Merge tag 'trace-printf-v6.13' of git://git.kernel.org/pub/scm/linux/kernel/git/trace...
[drm/drm-misc.git] / drivers / cxl / core / region.c
blob70d0a017e99c85c71f410e405970b2c6cdc47b67
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2022 Intel Corporation. All rights reserved. */
3 #include <linux/memregion.h>
4 #include <linux/genalloc.h>
5 #include <linux/device.h>
6 #include <linux/module.h>
7 #include <linux/memory.h>
8 #include <linux/slab.h>
9 #include <linux/uuid.h>
10 #include <linux/sort.h>
11 #include <linux/idr.h>
12 #include <linux/memory-tiers.h>
13 #include <cxlmem.h>
14 #include <cxl.h>
15 #include "core.h"
17 /**
18 * DOC: cxl core region
20 * CXL Regions represent mapped memory capacity in system physical address
21 * space. Whereas the CXL Root Decoders identify the bounds of potential CXL
22 * Memory ranges, Regions represent the active mapped capacity by the HDM
23 * Decoder Capability structures throughout the Host Bridges, Switches, and
24 * Endpoints in the topology.
26 * Region configuration has ordering constraints. UUID may be set at any time
27 * but is only visible for persistent regions.
28 * 1. Interleave granularity
29 * 2. Interleave size
30 * 3. Decoder targets
33 static struct cxl_region *to_cxl_region(struct device *dev);
35 #define __ACCESS_ATTR_RO(_level, _name) { \
36 .attr = { .name = __stringify(_name), .mode = 0444 }, \
37 .show = _name##_access##_level##_show, \
40 #define ACCESS_DEVICE_ATTR_RO(level, name) \
41 struct device_attribute dev_attr_access##level##_##name = __ACCESS_ATTR_RO(level, name)
43 #define ACCESS_ATTR_RO(level, attrib) \
44 static ssize_t attrib##_access##level##_show(struct device *dev, \
45 struct device_attribute *attr, \
46 char *buf) \
47 { \
48 struct cxl_region *cxlr = to_cxl_region(dev); \
50 if (cxlr->coord[level].attrib == 0) \
51 return -ENOENT; \
53 return sysfs_emit(buf, "%u\n", cxlr->coord[level].attrib); \
54 } \
55 static ACCESS_DEVICE_ATTR_RO(level, attrib)
57 ACCESS_ATTR_RO(0, read_bandwidth);
58 ACCESS_ATTR_RO(0, read_latency);
59 ACCESS_ATTR_RO(0, write_bandwidth);
60 ACCESS_ATTR_RO(0, write_latency);
62 #define ACCESS_ATTR_DECLARE(level, attrib) \
63 (&dev_attr_access##level##_##attrib.attr)
65 static struct attribute *access0_coordinate_attrs[] = {
66 ACCESS_ATTR_DECLARE(0, read_bandwidth),
67 ACCESS_ATTR_DECLARE(0, write_bandwidth),
68 ACCESS_ATTR_DECLARE(0, read_latency),
69 ACCESS_ATTR_DECLARE(0, write_latency),
70 NULL
73 ACCESS_ATTR_RO(1, read_bandwidth);
74 ACCESS_ATTR_RO(1, read_latency);
75 ACCESS_ATTR_RO(1, write_bandwidth);
76 ACCESS_ATTR_RO(1, write_latency);
78 static struct attribute *access1_coordinate_attrs[] = {
79 ACCESS_ATTR_DECLARE(1, read_bandwidth),
80 ACCESS_ATTR_DECLARE(1, write_bandwidth),
81 ACCESS_ATTR_DECLARE(1, read_latency),
82 ACCESS_ATTR_DECLARE(1, write_latency),
83 NULL
86 #define ACCESS_VISIBLE(level) \
87 static umode_t cxl_region_access##level##_coordinate_visible( \
88 struct kobject *kobj, struct attribute *a, int n) \
89 { \
90 struct device *dev = kobj_to_dev(kobj); \
91 struct cxl_region *cxlr = to_cxl_region(dev); \
93 if (a == &dev_attr_access##level##_read_latency.attr && \
94 cxlr->coord[level].read_latency == 0) \
95 return 0; \
97 if (a == &dev_attr_access##level##_write_latency.attr && \
98 cxlr->coord[level].write_latency == 0) \
99 return 0; \
101 if (a == &dev_attr_access##level##_read_bandwidth.attr && \
102 cxlr->coord[level].read_bandwidth == 0) \
103 return 0; \
105 if (a == &dev_attr_access##level##_write_bandwidth.attr && \
106 cxlr->coord[level].write_bandwidth == 0) \
107 return 0; \
109 return a->mode; \
112 ACCESS_VISIBLE(0);
113 ACCESS_VISIBLE(1);
115 static const struct attribute_group cxl_region_access0_coordinate_group = {
116 .name = "access0",
117 .attrs = access0_coordinate_attrs,
118 .is_visible = cxl_region_access0_coordinate_visible,
121 static const struct attribute_group *get_cxl_region_access0_group(void)
123 return &cxl_region_access0_coordinate_group;
126 static const struct attribute_group cxl_region_access1_coordinate_group = {
127 .name = "access1",
128 .attrs = access1_coordinate_attrs,
129 .is_visible = cxl_region_access1_coordinate_visible,
132 static const struct attribute_group *get_cxl_region_access1_group(void)
134 return &cxl_region_access1_coordinate_group;
137 static ssize_t uuid_show(struct device *dev, struct device_attribute *attr,
138 char *buf)
140 struct cxl_region *cxlr = to_cxl_region(dev);
141 struct cxl_region_params *p = &cxlr->params;
142 ssize_t rc;
144 rc = down_read_interruptible(&cxl_region_rwsem);
145 if (rc)
146 return rc;
147 if (cxlr->mode != CXL_DECODER_PMEM)
148 rc = sysfs_emit(buf, "\n");
149 else
150 rc = sysfs_emit(buf, "%pUb\n", &p->uuid);
151 up_read(&cxl_region_rwsem);
153 return rc;
156 static int is_dup(struct device *match, void *data)
158 struct cxl_region_params *p;
159 struct cxl_region *cxlr;
160 uuid_t *uuid = data;
162 if (!is_cxl_region(match))
163 return 0;
165 lockdep_assert_held(&cxl_region_rwsem);
166 cxlr = to_cxl_region(match);
167 p = &cxlr->params;
169 if (uuid_equal(&p->uuid, uuid)) {
170 dev_dbg(match, "already has uuid: %pUb\n", uuid);
171 return -EBUSY;
174 return 0;
177 static ssize_t uuid_store(struct device *dev, struct device_attribute *attr,
178 const char *buf, size_t len)
180 struct cxl_region *cxlr = to_cxl_region(dev);
181 struct cxl_region_params *p = &cxlr->params;
182 uuid_t temp;
183 ssize_t rc;
185 if (len != UUID_STRING_LEN + 1)
186 return -EINVAL;
188 rc = uuid_parse(buf, &temp);
189 if (rc)
190 return rc;
192 if (uuid_is_null(&temp))
193 return -EINVAL;
195 rc = down_write_killable(&cxl_region_rwsem);
196 if (rc)
197 return rc;
199 if (uuid_equal(&p->uuid, &temp))
200 goto out;
202 rc = -EBUSY;
203 if (p->state >= CXL_CONFIG_ACTIVE)
204 goto out;
206 rc = bus_for_each_dev(&cxl_bus_type, NULL, &temp, is_dup);
207 if (rc < 0)
208 goto out;
210 uuid_copy(&p->uuid, &temp);
211 out:
212 up_write(&cxl_region_rwsem);
214 if (rc)
215 return rc;
216 return len;
218 static DEVICE_ATTR_RW(uuid);
220 static struct cxl_region_ref *cxl_rr_load(struct cxl_port *port,
221 struct cxl_region *cxlr)
223 return xa_load(&port->regions, (unsigned long)cxlr);
226 static int cxl_region_invalidate_memregion(struct cxl_region *cxlr)
228 if (!cpu_cache_has_invalidate_memregion()) {
229 if (IS_ENABLED(CONFIG_CXL_REGION_INVALIDATION_TEST)) {
230 dev_info_once(
231 &cxlr->dev,
232 "Bypassing cpu_cache_invalidate_memregion() for testing!\n");
233 return 0;
234 } else {
235 dev_WARN(&cxlr->dev,
236 "Failed to synchronize CPU cache state\n");
237 return -ENXIO;
241 cpu_cache_invalidate_memregion(IORES_DESC_CXL);
242 return 0;
245 static void cxl_region_decode_reset(struct cxl_region *cxlr, int count)
247 struct cxl_region_params *p = &cxlr->params;
248 int i;
251 * Before region teardown attempt to flush, evict any data cached for
252 * this region, or scream loudly about missing arch / platform support
253 * for CXL teardown.
255 cxl_region_invalidate_memregion(cxlr);
257 for (i = count - 1; i >= 0; i--) {
258 struct cxl_endpoint_decoder *cxled = p->targets[i];
259 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
260 struct cxl_port *iter = cxled_to_port(cxled);
261 struct cxl_dev_state *cxlds = cxlmd->cxlds;
262 struct cxl_ep *ep;
264 if (cxlds->rcd)
265 goto endpoint_reset;
267 while (!is_cxl_root(to_cxl_port(iter->dev.parent)))
268 iter = to_cxl_port(iter->dev.parent);
270 for (ep = cxl_ep_load(iter, cxlmd); iter;
271 iter = ep->next, ep = cxl_ep_load(iter, cxlmd)) {
272 struct cxl_region_ref *cxl_rr;
273 struct cxl_decoder *cxld;
275 cxl_rr = cxl_rr_load(iter, cxlr);
276 cxld = cxl_rr->decoder;
277 if (cxld->reset)
278 cxld->reset(cxld);
279 set_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags);
282 endpoint_reset:
283 cxled->cxld.reset(&cxled->cxld);
284 set_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags);
287 /* all decoders associated with this region have been torn down */
288 clear_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags);
291 static int commit_decoder(struct cxl_decoder *cxld)
293 struct cxl_switch_decoder *cxlsd = NULL;
295 if (cxld->commit)
296 return cxld->commit(cxld);
298 if (is_switch_decoder(&cxld->dev))
299 cxlsd = to_cxl_switch_decoder(&cxld->dev);
301 if (dev_WARN_ONCE(&cxld->dev, !cxlsd || cxlsd->nr_targets > 1,
302 "->commit() is required\n"))
303 return -ENXIO;
304 return 0;
307 static int cxl_region_decode_commit(struct cxl_region *cxlr)
309 struct cxl_region_params *p = &cxlr->params;
310 int i, rc = 0;
312 for (i = 0; i < p->nr_targets; i++) {
313 struct cxl_endpoint_decoder *cxled = p->targets[i];
314 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
315 struct cxl_region_ref *cxl_rr;
316 struct cxl_decoder *cxld;
317 struct cxl_port *iter;
318 struct cxl_ep *ep;
320 /* commit bottom up */
321 for (iter = cxled_to_port(cxled); !is_cxl_root(iter);
322 iter = to_cxl_port(iter->dev.parent)) {
323 cxl_rr = cxl_rr_load(iter, cxlr);
324 cxld = cxl_rr->decoder;
325 rc = commit_decoder(cxld);
326 if (rc)
327 break;
330 if (rc) {
331 /* programming @iter failed, teardown */
332 for (ep = cxl_ep_load(iter, cxlmd); ep && iter;
333 iter = ep->next, ep = cxl_ep_load(iter, cxlmd)) {
334 cxl_rr = cxl_rr_load(iter, cxlr);
335 cxld = cxl_rr->decoder;
336 if (cxld->reset)
337 cxld->reset(cxld);
340 cxled->cxld.reset(&cxled->cxld);
341 goto err;
345 return 0;
347 err:
348 /* undo the targets that were successfully committed */
349 cxl_region_decode_reset(cxlr, i);
350 return rc;
353 static ssize_t commit_store(struct device *dev, struct device_attribute *attr,
354 const char *buf, size_t len)
356 struct cxl_region *cxlr = to_cxl_region(dev);
357 struct cxl_region_params *p = &cxlr->params;
358 bool commit;
359 ssize_t rc;
361 rc = kstrtobool(buf, &commit);
362 if (rc)
363 return rc;
365 rc = down_write_killable(&cxl_region_rwsem);
366 if (rc)
367 return rc;
369 /* Already in the requested state? */
370 if (commit && p->state >= CXL_CONFIG_COMMIT)
371 goto out;
372 if (!commit && p->state < CXL_CONFIG_COMMIT)
373 goto out;
375 /* Not ready to commit? */
376 if (commit && p->state < CXL_CONFIG_ACTIVE) {
377 rc = -ENXIO;
378 goto out;
382 * Invalidate caches before region setup to drop any speculative
383 * consumption of this address space
385 rc = cxl_region_invalidate_memregion(cxlr);
386 if (rc)
387 goto out;
389 if (commit) {
390 rc = cxl_region_decode_commit(cxlr);
391 if (rc == 0)
392 p->state = CXL_CONFIG_COMMIT;
393 } else {
394 p->state = CXL_CONFIG_RESET_PENDING;
395 up_write(&cxl_region_rwsem);
396 device_release_driver(&cxlr->dev);
397 down_write(&cxl_region_rwsem);
400 * The lock was dropped, so need to revalidate that the reset is
401 * still pending.
403 if (p->state == CXL_CONFIG_RESET_PENDING) {
404 cxl_region_decode_reset(cxlr, p->interleave_ways);
405 p->state = CXL_CONFIG_ACTIVE;
409 out:
410 up_write(&cxl_region_rwsem);
412 if (rc)
413 return rc;
414 return len;
417 static ssize_t commit_show(struct device *dev, struct device_attribute *attr,
418 char *buf)
420 struct cxl_region *cxlr = to_cxl_region(dev);
421 struct cxl_region_params *p = &cxlr->params;
422 ssize_t rc;
424 rc = down_read_interruptible(&cxl_region_rwsem);
425 if (rc)
426 return rc;
427 rc = sysfs_emit(buf, "%d\n", p->state >= CXL_CONFIG_COMMIT);
428 up_read(&cxl_region_rwsem);
430 return rc;
432 static DEVICE_ATTR_RW(commit);
434 static umode_t cxl_region_visible(struct kobject *kobj, struct attribute *a,
435 int n)
437 struct device *dev = kobj_to_dev(kobj);
438 struct cxl_region *cxlr = to_cxl_region(dev);
441 * Support tooling that expects to find a 'uuid' attribute for all
442 * regions regardless of mode.
