xref: /linux/drivers/cxl/core/memdev.c (revision 2d7f3d1a5866705be2393150e1ffdf67030ab88d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2020 Intel Corporation. */
3 
4 #include <linux/io-64-nonatomic-lo-hi.h>
5 #include <linux/firmware.h>
6 #include <linux/device.h>
7 #include <linux/slab.h>
8 #include <linux/idr.h>
9 #include <linux/pci.h>
10 #include <cxlmem.h>
11 #include "trace.h"
12 #include "core.h"
13 
14 static DECLARE_RWSEM(cxl_memdev_rwsem);
15 
16 /*
17  * An entire PCI topology full of devices should be enough for any
18  * config
19  */
20 #define CXL_MEM_MAX_DEVS 65536
21 
22 static int cxl_mem_major;
23 static DEFINE_IDA(cxl_memdev_ida);
24 
25 static void cxl_memdev_release(struct device *dev)
26 {
27 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
28 
29 	ida_free(&cxl_memdev_ida, cxlmd->id);
30 	kfree(cxlmd);
31 }
32 
33 static char *cxl_memdev_devnode(const struct device *dev, umode_t *mode, kuid_t *uid,
34 				kgid_t *gid)
35 {
36 	return kasprintf(GFP_KERNEL, "cxl/%s", dev_name(dev));
37 }
38 
39 static ssize_t firmware_version_show(struct device *dev,
40 				     struct device_attribute *attr, char *buf)
41 {
42 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
43 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
44 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
45 
46 	if (!mds)
47 		return sysfs_emit(buf, "\n");
48 	return sysfs_emit(buf, "%.16s\n", mds->firmware_version);
49 }
50 static DEVICE_ATTR_RO(firmware_version);
51 
52 static ssize_t payload_max_show(struct device *dev,
53 				struct device_attribute *attr, char *buf)
54 {
55 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
56 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
57 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
58 
59 	if (!mds)
60 		return sysfs_emit(buf, "\n");
61 	return sysfs_emit(buf, "%zu\n", mds->payload_size);
62 }
63 static DEVICE_ATTR_RO(payload_max);
64 
65 static ssize_t label_storage_size_show(struct device *dev,
66 				       struct device_attribute *attr, char *buf)
67 {
68 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
69 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
70 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
71 
72 	if (!mds)
73 		return sysfs_emit(buf, "\n");
74 	return sysfs_emit(buf, "%zu\n", mds->lsa_size);
75 }
76 static DEVICE_ATTR_RO(label_storage_size);
77 
78 static ssize_t ram_size_show(struct device *dev, struct device_attribute *attr,
79 			     char *buf)
80 {
81 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
82 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
83 	unsigned long long len = resource_size(&cxlds->ram_res);
84 
85 	return sysfs_emit(buf, "%#llx\n", len);
86 }
87 
88 static struct device_attribute dev_attr_ram_size =
89 	__ATTR(size, 0444, ram_size_show, NULL);
90 
91 static ssize_t pmem_size_show(struct device *dev, struct device_attribute *attr,
92 			      char *buf)
93 {
94 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
95 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
96 	unsigned long long len = resource_size(&cxlds->pmem_res);
97 
98 	return sysfs_emit(buf, "%#llx\n", len);
99 }
100 
101 static struct device_attribute dev_attr_pmem_size =
102 	__ATTR(size, 0444, pmem_size_show, NULL);
103 
104 static ssize_t serial_show(struct device *dev, struct device_attribute *attr,
105 			   char *buf)
106 {
107 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
108 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
109 
110 	return sysfs_emit(buf, "%#llx\n", cxlds->serial);
111 }
112 static DEVICE_ATTR_RO(serial);
113 
114 static ssize_t numa_node_show(struct device *dev, struct device_attribute *attr,
115 			      char *buf)
116 {
117 	return sysfs_emit(buf, "%d\n", dev_to_node(dev));
118 }
119 static DEVICE_ATTR_RO(numa_node);
120 
121 static ssize_t security_state_show(struct device *dev,
122 				   struct device_attribute *attr,
123 				   char *buf)
124 {
125 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
126 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
127 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
128 	unsigned long state = mds->security.state;
129 	int rc = 0;
130 
131 	/* sync with latest submission state */
132 	mutex_lock(&mds->mbox_mutex);
133 	if (mds->security.sanitize_active)
134 		rc = sysfs_emit(buf, "sanitize\n");
135 	mutex_unlock(&mds->mbox_mutex);
136 	if (rc)
137 		return rc;
138 
139 	if (!(state & CXL_PMEM_SEC_STATE_USER_PASS_SET))
140 		return sysfs_emit(buf, "disabled\n");
141 	if (state & CXL_PMEM_SEC_STATE_FROZEN ||
