• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2020 Intel Corporation. All rights reserved. */
3 #include <linux/io-64-nonatomic-lo-hi.h>
4 #include <linux/moduleparam.h>
5 #include <linux/module.h>
6 #include <linux/delay.h>
7 #include <linux/sizes.h>
8 #include <linux/mutex.h>
9 #include <linux/list.h>
10 #include <linux/pci.h>
11 #include <linux/pci-doe.h>
12 #include <linux/io.h>
13 #include "cxlmem.h"
14 #include "cxlpci.h"
15 #include "cxl.h"
16 
17 /**
18  * DOC: cxl pci
19  *
20  * This implements the PCI exclusive functionality for a CXL device as it is
21  * defined by the Compute Express Link specification. CXL devices may surface
22  * certain functionality even if it isn't CXL enabled. While this driver is
23  * focused around the PCI specific aspects of a CXL device, it binds to the
24  * specific CXL memory device class code, and therefore the implementation of
25  * cxl_pci is focused around CXL memory devices.
26  *
27  * The driver has several responsibilities, mainly:
28  *  - Create the memX device and register on the CXL bus.
29  *  - Enumerate device's register interface and map them.
30  *  - Registers nvdimm bridge device with cxl_core.
31  *  - Registers a CXL mailbox with cxl_core.
32  */
33 
34 #define cxl_doorbell_busy(cxlds)                                                \
35 	(readl((cxlds)->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET) &                  \
36 	 CXLDEV_MBOX_CTRL_DOORBELL)
37 
38 /* CXL 2.0 - 8.2.8.4 */
39 #define CXL_MAILBOX_TIMEOUT_MS (2 * HZ)
40 
41 /*
42  * CXL 2.0 ECN "Add Mailbox Ready Time" defines a capability field to
43  * dictate how long to wait for the mailbox to become ready. The new
44  * field allows the device to tell software the amount of time to wait
45  * before mailbox ready. This field per the spec theoretically allows
46  * for up to 255 seconds. 255 seconds is unreasonably long, its longer
47  * than the maximum SATA port link recovery wait. Default to 60 seconds
48  * until someone builds a CXL device that needs more time in practice.
49  */
50 static unsigned short mbox_ready_timeout = 60;
51 module_param(mbox_ready_timeout, ushort, 0644);
52 MODULE_PARM_DESC(mbox_ready_timeout, "seconds to wait for mailbox ready");
53 
cxl_pci_mbox_wait_for_doorbell(struct cxl_dev_state * cxlds)54 static int cxl_pci_mbox_wait_for_doorbell(struct cxl_dev_state *cxlds)
55 {
56 	const unsigned long start = jiffies;
57 	unsigned long end = start;
58 
59 	while (cxl_doorbell_busy(cxlds)) {
60 		end = jiffies;
61 
62 		if (time_after(end, start + CXL_MAILBOX_TIMEOUT_MS)) {
63 			/* Check again in case preempted before timeout test */
64 			if (!cxl_doorbell_busy(cxlds))
65 				break;
66 			return -ETIMEDOUT;
67 		}
68 		cpu_relax();
69 	}
70 
71 	dev_dbg(cxlds->dev, "Doorbell wait took %dms",
72 		jiffies_to_msecs(end) - jiffies_to_msecs(start));
73 	return 0;
74 }
75 
76 #define cxl_err(dev, status, msg)                                        \
77 	dev_err_ratelimited(dev, msg ", device state %s%s\n",                  \
78 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
79 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
80 
81 #define cxl_cmd_err(dev, cmd, status, msg)                               \
82 	dev_err_ratelimited(dev, msg " (opcode: %#x), device state %s%s\n",    \
83 			    (cmd)->opcode,                                     \
84 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
85 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
86 
87 /**
88  * __cxl_pci_mbox_send_cmd() - Execute a mailbox command
89  * @cxlds: The device state to communicate with.
90  * @mbox_cmd: Command to send to the memory device.
91  *
92  * Context: Any context. Expects mbox_mutex to be held.
93  * Return: -ETIMEDOUT if timeout occurred waiting for completion. 0 on success.
94  *         Caller should check the return code in @mbox_cmd to make sure it
95  *         succeeded.
96  *
97  * This is a generic form of the CXL mailbox send command thus only using the
98  * registers defined by the mailbox capability ID - CXL 2.0 8.2.8.4. Memory
99  * devices, and perhaps other types of CXL devices may have further information
100  * available upon error conditions. Driver facilities wishing to send mailbox
101  * commands should use the wrapper command.
