xref: /linux/drivers/cxl/pci.c (revision 06a130e42a5bfc84795464bff023bff4c16f58c5)
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/mailbox.h>
15 #include "cxlmem.h"
16 #include "cxlpci.h"
17 #include "cxl.h"
18 #include "pmu.h"
19 
20 /**
21  * DOC: cxl pci
22  *
23  * This implements the PCI exclusive functionality for a CXL device as it is
24  * defined by the Compute Express Link specification. CXL devices may surface
25  * certain functionality even if it isn't CXL enabled. While this driver is
26  * focused around the PCI specific aspects of a CXL device, it binds to the
27  * specific CXL memory device class code, and therefore the implementation of
28  * cxl_pci is focused around CXL memory devices.
29  *
30  * The driver has several responsibilities, mainly:
31  *  - Create the memX device and register on the CXL bus.
32  *  - Enumerate device's register interface and map them.
33  *  - Registers nvdimm bridge device with cxl_core.
34  *  - Registers a CXL mailbox with cxl_core.
35  */
36 
37 #define cxl_doorbell_busy(cxlds)                                                \
38 	(readl((cxlds)->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET) &                  \
39 	 CXLDEV_MBOX_CTRL_DOORBELL)
40 
41 /* CXL 2.0 - 8.2.8.4 */
42 #define CXL_MAILBOX_TIMEOUT_MS (2 * HZ)
43 
44 /*
45  * CXL 2.0 ECN "Add Mailbox Ready Time" defines a capability field to
46  * dictate how long to wait for the mailbox to become ready. The new
47  * field allows the device to tell software the amount of time to wait
48  * before mailbox ready. This field per the spec theoretically allows
49  * for up to 255 seconds. 255 seconds is unreasonably long, its longer
50  * than the maximum SATA port link recovery wait. Default to 60 seconds
51  * until someone builds a CXL device that needs more time in practice.
52  */
53 static unsigned short mbox_ready_timeout = 60;
54 module_param(mbox_ready_timeout, ushort, 0644);
55 MODULE_PARM_DESC(mbox_ready_timeout, "seconds to wait for mailbox ready");
56 
57 static int cxl_pci_mbox_wait_for_doorbell(struct cxl_dev_state *cxlds)
58 {
59 	const unsigned long start = jiffies;
60 	unsigned long end = start;
61 
62 	while (cxl_doorbell_busy(cxlds)) {
63 		end = jiffies;
64 
65 		if (time_after(end, start + CXL_MAILBOX_TIMEOUT_MS)) {
66 			/* Check again in case preempted before timeout test */
67 			if (!cxl_doorbell_busy(cxlds))
68 				break;
69 			return -ETIMEDOUT;
70 		}
71 		cpu_relax();
72 	}
73 
74 	dev_dbg(cxlds->dev, "Doorbell wait took %dms",
75 		jiffies_to_msecs(end) - jiffies_to_msecs(start));
76 	return 0;
77 }
78 
79 #define cxl_err(dev, status, msg)                                        \
80 	dev_err_ratelimited(dev, msg ", device state %s%s\n",                  \
81 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
82 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
83 
84 #define cxl_cmd_err(dev, cmd, status, msg)                               \
85 	dev_err_ratelimited(dev, msg " (opcode: %#x), device state %s%s\n",    \
86 			    (cmd)->opcode,                                     \
87 			    status & CXLMDEV_DEV_FATAL ? " fatal" : "",        \
88 			    status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
89 
90 /*
91  * Threaded irq dev_id's must be globally unique.  cxl_dev_id provides a unique
92  * wrapper object for each irq within the same cxlds.
93  */
94 struct cxl_dev_id {
95 	struct cxl_dev_state *cxlds;
96 };
97 
98 static int cxl_request_irq(struct cxl_dev_state *cxlds, int irq,
99 			   irq_handler_t thread_fn)
100 {
101 	struct device *dev = cxlds->dev;
102 	struct cxl_dev_id *dev_id;
103 
104 	dev_id = devm_kzalloc(dev, sizeof(*dev_id), GFP_KERNEL);
105 	if (!dev_id)
106 		return -ENOMEM;
107 	dev_id->cxlds = cxlds;
108 
109 	return devm_request_threaded_irq(dev, irq, NULL, thread_fn,
110 					 IRQF_SHARED | IRQF_ONESHOT, NULL,
111 					 dev_id);
112 }
113 
114 static bool cxl_mbox_background_complete(struct cxl_dev_state *cxlds)
115 {
116 	u64 reg;
117 
118 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
119 	return FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_PCT_MASK, reg) == 100;
120 }
121 
122 static irqreturn_t cxl_pci_mbox_irq(int irq, void *id)
123 {
124 	u64 reg;
125 	u16 opcode;
126 	struct cxl_dev_id *dev_id = id;
127 	struct cxl_dev_state *cxlds = dev_id->cxlds;
128 	struct cxl_mailbox *cxl_mbox = &cxlds->cxl_mbox;
129 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
130 
131 	if (!cxl_mbox_background_complete(cxlds))
132 		return IRQ_NONE;
133 
134 	reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
135 	opcode = FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_OPCODE_MASK, reg);
136 	if (opcode == CXL_MBOX_OP_SANITIZE) {
137 		mutex_lock(&cxl_mbox->mbox_mutex);
138 		if (mds->security.sanitize_node)
139 			mod_delayed_work(system_wq, &mds->security.poll_dwork, 0);
140 		mutex_unlock(&cxl_mbox->mbox_mutex);
141 	} else {
142 		/* short-circuit the wait in __cxl_pci_mbox_send_cmd() */
143 		rcuwait_wake_up(&cxl_mbox->mbox_wait);
144 	}
145 
146 	return IRQ_HANDLED;
147 }
148 
149 /*
150  * Sanitization operation polling mode.
