xref: /linux/drivers/cxl/core/pci.c (revision c7546e2c3cb739a3c1a2f5acaf9bb629d401afe5)
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 cxl_set_mem_enable(struct cxl_dev_state *cxlds, u16 val)
215 {
216 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
217 	int d = cxlds->cxl_dvsec;
218 	u16 ctrl;
219 	int rc;
220 
221 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, &ctrl);
222 	if (rc < 0)
223 		return rc;
224 
225 	if ((ctrl & CXL_DVSEC_MEM_ENABLE) == val)
226 		return 1;
227 	ctrl &= ~CXL_DVSEC_MEM_ENABLE;
228 	ctrl |= val;
229 
230 	rc = pci_write_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, ctrl);
231 	if (rc < 0)
232 		return rc;
233 
234 	return 0;
235 }
236 
237 static void clear_mem_enable(void *cxlds)
238 {
239 	cxl_set_mem_enable(cxlds, 0);
240 }
241 
242 static int devm_cxl_enable_mem(struct device *host, struct cxl_dev_state *cxlds)
243 {
244 	int rc;
245 
246 	rc = cxl_set_mem_enable(cxlds, CXL_DVSEC_MEM_ENABLE);
247 	if (rc < 0)
248 		return rc;
249 	if (rc > 0)
250 		return 0;
251 	return devm_add_action_or_reset(host, clear_mem_enable, cxlds);
252 }
253 
254 /* require dvsec ranges to be covered by a locked platform window */
255 static int dvsec_range_allowed(struct device *dev, void *arg)
256 {
257 	struct range *dev_range = arg;
258 	struct cxl_decoder *cxld;
259 
260 	if (!is_root_decoder(dev))
261 		return 0;
262 
263 	cxld = to_cxl_decoder(dev);
264 
265 	if (!(cxld->flags & CXL_DECODER_F_RAM))
266 		return 0;
267 
268 	return range_contains(&cxld->hpa_range, dev_range);
269 }
270 
271 static void disable_hdm(void *_cxlhdm)
272 {
273 	u32 global_ctrl;
274 	struct cxl_hdm *cxlhdm = _cxlhdm;
275 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
276 
277 	global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
278 	writel(global_ctrl & ~CXL_HDM_DECODER_ENABLE,
279 	       hdm + CXL_HDM_DECODER_CTRL_OFFSET);
280 }
281 
282 static int devm_cxl_enable_hdm(struct device *host, struct cxl_hdm *cxlhdm)
283 {
284 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
285 	u32 global_ctrl;
286 
287 	global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
288 	writel(global_ctrl | CXL_HDM_DECODER_ENABLE,
289 	       hdm + CXL_HDM_DECODER_CTRL_OFFSET);
290 
291 	return devm_add_action_or_reset(host, disable_hdm, cxlhdm);
292 }
293 
294 int cxl_dvsec_rr_decode(struct device *dev, struct cxl_port *port,
295 			struct cxl_endpoint_dvsec_info *info)
296 {
297 	struct pci_dev *pdev = to_pci_dev(dev);
298 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
299 	int hdm_count, rc, i, ranges = 0;
300 	int d = cxlds->cxl_dvsec;
301 	u16 cap, ctrl;
302 
303 	if (!d) {
304 		dev_dbg(dev, "No DVSEC Capability\n");
305 		return -ENXIO;
306 	}
307 
308 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CAP_OFFSET, &cap);
309 	if (rc)
310 		return rc;
311 
312 	if (!(cap & CXL_DVSEC_MEM_CAPABLE)) {
313 		dev_dbg(dev, "Not MEM Capable\n");
314 		return -ENXIO;
315 	}
316 
317 	/*
318 	 * It is not allowed by spec for MEM.capable to be set and have 0 legacy
319 	 * HDM decoders (values > 2 are also undefined as of CXL 2.0). As this
320 	 * driver is for a spec defined class code which must be CXL.mem
321 	 * capable, there is no point in continuing to enable CXL.mem.