102  *
103  * The CXL spec allows for up to two mailboxes. The intention is for the primary
104  * mailbox to be OS controlled and the secondary mailbox to be used by system
105  * firmware. This allows the OS and firmware to communicate with the device and
106  * not need to coordinate with each other. The driver only uses the primary
107  * mailbox.
108  */
__cxl_pci_mbox_send_cmd(struct cxl_dev_state * cxlds,struct cxl_mbox_cmd * mbox_cmd)109 static int __cxl_pci_mbox_send_cmd(struct cxl_dev_state *cxlds,
110 				   struct cxl_mbox_cmd *mbox_cmd)
111 {
112 	void __iomem *payload = cxlds->regs.mbox + CXLDEV_MBOX_PAYLOAD_OFFSET;
113 	struct device *dev = cxlds->dev;
114 	u64 cmd_reg, status_reg;
115 	size_t out_len;
116 	int rc;
117 
118 	lockdep_assert_held(&cxlds->mbox_mutex);
119 
120 	/*
121 	 * Here are the steps from 8.2.8.4 of the CXL 2.0 spec.
122 	 *   1. Caller reads MB Control Register to verify doorbell is clear
123 	 *   2. Caller writes Command Register
124 	 *   3. Caller writes Command Payload Registers if input payload is non-empty
125 	 *   4. Caller writes MB Control Register to set doorbell
126 	 *   5. Caller either polls for doorbell to be clear or waits for interrupt if configured
127 	 *   6. Caller reads MB Status Register to fetch Return code
128 	 *   7. If command successful, Caller reads Command Register to get Payload Length
129 	 *   8. If output payload is non-empty, host reads Command Payload Registers
130 	 *
131 	 * Hardware is free to do whatever it wants before the doorbell is rung,
132 	 * and isn't allowed to change anything after it clears the doorbell. As
133 	 * such, steps 2 and 3 can happen in any order, and steps 6, 7, 8 can
134 	 * also happen in any order (though some orders might not make sense).
135 	 */
136 
137 	/* #1 */
138 	if (cxl_doorbell_busy(cxlds)) {
139 		u64 md_status =
140 			readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
141 
142 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status,
143 			    "mailbox queue busy");
144 		return -EBUSY;
145 	}
146 
147 	cmd_reg = FIELD_PREP(CXLDEV_MBOX_CMD_COMMAND_OPCODE_MASK,
148 			     mbox_cmd->opcode);
149 	if (mbox_cmd->size_in) {
150 		if (WARN_ON(!mbox_cmd->payload_in))
151 			return -EINVAL;
152 
153 		cmd_reg |= FIELD_PREP(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK,
154 				      mbox_cmd->size_in);
155 		memcpy_toio(payload, mbox_cmd->payload_in, mbox_cmd->size_in);
156 	}
157 
158 	/* #2, #3 */
159 	writeq(cmd_reg, cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
160 
161 	/* #4 */
162 	dev_dbg(dev, "Sending command\n");
163 	writel(CXLDEV_MBOX_CTRL_DOORBELL,
164 	       cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
165 
166 	/* #5 */
167 	rc = cxl_pci_mbox_wait_for_doorbell(cxlds);
168 	if (rc == -ETIMEDOUT) {
169 		u64 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
170 
171 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status, "mailbox timeout");
172 		return rc;
173 	}
174 
175 	/* #6 */
176 	status_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_STATUS_OFFSET);
177 	mbox_cmd->return_code =
178 		FIELD_GET(CXLDEV_MBOX_STATUS_RET_CODE_MASK, status_reg);
179 
180 	if (mbox_cmd->return_code != CXL_MBOX_CMD_RC_SUCCESS) {
181 		dev_dbg(dev, "Mailbox operation had an error: %s\n",
182 			cxl_mbox_cmd_rc2str(mbox_cmd));
183 		return 0; /* completed but caller must check return_code */
184 	}
185 
186 	/* #7 */
187 	cmd_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
188 	out_len = FIELD_GET(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK, cmd_reg);
189 
190 	/* #8 */
191 	if (out_len && mbox_cmd->payload_out) {
192 		/*
193 		 * Sanitize the copy. If hardware misbehaves, out_len per the
194 		 * spec can actually be greater than the max allowed size (21
195 		 * bits available but spec defined 1M max). The caller also may
196 		 * have requested less data than the hardware supplied even
197 		 * within spec.
