xref: /linux/drivers/cxl/pci.c (revision 40d269c000bda9fcd276a0412a9cebd3f6e344c5)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2020 Intel Corporation. All rights reserved. */
3 #include <asm-generic/unaligned.h>
4 #include <linux/io-64-nonatomic-lo-hi.h>
5 #include <linux/moduleparam.h>
6 #include <linux/module.h>
7 #include <linux/delay.h>
8 #include <linux/sizes.h>
9 #include <linux/mutex.h>
10 #include <linux/list.h>
11 #include <linux/pci.h>
12 #include <linux/aer.h>
13 #include <linux/io.h>
14 #include "cxlmem.h"
15 #include "cxlpci.h"
16 #include "cxl.h"
17 #include "pmu.h"
18 
19 /**
20  * DOC: cxl pci
21  *
22  * This implements the PCI exclusive functionality for a CXL device as it is
23  * defined by the Compute Express Link specification. CXL devices may surface
24  * certain functionality even if it isn't CXL enabled. While this driver is
25  * focused around the PCI specific aspects of a CXL device, it binds to the
26  * specific CXL memory device class code, and therefore the implementation of
27  * cxl_pci is focused around CXL memory devices.
28  *
29  * The driver has several responsibilities, mainly:
30  *  - Create the memX device and register on the CXL bus.
31  *  - Enumerate device's register interface and map them.
32  *  - Registers nvdimm bridge device with cxl_core.
33  *  - Registers a CXL mailbox with cxl_core.
34  */
35 
36 #define cxl_doorbell_busy(cxlds)                                                \
37 	(readl((cxlds)->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET) &                  \
38 	 CXLDEV_MBOX_CTRL_DOORBELL)
39 
40 /* CXL 2.0 - 8.2.8.4 */
41 #define CXL_MAILBOX_TIMEOUT_MS (2 * HZ)
42 
43 /*
44  * CXL 2.0 ECN "Add Mailbox Ready Time" defines a capability field to
45  * dictate how long to wait for the mailbox to become ready. The new
46  * field allows the device to tell software the amount of time to wait
47  * before mailbox ready. This field per the spec theoretically allows
48  * for up to 255 seconds. 255 seconds is unreasonably long, its longer
49  * than the maximum SATA port link recovery wait. Default to 60 seconds
50  * until someone builds a CXL device that needs more time in practice.
51  */
52 static unsigned short mbox_ready_timeout = 60;
53 module_param(mbox_ready_timeout, ushort, 0644);
54 MODULE_PARM_DESC(mbox_ready_timeout, "seconds to wait for mailbox ready");
55 
56 static int cxl_pci_mbox_wait_for_doorbell(struct cxl_dev_state *cxlds)
57 {
58 	const unsigned long start = jiffies;
59 	unsigned long end = start;
60 
61 	while (cxl_doorbell_busy(cxlds)) {
62 		end = jiffies;
63 
64 		if (time_after(end, start + CXL_MAILBOX_TIMEOUT_MS)) {
65 			/* Check again in case preempted before timeout test */
66 			if (!cxl_doorbell_busy(cxlds))
67 				break;
68 			return -ETIMEDOUT;
69 		}
70 		cpu_relax();
71 	}
72 
73 	dev_dbg(cxlds->dev, "Doorbell wait took %dms",
74 		jiffies_to_msecs(end) - jiffies_to_msecs(start));
75 	return 0;
76 }
77 
78 #define cxl_err(dev, status, msg)                                        \
79 	dev_err_ratelimited(dev, msg ", device state %s%s\n",                  \
80 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
81 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
82 
83 #define cxl_cmd_err(dev, cmd, status, msg)                               \
84 	dev_err_ratelimited(dev, msg " (opcode: %#x), device state %s%s\n",    \
85 			    (cmd)->opcode,                                     \
86 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
87 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
88 
89 /*
90  * Threaded irq dev_id's must be globally unique.  cxl_dev_id provides a unique
91  * wrapper object for each irq within the same cxlds.
