGINE_MAX_INSTANCE);
258 	memirq_assert(memirq, memirq->bo);
259 
260 	instance = hw_reports_to_instance_zero(memirq) ? instance : 0;
261 	return xe_bo_ggtt_addr(memirq->bo) + XE_MEMIRQ_SOURCE_OFFSET(instance);
262 }
263 
264 /**
265  * xe_memirq_source_ptr - Get GGTT's offset of the `Interrupt Source Report Page`_.
266  * @memirq: the &xe_memirq to query
267  * @hwe: the hw engine for which we want the report page
268  *
269  * Shall be called when `Memory Based Interrupts`_ are used
270  * and xe_memirq_init() didn't fail.
271  *
272  * Return: GGTT's offset of the `Interrupt Source Report Page`_.
273  */
xe_memirq_source_ptr(struct xe_memirq * memirq,struct xe_hw_engine * hwe)274 u32 xe_memirq_source_ptr(struct xe_memirq *memirq, struct xe_hw_engine *hwe)
275 {
276 	memirq_assert(memirq, xe_device_uses_memirq(memirq_to_xe(memirq)));
277 
278 	return __memirq_source_page(memirq, hwe->instance);
279 }
280 
__memirq_status_page(struct xe_memirq * memirq,u16 instance)281 static u32 __memirq_status_page(struct xe_memirq *memirq, u16 instance)
282 {
283 	memirq_assert(memirq, instance <= XE_HW_ENGINE_MAX_INSTANCE);
284 	memirq_assert(memirq, memirq->bo);
285 
286 	instance = hw_reports_to_instance_zero(memirq) ? instance : 0;
287 	return xe_bo_ggtt_addr(memirq->bo) + XE_MEMIRQ_STATUS_OFFSET(instance);
288 }
289 
290 /**
291  * xe_memirq_status_ptr - Get GGTT's offset of the `Interrupt Status Report Page`_.
292  * @memirq: the &xe_memirq to query
293  * @hwe: the hw engine for which we want the report page
294  *
295  * Shall be called when `Memory Based Interrupts`_ are used
296  * and xe_memirq_init() didn't fail.
297  *
298  * Return: GGTT's offset of the `Interrupt Status Report Page`_.
299  */
xe_memirq_status_ptr(struct xe_memirq * memirq,struct xe_hw_engine * hwe)300 u32 xe_memirq_status_ptr(struct xe_memirq *memirq, struct xe_hw_engine *hwe)
301 {
302 	memirq_assert(memirq, xe_device_uses_memirq(memirq_to_xe(memirq)));
303 
304 	return __memirq_status_page(memirq, hwe->instance);
305 }
306 
307 /**
308  * xe_memirq_enable_ptr - Get GGTT's offset of the Interrupt Enable Mask.
309  * @memirq: the &xe_memirq to query
310  *
311  * Shall be called when `Memory Based Interrupts`_ are used
312  * and xe_memirq_init() didn't fail.
313  *
314  * Return: GGTT's offset of the Interrupt Enable Mask.
315  */
xe_memirq_enable_ptr(struct xe_memirq * memirq)316 u32 xe_memirq_enable_ptr(struct xe_memirq *memirq)
317 {
318 	memirq_assert(memirq, xe_device_uses_memirq(memirq_to_xe(memirq)));
319 	memirq_assert(memirq, memirq->bo);
320 
321 	return xe_bo_ggtt_addr(memirq->bo) + XE_MEMIRQ_ENABLE_OFFSET;
322 }
323 
324 /**
325  * xe_memirq_init_guc - Prepare GuC for `Memory Based Interrupts`_.
326  * @memirq: the &xe_memirq
327  * @guc: the &xe_guc to setup
328  *
329  * Register `Interrupt Source Report Page`_ and `Interrupt Status Report Page`_
330  * to be used by the GuC when `Memory Based Interrupts`_ are required.
331  *
332  * Shall be called when `Memory Based Interrupts`_ are used
333  * and xe_memirq_init() didn't fail.
334  *
335  * Return: 0 on success or a negative error code on failure.