198 		 */
199 		size_t n = min3(mbox_cmd->size_out, cxlds->payload_size, out_len);
200 
201 		memcpy_fromio(mbox_cmd->payload_out, payload, n);
202 		mbox_cmd->size_out = n;
203 	} else {
204 		mbox_cmd->size_out = 0;
205 	}
206 
207 	return 0;
208 }
209 
cxl_pci_mbox_send(struct cxl_dev_state * cxlds,struct cxl_mbox_cmd * cmd)210 static int cxl_pci_mbox_send(struct cxl_dev_state *cxlds, struct cxl_mbox_cmd *cmd)
211 {
212 	int rc;
213 
214 	mutex_lock_io(&cxlds->mbox_mutex);
215 	rc = __cxl_pci_mbox_send_cmd(cxlds, cmd);
216 	mutex_unlock(&cxlds->mbox_mutex);
217 
218 	return rc;
219 }
220 
cxl_pci_setup_mailbox(struct cxl_dev_state * cxlds)221 static int cxl_pci_setup_mailbox(struct cxl_dev_state *cxlds)
222 {
223 	const int cap = readl(cxlds->regs.mbox + CXLDEV_MBOX_CAPS_OFFSET);
224 	unsigned long timeout;
225 	u64 md_status;
226 
227 	timeout = jiffies + mbox_ready_timeout * HZ;
228 	do {
229 		md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
230 		if (md_status & CXLMDEV_MBOX_IF_READY)
231 			break;
232 		if (msleep_interruptible(100))
233 			break;
234 	} while (!time_after(jiffies, timeout));
235 
236 	if (!(md_status & CXLMDEV_MBOX_IF_READY)) {
237 		cxl_err(cxlds->dev, md_status,
238 			"timeout awaiting mailbox ready");
239 		return -ETIMEDOUT;
240 	}
241 
242 	/*
243 	 * A command may be in flight from a previous driver instance,
244 	 * think kexec, do one doorbell wait so that
245 	 * __cxl_pci_mbox_send_cmd() can assume that it is the only
246 	 * source for future doorbell busy events.
247 	 */
248 	if (cxl_pci_mbox_wait_for_doorbell(cxlds) != 0) {
249 		cxl_err(cxlds->dev, md_status, "timeout awaiting mailbox idle");
250 		return -ETIMEDOUT;
251 	}
252 
253 	cxlds->mbox_send = cxl_pci_mbox_send;
254 	cxlds->payload_size =
255 		1 << FIELD_GET(CXLDEV_MBOX_CAP_PAYLOAD_SIZE_MASK, cap);
256 
257 	/*
258 	 * CXL 2.0 8.2.8.4.3 Mailbox Capabilities Register
259 	 *
260 	 * If the size is too small, mandatory commands will not work and so
261 	 * there's no point in going forward. If the size is too large, there's
262 	 * no harm is soft limiting it.
263 	 */
264 	cxlds->payload_size = min_t(size_t, cxlds->payload_size, SZ_1M);
265 	if (cxlds->payload_size < 256) {
266 		dev_err(cxlds->dev, "Mailbox is too small (%zub)",
267 			cxlds->payload_size);
268 		return -ENXIO;
269 	}
270 
271 	dev_dbg(cxlds->dev, "Mailbox payload sized %zu",
272 		cxlds->payload_size);
273 
274 	return 0;
275 }
276 
cxl_map_regblock(struct pci_dev * pdev,struct cxl_register_map * map)277 static int cxl_map_regblock(struct pci_dev *pdev, struct cxl_register_map *map)
278 {
279 	void __iomem *addr;
280 	int bar = map->barno;
281 	struct device *dev = &pdev->dev;
282 	resource_size_t offset = map->block_offset;
283 
284 	/* Basic sanity check that BAR is big enough */
285 	if (pci_resource_len(pdev, bar) < offset) {
286 		dev_err(dev, "BAR%d: %pr: too small (offset: %pa)\n", bar,
287 			&pdev->resource[bar], &offset);
288 		return -ENXIO;
289 	}
290 
291 	addr = pci_iomap(pdev, bar, 0);
292 	if (!addr) {
293 		dev_err(dev, "failed to map registers\n");
294 		return -ENOMEM;
295 	}
296 
297 	dev_dbg(dev, "Mapped CXL Memory Device resource bar %u @ %pa\n",
298 		bar, &offset);
299 
300 	map->base = addr + map->block_offset;
301 	return 0;
302 }
303 