142 	    state & CXL_PMEM_SEC_STATE_MASTER_PLIMIT ||
143 	    state & CXL_PMEM_SEC_STATE_USER_PLIMIT)
144 		return sysfs_emit(buf, "frozen\n");
145 	if (state & CXL_PMEM_SEC_STATE_LOCKED)
146 		return sysfs_emit(buf, "locked\n");
147 	else
148 		return sysfs_emit(buf, "unlocked\n");
149 }
150 static struct device_attribute dev_attr_security_state =
151 	__ATTR(state, 0444, security_state_show, NULL);
152 
153 static ssize_t security_sanitize_store(struct device *dev,
154 				       struct device_attribute *attr,
155 				       const char *buf, size_t len)
156 {
157 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
158 	bool sanitize;
159 	ssize_t rc;
160 
161 	if (kstrtobool(buf, &sanitize) || !sanitize)
162 		return -EINVAL;
163 
164 	rc = cxl_mem_sanitize(cxlmd, CXL_MBOX_OP_SANITIZE);
165 	if (rc)
166 		return rc;
167 
168 	return len;
169 }
170 static struct device_attribute dev_attr_security_sanitize =
171 	__ATTR(sanitize, 0200, NULL, security_sanitize_store);
172 
173 static ssize_t security_erase_store(struct device *dev,
174 				    struct device_attribute *attr,
175 				    const char *buf, size_t len)
176 {
177 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
178 	ssize_t rc;
179 	bool erase;
180 
181 	if (kstrtobool(buf, &erase) || !erase)
182 		return -EINVAL;
183 
184 	rc = cxl_mem_sanitize(cxlmd, CXL_MBOX_OP_SECURE_ERASE);
185 	if (rc)
186 		return rc;
187 
188 	return len;
189 }
190 static struct device_attribute dev_attr_security_erase =
191 	__ATTR(erase, 0200, NULL, security_erase_store);
192 
193 static int cxl_get_poison_by_memdev(struct cxl_memdev *cxlmd)
194 {
195 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
196 	u64 offset, length;
197 	int rc = 0;
198 
199 	/* CXL 3.0 Spec 8.2.9.8.4.1 Separate pmem and ram poison requests */
200 	if (resource_size(&cxlds->pmem_res)) {
201 		offset = cxlds->pmem_res.start;
202 		length = resource_size(&cxlds->pmem_res);
203 		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
204 		if (rc)
205 			return rc;
206 	}
207 	if (resource_size(&cxlds->ram_res)) {
208 		offset = cxlds->ram_res.start;
209 		length = resource_size(&cxlds->ram_res);
210 		rc = cxl_mem_get_poison(cxlmd, offset, length, NULL);
211 		/*
212 		 * Invalid Physical Address is not an error for
213 		 * volatile addresses. Device support is optional.
214 		 */
215 		if (rc == -EFAULT)
216 			rc = 0;
217 	}
218 	return rc;
219 }
220 
221 int cxl_trigger_poison_list(struct cxl_memdev *cxlmd)
222 {
223 	struct cxl_port *port;
224 	int rc;
225 
226 	port = cxlmd->endpoint;
227 	if (!port || !is_cxl_endpoint(port))
228 		return -EINVAL;
229 
230 	rc = down_read_interruptible(&cxl_region_rwsem);
231 	if (rc)
232 		return rc;
233 
234 	rc = down_read_interruptible(&cxl_dpa_rwsem);
235 	if (rc) {
236 		up_read(&cxl_region_rwsem);
237 		return rc;
238 	}
239 
240 	if (cxl_num_decoders_committed(port) == 0) {
241 		/* No regions mapped to this memdev */
242 		rc = cxl_get_poison_by_memdev(cxlmd);
243 	} else {
244 		/* Regions mapped, collect poison by endpoint */
245 		rc =  cxl_get_poison_by_endpoint(port);
246 	}
247 	up_read(&cxl_dpa_rwsem);
248 	up_read(&cxl_region_rwsem);
249 
250 	return rc;
251 }
252 EXPORT_SYMBOL_NS_GPL(cxl_trigger_poison_list, CXL);
253 
254 struct cxl_dpa_to_region_context {
255 	struct cxl_region *cxlr;
256 	u64 dpa;
257 };
258 
259 static int __cxl_dpa_to_region(struct device *dev, void *arg)
260 {
261 	struct cxl_dpa_to_region_context *ctx = arg;
262 	struct cxl_endpoint_decoder *cxled;
263 	u64 dpa = ctx->dpa;
264 
265 	if (!is_endpoint_decoder(dev))
266 		return 0;
267 
268 	cxled = to_cxl_endpoint_decoder(dev);
269 	if (!cxled->dpa_res || !resource_size(cxled->dpa_res))
270 		return 0;
271 
272 	if (dpa > cxled->dpa_res->end || dpa < cxled->dpa_res->start)
273 		return 0;
274 
275 	dev_dbg(dev, "dpa:0x%llx mapped in region:%s\n", dpa,
276 		dev_name(&cxled->cxld.region->dev));
277 
278 	ctx->cxlr = cxled->cxld.region;
279 
280 	return 1;
281 }
282 
283 static struct cxl_region *cxl_dpa_to_region(struct cxl_memdev *cxlmd, u64 dpa)
284 {
285 	struct cxl_dpa_to_region_context ctx;
286 	struct cxl_port *port;
287 
288 	ctx = (struct cxl_dpa_to_region_context) {
289 		.dpa = dpa,
290 	};
291 	port = cxlmd->endpoint;
292 	if (port && is_cxl_endpoint(port) && cxl_num_decoders_committed(port))
293 		device_for_each_child(&port->dev, &ctx, __cxl_dpa_to_region);
294 
295 	return ctx.cxlr;
296 }
297 
298 static int cxl_validate_poison_dpa(struct cxl_memdev *cxlmd, u64 dpa)
299 {
300 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
301 
