xref: /linux/drivers/cxl/core/pci.c (revision 71dfa617ea9f18e4585fe78364217cd32b1fc382)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2021 Intel Corporation. All rights reserved. */
3 #include <linux/units.h>
4 #include <linux/io-64-nonatomic-lo-hi.h>
5 #include <linux/device.h>
6 #include <linux/delay.h>
7 #include <linux/pci.h>
8 #include <linux/pci-doe.h>
9 #include <linux/aer.h>
10 #include <cxlpci.h>
11 #include <cxlmem.h>
12 #include <cxl.h>
13 #include "core.h"
14 #include "trace.h"
15 
16 /**
17  * DOC: cxl core pci
18  *
19  * Compute Express Link protocols are layered on top of PCIe. CXL core provides
20  * a set of helpers for CXL interactions which occur via PCIe.
21  */
22 
23 static unsigned short media_ready_timeout = 60;
24 module_param(media_ready_timeout, ushort, 0644);
25 MODULE_PARM_DESC(media_ready_timeout, "seconds to wait for media ready");
26 
27 struct cxl_walk_context {
28 	struct pci_bus *bus;
29 	struct cxl_port *port;
30 	int type;
31 	int error;
32 	int count;
33 };
34 
35 static int match_add_dports(struct pci_dev *pdev, void *data)
36 {
37 	struct cxl_walk_context *ctx = data;
38 	struct cxl_port *port = ctx->port;
39 	int type = pci_pcie_type(pdev);
40 	struct cxl_register_map map;
41 	struct cxl_dport *dport;
42 	u32 lnkcap, port_num;
43 	int rc;
44 
45 	if (pdev->bus != ctx->bus)
46 		return 0;
47 	if (!pci_is_pcie(pdev))
48 		return 0;
49 	if (type != ctx->type)
50 		return 0;
51 	if (pci_read_config_dword(pdev, pci_pcie_cap(pdev) + PCI_EXP_LNKCAP,
52 				  &lnkcap))
53 		return 0;
54 
55 	rc = cxl_find_regblock(pdev, CXL_REGLOC_RBI_COMPONENT, &map);
56 	if (rc)
57 		dev_dbg(&port->dev, "failed to find component registers\n");
58 
59 	port_num = FIELD_GET(PCI_EXP_LNKCAP_PN, lnkcap);
60 	dport = devm_cxl_add_dport(port, &pdev->dev, port_num, map.resource);
61 	if (IS_ERR(dport)) {
62 		ctx->error = PTR_ERR(dport);
63 		return PTR_ERR(dport);
64 	}
65 	ctx->count++;
66 
67 	return 0;
68 }
69 
70 /**
71  * devm_cxl_port_enumerate_dports - enumerate downstream ports of the upstream port
72  * @port: cxl_port whose ->uport_dev is the upstream of dports to be enumerated
73  *
74  * Returns a positive number of dports enumerated or a negative error
75  * code.
76  */
77 int devm_cxl_port_enumerate_dports(struct cxl_port *port)
78 {
79 	struct pci_bus *bus = cxl_port_to_pci_bus(port);
80 	struct cxl_walk_context ctx;
81 	int type;
82 
83 	if (!bus)
84 		return -ENXIO;
85 
86 	if (pci_is_root_bus(bus))
87 		type = PCI_EXP_TYPE_ROOT_PORT;
88 	else
89 		type = PCI_EXP_TYPE_DOWNSTREAM;
90 
91 	ctx = (struct cxl_walk_context) {
92 		.port = port,
93 		.bus = bus,
94 		.type = type,
95 	};
96 	pci_walk_bus(bus, match_add_dports, &ctx);
97 
98 	if (ctx.count == 0)
99 		return -ENODEV;
100 	if (ctx.error)
101 		return ctx.error;
102 	return ctx.count;
103 }
104 EXPORT_SYMBOL_NS_GPL(devm_cxl_port_enumerate_dports, CXL);
105 
106 static int cxl_dvsec_mem_range_valid(struct cxl_dev_state *cxlds, int id)
107 {
108 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
109 	int d = cxlds->cxl_dvsec;
110 	bool valid = false;
111 	int rc, i;
112 	u32 temp;
113 
114 	if (id > CXL_DVSEC_RANGE_MAX)
115 		return -EINVAL;
116 
117 	/* Check MEM INFO VALID bit first, give up after 1s */
118 	i = 1;
119 	do {
120 		rc = pci_read_config_dword(pdev,
121 					   d + CXL_DVSEC_RANGE_SIZE_LOW(id),
122 					   &temp);
123 		if (rc)
124 			return rc;
125 
126 		valid = FIELD_GET(CXL_DVSEC_MEM_INFO_VALID, temp);
127 		if (valid)
128 			break;
129 		msleep(1000);
130 	} while (i--);
131 
132 	if (!valid) {
133 		dev_err(&pdev->dev,
134 			"Timeout awaiting memory range %d valid after 1s.\n",
135 			id);
136 		return -ETIMEDOUT;
137 	}
138 
139 	return 0;
140 }
141 
142 static int cxl_dvsec_mem_range_active(struct cxl_dev_state *cxlds, int id)
143 {
144 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
145 	int d = cxlds->cxl_dvsec;
146 	bool active = false;
147 	int rc, i;
148 	u32 temp;
149 
150 	if (id > CXL_DVSEC_RANGE_MAX)
151 		return -EINVAL;
152 
153 	/* Check MEM ACTIVE bit, up to 60s timeout by default */
154 	for (i = media_ready_timeout; i; i--) {
155 		rc = pci_read_config_dword(
156 			pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(id), &temp);
157 		if (rc)
158 			return rc;
159 
160 		active = FIELD_GET(CXL_DVSEC_MEM_ACTIVE, temp);
161 		if (active)
162 			break;
163 		msleep(1000);
164 	}
165 
166 	if (!active) {
167 		dev_err(&pdev->dev,
168 			"timeout awaiting memory active after %d seconds\n",
169 			media_ready_timeout);
170 		return -ETIMEDOUT;
171 	}
172 
173 	return 0;
174 }
175 
176 /*
177  * Wait up to @media_ready_timeout for the device to report memory
178  * active.
