xref: /linux/drivers/cxl/core/pci.c (revision c4101e55974cc7d835fbd2d8e01553a3f61e9e75)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2021 Intel Corporation. All rights reserved. */
3 #include <linux/io-64-nonatomic-lo-hi.h>
4 #include <linux/device.h>
5 #include <linux/delay.h>
6 #include <linux/pci.h>
7 #include <linux/pci-doe.h>
8 #include <linux/aer.h>
9 #include <cxlpci.h>
10 #include <cxlmem.h>
11 #include <cxl.h>
12 #include "core.h"
13 #include "trace.h"
14 
15 /**
16  * DOC: cxl core pci
17  *
18  * Compute Express Link protocols are layered on top of PCIe. CXL core provides
19  * a set of helpers for CXL interactions which occur via PCIe.
20  */
21 
22 static unsigned short media_ready_timeout = 60;
23 module_param(media_ready_timeout, ushort, 0644);
24 MODULE_PARM_DESC(media_ready_timeout, "seconds to wait for media ready");
25 
26 struct cxl_walk_context {
27 	struct pci_bus *bus;
28 	struct cxl_port *port;
29 	int type;
30 	int error;
31 	int count;
32 };
33 
34 static int match_add_dports(struct pci_dev *pdev, void *data)
35 {
36 	struct cxl_walk_context *ctx = data;
37 	struct cxl_port *port = ctx->port;
38 	int type = pci_pcie_type(pdev);
39 	struct cxl_register_map map;
40 	struct cxl_dport *dport;
41 	u32 lnkcap, port_num;
42 	int rc;
43 
44 	if (pdev->bus != ctx->bus)
45 		return 0;
46 	if (!pci_is_pcie(pdev))
47 		return 0;
48 	if (type != ctx->type)
49 		return 0;
50 	if (pci_read_config_dword(pdev, pci_pcie_cap(pdev) + PCI_EXP_LNKCAP,
51 				  &lnkcap))
52 		return 0;
53 
54 	rc = cxl_find_regblock(pdev, CXL_REGLOC_RBI_COMPONENT, &map);
55 	if (rc)
56 		dev_dbg(&port->dev, "failed to find component registers\n");
57 
58 	port_num = FIELD_GET(PCI_EXP_LNKCAP_PN, lnkcap);
59 	dport = devm_cxl_add_dport(port, &pdev->dev, port_num, map.resource);
60 	if (IS_ERR(dport)) {
61 		ctx->error = PTR_ERR(dport);
62 		return PTR_ERR(dport);
63 	}
64 	ctx->count++;
65 
66 	return 0;
67 }
68 
69 /**
70  * devm_cxl_port_enumerate_dports - enumerate downstream ports of the upstream port
71  * @port: cxl_port whose ->uport_dev is the upstream of dports to be enumerated
72  *
73  * Returns a positive number of dports enumerated or a negative error
74  * code.
75  */
76 int devm_cxl_port_enumerate_dports(struct cxl_port *port)
77 {
78 	struct pci_bus *bus = cxl_port_to_pci_bus(port);
79 	struct cxl_walk_context ctx;
80 	int type;
81 
82 	if (!bus)
83 		return -ENXIO;
84 
85 	if (pci_is_root_bus(bus))
86 		type = PCI_EXP_TYPE_ROOT_PORT;
87 	else
88 		type = PCI_EXP_TYPE_DOWNSTREAM;
89 
90 	ctx = (struct cxl_walk_context) {
91 		.port = port,
92 		.bus = bus,
93 		.type = type,
94 	};
95 	pci_walk_bus(bus, match_add_dports, &ctx);
96 
97 	if (ctx.count == 0)
98 		return -ENODEV;
99 	if (ctx.error)
100 		return ctx.error;
101 	return ctx.count;
102 }
103 EXPORT_SYMBOL_NS_GPL(devm_cxl_port_enumerate_dports, CXL);
104 
105 static int cxl_dvsec_mem_range_valid(struct cxl_dev_state *cxlds, int id)
106 {
107 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
108 	int d = cxlds->cxl_dvsec;
109 	bool valid = false;
110 	int rc, i;
111 	u32 temp;
112 
113 	if (id > CXL_DVSEC_RANGE_MAX)
114 		return -EINVAL;
115 
116 	/* Check MEM INFO VALID bit first, give up after 1s */
117 	i = 1;
118 	do {
119 		rc = pci_read_config_dword(pdev,
120 					   d + CXL_DVSEC_RANGE_SIZE_LOW(id),