302 	if (!IS_ENABLED(CONFIG_DEBUG_FS))
303 		return 0;
304 
305 	if (!resource_size(&cxlds->dpa_res)) {
306 		dev_dbg(cxlds->dev, "device has no dpa resource\n");
307 		return -EINVAL;
308 	}
309 	if (dpa < cxlds->dpa_res.start || dpa > cxlds->dpa_res.end) {
310 		dev_dbg(cxlds->dev, "dpa:0x%llx not in resource:%pR\n",
311 			dpa, &cxlds->dpa_res);
312 		return -EINVAL;
313 	}
314 	if (!IS_ALIGNED(dpa, 64)) {
315 		dev_dbg(cxlds->dev, "dpa:0x%llx is not 64-byte aligned\n", dpa);
316 		return -EINVAL;
317 	}
318 
319 	return 0;
320 }
321 
322 int cxl_inject_poison(struct cxl_memdev *cxlmd, u64 dpa)
323 {
324 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlmd->cxlds);
325 	struct cxl_mbox_inject_poison inject;
326 	struct cxl_poison_record record;
327 	struct cxl_mbox_cmd mbox_cmd;
328 	struct cxl_region *cxlr;
329 	int rc;
330 
331 	if (!IS_ENABLED(CONFIG_DEBUG_FS))
332 		return 0;
333 
334 	rc = down_read_interruptible(&cxl_region_rwsem);
335 	if (rc)
336 		return rc;
337 
338 	rc = down_read_interruptible(&cxl_dpa_rwsem);
339 	if (rc) {
340 		up_read(&cxl_region_rwsem);
341 		return rc;
342 	}
343 
344 	rc = cxl_validate_poison_dpa(cxlmd, dpa);
345 	if (rc)
346 		goto out;
347 
348 	inject.address = cpu_to_le64(dpa);
349 	mbox_cmd = (struct cxl_mbox_cmd) {
350 		.opcode = CXL_MBOX_OP_INJECT_POISON,
351 		.size_in = sizeof(inject),
352 		.payload_in = &inject,
353 	};
354 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
355 	if (rc)
356 		goto out;
357 
358 	cxlr = cxl_dpa_to_region(cxlmd, dpa);
359 	if (cxlr)
360 		dev_warn_once(mds->cxlds.dev,
361 			      "poison inject dpa:%#llx region: %s\n", dpa,
362 			      dev_name(&cxlr->dev));
363 
364 	record = (struct cxl_poison_record) {
365 		.address = cpu_to_le64(dpa),
366 		.length = cpu_to_le32(1),
367 	};
368 	trace_cxl_poison(cxlmd, cxlr, &record, 0, 0, CXL_POISON_TRACE_INJECT);
369 out:
370 	up_read(&cxl_dpa_rwsem);
371 	up_read(&cxl_region_rwsem);
372 
373 	return rc;
374 }
375 EXPORT_SYMBOL_NS_GPL(cxl_inject_poison, CXL);
376 
377 int cxl_clear_poison(struct cxl_memdev *cxlmd, u64 dpa)
378 {
379 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlmd->cxlds);
380 	struct cxl_mbox_clear_poison clear;
381 	struct cxl_poison_record record;
382 	struct cxl_mbox_cmd mbox_cmd;
383 	struct cxl_region *cxlr;
384 	int rc;
385 
386 	if (!IS_ENABLED(CONFIG_DEBUG_FS))
387 		return 0;
388 
389 	rc = down_read_interruptible(&cxl_region_rwsem);
390 	if (rc)
391 		return rc;
392 
393 	rc = down_read_interruptible(&cxl_dpa_rwsem);
394 	if (rc) {
395 		up_read(&cxl_region_rwsem);
396 		return rc;
397 	}
398 
399 	rc = cxl_validate_poison_dpa(cxlmd, dpa);
400 	if (rc)
401 		goto out;
402 
403 	/*
404 	 * In CXL 3.0 Spec 8.2.9.8.4.3, the Clear Poison mailbox command
405 	 * is defined to accept 64 bytes of write-data, along with the
406 	 * address to clear. This driver uses zeroes as write-data.
407 	 */
408 	clear = (struct cxl_mbox_clear_poison) {
409 		.address = cpu_to_le64(dpa)
410 	};
411 
412 	mbox_cmd = (struct cxl_mbox_cmd) {
413 		.opcode = CXL_MBOX_OP_CLEAR_POISON,
414 		.size_in = sizeof(clear),
415 		.payload_in = &clear,
416 	};
417 
418 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
419 	if (rc)
420 		goto out;
421 
422 	cxlr = cxl_dpa_to_region(cxlmd, dpa);
423 	if (cxlr)
424 		dev_warn_once(mds->cxlds.dev,
425 			      "poison clear dpa:%#llx region: %s\n", dpa,
426 			      dev_name(&cxlr->dev));
427 
428 	record = (struct cxl_poison_record) {
429 		.address = cpu_to_le64(dpa),
430 		.length = cpu_to_le32(1),
431 	};
432 	trace_cxl_poison(cxlmd, cxlr, &record, 0, 0, CXL_POISON_TRACE_CLEAR);
433 out:
434 	up_read(&cxl_dpa_rwsem);
435 	up_read(&cxl_region_rwsem);
436 
437 	return rc;
438 }
439 EXPORT_SYMBOL_NS_GPL(cxl_clear_poison, CXL);
440 
441 static struct attribute *cxl_memdev_attributes[] = {
442 	&dev_attr_serial.attr,
443 	&dev_attr_firmware_version.attr,
444 	&dev_attr_payload_max.attr,
445 	&dev_attr_label_storage_size.attr,
446 	&dev_attr_numa_node.attr,
447 	NULL,
448 };
449 
450 static ssize_t pmem_qos_class_show(struct device *dev,
451 				   struct device_attribute *attr, char *buf)
452 {
453 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
454 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
455 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
456 
457 	return sysfs_emit(buf, "%d\n", mds->pmem_perf.qos_class);
458 }
459 
460 static struct device_attribute dev_attr_pmem_qos_class =
461 	__ATTR(qos_class, 0444, pmem_qos_class_show, NULL);
462 
463 static struct attribute *cxl_memdev_pmem_attributes[] = {
464 	&dev_attr_pmem_size.attr,
