xref: /openbmc/linux/drivers/cxl/pci.c (revision 18afb028)
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/aer.h>
12 #include <linux/io.h>
13 #include "cxlmem.h"
14 #include "cxlpci.h"
15 #include "cxl.h"
16 #include "pmu.h"
17 
18 /**
19  * DOC: cxl pci
20  *
21  * This implements the PCI exclusive functionality for a CXL device as it is
22  * defined by the Compute Express Link specification. CXL devices may surface
23  * certain functionality even if it isn't CXL enabled. While this driver is
24  * focused around the PCI specific aspects of a CXL device, it binds to the
25  * specific CXL memory device class code, and therefore the implementation of
26  * cxl_pci is focused around CXL memory devices.
27  *
28  * The driver has several responsibilities, mainly:
29  *  - Create the memX device and register on the CXL bus.
30  *  - Enumerate device's register interface and map them.
31  *  - Registers nvdimm bridge device with cxl_core.
32  *  - Registers a CXL mailbox with cxl_core.
33  */
34 
35 #define cxl_doorbell_busy(cxlds)                                                \
36 	(readl((cxlds)->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET) &                  \
37 	 CXLDEV_MBOX_CTRL_DOORBELL)
38 
39 /* CXL 2.0 - 8.2.8.4 */
40 #define CXL_MAILBOX_TIMEOUT_MS (2 * HZ)
41 
42 /*
43  * CXL 2.0 ECN "Add Mailbox Ready Time" defines a capability field to
44  * dictate how long to wait for the mailbox to become ready. The new
45  * field allows the device to tell software the amount of time to wait
46  * before mailbox ready. This field per the spec theoretically allows
47  * for up to 255 seconds. 255 seconds is unreasonably long, its longer
48  * than the maximum SATA port link recovery wait. Default to 60 seconds
49  * until someone builds a CXL device that needs more time in practice.
50  */
51 static unsigned short mbox_ready_timeout = 60;
52 module_param(mbox_ready_timeout, ushort, 0644);
53 MODULE_PARM_DESC(mbox_ready_timeout, "seconds to wait for mailbox ready");
54 
55 static int cxl_pci_mbox_wait_for_doorbell(struct cxl_dev_state *cxlds)
56 {
57 	const unsigned long start = jiffies;
58 	unsigned long end = start;
59 
60 	while (cxl_doorbell_busy(cxlds)) {
61 		end = jiffies;
62 
63 		if (time_after(end, start + CXL_MAILBOX_TIMEOUT_MS)) {
64 			/* Check again in case preempted before timeout test */
65 			if (!cxl_doorbell_busy(cxlds))
66 				break;
67 			return -ETIMEDOUT;
68 		}
69 		cpu_relax();
70 	}
71 
72 	dev_dbg(cxlds->dev, "Doorbell wait took %dms",
73 		jiffies_to_msecs(end) - jiffies_to_msecs(start));
74 	return 0;
75 }
76 
77 #define cxl_err(dev, status, msg)                                        \
78 	dev_err_ratelimited(dev, msg ", device state %s%s\n",                  \
79 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
80 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
81 
82 #define cxl_cmd_err(dev, cmd, status, msg)                               \
83 	dev_err_ratelimited(dev, msg " (opcode: %#x), device state %s%s\n",    \
84 			    (cmd)->opcode,                                     \
85 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
86 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
87 
88 struct cxl_dev_id {
89 	struct cxl_dev_state *cxlds;
90 };
91 
92 static int cxl_request_irq(struct cxl_dev_state *cxlds, int irq,
93 			   irq_handler_t handler, irq_handler_t thread_fn)
94 {
95 	struct device *dev = cxlds->dev;
96 	struct cxl_dev_id *dev_id;
97 
98 	/* dev_id must be globally unique and must contain the cxlds */
99 	dev_id = devm_kzalloc(dev, sizeof(*dev_id), GFP_KERNEL);
100 	if (!dev_id)
101 		return -ENOMEM;
102 	dev_id->cxlds = cxlds;
103 
104 	return devm_request_threaded_irq(dev, irq, handler, thread_fn,
105 					 IRQF_SHARED | IRQF_ONESHOT,
106 					 NULL, dev_id);
107 }
108 
109 static bool cxl_mbox_background_complete(struct cxl_dev_state *cxlds)
110 {
111 	u64 reg;
112 
113 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
114 	return FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_PCT_MASK, reg) == 100;
115 }
116 
117 static irqreturn_t cxl_pci_mbox_irq(int irq, void *id)
118 {
119 	u64 reg;
120 	u16 opcode;
121 	struct cxl_dev_id *dev_id = id;
122 	struct cxl_dev_state *cxlds = dev_id->cxlds;
123 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
124 
125 	if (!cxl_mbox_background_complete(cxlds))
126 		return IRQ_NONE;
127 
128 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
129 	opcode = FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_OPCODE_MASK, reg);
130 	if (opcode == CXL_MBOX_OP_SANITIZE) {
131 		if (mds->security.sanitize_node)
132 			sysfs_notify_dirent(mds->security.sanitize_node);
133 
134 		dev_dbg(cxlds->dev, "Sanitization operation ended\n");
135 	} else {
136 		/* short-circuit the wait in __cxl_pci_mbox_send_cmd() */
137 		rcuwait_wake_up(&mds->mbox_wait);
138 	}
139 
140 	return IRQ_HANDLED;
141 }
142 
143 /*
144  * Sanitization operation polling mode.
