xref: /linux/drivers/cxl/core/cdat.c (revision 78c3925c048c752334873f56c3a3d1c9d53e0416)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2023 Intel Corporation. All rights reserved. */
3 #include <linux/acpi.h>
4 #include <linux/xarray.h>
5 #include <linux/fw_table.h>
6 #include <linux/node.h>
7 #include <linux/overflow.h>
8 #include "cxlpci.h"
9 #include "cxlmem.h"
10 #include "core.h"
11 #include "cxl.h"
12 #include "core.h"
13 
14 struct dsmas_entry {
15 	struct range dpa_range;
16 	u8 handle;
17 	struct access_coordinate coord;
18 
19 	int entries;
20 	int qos_class;
21 };
22 
23 static int cdat_dsmas_handler(union acpi_subtable_headers *header, void *arg,
24 			      const unsigned long end)
25 {
26 	struct acpi_cdat_header *hdr = &header->cdat;
27 	struct acpi_cdat_dsmas *dsmas;
28 	int size = sizeof(*hdr) + sizeof(*dsmas);
29 	struct xarray *dsmas_xa = arg;
30 	struct dsmas_entry *dent;
31 	u16 len;
32 	int rc;
33 
34 	len = le16_to_cpu((__force __le16)hdr->length);
35 	if (len != size || (unsigned long)hdr + len > end) {
36 		pr_warn("Malformed DSMAS table length: (%u:%u)\n", size, len);
37 		return -EINVAL;
38 	}
39 
40 	/* Skip common header */
41 	dsmas = (struct acpi_cdat_dsmas *)(hdr + 1);
42 
43 	dent = kzalloc(sizeof(*dent), GFP_KERNEL);
44 	if (!dent)
45 		return -ENOMEM;
46 
47 	dent->handle = dsmas->dsmad_handle;
48 	dent->dpa_range.start = le64_to_cpu((__force __le64)dsmas->dpa_base_address);
49 	dent->dpa_range.end = le64_to_cpu((__force __le64)dsmas->dpa_base_address) +
50 			      le64_to_cpu((__force __le64)dsmas->dpa_length) - 1;
51 
52 	rc = xa_insert(dsmas_xa, dent->handle, dent, GFP_KERNEL);
53 	if (rc) {
54 		kfree(dent);
55 		return rc;
56 	}
57 
58 	return 0;
59 }
60 
61 static void cxl_access_coordinate_set(struct access_coordinate *coord,
62 				      int access, unsigned int val)
63 {
64 	switch (access) {
65 	case ACPI_HMAT_ACCESS_LATENCY:
66 		coord->read_latency = val;
67 		coord->write_latency = val;
68 		break;
69 	case ACPI_HMAT_READ_LATENCY:
70 		coord->read_latency = val;
71 		break;
72 	case ACPI_HMAT_WRITE_LATENCY:
73 		coord->write_latency = val;
74 		break;
75 	case ACPI_HMAT_ACCESS_BANDWIDTH:
76 		coord->read_bandwidth = val;
77 		coord->write_bandwidth = val;
78 		break;
79 	case ACPI_HMAT_READ_BANDWIDTH:
80 		coord->read_bandwidth = val;
81 		break;
82 	case ACPI_HMAT_WRITE_BANDWIDTH:
83 		coord->write_bandwidth = val;
84 		break;
85 	}
86 }
87 
88 static int cdat_dslbis_handler(union acpi_subtable_headers *header, void *arg,
89 			       const unsigned long end)
90 {
91 	struct acpi_cdat_header *hdr = &header->cdat;
92 	struct acpi_cdat_dslbis *dslbis;
93 	int size = sizeof(*hdr) + sizeof(*dslbis);
94 	struct xarray *dsmas_xa = arg;
95 	struct dsmas_entry *dent;
96 	__le64 le_base;
97 	__le16 le_val;
98 	u64 val;
99 	u16 len;
100 	int rc;
101 
102 	len = le16_to_cpu((__force __le16)hdr->length);