444 if (a == &dev_attr_uuid.attr && cxlr->mode != CXL_DECODER_PMEM)
445 return 0444;
446 return a->mode;
449 static ssize_t interleave_ways_show(struct device *dev,
450 struct device_attribute *attr, char *buf)
452 struct cxl_region *cxlr = to_cxl_region(dev);
453 struct cxl_region_params *p = &cxlr->params;
454 ssize_t rc;
456 rc = down_read_interruptible(&cxl_region_rwsem);
457 if (rc)
458 return rc;
459 rc = sysfs_emit(buf, "%d\n", p->interleave_ways);
460 up_read(&cxl_region_rwsem);
462 return rc;
465 static const struct attribute_group *get_cxl_region_target_group(void);
467 static ssize_t interleave_ways_store(struct device *dev,
468 struct device_attribute *attr,
469 const char *buf, size_t len)
471 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(dev->parent);
472 struct cxl_decoder *cxld = &cxlrd->cxlsd.cxld;
473 struct cxl_region *cxlr = to_cxl_region(dev);
474 struct cxl_region_params *p = &cxlr->params;
475 unsigned int val, save;
476 int rc;
477 u8 iw;
479 rc = kstrtouint(buf, 0, &val);
480 if (rc)
481 return rc;
483 rc = ways_to_eiw(val, &iw);
484 if (rc)
485 return rc;
488 * Even for x3, x6, and x12 interleaves the region interleave must be a
489 * power of 2 multiple of the host bridge interleave.
491 if (!is_power_of_2(val / cxld->interleave_ways) ||
492 (val % cxld->interleave_ways)) {
493 dev_dbg(&cxlr->dev, "invalid interleave: %d\n", val);
494 return -EINVAL;
497 rc = down_write_killable(&cxl_region_rwsem);
498 if (rc)
499 return rc;
500 if (p->state >= CXL_CONFIG_INTERLEAVE_ACTIVE) {
501 rc = -EBUSY;
502 goto out;
505 save = p->interleave_ways;
506 p->interleave_ways = val;
507 rc = sysfs_update_group(&cxlr->dev.kobj, get_cxl_region_target_group());
508 if (rc)
509 p->interleave_ways = save;
510 out:
511 up_write(&cxl_region_rwsem);
512 if (rc)
513 return rc;
514 return len;
516 static DEVICE_ATTR_RW(interleave_ways);
518 static ssize_t interleave_granularity_show(struct device *dev,
519 struct device_attribute *attr,
520 char *buf)
522 struct cxl_region *cxlr = to_cxl_region(dev);
523 struct cxl_region_params *p = &cxlr->params;
524 ssize_t rc;
526 rc = down_read_interruptible(&cxl_region_rwsem);
527 if (rc)
528 return rc;
529 rc = sysfs_emit(buf, "%d\n", p->interleave_granularity);
530 up_read(&cxl_region_rwsem);
532 return rc;
535 static ssize_t interleave_granularity_store(struct device *dev,
536 struct device_attribute *attr,
537 const char *buf, size_t len)
539 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(dev->parent);
540 struct cxl_decoder *cxld = &cxlrd->cxlsd.cxld;
541 struct cxl_region *cxlr = to_cxl_region(dev);
542 struct cxl_region_params *p = &cxlr->params;
543 int rc, val;
544 u16 ig;
546 rc = kstrtoint(buf, 0, &val);
547 if (rc)
548 return rc;
550 rc = granularity_to_eig(val, &ig);
551 if (rc)
552 return rc;
555 * When the host-bridge is interleaved, disallow region granularity !=
556 * root granularity. Regions with a granularity less than the root
557 * interleave result in needing multiple endpoints to support a single
558 * slot in the interleave (possible to support in the future). Regions
559 * with a granularity greater than the root interleave result in invalid
560 * DPA translations (invalid to support).
562 if (cxld->interleave_ways > 1 && val != cxld->interleave_granularity)
563 return -EINVAL;
565 rc = down_write_killable(&cxl_region_rwsem);
566 if (rc)
567 return rc;
568 if (p->state >= CXL_CONFIG_INTERLEAVE_ACTIVE) {
569 rc = -EBUSY;
570 goto out;
573 p->interleave_granularity = val;
574 out:
575 up_write(&cxl_region_rwsem);
576 if (rc)
577 return rc;
578 return len;
580 static DEVICE_ATTR_RW(interleave_granularity);
582 static ssize_t resource_show(struct device *dev, struct device_attribute *attr,
583 char *buf)
585 struct cxl_region *cxlr = to_cxl_region(dev);
586 struct cxl_region_params *p = &cxlr->params;
587 u64 resource = -1ULL;
588 ssize_t rc;
590 rc = down_read_interruptible(&cxl_region_rwsem);
591 if (rc)
592 return rc;
593 if (p->res)
594 resource = p->res->start;
595 rc = sysfs_emit(buf, "%#llx\n", resource);
596 up_read(&cxl_region_rwsem);
598 return rc;
600 static DEVICE_ATTR_RO(resource);
602 static ssize_t mode_show(struct device *dev, struct device_attribute *attr,
603 char *buf)
605 struct cxl_region *cxlr = to_cxl_region(dev);
607 return sysfs_emit(buf, "%s\n", cxl_decoder_mode_name(cxlr->mode));
609 static DEVICE_ATTR_RO(mode);
611 static int alloc_hpa(struct cxl_region *cxlr, resource_size_t size)
613 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(cxlr->dev.parent);
614 struct cxl_region_params *p = &cxlr->params;
615 struct resource *res;
616 u64 remainder = 0;
618 lockdep_assert_held_write(&cxl_region_rwsem);
620 /* Nothing to do... */
621 if (p->res && resource_size(p->res) == size)
622 return 0;
624 /* To change size the old size must be freed first */
625 if (p->res)
626 return -EBUSY;
628 if (p->state >= CXL_CONFIG_INTERLEAVE_ACTIVE)
629 return -EBUSY;
631 /* ways, granularity and uuid (if PMEM) need to be set before HPA */
632 if (!p->interleave_ways || !p->interleave_granularity ||
633 (cxlr->mode == CXL_DECODER_PMEM && uuid_is_null(&p->uuid)))
634 return -ENXIO;
636 div64_u64_rem(size, (u64)SZ_256M * p->interleave_ways, &remainder);
637 if (remainder)
638 return -EINVAL;
640 res = alloc_free_mem_region(cxlrd->res, size, SZ_256M,
641 dev_name(&cxlr->dev));
642 if (IS_ERR(res)) {
643 dev_dbg(&cxlr->dev,
644 "HPA allocation error (%ld) for size:%pap in %s %pr\n",
645 PTR_ERR(res), &size, cxlrd->res->name, cxlrd->res);
646 return PTR_ERR(res);
649 p->res = res;
650 p->state = CXL_CONFIG_INTERLEAVE_ACTIVE;
652 return 0;
655 static void cxl_region_iomem_release(struct cxl_region *cxlr)
657 struct cxl_region_params *p = &cxlr->params;
659 if (device_is_registered(&cxlr->dev))
660 lockdep_assert_held_write(&cxl_region_rwsem);
661 if (p->res) {
663 * Autodiscovered regions may not have been able to insert their
664 * resource.
666 if (p->res->parent)
667 remove_resource(p->res);
668 kfree(p->res);
669 p->res = NULL;
673 static int free_hpa(struct cxl_region *cxlr)
675 struct cxl_region_params *p = &cxlr->params;
677 lockdep_assert_held_write(&cxl_region_rwsem);
679 if (!p->res)
680 return 0;
682 if (p->state >= CXL_CONFIG_ACTIVE)
683 return -EBUSY;
685 cxl_region_iomem_release(cxlr);
686 p->state = CXL_CONFIG_IDLE;
687 return 0;
690 static ssize_t size_store(struct device *dev, struct device_attribute *attr,
691 const char *buf, size_t len)
693 struct cxl_region *cxlr = to_cxl_region(dev);
694 u64 val;
695 int rc;
697 rc = kstrtou64(buf, 0, &val);
698 if (rc)
699 return rc;
701 rc = down_write_killable(&cxl_region_rwsem);
702 if (rc)
703 return rc;
705 if (val)
706 rc = alloc_hpa(cxlr, val);
707 else
708 rc = free_hpa(cxlr);
709 up_write(&cxl_region_rwsem);
711 if (rc)
712 return rc;
714 return len;
717 static ssize_t size_show(struct device *dev, struct device_attribute *attr,
718 char *buf)
720 struct cxl_region *cxlr = to_cxl_region(dev);
721 struct cxl_region_params *p = &cxlr->params;
722 u64 size = 0;
723 ssize_t rc;
725 rc = down_read_interruptible(&cxl_region_rwsem);
726 if (rc)
727 return rc;
728 if (p->res)
729 size = resource_size(p->res);
730 rc = sysfs_emit(buf, "%#llx\n", size);
731 up_read(&cxl_region_rwsem);
733 return rc;
735 static DEVICE_ATTR_RW(size);
737 static struct attribute *cxl_region_attrs[] = {
738 &dev_attr_uuid.attr,
739 &dev_attr_commit.attr,
740 &dev_attr_interleave_ways.attr,
741 &dev_attr_interleave_granularity.attr,
742 &dev_attr_resource.attr,
743 &dev_attr_size.attr,
744 &dev_attr_mode.attr,
745 NULL,
748 static const struct attribute_group cxl_region_group = {
749 .attrs = cxl_region_attrs,
750 .is_visible = cxl_region_visible,
753 static size_t show_targetN(struct cxl_region *cxlr, char *buf, int pos)
755 struct cxl_region_params *p = &cxlr->params;
756 struct cxl_endpoint_decoder *cxled;
757 int rc;
759 rc = down_read_interruptible(&cxl_region_rwsem);
760 if (rc)
761 return rc;
763 if (pos >= p->interleave_ways) {
764 dev_dbg(&cxlr->dev, "position %d out of range %d\n", pos,
765 p->interleave_ways);
766 rc = -ENXIO;
767 goto out;
770 cxled = p->targets[pos];
771 if (!cxled)
772 rc = sysfs_emit(buf, "\n");
773 else
774 rc = sysfs_emit(buf, "%s\n", dev_name(&cxled->cxld.dev));
775 out:
776 up_read(&cxl_region_rwsem);
778 return rc;
781 static int check_commit_order(struct device *dev, const void *data)
783 struct cxl_decoder *cxld = to_cxl_decoder(dev);
786 * if port->commit_end is not the only free decoder, then out of
787 * order shutdown has occurred, block further allocations until
788 * that is resolved
790 if (((cxld->flags & CXL_DECODER_F_ENABLE) == 0))
791 return -EBUSY;
792 return 0;
795 static int match_free_decoder(struct device *dev, void *data)
797 struct cxl_port *port = to_cxl_port(dev->parent);
798 struct cxl_decoder *cxld;
799 int rc;
801 if (!is_switch_decoder(dev))
802 return 0;
804 cxld = to_cxl_decoder(dev);
806 if (cxld->id != port->commit_end + 1)
807 return 0;
809 if (cxld->region) {
810 dev_dbg(dev->parent,
811 "next decoder to commit (%s) is already reserved (%s)\n",
812 dev_name(dev), dev_name(&cxld->region->dev));
813 return 0;
816 rc = device_for_each_child_reverse_from(dev->parent, dev, NULL,
817 check_commit_order);
818 if (rc) {
819 dev_dbg(dev->parent,
820 "unable to allocate %s due to out of order shutdown\n",
821 dev_name(dev));
822 return 0;
824 return 1;
827 static int match_auto_decoder(struct device *dev, void *data)
829 struct cxl_region_params *p = data;
830 struct cxl_decoder *cxld;
831 struct range *r;
833 if (!is_switch_decoder(dev))
834 return 0;
836 cxld = to_cxl_decoder(dev);
837 r = &cxld->hpa_range;
839 if (p->res && p->res->start == r->start && p->res->end == r->end)
840 return 1;
842 return 0;
845 static struct cxl_decoder *
846 cxl_region_find_decoder(struct cxl_port *port,
847 struct cxl_endpoint_decoder *cxled,
848 struct cxl_region *cxlr)
850 struct device *dev;
852 if (port == cxled_to_port(cxled))
853 return &cxled->cxld;
855 if (test_bit(CXL_REGION_F_AUTO, &cxlr->flags))
856 dev = device_find_child(&port->dev, &cxlr->params,
857 match_auto_decoder);
858 else
859 dev = device_find_child(&port->dev, NULL, match_free_decoder);
860 if (!dev)
861 return NULL;
863 * This decoder is pinned registered as long as the endpoint decoder is
864 * registered, and endpoint decoder unregistration holds the
865 * cxl_region_rwsem over unregister events, so no need to hold on to
866 * this extra reference.
868 put_device(dev);
869 return to_cxl_decoder(dev);
872 static bool auto_order_ok(struct cxl_port *port, struct cxl_region *cxlr_iter,
873 struct cxl_decoder *cxld)
875 struct cxl_region_ref *rr = cxl_rr_load(port, cxlr_iter);
876 struct cxl_decoder *cxld_iter = rr->decoder;
879 * Allow the out of order assembly of auto-discovered regions.
880 * Per CXL Spec 3.1 8.2.4.20.12 software must commit decoders
881 * in HPA order. Confirm that the decoder with the lesser HPA
882 * starting address has the lesser id.