145  */
146 static void cxl_mbox_sanitize_work(struct work_struct *work)
147 {
148 	struct cxl_memdev_state *mds =
149 		container_of(work, typeof(*mds), security.poll_dwork.work);
150 	struct cxl_dev_state *cxlds = &mds->cxlds;
151 
152 	mutex_lock(&mds->mbox_mutex);
153 	if (cxl_mbox_background_complete(cxlds)) {
154 		mds->security.poll_tmo_secs = 0;
155 		put_device(cxlds->dev);
156 
157 		if (mds->security.sanitize_node)
158 			sysfs_notify_dirent(mds->security.sanitize_node);
159 
160 		dev_dbg(cxlds->dev, "Sanitization operation ended\n");
161 	} else {
162 		int timeout = mds->security.poll_tmo_secs + 10;
163 
164 		mds->security.poll_tmo_secs = min(15 * 60, timeout);
165 		queue_delayed_work(system_wq, &mds->security.poll_dwork,
166 				   timeout * HZ);
167 	}
168 	mutex_unlock(&mds->mbox_mutex);
169 }
170 
171 /**
172  * __cxl_pci_mbox_send_cmd() - Execute a mailbox command
173  * @mds: The memory device driver data
174  * @mbox_cmd: Command to send to the memory device.
175  *
176  * Context: Any context. Expects mbox_mutex to be held.
177  * Return: -ETIMEDOUT if timeout occurred waiting for completion. 0 on success.
178  *         Caller should check the return code in @mbox_cmd to make sure it
179  *         succeeded.
180  *
181  * This is a generic form of the CXL mailbox send command thus only using the
182  * registers defined by the mailbox capability ID - CXL 2.0 8.2.8.4. Memory
183  * devices, and perhaps other types of CXL devices may have further information
184  * available upon error conditions. Driver facilities wishing to send mailbox
185  * commands should use the wrapper command.
186  *
187  * The CXL spec allows for up to two mailboxes. The intention is for the primary
188  * mailbox to be OS controlled and the secondary mailbox to be used by system
189  * firmware. This allows the OS and firmware to communicate with the device and
190  * not need to coordinate with each other. The driver only uses the primary
191  * mailbox.
192  */
193 static int __cxl_pci_mbox_send_cmd(struct cxl_memdev_state *mds,
194 				   struct cxl_mbox_cmd *mbox_cmd)
195 {
196 	struct cxl_dev_state *cxlds = &mds->cxlds;
197 	void __iomem *payload = cxlds->regs.mbox + CXLDEV_MBOX_PAYLOAD_OFFSET;
198 	struct device *dev = cxlds->dev;
199 	u64 cmd_reg, status_reg;
200 	size_t out_len;
201 	int rc;
202 
203 	lockdep_assert_held(&mds->mbox_mutex);
204 
205 	/*
206 	 * Here are the steps from 8.2.8.4 of the CXL 2.0 spec.