465 	&dev_attr_pmem_qos_class.attr,
466 	NULL,
467 };
468 
469 static ssize_t ram_qos_class_show(struct device *dev,
470 				  struct device_attribute *attr, char *buf)
471 {
472 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
473 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
474 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
475 
476 	return sysfs_emit(buf, "%d\n", mds->ram_perf.qos_class);
477 }
478 
479 static struct device_attribute dev_attr_ram_qos_class =
480 	__ATTR(qos_class, 0444, ram_qos_class_show, NULL);
481 
482 static struct attribute *cxl_memdev_ram_attributes[] = {
483 	&dev_attr_ram_size.attr,
484 	&dev_attr_ram_qos_class.attr,
485 	NULL,
486 };
487 
488 static struct attribute *cxl_memdev_security_attributes[] = {
489 	&dev_attr_security_state.attr,
490 	&dev_attr_security_sanitize.attr,
491 	&dev_attr_security_erase.attr,
492 	NULL,
493 };
494 
495 static umode_t cxl_memdev_visible(struct kobject *kobj, struct attribute *a,
496 				  int n)
497 {
498 	if (!IS_ENABLED(CONFIG_NUMA) && a == &dev_attr_numa_node.attr)
499 		return 0;
500 	return a->mode;
501 }
502 
503 static struct attribute_group cxl_memdev_attribute_group = {
504 	.attrs = cxl_memdev_attributes,
505 	.is_visible = cxl_memdev_visible,
506 };
507 
508 static umode_t cxl_ram_visible(struct kobject *kobj, struct attribute *a, int n)
509 {
510 	struct device *dev = kobj_to_dev(kobj);
511 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
512 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlmd->cxlds);
513 
514 	if (a == &dev_attr_ram_qos_class.attr)
515 		if (mds->ram_perf.qos_class == CXL_QOS_CLASS_INVALID)
516 			return 0;
517 
518 	return a->mode;
519 }
520 
521 static struct attribute_group cxl_memdev_ram_attribute_group = {
522 	.name = "ram",
523 	.attrs = cxl_memdev_ram_attributes,
524 	.is_visible = cxl_ram_visible,
525 };
526 
527 static umode_t cxl_pmem_visible(struct kobject *kobj, struct attribute *a, int n)
528 {
529 	struct device *dev = kobj_to_dev(kobj);
530 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
531 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlmd->cxlds);
532 
533 	if (a == &dev_attr_pmem_qos_class.attr)
534 		if (mds->pmem_perf.qos_class == CXL_QOS_CLASS_INVALID)
535 			return 0;
536 
537 	return a->mode;
538 }
539 
540 static struct attribute_group cxl_memdev_pmem_attribute_group = {
541 	.name = "pmem",
542 	.attrs = cxl_memdev_pmem_attributes,
543 	.is_visible = cxl_pmem_visible,
544 };
545 
546 static umode_t cxl_memdev_security_visible(struct kobject *kobj,
547 					   struct attribute *a, int n)
548 {
549 	struct device *dev = kobj_to_dev(kobj);
550 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
551 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlmd->cxlds);
552 
553 	if (a == &dev_attr_security_sanitize.attr &&
554 	    !test_bit(CXL_SEC_ENABLED_SANITIZE, mds->security.enabled_cmds))
555 		return 0;
556 
557 	if (a == &dev_attr_security_erase.attr &&
558 	    !test_bit(CXL_SEC_ENABLED_SECURE_ERASE, mds->security.enabled_cmds))
559 		return 0;
560 
561 	return a->mode;
562 }
563 
564 static struct attribute_group cxl_memdev_security_attribute_group = {
565 	.name = "security",
566 	.attrs = cxl_memdev_security_attributes,
567 	.is_visible = cxl_memdev_security_visible,
568 };
569 
570 static const struct attribute_group *cxl_memdev_attribute_groups[] = {
571 	&cxl_memdev_attribute_group,
572 	&cxl_memdev_ram_attribute_group,
573 	&cxl_memdev_pmem_attribute_group,
574 	&cxl_memdev_security_attribute_group,
575 	NULL,
576 };
577 
578 void cxl_memdev_update_perf(struct cxl_memdev *cxlmd)
579 {
580 	sysfs_update_group(&cxlmd->dev.kobj, &cxl_memdev_ram_attribute_group);
581 	sysfs_update_group(&cxlmd->dev.kobj, &cxl_memdev_pmem_attribute_group);
582 }
583 EXPORT_SYMBOL_NS_GPL(cxl_memdev_update_perf, CXL);
584 
585 static const struct device_type cxl_memdev_type = {
586 	.name = "cxl_memdev",
587 	.release = cxl_memdev_release,
588 	.devnode = cxl_memdev_devnode,
589 	.groups = cxl_memdev_attribute_groups,
590 };
591 
592 bool is_cxl_memdev(const struct device *dev)
593 {
594 	return dev->type == &cxl_memdev_type;
595 }
596 EXPORT_SYMBOL_NS_GPL(is_cxl_memdev, CXL);
597 
598 /**
599  * set_exclusive_cxl_commands() - atomically disable user cxl commands
600  * @mds: The device state to operate on
601  * @cmds: bitmap of commands to mark exclusive
602  *
603  * Grab the cxl_memdev_rwsem in write mode to flush in-flight
604  * invocations of the ioctl path and then disable future execution of
605  * commands with the command ids set in @cmds.