179  */
180 int cxl_await_media_ready(struct cxl_dev_state *cxlds)
181 {
182 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
183 	int d = cxlds->cxl_dvsec;
184 	int rc, i, hdm_count;
185 	u64 md_status;
186 	u16 cap;
187 
188 	rc = pci_read_config_word(pdev,
189 				  d + CXL_DVSEC_CAP_OFFSET, &cap);
190 	if (rc)
191 		return rc;
192 
193 	hdm_count = FIELD_GET(CXL_DVSEC_HDM_COUNT_MASK, cap);
194 	for (i = 0; i < hdm_count; i++) {
195 		rc = cxl_dvsec_mem_range_valid(cxlds, i);
196 		if (rc)
197 			return rc;
198 	}
199 
200 	for (i = 0; i < hdm_count; i++) {
201 		rc = cxl_dvsec_mem_range_active(cxlds, i);
202 		if (rc)
203 			return rc;
204 	}
205 
206 	md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
207 	if (!CXLMDEV_READY(md_status))
208 		return -EIO;
209 
210 	return 0;
211 }
212 EXPORT_SYMBOL_NS_GPL(cxl_await_media_ready, CXL);
213 
214 static int wait_for_valid(struct pci_dev *pdev, int d)
215 {
216 	u32 val;
217 	int rc;
218 
219 	/*
220 	 * Memory_Info_Valid: When set, indicates that the CXL Range 1 Size high
221 	 * and Size Low registers are valid. Must be set within 1 second of
222 	 * deassertion of reset to CXL device. Likely it is already set by the
223 	 * time this runs, but otherwise give a 1.5 second timeout in case of
224 	 * clock skew.
225 	 */
226 	rc = pci_read_config_dword(pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(0), &val);
227 	if (rc)
228 		return rc;
229 
230 	if (val & CXL_DVSEC_MEM_INFO_VALID)
231 		return 0;
232 
233 	msleep(1500);
234 
235 	rc = pci_read_config_dword(pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(0), &val);
236 	if (rc)
237 		return rc;
238 
239 	if (val & CXL_DVSEC_MEM_INFO_VALID)
240 		return 0;
241 
242 	return -ETIMEDOUT;
243 }
244 
245 static int cxl_set_mem_enable(struct cxl_dev_state *cxlds, u16 val)
246 {
247 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
248 	int d = cxlds->cxl_dvsec;
249 	u16 ctrl;
250 	int rc;
251 
252 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, &ctrl);
253 	if (rc < 0)
254 		return rc;
255 
256 	if ((ctrl & CXL_DVSEC_MEM_ENABLE) == val)
257 		return 1;
258 	ctrl &= ~CXL_DVSEC_MEM_ENABLE;
259 	ctrl |= val;
260 
261 	rc = pci_write_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, ctrl);
262 	if (rc < 0)
263 		return rc;
264 
265 	return 0;
266 }
267 
268 static void clear_mem_enable(void *cxlds)
269 {
270 	cxl_set_mem_enable(cxlds, 0);
271 }
272 
273 static int devm_cxl_enable_mem(struct device *host, struct cxl_dev_state *cxlds)
274 {
275 	int rc;
276 
277 	rc = cxl_set_mem_enable(cxlds, CXL_DVSEC_MEM_ENABLE);
278 	if (rc < 0)
279 		return rc;
280 	if (rc > 0)
281 		return 0;
282 	return devm_add_action_or_reset(host, clear_mem_enable, cxlds);
283 }
284 
285 /* require dvsec ranges to be covered by a locked platform window */
286 static int dvsec_range_allowed(struct device *dev, void *arg)
287 {
288 	struct range *dev_range = arg;
289 	struct cxl_decoder *cxld;
290 
291 	if (!is_root_decoder(dev))
292 		return 0;
293 
294 	cxld = to_cxl_decoder(dev);
295 
296 	if (!(cxld->flags & CXL_DECODER_F_RAM))
297 		return 0;
298 
299 	return range_contains(&cxld->hpa_range, dev_range);
300 }
301 
302 static void disable_hdm(void *_cxlhdm)
303 {
304 	u32 global_ctrl;
305 	struct cxl_hdm *cxlhdm = _cxlhdm;
306 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
307 
308 	global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
309 	writel(global_ctrl & ~CXL_HDM_DECODER_ENABLE,
310 	       hdm + CXL_HDM_DECODER_CTRL_OFFSET);
311 }
312 
313 static int devm_cxl_enable_hdm(struct device *host, struct cxl_hdm *cxlhdm)
314 {
315 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
316 	u32 global_ctrl;
317 
318 	global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