207 	 *   1. Caller reads MB Control Register to verify doorbell is clear
208 	 *   2. Caller writes Command Register
209 	 *   3. Caller writes Command Payload Registers if input payload is non-empty
210 	 *   4. Caller writes MB Control Register to set doorbell
211 	 *   5. Caller either polls for doorbell to be clear or waits for interrupt if configured
212 	 *   6. Caller reads MB Status Register to fetch Return code
213 	 *   7. If command successful, Caller reads Command Register to get Payload Length
214 	 *   8. If output payload is non-empty, host reads Command Payload Registers
215 	 *
216 	 * Hardware is free to do whatever it wants before the doorbell is rung,
217 	 * and isn't allowed to change anything after it clears the doorbell. As
218 	 * such, steps 2 and 3 can happen in any order, and steps 6, 7, 8 can
219 	 * also happen in any order (though some orders might not make sense).
220 	 */
221 
222 	/* #1 */
223 	if (cxl_doorbell_busy(cxlds)) {
224 		u64 md_status =
225 			readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
226 
227 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status,
228 			    "mailbox queue busy");
229 		return -EBUSY;
230 	}
231 
232 	/*
233 	 * With sanitize polling, hardware might be done and the poller still
234 	 * not be in sync. Ensure no new command comes in until so. Keep the
235 	 * hardware semantics and only allow device health status.
236 	 */
237 	if (mds->security.poll_tmo_secs > 0) {
238 		if (mbox_cmd->opcode != CXL_MBOX_OP_GET_HEALTH_INFO)
239 			return -EBUSY;
240 	}
241 
242 	cmd_reg = FIELD_PREP(CXLDEV_MBOX_CMD_COMMAND_OPCODE_MASK,
243 			     mbox_cmd->opcode);
244 	if (mbox_cmd->size_in) {
245 		if (WARN_ON(!mbox_cmd->payload_in))
246 			return -EINVAL;
247 
248 		cmd_reg |= FIELD_PREP(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK,
249 				      mbox_cmd->size_in);
250 		memcpy_toio(payload, mbox_cmd->payload_in, mbox_cmd->size_in);
251 	}
252 
253 	/* #2, #3 */
254 	writeq(cmd_reg, cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
255 
256 	/* #4 */
257 	dev_dbg(dev, "Sending command: 0x%04x\n", mbox_cmd->opcode);
258 	writel(CXLDEV_MBOX_CTRL_DOORBELL,
259 	       cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
260 
261 	/* #5 */
262 	rc = cxl_pci_mbox_wait_for_doorbell(cxlds);
263 	if (rc == -ETIMEDOUT) {
264 		u64 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
265 
266 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status, "mailbox timeout");
267 		return rc;
268 	}
269 
270 	/* #6 */
271 	status_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_STATUS_OFFSET);
272 	mbox_cmd->return_code =
273 		FIELD_GET(CXLDEV_MBOX_STATUS_RET_CODE_MASK, status_reg);
274 
275 	/*
276 	 * Handle the background command in a synchronous manner.
277 	 *
278 	 * All other mailbox commands will serialize/queue on the mbox_mutex,
279 	 * which we currently hold. Furthermore this also guarantees that
280 	 * cxl_mbox_background_complete() checks are safe amongst each other,
281 	 * in that no new bg operation can occur in between.
282 	 *
283 	 * Background operations are timesliced in accordance with the nature
284 	 * of the command. In the event of timeout, the mailbox state is
285 	 * indeterminate until the next successful command submission and the
286 	 * driver can get back in sync with the hardware state.
287 	 */
288 	if (mbox_cmd->return_code == CXL_MBOX_CMD_RC_BACKGROUND) {
289 		u64 bg_status_reg;
290 		int i, timeout;
291 
292 		/*
293 		 * Sanitization is a special case which monopolizes the device
294 		 * and cannot be timesliced. Handle asynchronously instead,
295 		 * and allow userspace to poll(2) for completion.