322 	 */
323 	hdm_count = FIELD_GET(CXL_DVSEC_HDM_COUNT_MASK, cap);
324 	if (!hdm_count || hdm_count > 2)
325 		return -EINVAL;
326 
327 	/*
328 	 * The current DVSEC values are moot if the memory capability is
329 	 * disabled, and they will remain moot after the HDM Decoder
330 	 * capability is enabled.
331 	 */
332 	rc = pci_read_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, &ctrl);
333 	if (rc)
334 		return rc;
335 
336 	info->mem_enabled = FIELD_GET(CXL_DVSEC_MEM_ENABLE, ctrl);
337 	if (!info->mem_enabled)
338 		return 0;
339 
340 	for (i = 0; i < hdm_count; i++) {
341 		u64 base, size;
342 		u32 temp;
343 
344 		rc = cxl_dvsec_mem_range_valid(cxlds, i);
345 		if (rc)
346 			return rc;
347 
348 		rc = pci_read_config_dword(
349 			pdev, d + CXL_DVSEC_RANGE_SIZE_HIGH(i), &temp);
350 		if (rc)
351 			return rc;
352 
353 		size = (u64)temp << 32;
354 
355 		rc = pci_read_config_dword(
356 			pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(i), &temp);
357 		if (rc)
358 			return rc;
359 
360 		size |= temp & CXL_DVSEC_MEM_SIZE_LOW_MASK;
361 		if (!size) {
362 			continue;
363 		}
364 
365 		rc = pci_read_config_dword(
366 			pdev, d + CXL_DVSEC_RANGE_BASE_HIGH(i), &temp);
367 		if (rc)
368 			return rc;
369 
370 		base = (u64)temp << 32;
371 
372 		rc = pci_read_config_dword(
373 			pdev, d + CXL_DVSEC_RANGE_BASE_LOW(i), &temp);
374 		if (rc)
375 			return rc;
376 
377 		base |= temp & CXL_DVSEC_MEM_BASE_LOW_MASK;
378 
379 		info->dvsec_range[ranges++] = (struct range) {
380 			.start = base,
381 			.end = base + size - 1
382 		};
383 	}
384 
385 	info->ranges = ranges;
386 
387 	return 0;
388 }
389 EXPORT_SYMBOL_NS_GPL(cxl_dvsec_rr_decode, CXL);
390 
391 /**
392  * cxl_hdm_decode_init() - Setup HDM decoding for the endpoint
393  * @cxlds: Device state
394  * @cxlhdm: Mapped HDM decoder Capability
395  * @info: Cached DVSEC range registers info
396  *
397  * Try to enable the endpoint's HDM Decoder Capability
398  */
399 int cxl_hdm_decode_init(struct cxl_dev_state *cxlds, struct cxl_hdm *cxlhdm,
400 			struct cxl_endpoint_dvsec_info *info)
401 {
402 	void __iomem *hdm = cxlhdm->regs.hdm_decoder;
403 	struct cxl_port *port = cxlhdm->port;
404 	struct device *dev = cxlds->dev;
405 	struct cxl_port *root;
406 	int i, rc, allowed;
407 	u32 global_ctrl = 0;
408 
409 	if (hdm)
410 		global_ctrl = readl(hdm + CXL_HDM_DECODER_CTRL_OFFSET);
411 
412 	/*
413 	 * If the HDM Decoder Capability is already enabled then assume
414 	 * that some other agent like platform firmware set it up.