319 	writel(global_ctrl | CXL_HDM_DECODER_ENABLE,
320 	       hdm + CXL_HDM_DECODER_CTRL_OFFSET);
321 
322 	return devm_add_action_or_reset(host, disable_hdm, cxlhdm);
323 }
324 
325 int cxl_dvsec_rr_decode(struct device *dev, int d,
326 			struct cxl_endpoint_dvsec_info *info)
327 {
328 	struct pci_dev *pdev = to_pci_dev(dev);
329 	int hdm_count, rc, i, ranges = 0;
330 	u16 cap, ctrl;
331 
332 	if (!d) {
333 		dev_dbg(dev, "No DVSEC Capability\n");
334 		return -ENXIO;
335 	}
336 
337 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CAP_OFFSET, &cap);
338 	if (rc)
339 		return rc;
340 
341 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, &ctrl);
342 	if (rc)
343 		return rc;
344 
345 	if (!(cap & CXL_DVSEC_MEM_CAPABLE)) {
346 		dev_dbg(dev, "Not MEM Capable\n");
347 		return -ENXIO;
348 	}
349 
350 	/*
351 	 * It is not allowed by spec for MEM.capable to be set and have 0 legacy
352 	 * HDM decoders (values > 2 are also undefined as of CXL 2.0). As this
353 	 * driver is for a spec defined class code which must be CXL.mem
354 	 * capable, there is no point in continuing to enable CXL.mem.
355 	 */
356 	hdm_count = FIELD_GET(CXL_DVSEC_HDM_COUNT_MASK, cap);
357 	if (!hdm_count || hdm_count > 2)
358 		return -EINVAL;
359 
360 	rc = wait_for_valid(pdev, d);
361 	if (rc) {
362 		dev_dbg(dev, "Failure awaiting MEM_INFO_VALID (%d)\n", rc);
363 		return rc;
364 	}
365 
366 	/*
367 	 * The current DVSEC values are moot if the memory capability is
368 	 * disabled, and they will remain moot after the HDM Decoder
369 	 * capability is enabled.
370 	 */
371 	info->mem_enabled = FIELD_GET(CXL_DVSEC_MEM_ENABLE, ctrl);
372 	if (!info->mem_enabled)
373 		return 0;
374 
375 	for (i = 0; i < hdm_count; i++) {
376 		u64 base, size;
377 		u32 temp;
378 
379 		rc = pci_read_config_dword(
380 			pdev, d + CXL_DVSEC_RANGE_SIZE_HIGH(i), &temp);
381 		if (rc)
382 			return rc;
383 
384 		size = (u64)temp << 32;
385 
386 		rc = pci_read_config_dword(
387 			pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(i), &temp);
388 		if (rc)
389 			return rc;
390 
391 		size |= temp & CXL_DVSEC_MEM_SIZE_LOW_MASK;
392 		if (!size) {
393 			info->dvsec_range[i] = (struct range) {
394 				.start = 0,
395 				.end = CXL_RESOURCE_NONE,
396 			};
397 			continue;
398 		}
399 
400 		rc = pci_read_config_dword(
401 			pdev, d + CXL_DVSEC_RANGE_BASE_HIGH(i), &temp);
402 		if (rc)
403 			return rc;
404 
405 		base = (u64)temp << 32;
406 
407 		rc = pci_read_config_dword(
408 			pdev, d + CXL_DVSEC_RANGE_BASE_LOW(i), &temp);
409 		if (rc)
410 			return rc;
411 
412 		base |= temp & CXL_DVSEC_MEM_BASE_LOW_MASK;
413 
414 		info->dvsec_range[i] = (struct range) {
415 			.start = base,
416 			.end = base + size - 1
417 		};
418 
419 		ranges++;
420 	}
421 
422 	info->ranges = ranges;
423 
424 	return 0;
425 }
426 EXPORT_SYMBOL_NS_GPL(cxl_dvsec_rr_decode, CXL);
427 
428 /**
429  * cxl_hdm_decode_init() - Setup HDM decoding for the endpoint
430  * @cxlds: Device state
431  * @cxlhdm: Mapped HDM decoder Capability
432  * @info: Cached DVSEC range registers info
433  *
434  * Try to enable the endpoint's HDM Decoder Capability
435  */
436 int cxl_hdm_decode_init(struct cxl_dev_state *cxlds, struct cxl_hdm *cxlhdm,
437 			struct cxl_endpoint_dvsec_info *info)
438 {
439 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
440 	struct cxl_port *port = cxlhdm->port;
441 	struct device *dev = cxlds->dev;
442 	struct cxl_port *root;
443 	int i, rc, allowed;
444 	u32 global_ctrl = 0;
445 
446 	if (hdm)
447 		global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
448 
449 	/*
450 	 * If the HDM Decoder Capability is already enabled then assume
451 	 * that some other agent like platform firmware set it up.