103 	if (len != size || (unsigned long)hdr + len > end) {
104 		pr_warn("Malformed DSLBIS table length: (%u:%u)\n", size, len);
105 		return -EINVAL;
106 	}
107 
108 	/* Skip common header */
109 	dslbis = (struct acpi_cdat_dslbis *)(hdr + 1);
110 
111 	/* Skip unrecognized data type */
112 	if (dslbis->data_type > ACPI_HMAT_WRITE_BANDWIDTH)
113 		return 0;
114 
115 	/* Not a memory type, skip */
116 	if ((dslbis->flags & ACPI_HMAT_MEMORY_HIERARCHY) != ACPI_HMAT_MEMORY)
117 		return 0;
118 
119 	dent = xa_load(dsmas_xa, dslbis->handle);
120 	if (!dent) {
121 		pr_warn("No matching DSMAS entry for DSLBIS entry.\n");
122 		return 0;
123 	}
124 
125 	le_base = (__force __le64)dslbis->entry_base_unit;
126 	le_val = (__force __le16)dslbis->entry[0];
127 	rc = check_mul_overflow(le64_to_cpu(le_base),
128 				le16_to_cpu(le_val), &val);
129 	if (rc)
130 		pr_warn("DSLBIS value overflowed.\n");
131 
132 	cxl_access_coordinate_set(&dent->coord, dslbis->data_type, val);
133 
134 	return 0;
135 }
136 
137 static int cdat_table_parse_output(int rc)
138 {
139 	if (rc < 0)
140 		return rc;
141 	if (rc == 0)
142 		return -ENOENT;
143 
144 	return 0;
145 }
146 
147 static int cxl_cdat_endpoint_process(struct cxl_port *port,
148 				     struct xarray *dsmas_xa)
149 {
150 	int rc;
151 
152 	rc = cdat_table_parse(ACPI_CDAT_TYPE_DSMAS, cdat_dsmas_handler,
153 			      dsmas_xa, port->cdat.table, port->cdat.length);
154 	rc = cdat_table_parse_output(rc);
155 	if (rc)
156 		return rc;
157 
158 	rc = cdat_table_parse(ACPI_CDAT_TYPE_DSLBIS, cdat_dslbis_handler,
159 			      dsmas_xa, port->cdat.table, port->cdat.length);
160 	return cdat_table_parse_output(rc);
161 }
162 
163 static int cxl_port_perf_data_calculate(struct cxl_port *port,
164 					struct xarray *dsmas_xa)
165 {
166 	struct access_coordinate ep_c;
167 	struct access_coordinate coord[ACCESS_COORDINATE_MAX];
168 	struct dsmas_entry *dent;
169 	int valid_entries = 0;
170 	unsigned long index;
171 	int rc;
172 
173 	rc = cxl_endpoint_get_perf_coordinates(port, &ep_c);
174 	if (rc) {
175 		dev_dbg(&port->dev, "Failed to retrieve ep perf coordinates.\n");
176 		return rc;
177 	}
178 
179 	rc = cxl_hb_get_perf_coordinates(port, coord);
180 	if (rc)  {
181 		dev_dbg(&port->dev, "Failed to retrieve hb perf coordinates.\n");
182 		return rc;
183 	}
184 
185 	struct cxl_root *cxl_root __free(put_cxl_root) = find_cxl_root(port);
186 
187 	if (!cxl_root)
188 		return -ENODEV;
189 
190 	if (!cxl_root->ops || !cxl_root->ops->qos_class)
191 		return -EOPNOTSUPP;
192 
193 	xa_for_each(dsmas_xa, index, dent) {
194 		int qos_class;
195 
196 		cxl_coordinates_combine(&dent->coord, &dent->coord, &ep_c);
197 		/*
198 		 * Keeping the host bridge coordinates separate from the dsmas
199 		 * coordinates in order to allow calculation of access class
200 		 * 0 and 1 for region later.