606  */
607 void set_exclusive_cxl_commands(struct cxl_memdev_state *mds,
608 				unsigned long *cmds)
609 {
610 	down_write(&cxl_memdev_rwsem);
611 	bitmap_or(mds->exclusive_cmds, mds->exclusive_cmds, cmds,
612 		  CXL_MEM_COMMAND_ID_MAX);
613 	up_write(&cxl_memdev_rwsem);
614 }
615 EXPORT_SYMBOL_NS_GPL(set_exclusive_cxl_commands, CXL);
616 
617 /**
618  * clear_exclusive_cxl_commands() - atomically enable user cxl commands
619  * @mds: The device state to modify
620  * @cmds: bitmap of commands to mark available for userspace
621  */
622 void clear_exclusive_cxl_commands(struct cxl_memdev_state *mds,
623 				  unsigned long *cmds)
624 {
625 	down_write(&cxl_memdev_rwsem);
626 	bitmap_andnot(mds->exclusive_cmds, mds->exclusive_cmds, cmds,
627 		      CXL_MEM_COMMAND_ID_MAX);
628 	up_write(&cxl_memdev_rwsem);
629 }
630 EXPORT_SYMBOL_NS_GPL(clear_exclusive_cxl_commands, CXL);
631 
632 static void cxl_memdev_shutdown(struct device *dev)
633 {
634 	struct cxl_memdev *cxlmd = to_cxl_memdev(dev);
635 
636 	down_write(&cxl_memdev_rwsem);
637 	cxlmd->cxlds = NULL;
638 	up_write(&cxl_memdev_rwsem);
639 }
640 
641 static void cxl_memdev_unregister(void *_cxlmd)
642 {
643 	struct cxl_memdev *cxlmd = _cxlmd;
644 	struct device *dev = &cxlmd->dev;
645 
646 	cdev_device_del(&cxlmd->cdev, dev);
647 	cxl_memdev_shutdown(dev);
648 	put_device(dev);
649 }
650 
651 static void detach_memdev(struct work_struct *work)
652 {
653 	struct cxl_memdev *cxlmd;
654 
655 	cxlmd = container_of(work, typeof(*cxlmd), detach_work);
656 	device_release_driver(&cxlmd->dev);
657 	put_device(&cxlmd->dev);
658 }
659 
660 static struct lock_class_key cxl_memdev_key;
661 
662 static struct cxl_memdev *cxl_memdev_alloc(struct cxl_dev_state *cxlds,
663 					   const struct file_operations *fops)
664 {
665 	struct cxl_memdev *cxlmd;
666 	struct device *dev;
667 	struct cdev *cdev;
668 	int rc;
669 
670 	cxlmd = kzalloc(sizeof(*cxlmd), GFP_KERNEL);
671 	if (!cxlmd)
672 		return ERR_PTR(-ENOMEM);
673 
674 	rc = ida_alloc_max(&cxl_memdev_ida, CXL_MEM_MAX_DEVS - 1, GFP_KERNEL);
675 	if (rc < 0)
676 		goto err;
677 	cxlmd->id = rc;
678 	cxlmd->depth = -1;
679 
680 	dev = &cxlmd->dev;
681 	device_initialize(dev);
682 	lockdep_set_class(&dev->mutex, &cxl_memdev_key);
683 	dev->parent = cxlds->dev;
684 	dev->bus = &cxl_bus_type;
685 	dev->devt = MKDEV(cxl_mem_major, cxlmd->id);
686 	dev->type = &cxl_memdev_type;
687 	device_set_pm_not_required(dev);
688 	INIT_WORK(&cxlmd->detach_work, detach_memdev);
689 
690 	cdev = &cxlmd->cdev;
691 	cdev_init(cdev, fops);
692 	return cxlmd;
693 
694 err:
695 	kfree(cxlmd);
696 	return ERR_PTR(rc);
697 }
698 
699 static long __cxl_memdev_ioctl(struct cxl_memdev *cxlmd, unsigned int cmd,
700 			       unsigned long arg)
701 {
702 	switch (cmd) {
703 	case CXL_MEM_QUERY_COMMANDS:
704 		return cxl_query_cmd(cxlmd, (void __user *)arg);
705 	case CXL_MEM_SEND_COMMAND:
706 		return cxl_send_cmd(cxlmd, (void __user *)arg);
707 	default:
708 		return -ENOTTY;
709 	}
710 }
711 
712 static long cxl_memdev_ioctl(struct file *file, unsigned int cmd,
713 			     unsigned long arg)
714 {
715 	struct cxl_memdev *cxlmd = file->private_data;
716 	struct cxl_dev_state *cxlds;
717 	int rc = -ENXIO;
718 
719 	down_read(&cxl_memdev_rwsem);
720 	cxlds = cxlmd->cxlds;
721 	if (cxlds && cxlds->type == CXL_DEVTYPE_CLASSMEM)
722 		rc = __cxl_memdev_ioctl(cxlmd, cmd, arg);
723 	up_read(&cxl_memdev_rwsem);
724 
725 	return rc;
726 }
727 
728 static int cxl_memdev_open(struct inode *inode, struct file *file)
729 {
730 	struct cxl_memdev *cxlmd =
731 		container_of(inode->i_cdev, typeof(*cxlmd), cdev);
732 
733 	get_device(&cxlmd->dev);
734 	file->private_data = cxlmd;
735 
736 	return 0;
737 }
738 
739 static int cxl_memdev_release_file(struct inode *inode, struct file *file)
740 {
741 	struct cxl_memdev *cxlmd =
742 		container_of(inode->i_cdev, typeof(*cxlmd), cdev);
743 
744 	put_device(&cxlmd->dev);
745 
746 	return 0;
747 }
748 
749 /**
750  * cxl_mem_get_fw_info - Get Firmware info
751  * @mds: The device data for the operation
752  *
753  * Retrieve firmware info for the device specified.
