xref: /linux/drivers/cxl/pci.c (revision a50eba1e778ad4da5b6f9ddbbf57dabbea59bc05)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2020 Intel Corporation. All rights reserved. */
3 #include <linux/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 <cxl/pci.h>
15 #include <cxl/mailbox.h>
16 #include "cxlmem.h"
17 #include "cxlpci.h"
18 #include "cxl.h"
19 #include "pmu.h"
20 
21 /**
22  * DOC: cxl pci
23  *
24  * This implements the PCI exclusive functionality for a CXL device as it is
25  * defined by the Compute Express Link specification. CXL devices may surface
26  * certain functionality even if it isn't CXL enabled. While this driver is
27  * focused around the PCI specific aspects of a CXL device, it binds to the
28  * specific CXL memory device class code, and therefore the implementation of
29  * cxl_pci is focused around CXL memory devices.
30  *
31  * The driver has several responsibilities, mainly:
32  *  - Create the memX device and register on the CXL bus.
33  *  - Enumerate device's register interface and map them.
34  *  - Registers nvdimm bridge device with cxl_core.
35  *  - Registers a CXL mailbox with cxl_core.
36  */
37 
38 #define cxl_doorbell_busy(cxlds)                                                \
39 	(readl((cxlds)->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET) &                  \
40 	 CXLDEV_MBOX_CTRL_DOORBELL)
41 
42 /* CXL 2.0 - 8.2.8.4 */
43 #define CXL_MAILBOX_TIMEOUT_MS (2 * HZ)
44 
45 /*
46  * CXL 2.0 ECN "Add Mailbox Ready Time" defines a capability field to
47  * dictate how long to wait for the mailbox to become ready. The new
48  * field allows the device to tell software the amount of time to wait
49  * before mailbox ready. This field per the spec theoretically allows
50  * for up to 255 seconds. 255 seconds is unreasonably long, its longer
51  * than the maximum SATA port link recovery wait. Default to 60 seconds
52  * until someone builds a CXL device that needs more time in practice.
53  */
54 static unsigned short mbox_ready_timeout = 60;
55 module_param(mbox_ready_timeout, ushort, 0644);
56 MODULE_PARM_DESC(mbox_ready_timeout, "seconds to wait for mailbox ready");
57 
58 static int cxl_pci_mbox_wait_for_doorbell(struct cxl_dev_state *cxlds)
59 {
60 	const unsigned long start = jiffies;
61 	unsigned long end = start;
62 
63 	while (cxl_doorbell_busy(cxlds)) {
64 		end = jiffies;
65 
66 		if (time_after(end, start + CXL_MAILBOX_TIMEOUT_MS)) {
67 			/* Check again in case preempted before timeout test */
68 			if (!cxl_doorbell_busy(cxlds))
69 				break;
70 			return -ETIMEDOUT;
71 		}
72 		cpu_relax();
73 	}
74 
75 	dev_dbg(cxlds->dev, "Doorbell wait took %dms",
76 		jiffies_to_msecs(end) - jiffies_to_msecs(start));
77 	return 0;
78 }
79 
80 #define cxl_err(dev, status, msg)                                        \
81 	dev_err_ratelimited(dev, msg ", device state %s%s\n",                  \
82 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
83 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
84 
85 #define cxl_cmd_err(dev, cmd, status, msg)                               \
86 	dev_err_ratelimited(dev, msg " (opcode: %#x), device state %s%s\n",    \
87 			    (cmd)->opcode,                                     \
88 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
89 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
90 
91 /*
92  * Threaded irq dev_id's must be globally unique.  cxl_dev_id provides a unique
93  * wrapper object for each irq within the same cxlds.
94  */
95 struct cxl_dev_id {
96 	struct cxl_dev_state *cxlds;
97 };
98 
99 static int cxl_request_irq(struct cxl_dev_state *cxlds, int irq,
100 			   irq_handler_t thread_fn)
101 {
102 	struct device *dev = cxlds->dev;
103 	struct cxl_dev_id *dev_id;
104 
105 	dev_id = devm_kzalloc(dev, sizeof(*dev_id), GFP_KERNEL);
106 	if (!dev_id)
107 		return -ENOMEM;
108 	dev_id->cxlds = cxlds;
109 
110 	return devm_request_threaded_irq(dev, irq, NULL, thread_fn,
111 					 IRQF_SHARED | IRQF_ONESHOT, NULL,
112 					 dev_id);
113 }
114 
115 static bool cxl_mbox_background_complete(struct cxl_dev_state *cxlds)
116 {
117 	u64 reg;
118 
119 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
120 	return FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_PCT_MASK, reg) == 100;
121 }
122 
123 static irqreturn_t cxl_pci_mbox_irq(int irq, void *id)
124 {
125 	u64 reg;
126 	u16 opcode;
127 	struct cxl_dev_id *dev_id = id;
128 	struct cxl_dev_state *cxlds = dev_id->cxlds;
129 	struct cxl_mailbox *cxl_mbox = &cxlds->cxl_mbox;
130 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
131 
132 	if (!cxl_mbox_background_complete(cxlds))
133 		return IRQ_NONE;
134 
135 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
136 	opcode = FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_OPCODE_MASK, reg);
137 	if (opcode == CXL_MBOX_OP_SANITIZE) {
138 		mutex_lock(&cxl_mbox->mbox_mutex);
139 		if (mds->security.sanitize_node)
140 			mod_delayed_work(system_percpu_wq, &mds->security.poll_dwork, 0);
141 		mutex_unlock(&cxl_mbox->mbox_mutex);
142 	} else {
143 		/* short-circuit the wait in __cxl_pci_mbox_send_cmd() */
144 		rcuwait_wake_up(&cxl_mbox->mbox_wait);
145 	}
146 
147 	return IRQ_HANDLED;
148 }
149 
150 /*
151  * Sanitization operation polling mode.