92  */
93 struct cxl_dev_id {
94 	struct cxl_dev_state *cxlds;
95 };
96 
97 static int cxl_request_irq(struct cxl_dev_state *cxlds, int irq,
98 			   irq_handler_t thread_fn)
99 {
100 	struct device *dev = cxlds->dev;
101 	struct cxl_dev_id *dev_id;
102 
103 	dev_id = devm_kzalloc(dev, sizeof(*dev_id), GFP_KERNEL);
104 	if (!dev_id)
105 		return -ENOMEM;
106 	dev_id->cxlds = cxlds;
107 
108 	return devm_request_threaded_irq(dev, irq, NULL, thread_fn,
109 					 IRQF_SHARED | IRQF_ONESHOT, NULL,
110 					 dev_id);
111 }
112 
113 static bool cxl_mbox_background_complete(struct cxl_dev_state *cxlds)
114 {
115 	u64 reg;
116 
117 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
118 	return FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_PCT_MASK, reg) == 100;
119 }
120 
121 static irqreturn_t cxl_pci_mbox_irq(int irq, void *id)
122 {
123 	u64 reg;
124 	u16 opcode;
125 	struct cxl_dev_id *dev_id = id;
126 	struct cxl_dev_state *cxlds = dev_id->cxlds;
127 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
128 
129 	if (!cxl_mbox_background_complete(cxlds))
130 		return IRQ_NONE;
131 
132 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
133 	opcode = FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_OPCODE_MASK, reg);
134 	if (opcode == CXL_MBOX_OP_SANITIZE) {
135 		mutex_lock(&mds->mbox_mutex);
136 		if (mds->security.sanitize_node)
137 			mod_delayed_work(system_wq, &mds->security.poll_dwork, 0);
138 		mutex_unlock(&mds->mbox_mutex);
139 	} else {
140 		/* short-circuit the wait in __cxl_pci_mbox_send_cmd() */
141 		rcuwait_wake_up(&mds->mbox_wait);
142 	}
143 
144 	return IRQ_HANDLED;
145 }
146 
147 /*
148  * Sanitization operation polling mode.
149  */
150 static void cxl_mbox_sanitize_work(struct work_struct *work)
151 {
152 	struct cxl_memdev_state *mds =
153 		container_of(work, typeof(*mds), security.poll_dwork.work);
154 	struct cxl_dev_state *cxlds = &mds->cxlds;
155 
156 	mutex_lock(&mds->mbox_mutex);
157 	if (cxl_mbox_background_complete(cxlds)) {
158 		mds->security.poll_tmo_secs = 0;
159 		if (mds->security.sanitize_node)
160 			sysfs_notify_dirent(mds->security.sanitize_node);
161 		mds->security.sanitize_active = false;
162 
163 		dev_dbg(cxlds->dev, "Sanitization operation ended\n");
164 	} else {
165 		int timeout = mds->security.poll_tmo_secs + 10;
166 
167 		mds->security.poll_tmo_secs = min(15 * 60, timeout);
168 		schedule_delayed_work(&mds->security.poll_dwork, timeout * HZ);
169 	}
170 	mutex_unlock(&mds->mbox_mutex);
171 }
172 
173 /**
174  * __cxl_pci_mbox_send_cmd() - Execute a mailbox command
175  * @mds: The memory device driver data
176  * @mbox_cmd: Command to send to the memory device.
177  *
178  * Context: Any context. Expects mbox_mutex to be held.
179  * Return: -ETIMEDOUT if timeout occurred waiting for completion. 0 on success.
180  *         Caller should check the return code in @mbox_cmd to make sure it
181  *         succeeded.
182  *
183  * This is a generic form of the CXL mailbox send command thus only using the
184  * registers defined by the mailbox capability ID - CXL 2.0 8.2.8.4. Memory
185  * devices, and perhaps other types of CXL devices may have further information
186  * available upon error conditions. Driver facilities wishing to send mailbox
187  * commands should use the wrapper command.
188  *
189  * The CXL spec allows for up to two mailboxes. The intention is for the primary
190  * mailbox to be OS controlled and the secondary mailbox to be used by system
191  * firmware. This allows the OS and firmware to communicate with the device and
192  * not need to coordinate with each other. The driver only uses the primary
193  * mailbox.
194  */
195 static int __cxl_pci_mbox_send_cmd(struct cxl_memdev_state *mds,
196 				   struct cxl_mbox_cmd *mbox_cmd)
197 {
198 	struct cxl_dev_state *cxlds = &mds->cxlds;
199 	void __iomem *payload = cxlds->regs.mbox + CXLDEV_MBOX_PAYLOAD_OFFSET;
200 	struct device *dev = cxlds->dev;
201 	u64 cmd_reg, status_reg;
202 	size_t out_len;
203 	int rc;
204 
205 	lockdep_assert_held(&mds->mbox_mutex);
206 
207 	/*
208 	 * Here are the steps from 8.2.8.4 of the CXL 2.0 spec.
209 	 *   1. Caller reads MB Control Register to verify doorbell is clear
210 	 *   2. Caller writes Command Register
211 	 *   3. Caller writes Command Payload Registers if input payload is non-empty
212 	 *   4. Caller writes MB Control Register to set doorbell
213 	 *   5. Caller either polls for doorbell to be clear or waits for interrupt if configured
214 	 *   6. Caller reads MB Status Register to fetch Return code
215 	 *   7. If command successful, Caller reads Command Register to get Payload Length
216 	 *   8. If output payload is non-empty, host reads Command Payload Registers
217 	 *
218 	 * Hardware is free to do whatever it wants before the doorbell is rung,
219 	 * and isn't allowed to change anything after it clears the doorbell. As
220 	 * such, steps 2 and 3 can happen in any order, and steps 6, 7, 8 can
221 	 * also happen in any order (though some orders might not make sense).