415 	 */
416 	if (global_ctrl & CXL_HDM_DECODER_ENABLE || (!hdm && info->mem_enabled))
417 		return devm_cxl_enable_mem(&port->dev, cxlds);
418 	else if (!hdm)
419 		return -ENODEV;
420 
421 	root = to_cxl_port(port->dev.parent);
422 	while (!is_cxl_root(root) && is_cxl_port(root->dev.parent))
423 		root = to_cxl_port(root->dev.parent);
424 	if (!is_cxl_root(root)) {
425 		dev_err(dev, "Failed to acquire root port for HDM enable\n");
426 		return -ENODEV;
427 	}
428 
429 	if (!info->mem_enabled) {
430 		rc = devm_cxl_enable_hdm(&port->dev, cxlhdm);
431 		if (rc)
432 			return rc;
433 
434 		return devm_cxl_enable_mem(&port->dev, cxlds);
435 	}
436 
437 	for (i = 0, allowed = 0; i < info->ranges; i++) {
438 		struct device *cxld_dev;
439 
440 		cxld_dev = device_find_child(&root->dev, &info->dvsec_range[i],
441 					     dvsec_range_allowed);
442 		if (!cxld_dev) {
443 			dev_dbg(dev, "DVSEC Range%d denied by platform\n", i);
444 			continue;
445 		}
446 		dev_dbg(dev, "DVSEC Range%d allowed by platform\n", i);
447 		put_device(cxld_dev);
448 		allowed++;
449 	}
450 
451 	if (!allowed) {
452 		dev_err(dev, "Range register decodes outside platform defined CXL ranges.\n");
453 		return -ENXIO;
454 	}
455 
456 	/*
457 	 * Per CXL 2.0 Section 8.1.3.8.3 and 8.1.3.8.4 DVSEC CXL Range 1 Base
458 	 * [High,Low] when HDM operation is enabled the range register values
459 	 * are ignored by the device, but the spec also recommends matching the
460 	 * DVSEC Range 1,2 to HDM Decoder Range 0,1. So, non-zero info->ranges
461 	 * are expected even though Linux does not require or maintain that
462 	 * match. If at least one DVSEC range is enabled and allowed, skip HDM
463 	 * Decoder Capability Enable.
464 	 */
465 	return 0;
466 }
467 EXPORT_SYMBOL_NS_GPL(cxl_hdm_decode_init, CXL);
468 
469 #define CXL_DOE_TABLE_ACCESS_REQ_CODE		0x000000ff
470 #define   CXL_DOE_TABLE_ACCESS_REQ_CODE_READ	0
471 #define CXL_DOE_TABLE_ACCESS_TABLE_TYPE		0x0000ff00
472 #define   CXL_DOE_TABLE_ACCESS_TABLE_TYPE_CDATA	0
473 #define CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE	0xffff0000
474 #define CXL_DOE_TABLE_ACCESS_LAST_ENTRY		0xffff
475 #define CXL_DOE_PROTOCOL_TABLE_ACCESS 2
476 
477 #define CDAT_DOE_REQ(entry_handle) cpu_to_le32				\
478 	(FIELD_PREP(CXL_DOE_TABLE_ACCESS_REQ_CODE,			\
479 		    CXL_DOE_TABLE_ACCESS_REQ_CODE_READ) |		\
480 	 FIELD_PREP(CXL_DOE_TABLE_ACCESS_TABLE_TYPE,			\
481 		    CXL_DOE_TABLE_ACCESS_TABLE_TYPE_CDATA) |		\
482 	 FIELD_PREP(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE, (entry_handle)))
483 
484 static int cxl_cdat_get_length(struct device *dev,
485 			       struct pci_doe_mb *doe_mb,
486 			       size_t *length)
487 {
488 	__le32 request = CDAT_DOE_REQ(0);
489 	__le32 response[2];
490 	int rc;
491 
492 	rc = pci_doe(doe_mb, PCI_VENDOR_ID_CXL,
493 		     CXL_DOE_PROTOCOL_TABLE_ACCESS,
494 		     &request, sizeof(request),
495 		     &response, sizeof(response));
496 	if (rc < 0) {
497 		dev_err(dev, "DOE failed: %d", rc);
498 		return rc;
499 	}
500 	if (rc < sizeof(response))
501 		return -EIO;
502 
503 	*length = le32_to_cpu(response[1]);
504 	dev_dbg(dev, "CDAT length %zu\n", *length);
505 
506 	return 0;
507 }