452 	 */
453 	if (global_ctrl & CXL_HDM_DECODER_ENABLE || (!hdm && info->mem_enabled))
454 		return devm_cxl_enable_mem(&port->dev, cxlds);
455 	else if (!hdm)
456 		return -ENODEV;
457 
458 	root = to_cxl_port(port->dev.parent);
459 	while (!is_cxl_root(root) && is_cxl_port(root->dev.parent))
460 		root = to_cxl_port(root->dev.parent);
461 	if (!is_cxl_root(root)) {
462 		dev_err(dev, "Failed to acquire root port for HDM enable\n");
463 		return -ENODEV;
464 	}
465 
466 	for (i = 0, allowed = 0; info->mem_enabled && i < info->ranges; i++) {
467 		struct device *cxld_dev;
468 
469 		cxld_dev = device_find_child(&root->dev, &info->dvsec_range[i],
470 					     dvsec_range_allowed);
471 		if (!cxld_dev) {
472 			dev_dbg(dev, "DVSEC Range%d denied by platform\n", i);
473 			continue;
474 		}
475 		dev_dbg(dev, "DVSEC Range%d allowed by platform\n", i);
476 		put_device(cxld_dev);
477 		allowed++;
478 	}
479 
480 	if (!allowed && info->mem_enabled) {
481 		dev_err(dev, "Range register decodes outside platform defined CXL ranges.\n");
482 		return -ENXIO;
483 	}
484 
485 	/*
486 	 * Per CXL 2.0 Section 8.1.3.8.3 and 8.1.3.8.4 DVSEC CXL Range 1 Base
487 	 * [High,Low] when HDM operation is enabled the range register values
488 	 * are ignored by the device, but the spec also recommends matching the
489 	 * DVSEC Range 1,2 to HDM Decoder Range 0,1. So, non-zero info->ranges
490 	 * are expected even though Linux does not require or maintain that
491 	 * match. If at least one DVSEC range is enabled and allowed, skip HDM
492 	 * Decoder Capability Enable.
493 	 */
494 	if (info->mem_enabled)
495 		return 0;
496 
497 	rc = devm_cxl_enable_hdm(&port->dev, cxlhdm);
498 	if (rc)
499 		return rc;
500 
501 	return devm_cxl_enable_mem(&port->dev, cxlds);
502 }
503 EXPORT_SYMBOL_NS_GPL(cxl_hdm_decode_init, CXL);
504 
505 #define CXL_DOE_TABLE_ACCESS_REQ_CODE		0x000000ff
506 #define   CXL_DOE_TABLE_ACCESS_REQ_CODE_READ	0
507 #define CXL_DOE_TABLE_ACCESS_TABLE_TYPE		0x0000ff00
508 #define   CXL_DOE_TABLE_ACCESS_TABLE_TYPE_CDATA	0
509 #define CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE	0xffff0000
510 #define CXL_DOE_TABLE_ACCESS_LAST_ENTRY		0xffff
511 #define CXL_DOE_PROTOCOL_TABLE_ACCESS 2
512 
513 #define CDAT_DOE_REQ(entry_handle) cpu_to_le32				\
514 	(FIELD_PREP(CXL_DOE_TABLE_ACCESS_REQ_CODE,			\
515 		    CXL_DOE_TABLE_ACCESS_REQ_CODE_READ) |		\
516 	 FIELD_PREP(CXL_DOE_TABLE_ACCESS_TABLE_TYPE,			\
517 		    CXL_DOE_TABLE_ACCESS_TABLE_TYPE_CDATA) |		\
518 	 FIELD_PREP(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE, (entry_handle)))
519 
520 static int cxl_cdat_get_length(struct device *dev,
521 			       struct pci_doe_mb *doe_mb,
522 			       size_t *length)
523 {
524 	__le32 request = CDAT_DOE_REQ(0);
525 	__le32 response[2];
526 	int rc;
527 
528 	rc = pci_doe(doe_mb, PCI_DVSEC_VENDOR_ID_CXL,
529 		     CXL_DOE_PROTOCOL_TABLE_ACCESS,
530 		     &request, sizeof(request),
531 		     &response, sizeof(response));
532 	if (rc < 0) {
533 		dev_err(dev, "DOE failed: %d", rc);
534 		return rc;
535 	}
536 	if (rc < sizeof(response))
537 		return -EIO;
538 
539 	*length = le32_to_cpu(response[1]);
540 	dev_dbg(dev, "CDAT length %zu\n", *length);
541 
542 	return 0;