884 dev_dbg(&cxld->dev, "check for HPA violation %s:%d < %s:%d\n",
885 dev_name(&cxld->dev), cxld->id,
886 dev_name(&cxld_iter->dev), cxld_iter->id);
888 if (cxld_iter->id > cxld->id)
889 return true;
891 return false;
894 static struct cxl_region_ref *
895 alloc_region_ref(struct cxl_port *port, struct cxl_region *cxlr,
896 struct cxl_endpoint_decoder *cxled)
898 struct cxl_region_params *p = &cxlr->params;
899 struct cxl_region_ref *cxl_rr, *iter;
900 unsigned long index;
901 int rc;
903 xa_for_each(&port->regions, index, iter) {
904 struct cxl_region_params *ip = &iter->region->params;
906 if (!ip->res || ip->res->start < p->res->start)
907 continue;
909 if (test_bit(CXL_REGION_F_AUTO, &cxlr->flags)) {
910 struct cxl_decoder *cxld;
912 cxld = cxl_region_find_decoder(port, cxled, cxlr);
913 if (auto_order_ok(port, iter->region, cxld))
914 continue;
916 dev_dbg(&cxlr->dev, "%s: HPA order violation %s:%pr vs %pr\n",
917 dev_name(&port->dev),
918 dev_name(&iter->region->dev), ip->res, p->res);
920 return ERR_PTR(-EBUSY);
923 cxl_rr = kzalloc(sizeof(*cxl_rr), GFP_KERNEL);
924 if (!cxl_rr)
925 return ERR_PTR(-ENOMEM);
926 cxl_rr->port = port;
927 cxl_rr->region = cxlr;
928 cxl_rr->nr_targets = 1;
929 xa_init(&cxl_rr->endpoints);
931 rc = xa_insert(&port->regions, (unsigned long)cxlr, cxl_rr, GFP_KERNEL);
932 if (rc) {
933 dev_dbg(&cxlr->dev,
934 "%s: failed to track region reference: %d\n",
935 dev_name(&port->dev), rc);
936 kfree(cxl_rr);
937 return ERR_PTR(rc);
940 return cxl_rr;
943 static void cxl_rr_free_decoder(struct cxl_region_ref *cxl_rr)
945 struct cxl_region *cxlr = cxl_rr->region;
946 struct cxl_decoder *cxld = cxl_rr->decoder;
948 if (!cxld)
949 return;
951 dev_WARN_ONCE(&cxlr->dev, cxld->region != cxlr, "region mismatch\n");
952 if (cxld->region == cxlr) {
953 cxld->region = NULL;
954 put_device(&cxlr->dev);
958 static void free_region_ref(struct cxl_region_ref *cxl_rr)
960 struct cxl_port *port = cxl_rr->port;
961 struct cxl_region *cxlr = cxl_rr->region;
963 cxl_rr_free_decoder(cxl_rr);
964 xa_erase(&port->regions, (unsigned long)cxlr);
965 xa_destroy(&cxl_rr->endpoints);
966 kfree(cxl_rr);
969 static int cxl_rr_ep_add(struct cxl_region_ref *cxl_rr,
970 struct cxl_endpoint_decoder *cxled)
972 int rc;
973 struct cxl_port *port = cxl_rr->port;
974 struct cxl_region *cxlr = cxl_rr->region;
975 struct cxl_decoder *cxld = cxl_rr->decoder;
976 struct cxl_ep *ep = cxl_ep_load(port, cxled_to_memdev(cxled));
978 if (ep) {
979 rc = xa_insert(&cxl_rr->endpoints, (unsigned long)cxled, ep,
980 GFP_KERNEL);
981 if (rc)
982 return rc;
984 cxl_rr->nr_eps++;
986 if (!cxld->region) {
987 cxld->region = cxlr;
988 get_device(&cxlr->dev);
991 return 0;
994 static int cxl_rr_alloc_decoder(struct cxl_port *port, struct cxl_region *cxlr,
995 struct cxl_endpoint_decoder *cxled,
996 struct cxl_region_ref *cxl_rr)
998 struct cxl_decoder *cxld;
1000 cxld = cxl_region_find_decoder(port, cxled, cxlr);
1001 if (!cxld) {
1002 dev_dbg(&cxlr->dev, "%s: no decoder available\n",
1003 dev_name(&port->dev));
1004 return -EBUSY;
1007 if (cxld->region) {
1008 dev_dbg(&cxlr->dev, "%s: %s already attached to %s\n",
1009 dev_name(&port->dev), dev_name(&cxld->dev),
1010 dev_name(&cxld->region->dev));
1011 return -EBUSY;
1015 * Endpoints should already match the region type, but backstop that
1016 * assumption with an assertion. Switch-decoders change mapping-type
1017 * based on what is mapped when they are assigned to a region.
1019 dev_WARN_ONCE(&cxlr->dev,
1020 port == cxled_to_port(cxled) &&
1021 cxld->target_type != cxlr->type,
1022 "%s:%s mismatch decoder type %d -> %d\n",
1023 dev_name(&cxled_to_memdev(cxled)->dev),
1024 dev_name(&cxld->dev), cxld->target_type, cxlr->type);
1025 cxld->target_type = cxlr->type;
1026 cxl_rr->decoder = cxld;
1027 return 0;
1031 * cxl_port_attach_region() - track a region's interest in a port by endpoint
1032 * @port: port to add a new region reference 'struct cxl_region_ref'
1033 * @cxlr: region to attach to @port
1034 * @cxled: endpoint decoder used to create or further pin a region reference
1035 * @pos: interleave position of @cxled in @cxlr
1037 * The attach event is an opportunity to validate CXL decode setup
1038 * constraints and record metadata needed for programming HDM decoders,
1039 * in particular decoder target lists.
1041 * The steps are:
1043 * - validate that there are no other regions with a higher HPA already
1044 * associated with @port
1045 * - establish a region reference if one is not already present
1047 * - additionally allocate a decoder instance that will host @cxlr on
1048 * @port
1050 * - pin the region reference by the endpoint
1051 * - account for how many entries in @port's target list are needed to
1052 * cover all of the added endpoints.
1054 static int cxl_port_attach_region(struct cxl_port *port,
1055 struct cxl_region *cxlr,
1056 struct cxl_endpoint_decoder *cxled, int pos)
1058 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
1059 struct cxl_ep *ep = cxl_ep_load(port, cxlmd);
1060 struct cxl_region_ref *cxl_rr;
1061 bool nr_targets_inc = false;
1062 struct cxl_decoder *cxld;
1063 unsigned long index;
1064 int rc = -EBUSY;
1066 lockdep_assert_held_write(&cxl_region_rwsem);
1068 cxl_rr = cxl_rr_load(port, cxlr);
1069 if (cxl_rr) {
1070 struct cxl_ep *ep_iter;
1071 int found = 0;
1074 * Walk the existing endpoints that have been attached to
1075 * @cxlr at @port and see if they share the same 'next' port
1076 * in the downstream direction. I.e. endpoints that share common
1077 * upstream switch.
1079 xa_for_each(&cxl_rr->endpoints, index, ep_iter) {
1080 if (ep_iter == ep)
1081 continue;
1082 if (ep_iter->next == ep->next) {
1083 found++;
1084 break;
1089 * New target port, or @port is an endpoint port that always
1090 * accounts its own local decode as a target.
1092 if (!found || !ep->next) {
1093 cxl_rr->nr_targets++;
1094 nr_targets_inc = true;
1096 } else {
1097 cxl_rr = alloc_region_ref(port, cxlr, cxled);
1098 if (IS_ERR(cxl_rr)) {
1099 dev_dbg(&cxlr->dev,
1100 "%s: failed to allocate region reference\n",
1101 dev_name(&port->dev));
1102 return PTR_ERR(cxl_rr);
1104 nr_targets_inc = true;
1106 rc = cxl_rr_alloc_decoder(port, cxlr, cxled, cxl_rr);
1107 if (rc)
1108 goto out_erase;
1110 cxld = cxl_rr->decoder;
1113 * the number of targets should not exceed the target_count
1114 * of the decoder
1116 if (is_switch_decoder(&cxld->dev)) {
1117 struct cxl_switch_decoder *cxlsd;
1119 cxlsd = to_cxl_switch_decoder(&cxld->dev);
1120 if (cxl_rr->nr_targets > cxlsd->nr_targets) {
1121 dev_dbg(&cxlr->dev,
1122 "%s:%s %s add: %s:%s @ %d overflows targets: %d\n",
1123 dev_name(port->uport_dev), dev_name(&port->dev),
1124 dev_name(&cxld->dev), dev_name(&cxlmd->dev),
1125 dev_name(&cxled->cxld.dev), pos,
1126 cxlsd->nr_targets);
1127 rc = -ENXIO;
1128 goto out_erase;
1132 rc = cxl_rr_ep_add(cxl_rr, cxled);
1133 if (rc) {
1134 dev_dbg(&cxlr->dev,
1135 "%s: failed to track endpoint %s:%s reference\n",
1136 dev_name(&port->dev), dev_name(&cxlmd->dev),
1137 dev_name(&cxld->dev));
1138 goto out_erase;
1141 dev_dbg(&cxlr->dev,
1142 "%s:%s %s add: %s:%s @ %d next: %s nr_eps: %d nr_targets: %d\n",
1143 dev_name(port->uport_dev), dev_name(&port->dev),
1144 dev_name(&cxld->dev), dev_name(&cxlmd->dev),
1145 dev_name(&cxled->cxld.dev), pos,
1146 ep ? ep->next ? dev_name(ep->next->uport_dev) :
1147 dev_name(&cxlmd->dev) :
1148 "none",
1149 cxl_rr->nr_eps, cxl_rr->nr_targets);
1151 return 0;
1152 out_erase:
1153 if (nr_targets_inc)
1154 cxl_rr->nr_targets--;
1155 if (cxl_rr->nr_eps == 0)
1156 free_region_ref(cxl_rr);
1157 return rc;
1160 static void cxl_port_detach_region(struct cxl_port *port,
1161 struct cxl_region *cxlr,
1162 struct cxl_endpoint_decoder *cxled)
1164 struct cxl_region_ref *cxl_rr;
1165 struct cxl_ep *ep = NULL;
1167 lockdep_assert_held_write(&cxl_region_rwsem);
1169 cxl_rr = cxl_rr_load(port, cxlr);
1170 if (!cxl_rr)
1171 return;
1174 * Endpoint ports do not carry cxl_ep references, and they
1175 * never target more than one endpoint by definition
1177 if (cxl_rr->decoder == &cxled->cxld)
1178 cxl_rr->nr_eps--;
1179 else
1180 ep = xa_erase(&cxl_rr->endpoints, (unsigned long)cxled);
1181 if (ep) {
1182 struct cxl_ep *ep_iter;
1183 unsigned long index;
1184 int found = 0;
1186 cxl_rr->nr_eps--;
1187 xa_for_each(&cxl_rr->endpoints, index, ep_iter) {
1188 if (ep_iter->next == ep->next) {
1189 found++;
1190 break;
1193 if (!found)
1194 cxl_rr->nr_targets--;
1197 if (cxl_rr->nr_eps == 0)
1198 free_region_ref(cxl_rr);
1201 static int check_last_peer(struct cxl_endpoint_decoder *cxled,
1202 struct cxl_ep *ep, struct cxl_region_ref *cxl_rr,
1203 int distance)
1205 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
1206 struct cxl_region *cxlr = cxl_rr->region;
1207 struct cxl_region_params *p = &cxlr->params;
1208 struct cxl_endpoint_decoder *cxled_peer;
1209 struct cxl_port *port = cxl_rr->port;
1210 struct cxl_memdev *cxlmd_peer;
1211 struct cxl_ep *ep_peer;
1212 int pos = cxled->pos;
1215 * If this position wants to share a dport with the last endpoint mapped
1216 * then that endpoint, at index 'position - distance', must also be
1217 * mapped by this dport.
1219 if (pos < distance) {
1220 dev_dbg(&cxlr->dev, "%s:%s: cannot host %s:%s at %d\n",
1221 dev_name(port->uport_dev), dev_name(&port->dev),
1222 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev), pos);
1223 return -ENXIO;
1225 cxled_peer = p->targets[pos - distance];
1226 cxlmd_peer = cxled_to_memdev(cxled_peer);
1227 ep_peer = cxl_ep_load(port, cxlmd_peer);
1228 if (ep->dport != ep_peer->dport) {
1229 dev_dbg(&cxlr->dev,
1230 "%s:%s: %s:%s pos %d mismatched peer %s:%s\n",
1231 dev_name(port->uport_dev), dev_name(&port->dev),
1232 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev), pos,
1233 dev_name(&cxlmd_peer->dev),
1234 dev_name(&cxled_peer->cxld.dev));
1235 return -ENXIO;
1238 return 0;
1241 static int check_interleave_cap(struct cxl_decoder *cxld, int iw, int ig)
1243 struct cxl_port *port = to_cxl_port(cxld->dev.parent);
1244 struct cxl_hdm *cxlhdm = dev_get_drvdata(&port->dev);
1245 unsigned int interleave_mask;
1246 u8 eiw;
1247 u16 eig;
1248 int high_pos, low_pos;
1250 if (!test_bit(iw, &cxlhdm->iw_cap_mask))
1251 return -ENXIO;
1253 * Per CXL specification r3.1(8.2.4.20.13 Decoder Protection),
1254 * if eiw < 8:
1255 * DPAOFFSET[51: eig + 8] = HPAOFFSET[51: eig + 8 + eiw]
1256 * DPAOFFSET[eig + 7: 0] = HPAOFFSET[eig + 7: 0]
1258 * when the eiw is 0, all the bits of HPAOFFSET[51: 0] are used, the
1259 * interleave bits are none.
1261 * if eiw >= 8:
1262 * DPAOFFSET[51: eig + 8] = HPAOFFSET[51: eig + eiw] / 3
1263 * DPAOFFSET[eig + 7: 0] = HPAOFFSET[eig + 7: 0]
1265 * when the eiw is 8, all the bits of HPAOFFSET[51: 0] are used, the
1266 * interleave bits are none.
1268 ways_to_eiw(iw, &eiw);
1269 if (eiw == 0 || eiw == 8)
1270 return 0;
1272 granularity_to_eig(ig, &eig);
1273 if (eiw > 8)
1274 high_pos = eiw + eig - 1;
1275 else
1276 high_pos = eiw + eig + 7;
1277 low_pos = eig + 8;
1278 interleave_mask = GENMASK(high_pos, low_pos);
1279 if (interleave_mask & ~cxlhdm->interleave_mask)
1280 return -ENXIO;
1282 return 0;
1285 static int cxl_port_setup_targets(struct cxl_port *port,
1286 struct cxl_region *cxlr,
1287 struct cxl_endpoint_decoder *cxled)
1289 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(cxlr->dev.parent);
1290 int parent_iw, parent_ig, ig, iw, rc, inc = 0, pos = cxled->pos;
1291 struct cxl_port *parent_port = to_cxl_port(port->dev.parent);
1292 struct cxl_region_ref *cxl_rr = cxl_rr_load(port, cxlr);
1293 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
1294 struct cxl_ep *ep = cxl_ep_load(port, cxlmd);
1295 struct cxl_region_params *p = &cxlr->params;
1296 struct cxl_decoder *cxld = cxl_rr->decoder;
1297 struct cxl_switch_decoder *cxlsd;
1298 u16 eig, peig;
1299 u8 eiw, peiw;
1302 * While root level decoders support x3, x6, x12, switch level
1303 * decoders only support powers of 2 up to x16.
1305 if (!is_power_of_2(cxl_rr->nr_targets)) {
1306 dev_dbg(&cxlr->dev, "%s:%s: invalid target count %d\n",
1307 dev_name(port->uport_dev), dev_name(&port->dev),
1308 cxl_rr->nr_targets);
1309 return -EINVAL;
1312 cxlsd = to_cxl_switch_decoder(&cxld->dev);
1313 if (cxl_rr->nr_targets_set) {
1314 int i, distance;
1317 * Passthrough decoders impose no distance requirements between
1318 * peers
1320 if (cxl_rr->nr_targets == 1)
1321 distance = 0;
1322 else
1323 distance = p->nr_targets / cxl_rr->nr_targets;
1324 for (i = 0; i < cxl_rr->nr_targets_set; i++)
1325 if (ep->dport == cxlsd->target[i]) {
1326 rc = check_last_peer(cxled, ep, cxl_rr,
1327 distance);
1328 if (rc)
1329 return rc;
1330 goto out_target_set;
1332 goto add_target;
1335 if (is_cxl_root(parent_port)) {
1337 * Root decoder IG is always set to value in CFMWS which
1338 * may be different than this region's IG. We can use the
1339 * region's IG here since interleave_granularity_store()
1340 * does not allow interleaved host-bridges with
1341 * root IG != region IG.