222 	 */
223 
224 	/* #1 */
225 	if (cxl_doorbell_busy(cxlds)) {
226 		u64 md_status =
227 			readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
228 
229 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status,
230 			    "mailbox queue busy");
231 		return -EBUSY;
232 	}
233 
234 	/*
235 	 * With sanitize polling, hardware might be done and the poller still
236 	 * not be in sync. Ensure no new command comes in until so. Keep the
237 	 * hardware semantics and only allow device health status.
238 	 */
239 	if (mds->security.poll_tmo_secs > 0) {
240 		if (mbox_cmd->opcode != CXL_MBOX_OP_GET_HEALTH_INFO)
241 			return -EBUSY;
242 	}
243 
244 	cmd_reg = FIELD_PREP(CXLDEV_MBOX_CMD_COMMAND_OPCODE_MASK,
245 			     mbox_cmd->opcode);
246 	if (mbox_cmd->size_in) {
247 		if (WARN_ON(!mbox_cmd->payload_in))
248 			return -EINVAL;
249 
250 		cmd_reg |= FIELD_PREP(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK,
251 				      mbox_cmd->size_in);
252 		memcpy_toio(payload, mbox_cmd->payload_in, mbox_cmd->size_in);
253 	}
254 
255 	/* #2, #3 */
256 	writeq(cmd_reg, cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
257 
258 	/* #4 */
259 	dev_dbg(dev, "Sending command: 0x%04x\n", mbox_cmd->opcode);
260 	writel(CXLDEV_MBOX_CTRL_DOORBELL,
261 	       cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
262 
263 	/* #5 */
264 	rc = cxl_pci_mbox_wait_for_doorbell(cxlds);
265 	if (rc == -ETIMEDOUT) {
266 		u64 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
267 
268 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status, "mailbox timeout");
269 		return rc;
270 	}
271 
272 	/* #6 */
273 	status_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_STATUS_OFFSET);
274 	mbox_cmd->return_code =
275 		FIELD_GET(CXLDEV_MBOX_STATUS_RET_CODE_MASK, status_reg);
276 
277 	/*
278 	 * Handle the background command in a synchronous manner.
279 	 *
280 	 * All other mailbox commands will serialize/queue on the mbox_mutex,
281 	 * which we currently hold. Furthermore this also guarantees that
282 	 * cxl_mbox_background_complete() checks are safe amongst each other,
283 	 * in that no new bg operation can occur in between.
284 	 *
285 	 * Background operations are timesliced in accordance with the nature
286 	 * of the command. In the event of timeout, the mailbox state is
287 	 * indeterminate until the next successful command submission and the
288 	 * driver can get back in sync with the hardware state.
289 	 */
290 	if (mbox_cmd->return_code == CXL_MBOX_CMD_RC_BACKGROUND) {
291 		u64 bg_status_reg;
292 		int i, timeout;
293 
294 		/*
295 		 * Sanitization is a special case which monopolizes the device
296 		 * and cannot be timesliced. Handle asynchronously instead,
297 		 * and allow userspace to poll(2) for completion.
298 		 */
299 		if (mbox_cmd->opcode == CXL_MBOX_OP_SANITIZE) {
300 			if (mds->security.sanitize_active)
301 				return -EBUSY;
302 
303 			/* give first timeout a second */
304 			timeout = 1;
305 			mds->security.poll_tmo_secs = timeout;
306 			mds->security.sanitize_active = true;
307 			schedule_delayed_work(&mds->security.poll_dwork,
308 					      timeout * HZ);
309 			dev_dbg(dev, "Sanitization operation started\n");
310 			goto success;
311 		}
312 
313 		dev_dbg(dev, "Mailbox background operation (0x%04x) started\n",
314 			mbox_cmd->opcode);
315 
316 		timeout = mbox_cmd->poll_interval_ms;
317 		for (i = 0; i < mbox_cmd->poll_count; i++) {
318 			if (rcuwait_wait_event_timeout(&mds->mbox_wait,
319 				       cxl_mbox_background_complete(cxlds),
320 				       TASK_UNINTERRUPTIBLE,
321 				       msecs_to_jiffies(timeout)) > 0)
322 				break;
323 		}
324 
325 		if (!cxl_mbox_background_complete(cxlds)) {
326 			dev_err(dev, "timeout waiting for background (%d ms)\n",
327 				timeout * mbox_cmd->poll_count);
328 			return -ETIMEDOUT;
329 		}
330 
331 		bg_status_reg = readq(cxlds->regs.mbox +
332 				      CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
333 		mbox_cmd->return_code =
334 			FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_RC_MASK,
335 				  bg_status_reg);
336 		dev_dbg(dev,
337 			"Mailbox background operation (0x%04x) completed\n",
338 			mbox_cmd->opcode);
339 	}
340 
341 	if (mbox_cmd->return_code != CXL_MBOX_CMD_RC_SUCCESS) {
342 		dev_dbg(dev, "Mailbox operation had an error: %s\n",
343 			cxl_mbox_cmd_rc2str(mbox_cmd));
344 		return 0; /* completed but caller must check return_code */
345 	}
346 
347 success:
348 	/* #7 */
349 	cmd_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
350 	out_len = FIELD_GET(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK, cmd_reg);
351 
352 	/* #8 */
353 	if (out_len && mbox_cmd->payload_out) {
354 		/*
355 		 * Sanitize the copy. If hardware misbehaves, out_len per the
356 		 * spec can actually be greater than the max allowed size (21
357 		 * bits available but spec defined 1M max). The caller also may
358 		 * have requested less data than the hardware supplied even
359 		 * within spec.