543 }
544 
545 static int cxl_cdat_read_table(struct device *dev,
546 			       struct pci_doe_mb *doe_mb,
547 			       struct cdat_doe_rsp *rsp, size_t *length)
548 {
549 	size_t received, remaining = *length;
550 	unsigned int entry_handle = 0;
551 	union cdat_data *data;
552 	__le32 saved_dw = 0;
553 
554 	do {
555 		__le32 request = CDAT_DOE_REQ(entry_handle);
556 		int rc;
557 
558 		rc = pci_doe(doe_mb, PCI_DVSEC_VENDOR_ID_CXL,
559 			     CXL_DOE_PROTOCOL_TABLE_ACCESS,
560 			     &request, sizeof(request),
561 			     rsp, sizeof(*rsp) + remaining);
562 		if (rc < 0) {
563 			dev_err(dev, "DOE failed: %d", rc);
564 			return rc;
565 		}
566 
567 		if (rc < sizeof(*rsp))
568 			return -EIO;
569 
570 		data = (union cdat_data *)rsp->data;
571 		received = rc - sizeof(*rsp);
572 
573 		if (entry_handle == 0) {
574 			if (received != sizeof(data->header))
575 				return -EIO;
576 		} else {
577 			if (received < sizeof(data->entry) ||
578 			    received != le16_to_cpu(data->entry.length))
579 				return -EIO;
580 		}
581 
582 		/* Get the CXL table access header entry handle */
583 		entry_handle = FIELD_GET(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE,
584 					 le32_to_cpu(rsp->doe_header));
585 
586 		/*
587 		 * Table Access Response Header overwrote the last DW of
588 		 * previous entry, so restore that DW
589 		 */
590 		rsp->doe_header = saved_dw;
591 		remaining -= received;
592 		rsp = (void *)rsp + received;
593 		saved_dw = rsp->doe_header;
594 	} while (entry_handle != CXL_DOE_TABLE_ACCESS_LAST_ENTRY);
595 
596 	/* Length in CDAT header may exceed concatenation of CDAT entries */
597 	*length -= remaining;
598 
599 	return 0;
600 }
601 
602 static unsigned char cdat_checksum(void *buf, size_t size)
603 {
604 	unsigned char sum, *data = buf;
605 	size_t i;
606 
607 	for (sum = 0, i = 0; i < size; i++)
608 		sum += data[i];
609 	return sum;
610 }
611 
612 /**
613  * read_cdat_data - Read the CDAT data on this port
614  * @port: Port to read data from
615  *
616  * This call will sleep waiting for responses from the DOE mailbox.
617  */
618 void read_cdat_data(struct cxl_port *port)
619 {
620 	struct device *uport = port->uport_dev;
621 	struct device *dev = &port->dev;
622 	struct pci_doe_mb *doe_mb;
623 	struct pci_dev *pdev = NULL;
624 	struct cxl_memdev *cxlmd;
625 	struct cdat_doe_rsp *buf;
626 	size_t table_length, length;
627 	int rc;
628 
629 	if (is_cxl_memdev(uport)) {
630 		struct device *host;
631 
632 		cxlmd = to_cxl_memdev(uport);
633 		host = cxlmd->dev.parent;
634 		if (dev_is_pci(host))
635 			pdev = to_pci_dev(host);
636 	} else if (dev_is_pci(uport)) {
637 		pdev = to_pci_dev(uport);
638 	}
639 
640 	if (!pdev)
641 		return;
642 
643 	doe_mb = pci_find_doe_mailbox(pdev, PCI_DVSEC_VENDOR_ID_CXL,
644 				      CXL_DOE_PROTOCOL_TABLE_ACCESS);
645 	if (!doe_mb) {
646 		dev_dbg(dev, "No CDAT mailbox\n");
647 		return;
648 	}
649 
650 	port->cdat_available = true;
651 
652 	if (cxl_cdat_get_length(dev, doe_mb, &length)) {
653 		dev_dbg(dev, "No CDAT length\n");
654 		return;
655 	}
656 
657 	/*
658 	 * The begin of the CDAT buffer needs space for additional 4
659 	 * bytes for the DOE header. Table data starts afterwards.