121 					   &temp);
122 		if (rc)
123 			return rc;
124 
125 		valid = FIELD_GET(CXL_DVSEC_MEM_INFO_VALID, temp);
126 		if (valid)
127 			break;
128 		msleep(1000);
129 	} while (i--);
130 
131 	if (!valid) {
132 		dev_err(&pdev->dev,
133 			"Timeout awaiting memory range %d valid after 1s.\n",
134 			id);
135 		return -ETIMEDOUT;
136 	}
137 
138 	return 0;
139 }
140 
141 static int cxl_dvsec_mem_range_active(struct cxl_dev_state *cxlds, int id)
142 {
143 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
144 	int d = cxlds->cxl_dvsec;
145 	bool active = false;
146 	int rc, i;
147 	u32 temp;
148 
149 	if (id > CXL_DVSEC_RANGE_MAX)
150 		return -EINVAL;
151 
152 	/* Check MEM ACTIVE bit, up to 60s timeout by default */
153 	for (i = media_ready_timeout; i; i--) {
154 		rc = pci_read_config_dword(
155 			pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(id), &temp);
156 		if (rc)
157 			return rc;
158 
159 		active = FIELD_GET(CXL_DVSEC_MEM_ACTIVE, temp);
160 		if (active)
161 			break;
162 		msleep(1000);
163 	}
164 
165 	if (!active) {
166 		dev_err(&pdev->dev,
167 			"timeout awaiting memory active after %d seconds\n",
168 			media_ready_timeout);
169 		return -ETIMEDOUT;
170 	}
171 
172 	return 0;
173 }
174 
175 /*
176  * Wait up to @media_ready_timeout for the device to report memory
177  * active.
178  */
179 int cxl_await_media_ready(struct cxl_dev_state *cxlds)
180 {
181 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
182 	int d = cxlds->cxl_dvsec;
183 	int rc, i, hdm_count;
184 	u64 md_status;
185 	u16 cap;
186 
187 	rc = pci_read_config_word(pdev,
188 				  d + CXL_DVSEC_CAP_OFFSET, &cap);
189 	if (rc)
190 		return rc;
191 
192 	hdm_count = FIELD_GET(CXL_DVSEC_HDM_COUNT_MASK, cap);
193 	for (i = 0; i < hdm_count; i++) {
194 		rc = cxl_dvsec_mem_range_valid(cxlds, i);
195 		if (rc)
196 			return rc;
197 	}
198 
199 	for (i = 0; i < hdm_count; i++) {
200 		rc = cxl_dvsec_mem_range_active(cxlds, i);
201 		if (rc)
202 			return rc;
203 	}
204 
205 	md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
206 	if (!CXLMDEV_READY(md_status))
207 		return -EIO;
208 
209 	return 0;
210 }
211 EXPORT_SYMBOL_NS_GPL(cxl_await_media_ready, CXL);
212 
213 static int wait_for_valid(struct pci_dev *pdev, int d)
214 {
215 	u32 val;
216 	int rc;
217 
218 	/*
219 	 * Memory_Info_Valid: When set, indicates that the CXL Range 1 Size high
220 	 * and Size Low registers are valid. Must be set within 1 second of
221 	 * deassertion of reset to CXL device. Likely it is already set by the
222 	 * time this runs, but otherwise give a 1.5 second timeout in case of
223 	 * clock skew.
224 	 */
225 	rc = pci_read_config_dword(pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(0), &val);
226 	if (rc)
227 		return rc;
228 
229 	if (val & CXL_DVSEC_MEM_INFO_VALID)
230 		return 0;
231 
232 	msleep(1500);
233 
234 	rc = pci_read_config_dword(pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(0), &val);
235 	if (rc)
236 		return rc;
237 
238 	if (val & CXL_DVSEC_MEM_INFO_VALID)
239 		return 0;
240 
241 	return -ETIMEDOUT;
242 }
243 
244 static int cxl_set_mem_enable(struct cxl_dev_state *cxlds, u16 val)
245 {
246 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
247 	int d = cxlds->cxl_dvsec;
248 	u16 ctrl;
249 	int rc;
250 
251 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, &ctrl);