508 
509 static int cxl_cdat_read_table(struct device *dev,
510 			       struct pci_doe_mb *doe_mb,
511 			       struct cdat_doe_rsp *rsp, size_t *length)
512 {
513 	size_t received, remaining = *length;
514 	unsigned int entry_handle = 0;
515 	union cdat_data *data;
516 	__le32 saved_dw = 0;
517 
518 	do {
519 		__le32 request = CDAT_DOE_REQ(entry_handle);
520 		int rc;
521 
522 		rc = pci_doe(doe_mb, PCI_VENDOR_ID_CXL,
523 			     CXL_DOE_PROTOCOL_TABLE_ACCESS,
524 			     &request, sizeof(request),
525 			     rsp, sizeof(*rsp) + remaining);
526 		if (rc < 0) {
527 			dev_err(dev, "DOE failed: %d", rc);
528 			return rc;
529 		}
530 
531 		if (rc < sizeof(*rsp))
532 			return -EIO;
533 
534 		data = (union cdat_data *)rsp->data;
535 		received = rc - sizeof(*rsp);
536 
537 		if (entry_handle == 0) {
538 			if (received != sizeof(data->header))
539 				return -EIO;
540 		} else {
541 			if (received < sizeof(data->entry) ||
542 			    received != le16_to_cpu(data->entry.length))
543 				return -EIO;
544 		}
545 
546 		/* Get the CXL table access header entry handle */
547 		entry_handle = FIELD_GET(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE,
548 					 le32_to_cpu(rsp->doe_header));
549 
550 		/*
551 		 * Table Access Response Header overwrote the last DW of
552 		 * previous entry, so restore that DW
553 		 */
554 		rsp->doe_header = saved_dw;
555 		remaining -= received;
556 		rsp = (void *)rsp + received;
557 		saved_dw = rsp->doe_header;
558 	} while (entry_handle != CXL_DOE_TABLE_ACCESS_LAST_ENTRY);
559 
560 	/* Length in CDAT header may exceed concatenation of CDAT entries */
561 	*length -= remaining;
562 
563 	return 0;
564 }
565 
566 static unsigned char cdat_checksum(void *buf, size_t size)
567 {
568 	unsigned char sum, *data = buf;
569 	size_t i;
570 
571 	for (sum = 0, i = 0; i < size; i++)
572 		sum += data[i];
573 	return sum;
574 }
575 
576 /**
577  * read_cdat_data - Read the CDAT data on this port
578  * @port: Port to read data from
579  *
580  * This call will sleep waiting for responses from the DOE mailbox.
581  */
582 void read_cdat_data(struct cxl_port *port)
583 {
584 	struct device *uport = port->uport_dev;
585 	struct device *dev = &port->dev;
586 	struct pci_doe_mb *doe_mb;
587 	struct pci_dev *pdev = NULL;
588 	struct cxl_memdev *cxlmd;
589 	struct cdat_doe_rsp *buf;
590 	size_t table_length, length;
591 	int rc;
592 
593 	if (is_cxl_memdev(uport)) {
594 		struct device *host;
595 
596 		cxlmd = to_cxl_memdev(uport);
597 		host = cxlmd->dev.parent;
598 		if (dev_is_pci(host))
599 			pdev = to_pci_dev(host);
600 	} else if (dev_is_pci(uport)) {
601 		pdev = to_pci_dev(uport);
602 	}
603 
604 	if (!pdev)
605 		return;
606 
607 	doe_mb = pci_find_doe_mailbox(pdev, PCI_VENDOR_ID_CXL,
608 				      CXL_DOE_PROTOCOL_TABLE_ACCESS);
609 	if (!doe_mb) {
610 		dev_dbg(dev, "No CDAT mailbox\n");
611 		return;
612 	}
613 
614 	port->cdat_available = true;
615 
616 	if (cxl_cdat_get_length(dev, doe_mb, &length)) {
617 		dev_dbg(dev, "No CDAT length\n");
618 		return;
619 	}
620 
621 	/*
622 	 * The begin of the CDAT buffer needs space for additional 4
623 	 * bytes for the DOE header. Table data starts afterwards.