660 	 */
661 	buf = devm_kzalloc(dev, sizeof(*buf) + length, GFP_KERNEL);
662 	if (!buf)
663 		goto err;
664 
665 	table_length = length;
666 
667 	rc = cxl_cdat_read_table(dev, doe_mb, buf, &length);
668 	if (rc)
669 		goto err;
670 
671 	if (table_length != length)
672 		dev_warn(dev, "Malformed CDAT table length (%zu:%zu), discarding trailing data\n",
673 			table_length, length);
674 
675 	if (cdat_checksum(buf->data, length))
676 		goto err;
677 
678 	port->cdat.table = buf->data;
679 	port->cdat.length = length;
680 
681 	return;
682 err:
683 	/* Don't leave table data allocated on error */
684 	devm_kfree(dev, buf);
685 	dev_err(dev, "Failed to read/validate CDAT.\n");
686 }
687 EXPORT_SYMBOL_NS_GPL(read_cdat_data, CXL);
688 
689 static void __cxl_handle_cor_ras(struct cxl_dev_state *cxlds,
690 				 void __iomem *ras_base)
691 {
692 	void __iomem *addr;
693 	u32 status;
694 
695 	if (!ras_base)
696 		return;
697 
698 	addr = ras_base + CXL_RAS_CORRECTABLE_STATUS_OFFSET;
699 	status = readl(addr);
700 	if (status & CXL_RAS_CORRECTABLE_STATUS_MASK) {
701 		writel(status & CXL_RAS_CORRECTABLE_STATUS_MASK, addr);
702 		trace_cxl_aer_correctable_error(cxlds->cxlmd, status);
703 	}
704 }
705 
706 static void cxl_handle_endpoint_cor_ras(struct cxl_dev_state *cxlds)
707 {
708 	return __cxl_handle_cor_ras(cxlds, cxlds->regs.ras);
709 }
710 
711 /* CXL spec rev3.0 8.2.4.16.1 */
712 static void header_log_copy(void __iomem *ras_base, u32 *log)
713 {
714 	void __iomem *addr;
715 	u32 *log_addr;
716 	int i, log_u32_size = CXL_HEADERLOG_SIZE / sizeof(u32);
717 
718 	addr = ras_base + CXL_RAS_HEADER_LOG_OFFSET;
719 	log_addr = log;
720 
721 	for (i = 0; i < log_u32_size; i++) {
722 		*log_addr = readl(addr);
723 		log_addr++;
724 		addr += sizeof(u32);
725 	}
726 }
727 
728 /*
729  * Log the state of the RAS status registers and prepare them to log the
730  * next error status. Return 1 if reset needed.
731  */
732 static bool __cxl_handle_ras(struct cxl_dev_state *cxlds,
733 				  void __iomem *ras_base)
734 {
735 	u32 hl[CXL_HEADERLOG_SIZE_U32];
736 	void __iomem *addr;
737 	u32 status;
738 	u32 fe;
739 
740 	if (!ras_base)
741 		return false;
742 
743 	addr = ras_base + CXL_RAS_UNCORRECTABLE_STATUS_OFFSET;
744 	status = readl(addr);
745 	if (!(status & CXL_RAS_UNCORRECTABLE_STATUS_MASK))
746 		return false;
747 
748 	/* If multiple errors, log header points to first error from ctrl reg */
749 	if (hweight32(status) > 1) {
750 		void __iomem *rcc_addr =
751 			ras_base + CXL_RAS_CAP_CONTROL_OFFSET;
752 
753 		fe = BIT(FIELD_GET(CXL_RAS_CAP_CONTROL_FE_MASK,
754 				   readl(rcc_addr)));
755 	} else {
756 		fe = status;
757 	}
758 
759 	header_log_copy(ras_base, hl);
760 	trace_cxl_aer_uncorrectable_error(cxlds->cxlmd, status, fe, hl);
761 	writel(status & CXL_RAS_UNCORRECTABLE_STATUS_MASK, addr);
762 
763 	return true;
764 }
765 
766 static bool cxl_handle_endpoint_ras(struct cxl_dev_state *cxlds)
767 {
768 	return __cxl_handle_ras(cxlds, cxlds->regs.ras);
769 }
770 
771 #ifdef CONFIG_PCIEAER_CXL
772 
773 static void cxl_dport_map_rch_aer(struct cxl_dport *dport)
774 {
775 	struct cxl_rcrb_info *ri = &dport->rcrb;
776 	void __iomem *dport_aer = NULL;
777 	resource_size_t aer_phys;
778 	struct device *host;
779 
780 	if (dport->rch && ri->aer_cap) {
781 		host = dport->reg_map.host;
782 		aer_phys = ri->aer_cap + ri->base;
783 		dport_aer = devm_cxl_iomap_block(host, aer_phys,
784 				sizeof(struct aer_capability_regs));
785 	}
786 
787 	dport->regs.dport_aer = dport_aer;
788 }
789 
790 static void cxl_dport_map_regs(struct cxl_dport *dport)
791 {
792 	struct cxl_register_map *map = &dport->reg_map;
793 	struct device *dev = dport->dport_dev;
794 
795 	if (!map->component_map.ras.valid)
796 		dev_dbg(dev, "RAS registers not found\n");
797 	else if (cxl_map_component_regs(map, &dport->regs.component,
798 					BIT(CXL_CM_CAP_CAP_ID_RAS)))
799 		dev_dbg(dev, "Failed to map RAS capability.\n");
800 
801 	if (dport->rch)
802 		cxl_dport_map_rch_aer(dport);
803 }
804 
805 static void cxl_disable_rch_root_ints(struct cxl_dport *dport)
806 {
807 	void __iomem *aer_base = dport->regs.dport_aer;
808 	struct pci_host_bridge *bridge;
809 	u32 aer_cmd_mask, aer_cmd;
810 
811 	if (!aer_base)
812 		return;
813 
814 	bridge = to_pci_host_bridge(dport->dport_dev);
815 
816 	/*
817 	 * Disable RCH root port command interrupts.