cxl_unmap_regblock(struct pci_dev * pdev,struct cxl_register_map * map)304 static void cxl_unmap_regblock(struct pci_dev *pdev,
305 			       struct cxl_register_map *map)
306 {
307 	pci_iounmap(pdev, map->base - map->block_offset);
308 	map->base = NULL;
309 }
310 
cxl_probe_regs(struct pci_dev * pdev,struct cxl_register_map * map)311 static int cxl_probe_regs(struct pci_dev *pdev, struct cxl_register_map *map)
312 {
313 	struct cxl_component_reg_map *comp_map;
314 	struct cxl_device_reg_map *dev_map;
315 	struct device *dev = &pdev->dev;
316 	void __iomem *base = map->base;
317 
318 	switch (map->reg_type) {
319 	case CXL_REGLOC_RBI_COMPONENT:
320 		comp_map = &map->component_map;
321 		cxl_probe_component_regs(dev, base, comp_map);
322 		if (!comp_map->hdm_decoder.valid) {
323 			dev_err(dev, "HDM decoder registers not found\n");
324 			return -ENXIO;
325 		}
326 
327 		dev_dbg(dev, "Set up component registers\n");
328 		break;
329 	case CXL_REGLOC_RBI_MEMDEV:
330 		dev_map = &map->device_map;
331 		cxl_probe_device_regs(dev, base, dev_map);
332 		if (!dev_map->status.valid || !dev_map->mbox.valid ||
333 		    !dev_map->memdev.valid) {
334 			dev_err(dev, "registers not found: %s%s%s\n",
335 				!dev_map->status.valid ? "status " : "",
336 				!dev_map->mbox.valid ? "mbox " : "",
337 				!dev_map->memdev.valid ? "memdev " : "");
338 			return -ENXIO;
339 		}
340 
341 		dev_dbg(dev, "Probing device registers...\n");
342 		break;
343 	default:
344 		break;
345 	}
346 
347 	return 0;
348 }
349 
cxl_map_regs(struct cxl_dev_state * cxlds,struct cxl_register_map * map)350 static int cxl_map_regs(struct cxl_dev_state *cxlds, struct cxl_register_map *map)
351 {
352 	struct device *dev = cxlds->dev;
353 	struct pci_dev *pdev = to_pci_dev(dev);
354 
355 	switch (map->reg_type) {
356 	case CXL_REGLOC_RBI_COMPONENT:
357 		cxl_map_component_regs(pdev, &cxlds->regs.component, map);
358 		dev_dbg(dev, "Mapping component registers...\n");
359 		break;
360 	case CXL_REGLOC_RBI_MEMDEV:
361 		cxl_map_device_regs(pdev, &cxlds->regs.device_regs, map);
362 		dev_dbg(dev, "Probing device registers...\n");
363 		break;
364 	default:
365 		break;
366 	}
367 
368 	return 0;
369 }
370 
cxl_setup_regs(struct pci_dev * pdev,enum cxl_regloc_type type,struct cxl_register_map * map)371 static int cxl_setup_regs(struct pci_dev *pdev, enum cxl_regloc_type type,
372 			  struct cxl_register_map *map)
373 {
374 	int rc;
375 
376 	rc = cxl_find_regblock(pdev, type, map);
377 	if (rc)
378 		return rc;
379 
380 	rc = cxl_map_regblock(pdev, map);
381 	if (rc)
382 		return rc;
383 
384 	rc = cxl_probe_regs(pdev, map);
385 	cxl_unmap_regblock(pdev, map);
386 
387 	return rc;
388 }
389 
cxl_pci_destroy_doe(void * mbs)390 static void cxl_pci_destroy_doe(void *mbs)
391 {
392 	xa_destroy(mbs);
393 }
394 
devm_cxl_pci_create_doe(struct cxl_dev_state * cxlds)395 static void devm_cxl_pci_create_doe(struct cxl_dev_state *cxlds)
396 {
397 	struct device *dev = cxlds->dev;
398 	struct pci_dev *pdev = to_pci_dev(dev);
399 	u16 off = 0;
400 
401 	xa_init(&cxlds->doe_mbs);
402 	if (devm_add_action(&pdev->dev, cxl_pci_destroy_doe, &cxlds->doe_mbs)) {
403 		dev_err(dev, "Failed to create XArray for DOE's\n");
404 		return;
405 	}
406 
407 	/*
408 	 * Mailbox creation is best effort.  Higher layers must determine if
409 	 * the lack of a mailbox for their protocol is a device failure or not.