296 		 */
297 		if (mbox_cmd->opcode == CXL_MBOX_OP_SANITIZE) {
298 			if (mds->security.poll) {
299 				/* hold the device throughout */
300 				get_device(cxlds->dev);
301 
302 				/* give first timeout a second */
303 				timeout = 1;
304 				mds->security.poll_tmo_secs = timeout;
305 				queue_delayed_work(system_wq,
306 						   &mds->security.poll_dwork,
307 						   timeout * HZ);
308 			}
309 
310 			dev_dbg(dev, "Sanitization operation started\n");
311 			goto success;
312 		}
313 
314 		dev_dbg(dev, "Mailbox background operation (0x%04x) started\n",
315 			mbox_cmd->opcode);
316 
317 		timeout = mbox_cmd->poll_interval_ms;
318 		for (i = 0; i < mbox_cmd->poll_count; i++) {
319 			if (rcuwait_wait_event_timeout(&mds->mbox_wait,
320 				       cxl_mbox_background_complete(cxlds),
321 				       TASK_UNINTERRUPTIBLE,
322 				       msecs_to_jiffies(timeout)) > 0)
323 				break;
324 		}
325 
326 		if (!cxl_mbox_background_complete(cxlds)) {
327 			dev_err(dev, "timeout waiting for background (%d ms)\n",
328 				timeout * mbox_cmd->poll_count);
329 			return -ETIMEDOUT;
330 		}
331 
332 		bg_status_reg = readq(cxlds->regs.mbox +
333 				      CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
334 		mbox_cmd->return_code =
335 			FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_RC_MASK,
336 				  bg_status_reg);
337 		dev_dbg(dev,
338 			"Mailbox background operation (0x%04x) completed\n",
339 			mbox_cmd->opcode);
340 	}
341 
342 	if (mbox_cmd->return_code != CXL_MBOX_CMD_RC_SUCCESS) {
343 		dev_dbg(dev, "Mailbox operation had an error: %s\n",
344 			cxl_mbox_cmd_rc2str(mbox_cmd));
345 		return 0; /* completed but caller must check return_code */
346 	}
347 
348 success:
349 	/* #7 */
350 	cmd_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
351 	out_len = FIELD_GET(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK, cmd_reg);
352 
353 	/* #8 */
354 	if (out_len && mbox_cmd->payload_out) {
355 		/*
356 		 * Sanitize the copy. If hardware misbehaves, out_len per the
357 		 * spec can actually be greater than the max allowed size (21
358 		 * bits available but spec defined 1M max). The caller also may
359 		 * have requested less data than the hardware supplied even
360 		 * within spec.
361 		 */
362 		size_t n;
363 
364 		n = min3(mbox_cmd->size_out, mds->payload_size, out_len);
365 		memcpy_fromio(mbox_cmd->payload_out, payload, n);
366 		mbox_cmd->size_out = n;
367 	} else {
368 		mbox_cmd->size_out = 0;
369 	}
370 
371 	return 0;
372 }
373 
374 static int cxl_pci_mbox_send(struct cxl_memdev_state *mds,
375 			     struct cxl_mbox_cmd *cmd)
376 {
377 	int rc;
378 
379 	mutex_lock_io(&mds->mbox_mutex);
380 	rc = __cxl_pci_mbox_send_cmd(mds, cmd);
381 	mutex_unlock(&mds->mbox_mutex);
382 
383 	return rc;
384 }
385 
386 static int cxl_pci_setup_mailbox(struct cxl_memdev_state *mds)
387 {
388 	struct cxl_dev_state *cxlds = &mds->cxlds;
389 	const int cap = readl(cxlds->regs.mbox + CXLDEV_MBOX_CAPS_OFFSET);
390 	struct device *dev = cxlds->dev;
391 	unsigned long timeout;
392 	u64 md_status;
393 
394 	timeout = jiffies + mbox_ready_timeout * HZ;
395 	do {
396 		md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
397 		if (md_status & CXLMDEV_MBOX_IF_READY)
398 			break;
399 		if (msleep_interruptible(100))
400 			break;
401 	} while (!time_after(jiffies, timeout));
402 
403 	if (!(md_status & CXLMDEV_MBOX_IF_READY)) {
404 		cxl_err(dev, md_status, "timeout awaiting mailbox ready");
405 		return -ETIMEDOUT;
406 	}
407 
408 	/*
409 	 * A command may be in flight from a previous driver instance,
410 	 * think kexec, do one doorbell wait so that
411 	 * __cxl_pci_mbox_send_cmd() can assume that it is the only
412 	 * source for future doorbell busy events.