201 		 */
202 		cxl_coordinates_combine(&coord[ACCESS_COORDINATE_CPU],
203 					&coord[ACCESS_COORDINATE_CPU],
204 					&dent->coord);
205 		dent->entries = 1;
206 		rc = cxl_root->ops->qos_class(cxl_root,
207 					      &coord[ACCESS_COORDINATE_CPU],
208 					      1, &qos_class);
209 		if (rc != 1)
210 			continue;
211 
212 		valid_entries++;
213 		dent->qos_class = qos_class;
214 	}
215 
216 	if (!valid_entries)
217 		return -ENOENT;
218 
219 	return 0;
220 }
221 
222 static void update_perf_entry(struct device *dev, struct dsmas_entry *dent,
223 			      struct cxl_dpa_perf *dpa_perf)
224 {
225 	dpa_perf->dpa_range = dent->dpa_range;
226 	dpa_perf->coord = dent->coord;
227 	dpa_perf->qos_class = dent->qos_class;
228 	dev_dbg(dev,
229 		"DSMAS: dpa: %#llx qos: %d read_bw: %d write_bw %d read_lat: %d write_lat: %d\n",
230 		dent->dpa_range.start, dpa_perf->qos_class,
231 		dent->coord.read_bandwidth, dent->coord.write_bandwidth,
232 		dent->coord.read_latency, dent->coord.write_latency);
233 }
234 
235 static void cxl_memdev_set_qos_class(struct cxl_dev_state *cxlds,
236 				     struct xarray *dsmas_xa)
237 {
238 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
239 	struct device *dev = cxlds->dev;
240 	struct range pmem_range = {
241 		.start = cxlds->pmem_res.start,
242 		.end = cxlds->pmem_res.end,
243 	};
244 	struct range ram_range = {
245 		.start = cxlds->ram_res.start,
246 		.end = cxlds->ram_res.end,
247 	};
248 	struct dsmas_entry *dent;
249 	unsigned long index;
250 
251 	xa_for_each(dsmas_xa, index, dent) {
252 		if (resource_size(&cxlds->ram_res) &&
253 		    range_contains(&ram_range, &dent->dpa_range))
254 			update_perf_entry(dev, dent, &mds->ram_perf);
255 		else if (resource_size(&cxlds->pmem_res) &&
256 			 range_contains(&pmem_range, &dent->dpa_range))
257 			update_perf_entry(dev, dent, &mds->pmem_perf);
258 		else
259 			dev_dbg(dev, "no partition for dsmas dpa: %#llx\n",
260 				dent->dpa_range.start);
261 	}
262 }
263 
264 static int match_cxlrd_qos_class(struct device *dev, void *data)
265 {
266 	int dev_qos_class = *(int *)data;
267 	struct cxl_root_decoder *cxlrd;
268 
269 	if (!is_root_decoder(dev))
270 		return 0;
271 
272 	cxlrd = to_cxl_root_decoder(dev);
273 	if (cxlrd->qos_class == CXL_QOS_CLASS_INVALID)
274 		return 0;
275 
276 	if (cxlrd->qos_class == dev_qos_class)
277 		return 1;
278 
279 	return 0;
280 }
281 
282 static void reset_dpa_perf(struct cxl_dpa_perf *dpa_perf)
283 {
284 	*dpa_perf = (struct cxl_dpa_perf) {
285 		.qos_class = CXL_QOS_CLASS_INVALID,
286 	};
287 }
288 
289 static bool cxl_qos_match(struct cxl_port *root_port,
290 			  struct cxl_dpa_perf *dpa_perf)
291 {
292 	if (dpa_perf->qos_class == CXL_QOS_CLASS_INVALID)
293 		return false;
294 
295 	if (!device_for_each_child(&root_port->dev, &dpa_perf->qos_class,
296 				   match_cxlrd_qos_class))
297 		return false;
298 
299 	return true;
300 }
301 
302 static int match_cxlrd_hb(struct device *dev, void *data)