252 	if (rc < 0)
253 		return rc;
254 
255 	if ((ctrl & CXL_DVSEC_MEM_ENABLE) == val)
256 		return 1;
257 	ctrl &= ~CXL_DVSEC_MEM_ENABLE;
258 	ctrl |= val;
259 
260 	rc = pci_write_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, ctrl);
261 	if (rc < 0)
262 		return rc;
263 
264 	return 0;
265 }
266 
267 static void clear_mem_enable(void *cxlds)
268 {
269 	cxl_set_mem_enable(cxlds, 0);
270 }
271 
272 static int devm_cxl_enable_mem(struct device *host, struct cxl_dev_state *cxlds)
273 {
274 	int rc;
275 
276 	rc = cxl_set_mem_enable(cxlds, CXL_DVSEC_MEM_ENABLE);
277 	if (rc < 0)
278 		return rc;
279 	if (rc > 0)
280 		return 0;
281 	return devm_add_action_or_reset(host, clear_mem_enable, cxlds);
282 }
283 
284 /* require dvsec ranges to be covered by a locked platform window */
285 static int dvsec_range_allowed(struct device *dev, void *arg)
286 {
287 	struct range *dev_range = arg;
288 	struct cxl_decoder *cxld;
289 
290 	if (!is_root_decoder(dev))
291 		return 0;
292 
293 	cxld = to_cxl_decoder(dev);
294 
295 	if (!(cxld->flags & CXL_DECODER_F_RAM))
296 		return 0;
297 
298 	return range_contains(&cxld->hpa_range, dev_range);
299 }
300 
301 static void disable_hdm(void *_cxlhdm)
302 {
303 	u32 global_ctrl;
304 	struct cxl_hdm *cxlhdm = _cxlhdm;
305 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
306 
307 	global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
308 	writel(global_ctrl & ~CXL_HDM_DECODER_ENABLE,
309 	       hdm + CXL_HDM_DECODER_CTRL_OFFSET);
310 }
311 
312 static int devm_cxl_enable_hdm(struct device *host, struct cxl_hdm *cxlhdm)
313 {
314 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
315 	u32 global_ctrl;
316 
317 	global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
318 	writel(global_ctrl | CXL_HDM_DECODER_ENABLE,
319 	       hdm + CXL_HDM_DECODER_CTRL_OFFSET);
320 
321 	return devm_add_action_or_reset(host, disable_hdm, cxlhdm);
322 }
323 
324 int cxl_dvsec_rr_decode(struct device *dev, int d,
325 			struct cxl_endpoint_dvsec_info *info)
326 {
327 	struct pci_dev *pdev = to_pci_dev(dev);
328 	int hdm_count, rc, i, ranges = 0;
329 	u16 cap, ctrl;
330 
331 	if (!d) {
332 		dev_dbg(dev, "No DVSEC Capability\n");
333 		return -ENXIO;
334 	}
335 
336 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CAP_OFFSET, &cap);
337 	if (rc)
338 		return rc;
339 
340 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, &ctrl);
341 	if (rc)
342 		return rc;
343 
344 	if (!(cap & CXL_DVSEC_MEM_CAPABLE)) {
345 		dev_dbg(dev, "Not MEM Capable\n");
346 		return -ENXIO;
347 	}
348 
349 	/*
350 	 * It is not allowed by spec for MEM.capable to be set and have 0 legacy
351 	 * HDM decoders (values > 2 are also undefined as of CXL 2.0). As this
352 	 * driver is for a spec defined class code which must be CXL.mem
353 	 * capable, there is no point in continuing to enable CXL.mem.
354 	 */
355 	hdm_count = FIELD_GET(CXL_DVSEC_HDM_COUNT_MASK, cap);
356 	if (!hdm_count || hdm_count > 2)
357 		return -EINVAL;
358 
359 	rc = wait_for_valid(pdev, d);
360 	if (rc) {
361 		dev_dbg(dev, "Failure awaiting MEM_INFO_VALID (%d)\n", rc);
362 		return rc;
363 	}
364 
365 	/*
366 	 * The current DVSEC values are moot if the memory capability is
367 	 * disabled, and they will remain moot after the HDM Decoder
368 	 * capability is enabled.