1343 parent_ig = p->interleave_granularity;
1344 parent_iw = cxlrd->cxlsd.cxld.interleave_ways;
1346 * For purposes of address bit routing, use power-of-2 math for
1347 * switch ports.
1349 if (!is_power_of_2(parent_iw))
1350 parent_iw /= 3;
1351 } else {
1352 struct cxl_region_ref *parent_rr;
1353 struct cxl_decoder *parent_cxld;
1355 parent_rr = cxl_rr_load(parent_port, cxlr);
1356 parent_cxld = parent_rr->decoder;
1357 parent_ig = parent_cxld->interleave_granularity;
1358 parent_iw = parent_cxld->interleave_ways;
1361 rc = granularity_to_eig(parent_ig, &peig);
1362 if (rc) {
1363 dev_dbg(&cxlr->dev, "%s:%s: invalid parent granularity: %d\n",
1364 dev_name(parent_port->uport_dev),
1365 dev_name(&parent_port->dev), parent_ig);
1366 return rc;
1369 rc = ways_to_eiw(parent_iw, &peiw);
1370 if (rc) {
1371 dev_dbg(&cxlr->dev, "%s:%s: invalid parent interleave: %d\n",
1372 dev_name(parent_port->uport_dev),
1373 dev_name(&parent_port->dev), parent_iw);
1374 return rc;
1377 iw = cxl_rr->nr_targets;
1378 rc = ways_to_eiw(iw, &eiw);
1379 if (rc) {
1380 dev_dbg(&cxlr->dev, "%s:%s: invalid port interleave: %d\n",
1381 dev_name(port->uport_dev), dev_name(&port->dev), iw);
1382 return rc;
1386 * Interleave granularity is a multiple of @parent_port granularity.
1387 * Multiplier is the parent port interleave ways.
1389 rc = granularity_to_eig(parent_ig * parent_iw, &eig);
1390 if (rc) {
1391 dev_dbg(&cxlr->dev,
1392 "%s: invalid granularity calculation (%d * %d)\n",
1393 dev_name(&parent_port->dev), parent_ig, parent_iw);
1394 return rc;
1397 rc = eig_to_granularity(eig, &ig);
1398 if (rc) {
1399 dev_dbg(&cxlr->dev, "%s:%s: invalid interleave: %d\n",
1400 dev_name(port->uport_dev), dev_name(&port->dev),
1401 256 << eig);
1402 return rc;
1405 if (iw > 8 || iw > cxlsd->nr_targets) {
1406 dev_dbg(&cxlr->dev,
1407 "%s:%s:%s: ways: %d overflows targets: %d\n",
1408 dev_name(port->uport_dev), dev_name(&port->dev),
1409 dev_name(&cxld->dev), iw, cxlsd->nr_targets);
1410 return -ENXIO;
1413 if (test_bit(CXL_REGION_F_AUTO, &cxlr->flags)) {
1414 if (cxld->interleave_ways != iw ||
1415 cxld->interleave_granularity != ig ||
1416 cxld->hpa_range.start != p->res->start ||
1417 cxld->hpa_range.end != p->res->end ||
1418 ((cxld->flags & CXL_DECODER_F_ENABLE) == 0)) {
1419 dev_err(&cxlr->dev,
1420 "%s:%s %s expected iw: %d ig: %d %pr\n",
1421 dev_name(port->uport_dev), dev_name(&port->dev),
1422 __func__, iw, ig, p->res);
1423 dev_err(&cxlr->dev,
1424 "%s:%s %s got iw: %d ig: %d state: %s %#llx:%#llx\n",
1425 dev_name(port->uport_dev), dev_name(&port->dev),
1426 __func__, cxld->interleave_ways,
1427 cxld->interleave_granularity,
1428 (cxld->flags & CXL_DECODER_F_ENABLE) ?
1429 "enabled" :
1430 "disabled",
1431 cxld->hpa_range.start, cxld->hpa_range.end);
1432 return -ENXIO;
1434 } else {
1435 rc = check_interleave_cap(cxld, iw, ig);
1436 if (rc) {
1437 dev_dbg(&cxlr->dev,
1438 "%s:%s iw: %d ig: %d is not supported\n",
1439 dev_name(port->uport_dev),
1440 dev_name(&port->dev), iw, ig);
1441 return rc;
1444 cxld->interleave_ways = iw;
1445 cxld->interleave_granularity = ig;
1446 cxld->hpa_range = (struct range) {
1447 .start = p->res->start,
1448 .end = p->res->end,
1451 dev_dbg(&cxlr->dev, "%s:%s iw: %d ig: %d\n", dev_name(port->uport_dev),
1452 dev_name(&port->dev), iw, ig);
1453 add_target:
1454 if (cxl_rr->nr_targets_set == cxl_rr->nr_targets) {
1455 dev_dbg(&cxlr->dev,
1456 "%s:%s: targets full trying to add %s:%s at %d\n",
1457 dev_name(port->uport_dev), dev_name(&port->dev),
1458 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev), pos);
1459 return -ENXIO;
1461 if (test_bit(CXL_REGION_F_AUTO, &cxlr->flags)) {
1462 if (cxlsd->target[cxl_rr->nr_targets_set] != ep->dport) {
1463 dev_dbg(&cxlr->dev, "%s:%s: %s expected %s at %d\n",
1464 dev_name(port->uport_dev), dev_name(&port->dev),
1465 dev_name(&cxlsd->cxld.dev),
1466 dev_name(ep->dport->dport_dev),
1467 cxl_rr->nr_targets_set);
1468 return -ENXIO;
1470 } else
1471 cxlsd->target[cxl_rr->nr_targets_set] = ep->dport;
1472 inc = 1;
1473 out_target_set:
1474 cxl_rr->nr_targets_set += inc;
1475 dev_dbg(&cxlr->dev, "%s:%s target[%d] = %s for %s:%s @ %d\n",
1476 dev_name(port->uport_dev), dev_name(&port->dev),
1477 cxl_rr->nr_targets_set - 1, dev_name(ep->dport->dport_dev),
1478 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev), pos);
1480 return 0;
1483 static void cxl_port_reset_targets(struct cxl_port *port,
1484 struct cxl_region *cxlr)
1486 struct cxl_region_ref *cxl_rr = cxl_rr_load(port, cxlr);
1487 struct cxl_decoder *cxld;
1490 * After the last endpoint has been detached the entire cxl_rr may now
1491 * be gone.
1493 if (!cxl_rr)
1494 return;
1495 cxl_rr->nr_targets_set = 0;
1497 cxld = cxl_rr->decoder;
1498 cxld->hpa_range = (struct range) {
1499 .start = 0,
1500 .end = -1,
1504 static void cxl_region_teardown_targets(struct cxl_region *cxlr)
1506 struct cxl_region_params *p = &cxlr->params;
1507 struct cxl_endpoint_decoder *cxled;
1508 struct cxl_dev_state *cxlds;
1509 struct cxl_memdev *cxlmd;
1510 struct cxl_port *iter;
1511 struct cxl_ep *ep;
1512 int i;
1515 * In the auto-discovery case skip automatic teardown since the
1516 * address space is already active
1518 if (test_bit(CXL_REGION_F_AUTO, &cxlr->flags))
1519 return;
1521 for (i = 0; i < p->nr_targets; i++) {
1522 cxled = p->targets[i];
1523 cxlmd = cxled_to_memdev(cxled);
1524 cxlds = cxlmd->cxlds;
1526 if (cxlds->rcd)
1527 continue;
1529 iter = cxled_to_port(cxled);
1530 while (!is_cxl_root(to_cxl_port(iter->dev.parent)))
1531 iter = to_cxl_port(iter->dev.parent);
1533 for (ep = cxl_ep_load(iter, cxlmd); iter;
1534 iter = ep->next, ep = cxl_ep_load(iter, cxlmd))
1535 cxl_port_reset_targets(iter, cxlr);
1539 static int cxl_region_setup_targets(struct cxl_region *cxlr)
1541 struct cxl_region_params *p = &cxlr->params;
1542 struct cxl_endpoint_decoder *cxled;
1543 struct cxl_dev_state *cxlds;
1544 int i, rc, rch = 0, vh = 0;
1545 struct cxl_memdev *cxlmd;
1546 struct cxl_port *iter;
1547 struct cxl_ep *ep;
1549 for (i = 0; i < p->nr_targets; i++) {
1550 cxled = p->targets[i];
1551 cxlmd = cxled_to_memdev(cxled);
1552 cxlds = cxlmd->cxlds;
1554 /* validate that all targets agree on topology */
1555 if (!cxlds->rcd) {
1556 vh++;
1557 } else {
1558 rch++;
1559 continue;
1562 iter = cxled_to_port(cxled);
1563 while (!is_cxl_root(to_cxl_port(iter->dev.parent)))
1564 iter = to_cxl_port(iter->dev.parent);
1567 * Descend the topology tree programming / validating
1568 * targets while looking for conflicts.
1570 for (ep = cxl_ep_load(iter, cxlmd); iter;
1571 iter = ep->next, ep = cxl_ep_load(iter, cxlmd)) {
1572 rc = cxl_port_setup_targets(iter, cxlr, cxled);
1573 if (rc) {
1574 cxl_region_teardown_targets(cxlr);
1575 return rc;
1580 if (rch && vh) {
1581 dev_err(&cxlr->dev, "mismatched CXL topologies detected\n");
1582 cxl_region_teardown_targets(cxlr);
1583 return -ENXIO;
1586 return 0;
1589 static int cxl_region_validate_position(struct cxl_region *cxlr,
1590 struct cxl_endpoint_decoder *cxled,
1591 int pos)
1593 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
1594 struct cxl_region_params *p = &cxlr->params;
1595 int i;
1597 if (pos < 0 || pos >= p->interleave_ways) {
1598 dev_dbg(&cxlr->dev, "position %d out of range %d\n", pos,
1599 p->interleave_ways);
1600 return -ENXIO;
1603 if (p->targets[pos] == cxled)
1604 return 0;
1606 if (p->targets[pos]) {
1607 struct cxl_endpoint_decoder *cxled_target = p->targets[pos];
1608 struct cxl_memdev *cxlmd_target = cxled_to_memdev(cxled_target);
1610 dev_dbg(&cxlr->dev, "position %d already assigned to %s:%s\n",
1611 pos, dev_name(&cxlmd_target->dev),
1612 dev_name(&cxled_target->cxld.dev));
1613 return -EBUSY;
1616 for (i = 0; i < p->interleave_ways; i++) {
1617 struct cxl_endpoint_decoder *cxled_target;
1618 struct cxl_memdev *cxlmd_target;
1620 cxled_target = p->targets[i];
1621 if (!cxled_target)
1622 continue;
1624 cxlmd_target = cxled_to_memdev(cxled_target);
1625 if (cxlmd_target == cxlmd) {
1626 dev_dbg(&cxlr->dev,
1627 "%s already specified at position %d via: %s\n",
1628 dev_name(&cxlmd->dev), pos,
1629 dev_name(&cxled_target->cxld.dev));
1630 return -EBUSY;
1634 return 0;
1637 static int cxl_region_attach_position(struct cxl_region *cxlr,
1638 struct cxl_root_decoder *cxlrd,
1639 struct cxl_endpoint_decoder *cxled,
1640 const struct cxl_dport *dport, int pos)
1642 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
1643 struct cxl_switch_decoder *cxlsd = &cxlrd->cxlsd;
1644 struct cxl_decoder *cxld = &cxlsd->cxld;
1645 int iw = cxld->interleave_ways;
1646 struct cxl_port *iter;
1647 int rc;
1649 if (dport != cxlrd->cxlsd.target[pos % iw]) {
1650 dev_dbg(&cxlr->dev, "%s:%s invalid target position for %s\n",
1651 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
1652 dev_name(&cxlrd->cxlsd.cxld.dev));
1653 return -ENXIO;
1656 for (iter = cxled_to_port(cxled); !is_cxl_root(iter);
1657 iter = to_cxl_port(iter->dev.parent)) {
1658 rc = cxl_port_attach_region(iter, cxlr, cxled, pos);
1659 if (rc)
1660 goto err;
1663 return 0;
1665 err:
1666 for (iter = cxled_to_port(cxled); !is_cxl_root(iter);
1667 iter = to_cxl_port(iter->dev.parent))
1668 cxl_port_detach_region(iter, cxlr, cxled);
1669 return rc;
1672 static int cxl_region_attach_auto(struct cxl_region *cxlr,
1673 struct cxl_endpoint_decoder *cxled, int pos)
1675 struct cxl_region_params *p = &cxlr->params;
1677 if (cxled->state != CXL_DECODER_STATE_AUTO) {
1678 dev_err(&cxlr->dev,
1679 "%s: unable to add decoder to autodetected region\n",
1680 dev_name(&cxled->cxld.dev));
1681 return -EINVAL;
1684 if (pos >= 0) {
1685 dev_dbg(&cxlr->dev, "%s: expected auto position, not %d\n",
1686 dev_name(&cxled->cxld.dev), pos);
1687 return -EINVAL;
1690 if (p->nr_targets >= p->interleave_ways) {
1691 dev_err(&cxlr->dev, "%s: no more target slots available\n",
1692 dev_name(&cxled->cxld.dev));
1693 return -ENXIO;
1697 * Temporarily record the endpoint decoder into the target array. Yes,
1698 * this means that userspace can view devices in the wrong position
1699 * before the region activates, and must be careful to understand when
1700 * it might be racing region autodiscovery.
1702 pos = p->nr_targets;
1703 p->targets[pos] = cxled;
1704 cxled->pos = pos;
1705 p->nr_targets++;
1707 return 0;
1710 static int cmp_interleave_pos(const void *a, const void *b)
1712 struct cxl_endpoint_decoder *cxled_a = *(typeof(cxled_a) *)a;
1713 struct cxl_endpoint_decoder *cxled_b = *(typeof(cxled_b) *)b;
1715 return cxled_a->pos - cxled_b->pos;
1718 static struct cxl_port *next_port(struct cxl_port *port)
1720 if (!port->parent_dport)
1721 return NULL;
1722 return port->parent_dport->port;
1725 static int match_switch_decoder_by_range(struct device *dev, void *data)
1727 struct cxl_switch_decoder *cxlsd;
1728 struct range *r1, *r2 = data;
1730 if (!is_switch_decoder(dev))
1731 return 0;
1733 cxlsd = to_cxl_switch_decoder(dev);
1734 r1 = &cxlsd->cxld.hpa_range;
1736 if (is_root_decoder(dev))
1737 return range_contains(r1, r2);
1738 return (r1->start == r2->start && r1->end == r2->end);
1741 static int find_pos_and_ways(struct cxl_port *port, struct range *range,
1742 int *pos, int *ways)
1744 struct cxl_switch_decoder *cxlsd;
1745 struct cxl_port *parent;
1746 struct device *dev;
1747 int rc = -ENXIO;
1749 parent = next_port(port);
1750 if (!parent)
1751 return rc;
1753 dev = device_find_child(&parent->dev, range,
1754 match_switch_decoder_by_range);
1755 if (!dev) {
1756 dev_err(port->uport_dev,
1757 "failed to find decoder mapping %#llx-%#llx\n",
1758 range->start, range->end);
1759 return rc;
1761 cxlsd = to_cxl_switch_decoder(dev);
1762 *ways = cxlsd->cxld.interleave_ways;
1764 for (int i = 0; i < *ways; i++) {
1765 if (cxlsd->target[i] == port->parent_dport) {
1766 *pos = i;
1767 rc = 0;
1768 break;
1771 put_device(dev);
1773 return rc;
1777 * cxl_calc_interleave_pos() - calculate an endpoint position in a region
1778 * @cxled: endpoint decoder member of given region
1780 * The endpoint position is calculated by traversing the topology from
1781 * the endpoint to the root decoder and iteratively applying this
1782 * calculation:
1784 * position = position * parent_ways + parent_pos;
1786 * ...where @position is inferred from switch and root decoder target lists.