413 	 */
414 	if (cxl_pci_mbox_wait_for_doorbell(cxlds) != 0) {
415 		cxl_err(dev, md_status, "timeout awaiting mailbox idle");
416 		return -ETIMEDOUT;
417 	}
418 
419 	mds->mbox_send = cxl_pci_mbox_send;
420 	mds->payload_size =
421 		1 << FIELD_GET(CXLDEV_MBOX_CAP_PAYLOAD_SIZE_MASK, cap);
422 
423 	/*
424 	 * CXL 2.0 8.2.8.4.3 Mailbox Capabilities Register
425 	 *
426 	 * If the size is too small, mandatory commands will not work and so
427 	 * there's no point in going forward. If the size is too large, there's
428 	 * no harm is soft limiting it.
429 	 */
430 	mds->payload_size = min_t(size_t, mds->payload_size, SZ_1M);
431 	if (mds->payload_size < 256) {
432 		dev_err(dev, "Mailbox is too small (%zub)",
433 			mds->payload_size);
434 		return -ENXIO;
435 	}
436 
437 	dev_dbg(dev, "Mailbox payload sized %zu", mds->payload_size);
438 
439 	rcuwait_init(&mds->mbox_wait);
440 
441 	if (cap & CXLDEV_MBOX_CAP_BG_CMD_IRQ) {
442 		u32 ctrl;
443 		int irq, msgnum;
444 		struct pci_dev *pdev = to_pci_dev(cxlds->dev);
445 
446 		msgnum = FIELD_GET(CXLDEV_MBOX_CAP_IRQ_MSGNUM_MASK, cap);
447 		irq = pci_irq_vector(pdev, msgnum);
448 		if (irq < 0)
449 			goto mbox_poll;
450 
451 		if (cxl_request_irq(cxlds, irq, cxl_pci_mbox_irq, NULL))
452 			goto mbox_poll;
453 
454 		/* enable background command mbox irq support */
455 		ctrl = readl(cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
456 		ctrl |= CXLDEV_MBOX_CTRL_BG_CMD_IRQ;
457 		writel(ctrl, cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
458 
459 		return 0;
460 	}
461 
462 mbox_poll:
463 	mds->security.poll = true;
464 	INIT_DELAYED_WORK(&mds->security.poll_dwork, cxl_mbox_sanitize_work);
465 
466 	dev_dbg(cxlds->dev, "Mailbox interrupts are unsupported");
467 	return 0;
468 }
469 
470 /*
471  * Assume that any RCIEP that emits the CXL memory expander class code
472  * is an RCD
473  */
474 static bool is_cxl_restricted(struct pci_dev *pdev)
475 {
476 	return pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END;
477 }
478 
479 static int cxl_rcrb_get_comp_regs(struct pci_dev *pdev,
480 				  struct cxl_register_map *map)
481 {
482 	struct cxl_port *port;
483 	struct cxl_dport *dport;
484 	resource_size_t component_reg_phys;
485 
486 	*map = (struct cxl_register_map) {
487 		.dev = &pdev->dev,
488 		.resource = CXL_RESOURCE_NONE,
489 	};
490 
491 	port = cxl_pci_find_port(pdev, &dport);
492 	if (!port)
493 		return -EPROBE_DEFER;
494 
495 	component_reg_phys = cxl_rcd_component_reg_phys(&pdev->dev, dport);
496 
497 	put_device(&port->dev);
498 
499 	if (component_reg_phys == CXL_RESOURCE_NONE)
500 		return -ENXIO;
501 
502 	map->resource = component_reg_phys;
503 	map->reg_type = CXL_REGLOC_RBI_COMPONENT;
504 	map->max_size = CXL_COMPONENT_REG_BLOCK_SIZE;
505 
506 	return 0;
507 }
508 
509 static int cxl_pci_setup_regs(struct pci_dev *pdev, enum cxl_regloc_type type,
510 			      struct cxl_register_map *map)
511 {
512 	int rc;
513 
514 	rc = cxl_find_regblock(pdev, type, map);
515 
516 	/*
517 	 * If the Register Locator DVSEC does not exist, check if it
518 	 * is an RCH and try to extract the Component Registers from
519 	 * an RCRB.