360 		 */
361 		size_t n;
362 
363 		n = min3(mbox_cmd->size_out, mds->payload_size, out_len);
364 		memcpy_fromio(mbox_cmd->payload_out, payload, n);
365 		mbox_cmd->size_out = n;
366 	} else {
367 		mbox_cmd->size_out = 0;
368 	}
369 
370 	return 0;
371 }
372 
373 static int cxl_pci_mbox_send(struct cxl_memdev_state *mds,
374 			     struct cxl_mbox_cmd *cmd)
375 {
376 	int rc;
377 
378 	mutex_lock_io(&mds->mbox_mutex);
379 	rc = __cxl_pci_mbox_send_cmd(mds, cmd);
380 	mutex_unlock(&mds->mbox_mutex);
381 
382 	return rc;
383 }
384 
385 static int cxl_pci_setup_mailbox(struct cxl_memdev_state *mds, bool irq_avail)
386 {
387 	struct cxl_dev_state *cxlds = &mds->cxlds;
388 	const int cap = readl(cxlds->regs.mbox + CXLDEV_MBOX_CAPS_OFFSET);
389 	struct device *dev = cxlds->dev;
390 	unsigned long timeout;
391 	int irq, msgnum;
392 	u64 md_status;
393 	u32 ctrl;
394 
395 	timeout = jiffies + mbox_ready_timeout * HZ;
396 	do {
397 		md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
398 		if (md_status & CXLMDEV_MBOX_IF_READY)
399 			break;
400 		if (msleep_interruptible(100))
401 			break;
402 	} while (!time_after(jiffies, timeout));
403 
404 	if (!(md_status & CXLMDEV_MBOX_IF_READY)) {
405 		cxl_err(dev, md_status, "timeout awaiting mailbox ready");
406 		return -ETIMEDOUT;
407 	}
408 
409 	/*
410 	 * A command may be in flight from a previous driver instance,
411 	 * think kexec, do one doorbell wait so that
412 	 * __cxl_pci_mbox_send_cmd() can assume that it is the only
413 	 * source for future doorbell busy events.
414 	 */
415 	if (cxl_pci_mbox_wait_for_doorbell(cxlds) != 0) {
416 		cxl_err(dev, md_status, "timeout awaiting mailbox idle");
417 		return -ETIMEDOUT;
418 	}
419 
420 	mds->mbox_send = cxl_pci_mbox_send;
421 	mds->payload_size =
422 		1 << FIELD_GET(CXLDEV_MBOX_CAP_PAYLOAD_SIZE_MASK, cap);
423 
424 	/*
425 	 * CXL 2.0 8.2.8.4.3 Mailbox Capabilities Register
426 	 *
427 	 * If the size is too small, mandatory commands will not work and so
428 	 * there's no point in going forward. If the size is too large, there's
429 	 * no harm is soft limiting it.
430 	 */
431 	mds->payload_size = min_t(size_t, mds->payload_size, SZ_1M);
432 	if (mds->payload_size < 256) {
433 		dev_err(dev, "Mailbox is too small (%zub)",
434 			mds->payload_size);
435 		return -ENXIO;
436 	}
437 
438 	dev_dbg(dev, "Mailbox payload sized %zu", mds->payload_size);
439 
440 	rcuwait_init(&mds->mbox_wait);
441 	INIT_DELAYED_WORK(&mds->security.poll_dwork, cxl_mbox_sanitize_work);
442 
443 	/* background command interrupts are optional */
444 	if (!(cap & CXLDEV_MBOX_CAP_BG_CMD_IRQ) || !irq_avail)
445 		return 0;
446 
447 	msgnum = FIELD_GET(CXLDEV_MBOX_CAP_IRQ_MSGNUM_MASK, cap);
448 	irq = pci_irq_vector(to_pci_dev(cxlds->dev), msgnum);
449 	if (irq < 0)
450 		return 0;
451 
452 	if (cxl_request_irq(cxlds, irq, cxl_pci_mbox_irq))
453 		return 0;
454 
455 	dev_dbg(cxlds->dev, "Mailbox interrupts enabled\n");
456 	/* enable background command mbox irq support */
457 	ctrl = readl(cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
458 	ctrl |= CXLDEV_MBOX_CTRL_BG_CMD_IRQ;
459 	writel(ctrl, cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
460 
461 	return 0;
462 }
463 
464 /*
465  * Assume that any RCIEP that emits the CXL memory expander class code
466  * is an RCD
467  */
468 static bool is_cxl_restricted(struct pci_dev *pdev)
469 {
470 	return pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END;