1788 * Return: position >= 0 on success
1789 * -ENXIO on failure
1791 static int cxl_calc_interleave_pos(struct cxl_endpoint_decoder *cxled)
1793 struct cxl_port *iter, *port = cxled_to_port(cxled);
1794 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
1795 struct range *range = &cxled->cxld.hpa_range;
1796 int parent_ways = 0, parent_pos = 0, pos = 0;
1797 int rc;
1800 * Example: the expected interleave order of the 4-way region shown
1801 * below is: mem0, mem2, mem1, mem3
1803 * root_port
1804 * / \
1805 * host_bridge_0 host_bridge_1
1806 * | | | |
1807 * mem0 mem1 mem2 mem3
1809 * In the example the calculator will iterate twice. The first iteration
1810 * uses the mem position in the host-bridge and the ways of the host-
1811 * bridge to generate the first, or local, position. The second
1812 * iteration uses the host-bridge position in the root_port and the ways
1813 * of the root_port to refine the position.
1815 * A trace of the calculation per endpoint looks like this:
1816 * mem0: pos = 0 * 2 + 0 mem2: pos = 0 * 2 + 0
1817 * pos = 0 * 2 + 0 pos = 0 * 2 + 1
1818 * pos: 0 pos: 1
1820 * mem1: pos = 0 * 2 + 1 mem3: pos = 0 * 2 + 1
1821 * pos = 1 * 2 + 0 pos = 1 * 2 + 1
1822 * pos: 2 pos = 3
1824 * Note that while this example is simple, the method applies to more
1825 * complex topologies, including those with switches.
1828 /* Iterate from endpoint to root_port refining the position */
1829 for (iter = port; iter; iter = next_port(iter)) {
1830 if (is_cxl_root(iter))
1831 break;
1833 rc = find_pos_and_ways(iter, range, &parent_pos, &parent_ways);
1834 if (rc)
1835 return rc;
1837 pos = pos * parent_ways + parent_pos;
1840 dev_dbg(&cxlmd->dev,
1841 "decoder:%s parent:%s port:%s range:%#llx-%#llx pos:%d\n",
1842 dev_name(&cxled->cxld.dev), dev_name(cxlmd->dev.parent),
1843 dev_name(&port->dev), range->start, range->end, pos);
1845 return pos;
1848 static int cxl_region_sort_targets(struct cxl_region *cxlr)
1850 struct cxl_region_params *p = &cxlr->params;
1851 int i, rc = 0;
1853 for (i = 0; i < p->nr_targets; i++) {
1854 struct cxl_endpoint_decoder *cxled = p->targets[i];
1856 cxled->pos = cxl_calc_interleave_pos(cxled);
1858 * Record that sorting failed, but still continue to calc
1859 * cxled->pos so that follow-on code paths can reliably
1860 * do p->targets[cxled->pos] to self-reference their entry.
1862 if (cxled->pos < 0)
1863 rc = -ENXIO;
1865 /* Keep the cxlr target list in interleave position order */
1866 sort(p->targets, p->nr_targets, sizeof(p->targets[0]),
1867 cmp_interleave_pos, NULL);
1869 dev_dbg(&cxlr->dev, "region sort %s\n", rc ? "failed" : "successful");
1870 return rc;
1873 static int cxl_region_attach(struct cxl_region *cxlr,
1874 struct cxl_endpoint_decoder *cxled, int pos)
1876 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(cxlr->dev.parent);
1877 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
1878 struct cxl_region_params *p = &cxlr->params;
1879 struct cxl_port *ep_port, *root_port;
1880 struct cxl_dport *dport;
1881 int rc = -ENXIO;
1883 rc = check_interleave_cap(&cxled->cxld, p->interleave_ways,
1884 p->interleave_granularity);
1885 if (rc) {
1886 dev_dbg(&cxlr->dev, "%s iw: %d ig: %d is not supported\n",
1887 dev_name(&cxled->cxld.dev), p->interleave_ways,
1888 p->interleave_granularity);
1889 return rc;
1892 if (cxled->mode != cxlr->mode) {
1893 dev_dbg(&cxlr->dev, "%s region mode: %d mismatch: %d\n",
1894 dev_name(&cxled->cxld.dev), cxlr->mode, cxled->mode);
1895 return -EINVAL;
1898 if (cxled->mode == CXL_DECODER_DEAD) {
1899 dev_dbg(&cxlr->dev, "%s dead\n", dev_name(&cxled->cxld.dev));
1900 return -ENODEV;
1903 /* all full of members, or interleave config not established? */
1904 if (p->state > CXL_CONFIG_INTERLEAVE_ACTIVE) {
1905 dev_dbg(&cxlr->dev, "region already active\n");
1906 return -EBUSY;
1907 } else if (p->state < CXL_CONFIG_INTERLEAVE_ACTIVE) {
1908 dev_dbg(&cxlr->dev, "interleave config missing\n");
1909 return -ENXIO;
1912 if (p->nr_targets >= p->interleave_ways) {
1913 dev_dbg(&cxlr->dev, "region already has %d endpoints\n",
1914 p->nr_targets);
1915 return -EINVAL;
1918 ep_port = cxled_to_port(cxled);
1919 root_port = cxlrd_to_port(cxlrd);
1920 dport = cxl_find_dport_by_dev(root_port, ep_port->host_bridge);
1921 if (!dport) {
1922 dev_dbg(&cxlr->dev, "%s:%s invalid target for %s\n",
1923 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
1924 dev_name(cxlr->dev.parent));
1925 return -ENXIO;
1928 if (cxled->cxld.target_type != cxlr->type) {
1929 dev_dbg(&cxlr->dev, "%s:%s type mismatch: %d vs %d\n",
1930 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
1931 cxled->cxld.target_type, cxlr->type);
1932 return -ENXIO;
1935 if (!cxled->dpa_res) {
1936 dev_dbg(&cxlr->dev, "%s:%s: missing DPA allocation.\n",
1937 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev));
1938 return -ENXIO;
1941 if (resource_size(cxled->dpa_res) * p->interleave_ways !=
1942 resource_size(p->res)) {
1943 dev_dbg(&cxlr->dev,
1944 "%s:%s: decoder-size-%#llx * ways-%d != region-size-%#llx\n",
1945 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
1946 (u64)resource_size(cxled->dpa_res), p->interleave_ways,
1947 (u64)resource_size(p->res));
1948 return -EINVAL;
1951 cxl_region_perf_data_calculate(cxlr, cxled);
1953 if (test_bit(CXL_REGION_F_AUTO, &cxlr->flags)) {
1954 int i;
1956 rc = cxl_region_attach_auto(cxlr, cxled, pos);
1957 if (rc)
1958 return rc;
1960 /* await more targets to arrive... */
1961 if (p->nr_targets < p->interleave_ways)
1962 return 0;
1965 * All targets are here, which implies all PCI enumeration that
1966 * affects this region has been completed. Walk the topology to
1967 * sort the devices into their relative region decode position.
1969 rc = cxl_region_sort_targets(cxlr);
1970 if (rc)
1971 return rc;
1973 for (i = 0; i < p->nr_targets; i++) {
1974 cxled = p->targets[i];
1975 ep_port = cxled_to_port(cxled);
1976 dport = cxl_find_dport_by_dev(root_port,
1977 ep_port->host_bridge);
1978 rc = cxl_region_attach_position(cxlr, cxlrd, cxled,
1979 dport, i);
1980 if (rc)
1981 return rc;
1984 rc = cxl_region_setup_targets(cxlr);
1985 if (rc)
1986 return rc;
1989 * If target setup succeeds in the autodiscovery case
1990 * then the region is already committed.
1992 p->state = CXL_CONFIG_COMMIT;
1993 cxl_region_shared_upstream_bandwidth_update(cxlr);
1995 return 0;
1998 rc = cxl_region_validate_position(cxlr, cxled, pos);
1999 if (rc)
2000 return rc;
2002 rc = cxl_region_attach_position(cxlr, cxlrd, cxled, dport, pos);
2003 if (rc)
2004 return rc;
2006 p->targets[pos] = cxled;
2007 cxled->pos = pos;
2008 p->nr_targets++;
2010 if (p->nr_targets == p->interleave_ways) {
2011 rc = cxl_region_setup_targets(cxlr);
2012 if (rc)
2013 return rc;
2014 p->state = CXL_CONFIG_ACTIVE;
2015 cxl_region_shared_upstream_bandwidth_update(cxlr);
2018 cxled->cxld.interleave_ways = p->interleave_ways;
2019 cxled->cxld.interleave_granularity = p->interleave_granularity;
2020 cxled->cxld.hpa_range = (struct range) {
2021 .start = p->res->start,
2022 .end = p->res->end,
2025 if (p->nr_targets != p->interleave_ways)
2026 return 0;
2029 * Test the auto-discovery position calculator function
2030 * against this successfully created user-defined region.
2031 * A fail message here means that this interleave config
2032 * will fail when presented as CXL_REGION_F_AUTO.
2034 for (int i = 0; i < p->nr_targets; i++) {
2035 struct cxl_endpoint_decoder *cxled = p->targets[i];
2036 int test_pos;
2038 test_pos = cxl_calc_interleave_pos(cxled);
2039 dev_dbg(&cxled->cxld.dev,
2040 "Test cxl_calc_interleave_pos(): %s test_pos:%d cxled->pos:%d\n",
2041 (test_pos == cxled->pos) ? "success" : "fail",
2042 test_pos, cxled->pos);
2045 return 0;
2048 static int cxl_region_detach(struct cxl_endpoint_decoder *cxled)
2050 struct cxl_port *iter, *ep_port = cxled_to_port(cxled);
2051 struct cxl_region *cxlr = cxled->cxld.region;
2052 struct cxl_region_params *p;
2053 int rc = 0;
2055 lockdep_assert_held_write(&cxl_region_rwsem);
2057 if (!cxlr)
2058 return 0;
2060 p = &cxlr->params;
2061 get_device(&cxlr->dev);
2063 if (p->state > CXL_CONFIG_ACTIVE) {
2064 cxl_region_decode_reset(cxlr, p->interleave_ways);
2065 p->state = CXL_CONFIG_ACTIVE;
2068 for (iter = ep_port; !is_cxl_root(iter);
2069 iter = to_cxl_port(iter->dev.parent))
2070 cxl_port_detach_region(iter, cxlr, cxled);
2072 if (cxled->pos < 0 || cxled->pos >= p->interleave_ways ||
2073 p->targets[cxled->pos] != cxled) {
2074 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
2076 dev_WARN_ONCE(&cxlr->dev, 1, "expected %s:%s at position %d\n",
2077 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
2078 cxled->pos);
2079 goto out;
2082 if (p->state == CXL_CONFIG_ACTIVE) {
2083 p->state = CXL_CONFIG_INTERLEAVE_ACTIVE;
2084 cxl_region_teardown_targets(cxlr);
2086 p->targets[cxled->pos] = NULL;
2087 p->nr_targets--;
2088 cxled->cxld.hpa_range = (struct range) {
2089 .start = 0,
2090 .end = -1,
2093 /* notify the region driver that one of its targets has departed */
2094 up_write(&cxl_region_rwsem);
2095 device_release_driver(&cxlr->dev);
2096 down_write(&cxl_region_rwsem);
2097 out:
2098 put_device(&cxlr->dev);
2099 return rc;
2102 void cxl_decoder_kill_region(struct cxl_endpoint_decoder *cxled)
2104 down_write(&cxl_region_rwsem);
2105 cxled->mode = CXL_DECODER_DEAD;
2106 cxl_region_detach(cxled);
2107 up_write(&cxl_region_rwsem);
2110 static int attach_target(struct cxl_region *cxlr,
2111 struct cxl_endpoint_decoder *cxled, int pos,
2112 unsigned int state)
2114 int rc = 0;
2116 if (state == TASK_INTERRUPTIBLE)
2117 rc = down_write_killable(&cxl_region_rwsem);
2118 else
2119 down_write(&cxl_region_rwsem);
2120 if (rc)
2121 return rc;
2123 down_read(&cxl_dpa_rwsem);
2124 rc = cxl_region_attach(cxlr, cxled, pos);
2125 up_read(&cxl_dpa_rwsem);
2126 up_write(&cxl_region_rwsem);
2127 return rc;
2130 static int detach_target(struct cxl_region *cxlr, int pos)
2132 struct cxl_region_params *p = &cxlr->params;
2133 int rc;
2135 rc = down_write_killable(&cxl_region_rwsem);
2136 if (rc)
2137 return rc;
2139 if (pos >= p->interleave_ways) {
2140 dev_dbg(&cxlr->dev, "position %d out of range %d\n", pos,
2141 p->interleave_ways);
2142 rc = -ENXIO;
2143 goto out;
2146 if (!p->targets[pos]) {
2147 rc = 0;
2148 goto out;
2151 rc = cxl_region_detach(p->targets[pos]);
2152 out:
2153 up_write(&cxl_region_rwsem);
2154 return rc;
2157 static size_t store_targetN(struct cxl_region *cxlr, const char *buf, int pos,
2158 size_t len)
2160 int rc;
2162 if (sysfs_streq(buf, "\n"))
2163 rc = detach_target(cxlr, pos);
2164 else {
2165 struct device *dev;
2167 dev = bus_find_device_by_name(&cxl_bus_type, NULL, buf);
2168 if (!dev)
2169 return -ENODEV;
2171 if (!is_endpoint_decoder(dev)) {
2172 rc = -EINVAL;
2173 goto out;
2176 rc = attach_target(cxlr, to_cxl_endpoint_decoder(dev), pos,
2177 TASK_INTERRUPTIBLE);
2178 out:
2179 put_device(dev);
2182 if (rc < 0)
2183 return rc;
2184 return len;
2187 #define TARGET_ATTR_RW(n) \
2188 static ssize_t target##n##_show( \
2189 struct device *dev, struct device_attribute *attr, char *buf) \
2191 return show_targetN(to_cxl_region(dev), buf, (n)); \
2193 static ssize_t target##n##_store(struct device *dev, \
2194 struct device_attribute *attr, \
2195 const char *buf, size_t len) \
2197 return store_targetN(to_cxl_region(dev), buf, (n), len); \
2199 static DEVICE_ATTR_RW(target##n)
2201 TARGET_ATTR_RW(0);
2202 TARGET_ATTR_RW(1);
2203 TARGET_ATTR_RW(2);
2204 TARGET_ATTR_RW(3);
2205 TARGET_ATTR_RW(4);
2206 TARGET_ATTR_RW(5);
2207 TARGET_ATTR_RW(6);
2208 TARGET_ATTR_RW(7);
2209 TARGET_ATTR_RW(8);
2210 TARGET_ATTR_RW(9);
2211 TARGET_ATTR_RW(10);
2212 TARGET_ATTR_RW(11);
2213 TARGET_ATTR_RW(12);
2214 TARGET_ATTR_RW(13);
2215 TARGET_ATTR_RW(14);
2216 TARGET_ATTR_RW(15);
2218 static struct attribute *target_attrs[] = {
2219 &dev_attr_target0.attr,
2220 &dev_attr_target1.attr,
2221 &dev_attr_target2.attr,
2222 &dev_attr_target3.attr,
2223 &dev_attr_target4.attr,
2224 &dev_attr_target5.attr,
2225 &dev_attr_target6.attr,
2226 &dev_attr_target7.attr,
2227 &dev_attr_target8.attr,
2228 &dev_attr_target9.attr,
2229 &dev_attr_target10.attr,
2230 &dev_attr_target11.attr,
2231 &dev_attr_target12.attr,
2232 &dev_attr_target13.attr,
2233 &dev_attr_target14.attr,
2234 &dev_attr_target15.attr,
2235 NULL,
2238 static umode_t cxl_region_target_visible(struct kobject *kobj,
2239 struct attribute *a, int n)
2241 struct device *dev = kobj_to_dev(kobj);
2242 struct cxl_region *cxlr = to_cxl_region(dev);
2243 struct cxl_region_params *p = &cxlr->params;
2245 if (n < p->interleave_ways)
2246 return a->mode;
2247 return 0;
2250 static const struct attribute_group cxl_region_target_group = {
2251 .attrs = target_attrs,
2252 .is_visible = cxl_region_target_visible,
2255 static const struct attribute_group *get_cxl_region_target_group(void)
2257 return &cxl_region_target_group;
2260 static const struct attribute_group *region_groups[] = {
2261 &cxl_base_attribute_group,
2262 &cxl_region_group,
2263 &cxl_region_target_group,
2264 &cxl_region_access0_coordinate_group,
2265 &cxl_region_access1_coordinate_group,
2266 NULL,
2269 static void cxl_region_release(struct device *dev)
2271 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(dev->parent);
2272 struct cxl_region *cxlr = to_cxl_region(dev);
2273 int id = atomic_read(&cxlrd->region_id);
2276 * Try to reuse the recently idled id rather than the cached
2277 * next id to prevent the region id space from increasing
2278 * unnecessarily.