336  */
xe_memirq_init_guc(struct xe_memirq * memirq,struct xe_guc * guc)337 int xe_memirq_init_guc(struct xe_memirq *memirq, struct xe_guc *guc)
338 {
339 	bool is_media = xe_gt_is_media_type(guc_to_gt(guc));
340 	u32 offset = is_media ? ilog2(INTR_MGUC) : ilog2(INTR_GUC);
341 	u32 source, status;
342 	int err;
343 
344 	memirq_assert(memirq, xe_device_uses_memirq(memirq_to_xe(memirq)));
345 
346 	source = __memirq_source_page(memirq, 0) + offset;
347 	status = __memirq_status_page(memirq, 0) + offset * SZ_16;
348 
349 	err = xe_guc_self_cfg64(guc, GUC_KLV_SELF_CFG_MEMIRQ_SOURCE_ADDR_KEY,
350 				source);
351 	if (unlikely(err))
352 		goto failed;
353 
354 	err = xe_guc_self_cfg64(guc, GUC_KLV_SELF_CFG_MEMIRQ_STATUS_ADDR_KEY,
355 				status);
356 	if (unlikely(err))
357 		goto failed;
358 
359 	return 0;
360 
361 failed:
362 	memirq_err(memirq, "Failed to setup report pages in %s (%pe)\n",
363 		   guc_name(guc), ERR_PTR(err));
364 	return err;
365 }
366 
367 /**
368  * xe_memirq_reset - Disable processing of `Memory Based Interrupts`_.
369  * @memirq: struct xe_memirq
370  *
371  * This is part of the driver IRQ setup flow.
372  *
373  * This function shall only be used on platforms that use
374  * `Memory Based Interrupts`_.
375  */
xe_memirq_reset(struct xe_memirq * memirq)376 void xe_memirq_reset(struct xe_memirq *memirq)
377 {
378 	memirq_assert(memirq, xe_device_uses_memirq(memirq_to_xe(memirq)));
379 
380 	if (memirq->bo)
381 		memirq_set_enable(memirq, false);
382 }
383 
384 /**
385  * xe_memirq_postinstall - Enable processing of `Memory Based Interrupts`_.
386  * @memirq: the &xe_memirq
387  *
388  * This is part of the driver IRQ setup flow.
389  *
390  * This function shall only be used on platforms that use
391  * `Memory Based Interrupts`_.
392  */
xe_memirq_postinstall(struct xe_memirq * memirq)393 void xe_memirq_postinstall(struct xe_memirq *memirq)
394 {
395 	memirq_assert(memirq, xe_device_uses_memirq(memirq_to_xe(memirq)));
396 
397 	if (memirq->bo)
398 		memirq_set_enable(memirq, true);
399 }
400 
memirq_received(struct xe_memirq * memirq,struct iosys_map * vector,u16 offset,const char * name)401 static bool memirq_received(struct xe_memirq *memirq, struct iosys_map *vector,
402 			    u16 offset, const char *name)
403 {
404 	u8 value;
405 
406 	value = iosys_map_rd(vector, offset, u8);
407 	if (value) {
408 		if (value != 0xff)
409 			memirq_err_ratelimited(memirq,
410 					       "Unexpected memirq value %#x from %s at %u\n",
411 					       value, name, offset);
412 		iosys_map_wr(vector, offset, u8, 0x00);
413 	}
414 
415 	return value;
416 }
417 
memirq_dispatch_engine(struct xe_memirq * memirq,struct iosys_map * status,struct xe_hw_engine * hwe)418 static void memirq_dispatch_engine(struct xe_memirq *memirq, struct iosys_map *status,
419 				   struct xe_hw_engine *hwe)
420 {
421 	memirq_debug(memirq, "STATUS %s %*ph\n", hwe->name, 16, status->vaddr);
422 
423 	if (memirq_received(memirq, status, ilog2(GT_RENDER_USER_INTERRUPT), hwe->name))
424 		xe_hw_engine_handle_irq(hwe, GT_RENDER_USER_INTERRUPT);
425 }
426 
memirq_dispatch_guc(struct xe_memirq * memirq,struct iosys_map * status,struct xe_guc * guc)427 static void memirq_dispatch_guc(struct xe_memirq *memirq, struct iosys_map *status,
428 				struct xe_guc *guc)
429 {
430 	const char *name = guc_name(guc);
431 
432 	memirq_debug(memirq, "STATUS %s %*ph\n", name, 16, status->vaddr);
433 
434 	if (memirq_received(memirq, status, ilog2(GUC_INTR_GUC2HOST), name))
435 		xe_guc_irq_handler(guc, GUC_INTR_GUC2HOST);
436 
437 	if (memirq_received(memirq, status, ilog2(GUC_INTR_SW_INT_0), name))
438 		xe_guc_irq_handler(guc, GUC_INTR_SW_INT_0);
439 }
440 
441 /**
442  * xe_memirq_hwe_handler - Check and process interrupts for a specific HW engine.