624 	 */
625 	buf = devm_kzalloc(dev, sizeof(*buf) + length, GFP_KERNEL);
626 	if (!buf)
627 		goto err;
628 
629 	table_length = length;
630 
631 	rc = cxl_cdat_read_table(dev, doe_mb, buf, &length);
632 	if (rc)
633 		goto err;
634 
635 	if (table_length != length)
636 		dev_warn(dev, "Malformed CDAT table length (%zu:%zu), discarding trailing data\n",
637 			table_length, length);
638 
639 	if (cdat_checksum(buf->data, length))
640 		goto err;
641 
642 	port->cdat.table = buf->data;
643 	port->cdat.length = length;
644 
645 	return;
646 err:
647 	/* Don't leave table data allocated on error */
648 	devm_kfree(dev, buf);
649 	dev_err(dev, "Failed to read/validate CDAT.\n");
650 }
651 EXPORT_SYMBOL_NS_GPL(read_cdat_data, CXL);
652 
653 static void __cxl_handle_cor_ras(struct cxl_dev_state *cxlds,
654 				 void __iomem *ras_base)
655 {
656 	void __iomem *addr;
657 	u32 status;
658 
659 	if (!ras_base)
660 		return;
661 
662 	addr = ras_base + CXL_RAS_CORRECTABLE_STATUS_OFFSET;
663 	status = readl(addr);
664 	if (status & CXL_RAS_CORRECTABLE_STATUS_MASK) {
665 		writel(status & CXL_RAS_CORRECTABLE_STATUS_MASK, addr);
666 		trace_cxl_aer_correctable_error(cxlds->cxlmd, status);
667 	}
668 }
669 
670 static void cxl_handle_endpoint_cor_ras(struct cxl_dev_state *cxlds)
671 {
672 	return __cxl_handle_cor_ras(cxlds, cxlds->regs.ras);
673 }
674 
675 /* CXL spec rev3.0 8.2.4.16.1 */
676 static void header_log_copy(void __iomem *ras_base, u32 *log)
677 {
678 	void __iomem *addr;
679 	u32 *log_addr;
680 	int i, log_u32_size = CXL_HEADERLOG_SIZE / sizeof(u32);
681 
682 	addr = ras_base + CXL_RAS_HEADER_LOG_OFFSET;
683 	log_addr = log;
684 
685 	for (i = 0; i < log_u32_size; i++) {
686 		*log_addr = readl(addr);
687 		log_addr++;
688 		addr += sizeof(u32);
689 	}
690 }
691 
692 /*
693  * Log the state of the RAS status registers and prepare them to log the
694  * next error status. Return 1 if reset needed.
695  */
696 static bool __cxl_handle_ras(struct cxl_dev_state *cxlds,
697 				  void __iomem *ras_base)
698 {
699 	u32 hl[CXL_HEADERLOG_SIZE_U32];
700 	void __iomem *addr;
701 	u32 status;
702 	u32 fe;
703 
704 	if (!ras_base)
705 		return false;
706 
707 	addr = ras_base + CXL_RAS_UNCORRECTABLE_STATUS_OFFSET;
708 	status = readl(addr);
709 	if (!(status & CXL_RAS_UNCORRECTABLE_STATUS_MASK))
710 		return false;
711 
712 	/* If multiple errors, log header points to first error from ctrl reg */
713 	if (hweight32(status) > 1) {
714 		void __iomem *rcc_addr =
715 			ras_base + CXL_RAS_CAP_CONTROL_OFFSET;
716 
717 		fe = BIT(FIELD_GET(CXL_RAS_CAP_CONTROL_FE_MASK,
718 				   readl(rcc_addr)));
719 	} else {
720 		fe = status;
721 	}
722 
723 	header_log_copy(ras_base, hl);
724 	trace_cxl_aer_uncorrectable_error(cxlds->cxlmd, status, fe, hl);
725 	writel(status & CXL_RAS_UNCORRECTABLE_STATUS_MASK, addr);
726 
727 	return true;
728 }
729 
730 static bool cxl_handle_endpoint_ras(struct cxl_dev_state *cxlds)
731 {
732 	return __cxl_handle_ras(cxlds, cxlds->regs.ras);
733 }
734 
735 #ifdef CONFIG_PCIEAER_CXL
736 
737 static void cxl_dport_map_rch_aer(struct cxl_dport *dport)
738 {
739 	resource_size_t aer_phys;
740 	struct device *host;
741 	u16 aer_cap;
742 
743 	aer_cap = cxl_rcrb_to_aer(dport->dport_dev, dport->rcrb.base);
744 	if (aer_cap) {
745 		host = dport->reg_map.host;
746 		aer_phys = aer_cap + dport->rcrb.base;
747 		dport->regs.dport_aer = devm_cxl_iomap_block(host, aer_phys,
748 						sizeof(struct aer_capability_regs));
749 	}
750 }
751 
752 static void cxl_dport_map_ras(struct cxl_dport *dport)
753 {
754 	struct cxl_register_map *map = &dport->reg_map;
755 	struct device *dev = dport->dport_dev;
756 
757 	if (!map->component_map.ras.valid)
758 		dev_dbg(dev, "RAS registers not found\n");
759 	else if (cxl_map_component_regs(map, &dport->regs.component,
760 					BIT(CXL_CM_CAP_CAP_ID_RAS)))
761 		dev_dbg(dev, "Failed to map RAS capability.\n");
762 }
763 
764 static void cxl_disable_rch_root_ints(struct cxl_dport *dport)
765 {
766 	void __iomem *aer_base = dport->regs.dport_aer;
767 	u32 aer_cmd_mask, aer_cmd;
768 
769 	if (!aer_base)
770 		return;
771 
772 	/*
773 	 * Disable RCH root port command interrupts.