471 }
472 
473 static int cxl_rcrb_get_comp_regs(struct pci_dev *pdev,
474 				  struct cxl_register_map *map)
475 {
476 	struct cxl_port *port;
477 	struct cxl_dport *dport;
478 	resource_size_t component_reg_phys;
479 
480 	*map = (struct cxl_register_map) {
481 		.host = &pdev->dev,
482 		.resource = CXL_RESOURCE_NONE,
483 	};
484 
485 	port = cxl_pci_find_port(pdev, &dport);
486 	if (!port)
487 		return -EPROBE_DEFER;
488 
489 	component_reg_phys = cxl_rcd_component_reg_phys(&pdev->dev, dport);
490 
491 	put_device(&port->dev);
492 
493 	if (component_reg_phys == CXL_RESOURCE_NONE)
494 		return -ENXIO;
495 
496 	map->resource = component_reg_phys;
497 	map->reg_type = CXL_REGLOC_RBI_COMPONENT;
498 	map->max_size = CXL_COMPONENT_REG_BLOCK_SIZE;
499 
500 	return 0;
501 }
502 
503 static int cxl_pci_setup_regs(struct pci_dev *pdev, enum cxl_regloc_type type,
504 			      struct cxl_register_map *map)
505 {
506 	int rc;
507 
508 	rc = cxl_find_regblock(pdev, type, map);
509 
510 	/*
511 	 * If the Register Locator DVSEC does not exist, check if it
512 	 * is an RCH and try to extract the Component Registers from
513 	 * an RCRB.
514 	 */
515 	if (rc && type == CXL_REGLOC_RBI_COMPONENT && is_cxl_restricted(pdev))
516 		rc = cxl_rcrb_get_comp_regs(pdev, map);
517 
518 	if (rc)
519 		return rc;
520 
521 	return cxl_setup_regs(map);
522 }
523 
524 static int cxl_pci_ras_unmask(struct pci_dev *pdev)
525 {
526 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
527 	void __iomem *addr;
528 	u32 orig_val, val, mask;
529 	u16 cap;
530 	int rc;
531 
532 	if (!cxlds->regs.ras) {
533 		dev_dbg(&pdev->dev, "No RAS registers.\n");
534 		return 0;
535 	}
536 
537 	/* BIOS has PCIe AER error control */
538 	if (!pcie_aer_is_native(pdev))
539 		return 0;
540 
541 	rc = pcie_capability_read_word(pdev, PCI_EXP_DEVCTL, &cap);
542 	if (rc)
543 		return rc;
544 
545 	if (cap & PCI_EXP_DEVCTL_URRE) {
546 		addr = cxlds->regs.ras + CXL_RAS_UNCORRECTABLE_MASK_OFFSET;
547 		orig_val = readl(addr);
548 
549 		mask = CXL_RAS_UNCORRECTABLE_MASK_MASK |
550 		       CXL_RAS_UNCORRECTABLE_MASK_F256B_MASK;
551 		val = orig_val & ~mask;
552 		writel(val, addr);
553 		dev_dbg(&pdev->dev,
554 			"Uncorrectable RAS Errors Mask: %#x -> %#x\n",
555 			orig_val, val);
556 	}
557 
558 	if (cap & PCI_EXP_DEVCTL_CERE) {
559 		addr = cxlds->regs.ras + CXL_RAS_CORRECTABLE_MASK_OFFSET;
560 		orig_val = readl(addr);
561 		val = orig_val & ~CXL_RAS_CORRECTABLE_MASK_MASK;
562 		writel(val, addr);
563 		dev_dbg(&pdev->dev, "Correctable RAS Errors Mask: %#x -> %#x\n",
564 			orig_val, val);
565 	}
566 
567 	return 0;
568 }
569 
570 static void free_event_buf(void *buf)
571 {
572 	kvfree(buf);
573 }
574 
575 /*
576  * There is a single buffer for reading event logs from the mailbox.  All logs
577  * share this buffer protected by the mds->event_log_lock.
578  */
579 static int cxl_mem_alloc_event_buf(struct cxl_memdev_state *mds)
580 {
581 	struct cxl_get_event_payload *buf;
582 
583 	buf = kvmalloc(mds->payload_size, GFP_KERNEL);
584 	if (!buf)
585 		return -ENOMEM;
586 	mds->event.buf = buf;
587 
588 	return devm_add_action_or_reset(mds->cxlds.dev, free_event_buf, buf);
589 }
590 
591 static bool cxl_alloc_irq_vectors(struct pci_dev *pdev)
592 {
593 	int nvecs;
594 
595 	/*
596 	 * Per CXL 3.0 3.1.1 CXL.io Endpoint a function on a CXL device must
597 	 * not generate INTx messages if that function participates in
598 	 * CXL.cache or CXL.mem.