754  *
755  * Return: 0 if no error: or the result of the mailbox command.
756  *
757  * See CXL-3.0 8.2.9.3.1 Get FW Info
758  */
759 static int cxl_mem_get_fw_info(struct cxl_memdev_state *mds)
760 {
761 	struct cxl_mbox_get_fw_info info;
762 	struct cxl_mbox_cmd mbox_cmd;
763 	int rc;
764 
765 	mbox_cmd = (struct cxl_mbox_cmd) {
766 		.opcode = CXL_MBOX_OP_GET_FW_INFO,
767 		.size_out = sizeof(info),
768 		.payload_out = &info,
769 	};
770 
771 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
772 	if (rc < 0)
773 		return rc;
774 
775 	mds->fw.num_slots = info.num_slots;
776 	mds->fw.cur_slot = FIELD_GET(CXL_FW_INFO_SLOT_INFO_CUR_MASK,
777 				       info.slot_info);
778 
779 	return 0;
780 }
781 
782 /**
783  * cxl_mem_activate_fw - Activate Firmware
784  * @mds: The device data for the operation
785  * @slot: slot number to activate
786  *
787  * Activate firmware in a given slot for the device specified.
788  *
789  * Return: 0 if no error: or the result of the mailbox command.
790  *
791  * See CXL-3.0 8.2.9.3.3 Activate FW
792  */
793 static int cxl_mem_activate_fw(struct cxl_memdev_state *mds, int slot)
794 {
795 	struct cxl_mbox_activate_fw activate;
796 	struct cxl_mbox_cmd mbox_cmd;
797 
798 	if (slot == 0 || slot > mds->fw.num_slots)
799 		return -EINVAL;
800 
801 	mbox_cmd = (struct cxl_mbox_cmd) {
802 		.opcode = CXL_MBOX_OP_ACTIVATE_FW,
803 		.size_in = sizeof(activate),
804 		.payload_in = &activate,
805 	};
806 
807 	/* Only offline activation supported for now */
808 	activate.action = CXL_FW_ACTIVATE_OFFLINE;
809 	activate.slot = slot;
810 
811 	return cxl_internal_send_cmd(mds, &mbox_cmd);
812 }
813 
814 /**
815  * cxl_mem_abort_fw_xfer - Abort an in-progress FW transfer
816  * @mds: The device data for the operation
817  *
818  * Abort an in-progress firmware transfer for the device specified.
819  *
820  * Return: 0 if no error: or the result of the mailbox command.
821  *
822  * See CXL-3.0 8.2.9.3.2 Transfer FW
823  */
824 static int cxl_mem_abort_fw_xfer(struct cxl_memdev_state *mds)
825 {
826 	struct cxl_mbox_transfer_fw *transfer;
827 	struct cxl_mbox_cmd mbox_cmd;
828 	int rc;
829 
830 	transfer = kzalloc(struct_size(transfer, data, 0), GFP_KERNEL);
831 	if (!transfer)
832 		return -ENOMEM;
833 
834 	/* Set a 1s poll interval and a total wait time of 30s */
835 	mbox_cmd = (struct cxl_mbox_cmd) {
836 		.opcode = CXL_MBOX_OP_TRANSFER_FW,
837 		.size_in = sizeof(*transfer),
838 		.payload_in = transfer,
839 		.poll_interval_ms = 1000,
840 		.poll_count = 30,
841 	};
842 
843 	transfer->action = CXL_FW_TRANSFER_ACTION_ABORT;
844 
845 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
846 	kfree(transfer);
847 	return rc;
848 }
849 
850 static void cxl_fw_cleanup(struct fw_upload *fwl)
851 {
852 	struct cxl_memdev_state *mds = fwl->dd_handle;
853 
854 	mds->fw.next_slot = 0;
855 }
856 
857 static int cxl_fw_do_cancel(struct fw_upload *fwl)
858 {
859 	struct cxl_memdev_state *mds = fwl->dd_handle;
860 	struct cxl_dev_state *cxlds = &mds->cxlds;
861 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
862 	int rc;
863 
864 	rc = cxl_mem_abort_fw_xfer(mds);
865 	if (rc < 0)
866 		dev_err(&cxlmd->dev, "Error aborting FW transfer: %d\n", rc);
867 
868 	return FW_UPLOAD_ERR_CANCELED;
869 }
870 
871 static enum fw_upload_err cxl_fw_prepare(struct fw_upload *fwl, const u8 *data,
872 					 u32 size)
873 {
874 	struct cxl_memdev_state *mds = fwl->dd_handle;
875 	struct cxl_mbox_transfer_fw *transfer;
876 
877 	if (!size)
878 		return FW_UPLOAD_ERR_INVALID_SIZE;
879 
880 	mds->fw.oneshot = struct_size(transfer, data, size) <
881 			    mds->payload_size;
882 
883 	if (cxl_mem_get_fw_info(mds))
884 		return FW_UPLOAD_ERR_HW_ERROR;
885 
886 	/*
887 	 * So far no state has been changed, hence no other cleanup is
888 	 * necessary. Simply return the cancelled status.