520 	 */
521 	if (rc && type == CXL_REGLOC_RBI_COMPONENT && is_cxl_restricted(pdev))
522 		rc = cxl_rcrb_get_comp_regs(pdev, map);
523 
524 	if (rc)
525 		return rc;
526 
527 	return cxl_setup_regs(map);
528 }
529 
530 static int cxl_pci_ras_unmask(struct pci_dev *pdev)
531 {
532 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
533 	void __iomem *addr;
534 	u32 orig_val, val, mask;
535 	u16 cap;
536 	int rc;
537 
538 	if (!cxlds->regs.ras) {
539 		dev_dbg(&pdev->dev, "No RAS registers.\n");
540 		return 0;
541 	}
542 
543 	/* BIOS has PCIe AER error control */
544 	if (!pcie_aer_is_native(pdev))
545 		return 0;
546 
547 	rc = pcie_capability_read_word(pdev, PCI_EXP_DEVCTL, &cap);
548 	if (rc)
549 		return rc;
550 
551 	if (cap & PCI_EXP_DEVCTL_URRE) {
552 		addr = cxlds->regs.ras + CXL_RAS_UNCORRECTABLE_MASK_OFFSET;
553 		orig_val = readl(addr);
554 
555 		mask = CXL_RAS_UNCORRECTABLE_MASK_MASK |
556 		       CXL_RAS_UNCORRECTABLE_MASK_F256B_MASK;
557 		val = orig_val & ~mask;
558 		writel(val, addr);
559 		dev_dbg(&pdev->dev,
560 			"Uncorrectable RAS Errors Mask: %#x -> %#x\n",
561 			orig_val, val);
562 	}
563 
564 	if (cap & PCI_EXP_DEVCTL_CERE) {
565 		addr = cxlds->regs.ras + CXL_RAS_CORRECTABLE_MASK_OFFSET;
566 		orig_val = readl(addr);
567 		val = orig_val & ~CXL_RAS_CORRECTABLE_MASK_MASK;
568 		writel(val, addr);
569 		dev_dbg(&pdev->dev, "Correctable RAS Errors Mask: %#x -> %#x\n",
570 			orig_val, val);
571 	}
572 
573 	return 0;
574 }
575 
576 static void free_event_buf(void *buf)
577 {
578 	kvfree(buf);
579 }
580 
581 /*
582  * There is a single buffer for reading event logs from the mailbox.  All logs
583  * share this buffer protected by the mds->event_log_lock.
584  */
585 static int cxl_mem_alloc_event_buf(struct cxl_memdev_state *mds)
586 {
587 	struct cxl_get_event_payload *buf;
588 
589 	buf = kvmalloc(mds->payload_size, GFP_KERNEL);
590 	if (!buf)
591 		return -ENOMEM;
592 	mds->event.buf = buf;
593 
594 	return devm_add_action_or_reset(mds->cxlds.dev, free_event_buf, buf);
595 }
596 
597 static int cxl_alloc_irq_vectors(struct pci_dev *pdev)
598 {
599 	int nvecs;
600 
601 	/*
602 	 * Per CXL 3.0 3.1.1 CXL.io Endpoint a function on a CXL device must
603 	 * not generate INTx messages if that function participates in
604 	 * CXL.cache or CXL.mem.
605 	 *
606 	 * Additionally pci_alloc_irq_vectors() handles calling
607 	 * pci_free_irq_vectors() automatically despite not being called
608 	 * pcim_*.  See pci_setup_msi_context().