599 	 *
600 	 * Additionally pci_alloc_irq_vectors() handles calling
601 	 * pci_free_irq_vectors() automatically despite not being called
602 	 * pcim_*.  See pci_setup_msi_context().
603 	 */
604 	nvecs = pci_alloc_irq_vectors(pdev, 1, CXL_PCI_DEFAULT_MAX_VECTORS,
605 				      PCI_IRQ_MSIX | PCI_IRQ_MSI);
606 	if (nvecs < 1) {
607 		dev_dbg(&pdev->dev, "Failed to alloc irq vectors: %d\n", nvecs);
608 		return false;
609 	}
610 	return true;
611 }
612 
613 static irqreturn_t cxl_event_thread(int irq, void *id)
614 {
615 	struct cxl_dev_id *dev_id = id;
616 	struct cxl_dev_state *cxlds = dev_id->cxlds;
617 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
618 	u32 status;
619 
620 	do {
621 		/*
622 		 * CXL 3.0 8.2.8.3.1: The lower 32 bits are the status;
623 		 * ignore the reserved upper 32 bits
624 		 */
625 		status = readl(cxlds->regs.status + CXLDEV_DEV_EVENT_STATUS_OFFSET);
626 		/* Ignore logs unknown to the driver */
627 		status &= CXLDEV_EVENT_STATUS_ALL;
628 		if (!status)
629 			break;
630 		cxl_mem_get_event_records(mds, status);
631 		cond_resched();
632 	} while (status);
633 
634 	return IRQ_HANDLED;
635 }
636 
637 static int cxl_event_req_irq(struct cxl_dev_state *cxlds, u8 setting)
638 {
639 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
640 	int irq;
641 
642 	if (FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting) != CXL_INT_MSI_MSIX)
643 		return -ENXIO;
644 
645 	irq =  pci_irq_vector(pdev,
646 			      FIELD_GET(CXLDEV_EVENT_INT_MSGNUM_MASK, setting));
647 	if (irq < 0)
648 		return irq;
649 
650 	return cxl_request_irq(cxlds, irq, cxl_event_thread);
651 }
652 
653 static int cxl_event_get_int_policy(struct cxl_memdev_state *mds,
654 				    struct cxl_event_interrupt_policy *policy)
655 {
656 	struct cxl_mbox_cmd mbox_cmd = {
657 		.opcode = CXL_MBOX_OP_GET_EVT_INT_POLICY,
658 		.payload_out = policy,
659 		.size_out = sizeof(*policy),
660 	};
661 	int rc;
662 
663 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
664 	if (rc < 0)
665 		dev_err(mds->cxlds.dev,
666 			"Failed to get event interrupt policy : %d", rc);
667 
668 	return rc;
669 }
670 
671 static int cxl_event_config_msgnums(struct cxl_memdev_state *mds,
672 				    struct cxl_event_interrupt_policy *policy)
673 {
674 	struct cxl_mbox_cmd mbox_cmd;
675 	int rc;
676 
677 	*policy = (struct cxl_event_interrupt_policy) {
678 		.info_settings = CXL_INT_MSI_MSIX,
679 		.warn_settings = CXL_INT_MSI_MSIX,
680 		.failure_settings = CXL_INT_MSI_MSIX,
681 		.fatal_settings = CXL_INT_MSI_MSIX,
682 	};
683 
684 	mbox_cmd = (struct cxl_mbox_cmd) {
685 		.opcode = CXL_MBOX_OP_SET_EVT_INT_POLICY,
686 		.payload_in = policy,
687 		.size_in = sizeof(*policy),
688 	};
689 
690 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
691 	if (rc < 0) {
692 		dev_err(mds->cxlds.dev, "Failed to set event interrupt policy : %d",
693 			rc);
694 		return rc;
695 	}
696 
697 	/* Retrieve final interrupt settings */
698 	return cxl_event_get_int_policy(mds, policy);
699 }
700 
701 static int cxl_event_irqsetup(struct cxl_memdev_state *mds)
702 {
703 	struct cxl_dev_state *cxlds = &mds->cxlds;
704 	struct cxl_event_interrupt_policy policy;
705 	int rc;
706 
707 	rc = cxl_event_config_msgnums(mds, &policy);
708 	if (rc)
709 		return rc;
710 
711 	rc = cxl_event_req_irq(cxlds, policy.info_settings);
712 	if (rc) {
713 		dev_err(cxlds->dev, "Failed to get interrupt for event Info log\n");
714 		return rc;
715 	}
716 
717 	rc = cxl_event_req_irq(cxlds, policy.warn_settings);
718 	if (rc) {
719 		dev_err(cxlds->dev, "Failed to get interrupt for event Warn log\n");
720 		return rc;
721 	}
722 
723 	rc = cxl_event_req_irq(cxlds, policy.failure_settings);
724 	if (rc) {
725 		dev_err(cxlds->dev, "Failed to get interrupt for event Failure log\n");
726 		return rc;
727 	}
728 
729 	rc = cxl_event_req_irq(cxlds, policy.fatal_settings);
730 	if (rc) {
731 		dev_err(cxlds->dev, "Failed to get interrupt for event Fatal log\n");
732 		return rc;
733 	}
734 
735 	return 0;
736 }
737 
738 static bool cxl_event_int_is_fw(u8 setting)
739 {
740 	u8 mode = FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting);
741 
742 	return mode == CXL_INT_FW;
743 }
744 
745 static int cxl_event_config(struct pci_host_bridge *host_bridge,
746 			    struct cxl_memdev_state *mds, bool irq_avail)
747 {
748 	struct cxl_event_interrupt_policy policy;
749 	int rc;
750 
751 	/*
752 	 * When BIOS maintains CXL error reporting control, it will process
753 	 * event records.  Only one agent can do so.