151  */
152 static void cxl_mbox_sanitize_work(struct work_struct *work)
153 {
154 	struct cxl_memdev_state *mds =
155 		container_of(work, typeof(*mds), security.poll_dwork.work);
156 	struct cxl_dev_state *cxlds = &mds->cxlds;
157 	struct cxl_mailbox *cxl_mbox = &cxlds->cxl_mbox;
158 
159 	mutex_lock(&cxl_mbox->mbox_mutex);
160 	if (cxl_mbox_background_complete(cxlds)) {
161 		mds->security.poll_tmo_secs = 0;
162 		if (mds->security.sanitize_node)
163 			sysfs_notify_dirent(mds->security.sanitize_node);
164 		mds->security.sanitize_active = false;
165 
166 		dev_dbg(cxlds->dev, "Sanitization operation ended\n");
167 	} else {
168 		int timeout = mds->security.poll_tmo_secs + 10;
169 
170 		mds->security.poll_tmo_secs = min(15 * 60, timeout);
171 		schedule_delayed_work(&mds->security.poll_dwork, timeout * HZ);
172 	}
173 	mutex_unlock(&cxl_mbox->mbox_mutex);
174 }
175 
176 /**
177  * __cxl_pci_mbox_send_cmd() - Execute a mailbox command
178  * @cxl_mbox: CXL mailbox context
179  * @mbox_cmd: Command to send to the memory device.
180  *
181  * Context: Any context. Expects mbox_mutex to be held.
182  * Return: -ETIMEDOUT if timeout occurred waiting for completion. 0 on success.
183  *         Caller should check the return code in @mbox_cmd to make sure it
184  *         succeeded.
185  *
186  * This is a generic form of the CXL mailbox send command thus only using the
187  * registers defined by the mailbox capability ID - CXL 2.0 8.2.8.4. Memory
188  * devices, and perhaps other types of CXL devices may have further information
189  * available upon error conditions. Driver facilities wishing to send mailbox
190  * commands should use the wrapper command.
191  *
192  * The CXL spec allows for up to two mailboxes. The intention is for the primary
193  * mailbox to be OS controlled and the secondary mailbox to be used by system
194  * firmware. This allows the OS and firmware to communicate with the device and
195  * not need to coordinate with each other. The driver only uses the primary
196  * mailbox.
197  */
198 static int __cxl_pci_mbox_send_cmd(struct cxl_mailbox *cxl_mbox,
199 				   struct cxl_mbox_cmd *mbox_cmd)
200 {
201 	struct cxl_dev_state *cxlds = mbox_to_cxlds(cxl_mbox);
202 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
203 	void __iomem *payload = cxlds->regs.mbox + CXLDEV_MBOX_PAYLOAD_OFFSET;
204 	struct device *dev = cxlds->dev;
205 	u64 cmd_reg, status_reg;
206 	size_t out_len;
207 	int rc;
208 
209 	lockdep_assert_held(&cxl_mbox->mbox_mutex);
210 
211 	/*
212 	 * Here are the steps from 8.2.8.4 of the CXL 2.0 spec.
213 	 *   1. Caller reads MB Control Register to verify doorbell is clear
214 	 *   2. Caller writes Command Register
215 	 *   3. Caller writes Command Payload Registers if input payload is non-empty
216 	 *   4. Caller writes MB Control Register to set doorbell
217 	 *   5. Caller either polls for doorbell to be clear or waits for interrupt if configured
218 	 *   6. Caller reads MB Status Register to fetch Return code
219 	 *   7. If command successful, Caller reads Command Register to get Payload Length
220 	 *   8. If output payload is non-empty, host reads Command Payload Registers
221 	 *
222 	 * Hardware is free to do whatever it wants before the doorbell is rung,
223 	 * and isn't allowed to change anything after it clears the doorbell. As
224 	 * such, steps 2 and 3 can happen in any order, and steps 6, 7, 8 can
225 	 * also happen in any order (though some orders might not make sense).