303 {
304 	struct device *host_bridge = data;
305 	struct cxl_switch_decoder *cxlsd;
306 	struct cxl_root_decoder *cxlrd;
307 
308 	if (!is_root_decoder(dev))
309 		return 0;
310 
311 	cxlrd = to_cxl_root_decoder(dev);
312 	cxlsd = &cxlrd->cxlsd;
313 
314 	guard(rwsem_read)(&cxl_region_rwsem);
315 	for (int i = 0; i < cxlsd->nr_targets; i++) {
316 		if (host_bridge == cxlsd->target[i]->dport_dev)
317 			return 1;
318 	}
319 
320 	return 0;
321 }
322 
323 static int cxl_qos_class_verify(struct cxl_memdev *cxlmd)
324 {
325 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
326 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
327 	struct cxl_port *root_port;
328 	int rc;
329 
330 	struct cxl_root *cxl_root __free(put_cxl_root) =
331 		find_cxl_root(cxlmd->endpoint);
332 
333 	if (!cxl_root)
334 		return -ENODEV;
335 
336 	root_port = &cxl_root->port;
337 
338 	/* Check that the QTG IDs are all sane between end device and root decoders */
339 	if (!cxl_qos_match(root_port, &mds->ram_perf))
340 		reset_dpa_perf(&mds->ram_perf);
341 	if (!cxl_qos_match(root_port, &mds->pmem_perf))
342 		reset_dpa_perf(&mds->pmem_perf);
343 
344 	/* Check to make sure that the device's host bridge is under a root decoder */
345 	rc = device_for_each_child(&root_port->dev,
346 				   cxlmd->endpoint->host_bridge, match_cxlrd_hb);
347 	if (!rc) {
348 		reset_dpa_perf(&mds->ram_perf);
349 		reset_dpa_perf(&mds->pmem_perf);
350 	}
351 
352 	return rc;
353 }
354 
355 static void discard_dsmas(struct xarray *xa)
356 {
357 	unsigned long index;
358 	void *ent;
359 
360 	xa_for_each(xa, index, ent) {
361 		xa_erase(xa, index);
362 		kfree(ent);
363 	}
364 	xa_destroy(xa);
365 }
366 DEFINE_FREE(dsmas, struct xarray *, if (_T) discard_dsmas(_T))
367 
368 void cxl_endpoint_parse_cdat(struct cxl_port *port)
369 {
370 	struct cxl_memdev *cxlmd = to_cxl_memdev(port->uport_dev);
371 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
372 	struct xarray __dsmas_xa;
373 	struct xarray *dsmas_xa __free(dsmas) = &__dsmas_xa;
374 	int rc;
375 
376 	xa_init(&__dsmas_xa);
377 	if (!port->cdat.table)
378 		return;
379 
380 	rc = cxl_cdat_endpoint_process(port, dsmas_xa);
381 	if (rc < 0) {
382 		dev_dbg(&port->dev, "Failed to parse CDAT: %d\n", rc);
383 		return;
384 	}
385 
386 	rc = cxl_port_perf_data_calculate(port, dsmas_xa);
387 	if (rc) {
388 		dev_dbg(&port->dev, "Failed to do perf coord calculations.\n");
389 		return;
390 	}
391 
392 	cxl_memdev_set_qos_class(cxlds, dsmas_xa);
393 	cxl_qos_class_verify(cxlmd);
394 	cxl_memdev_update_perf(cxlmd);
395 }
396 EXPORT_SYMBOL_NS_GPL(cxl_endpoint_parse_cdat, CXL);
397 
398 static int cdat_sslbis_handler(union acpi_subtable_headers *header, void *arg,
399 			       const unsigned long end)
400 {
401 	struct acpi_cdat_sslbis_table {
402 		struct acpi_cdat_header header;
403 		struct acpi_cdat_sslbis sslbis_header;
404 		struct acpi_cdat_sslbe entries[];