754 	 */
755 	if (!host_bridge->native_cxl_error)
756 		return 0;
757 
758 	if (!irq_avail) {
759 		dev_info(mds->cxlds.dev, "No interrupt support, disable event processing.\n");
760 		return 0;
761 	}
762 
763 	rc = cxl_mem_alloc_event_buf(mds);
764 	if (rc)
765 		return rc;
766 
767 	rc = cxl_event_get_int_policy(mds, &policy);
768 	if (rc)
769 		return rc;
770 
771 	if (cxl_event_int_is_fw(policy.info_settings) ||
772 	    cxl_event_int_is_fw(policy.warn_settings) ||
773 	    cxl_event_int_is_fw(policy.failure_settings) ||
774 	    cxl_event_int_is_fw(policy.fatal_settings)) {
775 		dev_err(mds->cxlds.dev,
776 			"FW still in control of Event Logs despite _OSC settings\n");
777 		return -EBUSY;
778 	}
779 
780 	rc = cxl_event_irqsetup(mds);
781 	if (rc)
782 		return rc;
783 
784 	cxl_mem_get_event_records(mds, CXLDEV_EVENT_STATUS_ALL);
785 
786 	return 0;
787 }
788 
789 static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
790 {
791 	struct pci_host_bridge *host_bridge = pci_find_host_bridge(pdev->bus);
792 	struct cxl_memdev_state *mds;
793 	struct cxl_dev_state *cxlds;
794 	struct cxl_register_map map;
795 	struct cxl_memdev *cxlmd;
796 	int i, rc, pmu_count;
797 	bool irq_avail;
798 
799 	/*
800 	 * Double check the anonymous union trickery in struct cxl_regs
801 	 * FIXME switch to struct_group()
802 	 */
803 	BUILD_BUG_ON(offsetof(struct cxl_regs, memdev) !=
804 		     offsetof(struct cxl_regs, device_regs.memdev));
805 
806 	rc = pcim_enable_device(pdev);
807 	if (rc)
808 		return rc;
809 	pci_set_master(pdev);
810 
811 	mds = cxl_memdev_state_create(&pdev->dev);
812 	if (IS_ERR(mds))
813 		return PTR_ERR(mds);
814 	cxlds = &mds->cxlds;
815 	pci_set_drvdata(pdev, cxlds);
816 
817 	cxlds->rcd = is_cxl_restricted(pdev);
818 	cxlds->serial = pci_get_dsn(pdev);
819 	cxlds->cxl_dvsec = pci_find_dvsec_capability(
820 		pdev, PCI_DVSEC_VENDOR_ID_CXL, CXL_DVSEC_PCIE_DEVICE);
821 	if (!cxlds->cxl_dvsec)
822 		dev_warn(&pdev->dev,
823 			 "Device DVSEC not present, skip CXL.mem init\n");
824 
825 	rc = cxl_pci_setup_regs(pdev, CXL_REGLOC_RBI_MEMDEV, &map);
826 	if (rc)
827 		return rc;
828 
829 	rc = cxl_map_device_regs(&map, &cxlds->regs.device_regs);
830 	if (rc)
831 		return rc;
832 
833 	/*
834 	 * If the component registers can't be found, the cxl_pci driver may
835 	 * still be useful for management functions so don't return an error.