2280 if (cxlr->id < id)
2281 if (atomic_try_cmpxchg(&cxlrd->region_id, &id, cxlr->id)) {
2282 memregion_free(id);
2283 goto out;
2286 memregion_free(cxlr->id);
2287 out:
2288 put_device(dev->parent);
2289 kfree(cxlr);
2292 const struct device_type cxl_region_type = {
2293 .name = "cxl_region",
2294 .release = cxl_region_release,
2295 .groups = region_groups
2298 bool is_cxl_region(struct device *dev)
2300 return dev->type == &cxl_region_type;
2302 EXPORT_SYMBOL_NS_GPL(is_cxl_region, CXL);
2304 static struct cxl_region *to_cxl_region(struct device *dev)
2306 if (dev_WARN_ONCE(dev, dev->type != &cxl_region_type,
2307 "not a cxl_region device\n"))
2308 return NULL;
2310 return container_of(dev, struct cxl_region, dev);
2313 static void unregister_region(void *_cxlr)
2315 struct cxl_region *cxlr = _cxlr;
2316 struct cxl_region_params *p = &cxlr->params;
2317 int i;
2319 device_del(&cxlr->dev);
2322 * Now that region sysfs is shutdown, the parameter block is now
2323 * read-only, so no need to hold the region rwsem to access the
2324 * region parameters.
2326 for (i = 0; i < p->interleave_ways; i++)
2327 detach_target(cxlr, i);
2329 cxl_region_iomem_release(cxlr);
2330 put_device(&cxlr->dev);
2333 static struct lock_class_key cxl_region_key;
2335 static struct cxl_region *cxl_region_alloc(struct cxl_root_decoder *cxlrd, int id)
2337 struct cxl_region *cxlr;
2338 struct device *dev;
2340 cxlr = kzalloc(sizeof(*cxlr), GFP_KERNEL);
2341 if (!cxlr) {
2342 memregion_free(id);
2343 return ERR_PTR(-ENOMEM);
2346 dev = &cxlr->dev;
2347 device_initialize(dev);
2348 lockdep_set_class(&dev->mutex, &cxl_region_key);
2349 dev->parent = &cxlrd->cxlsd.cxld.dev;
2351 * Keep root decoder pinned through cxl_region_release to fixup
2352 * region id allocations
2354 get_device(dev->parent);
2355 device_set_pm_not_required(dev);
2356 dev->bus = &cxl_bus_type;
2357 dev->type = &cxl_region_type;
2358 cxlr->id = id;
2360 return cxlr;
2363 static bool cxl_region_update_coordinates(struct cxl_region *cxlr, int nid)
2365 int cset = 0;
2366 int rc;
2368 for (int i = 0; i < ACCESS_COORDINATE_MAX; i++) {
2369 if (cxlr->coord[i].read_bandwidth) {
2370 rc = 0;
2371 if (cxl_need_node_perf_attrs_update(nid))
2372 node_set_perf_attrs(nid, &cxlr->coord[i], i);
2373 else
2374 rc = cxl_update_hmat_access_coordinates(nid, cxlr, i);
2376 if (rc == 0)
2377 cset++;
2381 if (!cset)
2382 return false;
2384 rc = sysfs_update_group(&cxlr->dev.kobj, get_cxl_region_access0_group());
2385 if (rc)
2386 dev_dbg(&cxlr->dev, "Failed to update access0 group\n");
2388 rc = sysfs_update_group(&cxlr->dev.kobj, get_cxl_region_access1_group());
2389 if (rc)
2390 dev_dbg(&cxlr->dev, "Failed to update access1 group\n");
2392 return true;
2395 static int cxl_region_perf_attrs_callback(struct notifier_block *nb,
2396 unsigned long action, void *arg)
2398 struct cxl_region *cxlr = container_of(nb, struct cxl_region,
2399 memory_notifier);
2400 struct memory_notify *mnb = arg;
2401 int nid = mnb->status_change_nid;
2402 int region_nid;
2404 if (nid == NUMA_NO_NODE || action != MEM_ONLINE)
2405 return NOTIFY_DONE;
2408 * No need to hold cxl_region_rwsem; region parameters are stable
2409 * within the cxl_region driver.
2411 region_nid = phys_to_target_node(cxlr->params.res->start);
2412 if (nid != region_nid)
2413 return NOTIFY_DONE;
2415 if (!cxl_region_update_coordinates(cxlr, nid))
2416 return NOTIFY_DONE;
2418 return NOTIFY_OK;
2421 static int cxl_region_calculate_adistance(struct notifier_block *nb,
2422 unsigned long nid, void *data)
2424 struct cxl_region *cxlr = container_of(nb, struct cxl_region,
2425 adist_notifier);
2426 struct access_coordinate *perf;
2427 int *adist = data;
2428 int region_nid;
2431 * No need to hold cxl_region_rwsem; region parameters are stable
2432 * within the cxl_region driver.
2434 region_nid = phys_to_target_node(cxlr->params.res->start);
2435 if (nid != region_nid)
2436 return NOTIFY_OK;
2438 perf = &cxlr->coord[ACCESS_COORDINATE_CPU];
2440 if (mt_perf_to_adistance(perf, adist))
2441 return NOTIFY_OK;
2443 return NOTIFY_STOP;
2447 * devm_cxl_add_region - Adds a region to a decoder
2448 * @cxlrd: root decoder
2449 * @id: memregion id to create, or memregion_free() on failure
2450 * @mode: mode for the endpoint decoders of this region
2451 * @type: select whether this is an expander or accelerator (type-2 or type-3)
2453 * This is the second step of region initialization. Regions exist within an
2454 * address space which is mapped by a @cxlrd.
2456 * Return: 0 if the region was added to the @cxlrd, else returns negative error
2457 * code. The region will be named "regionZ" where Z is the unique region number.
2459 static struct cxl_region *devm_cxl_add_region(struct cxl_root_decoder *cxlrd,
2460 int id,
2461 enum cxl_decoder_mode mode,
2462 enum cxl_decoder_type type)
2464 struct cxl_port *port = to_cxl_port(cxlrd->cxlsd.cxld.dev.parent);
2465 struct cxl_region *cxlr;
2466 struct device *dev;
2467 int rc;
2469 cxlr = cxl_region_alloc(cxlrd, id);
2470 if (IS_ERR(cxlr))
2471 return cxlr;
2472 cxlr->mode = mode;
2473 cxlr->type = type;
2475 dev = &cxlr->dev;
2476 rc = dev_set_name(dev, "region%d", id);
2477 if (rc)
2478 goto err;
2480 rc = device_add(dev);
2481 if (rc)
2482 goto err;
2484 rc = devm_add_action_or_reset(port->uport_dev, unregister_region, cxlr);
2485 if (rc)
2486 return ERR_PTR(rc);
2488 dev_dbg(port->uport_dev, "%s: created %s\n",
2489 dev_name(&cxlrd->cxlsd.cxld.dev), dev_name(dev));
2490 return cxlr;
2492 err:
2493 put_device(dev);
2494 return ERR_PTR(rc);
2497 static ssize_t __create_region_show(struct cxl_root_decoder *cxlrd, char *buf)
2499 return sysfs_emit(buf, "region%u\n", atomic_read(&cxlrd->region_id));
2502 static ssize_t create_pmem_region_show(struct device *dev,
2503 struct device_attribute *attr, char *buf)
2505 return __create_region_show(to_cxl_root_decoder(dev), buf);
2508 static ssize_t create_ram_region_show(struct device *dev,
2509 struct device_attribute *attr, char *buf)
2511 return __create_region_show(to_cxl_root_decoder(dev), buf);
2514 static struct cxl_region *__create_region(struct cxl_root_decoder *cxlrd,
2515 enum cxl_decoder_mode mode, int id)
2517 int rc;
2519 switch (mode) {
2520 case CXL_DECODER_RAM:
2521 case CXL_DECODER_PMEM:
2522 break;
2523 default:
2524 dev_err(&cxlrd->cxlsd.cxld.dev, "unsupported mode %d\n", mode);
2525 return ERR_PTR(-EINVAL);
2528 rc = memregion_alloc(GFP_KERNEL);
2529 if (rc < 0)
2530 return ERR_PTR(rc);
2532 if (atomic_cmpxchg(&cxlrd->region_id, id, rc) != id) {
2533 memregion_free(rc);
2534 return ERR_PTR(-EBUSY);
2537 return devm_cxl_add_region(cxlrd, id, mode, CXL_DECODER_HOSTONLYMEM);
2540 static ssize_t create_region_store(struct device *dev, const char *buf,
2541 size_t len, enum cxl_decoder_mode mode)
2543 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(dev);
2544 struct cxl_region *cxlr;
2545 int rc, id;
2547 rc = sscanf(buf, "region%d\n", &id);
2548 if (rc != 1)
2549 return -EINVAL;
2551 cxlr = __create_region(cxlrd, mode, id);
2552 if (IS_ERR(cxlr))
2553 return PTR_ERR(cxlr);
2555 return len;
2558 static ssize_t create_pmem_region_store(struct device *dev,
2559 struct device_attribute *attr,
2560 const char *buf, size_t len)
2562 return create_region_store(dev, buf, len, CXL_DECODER_PMEM);
2564 DEVICE_ATTR_RW(create_pmem_region);
2566 static ssize_t create_ram_region_store(struct device *dev,
2567 struct device_attribute *attr,
2568 const char *buf, size_t len)
2570 return create_region_store(dev, buf, len, CXL_DECODER_RAM);
2572 DEVICE_ATTR_RW(create_ram_region);
2574 static ssize_t region_show(struct device *dev, struct device_attribute *attr,
2575 char *buf)
2577 struct cxl_decoder *cxld = to_cxl_decoder(dev);
2578 ssize_t rc;
2580 rc = down_read_interruptible(&cxl_region_rwsem);
2581 if (rc)
2582 return rc;
2584 if (cxld->region)
2585 rc = sysfs_emit(buf, "%s\n", dev_name(&cxld->region->dev));
2586 else
2587 rc = sysfs_emit(buf, "\n");
2588 up_read(&cxl_region_rwsem);
2590 return rc;
2592 DEVICE_ATTR_RO(region);
2594 static struct cxl_region *
2595 cxl_find_region_by_name(struct cxl_root_decoder *cxlrd, const char *name)
2597 struct cxl_decoder *cxld = &cxlrd->cxlsd.cxld;
2598 struct device *region_dev;
2600 region_dev = device_find_child_by_name(&cxld->dev, name);
2601 if (!region_dev)
2602 return ERR_PTR(-ENODEV);
2604 return to_cxl_region(region_dev);
2607 static ssize_t delete_region_store(struct device *dev,
2608 struct device_attribute *attr,
2609 const char *buf, size_t len)
2611 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(dev);
2612 struct cxl_port *port = to_cxl_port(dev->parent);
2613 struct cxl_region *cxlr;
2615 cxlr = cxl_find_region_by_name(cxlrd, buf);
2616 if (IS_ERR(cxlr))
2617 return PTR_ERR(cxlr);
2619 devm_release_action(port->uport_dev, unregister_region, cxlr);
2620 put_device(&cxlr->dev);
2622 return len;
2624 DEVICE_ATTR_WO(delete_region);
2626 static void cxl_pmem_region_release(struct device *dev)
2628 struct cxl_pmem_region *cxlr_pmem = to_cxl_pmem_region(dev);
2629 int i;
2631 for (i = 0; i < cxlr_pmem->nr_mappings; i++) {
2632 struct cxl_memdev *cxlmd = cxlr_pmem->mapping[i].cxlmd;
2634 put_device(&cxlmd->dev);
2637 kfree(cxlr_pmem);
2640 static const struct attribute_group *cxl_pmem_region_attribute_groups[] = {
2641 &cxl_base_attribute_group,
2642 NULL,
2645 const struct device_type cxl_pmem_region_type = {
2646 .name = "cxl_pmem_region",
2647 .release = cxl_pmem_region_release,
2648 .groups = cxl_pmem_region_attribute_groups,
2651 bool is_cxl_pmem_region(struct device *dev)
2653 return dev->type == &cxl_pmem_region_type;
2655 EXPORT_SYMBOL_NS_GPL(is_cxl_pmem_region, CXL);
2657 struct cxl_pmem_region *to_cxl_pmem_region(struct device *dev)
2659 if (dev_WARN_ONCE(dev, !is_cxl_pmem_region(dev),
2660 "not a cxl_pmem_region device\n"))
2661 return NULL;
2662 return container_of(dev, struct cxl_pmem_region, dev);
2664 EXPORT_SYMBOL_NS_GPL(to_cxl_pmem_region, CXL);
2666 struct cxl_poison_context {
2667 struct cxl_port *port;
2668 enum cxl_decoder_mode mode;
2669 u64 offset;
2672 static int cxl_get_poison_unmapped(struct cxl_memdev *cxlmd,
2673 struct cxl_poison_context *ctx)
2675 struct cxl_dev_state *cxlds = cxlmd->cxlds;
2676 u64 offset, length;
2677 int rc = 0;
2680 * Collect poison for the remaining unmapped resources
2681 * after poison is collected by committed endpoints.