774 	 * CXL 3.0 12.2.1.1 - RCH Downstream Port-detected Errors
775 	 *
776 	 * This sequence may not be necessary. CXL spec states disabling
777 	 * the root cmd register's interrupts is required. But, PCI spec
778 	 * shows these are disabled by default on reset.
779 	 */
780 	aer_cmd_mask = (PCI_ERR_ROOT_CMD_COR_EN |
781 			PCI_ERR_ROOT_CMD_NONFATAL_EN |
782 			PCI_ERR_ROOT_CMD_FATAL_EN);
783 	aer_cmd = readl(aer_base + PCI_ERR_ROOT_COMMAND);
784 	aer_cmd &= ~aer_cmd_mask;
785 	writel(aer_cmd, aer_base + PCI_ERR_ROOT_COMMAND);
786 }
787 
788 /**
789  * cxl_dport_init_ras_reporting - Setup CXL RAS report on this dport
790  * @dport: the cxl_dport that needs to be initialized
791  * @host: host device for devm operations
792  */
793 void cxl_dport_init_ras_reporting(struct cxl_dport *dport, struct device *host)
794 {
795 	dport->reg_map.host = host;
796 	cxl_dport_map_ras(dport);
797 
798 	if (dport->rch) {
799 		struct pci_host_bridge *host_bridge = to_pci_host_bridge(dport->dport_dev);
800 
801 		if (!host_bridge->native_aer)
802 			return;
803 
804 		cxl_dport_map_rch_aer(dport);
805 		cxl_disable_rch_root_ints(dport);
806 	}
807 }
808 EXPORT_SYMBOL_NS_GPL(cxl_dport_init_ras_reporting, CXL);
809 
810 static void cxl_handle_rdport_cor_ras(struct cxl_dev_state *cxlds,
811 					  struct cxl_dport *dport)
812 {
813 	return __cxl_handle_cor_ras(cxlds, dport->regs.ras);
814 }
815 
816 static bool cxl_handle_rdport_ras(struct cxl_dev_state *cxlds,
817 				       struct cxl_dport *dport)
818 {
819 	return __cxl_handle_ras(cxlds, dport->regs.ras);
820 }
821 
822 /*
823  * Copy the AER capability registers using 32 bit read accesses.
824  * This is necessary because RCRB AER capability is MMIO mapped. Clear the
825  * status after copying.