889 	 */
890 	if (test_and_clear_bit(CXL_FW_CANCEL, mds->fw.state))
891 		return FW_UPLOAD_ERR_CANCELED;
892 
893 	return FW_UPLOAD_ERR_NONE;
894 }
895 
896 static enum fw_upload_err cxl_fw_write(struct fw_upload *fwl, const u8 *data,
897 				       u32 offset, u32 size, u32 *written)
898 {
899 	struct cxl_memdev_state *mds = fwl->dd_handle;
900 	struct cxl_dev_state *cxlds = &mds->cxlds;
901 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
902 	struct cxl_mbox_transfer_fw *transfer;
903 	struct cxl_mbox_cmd mbox_cmd;
904 	u32 cur_size, remaining;
905 	size_t size_in;
906 	int rc;
907 
908 	*written = 0;
909 
910 	/* Offset has to be aligned to 128B (CXL-3.0 8.2.9.3.2 Table 8-57) */
911 	if (!IS_ALIGNED(offset, CXL_FW_TRANSFER_ALIGNMENT)) {
912 		dev_err(&cxlmd->dev,
913 			"misaligned offset for FW transfer slice (%u)\n",
914 			offset);
915 		return FW_UPLOAD_ERR_RW_ERROR;
916 	}
917 
918 	/*
919 	 * Pick transfer size based on mds->payload_size @size must bw 128-byte
920 	 * aligned, ->payload_size is a power of 2 starting at 256 bytes, and
921 	 * sizeof(*transfer) is 128.  These constraints imply that @cur_size
922 	 * will always be 128b aligned.
923 	 */
924 	cur_size = min_t(size_t, size, mds->payload_size - sizeof(*transfer));
925 
926 	remaining = size - cur_size;
927 	size_in = struct_size(transfer, data, cur_size);
928 
929 	if (test_and_clear_bit(CXL_FW_CANCEL, mds->fw.state))
930 		return cxl_fw_do_cancel(fwl);
931 
932 	/*
933 	 * Slot numbers are 1-indexed
934 	 * cur_slot is the 0-indexed next_slot (i.e. 'cur_slot - 1 + 1')
935 	 * Check for rollover using modulo, and 1-index it by adding 1
936 	 */
937 	mds->fw.next_slot = (mds->fw.cur_slot % mds->fw.num_slots) + 1;
938 
939 	/* Do the transfer via mailbox cmd */
940 	transfer = kzalloc(size_in, GFP_KERNEL);
941 	if (!transfer)
942 		return FW_UPLOAD_ERR_RW_ERROR;
943 
944 	transfer->offset = cpu_to_le32(offset / CXL_FW_TRANSFER_ALIGNMENT);
945 	memcpy(transfer->data, data + offset, cur_size);
946 	if (mds->fw.oneshot) {
947 		transfer->action = CXL_FW_TRANSFER_ACTION_FULL;
948 		transfer->slot = mds->fw.next_slot;
949 	} else {
950 		if (offset == 0) {
951 			transfer->action = CXL_FW_TRANSFER_ACTION_INITIATE;
952 		} else if (remaining == 0) {
953 			transfer->action = CXL_FW_TRANSFER_ACTION_END;
954 			transfer->slot = mds->fw.next_slot;
955 		} else {
956 			transfer->action = CXL_FW_TRANSFER_ACTION_CONTINUE;
957 		}
958 	}
959 
960 	mbox_cmd = (struct cxl_mbox_cmd) {
961 		.opcode = CXL_MBOX_OP_TRANSFER_FW,
962 		.size_in = size_in,
963 		.payload_in = transfer,
964 		.poll_interval_ms = 1000,
965 		.poll_count = 30,
966 	};
967 
968 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
969 	if (rc < 0) {
970 		rc = FW_UPLOAD_ERR_RW_ERROR;
971 		goto out_free;
972 	}
973 
974 	*written = cur_size;
975 
976 	/* Activate FW if oneshot or if the last slice was written */
977 	if (mds->fw.oneshot || remaining == 0) {
978 		dev_dbg(&cxlmd->dev, "Activating firmware slot: %d\n",
979 			mds->fw.next_slot);
980 		rc = cxl_mem_activate_fw(mds, mds->fw.next_slot);
981 		if (rc < 0) {
982 			dev_err(&cxlmd->dev, "Error activating firmware: %d\n",
983 				rc);
984 			rc = FW_UPLOAD_ERR_HW_ERROR;
985 			goto out_free;
986 		}
987 	}
988 
989 	rc = FW_UPLOAD_ERR_NONE;
990 
991 out_free:
992 	kfree(transfer);
993 	return rc;
994 }
995 
996 static enum fw_upload_err cxl_fw_poll_complete(struct fw_upload *fwl)
997 {
998 	struct cxl_memdev_state *mds = fwl->dd_handle;
999 
1000 	/*
1001 	 * cxl_internal_send_cmd() handles background operations synchronously.
1002 	 * No need to wait for completions here - any errors would've been
1003 	 * reported and handled during the ->write() call(s).
1004 	 * Just check if a cancel request was received, and return success.