152  */
153 static void cxl_mbox_sanitize_work(struct work_struct *work)
154 {
155 	struct cxl_memdev_state *mds =
156 		container_of(work, typeof(*mds), security.poll_dwork.work);
157 	struct cxl_dev_state *cxlds = &mds->cxlds;
158 	struct cxl_mailbox *cxl_mbox = &cxlds->cxl_mbox;
159 
160 	mutex_lock(&cxl_mbox->mbox_mutex);
161 	if (cxl_mbox_background_complete(cxlds)) {
162 		mds->security.poll_tmo_secs = 0;
163 		if (mds->security.sanitize_node)
164 			sysfs_notify_dirent(mds->security.sanitize_node);
165 		mds->security.sanitize_active = false;
166 
167 		dev_dbg(cxlds->dev, "Sanitization operation ended\n");
168 	} else {
169 		int timeout = mds->security.poll_tmo_secs + 10;
170 
171 		mds->security.poll_tmo_secs = min(15 * 60, timeout);
172 		schedule_delayed_work(&mds->security.poll_dwork, timeout * HZ);
173 	}
174 	mutex_unlock(&cxl_mbox->mbox_mutex);
175 }
176 
177 /**
178  * __cxl_pci_mbox_send_cmd() - Execute a mailbox command
179  * @cxl_mbox: CXL mailbox context
180  * @mbox_cmd: Command to send to the memory device.
181  *
182  * Context: Any context. Expects mbox_mutex to be held.
183  * Return: -ETIMEDOUT if timeout occurred waiting for completion. 0 on success.
184  *         Caller should check the return code in @mbox_cmd to make sure it
185  *         succeeded.
186  *
187  * This is a generic form of the CXL mailbox send command thus only using the
188  * registers defined by the mailbox capability ID - CXL 2.0 8.2.8.4. Memory
189  * devices, and perhaps other types of CXL devices may have further information
190  * available upon error conditions. Driver facilities wishing to send mailbox
191  * commands should use the wrapper command.
192  *
193  * The CXL spec allows for up to two mailboxes. The intention is for the primary
194  * mailbox to be OS controlled and the secondary mailbox to be used by system
195  * firmware. This allows the OS and firmware to communicate with the device and
196  * not need to coordinate with each other. The driver only uses the primary
197  * mailbox.
198  */
199 static int __cxl_pci_mbox_send_cmd(struct cxl_mailbox *cxl_mbox,
200 				   struct cxl_mbox_cmd *mbox_cmd)
201 {
202 	struct cxl_dev_state *cxlds = mbox_to_cxlds(cxl_mbox);
203 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
204 	void __iomem *payload = cxlds->regs.mbox + CXLDEV_MBOX_PAYLOAD_OFFSET;
205 	struct device *dev = cxlds->dev;
206 	u64 cmd_reg, status_reg;
207 	size_t out_len;
208 	int rc;
209 
210 	lockdep_assert_held(&cxl_mbox->mbox_mutex);
211 
212 	/*
213 	 * Here are the steps from 8.2.8.4 of the CXL 2.0 spec.
214 	 *   1. Caller reads MB Control Register to verify doorbell is clear
215 	 *   2. Caller writes Command Register
216 	 *   3. Caller writes Command Payload Registers if input payload is non-empty
217 	 *   4. Caller writes MB Control Register to set doorbell
218 	 *   5. Caller either polls for doorbell to be clear or waits for interrupt if configured
219 	 *   6. Caller reads MB Status Register to fetch Return code
220 	 *   7. If command successful, Caller reads Command Register to get Payload Length
221 	 *   8. If output payload is non-empty, host reads Command Payload Registers
222 	 *
223 	 * Hardware is free to do whatever it wants before the doorbell is rung,
224 	 * and isn't allowed to change anything after it clears the doorbell. As
225 	 * such, steps 2 and 3 can happen in any order, and steps 6, 7, 8 can
226 	 * also happen in any order (though some orders might not make sense).
227 	 */
228 
229 	/* #1 */
230 	if (cxl_doorbell_busy(cxlds)) {
231 		u64 md_status =
232 			readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
233 
234 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status,
235 			    "mailbox queue busy");
236 		return -EBUSY;
237 	}
238 
239 	/*
240 	 * With sanitize polling, hardware might be done and the poller still
241 	 * not be in sync. Ensure no new command comes in until so. Keep the
242 	 * hardware semantics and only allow device health status.