2683 * Knowing that PMEM must always follow RAM, get poison
2684 * for unmapped resources based on the last decoder's mode:
2685 * ram: scan remains of ram range, then any pmem range
2686 * pmem: scan remains of pmem range
2689 if (ctx->mode == CXL_DECODER_RAM) {
2690 offset = ctx->offset;
2691 length = resource_size(&cxlds->ram_res) - offset;
2692 rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
2693 if (rc == -EFAULT)
2694 rc = 0;
2695 if (rc)
2696 return rc;
2698 if (ctx->mode == CXL_DECODER_PMEM) {
2699 offset = ctx->offset;
2700 length = resource_size(&cxlds->dpa_res) - offset;
2701 if (!length)
2702 return 0;
2703 } else if (resource_size(&cxlds->pmem_res)) {
2704 offset = cxlds->pmem_res.start;
2705 length = resource_size(&cxlds->pmem_res);
2706 } else {
2707 return 0;
2710 return cxl_mem_get_poison(cxlmd, offset, length, NULL);
2713 static int poison_by_decoder(struct device *dev, void *arg)
2715 struct cxl_poison_context *ctx = arg;
2716 struct cxl_endpoint_decoder *cxled;
2717 struct cxl_memdev *cxlmd;
2718 u64 offset, length;
2719 int rc = 0;
2721 if (!is_endpoint_decoder(dev))
2722 return rc;
2724 cxled = to_cxl_endpoint_decoder(dev);
2725 if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
2726 return rc;
2729 * Regions are only created with single mode decoders: pmem or ram.
2730 * Linux does not support mixed mode decoders. This means that
2731 * reading poison per endpoint decoder adheres to the requirement
2732 * that poison reads of pmem and ram must be separated.
2733 * CXL 3.0 Spec 8.2.9.8.4.1
2735 if (cxled->mode == CXL_DECODER_MIXED) {
2736 dev_dbg(dev, "poison list read unsupported in mixed mode\n");
2737 return rc;
2740 cxlmd = cxled_to_memdev(cxled);
2741 if (cxled->skip) {
2742 offset = cxled->dpa_res->start - cxled->skip;
2743 length = cxled->skip;
2744 rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
2745 if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
2746 rc = 0;
2747 if (rc)
2748 return rc;
2751 offset = cxled->dpa_res->start;
2752 length = cxled->dpa_res->end - offset + 1;
2753 rc = cxl_mem_get_poison(cxlmd, offset, length, cxled->cxld.region);
2754 if (rc == -EFAULT && cxled->mode == CXL_DECODER_RAM)
2755 rc = 0;
2756 if (rc)
2757 return rc;
2759 /* Iterate until commit_end is reached */
2760 if (cxled->cxld.id == ctx->port->commit_end) {
2761 ctx->offset = cxled->dpa_res->end + 1;
2762 ctx->mode = cxled->mode;
2763 return 1;
2766 return 0;
2769 int cxl_get_poison_by_endpoint(struct cxl_port *port)
2771 struct cxl_poison_context ctx;
2772 int rc = 0;
2774 ctx = (struct cxl_poison_context) {
2775 .port = port
2778 rc = device_for_each_child(&port->dev, &ctx, poison_by_decoder);
2779 if (rc == 1)
2780 rc = cxl_get_poison_unmapped(to_cxl_memdev(port->uport_dev),
2781 &ctx);
2783 return rc;
2786 struct cxl_dpa_to_region_context {
2787 struct cxl_region *cxlr;
2788 u64 dpa;
2791 static int __cxl_dpa_to_region(struct device *dev, void *arg)
2793 struct cxl_dpa_to_region_context *ctx = arg;
2794 struct cxl_endpoint_decoder *cxled;
2795 struct cxl_region *cxlr;
2796 u64 dpa = ctx->dpa;
2798 if (!is_endpoint_decoder(dev))
2799 return 0;
2801 cxled = to_cxl_endpoint_decoder(dev);
2802 if (!cxled || !cxled->dpa_res || !resource_size(cxled->dpa_res))
2803 return 0;
2805 if (dpa > cxled->dpa_res->end || dpa < cxled->dpa_res->start)
2806 return 0;
2809 * Stop the region search (return 1) when an endpoint mapping is
2810 * found. The region may not be fully constructed so offering
2811 * the cxlr in the context structure is not guaranteed.
2813 cxlr = cxled->cxld.region;
2814 if (cxlr)
2815 dev_dbg(dev, "dpa:0x%llx mapped in region:%s\n", dpa,
2816 dev_name(&cxlr->dev));
2817 else
2818 dev_dbg(dev, "dpa:0x%llx mapped in endpoint:%s\n", dpa,
2819 dev_name(dev));
2821 ctx->cxlr = cxlr;
2823 return 1;
2826 struct cxl_region *cxl_dpa_to_region(const struct cxl_memdev *cxlmd, u64 dpa)
2828 struct cxl_dpa_to_region_context ctx;
2829 struct cxl_port *port;
2831 ctx = (struct cxl_dpa_to_region_context) {
2832 .dpa = dpa,
2834 port = cxlmd->endpoint;
2835 if (port && is_cxl_endpoint(port) && cxl_num_decoders_committed(port))
2836 device_for_each_child(&port->dev, &ctx, __cxl_dpa_to_region);
2838 return ctx.cxlr;
2841 static bool cxl_is_hpa_in_chunk(u64 hpa, struct cxl_region *cxlr, int pos)
2843 struct cxl_region_params *p = &cxlr->params;
2844 int gran = p->interleave_granularity;
2845 int ways = p->interleave_ways;
2846 u64 offset;
2848 /* Is the hpa in an expected chunk for its pos(-ition) */
2849 offset = hpa - p->res->start;
2850 offset = do_div(offset, gran * ways);
2851 if ((offset >= pos * gran) && (offset < (pos + 1) * gran))
2852 return true;
2854 dev_dbg(&cxlr->dev,
2855 "Addr trans fail: hpa 0x%llx not in expected chunk\n", hpa);
2857 return false;
2860 u64 cxl_dpa_to_hpa(struct cxl_region *cxlr, const struct cxl_memdev *cxlmd,
2861 u64 dpa)
2863 struct cxl_root_decoder *cxlrd = to_cxl_root_decoder(cxlr->dev.parent);
2864 u64 dpa_offset, hpa_offset, bits_upper, mask_upper, hpa;
2865 struct cxl_region_params *p = &cxlr->params;
2866 struct cxl_endpoint_decoder *cxled = NULL;
2867 u16 eig = 0;
2868 u8 eiw = 0;
2869 int pos;
2871 for (int i = 0; i < p->nr_targets; i++) {
2872 cxled = p->targets[i];
2873 if (cxlmd == cxled_to_memdev(cxled))
2874 break;
2876 if (!cxled || cxlmd != cxled_to_memdev(cxled))
2877 return ULLONG_MAX;
2879 pos = cxled->pos;
2880 ways_to_eiw(p->interleave_ways, &eiw);
2881 granularity_to_eig(p->interleave_granularity, &eig);
2884 * The device position in the region interleave set was removed
2885 * from the offset at HPA->DPA translation. To reconstruct the
2886 * HPA, place the 'pos' in the offset.
2888 * The placement of 'pos' in the HPA is determined by interleave
2889 * ways and granularity and is defined in the CXL Spec 3.0 Section
2890 * 8.2.4.19.13 Implementation Note: Device Decode Logic
2893 /* Remove the dpa base */
2894 dpa_offset = dpa - cxl_dpa_resource_start(cxled);
2896 mask_upper = GENMASK_ULL(51, eig + 8);
2898 if (eiw < 8) {
2899 hpa_offset = (dpa_offset & mask_upper) << eiw;
2900 hpa_offset |= pos << (eig + 8);
2901 } else {
2902 bits_upper = (dpa_offset & mask_upper) >> (eig + 8);
2903 bits_upper = bits_upper * 3;
2904 hpa_offset = ((bits_upper << (eiw - 8)) + pos) << (eig + 8);
2907 /* The lower bits remain unchanged */
2908 hpa_offset |= dpa_offset & GENMASK_ULL(eig + 7, 0);
2910 /* Apply the hpa_offset to the region base address */
2911 hpa = hpa_offset + p->res->start;
2913 /* Root decoder translation overrides typical modulo decode */
2914 if (cxlrd->hpa_to_spa)
2915 hpa = cxlrd->hpa_to_spa(cxlrd, hpa);
2917 if (hpa < p->res->start || hpa > p->res->end) {
2918 dev_dbg(&cxlr->dev,
2919 "Addr trans fail: hpa 0x%llx not in region\n", hpa);
2920 return ULLONG_MAX;
2923 /* Simple chunk check, by pos & gran, only applies to modulo decodes */
2924 if (!cxlrd->hpa_to_spa && (!cxl_is_hpa_in_chunk(hpa, cxlr, pos)))
2925 return ULLONG_MAX;
2927 return hpa;
2930 static struct lock_class_key cxl_pmem_region_key;
2932 static int cxl_pmem_region_alloc(struct cxl_region *cxlr)
2934 struct cxl_region_params *p = &cxlr->params;
2935 struct cxl_nvdimm_bridge *cxl_nvb;
2936 struct device *dev;
2937 int i;
2939 guard(rwsem_read)(&cxl_region_rwsem);
2940 if (p->state != CXL_CONFIG_COMMIT)
2941 return -ENXIO;
2943 struct cxl_pmem_region *cxlr_pmem __free(kfree) =
2944 kzalloc(struct_size(cxlr_pmem, mapping, p->nr_targets), GFP_KERNEL);
2945 if (!cxlr_pmem)
2946 return -ENOMEM;
2948 cxlr_pmem->hpa_range.start = p->res->start;
2949 cxlr_pmem->hpa_range.end = p->res->end;
2951 /* Snapshot the region configuration underneath the cxl_region_rwsem */
2952 cxlr_pmem->nr_mappings = p->nr_targets;
2953 for (i = 0; i < p->nr_targets; i++) {
2954 struct cxl_endpoint_decoder *cxled = p->targets[i];
2955 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
2956 struct cxl_pmem_region_mapping *m = &cxlr_pmem->mapping[i];
2959 * Regions never span CXL root devices, so by definition the
2960 * bridge for one device is the same for all.
2962 if (i == 0) {
2963 cxl_nvb = cxl_find_nvdimm_bridge(cxlmd->endpoint);
2964 if (!cxl_nvb)
2965 return -ENODEV;
2966 cxlr->cxl_nvb = cxl_nvb;
2968 m->cxlmd = cxlmd;
2969 get_device(&cxlmd->dev);
2970 m->start = cxled->dpa_res->start;
2971 m->size = resource_size(cxled->dpa_res);
2972 m->position = i;
2975 dev = &cxlr_pmem->dev;
2976 device_initialize(dev);
2977 lockdep_set_class(&dev->mutex, &cxl_pmem_region_key);
2978 device_set_pm_not_required(dev);
2979 dev->parent = &cxlr->dev;
2980 dev->bus = &cxl_bus_type;
2981 dev->type = &cxl_pmem_region_type;
2982 cxlr_pmem->cxlr = cxlr;
2983 cxlr->cxlr_pmem = no_free_ptr(cxlr_pmem);
2985 return 0;
2988 static void cxl_dax_region_release(struct device *dev)
2990 struct cxl_dax_region *cxlr_dax = to_cxl_dax_region(dev);
2992 kfree(cxlr_dax);
2995 static const struct attribute_group *cxl_dax_region_attribute_groups[] = {
2996 &cxl_base_attribute_group,
2997 NULL,
3000 const struct device_type cxl_dax_region_type = {
3001 .name = "cxl_dax_region",
3002 .release = cxl_dax_region_release,
3003 .groups = cxl_dax_region_attribute_groups,
3006 static bool is_cxl_dax_region(struct device *dev)
3008 return dev->type == &cxl_dax_region_type;
3011 struct cxl_dax_region *to_cxl_dax_region(struct device *dev)
3013 if (dev_WARN_ONCE(dev, !is_cxl_dax_region(dev),
3014 "not a cxl_dax_region device\n"))
3015 return NULL;
3016 return container_of(dev, struct cxl_dax_region, dev);
3018 EXPORT_SYMBOL_NS_GPL(to_cxl_dax_region, CXL);
3020 static struct lock_class_key cxl_dax_region_key;
3022 static struct cxl_dax_region *cxl_dax_region_alloc(struct cxl_region *cxlr)
3024 struct cxl_region_params *p = &cxlr->params;
3025 struct cxl_dax_region *cxlr_dax;
3026 struct device *dev;
3028 down_read(&cxl_region_rwsem);
3029 if (p->state != CXL_CONFIG_COMMIT) {
3030 cxlr_dax = ERR_PTR(-ENXIO);
3031 goto out;
3034 cxlr_dax = kzalloc(sizeof(*cxlr_dax), GFP_KERNEL);
3035 if (!cxlr_dax) {
3036 cxlr_dax = ERR_PTR(-ENOMEM);
3037 goto out;
3040 cxlr_dax->hpa_range.start = p->res->start;
3041 cxlr_dax->hpa_range.end = p->res->end;
3043 dev = &cxlr_dax->dev;
3044 cxlr_dax->cxlr = cxlr;
3045 device_initialize(dev);
3046 lockdep_set_class(&dev->mutex, &cxl_dax_region_key);
3047 device_set_pm_not_required(dev);
3048 dev->parent = &cxlr->dev;
3049 dev->bus = &cxl_bus_type;
3050 dev->type = &cxl_dax_region_type;
3051 out:
3052 up_read(&cxl_region_rwsem);
3054 return cxlr_dax;
3057 static void cxlr_pmem_unregister(void *_cxlr_pmem)
3059 struct cxl_pmem_region *cxlr_pmem = _cxlr_pmem;
3060 struct cxl_region *cxlr = cxlr_pmem->cxlr;
3061 struct cxl_nvdimm_bridge *cxl_nvb = cxlr->cxl_nvb;
3064 * Either the bridge is in ->remove() context under the device_lock(),
3065 * or cxlr_release_nvdimm() is cancelling the bridge's release action
3066 * for @cxlr_pmem and doing it itself (while manually holding the bridge
3067 * lock).