1005 	 */
1006 	if (test_and_clear_bit(CXL_FW_CANCEL, mds->fw.state))
1007 		return cxl_fw_do_cancel(fwl);
1008 
1009 	return FW_UPLOAD_ERR_NONE;
1010 }
1011 
1012 static void cxl_fw_cancel(struct fw_upload *fwl)
1013 {
1014 	struct cxl_memdev_state *mds = fwl->dd_handle;
1015 
1016 	set_bit(CXL_FW_CANCEL, mds->fw.state);
1017 }
1018 
1019 static const struct fw_upload_ops cxl_memdev_fw_ops = {
1020         .prepare = cxl_fw_prepare,
1021         .write = cxl_fw_write,
1022         .poll_complete = cxl_fw_poll_complete,
1023         .cancel = cxl_fw_cancel,
1024         .cleanup = cxl_fw_cleanup,
1025 };
1026 
1027 static void cxl_remove_fw_upload(void *fwl)
1028 {
1029 	firmware_upload_unregister(fwl);
1030 }
1031 
1032 int devm_cxl_setup_fw_upload(struct device *host, struct cxl_memdev_state *mds)
1033 {
1034 	struct cxl_dev_state *cxlds = &mds->cxlds;
1035 	struct device *dev = &cxlds->cxlmd->dev;
1036 	struct fw_upload *fwl;
1037 
1038 	if (!test_bit(CXL_MEM_COMMAND_ID_GET_FW_INFO, mds->enabled_cmds))
1039 		return 0;
1040 
1041 	fwl = firmware_upload_register(THIS_MODULE, dev, dev_name(dev),
1042 				       &cxl_memdev_fw_ops, mds);
1043 	if (IS_ERR(fwl))
1044 		return PTR_ERR(fwl);
1045 	return devm_add_action_or_reset(host, cxl_remove_fw_upload, fwl);
1046 }
1047 EXPORT_SYMBOL_NS_GPL(devm_cxl_setup_fw_upload, CXL);
1048 
1049 static const struct file_operations cxl_memdev_fops = {
1050 	.owner = THIS_MODULE,
1051 	.unlocked_ioctl = cxl_memdev_ioctl,
1052 	.open = cxl_memdev_open,
1053 	.release = cxl_memdev_release_file,
1054 	.compat_ioctl = compat_ptr_ioctl,
1055 	.llseek = noop_llseek,
1056 };
1057 
1058 struct cxl_memdev *devm_cxl_add_memdev(struct device *host,
1059 				       struct cxl_dev_state *cxlds)
1060 {
1061 	struct cxl_memdev *cxlmd;
1062 	struct device *dev;
1063 	struct cdev *cdev;
1064 	int rc;
1065 
1066 	cxlmd = cxl_memdev_alloc(cxlds, &cxl_memdev_fops);
1067 	if (IS_ERR(cxlmd))
1068 		return cxlmd;
1069 
1070 	dev = &cxlmd->dev;
1071 	rc = dev_set_name(dev, "mem%d", cxlmd->id);
1072 	if (rc)
1073 		goto err;
1074 
1075 	/*
1076 	 * Activate ioctl operations, no cxl_memdev_rwsem manipulation
1077 	 * needed as this is ordered with cdev_add() publishing the device.
1078 	 */
1079 	cxlmd->cxlds = cxlds;
1080 	cxlds->cxlmd = cxlmd;
1081 
1082 	cdev = &cxlmd->cdev;
1083 	rc = cdev_device_add(cdev, dev);
1084 	if (rc)
1085 		goto err;
1086 
1087 	rc = devm_add_action_or_reset(host, cxl_memdev_unregister, cxlmd);
1088 	if (rc)
1089 		return ERR_PTR(rc);
1090 	return cxlmd;
1091 
1092 err:
1093 	/*
1094 	 * The cdev was briefly live, shutdown any ioctl operations that
1095 	 * saw that state.
1096 	 */
1097 	cxl_memdev_shutdown(dev);
1098 	put_device(dev);
1099 	return ERR_PTR(rc);
1100 }
1101 EXPORT_SYMBOL_NS_GPL(devm_cxl_add_memdev, CXL);
1102 
1103 static void sanitize_teardown_notifier(void *data)
1104 {
1105 	struct cxl_memdev_state *mds = data;
1106 	struct kernfs_node *state;
1107 
1108 	/*
1109 	 * Prevent new irq triggered invocations of the workqueue and
1110 	 * flush inflight invocations.
1111 	 */
1112 	mutex_lock(&mds->mbox_mutex);
1113 	state = mds->security.sanitize_node;
1114 	mds->security.sanitize_node = NULL;
1115 	mutex_unlock(&mds->mbox_mutex);
1116 
1117 	cancel_delayed_work_sync(&mds->security.poll_dwork);
1118 	sysfs_put(state);
1119 }
1120 
1121 int devm_cxl_sanitize_setup_notifier(struct device *host,
1122 				     struct cxl_memdev *cxlmd)
1123 {
1124 	struct cxl_dev_state *cxlds = cxlmd->cxlds;
1125 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
1126 	struct kernfs_node *sec;
1127 
1128 	if (!test_bit(CXL_SEC_ENABLED_SANITIZE, mds->security.enabled_cmds))
1129 		return 0;
1130 
1131 	/*
1132 	 * Note, the expectation is that @cxlmd would have failed to be
1133 	 * created if these sysfs_get_dirent calls fail.
1134 	 */
1135 	sec = sysfs_get_dirent(cxlmd->dev.kobj.sd, "security");
1136 	if (!sec)
1137 		return -ENOENT;
1138 	mds->security.sanitize_node = sysfs_get_dirent(sec, "state");
1139 	sysfs_put(sec);
1140 	if (!mds->security.sanitize_node)
1141 		return -ENOENT;
1142 
1143 	return devm_add_action_or_reset(host, sanitize_teardown_notifier, mds);
1144 }
1145 EXPORT_SYMBOL_NS_GPL(devm_cxl_sanitize_setup_notifier, CXL);
1146 
1147 __init int cxl_memdev_init(void)
1148 {
1149 	dev_t devt;
1150 	int rc;
1151 
1152 	rc = alloc_chrdev_region(&devt, 0, CXL_MEM_MAX_DEVS, "cxl");
1153 	if (rc)
1154 		return rc;
1155 
1156 	cxl_mem_major = MAJOR(devt);
1157 
1158 	return 0;
1159 }
1160 
1161 void cxl_memdev_exit(void)
1162 {
1163 	unregister_chrdev_region(MKDEV(cxl_mem_major, 0), CXL_MEM_MAX_DEVS);
1164 }
1165