818 	 * CXL 3.0 12.2.1.1 - RCH Downstream Port-detected Errors
819 	 *
820 	 * This sequence may not be necessary. CXL spec states disabling
821 	 * the root cmd register's interrupts is required. But, PCI spec
822 	 * shows these are disabled by default on reset.
823 	 */
824 	if (bridge->native_aer) {
825 		aer_cmd_mask = (PCI_ERR_ROOT_CMD_COR_EN |
826 				PCI_ERR_ROOT_CMD_NONFATAL_EN |
827 				PCI_ERR_ROOT_CMD_FATAL_EN);
828 		aer_cmd = readl(aer_base + PCI_ERR_ROOT_COMMAND);
829 		aer_cmd &= ~aer_cmd_mask;
830 		writel(aer_cmd, aer_base + PCI_ERR_ROOT_COMMAND);
831 	}
832 }
833 
834 void cxl_setup_parent_dport(struct device *host, struct cxl_dport *dport)
835 {
836 	struct device *dport_dev = dport->dport_dev;
837 	struct pci_host_bridge *host_bridge;
838 
839 	host_bridge = to_pci_host_bridge(dport_dev);
840 	if (host_bridge->native_aer)
841 		dport->rcrb.aer_cap = cxl_rcrb_to_aer(dport_dev, dport->rcrb.base);
842 
843 	dport->reg_map.host = host;
844 	cxl_dport_map_regs(dport);
845 
846 	if (dport->rch)
847 		cxl_disable_rch_root_ints(dport);
848 }
849 EXPORT_SYMBOL_NS_GPL(cxl_setup_parent_dport, CXL);
850 
851 static void cxl_handle_rdport_cor_ras(struct cxl_dev_state *cxlds,
852 					  struct cxl_dport *dport)
853 {
854 	return __cxl_handle_cor_ras(cxlds, dport->regs.ras);
855 }
856 
857 static bool cxl_handle_rdport_ras(struct cxl_dev_state *cxlds,
858 				       struct cxl_dport *dport)
859 {
860 	return __cxl_handle_ras(cxlds, dport->regs.ras);
861 }
862 
863 /*
864  * Copy the AER capability registers using 32 bit read accesses.
865  * This is necessary because RCRB AER capability is MMIO mapped. Clear the
866  * status after copying.
867  *
868  * @aer_base: base address of AER capability block in RCRB
869  * @aer_regs: destination for copying AER capability
870  */
871 static bool cxl_rch_get_aer_info(void __iomem *aer_base,
872 				 struct aer_capability_regs *aer_regs)
873 {
874 	int read_cnt = sizeof(struct aer_capability_regs) / sizeof(u32);
875 	u32 *aer_regs_buf = (u32 *)aer_regs;
876 	int n;
877 
878 	if (!aer_base)
879 		return false;
880 
881 	/* Use readl() to guarantee 32-bit accesses */
882 	for (n = 0; n < read_cnt; n++)
883 		aer_regs_buf[n] = readl(aer_base + n * sizeof(u32));
884 
885 	writel(aer_regs->uncor_status, aer_base + PCI_ERR_UNCOR_STATUS);
886 	writel(aer_regs->cor_status, aer_base + PCI_ERR_COR_STATUS);
887 
888 	return true;
889 }
890 
891 /* Get AER severity. Return false if there is no error. */
892 static bool cxl_rch_get_aer_severity(struct aer_capability_regs *aer_regs,
893 				     int *severity)
894 {
895 	if (aer_regs->uncor_status & ~aer_regs->uncor_mask) {
896 		if (aer_regs->uncor_status & PCI_ERR_ROOT_FATAL_RCV)
897 			*severity = AER_FATAL;
898 		else
899 			*severity = AER_NONFATAL;
900 		return true;
901 	}
902 
903 	if (aer_regs->cor_status & ~aer_regs->cor_mask) {
904 		*severity = AER_CORRECTABLE;
905 		return true;
906 	}
907 
908 	return false;
909 }
910 
911 static void cxl_handle_rdport_errors(struct cxl_dev_state *cxlds)
912 {
913 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
914 	struct aer_capability_regs aer_regs;
915 	struct cxl_dport *dport;
916 	struct cxl_port *port;
917 	int severity;
918 
919 	port = cxl_pci_find_port(pdev, &dport);
920 	if (!port)
921 		return;
922 
923 	put_device(&port->dev);
924 
925 	if (!cxl_rch_get_aer_info(dport->regs.dport_aer, &aer_regs))
926 		return;
927 
928 	if (!cxl_rch_get_aer_severity(&aer_regs, &severity))
929 		return;
930 
931 	pci_print_aer(pdev, severity, &aer_regs);