243 	 */
244 	if (mds->security.poll_tmo_secs > 0) {
245 		if (mbox_cmd->opcode != CXL_MBOX_OP_GET_HEALTH_INFO)
246 			return -EBUSY;
247 	}
248 
249 	cmd_reg = FIELD_PREP(CXLDEV_MBOX_CMD_COMMAND_OPCODE_MASK,
250 			     mbox_cmd->opcode);
251 	if (mbox_cmd->size_in) {
252 		if (WARN_ON(!mbox_cmd->payload_in))
253 			return -EINVAL;
254 
255 		cmd_reg |= FIELD_PREP(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK,
256 				      mbox_cmd->size_in);
257 		memcpy_toio(payload, mbox_cmd->payload_in, mbox_cmd->size_in);
258 	}
259 
260 	/* #2, #3 */
261 	writeq(cmd_reg, cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
262 
263 	/* #4 */
264 	dev_dbg(dev, "Sending command: 0x%04x\n", mbox_cmd->opcode);
265 	writel(CXLDEV_MBOX_CTRL_DOORBELL,
266 	       cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
267 
268 	/* #5 */
269 	rc = cxl_pci_mbox_wait_for_doorbell(cxlds);
270 	if (rc == -ETIMEDOUT) {
271 		u64 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
272 
273 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status, "mailbox timeout");
274 		return rc;
275 	}
276 
277 	/* #6 */
278 	status_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_STATUS_OFFSET);
279 	mbox_cmd->return_code =
280 		FIELD_GET(CXLDEV_MBOX_STATUS_RET_CODE_MASK, status_reg);
281 
282 	/*
283 	 * Handle the background command in a synchronous manner.
284 	 *
285 	 * All other mailbox commands will serialize/queue on the mbox_mutex,
286 	 * which we currently hold. Furthermore this also guarantees that
287 	 * cxl_mbox_background_complete() checks are safe amongst each other,
288 	 * in that no new bg operation can occur in between.
289 	 *
290 	 * Background operations are timesliced in accordance with the nature
291 	 * of the command. In the event of timeout, the mailbox state is
292 	 * indeterminate until the next successful command submission and the
293 	 * driver can get back in sync with the hardware state.
294 	 */
295 	if (mbox_cmd->return_code == CXL_MBOX_CMD_RC_BACKGROUND) {
296 		u64 bg_status_reg;
297 		int i, timeout;
298 
299 		/*
300 		 * Sanitization is a special case which monopolizes the device
301 		 * and cannot be timesliced. Handle asynchronously instead,
302 		 * and allow userspace to poll(2) for completion.
303 		 */
304 		if (mbox_cmd->opcode == CXL_MBOX_OP_SANITIZE) {
305 			if (mds->security.sanitize_active)
306 				return -EBUSY;
307 
308 			/* give first timeout a second */
309 			timeout = 1;
310 			mds->security.poll_tmo_secs = timeout;
311 			mds->security.sanitize_active = true;
312 			schedule_delayed_work(&mds->security.poll_dwork,
313 					      timeout * HZ);
314 			dev_dbg(dev, "Sanitization operation started\n");
315 			goto success;
316 		}
317 
318 		dev_dbg(dev, "Mailbox background operation (0x%04x) started\n",
319 			mbox_cmd->opcode);
320 
321 		timeout = mbox_cmd->poll_interval_ms;
322 		for (i = 0; i < mbox_cmd->poll_count; i++) {
323 			if (rcuwait_wait_event_timeout(&cxl_mbox->mbox_wait,
324 						       cxl_mbox_background_complete(cxlds),
325 						       TASK_UNINTERRUPTIBLE,
326 						       msecs_to_jiffies(timeout)) > 0)
327 				break;
328 		}
329 
330 		if (!cxl_mbox_background_complete(cxlds)) {
331 			dev_err(dev, "timeout waiting for background (%d ms)\n",
332 				timeout * mbox_cmd->poll_count);
333 			return -ETIMEDOUT;
334 		}
335 
336 		bg_status_reg = readq(cxlds->regs.mbox +
337 				      CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
338 		mbox_cmd->return_code =
339 			FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_RC_MASK,
340 				  bg_status_reg);
341 		dev_dbg(dev,
342 			"Mailbox background operation (0x%04x) completed\n",
343 			mbox_cmd->opcode);
344 	}
345 
346 	if (mbox_cmd->return_code != CXL_MBOX_CMD_RC_SUCCESS) {
347 		dev_dbg(dev, "Mailbox operation had an error: %s\n",
348 			cxl_mbox_cmd_rc2str(mbox_cmd));
349 		return 0; /* completed but caller must check return_code */
350 	}
351 
352 success:
353 	/* #7 */
354 	cmd_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
355 	out_len = FIELD_GET(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK, cmd_reg);
356 
357 	/* #8 */
358 	if (out_len && mbox_cmd->payload_out) {
359 		/*
360 		 * Sanitize the copy. If hardware misbehaves, out_len per the
361 		 * spec can actually be greater than the max allowed size (21
362 		 * bits available but spec defined 1M max). The caller also may
363 		 * have requested less data than the hardware supplied even
364 		 * within spec.