3069 device_lock_assert(&cxl_nvb->dev);
3070 cxlr->cxlr_pmem = NULL;
3071 cxlr_pmem->cxlr = NULL;
3072 device_unregister(&cxlr_pmem->dev);
3075 static void cxlr_release_nvdimm(void *_cxlr)
3077 struct cxl_region *cxlr = _cxlr;
3078 struct cxl_nvdimm_bridge *cxl_nvb = cxlr->cxl_nvb;
3080 scoped_guard(device, &cxl_nvb->dev) {
3081 if (cxlr->cxlr_pmem)
3082 devm_release_action(&cxl_nvb->dev, cxlr_pmem_unregister,
3083 cxlr->cxlr_pmem);
3085 cxlr->cxl_nvb = NULL;
3086 put_device(&cxl_nvb->dev);
3090 * devm_cxl_add_pmem_region() - add a cxl_region-to-nd_region bridge
3091 * @cxlr: parent CXL region for this pmem region bridge device
3093 * Return: 0 on success negative error code on failure.
3095 static int devm_cxl_add_pmem_region(struct cxl_region *cxlr)
3097 struct cxl_pmem_region *cxlr_pmem;
3098 struct cxl_nvdimm_bridge *cxl_nvb;
3099 struct device *dev;
3100 int rc;
3102 rc = cxl_pmem_region_alloc(cxlr);
3103 if (rc)
3104 return rc;
3105 cxlr_pmem = cxlr->cxlr_pmem;
3106 cxl_nvb = cxlr->cxl_nvb;
3108 dev = &cxlr_pmem->dev;
3109 rc = dev_set_name(dev, "pmem_region%d", cxlr->id);
3110 if (rc)
3111 goto err;
3113 rc = device_add(dev);
3114 if (rc)
3115 goto err;
3117 dev_dbg(&cxlr->dev, "%s: register %s\n", dev_name(dev->parent),
3118 dev_name(dev));
3120 scoped_guard(device, &cxl_nvb->dev) {
3121 if (cxl_nvb->dev.driver)
3122 rc = devm_add_action_or_reset(&cxl_nvb->dev,
3123 cxlr_pmem_unregister,
3124 cxlr_pmem);
3125 else
3126 rc = -ENXIO;
3129 if (rc)
3130 goto err_bridge;
3132 /* @cxlr carries a reference on @cxl_nvb until cxlr_release_nvdimm */
3133 return devm_add_action_or_reset(&cxlr->dev, cxlr_release_nvdimm, cxlr);
3135 err:
3136 put_device(dev);
3137 err_bridge:
3138 put_device(&cxl_nvb->dev);
3139 cxlr->cxl_nvb = NULL;
3140 return rc;
3143 static void cxlr_dax_unregister(void *_cxlr_dax)
3145 struct cxl_dax_region *cxlr_dax = _cxlr_dax;
3147 device_unregister(&cxlr_dax->dev);
3150 static int devm_cxl_add_dax_region(struct cxl_region *cxlr)
3152 struct cxl_dax_region *cxlr_dax;
3153 struct device *dev;
3154 int rc;
3156 cxlr_dax = cxl_dax_region_alloc(cxlr);
3157 if (IS_ERR(cxlr_dax))
3158 return PTR_ERR(cxlr_dax);
3160 dev = &cxlr_dax->dev;
3161 rc = dev_set_name(dev, "dax_region%d", cxlr->id);
3162 if (rc)
3163 goto err;
3165 rc = device_add(dev);
3166 if (rc)
3167 goto err;
3169 dev_dbg(&cxlr->dev, "%s: register %s\n", dev_name(dev->parent),
3170 dev_name(dev));
3172 return devm_add_action_or_reset(&cxlr->dev, cxlr_dax_unregister,
3173 cxlr_dax);
3174 err:
3175 put_device(dev);
3176 return rc;
3179 static int match_root_decoder_by_range(struct device *dev, void *data)
3181 struct range *r1, *r2 = data;
3182 struct cxl_root_decoder *cxlrd;
3184 if (!is_root_decoder(dev))
3185 return 0;
3187 cxlrd = to_cxl_root_decoder(dev);
3188 r1 = &cxlrd->cxlsd.cxld.hpa_range;
3189 return range_contains(r1, r2);
3192 static int match_region_by_range(struct device *dev, void *data)
3194 struct cxl_region_params *p;
3195 struct cxl_region *cxlr;
3196 struct range *r = data;
3197 int rc = 0;
3199 if (!is_cxl_region(dev))
3200 return 0;
3202 cxlr = to_cxl_region(dev);
3203 p = &cxlr->params;
3205 down_read(&cxl_region_rwsem);
3206 if (p->res && p->res->start == r->start && p->res->end == r->end)
3207 rc = 1;
3208 up_read(&cxl_region_rwsem);
3210 return rc;
3213 /* Establish an empty region covering the given HPA range */
3214 static struct cxl_region *construct_region(struct cxl_root_decoder *cxlrd,
3215 struct cxl_endpoint_decoder *cxled)
3217 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
3218 struct cxl_port *port = cxlrd_to_port(cxlrd);
3219 struct range *hpa = &cxled->cxld.hpa_range;
3220 struct cxl_region_params *p;
3221 struct cxl_region *cxlr;
3222 struct resource *res;
3223 int rc;
3225 do {
3226 cxlr = __create_region(cxlrd, cxled->mode,
3227 atomic_read(&cxlrd->region_id));
3228 } while (IS_ERR(cxlr) && PTR_ERR(cxlr) == -EBUSY);
3230 if (IS_ERR(cxlr)) {
3231 dev_err(cxlmd->dev.parent,
3232 "%s:%s: %s failed assign region: %ld\n",
3233 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
3234 __func__, PTR_ERR(cxlr));
3235 return cxlr;
3238 down_write(&cxl_region_rwsem);
3239 p = &cxlr->params;
3240 if (p->state >= CXL_CONFIG_INTERLEAVE_ACTIVE) {
3241 dev_err(cxlmd->dev.parent,
3242 "%s:%s: %s autodiscovery interrupted\n",
3243 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
3244 __func__);
3245 rc = -EBUSY;
3246 goto err;
3249 set_bit(CXL_REGION_F_AUTO, &cxlr->flags);
3251 res = kmalloc(sizeof(*res), GFP_KERNEL);
3252 if (!res) {
3253 rc = -ENOMEM;
3254 goto err;
3257 *res = DEFINE_RES_MEM_NAMED(hpa->start, range_len(hpa),
3258 dev_name(&cxlr->dev));
3259 rc = insert_resource(cxlrd->res, res);
3260 if (rc) {
3262 * Platform-firmware may not have split resources like "System
3263 * RAM" on CXL window boundaries see cxl_region_iomem_release()
3265 dev_warn(cxlmd->dev.parent,
3266 "%s:%s: %s %s cannot insert resource\n",
3267 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev),
3268 __func__, dev_name(&cxlr->dev));
3271 p->res = res;
3272 p->interleave_ways = cxled->cxld.interleave_ways;
3273 p->interleave_granularity = cxled->cxld.interleave_granularity;
3274 p->state = CXL_CONFIG_INTERLEAVE_ACTIVE;
3276 rc = sysfs_update_group(&cxlr->dev.kobj, get_cxl_region_target_group());
3277 if (rc)
3278 goto err;
3280 dev_dbg(cxlmd->dev.parent, "%s:%s: %s %s res: %pr iw: %d ig: %d\n",
3281 dev_name(&cxlmd->dev), dev_name(&cxled->cxld.dev), __func__,
3282 dev_name(&cxlr->dev), p->res, p->interleave_ways,
3283 p->interleave_granularity);
3285 /* ...to match put_device() in cxl_add_to_region() */
3286 get_device(&cxlr->dev);
3287 up_write(&cxl_region_rwsem);
3289 return cxlr;
3291 err:
3292 up_write(&cxl_region_rwsem);
3293 devm_release_action(port->uport_dev, unregister_region, cxlr);
3294 return ERR_PTR(rc);
3297 int cxl_add_to_region(struct cxl_port *root, struct cxl_endpoint_decoder *cxled)
3299 struct cxl_memdev *cxlmd = cxled_to_memdev(cxled);
3300 struct range *hpa = &cxled->cxld.hpa_range;
3301 struct cxl_decoder *cxld = &cxled->cxld;
3302 struct device *cxlrd_dev, *region_dev;
3303 struct cxl_root_decoder *cxlrd;
3304 struct cxl_region_params *p;
3305 struct cxl_region *cxlr;
3306 bool attach = false;
3307 int rc;
3309 cxlrd_dev = device_find_child(&root->dev, &cxld->hpa_range,
3310 match_root_decoder_by_range);
3311 if (!cxlrd_dev) {
3312 dev_err(cxlmd->dev.parent,
3313 "%s:%s no CXL window for range %#llx:%#llx\n",
3314 dev_name(&cxlmd->dev), dev_name(&cxld->dev),
3315 cxld->hpa_range.start, cxld->hpa_range.end);
3316 return -ENXIO;
3319 cxlrd = to_cxl_root_decoder(cxlrd_dev);
3322 * Ensure that if multiple threads race to construct_region() for @hpa
3323 * one does the construction and the others add to that.
3325 mutex_lock(&cxlrd->range_lock);
3326 region_dev = device_find_child(&cxlrd->cxlsd.cxld.dev, hpa,
3327 match_region_by_range);
3328 if (!region_dev) {
3329 cxlr = construct_region(cxlrd, cxled);
3330 region_dev = &cxlr->dev;
3331 } else
3332 cxlr = to_cxl_region(region_dev);
3333 mutex_unlock(&cxlrd->range_lock);
3335 rc = PTR_ERR_OR_ZERO(cxlr);
3336 if (rc)
3337 goto out;
3339 attach_target(cxlr, cxled, -1, TASK_UNINTERRUPTIBLE);
3341 down_read(&cxl_region_rwsem);
3342 p = &cxlr->params;
3343 attach = p->state == CXL_CONFIG_COMMIT;
3344 up_read(&cxl_region_rwsem);
3346 if (attach) {
3348 * If device_attach() fails the range may still be active via
3349 * the platform-firmware memory map, otherwise the driver for
3350 * regions is local to this file, so driver matching can't fail.
3352 if (device_attach(&cxlr->dev) < 0)
3353 dev_err(&cxlr->dev, "failed to enable, range: %pr\n",
3354 p->res);
3357 put_device(region_dev);
3358 out:
3359 put_device(cxlrd_dev);
3360 return rc;
3362 EXPORT_SYMBOL_NS_GPL(cxl_add_to_region, CXL);
3364 static int is_system_ram(struct resource *res, void *arg)
3366 struct cxl_region *cxlr = arg;
3367 struct cxl_region_params *p = &cxlr->params;
3369 dev_dbg(&cxlr->dev, "%pr has System RAM: %pr\n", p->res, res);
3370 return 1;
3373 static void shutdown_notifiers(void *_cxlr)
3375 struct cxl_region *cxlr = _cxlr;
3377 unregister_memory_notifier(&cxlr->memory_notifier);
3378 unregister_mt_adistance_algorithm(&cxlr->adist_notifier);
3381 static int cxl_region_probe(struct device *dev)
3383 struct cxl_region *cxlr = to_cxl_region(dev);
3384 struct cxl_region_params *p = &cxlr->params;
3385 int rc;
3387 rc = down_read_interruptible(&cxl_region_rwsem);
3388 if (rc) {
3389 dev_dbg(&cxlr->dev, "probe interrupted\n");
3390 return rc;
3393 if (p->state < CXL_CONFIG_COMMIT) {
3394 dev_dbg(&cxlr->dev, "config state: %d\n", p->state);
3395 rc = -ENXIO;
3396 goto out;
3399 if (test_bit(CXL_REGION_F_NEEDS_RESET, &cxlr->flags)) {
3400 dev_err(&cxlr->dev,
3401 "failed to activate, re-commit region and retry\n");
3402 rc = -ENXIO;
3403 goto out;
3407 * From this point on any path that changes the region's state away from
3408 * CXL_CONFIG_COMMIT is also responsible for releasing the driver.
3410 out:
3411 up_read(&cxl_region_rwsem);
3413 if (rc)
3414 return rc;
3416 cxlr->memory_notifier.notifier_call = cxl_region_perf_attrs_callback;
3417 cxlr->memory_notifier.priority = CXL_CALLBACK_PRI;
3418 register_memory_notifier(&cxlr->memory_notifier);
3420 cxlr->adist_notifier.notifier_call = cxl_region_calculate_adistance;
3421 cxlr->adist_notifier.priority = 100;
3422 register_mt_adistance_algorithm(&cxlr->adist_notifier);
3424 rc = devm_add_action_or_reset(&cxlr->dev, shutdown_notifiers, cxlr);
3425 if (rc)
3426 return rc;
3428 switch (cxlr->mode) {
3429 case CXL_DECODER_PMEM:
3430 return devm_cxl_add_pmem_region(cxlr);
3431 case CXL_DECODER_RAM:
3433 * The region can not be manged by CXL if any portion of
3434 * it is already online as 'System RAM'
3436 if (walk_iomem_res_desc(IORES_DESC_NONE,
3437 IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY,
3438 p->res->start, p->res->end, cxlr,
3439 is_system_ram) > 0)
3440 return 0;
3441 return devm_cxl_add_dax_region(cxlr);
3442 default:
3443 dev_dbg(&cxlr->dev, "unsupported region mode: %d\n",
3444 cxlr->mode);
3445 return -ENXIO;
3449 static struct cxl_driver cxl_region_driver = {
3450 .name = "cxl_region",
3451 .probe = cxl_region_probe,
3452 .id = CXL_DEVICE_REGION,
3455 int cxl_region_init(void)
3457 return cxl_driver_register(&cxl_region_driver);
3460 void cxl_region_exit(void)
3462 cxl_driver_unregister(&cxl_region_driver);
3465 MODULE_IMPORT_NS(CXL);
3466 MODULE_IMPORT_NS(DEVMEM);
3467 MODULE_ALIAS_CXL(CXL_DEVICE_REGION);