369 	 */
370 	info->mem_enabled = FIELD_GET(CXL_DVSEC_MEM_ENABLE, ctrl);
371 	if (!info->mem_enabled)
372 		return 0;
373 
374 	for (i = 0; i < hdm_count; i++) {
375 		u64 base, size;
376 		u32 temp;
377 
378 		rc = pci_read_config_dword(
379 			pdev, d + CXL_DVSEC_RANGE_SIZE_HIGH(i), &temp);
380 		if (rc)
381 			return rc;
382 
383 		size = (u64)temp << 32;
384 
385 		rc = pci_read_config_dword(
386 			pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(i), &temp);
387 		if (rc)
388 			return rc;
389 
390 		size |= temp & CXL_DVSEC_MEM_SIZE_LOW_MASK;
391 		if (!size) {
392 			info->dvsec_range[i] = (struct range) {
393 				.start = 0,
394 				.end = CXL_RESOURCE_NONE,
395 			};
396 			continue;
397 		}
398 
399 		rc = pci_read_config_dword(
400 			pdev, d + CXL_DVSEC_RANGE_BASE_HIGH(i), &temp);
401 		if (rc)
402 			return rc;
403 
404 		base = (u64)temp << 32;
405 
406 		rc = pci_read_config_dword(
407 			pdev, d + CXL_DVSEC_RANGE_BASE_LOW(i), &temp);
408 		if (rc)
409 			return rc;
410 
411 		base |= temp & CXL_DVSEC_MEM_BASE_LOW_MASK;
412 
413 		info->dvsec_range[i] = (struct range) {
414 			.start = base,
415 			.end = base + size - 1
416 		};
417 
418 		ranges++;
419 	}
420 
421 	info->ranges = ranges;
422 
423 	return 0;
424 }
425 EXPORT_SYMBOL_NS_GPL(cxl_dvsec_rr_decode, CXL);
426 
427 /**
428  * cxl_hdm_decode_init() - Setup HDM decoding for the endpoint
429  * @cxlds: Device state
430  * @cxlhdm: Mapped HDM decoder Capability
431  * @info: Cached DVSEC range registers info
432  *
433  * Try to enable the endpoint's HDM Decoder Capability
434  */
435 int cxl_hdm_decode_init(struct cxl_dev_state *cxlds, struct cxl_hdm *cxlhdm,
436 			struct cxl_endpoint_dvsec_info *info)
437 {
438 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
439 	struct cxl_port *port = cxlhdm->port;
440 	struct device *dev = cxlds->dev;
441 	struct cxl_port *root;
442 	int i, rc, allowed;
443 	u32 global_ctrl = 0;
444 
445 	if (hdm)
446 		global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
447 
448 	/*
449 	 * If the HDM Decoder Capability is already enabled then assume
450 	 * that some other agent like platform firmware set it up.
451 	 */
452 	if (global_ctrl & CXL_HDM_DECODER_ENABLE || (!hdm && info->mem_enabled))
453 		return devm_cxl_enable_mem(&port->dev, cxlds);
454 	else if (!hdm)
455 		return -ENODEV;
456 
457 	root = to_cxl_port(port->dev.parent);
458 	while (!is_cxl_root(root) && is_cxl_port(root->dev.parent))
459 		root = to_cxl_port(root->dev.parent);
460 	if (!is_cxl_root(root)) {
461 		dev_err(dev, "Failed to acquire root port for HDM enable\n");
462 		return -ENODEV;
463 	}
464 
465 	for (i = 0, allowed = 0; info->mem_enabled && i < info->ranges; i++) {
466 		struct device *cxld_dev;
467 
468 		cxld_dev = device_find_child(&root->dev, &info->dvsec_range[i],
469 					     dvsec_range_allowed);
470 		if (!cxld_dev) {
471 			dev_dbg(dev, "DVSEC Range%d denied by platform\n", i);
472 			continue;
473 		}
474 		dev_dbg(dev, "DVSEC Range%d allowed by platform\n", i);
475 		put_device(cxld_dev);
476 		allowed++;
477 	}
478 
479 	if (!allowed) {
480 		cxl_set_mem_enable(cxlds, 0);
481 		info->mem_enabled = 0;
482 	}
483 
484 	/*
485 	 * Per CXL 2.0 Section 8.1.3.8.3 and 8.1.3.8.4 DVSEC CXL Range 1 Base
486 	 * [High,Low] when HDM operation is enabled the range register values
487 	 * are ignored by the device, but the spec also recommends matching the
488 	 * DVSEC Range 1,2 to HDM Decoder Range 0,1. So, non-zero info->ranges
489 	 * are expected even though Linux does not require or maintain that
490 	 * match. If at least one DVSEC range is enabled and allowed, skip HDM
491 	 * Decoder Capability Enable.