365 		 */
366 		size_t n;
367 
368 		n = min3(mbox_cmd->size_out, cxl_mbox->payload_size, out_len);
369 		memcpy_fromio(mbox_cmd->payload_out, payload, n);
370 		mbox_cmd->size_out = n;
371 	} else {
372 		mbox_cmd->size_out = 0;
373 	}
374 
375 	return 0;
376 }
377 
378 static int cxl_pci_mbox_send(struct cxl_mailbox *cxl_mbox,
379 			     struct cxl_mbox_cmd *cmd)
380 {
381 	int rc;
382 
383 	mutex_lock(&cxl_mbox->mbox_mutex);
384 	rc = __cxl_pci_mbox_send_cmd(cxl_mbox, cmd);
385 	mutex_unlock(&cxl_mbox->mbox_mutex);
386 
387 	return rc;
388 }
389 
390 static int cxl_pci_setup_mailbox(struct cxl_memdev_state *mds, bool irq_avail)
391 {
392 	struct cxl_dev_state *cxlds = &mds->cxlds;
393 	struct cxl_mailbox *cxl_mbox = &cxlds->cxl_mbox;
394 	const int cap = readl(cxlds->regs.mbox + CXLDEV_MBOX_CAPS_OFFSET);
395 	struct device *dev = cxlds->dev;
396 	unsigned long timeout;
397 	int irq, msgnum;
398 	u64 md_status;
399 	u32 ctrl;
400 
401 	timeout = jiffies + mbox_ready_timeout * HZ;
402 	do {
403 		md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
404 		if (md_status & CXLMDEV_MBOX_IF_READY)
405 			break;
406 		if (msleep_interruptible(100))
407 			break;
408 	} while (!time_after(jiffies, timeout));
409 
410 	if (!(md_status & CXLMDEV_MBOX_IF_READY)) {
411 		cxl_err(dev, md_status, "timeout awaiting mailbox ready");
412 		return -ETIMEDOUT;
413 	}
414 
415 	/*
416 	 * A command may be in flight from a previous driver instance,
417 	 * think kexec, do one doorbell wait so that
418 	 * __cxl_pci_mbox_send_cmd() can assume that it is the only
419 	 * source for future doorbell busy events.
420 	 */
421 	if (cxl_pci_mbox_wait_for_doorbell(cxlds) != 0) {
422 		cxl_err(dev, md_status, "timeout awaiting mailbox idle");
423 		return -ETIMEDOUT;
424 	}
425 
426 	cxl_mbox->mbox_send = cxl_pci_mbox_send;
427 	cxl_mbox->payload_size =
428 		1 << FIELD_GET(CXLDEV_MBOX_CAP_PAYLOAD_SIZE_MASK, cap);
429 
430 	/*
431 	 * CXL 2.0 8.2.8.4.3 Mailbox Capabilities Register
432 	 *
433 	 * If the size is too small, mandatory commands will not work and so
434 	 * there's no point in going forward. If the size is too large, there's
435 	 * no harm is soft limiting it.
436 	 */
437 	cxl_mbox->payload_size = min_t(size_t, cxl_mbox->payload_size, SZ_1M);
438 	if (cxl_mbox->payload_size < 256) {
439 		dev_err(dev, "Mailbox is too small (%zub)",
440 			cxl_mbox->payload_size);
441 		return -ENXIO;
442 	}
443 
444 	dev_dbg(dev, "Mailbox payload sized %zu", cxl_mbox->payload_size);
445 
446 	INIT_DELAYED_WORK(&mds->security.poll_dwork, cxl_mbox_sanitize_work);
447 
448 	/* background command interrupts are optional */
449 	if (!(cap & CXLDEV_MBOX_CAP_BG_CMD_IRQ) || !irq_avail)
450 		return 0;
451 
452 	msgnum = FIELD_GET(CXLDEV_MBOX_CAP_IRQ_MSGNUM_MASK, cap);
453 	irq = pci_irq_vector(to_pci_dev(cxlds->dev), msgnum);
454 	if (irq < 0)
455 		return 0;
456 
457 	if (cxl_request_irq(cxlds, irq, cxl_pci_mbox_irq))
458 		return 0;
459 
460 	dev_dbg(cxlds->dev, "Mailbox interrupts enabled\n");
461 	/* enable background command mbox irq support */
462 	ctrl = readl(cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
463 	ctrl |= CXLDEV_MBOX_CTRL_BG_CMD_IRQ;
464 	writel(ctrl, cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
465 
466 	return 0;
467 }
468 
469 static void free_event_buf(void *buf)
470 {
471 	kvfree(buf);
472 }
473 
474 /*
475  * There is a single buffer for reading event logs from the mailbox.  All logs
476  * share this buffer protected by the mds->event_log_lock.
477  */
478 static int cxl_mem_alloc_event_buf(struct cxl_memdev_state *mds)
479 {
480 	struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
481 	struct cxl_get_event_payload *buf;
482 
483 	buf = kvmalloc(cxl_mbox->payload_size, GFP_KERNEL);
484 	if (!buf)
485 		return -ENOMEM;
486 	mds->event.buf = buf;
487 
488 	return devm_add_action_or_reset(mds->cxlds.dev, free_event_buf, buf);
489 }
490 
491 static bool cxl_alloc_irq_vectors(struct pci_dev *pdev)
492 {
493 	int nvecs;
494 
495 	/*
496 	 * Per CXL 3.0 3.1.1 CXL.io Endpoint a function on a CXL device must
497 	 * not generate INTx messages if that function participates in
498 	 * CXL.cache or CXL.mem.
499 	 *
500 	 * Additionally pci_alloc_irq_vectors() handles calling
501 	 * pci_free_irq_vectors() automatically despite not being called
502 	 * pcim_*.  See pci_setup_msi_context().