932 
933 	if (severity == AER_CORRECTABLE)
934 		cxl_handle_rdport_cor_ras(cxlds, dport);
935 	else
936 		cxl_handle_rdport_ras(cxlds, dport);
937 }
938 
939 #else
940 static void cxl_handle_rdport_errors(struct cxl_dev_state *cxlds) { }
941 #endif
942 
943 void cxl_cor_error_detected(struct pci_dev *pdev)
944 {
945 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
946 	struct device *dev = &cxlds->cxlmd->dev;
947 
948 	scoped_guard(device, dev) {
949 		if (!dev->driver) {
950 			dev_warn(&pdev->dev,
951 				 "%s: memdev disabled, abort error handling\n",
952 				 dev_name(dev));
953 			return;
954 		}
955 
956 		if (cxlds->rcd)
957 			cxl_handle_rdport_errors(cxlds);
958 
959 		cxl_handle_endpoint_cor_ras(cxlds);
960 	}
961 }
962 EXPORT_SYMBOL_NS_GPL(cxl_cor_error_detected, CXL);
963 
964 pci_ers_result_t cxl_error_detected(struct pci_dev *pdev,
965 				    pci_channel_state_t state)
966 {
967 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
968 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
969 	struct device *dev = &cxlmd->dev;
970 	bool ue;
971 
972 	scoped_guard(device, dev) {
973 		if (!dev->driver) {
974 			dev_warn(&pdev->dev,
975 				 "%s: memdev disabled, abort error handling\n",
976 				 dev_name(dev));
977 			return PCI_ERS_RESULT_DISCONNECT;
978 		}
979 
980 		if (cxlds->rcd)
981 			cxl_handle_rdport_errors(cxlds);
982 		/*
983 		 * A frozen channel indicates an impending reset which is fatal to
984 		 * CXL.mem operation, and will likely crash the system. On the off
985 		 * chance the situation is recoverable dump the status of the RAS
986 		 * capability registers and bounce the active state of the memdev.
987 		 */
988 		ue = cxl_handle_endpoint_ras(cxlds);
989 	}
990 
991 
992 	switch (state) {
993 	case pci_channel_io_normal:
994 		if (ue) {
995 			device_release_driver(dev);
996 			return PCI_ERS_RESULT_NEED_RESET;
997 		}
998 		return PCI_ERS_RESULT_CAN_RECOVER;
999 	case pci_channel_io_frozen:
1000 		dev_warn(&pdev->dev,
1001 			 "%s: frozen state error detected, disable CXL.mem\n",
1002 			 dev_name(dev));
1003 		device_release_driver(dev);
1004 		return PCI_ERS_RESULT_NEED_RESET;
1005 	case pci_channel_io_perm_failure:
1006 		dev_warn(&pdev->dev,
1007 			 "failure state error detected, request disconnect\n");
1008 		return PCI_ERS_RESULT_DISCONNECT;
1009 	}
1010 	return PCI_ERS_RESULT_NEED_RESET;
1011 }
1012 EXPORT_SYMBOL_NS_GPL(cxl_error_detected, CXL);
1013 
1014 static int cxl_flit_size(struct pci_dev *pdev)
1015 {
1016 	if (cxl_pci_flit_256(pdev))
1017 		return 256;
1018 
1019 	return 68;
1020 }
1021 
1022 /**
1023  * cxl_pci_get_latency - calculate the link latency for the PCIe link
1024  * @pdev: PCI device
1025  *
1026  * return: calculated latency or 0 for no latency
1027  *
1028  * CXL Memory Device SW Guide v1.0 2.11.4 Link latency calculation
1029  * Link latency = LinkPropagationLatency + FlitLatency + RetimerLatency
1030  * LinkProgationLatency is negligible, so 0 will be used
1031  * RetimerLatency is assumed to be negligible and 0 will be used
1032  * FlitLatency = FlitSize / LinkBandwidth
1033  * FlitSize is defined by spec. CXL rev3.0 4.2.1.
1034  * 68B flit is used up to 32GT/s. >32GT/s, 256B flit size is used.
1035  * The FlitLatency is converted to picoseconds.
1036  */
1037 long cxl_pci_get_latency(struct pci_dev *pdev)
1038 {
1039 	long bw;
1040 
1041 	bw = pcie_link_speed_mbps(pdev);
1042 	if (bw < 0)
1043 		return 0;
1044 	bw /= BITS_PER_BYTE;
1045 
1046 	return cxl_flit_size(pdev) * MEGA / bw;
1047 }
1048