826  *
827  * @aer_base: base address of AER capability block in RCRB
828  * @aer_regs: destination for copying AER capability
829  */
830 static bool cxl_rch_get_aer_info(void __iomem *aer_base,
831 				 struct aer_capability_regs *aer_regs)
832 {
833 	int read_cnt = sizeof(struct aer_capability_regs) / sizeof(u32);
834 	u32 *aer_regs_buf = (u32 *)aer_regs;
835 	int n;
836 
837 	if (!aer_base)
838 		return false;
839 
840 	/* Use readl() to guarantee 32-bit accesses */
841 	for (n = 0; n < read_cnt; n++)
842 		aer_regs_buf[n] = readl(aer_base + n * sizeof(u32));
843 
844 	writel(aer_regs->uncor_status, aer_base + PCI_ERR_UNCOR_STATUS);
845 	writel(aer_regs->cor_status, aer_base + PCI_ERR_COR_STATUS);
846 
847 	return true;
848 }
849 
850 /* Get AER severity. Return false if there is no error. */
851 static bool cxl_rch_get_aer_severity(struct aer_capability_regs *aer_regs,
852 				     int *severity)
853 {
854 	if (aer_regs->uncor_status & ~aer_regs->uncor_mask) {
855 		if (aer_regs->uncor_status & PCI_ERR_ROOT_FATAL_RCV)
856 			*severity = AER_FATAL;
857 		else
858 			*severity = AER_NONFATAL;
859 		return true;
860 	}
861 
862 	if (aer_regs->cor_status & ~aer_regs->cor_mask) {
863 		*severity = AER_CORRECTABLE;
864 		return true;
865 	}
866 
867 	return false;
868 }
869 
870 static void cxl_handle_rdport_errors(struct cxl_dev_state *cxlds)
871 {
872 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
873 	struct aer_capability_regs aer_regs;
874 	struct cxl_dport *dport;
875 	int severity;
876 
877 	struct cxl_port *port __free(put_cxl_port) =
878 		cxl_pci_find_port(pdev, &dport);
879 	if (!port)
880 		return;
881 
882 	if (!cxl_rch_get_aer_info(dport->regs.dport_aer, &aer_regs))
883 		return;
884 
885 	if (!cxl_rch_get_aer_severity(&aer_regs, &severity))
886 		return;
887 
888 	pci_print_aer(pdev, severity, &aer_regs);
889 
890 	if (severity == AER_CORRECTABLE)
891 		cxl_handle_rdport_cor_ras(cxlds, dport);
892 	else
893 		cxl_handle_rdport_ras(cxlds, dport);
894 }
895 
896 #else
897 static void cxl_handle_rdport_errors(struct cxl_dev_state *cxlds) { }
898 #endif
899 
900 void cxl_cor_error_detected(struct pci_dev *pdev)
901 {
902 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
903 	struct device *dev = &cxlds->cxlmd->dev;
904 
905 	scoped_guard(device, dev) {
906 		if (!dev->driver) {
907 			dev_warn(&pdev->dev,
908 				 "%s: memdev disabled, abort error handling\n",
909 				 dev_name(dev));
910 			return;
911 		}
912 
913 		if (cxlds->rcd)
914 			cxl_handle_rdport_errors(cxlds);
915 
916 		cxl_handle_endpoint_cor_ras(cxlds);
917 	}
918 }
919 EXPORT_SYMBOL_NS_GPL(cxl_cor_error_detected, CXL);
920 
921 pci_ers_result_t cxl_error_detected(struct pci_dev *pdev,
922 				    pci_channel_state_t state)
923 {
924 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
925 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
926 	struct device *dev = &cxlmd->dev;
927 	bool ue;
928 
929 	scoped_guard(device, dev) {
930 		if (!dev->driver) {
931 			dev_warn(&pdev->dev,
932 				 "%s: memdev disabled, abort error handling\n",
933 				 dev_name(dev));
934 			return PCI_ERS_RESULT_DISCONNECT;
935 		}
936 
937 		if (cxlds->rcd)
938 			cxl_handle_rdport_errors(cxlds);
939 		/*
940 		 * A frozen channel indicates an impending reset which is fatal to
941 		 * CXL.mem operation, and will likely crash the system. On the off
942 		 * chance the situation is recoverable dump the status of the RAS
943 		 * capability registers and bounce the active state of the memdev.