492 	 */
493 	if (info->mem_enabled)
494 		return 0;
495 
496 	rc = devm_cxl_enable_hdm(&port->dev, cxlhdm);
497 	if (rc)
498 		return rc;
499 
500 	return devm_cxl_enable_mem(&port->dev, cxlds);
501 }
502 EXPORT_SYMBOL_NS_GPL(cxl_hdm_decode_init, CXL);
503 
504 #define CXL_DOE_TABLE_ACCESS_REQ_CODE		0x000000ff
505 #define   CXL_DOE_TABLE_ACCESS_REQ_CODE_READ	0
506 #define CXL_DOE_TABLE_ACCESS_TABLE_TYPE		0x0000ff00
507 #define   CXL_DOE_TABLE_ACCESS_TABLE_TYPE_CDATA	0
508 #define CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE	0xffff0000
509 #define CXL_DOE_TABLE_ACCESS_LAST_ENTRY		0xffff
510 #define CXL_DOE_PROTOCOL_TABLE_ACCESS 2
511 
512 #define CDAT_DOE_REQ(entry_handle) cpu_to_le32				\
513 	(FIELD_PREP(CXL_DOE_TABLE_ACCESS_REQ_CODE,			\
514 		    CXL_DOE_TABLE_ACCESS_REQ_CODE_READ) |		\
515 	 FIELD_PREP(CXL_DOE_TABLE_ACCESS_TABLE_TYPE,			\
516 		    CXL_DOE_TABLE_ACCESS_TABLE_TYPE_CDATA) |		\
517 	 FIELD_PREP(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE, (entry_handle)))
518 
519 static int cxl_cdat_get_length(struct device *dev,
520 			       struct pci_doe_mb *cdat_doe,
521 			       size_t *length)
522 {
523 	__le32 request = CDAT_DOE_REQ(0);
524 	__le32 response[2];
525 	int rc;
526 
527 	rc = pci_doe(cdat_doe, PCI_DVSEC_VENDOR_ID_CXL,
528 		     CXL_DOE_PROTOCOL_TABLE_ACCESS,
529 		     &request, sizeof(request),
530 		     &response, sizeof(response));
531 	if (rc < 0) {
532 		dev_err(dev, "DOE failed: %d", rc);
533 		return rc;
534 	}
535 	if (rc < sizeof(response))
536 		return -EIO;
537 
538 	*length = le32_to_cpu(response[1]);
539 	dev_dbg(dev, "CDAT length %zu\n", *length);
540 
541 	return 0;
542 }
543 
544 static int cxl_cdat_read_table(struct device *dev,
545 			       struct pci_doe_mb *cdat_doe,
546 			       void *cdat_table, size_t *cdat_length)
547 {
548 	size_t length = *cdat_length + sizeof(__le32);
549 	__le32 *data = cdat_table;
550 	int entry_handle = 0;
551 	__le32 saved_dw = 0;
552 
553 	do {
554 		__le32 request = CDAT_DOE_REQ(entry_handle);
555 		struct cdat_entry_header *entry;
556 		size_t entry_dw;
557 		int rc;
558 
559 		rc = pci_doe(cdat_doe, PCI_DVSEC_VENDOR_ID_CXL,
560 			     CXL_DOE_PROTOCOL_TABLE_ACCESS,
561 			     &request, sizeof(request),
562 			     data, length);
563 		if (rc < 0) {
564 			dev_err(dev, "DOE failed: %d", rc);
565 			return rc;
566 		}
567 
568 		/* 1 DW Table Access Response Header + CDAT entry */
569 		entry = (struct cdat_entry_header *)(data + 1);
570 		if ((entry_handle == 0 &&
571 		     rc != sizeof(__le32) + sizeof(struct cdat_header)) ||
572 		    (entry_handle > 0 &&
573 		     (rc < sizeof(__le32) + sizeof(*entry) ||
574 		      rc != sizeof(__le32) + le16_to_cpu(entry->length))))
575 			return -EIO;
576 
577 		/* Get the CXL table access header entry handle */
578 		entry_handle = FIELD_GET(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE,
579 					 le32_to_cpu(data[0]));
580 		entry_dw = rc / sizeof(__le32);
581 		/* Skip Header */
582 		entry_dw -= 1;
583 		/*
584 		 * Table Access Response Header overwrote the last DW of
585 		 * previous entry, so restore that DW
586 		 */
587 		*data = saved_dw;
588 		length -= entry_dw * sizeof(__le32);
589 		data += entry_dw;
590 		saved_dw = *data;
591 	} while (entry_handle != CXL_DOE_TABLE_ACCESS_LAST_ENTRY);
592 
593 	/* Length in CDAT header may exceed concatenation of CDAT entries */
594 	*cdat_length -= length - sizeof(__le32);
595 
596 	return 0;
597 }
598 
599 static unsigned char cdat_checksum(void *buf, size_t size)
600 {
601 	unsigned char sum, *data = buf;
602 	size_t i;
603 
604 	for (sum = 0, i = 0; i < size; i++)
605 		sum += data[i];
606 	return sum;
607 }
608 
609 /**
610  * read_cdat_data - Read the CDAT data on this port
611  * @port: Port to read data from
612  *
613  * This call will sleep waiting for responses from the DOE mailbox.