410 	 */
411 	pci_doe_for_each_off(pdev, off) {
412 		struct pci_doe_mb *doe_mb;
413 
414 		doe_mb = pcim_doe_create_mb(pdev, off);
415 		if (IS_ERR(doe_mb)) {
416 			dev_err(dev, "Failed to create MB object for MB @ %x\n",
417 				off);
418 			continue;
419 		}
420 
421 		if (xa_insert(&cxlds->doe_mbs, off, doe_mb, GFP_KERNEL)) {
422 			dev_err(dev, "xa_insert failed to insert MB @ %x\n",
423 				off);
424 			continue;
425 		}
426 
427 		dev_dbg(dev, "Created DOE mailbox @%x\n", off);
428 	}
429 }
430 
cxl_pci_probe(struct pci_dev * pdev,const struct pci_device_id * id)431 static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
432 {
433 	struct cxl_register_map map;
434 	struct cxl_memdev *cxlmd;
435 	struct cxl_dev_state *cxlds;
436 	int rc;
437 
438 	/*
439 	 * Double check the anonymous union trickery in struct cxl_regs
440 	 * FIXME switch to struct_group()
441 	 */
442 	BUILD_BUG_ON(offsetof(struct cxl_regs, memdev) !=
443 		     offsetof(struct cxl_regs, device_regs.memdev));
444 
445 	rc = pcim_enable_device(pdev);
446 	if (rc)
447 		return rc;
448 
449 	cxlds = cxl_dev_state_create(&pdev->dev);
450 	if (IS_ERR(cxlds))
451 		return PTR_ERR(cxlds);
452 
453 	cxlds->serial = pci_get_dsn(pdev);
454 	cxlds->cxl_dvsec = pci_find_dvsec_capability(
455 		pdev, PCI_DVSEC_VENDOR_ID_CXL, CXL_DVSEC_PCIE_DEVICE);
456 	if (!cxlds->cxl_dvsec)
457 		dev_warn(&pdev->dev,
458 			 "Device DVSEC not present, skip CXL.mem init\n");
459 
460 	rc = cxl_setup_regs(pdev, CXL_REGLOC_RBI_MEMDEV, &map);
461 	if (rc)
462 		return rc;
463 
464 	rc = cxl_map_regs(cxlds, &map);
465 	if (rc)
466 		return rc;
467 
468 	/*
469 	 * If the component registers can't be found, the cxl_pci driver may
470 	 * still be useful for management functions so don't return an error.
471 	 */
472 	cxlds->component_reg_phys = CXL_RESOURCE_NONE;
473 	rc = cxl_setup_regs(pdev, CXL_REGLOC_RBI_COMPONENT, &map);
474 	if (rc)
475 		dev_warn(&pdev->dev, "No component registers (%d)\n", rc);
476 
477 	cxlds->component_reg_phys = cxl_regmap_to_base(pdev, &map);
478 
479 	devm_cxl_pci_create_doe(cxlds);
480 
481 	rc = cxl_pci_setup_mailbox(cxlds);
482 	if (rc)
483 		return rc;
484 
485 	rc = cxl_enumerate_cmds(cxlds);
486 	if (rc)
487 		return rc;
488 
489 	rc = cxl_dev_state_identify(cxlds);
490 	if (rc)
491 		return rc;
492 
493 	rc = cxl_mem_create_range_info(cxlds);
494 	if (rc)
495 		return rc;
496 
497 	cxlmd = devm_cxl_add_memdev(cxlds);
498 	if (IS_ERR(cxlmd))
499 		return PTR_ERR(cxlmd);
500 
501 	if (resource_size(&cxlds->pmem_res) && IS_ENABLED(CONFIG_CXL_PMEM))
502 		rc = devm_cxl_add_nvdimm(&pdev->dev, cxlmd);
503 
504 	return rc;
505 }
506 
507 static const struct pci_device_id cxl_mem_pci_tbl[] = {
508 	/* PCI class code for CXL.mem Type-3 Devices */
509 	{ PCI_DEVICE_CLASS((PCI_CLASS_MEMORY_CXL << 8 | CXL_MEMORY_PROGIF), ~0)},
510 	{ /* terminate list */ },
511 };
512 MODULE_DEVICE_TABLE(pci, cxl_mem_pci_tbl);
513 
514 static struct pci_driver cxl_pci_driver = {
515 	.name			= KBUILD_MODNAME,
516 	.id_table		= cxl_mem_pci_tbl,
517 	.probe			= cxl_pci_probe,
518 	.driver	= {
519 		.probe_type	= PROBE_PREFER_ASYNCHRONOUS,
520 	},
521 };
522 
523 MODULE_LICENSE("GPL v2");
524 module_pci_driver(cxl_pci_driver);
525 MODULE_IMPORT_NS(CXL);
526