944 		 */
945 		ue = cxl_handle_endpoint_ras(cxlds);
946 	}
947 
948 
949 	switch (state) {
950 	case pci_channel_io_normal:
951 		if (ue) {
952 			device_release_driver(dev);
953 			return PCI_ERS_RESULT_NEED_RESET;
954 		}
955 		return PCI_ERS_RESULT_CAN_RECOVER;
956 	case pci_channel_io_frozen:
957 		dev_warn(&pdev->dev,
958 			 "%s: frozen state error detected, disable CXL.mem\n",
959 			 dev_name(dev));
960 		device_release_driver(dev);
961 		return PCI_ERS_RESULT_NEED_RESET;
962 	case pci_channel_io_perm_failure:
963 		dev_warn(&pdev->dev,
964 			 "failure state error detected, request disconnect\n");
965 		return PCI_ERS_RESULT_DISCONNECT;
966 	}
967 	return PCI_ERS_RESULT_NEED_RESET;
968 }
969 EXPORT_SYMBOL_NS_GPL(cxl_error_detected, CXL);
970 
971 static int cxl_flit_size(struct pci_dev *pdev)
972 {
973 	if (cxl_pci_flit_256(pdev))
974 		return 256;
975 
976 	return 68;
977 }
978 
979 /**
980  * cxl_pci_get_latency - calculate the link latency for the PCIe link
981  * @pdev: PCI device
982  *
983  * return: calculated latency or 0 for no latency
984  *
985  * CXL Memory Device SW Guide v1.0 2.11.4 Link latency calculation
986  * Link latency = LinkPropagationLatency + FlitLatency + RetimerLatency
987  * LinkProgationLatency is negligible, so 0 will be used
988  * RetimerLatency is assumed to be negligible and 0 will be used
989  * FlitLatency = FlitSize / LinkBandwidth
990  * FlitSize is defined by spec. CXL rev3.0 4.2.1.
991  * 68B flit is used up to 32GT/s. >32GT/s, 256B flit size is used.
992  * The FlitLatency is converted to picoseconds.
993  */
994 long cxl_pci_get_latency(struct pci_dev *pdev)
995 {
996 	long bw;
997 
998 	bw = pcie_link_speed_mbps(pdev);
999 	if (bw < 0)
1000 		return 0;
1001 	bw /= BITS_PER_BYTE;
1002 
1003 	return cxl_flit_size(pdev) * MEGA / bw;
1004 }
1005 
1006 static int __cxl_endpoint_decoder_reset_detected(struct device *dev, void *data)
1007 {
1008 	struct cxl_port *port = data;
1009 	struct cxl_decoder *cxld;
1010 	struct cxl_hdm *cxlhdm;
1011 	void __iomem *hdm;
1012 	u32 ctrl;
1013 
1014 	if (!is_endpoint_decoder(dev))
1015 		return 0;
1016 
1017 	cxld = to_cxl_decoder(dev);
1018 	if ((cxld->flags & CXL_DECODER_F_ENABLE) == 0)
1019 		return 0;
1020 
1021 	cxlhdm = dev_get_drvdata(&port->dev);
1022 	hdm = cxlhdm->regs.hdm_decoder;
1023 	ctrl = readl(hdm + CXL_HDM_DECODER0_CTRL_OFFSET(cxld->id));
1024 
1025 	return !FIELD_GET(CXL_HDM_DECODER0_CTRL_COMMITTED, ctrl);
1026 }
1027 
1028 bool cxl_endpoint_decoder_reset_detected(struct cxl_port *port)
1029 {
1030 	return device_for_each_child(&port->dev, port,
1031 				     __cxl_endpoint_decoder_reset_detected);
1032 }
1033 EXPORT_SYMBOL_NS_GPL(cxl_endpoint_decoder_reset_detected, CXL);
1034 
1035 int cxl_pci_get_bandwidth(struct pci_dev *pdev, struct access_coordinate *c)
1036 {
1037 	int speed, bw;
1038 	u16 lnksta;
1039 	u32 width;
1040 
1041 	speed = pcie_link_speed_mbps(pdev);
1042 	if (speed < 0)
1043 		return speed;
1044 	speed /= BITS_PER_BYTE;
1045 
1046 	pcie_capability_read_word(pdev, PCI_EXP_LNKSTA, &lnksta);
1047 	width = FIELD_GET(PCI_EXP_LNKSTA_NLW, lnksta);
1048 	bw = speed * width;
1049 
1050 	for (int i = 0; i < ACCESS_COORDINATE_MAX; i++) {
1051 		c[i].read_bandwidth = bw;
1052 		c[i].write_bandwidth = bw;
1053 	}
1054 
1055 	return 0;
1056 }
1057