503 	 */
504 	nvecs = pci_alloc_irq_vectors(pdev, 1, CXL_PCI_DEFAULT_MAX_VECTORS,
505 				      PCI_IRQ_MSIX | PCI_IRQ_MSI);
506 	if (nvecs < 1) {
507 		dev_dbg(&pdev->dev, "Failed to alloc irq vectors: %d\n", nvecs);
508 		return false;
509 	}
510 	return true;
511 }
512 
513 static irqreturn_t cxl_event_thread(int irq, void *id)
514 {
515 	struct cxl_dev_id *dev_id = id;
516 	struct cxl_dev_state *cxlds = dev_id->cxlds;
517 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
518 	u32 status;
519 
520 	do {
521 		/*
522 		 * CXL 3.0 8.2.8.3.1: The lower 32 bits are the status;
523 		 * ignore the reserved upper 32 bits
524 		 */
525 		status = readl(cxlds->regs.status + CXLDEV_DEV_EVENT_STATUS_OFFSET);
526 		/* Ignore logs unknown to the driver */
527 		status &= CXLDEV_EVENT_STATUS_ALL;
528 		if (!status)
529 			break;
530 		cxl_mem_get_event_records(mds, status);
531 		cond_resched();
532 	} while (status);
533 
534 	return IRQ_HANDLED;
535 }
536 
537 static int cxl_event_req_irq(struct cxl_dev_state *cxlds, u8 setting)
538 {
539 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
540 	int irq;
541 
542 	if (FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting) != CXL_INT_MSI_MSIX)
543 		return -ENXIO;
544 
545 	irq =  pci_irq_vector(pdev,
546 			      FIELD_GET(CXLDEV_EVENT_INT_MSGNUM_MASK, setting));
547 	if (irq < 0)
548 		return irq;
549 
550 	return cxl_request_irq(cxlds, irq, cxl_event_thread);
551 }
552 
553 static int cxl_event_get_int_policy(struct cxl_memdev_state *mds,
554 				    struct cxl_event_interrupt_policy *policy)
555 {
556 	struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
557 	struct cxl_mbox_cmd mbox_cmd = {
558 		.opcode = CXL_MBOX_OP_GET_EVT_INT_POLICY,
559 		.payload_out = policy,
560 		.size_out = sizeof(*policy),
561 	};
562 	int rc;
563 
564 	rc = cxl_internal_send_cmd(cxl_mbox, &mbox_cmd);
565 	if (rc < 0)
566 		dev_err(mds->cxlds.dev,
567 			"Failed to get event interrupt policy : %d", rc);
568 
569 	return rc;
570 }
571 
572 static int cxl_event_config_msgnums(struct cxl_memdev_state *mds,
573 				    struct cxl_event_interrupt_policy *policy)
574 {
575 	struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
576 	struct cxl_mbox_cmd mbox_cmd;
577 	int rc;
578 
579 	*policy = (struct cxl_event_interrupt_policy) {
580 		.info_settings = CXL_INT_MSI_MSIX,
581 		.warn_settings = CXL_INT_MSI_MSIX,
582 		.failure_settings = CXL_INT_MSI_MSIX,
583 		.fatal_settings = CXL_INT_MSI_MSIX,
584 	};
585 
586 	mbox_cmd = (struct cxl_mbox_cmd) {
587 		.opcode = CXL_MBOX_OP_SET_EVT_INT_POLICY,
588 		.payload_in = policy,
589 		.size_in = sizeof(*policy),
590 	};
591 
592 	rc = cxl_internal_send_cmd(cxl_mbox, &mbox_cmd);
593 	if (rc < 0) {
594 		dev_err(mds->cxlds.dev, "Failed to set event interrupt policy : %d",
595 			rc);
596 		return rc;
597 	}
598 
599 	/* Retrieve final interrupt settings */
600 	return cxl_event_get_int_policy(mds, policy);
601 }
602 
603 static int cxl_event_irqsetup(struct cxl_memdev_state *mds)
604 {
605 	struct cxl_dev_state *cxlds = &mds->cxlds;
606 	struct cxl_event_interrupt_policy policy;
607 	int rc;
608 
609 	rc = cxl_event_config_msgnums(mds, &policy);
610 	if (rc)
611 		return rc;
612 
613 	rc = cxl_event_req_irq(cxlds, policy.info_settings);
614 	if (rc) {
615 		dev_err(cxlds->dev, "Failed to get interrupt for event Info log\n");
616 		return rc;
617 	}
618 
619 	rc = cxl_event_req_irq(cxlds, policy.warn_settings);
620 	if (rc) {
621 		dev_err(cxlds->dev, "Failed to get interrupt for event Warn log\n");
622 		return rc;
623 	}
624 
625 	rc = cxl_event_req_irq(cxlds, policy.failure_settings);
626 	if (rc) {
627 		dev_err(cxlds->dev, "Failed to get interrupt for event Failure log\n");
628 		return rc;
629 	}
630 
631 	rc = cxl_event_req_irq(cxlds, policy.fatal_settings);
632 	if (rc) {
633 		dev_err(cxlds->dev, "Failed to get interrupt for event Fatal log\n");
634 		return rc;
635 	}
636 
637 	return 0;
638 }
639 
640 static bool cxl_event_int_is_fw(u8 setting)
641 {
642 	u8 mode = FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting);
643 
644 	return mode == CXL_INT_FW;
645 }
646 
647 static int cxl_event_config(struct pci_host_bridge *host_bridge,
648 			    struct cxl_memdev_state *mds, bool irq_avail)
649 {
650 	struct cxl_event_interrupt_policy policy;
651 	int rc;
652 
653 	/*
654 	 * When BIOS maintains CXL error reporting control, it will process
655 	 * event records.  Only one agent can do so.