614  */
615 void read_cdat_data(struct cxl_port *port)
616 {
617 	struct device *uport = port->uport_dev;
618 	struct device *dev = &port->dev;
619 	struct pci_doe_mb *cdat_doe;
620 	struct pci_dev *pdev = NULL;
621 	struct cxl_memdev *cxlmd;
622 	size_t cdat_length;
623 	void *cdat_table, *cdat_buf;
624 	int rc;
625 
626 	if (is_cxl_memdev(uport)) {
627 		struct device *host;
628 
629 		cxlmd = to_cxl_memdev(uport);
630 		host = cxlmd->dev.parent;
631 		if (dev_is_pci(host))
632 			pdev = to_pci_dev(host);
633 	} else if (dev_is_pci(uport)) {
634 		pdev = to_pci_dev(uport);
635 	}
636 
637 	if (!pdev)
638 		return;
639 
640 	cdat_doe = pci_find_doe_mailbox(pdev, PCI_DVSEC_VENDOR_ID_CXL,
641 					CXL_DOE_PROTOCOL_TABLE_ACCESS);
642 	if (!cdat_doe) {
643 		dev_dbg(dev, "No CDAT mailbox\n");
644 		return;
645 	}
646 
647 	port->cdat_available = true;
648 
649 	if (cxl_cdat_get_length(dev, cdat_doe, &cdat_length)) {
650 		dev_dbg(dev, "No CDAT length\n");
651 		return;
652 	}
653 
654 	cdat_buf = devm_kzalloc(dev, cdat_length + sizeof(__le32), GFP_KERNEL);
655 	if (!cdat_buf)
656 		return;
657 
658 	rc = cxl_cdat_read_table(dev, cdat_doe, cdat_buf, &cdat_length);
659 	if (rc)
660 		goto err;
661 
662 	cdat_table = cdat_buf + sizeof(__le32);
663 	if (cdat_checksum(cdat_table, cdat_length))
664 		goto err;
665 
666 	port->cdat.table = cdat_table;
667 	port->cdat.length = cdat_length;
668 	return;
669 
670 err:
671 	/* Don't leave table data allocated on error */
672 	devm_kfree(dev, cdat_buf);
673 	dev_err(dev, "Failed to read/validate CDAT.\n");
674 }
675 EXPORT_SYMBOL_NS_GPL(read_cdat_data, CXL);
676 
677 static void __cxl_handle_cor_ras(struct cxl_dev_state *cxlds,
678 				 void __iomem *ras_base)
679 {
680 	void __iomem *addr;
681 	u32 status;
682 
683 	if (!ras_base)
684 		return;
685 
686 	addr = ras_base + CXL_RAS_CORRECTABLE_STATUS_OFFSET;
687 	status = readl(addr);
688 	if (status & CXL_RAS_CORRECTABLE_STATUS_MASK) {
689 		writel(status & CXL_RAS_CORRECTABLE_STATUS_MASK, addr);
690 		trace_cxl_aer_correctable_error(cxlds->cxlmd, status);
691 	}
692 }
693 
694 static void cxl_handle_endpoint_cor_ras(struct cxl_dev_state *cxlds)
695 {
696 	return __cxl_handle_cor_ras(cxlds, cxlds->regs.ras);
697 }
698 
699 /* CXL spec rev3.0 8.2.4.16.1 */
700 static void header_log_copy(void __iomem *ras_base, u32 *log)
701 {
702 	void __iomem *addr;
703 	u32 *log_addr;
704 	int i, log_u32_size = CXL_HEADERLOG_SIZE / sizeof(u32);
705 
706 	addr = ras_base + CXL_RAS_HEADER_LOG_OFFSET;
707 	log_addr = log;
708 
709 	for (i = 0; i < log_u32_size; i++) {
710 		*log_addr = readl(addr);
711 		log_addr++;
712 		addr += sizeof(u32);
713 	}
714 }
715 
716 /*
717  * Log the state of the RAS status registers and prepare them to log the
718  * next error status. Return 1 if reset needed.