443  * @memirq: the &xe_memirq
444  * @hwe: the hw engine to process
445  *
446  * This function reads and dispatches `Memory Based Interrupts` for the provided HW engine.
447  */
xe_memirq_hwe_handler(struct xe_memirq * memirq,struct xe_hw_engine * hwe)448 void xe_memirq_hwe_handler(struct xe_memirq *memirq, struct xe_hw_engine *hwe)
449 {
450 	u16 offset = hwe->irq_offset;
451 	u16 instance = hw_reports_to_instance_zero(memirq) ? hwe->instance : 0;
452 	struct iosys_map src_offset = IOSYS_MAP_INIT_OFFSET(&memirq->bo->vmap,
453 							    XE_MEMIRQ_SOURCE_OFFSET(instance));
454 
455 	if (memirq_received(memirq, &src_offset, offset, "SRC")) {
456 		struct iosys_map status_offset =
457 			IOSYS_MAP_INIT_OFFSET(&memirq->bo->vmap,
458 					      XE_MEMIRQ_STATUS_OFFSET(instance) + offset * SZ_16);
459 		memirq_dispatch_engine(memirq, &status_offset, hwe);
460 	}
461 }
462 
463 /**
464  * xe_memirq_handler - The `Memory Based Interrupts`_ Handler.
465  * @memirq: the &xe_memirq
466  *
467  * This function reads and dispatches `Memory Based Interrupts`.
468  */
xe_memirq_handler(struct xe_memirq * memirq)469 void xe_memirq_handler(struct xe_memirq *memirq)
470 {
471 	struct xe_device *xe = memirq_to_xe(memirq);
472 	struct xe_tile *tile = memirq_to_tile(memirq);
473 	struct xe_hw_engine *hwe;
474 	enum xe_hw_engine_id id;
475 	struct iosys_map map;
476 	unsigned int gtid;
477 	struct xe_gt *gt;
478 
479 	if (!memirq->bo)
480 		return;
481 
482 	memirq_assert(memirq, !memirq->source.is_iomem);
483 	memirq_debug(memirq, "SOURCE %*ph\n", 32, memirq->source.vaddr);
484 	memirq_debug(memirq, "SOURCE %*ph\n", 32, memirq->source.vaddr + 32);
485 
486 	for_each_gt(gt, xe, gtid) {
487 		if (gt->tile != tile)
488 			continue;
489 
490 		for_each_hw_engine(hwe, gt, id)
491 			xe_memirq_hwe_handler(memirq, hwe);
492 	}
493 
494 	/* GuC and media GuC (if present) must be checked separately */
495 
496 	if (memirq_received(memirq, &memirq->source, ilog2(INTR_GUC), "SRC")) {
497 		map = IOSYS_MAP_INIT_OFFSET(&memirq->status, ilog2(INTR_GUC) * SZ_16);
498 		memirq_dispatch_guc(memirq, &map, &tile->primary_gt->uc.guc);
499 	}
500 
501 	if (!tile->media_gt)
502 		return;
503 
504 	if (memirq_received(memirq, &memirq->source, ilog2(INTR_MGUC), "SRC")) {
505 		map = IOSYS_MAP_INIT_OFFSET(&memirq->status, ilog2(INTR_MGUC) * SZ_16);
506 		memirq_dispatch_guc(memirq, &map, &tile->media_gt->uc.guc);
507 	}
508 }
509