656 	 */
657 	if (!host_bridge->native_cxl_error)
658 		return 0;
659 
660 	if (!irq_avail) {
661 		dev_info(mds->cxlds.dev, "No interrupt support, disable event processing.\n");
662 		return 0;
663 	}
664 
665 	rc = cxl_event_get_int_policy(mds, &policy);
666 	if (rc)
667 		return rc;
668 
669 	if (cxl_event_int_is_fw(policy.info_settings) ||
670 	    cxl_event_int_is_fw(policy.warn_settings) ||
671 	    cxl_event_int_is_fw(policy.failure_settings) ||
672 	    cxl_event_int_is_fw(policy.fatal_settings)) {
673 		dev_err(mds->cxlds.dev,
674 			"FW still in control of Event Logs despite _OSC settings\n");
675 		return -EBUSY;
676 	}
677 
678 	rc = cxl_mem_alloc_event_buf(mds);
679 	if (rc)
680 		return rc;
681 
682 	rc = cxl_event_irqsetup(mds);
683 	if (rc)
684 		return rc;
685 
686 	cxl_mem_get_event_records(mds, CXLDEV_EVENT_STATUS_ALL);
687 
688 	return 0;
689 }
690 
691 static int cxl_pci_type3_init_mailbox(struct cxl_dev_state *cxlds)
692 {
693 	int rc;
694 
695 	/*
696 	 * Fail the init if there's no mailbox. For a type3 this is out of spec.
697 	 */
698 	if (!cxlds->reg_map.device_map.mbox.valid)
699 		return -ENODEV;
700 
701 	rc = cxl_mailbox_init(&cxlds->cxl_mbox, cxlds->dev);
702 	if (rc)
703 		return rc;
704 
705 	return 0;
706 }
707 
708 static ssize_t rcd_pcie_cap_emit(struct device *dev, u16 offset, char *buf, size_t width)
709 {
710 	struct cxl_dev_state *cxlds = dev_get_drvdata(dev);
711 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
712 	struct device *root_dev;
713 	struct cxl_dport *dport;
714 	struct cxl_port *root __free(put_cxl_port) =
715 		cxl_mem_find_port(cxlmd, &dport);
716 
717 	if (!root)
718 		return -ENXIO;
719 
720 	root_dev = root->uport_dev;
721 	if (!root_dev)
722 		return -ENXIO;
723 
724 	if (!dport->regs.rcd_pcie_cap)
725 		return -ENXIO;
726 
727 	guard(device)(root_dev);
728 	if (!root_dev->driver)
729 		return -ENXIO;
730 
731 	switch (width) {
732 	case 2:
733 		return sysfs_emit(buf, "%#x\n",
734 				  readw(dport->regs.rcd_pcie_cap + offset));
735 	case 4:
736 		return sysfs_emit(buf, "%#x\n",
737 				  readl(dport->regs.rcd_pcie_cap + offset));
738 	default:
739 		return -EINVAL;
740 	}
741 }
742 
743 static ssize_t rcd_link_cap_show(struct device *dev,
744 				 struct device_attribute *attr, char *buf)
745 {
746 	return rcd_pcie_cap_emit(dev, PCI_EXP_LNKCAP, buf, sizeof(u32));
747 }
748 static DEVICE_ATTR_RO(rcd_link_cap);
749 
750 static ssize_t rcd_link_ctrl_show(struct device *dev,
751 				  struct device_attribute *attr, char *buf)
752 {
753 	return rcd_pcie_cap_emit(dev, PCI_EXP_LNKCTL, buf, sizeof(u16));
754 }
755 static DEVICE_ATTR_RO(rcd_link_ctrl);
756 
757 static ssize_t rcd_link_status_show(struct device *dev,
758 				    struct device_attribute *attr, char *buf)
759 {
760 	return rcd_pcie_cap_emit(dev, PCI_EXP_LNKSTA, buf, sizeof(u16));
761 }
762 static DEVICE_ATTR_RO(rcd_link_status);
763 
764 static struct attribute *cxl_rcd_attrs[] = {
765 	&dev_attr_rcd_link_cap.attr,
766 	&dev_attr_rcd_link_ctrl.attr,
767 	&dev_attr_rcd_link_status.attr,
768 	NULL
769 };
770 
771 static umode_t cxl_rcd_visible(struct kobject *kobj, struct attribute *a, int n)
772 {
773 	struct device *dev = kobj_to_dev(kobj);
774 	struct pci_dev *pdev = to_pci_dev(dev);
775 
776 	if (is_cxl_restricted(pdev))
777 		return a->mode;
778 
779 	return 0;
780 }
781 
782 static struct attribute_group cxl_rcd_group = {
783 	.attrs = cxl_rcd_attrs,
784 	.is_visible = cxl_rcd_visible,
785 };
786 __ATTRIBUTE_GROUPS(cxl_rcd);
787 
788 static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
789 {
790 	struct pci_host_bridge *host_bridge = pci_find_host_bridge(pdev->bus);
791 	struct cxl_dpa_info range_info = { 0 };
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 rc, pmu_count;
797 	unsigned int i;
798 	bool irq_avail;
799 	u16 dvsec;
800 
801 	rc = pcim_enable_device(pdev);
802 	if (rc)
803 		return rc;
804 	pci_set_master(pdev);
805 
806 	dvsec = pci_find_dvsec_capability(pdev, PCI_VENDOR_ID_CXL,
807 					  PCI_DVSEC_CXL_DEVICE);
808 	if (!dvsec)
809 		pci_warn(pdev, "Device DVSEC not present, skip CXL.mem init\n");
810 
811 	mds = cxl_memdev_state_create(&pdev->dev, pci_get_dsn(pdev), dvsec);
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 
819 	rc = cxl_pci_setup_regs(pdev, CXL_REGLOC_RBI_MEMDEV, &map);
820 	if (rc)
821 		return rc;
822 
823 	rc = cxl_map_device_regs(&map, &cxlds->regs);
824 	if (rc)
825 		return rc;
826 
827 	/*
828 	 * If the component registers can't be found, the cxl_pci driver may
829 	 * still be useful for management functions so don't return an error.