836 	 */
837 	rc = cxl_pci_setup_regs(pdev, CXL_REGLOC_RBI_COMPONENT,
838 				&cxlds->reg_map);
839 	if (rc)
840 		dev_warn(&pdev->dev, "No component registers (%d)\n", rc);
841 	else if (!cxlds->reg_map.component_map.ras.valid)
842 		dev_dbg(&pdev->dev, "RAS registers not found\n");
843 
844 	rc = cxl_map_component_regs(&cxlds->reg_map, &cxlds->regs.component,
845 				    BIT(CXL_CM_CAP_CAP_ID_RAS));
846 	if (rc)
847 		dev_dbg(&pdev->dev, "Failed to map RAS capability.\n");
848 
849 	rc = cxl_await_media_ready(cxlds);
850 	if (rc == 0)
851 		cxlds->media_ready = true;
852 	else
853 		dev_warn(&pdev->dev, "Media not active (%d)\n", rc);
854 
855 	irq_avail = cxl_alloc_irq_vectors(pdev);
856 
857 	rc = cxl_pci_setup_mailbox(mds, irq_avail);
858 	if (rc)
859 		return rc;
860 
861 	rc = cxl_enumerate_cmds(mds);
862 	if (rc)
863 		return rc;
864 
865 	rc = cxl_set_timestamp(mds);
866 	if (rc)
867 		return rc;
868 
869 	rc = cxl_poison_state_init(mds);
870 	if (rc)
871 		return rc;
872 
873 	rc = cxl_dev_state_identify(mds);
874 	if (rc)
875 		return rc;
876 
877 	rc = cxl_mem_create_range_info(mds);
878 	if (rc)
879 		return rc;
880 
881 	cxlmd = devm_cxl_add_memdev(&pdev->dev, cxlds);
882 	if (IS_ERR(cxlmd))
883 		return PTR_ERR(cxlmd);
884 
885 	rc = devm_cxl_setup_fw_upload(&pdev->dev, mds);
886 	if (rc)
887 		return rc;
888 
889 	rc = devm_cxl_sanitize_setup_notifier(&pdev->dev, cxlmd);
890 	if (rc)
891 		return rc;
892 
893 	pmu_count = cxl_count_regblock(pdev, CXL_REGLOC_RBI_PMU);
894 	for (i = 0; i < pmu_count; i++) {
895 		struct cxl_pmu_regs pmu_regs;
896 
897 		rc = cxl_find_regblock_instance(pdev, CXL_REGLOC_RBI_PMU, &map, i);
898 		if (rc) {
899 			dev_dbg(&pdev->dev, "Could not find PMU regblock\n");
900 			break;
901 		}
902 
903 		rc = cxl_map_pmu_regs(&map, &pmu_regs);
904 		if (rc) {
905 			dev_dbg(&pdev->dev, "Could not map PMU regs\n");
906 			break;
907 		}
908 
909 		rc = devm_cxl_pmu_add(cxlds->dev, &pmu_regs, cxlmd->id, i, CXL_PMU_MEMDEV);
910 		if (rc) {
911 			dev_dbg(&pdev->dev, "Could not add PMU instance\n");
912 			break;
913 		}
914 	}
915 
916 	rc = cxl_event_config(host_bridge, mds, irq_avail);
917 	if (rc)
918 		return rc;
919 
920 	rc = cxl_pci_ras_unmask(pdev);
921 	if (rc)
922 		dev_dbg(&pdev->dev, "No RAS reporting unmasked\n");
923 
924 	pci_save_state(pdev);
925 
926 	return rc;
927 }
928 
929 static const struct pci_device_id cxl_mem_pci_tbl[] = {
930 	/* PCI class code for CXL.mem Type-3 Devices */
931 	{ PCI_DEVICE_CLASS((PCI_CLASS_MEMORY_CXL << 8 | CXL_MEMORY_PROGIF), ~0)},
932 	{ /* terminate list */ },
933 };
934 MODULE_DEVICE_TABLE(pci, cxl_mem_pci_tbl);
935 
936 static pci_ers_result_t cxl_slot_reset(struct pci_dev *pdev)
937 {
938 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
939 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
940 	struct device *dev = &cxlmd->dev;
941 
942 	dev_info(&pdev->dev, "%s: restart CXL.mem after slot reset\n",
943 		 dev_name(dev));
944 	pci_restore_state(pdev);
945 	if (device_attach(dev) <= 0)
946 		return PCI_ERS_RESULT_DISCONNECT;
947 	return PCI_ERS_RESULT_RECOVERED;
948 }
949 
950 static void cxl_error_resume(struct pci_dev *pdev)
951 {
952 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
953 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
954 	struct device *dev = &cxlmd->dev;
955 
956 	dev_info(&pdev->dev, "%s: error resume %s\n", dev_name(dev),
957 		 dev->driver ? "successful" : "failed");
958 }
959 
960 static const struct pci_error_handlers cxl_error_handlers = {
961 	.error_detected	= cxl_error_detected,
962 	.slot_reset	= cxl_slot_reset,
963 	.resume		= cxl_error_resume,
964 	.cor_error_detected	= cxl_cor_error_detected,
965 };
966 
967 static struct pci_driver cxl_pci_driver = {
968 	.name			= KBUILD_MODNAME,
969 	.id_table		= cxl_mem_pci_tbl,
970 	.probe			= cxl_pci_probe,
971 	.err_handler		= &cxl_error_handlers,
972 	.driver	= {
973 		.probe_type	= PROBE_PREFER_ASYNCHRONOUS,
974 	},
975 };
976 
977 module_pci_driver(cxl_pci_driver);
978 MODULE_LICENSE("GPL v2");
979 MODULE_IMPORT_NS(CXL);
980