405 	} *tbl = (struct acpi_cdat_sslbis_table *)header;
406 	int size = sizeof(header->cdat) + sizeof(tbl->sslbis_header);
407 	struct acpi_cdat_sslbis *sslbis;
408 	struct cxl_port *port = arg;
409 	struct device *dev = &port->dev;
410 	int remain, entries, i;
411 	u16 len;
412 
413 	len = le16_to_cpu((__force __le16)header->cdat.length);
414 	remain = len - size;
415 	if (!remain || remain % sizeof(tbl->entries[0]) ||
416 	    (unsigned long)header + len > end) {
417 		dev_warn(dev, "Malformed SSLBIS table length: (%u)\n", len);
418 		return -EINVAL;
419 	}
420 
421 	sslbis = &tbl->sslbis_header;
422 	/* Unrecognized data type, we can skip */
423 	if (sslbis->data_type > ACPI_HMAT_WRITE_BANDWIDTH)
424 		return 0;
425 
426 	entries = remain / sizeof(tbl->entries[0]);
427 	if (struct_size(tbl, entries, entries) != len)
428 		return -EINVAL;
429 
430 	for (i = 0; i < entries; i++) {
431 		u16 x = le16_to_cpu((__force __le16)tbl->entries[i].portx_id);
432 		u16 y = le16_to_cpu((__force __le16)tbl->entries[i].porty_id);
433 		__le64 le_base;
434 		__le16 le_val;
435 		struct cxl_dport *dport;
436 		unsigned long index;
437 		u16 dsp_id;
438 		u64 val;
439 
440 		switch (x) {
441 		case ACPI_CDAT_SSLBIS_US_PORT:
442 			dsp_id = y;
443 			break;
444 		case ACPI_CDAT_SSLBIS_ANY_PORT:
445 			switch (y) {
446 			case ACPI_CDAT_SSLBIS_US_PORT:
447 				dsp_id = x;
448 				break;
449 			case ACPI_CDAT_SSLBIS_ANY_PORT:
450 				dsp_id = ACPI_CDAT_SSLBIS_ANY_PORT;
451 				break;
452 			default:
453 				dsp_id = y;
454 				break;
455 			}
456 			break;
457 		default:
458 			dsp_id = x;
459 			break;
460 		}
461 
462 		le_base = (__force __le64)tbl->sslbis_header.entry_base_unit;
463 		le_val = (__force __le16)tbl->entries[i].latency_or_bandwidth;
464 
465 		if (check_mul_overflow(le64_to_cpu(le_base),
466 				       le16_to_cpu(le_val), &val))
467 			dev_warn(dev, "SSLBIS value overflowed!\n");
468 
469 		xa_for_each(&port->dports, index, dport) {
470 			if (dsp_id == ACPI_CDAT_SSLBIS_ANY_PORT ||
471 			    dsp_id == dport->port_id)
472 				cxl_access_coordinate_set(&dport->sw_coord,
473 							  sslbis->data_type,
474 							  val);
475 		}
476 	}
477 
478 	return 0;
479 }
480 
481 void cxl_switch_parse_cdat(struct cxl_port *port)
482 {
483 	int rc;
484 
485 	if (!port->cdat.table)
486 		return;
487 
488 	rc = cdat_table_parse(ACPI_CDAT_TYPE_SSLBIS, cdat_sslbis_handler,
489 			      port, port->cdat.table, port->cdat.length);
490 	rc = cdat_table_parse_output(rc);
491 	if (rc)
492 		dev_dbg(&port->dev, "Failed to parse SSLBIS: %d\n", rc);
493 }
494 EXPORT_SYMBOL_NS_GPL(cxl_switch_parse_cdat, CXL);
495 
496 /**
497  * cxl_coordinates_combine - Combine the two input coordinates
498  *
499  * @out: Output coordinate of c1 and c2 combined
500  * @c1: input coordinates
501  * @c2: input coordinates
502  */
503 void cxl_coordinates_combine(struct access_coordinate *out,
504 			     struct access_coordinate *c1,