719  */
720 static bool __cxl_handle_ras(struct cxl_dev_state *cxlds,
721 				  void __iomem *ras_base)
722 {
723 	u32 hl[CXL_HEADERLOG_SIZE_U32];
724 	void __iomem *addr;
725 	u32 status;
726 	u32 fe;
727 
728 	if (!ras_base)
729 		return false;
730 
731 	addr = ras_base + CXL_RAS_UNCORRECTABLE_STATUS_OFFSET;
732 	status = readl(addr);
733 	if (!(status & CXL_RAS_UNCORRECTABLE_STATUS_MASK))
734 		return false;
735 
736 	/* If multiple errors, log header points to first error from ctrl reg */
737 	if (hweight32(status) > 1) {
738 		void __iomem *rcc_addr =
739 			ras_base + CXL_RAS_CAP_CONTROL_OFFSET;
740 
741 		fe = BIT(FIELD_GET(CXL_RAS_CAP_CONTROL_FE_MASK,
742 				   readl(rcc_addr)));
743 	} else {
744 		fe = status;
745 	}
746 
747 	header_log_copy(ras_base, hl);
748 	trace_cxl_aer_uncorrectable_error(cxlds->cxlmd, status, fe, hl);
749 	writel(status & CXL_RAS_UNCORRECTABLE_STATUS_MASK, addr);
750 
751 	return true;
752 }
753 
754 static bool cxl_handle_endpoint_ras(struct cxl_dev_state *cxlds)
755 {
756 	return __cxl_handle_ras(cxlds, cxlds->regs.ras);
757 }
758 
759 #ifdef CONFIG_PCIEAER_CXL
760 
761 static void cxl_dport_map_rch_aer(struct cxl_dport *dport)
762 {
763 	struct cxl_rcrb_info *ri = &dport->rcrb;
764 	void __iomem *dport_aer = NULL;
765 	resource_size_t aer_phys;
766 	struct device *host;
767 
768 	if (dport->rch && ri->aer_cap) {
769 		host = dport->reg_map.host;
770 		aer_phys = ri->aer_cap + ri->base;
771 		dport_aer = devm_cxl_iomap_block(host, aer_phys,
772 				sizeof(struct aer_capability_regs));
773 	}
774 
775 	dport->regs.dport_aer = dport_aer;
776 }
777 
778 static void cxl_dport_map_regs(struct cxl_dport *dport)
779 {
780 	struct cxl_register_map *map = &dport->reg_map;
781 	struct device *dev = dport->dport_dev;
782 
783 	if (!map->component_map.ras.valid)
784 		dev_dbg(dev, "RAS registers not found\n");
785 	else if (cxl_map_component_regs(map, &dport->regs.component,
786 					BIT(CXL_CM_CAP_CAP_ID_RAS)))
787 		dev_dbg(dev, "Failed to map RAS capability.\n");
788 
789 	if (dport->rch)
790 		cxl_dport_map_rch_aer(dport);
791 }
792 
793 static void cxl_disable_rch_root_ints(struct cxl_dport *dport)
794 {
795 	void __iomem *aer_base = dport->regs.dport_aer;
796 	struct pci_host_bridge *bridge;
797 	u32 aer_cmd_mask, aer_cmd;
798 
799 	if (!aer_base)
800 		return;
801 
802 	bridge = to_pci_host_bridge(dport->dport_dev);
803 
804 	/*
805 	 * Disable RCH root port command interrupts.
806 	 * CXL 3.0 12.2.1.1 - RCH Downstream Port-detected Errors
807 	 *
808 	 * This sequence may not be necessary. CXL spec states disabling
809 	 * the root cmd register's interrupts is required. But, PCI spec
810 	 * shows these are disabled by default on reset.
811 	 */
812 	if (bridge->native_aer) {
813 		aer_cmd_mask = (PCI_ERR_ROOT_CMD_COR_EN |
814 				PCI_ERR_ROOT_CMD_NONFATAL_EN |
815 				PCI_ERR_ROOT_CMD_FATAL_EN);
816 		aer_cmd = readl(aer_base + PCI_ERR_ROOT_COMMAND);
817 		aer_cmd &= ~aer_cmd_mask;
818 		writel(aer_cmd, aer_base + PCI_ERR_ROOT_COMMAND);
819 	}
820 }
821 
822 void cxl_setup_parent_dport(struct device *host, struct cxl_dport *dport)
823 {
824 	struct device *dport_dev = dport->dport_dev;
825 	struct pci_host_bridge *host_bridge;
826 
827 	host_bridge = to_pci_host_bridge(dport_dev);
828 	if (host_bridge->native_aer)
829 		dport->rcrb.aer_cap = cxl_rcrb_to_aer(dport_dev, dport->rcrb.base);
830 
831 	dport->reg_map.host = host;
832 	cxl_dport_map_regs(dport);
833 
834 	if (dport->rch)
835 		cxl_disable_rch_root_ints(dport);
836 }
837 EXPORT_SYMBOL_NS_GPL(cxl_setup_parent_dport, CXL);
838 
839 static void cxl_handle_rdport_cor_ras(struct cxl_dev_state *cxlds,
840 					  struct cxl_dport *dport)
841 {
842 	return __cxl_handle_cor_ras(cxlds, dport->regs.ras);
843 }
844 
845 static bool cxl_handle_rdport_ras(struct cxl_dev_state *cxlds,
846 				       struct cxl_dport *dport)
847 {
848 	return __cxl_handle_ras(cxlds, dport->regs.ras);
849 }
850 
851 /*
852  * Copy the AER capability registers using 32 bit read accesses.
853  * This is necessary because RCRB AER capability is MMIO mapped. Clear the
854  * status after copying.