830 	 */
831 	rc = cxl_pci_setup_regs(pdev, CXL_REGLOC_RBI_COMPONENT,
832 				&cxlds->reg_map);
833 	if (rc)
834 		dev_warn(&pdev->dev, "No component registers (%d)\n", rc);
835 	else if (!cxlds->reg_map.component_map.ras.valid)
836 		dev_dbg(&pdev->dev, "RAS registers not found\n");
837 
838 	rc = cxl_pci_type3_init_mailbox(cxlds);
839 	if (rc)
840 		return rc;
841 
842 	rc = cxl_await_media_ready(cxlds);
843 	if (rc == 0)
844 		cxlds->media_ready = true;
845 	else
846 		dev_warn(&pdev->dev, "Media not active (%d)\n", rc);
847 
848 	irq_avail = cxl_alloc_irq_vectors(pdev);
849 
850 	rc = cxl_pci_setup_mailbox(mds, irq_avail);
851 	if (rc)
852 		return rc;
853 
854 	rc = cxl_enumerate_cmds(mds);
855 	if (rc)
856 		return rc;
857 
858 	rc = cxl_set_timestamp(mds);
859 	if (rc)
860 		return rc;
861 
862 	rc = cxl_poison_state_init(mds);
863 	if (rc)
864 		return rc;
865 
866 	rc = cxl_dev_state_identify(mds);
867 	if (rc)
868 		return rc;
869 
870 	rc = cxl_mem_dpa_fetch(mds, &range_info);
871 	if (rc)
872 		return rc;
873 
874 	rc = cxl_dpa_setup(cxlds, &range_info);
875 	if (rc)
876 		return rc;
877 
878 	rc = devm_cxl_setup_features(cxlds);
879 	if (rc)
880 		dev_dbg(&pdev->dev, "No CXL Features discovered\n");
881 
882 	cxlmd = devm_cxl_add_classdev(cxlds);
883 	if (IS_ERR(cxlmd))
884 		return PTR_ERR(cxlmd);
885 
886 	rc = devm_cxl_setup_fw_upload(&pdev->dev, mds);
887 	if (rc)
888 		return rc;
889 
890 	rc = devm_cxl_sanitize_setup_notifier(&pdev->dev, cxlmd);
891 	if (rc)
892 		return rc;
893 
894 	rc = devm_cxl_setup_fwctl(&pdev->dev, cxlmd);
895 	if (rc)
896 		dev_dbg(&pdev->dev, "No CXL FWCTL setup\n");
897 
898 	pmu_count = cxl_count_regblock(pdev, CXL_REGLOC_RBI_PMU);
899 	if (pmu_count < 0)
900 		return pmu_count;
901 
902 	for (i = 0; i < pmu_count; i++) {
903 		struct cxl_pmu_regs pmu_regs;
904 
905 		rc = cxl_find_regblock_instance(pdev, CXL_REGLOC_RBI_PMU, &map, i);
906 		if (rc) {
907 			dev_dbg(&pdev->dev, "Could not find PMU regblock\n");
908 			break;
909 		}
910 
911 		rc = cxl_map_pmu_regs(&map, &pmu_regs);
912 		if (rc) {
913 			dev_dbg(&pdev->dev, "Could not map PMU regs\n");
914 			break;
915 		}
916 
917 		rc = devm_cxl_pmu_add(cxlds->dev, &pmu_regs, cxlmd->id, i, CXL_PMU_MEMDEV);
918 		if (rc) {
919 			dev_dbg(&pdev->dev, "Could not add PMU instance\n");
920 			break;
921 		}
922 	}
923 
924 	rc = cxl_event_config(host_bridge, mds, irq_avail);
925 	if (rc)
926 		return rc;
927 
928 	pci_save_state(pdev);
929 
930 	return rc;
931 }
932 
933 static const struct pci_device_id cxl_mem_pci_tbl[] = {
934 	/* PCI class code for CXL.mem Type-3 Devices */
935 	{ PCI_DEVICE_CLASS((PCI_CLASS_MEMORY_CXL << 8 | CXL_MEMORY_PROGIF), ~0)},
936 	{ /* terminate list */ },
937 };
938 MODULE_DEVICE_TABLE(pci, cxl_mem_pci_tbl);
939 
940 static pci_ers_result_t cxl_slot_reset(struct pci_dev *pdev)
941 {
942 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
943 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
944 	struct device *dev = &cxlmd->dev;
945 
946 	dev_info(&pdev->dev, "%s: restart CXL.mem after slot reset\n",
947 		 dev_name(dev));
948 	pci_restore_state(pdev);
949 	if (device_attach(dev) <= 0)
950 		return PCI_ERS_RESULT_DISCONNECT;
951 	return PCI_ERS_RESULT_RECOVERED;
952 }
953 
954 static void cxl_error_resume(struct pci_dev *pdev)
955 {
956 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
957 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
958 	struct device *dev = &cxlmd->dev;
959 
960 	dev_info(&pdev->dev, "%s: error resume %s\n", dev_name(dev),
961 		 dev->driver ? "successful" : "failed");
962 }
963 
964 static int cxl_endpoint_decoder_clear_reset_flags(struct device *dev, void *data)
965 {
966 	struct cxl_endpoint_decoder *cxled;
967 
968 	if (!is_endpoint_decoder(dev))
969 		return 0;
970 
971 	cxled = to_cxl_endpoint_decoder(dev);
972 	cxled->cxld.flags &= ~CXL_DECODER_F_RESET_MASK;
973 
974 	return 0;
975 }
976 
977 static void cxl_reset_done(struct pci_dev *pdev)
978 {
979 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
980 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
981 	struct device *dev = &pdev->dev;
982 
983 	/*
984 	 * FLR does not expect to touch the HDM decoders and related
985 	 * registers.  SBR, however, will wipe all device configurations.