505 			     struct access_coordinate *c2)
506 {
507 		if (c1->write_bandwidth && c2->write_bandwidth)
508 			out->write_bandwidth = min(c1->write_bandwidth,
509 						   c2->write_bandwidth);
510 		out->write_latency = c1->write_latency + c2->write_latency;
511 
512 		if (c1->read_bandwidth && c2->read_bandwidth)
513 			out->read_bandwidth = min(c1->read_bandwidth,
514 						  c2->read_bandwidth);
515 		out->read_latency = c1->read_latency + c2->read_latency;
516 }
517 
518 MODULE_IMPORT_NS(CXL);
519 
520 void cxl_region_perf_data_calculate(struct cxl_region *cxlr,
521 				    struct cxl_endpoint_decoder *cxled)
522 {
523 	struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
524 	struct cxl_port *port = cxlmd->endpoint;
525 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
526 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
527 	struct access_coordinate hb_coord[ACCESS_COORDINATE_MAX];
528 	struct access_coordinate coord;
529 	struct range dpa = {
530 			.start = cxled->dpa_res->start,
531 			.end = cxled->dpa_res->end,
532 	};
533 	struct cxl_dpa_perf *perf;
534 	int rc;
535 
536 	switch (cxlr->mode) {
537 	case CXL_DECODER_RAM:
538 		perf = &mds->ram_perf;
539 		break;
540 	case CXL_DECODER_PMEM:
541 		perf = &mds->pmem_perf;
542 		break;
543 	default:
544 		return;
545 	}
546 
547 	lockdep_assert_held(&cxl_dpa_rwsem);
548 
549 	if (!range_contains(&perf->dpa_range, &dpa))
550 		return;
551 
552 	rc = cxl_hb_get_perf_coordinates(port, hb_coord);
553 	if (rc)  {
554 		dev_dbg(&port->dev, "Failed to retrieve hb perf coordinates.\n");
555 		return;
556 	}
557 
558 	for (int i = 0; i < ACCESS_COORDINATE_MAX; i++) {
559 		/* Pickup the host bridge coords */
560 		cxl_coordinates_combine(&coord, &hb_coord[i], &perf->coord);
561 
562 		/* Get total bandwidth and the worst latency for the cxl region */
563 		cxlr->coord[i].read_latency = max_t(unsigned int,
564 						    cxlr->coord[i].read_latency,
565 						    coord.read_latency);
566 		cxlr->coord[i].write_latency = max_t(unsigned int,
567 						     cxlr->coord[i].write_latency,
568 						     coord.write_latency);
569 		cxlr->coord[i].read_bandwidth += coord.read_bandwidth;
570 		cxlr->coord[i].write_bandwidth += coord.write_bandwidth;
571 
572 		/*
573 		 * Convert latency to nanosec from picosec to be consistent
574 		 * with the resulting latency coordinates computed by the
575 		 * HMAT_REPORTING code.
576 		 */
577 		cxlr->coord[i].read_latency =
578 			DIV_ROUND_UP(cxlr->coord[i].read_latency, 1000);
579 		cxlr->coord[i].write_latency =
580 			DIV_ROUND_UP(cxlr->coord[i].write_latency, 1000);
581 	}
582 }
583 
584 int cxl_update_hmat_access_coordinates(int nid, struct cxl_region *cxlr,
585 				       enum access_coordinate_class access)
586 {
587 	return hmat_update_target_coordinates(nid, &cxlr->coord[access], access);
588 }
589 
590 bool cxl_need_node_perf_attrs_update(int nid)
591 {
592 	return !acpi_node_backed_by_real_pxm(nid);
593 }
594