855  *
856  * @aer_base: base address of AER capability block in RCRB
857  * @aer_regs: destination for copying AER capability
858  */
859 static bool cxl_rch_get_aer_info(void __iomem *aer_base,
860 				 struct aer_capability_regs *aer_regs)
861 {
862 	int read_cnt = sizeof(struct aer_capability_regs) / sizeof(u32);
863 	u32 *aer_regs_buf = (u32 *)aer_regs;
864 	int n;
865 
866 	if (!aer_base)
867 		return false;
868 
869 	/* Use readl() to guarantee 32-bit accesses */
870 	for (n = 0; n < read_cnt; n++)
871 		aer_regs_buf[n] = readl(aer_base + n * sizeof(u32));
872 
873 	writel(aer_regs->uncor_status, aer_base + PCI_ERR_UNCOR_STATUS);
874 	writel(aer_regs->cor_status, aer_base + PCI_ERR_COR_STATUS);
875 
876 	return true;
877 }
878 
879 /* Get AER severity. Return false if there is no error. */
880 static bool cxl_rch_get_aer_severity(struct aer_capability_regs *aer_regs,
881 				     int *severity)
882 {
883 	if (aer_regs->uncor_status & ~aer_regs->uncor_mask) {
884 		if (aer_regs->uncor_status & PCI_ERR_ROOT_FATAL_RCV)
885 			*severity = AER_FATAL;
886 		else
887 			*severity = AER_NONFATAL;
888 		return true;
889 	}
890 
891 	if (aer_regs->cor_status & ~aer_regs->cor_mask) {
892 		*severity = AER_CORRECTABLE;
893 		return true;
894 	}
895 
896 	return false;
897 }
898 
899 static void cxl_handle_rdport_errors(struct cxl_dev_state *cxlds)
900 {
901 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
902 	struct aer_capability_regs aer_regs;
903 	struct cxl_dport *dport;
904 	struct cxl_port *port;
905 	int severity;
906 
907 	port = cxl_pci_find_port(pdev, &dport);
908 	if (!port)
909 		return;
910 
911 	put_device(&port->dev);
912 
913 	if (!cxl_rch_get_aer_info(dport->regs.dport_aer, &aer_regs))
914 		return;
915 
916 	if (!cxl_rch_get_aer_severity(&aer_regs, &severity))
917 		return;
918 
919 	pci_print_aer(pdev, severity, &aer_regs);
920 
921 	if (severity == AER_CORRECTABLE)
922 		cxl_handle_rdport_cor_ras(cxlds, dport);
923 	else
924 		cxl_handle_rdport_ras(cxlds, dport);
925 }
926 
927 #else
928 static void cxl_handle_rdport_errors(struct cxl_dev_state *cxlds) { }
929 #endif
930 
931 void cxl_cor_error_detected(struct pci_dev *pdev)
932 {
933 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
934 
935 	if (cxlds->rcd)
936 		cxl_handle_rdport_errors(cxlds);
937 
938 	cxl_handle_endpoint_cor_ras(cxlds);
939 }
940 EXPORT_SYMBOL_NS_GPL(cxl_cor_error_detected, CXL);
941 
942 pci_ers_result_t cxl_error_detected(struct pci_dev *pdev,
943 				    pci_channel_state_t state)
944 {
945 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
946 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
947 	struct device *dev = &cxlmd->dev;
948 	bool ue;
949 
950 	if (cxlds->rcd)
951 		cxl_handle_rdport_errors(cxlds);
952 
953 	/*
954 	 * A frozen channel indicates an impending reset which is fatal to
955 	 * CXL.mem operation, and will likely crash the system. On the off
956 	 * chance the situation is recoverable dump the status of the RAS
957 	 * capability registers and bounce the active state of the memdev.
958 	 */
959 	ue = cxl_handle_endpoint_ras(cxlds);
960 
961 	switch (state) {
962 	case pci_channel_io_normal:
963 		if (ue) {
964 			device_release_driver(dev);
965 			return PCI_ERS_RESULT_NEED_RESET;
966 		}
967 		return PCI_ERS_RESULT_CAN_RECOVER;
968 	case pci_channel_io_frozen:
969 		dev_warn(&pdev->dev,
970 			 "%s: frozen state error detected, disable CXL.mem\n",
971 			 dev_name(dev));
972 		device_release_driver(dev);
973 		return PCI_ERS_RESULT_NEED_RESET;
974 	case pci_channel_io_perm_failure:
975 		dev_warn(&pdev->dev,
976 			 "failure state error detected, request disconnect\n");
977 		return PCI_ERS_RESULT_DISCONNECT;
978 	}
979 	return PCI_ERS_RESULT_NEED_RESET;
980 }
981 EXPORT_SYMBOL_NS_GPL(cxl_error_detected, CXL);
982