609 	 */
610 	nvecs = pci_alloc_irq_vectors(pdev, 1, CXL_PCI_DEFAULT_MAX_VECTORS,
611 				      PCI_IRQ_MSIX | PCI_IRQ_MSI);
612 	if (nvecs < 1) {
613 		dev_dbg(&pdev->dev, "Failed to alloc irq vectors: %d\n", nvecs);
614 		return -ENXIO;
615 	}
616 	return 0;
617 }
618 
619 static irqreturn_t cxl_event_thread(int irq, void *id)
620 {
621 	struct cxl_dev_id *dev_id = id;
622 	struct cxl_dev_state *cxlds = dev_id->cxlds;
623 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
624 	u32 status;
625 
626 	do {
627 		/*
628 		 * CXL 3.0 8.2.8.3.1: The lower 32 bits are the status;
629 		 * ignore the reserved upper 32 bits
630 		 */
631 		status = readl(cxlds->regs.status + CXLDEV_DEV_EVENT_STATUS_OFFSET);
632 		/* Ignore logs unknown to the driver */
633 		status &= CXLDEV_EVENT_STATUS_ALL;
634 		if (!status)
635 			break;
636 		cxl_mem_get_event_records(mds, status);
637 		cond_resched();
638 	} while (status);
639 
640 	return IRQ_HANDLED;
641 }
642 
643 static int cxl_event_req_irq(struct cxl_dev_state *cxlds, u8 setting)
644 {
645 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
646 	int irq;
647 
648 	if (FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting) != CXL_INT_MSI_MSIX)
649 		return -ENXIO;
650 
651 	irq =  pci_irq_vector(pdev,
652 			      FIELD_GET(CXLDEV_EVENT_INT_MSGNUM_MASK, setting));
653 	if (irq < 0)
654 		return irq;
655 
656 	return cxl_request_irq(cxlds, irq, NULL, cxl_event_thread);
657 }
658 
659 static int cxl_event_get_int_policy(struct cxl_memdev_state *mds,
660 				    struct cxl_event_interrupt_policy *policy)
661 {
662 	struct cxl_mbox_cmd mbox_cmd = {
663 		.opcode = CXL_MBOX_OP_GET_EVT_INT_POLICY,
664 		.payload_out = policy,
665 		.size_out = sizeof(*policy),
666 	};
667 	int rc;
668 
669 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
670 	if (rc < 0)
671 		dev_err(mds->cxlds.dev,
672 			"Failed to get event interrupt policy : %d", rc);
673 
674 	return rc;
675 }
676 
677 static int cxl_event_config_msgnums(struct cxl_memdev_state *mds,
678 				    struct cxl_event_interrupt_policy *policy)
679 {
680 	struct cxl_mbox_cmd mbox_cmd;
681 	int rc;
682 
683 	*policy = (struct cxl_event_interrupt_policy) {
684 		.info_settings = CXL_INT_MSI_MSIX,
685 		.warn_settings = CXL_INT_MSI_MSIX,
686 		.failure_settings = CXL_INT_MSI_MSIX,
687 		.fatal_settings = CXL_INT_MSI_MSIX,
688 	};
689 
690 	mbox_cmd = (struct cxl_mbox_cmd) {
691 		.opcode = CXL_MBOX_OP_SET_EVT_INT_POLICY,
692 		.payload_in = policy,
693 		.size_in = sizeof(*policy),
694 	};
695 
696 	rc = cxl_internal_send_cmd(mds, &mbox_cmd);
697 	if (rc < 0) {
698 		dev_err(mds->cxlds.dev, "Failed to set event interrupt policy : %d",
699 			rc);
700 		return rc;
701 	}
702 
703 	/* Retrieve final interrupt settings */
704 	return cxl_event_get_int_policy(mds, policy);
705 }
706 
707 static int cxl_event_irqsetup(struct cxl_memdev_state *mds)
708 {
709 	struct cxl_dev_state *cxlds = &mds->cxlds;
710 	struct cxl_event_interrupt_policy policy;
711 	int rc;
712 
713 	rc = cxl_event_config_msgnums(mds, &policy);
714 	if (rc)
715 		return rc;
716 
717 	rc = cxl_event_req_irq(cxlds, policy.info_settings);
718 	if (rc) {
719 		dev_err(cxlds->dev, "Failed to get interrupt for event Info log\n");
720 		return rc;
721 	}
722 
723 	rc = cxl_event_req_irq(cxlds, policy.warn_settings);
724 	if (rc) {
725 		dev_err(cxlds->dev, "Failed to get interrupt for event Warn log\n");
726 		return rc;
727 	}
728 
729 	rc = cxl_event_req_irq(cxlds, policy.failure_settings);
730 	if (rc) {
731 		dev_err(cxlds->dev, "Failed to get interrupt for event Failure log\n");
732 		return rc;
733 	}
734 
735 	rc = cxl_event_req_irq(cxlds, policy.fatal_settings);
736 	if (rc) {
737 		dev_err(cxlds->dev, "Failed to get interrupt for event Fatal log\n");
738 		return rc;
739 	}
740 
741 	return 0;
742 }
743 
744 static bool cxl_event_int_is_fw(u8 setting)
745 {
746 	u8 mode = FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting);
747 
748 	return mode == CXL_INT_FW;
749 }
750 
751 static int cxl_event_config(struct pci_host_bridge *host_bridge,
752 			    struct cxl_memdev_state *mds)
753 {
754 	struct cxl_event_interrupt_policy policy;
755 	int rc;
756 
757 	/*
758 	 * When BIOS maintains CXL error reporting control, it will process
759 	 * event records.  Only one agent can do so.