986 	 * Issue a warning if there was an active decoder before the reset
987 	 * that no longer exists.
988 	 */
989 	guard(device)(&cxlmd->dev);
990 	if (!cxlmd->dev.driver)
991 		return;
992 
993 	if (cxlmd->endpoint &&
994 	    cxl_endpoint_decoder_reset_detected(cxlmd->endpoint)) {
995 		device_for_each_child(&cxlmd->endpoint->dev, NULL,
996 				      cxl_endpoint_decoder_clear_reset_flags);
997 
998 		dev_crit(dev, "SBR happened without memory regions removal.\n");
999 		dev_crit(dev, "System may be unstable if regions hosted system memory.\n");
1000 		add_taint(TAINT_USER, LOCKDEP_STILL_OK);
1001 	}
1002 }
1003 
1004 static const struct pci_error_handlers cxl_error_handlers = {
1005 	.error_detected	= cxl_error_detected,
1006 	.slot_reset	= cxl_slot_reset,
1007 	.resume		= cxl_error_resume,
1008 	.cor_error_detected	= cxl_cor_error_detected,
1009 	.reset_done	= cxl_reset_done,
1010 };
1011 
1012 static struct pci_driver cxl_pci_driver = {
1013 	.name			= KBUILD_MODNAME,
1014 	.id_table		= cxl_mem_pci_tbl,
1015 	.probe			= cxl_pci_probe,
1016 	.err_handler		= &cxl_error_handlers,
1017 	.dev_groups		= cxl_rcd_groups,
1018 	.driver	= {
1019 		.probe_type	= PROBE_PREFER_ASYNCHRONOUS,
1020 	},
1021 };
1022 
1023 #define CXL_EVENT_HDR_FLAGS_REC_SEVERITY GENMASK(1, 0)
1024 static void cxl_handle_cper_event(enum cxl_event_type ev_type,
1025 				  struct cxl_cper_event_rec *rec)
1026 {
1027 	struct cper_cxl_event_devid *device_id = &rec->hdr.device_id;
1028 	struct pci_dev *pdev __free(pci_dev_put) = NULL;
1029 	enum cxl_event_log_type log_type;
1030 	struct cxl_dev_state *cxlds;
1031 	unsigned int devfn;
1032 	u32 hdr_flags;
1033 
1034 	pr_debug("CPER event %d for device %u:%u:%u.%u\n", ev_type,
1035 		 device_id->segment_num, device_id->bus_num,
1036 		 device_id->device_num, device_id->func_num);
1037 
1038 	devfn = PCI_DEVFN(device_id->device_num, device_id->func_num);
1039 	pdev = pci_get_domain_bus_and_slot(device_id->segment_num,
1040 					   device_id->bus_num, devfn);
1041 	if (!pdev)
1042 		return;
1043 
1044 	guard(device)(&pdev->dev);
1045 	if (pdev->driver != &cxl_pci_driver)
1046 		return;
1047 
1048 	cxlds = pci_get_drvdata(pdev);
1049 	if (!cxlds)
1050 		return;
1051 
1052 	/* Fabricate a log type */
1053 	hdr_flags = get_unaligned_le24(rec->event.generic.hdr.flags);
1054 	log_type = FIELD_GET(CXL_EVENT_HDR_FLAGS_REC_SEVERITY, hdr_flags);
1055 
1056 	cxl_event_trace_record(cxlds->cxlmd, log_type, ev_type,
1057 			       &uuid_null, &rec->event);
1058 }
1059 
1060 static void cxl_cper_work_fn(struct work_struct *work)
1061 {
1062 	struct cxl_cper_work_data wd;
1063 
1064 	while (cxl_cper_kfifo_get(&wd))
1065 		cxl_handle_cper_event(wd.event_type, &wd.rec);
1066 }
1067 static DECLARE_WORK(cxl_cper_work, cxl_cper_work_fn);
1068 
1069 static int __init cxl_pci_driver_init(void)
1070 {
1071 	int rc;
1072 
1073 	rc = pci_register_driver(&cxl_pci_driver);
1074 	if (rc)
1075 		return rc;
1076 
1077 	rc = cxl_cper_register_work(&cxl_cper_work);
1078 	if (rc)
1079 		pci_unregister_driver(&cxl_pci_driver);
1080 
1081 	return rc;
1082 }
1083 
1084 static void __exit cxl_pci_driver_exit(void)
1085 {
1086 	cxl_cper_unregister_work(&cxl_cper_work);
1087 	cancel_work_sync(&cxl_cper_work);
1088 	pci_unregister_driver(&cxl_pci_driver);
1089 }
1090 
1091 module_init(cxl_pci_driver_init);
1092 module_exit(cxl_pci_driver_exit);
1093 MODULE_DESCRIPTION("CXL: PCI manageability");
1094 MODULE_LICENSE("GPL v2");
1095 MODULE_IMPORT_NS("CXL");
1096