760 	 */
761 	if (!host_bridge->native_cxl_error)
762 		return 0;
763 
764 	rc = cxl_mem_alloc_event_buf(mds);
765 	if (rc)
766 		return rc;
767 
768 	rc = cxl_event_get_int_policy(mds, &policy);
769 	if (rc)
770 		return rc;
771 
772 	if (cxl_event_int_is_fw(policy.info_settings) ||
773 	    cxl_event_int_is_fw(policy.warn_settings) ||
774 	    cxl_event_int_is_fw(policy.failure_settings) ||
775 	    cxl_event_int_is_fw(policy.fatal_settings)) {
776 		dev_err(mds->cxlds.dev,
777 			"FW still in control of Event Logs despite _OSC settings\n");
778 		return -EBUSY;
779 	}
780 
781 	rc = cxl_event_irqsetup(mds);
782 	if (rc)
783 		return rc;
784 
785 	cxl_mem_get_event_records(mds, CXLDEV_EVENT_STATUS_ALL);
786 
787 	return 0;
788 }
789 
790 static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
791 {
792 	struct pci_host_bridge *host_bridge = pci_find_host_bridge(pdev->bus);
793 	struct cxl_memdev_state *mds;
794 	struct cxl_dev_state *cxlds;
795 	struct cxl_register_map map;
796 	struct cxl_memdev *cxlmd;
797 	int i, rc, pmu_count;
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 	cxlds->component_reg_phys = CXL_RESOURCE_NONE;
838 	rc = cxl_pci_setup_regs(pdev, CXL_REGLOC_RBI_COMPONENT, &map);
839 	if (rc)
840 		dev_warn(&pdev->dev, "No component registers (%d)\n", rc);
841 	else if (!map.component_map.ras.valid)
842 		dev_dbg(&pdev->dev, "RAS registers not found\n");
843 
844 	cxlds->component_reg_phys = map.resource;
845 
846 	rc = cxl_map_component_regs(&map, &cxlds->regs.component,
847 				    BIT(CXL_CM_CAP_CAP_ID_RAS));
848 	if (rc)
849 		dev_dbg(&pdev->dev, "Failed to map RAS capability.\n");
850 
851 	rc = cxl_await_media_ready(cxlds);
852 	if (rc == 0)
853 		cxlds->media_ready = true;
854 	else
855 		dev_warn(&pdev->dev, "Media not active (%d)\n", rc);
856 
857 	rc = cxl_alloc_irq_vectors(pdev);
858 	if (rc)
859 		return rc;
860 
861 	rc = cxl_pci_setup_mailbox(mds);
862 	if (rc)
863 		return rc;
864 
865 	rc = cxl_enumerate_cmds(mds);
866 	if (rc)
867 		return rc;
868 
869 	rc = cxl_set_timestamp(mds);
870 	if (rc)
871 		return rc;
872 
873 	rc = cxl_poison_state_init(mds);
874 	if (rc)
875 		return rc;
876 
877 	rc = cxl_dev_state_identify(mds);
878 	if (rc)
879 		return rc;
880 
881 	rc = cxl_mem_create_range_info(mds);
882 	if (rc)
883 		return rc;
884 
885 	cxlmd = devm_cxl_add_memdev(cxlds);
886 	if (IS_ERR(cxlmd))
887 		return PTR_ERR(cxlmd);
888 
889 	rc = cxl_memdev_setup_fw_upload(mds);
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(pdev, &pmu_regs, &map);
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);
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_LICENSE("GPL v2");
978 module_pci_driver(cxl_pci_driver);
979 MODULE_IMPORT_NS(CXL);
980