226 	 */
227 
228 	/* #1 */
229 	if (cxl_doorbell_busy(cxlds)) {
230 		u64 md_status =
231 			readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
232 
233 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status,
234 			    "mailbox queue busy");
235 		return -EBUSY;
236 	}
237 
238 	/*
239 	 * With sanitize polling, hardware might be done and the poller still
240 	 * not be in sync. Ensure no new command comes in until so. Keep the
241 	 * hardware semantics and only allow device health status.
242 	 */
243 	if (mds->security.poll_tmo_secs > 0) {
244 		if (mbox_cmd->opcode != CXL_MBOX_OP_GET_HEALTH_INFO)
245 			return -EBUSY;
246 	}
247 
248 	cmd_reg = FIELD_PREP(CXLDEV_MBOX_CMD_COMMAND_OPCODE_MASK,
249 			     mbox_cmd->opcode);
250 	if (mbox_cmd->size_in) {
251 		if (WARN_ON(!mbox_cmd->payload_in))
252 			return -EINVAL;
253 
254 		cmd_reg |= FIELD_PREP(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK,
255 				      mbox_cmd->size_in);
256 		memcpy_toio(payload, mbox_cmd->payload_in, mbox_cmd->size_in);
257 	}
258 
259 	/* #2, #3 */
260 	writeq(cmd_reg, cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
261 
262 	/* #4 */
263 	dev_dbg(dev, "Sending command: 0x%04x\n", mbox_cmd->opcode);
264 	writel(CXLDEV_MBOX_CTRL_DOORBELL,
265 	       cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
266 
267 	/* #5 */
268 	rc = cxl_pci_mbox_wait_for_doorbell(cxlds);
269 	if (rc == -ETIMEDOUT) {
270 		u64 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
271 
272 		cxl_cmd_err(cxlds->dev, mbox_cmd, md_status, "mailbox timeout");
273 		return rc;
274 	}
275 
276 	/* #6 */
277 	status_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_STATUS_OFFSET);
278 	mbox_cmd->return_code =
279 		FIELD_GET(CXLDEV_MBOX_STATUS_RET_CODE_MASK, status_reg);
280 
281 	/*
282 	 * Handle the background command in a synchronous manner.
283 	 *
284 	 * All other mailbox commands will serialize/queue on the mbox_mutex,
285 	 * which we currently hold. Furthermore this also guarantees that
286 	 * cxl_mbox_background_complete() checks are safe amongst each other,
287 	 * in that no new bg operation can occur in between.
288 	 *
289 	 * Background operations are timesliced in accordance with the nature
290 	 * of the command. In the event of timeout, the mailbox state is
291 	 * indeterminate until the next successful command submission and the
292 	 * driver can get back in sync with the hardware state.
293 	 */
294 	if (mbox_cmd->return_code == CXL_MBOX_CMD_RC_BACKGROUND) {
295 		u64 bg_status_reg;
296 		int i, timeout;
297 
298 		/*
299 		 * Sanitization is a special case which monopolizes the device
300 		 * and cannot be timesliced. Handle asynchronously instead,
301 		 * and allow userspace to poll(2) for completion.
302 		 */
303 		if (mbox_cmd->opcode == CXL_MBOX_OP_SANITIZE) {
304 			if (mds->security.sanitize_active)
305 				return -EBUSY;
306 
307 			/* give first timeout a second */
308 			timeout = 1;
309 			mds->security.poll_tmo_secs = timeout;
310 			mds->security.sanitize_active = true;
311 			schedule_delayed_work(&mds->security.poll_dwork,
312 					      timeout * HZ);
313 			dev_dbg(dev, "Sanitization operation started\n");
314 			goto success;
315 		}
316 
317 		dev_dbg(dev, "Mailbox background operation (0x%04x) started\n",
318 			mbox_cmd->opcode);
319 
320 		timeout = mbox_cmd->poll_interval_ms;
321 		for (i = 0; i < mbox_cmd->poll_count; i++) {
322 			if (rcuwait_wait_event_timeout(&cxl_mbox->mbox_wait,
323 						       cxl_mbox_background_complete(cxlds),
324 						       TASK_UNINTERRUPTIBLE,
325 						       msecs_to_jiffies(timeout)) > 0)
326 				break;
327 		}
328 
329 		if (!cxl_mbox_background_complete(cxlds)) {
330 			dev_err(dev, "timeout waiting for background (%d ms)\n",
331 				timeout * mbox_cmd->poll_count);
332 			return -ETIMEDOUT;
333 		}
334 
335 		bg_status_reg = readq(cxlds->regs.mbox +
336 				      CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
337 		mbox_cmd->return_code =
338 			FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_RC_MASK,
339 				  bg_status_reg);
340 		dev_dbg(dev,
341 			"Mailbox background operation (0x%04x) completed\n",
342 			mbox_cmd->opcode);
343 	}
344 
345 	if (mbox_cmd->return_code != CXL_MBOX_CMD_RC_SUCCESS) {
346 		dev_dbg(dev, "Mailbox operation had an error: %s\n",
347 			cxl_mbox_cmd_rc2str(mbox_cmd));
348 		return 0; /* completed but caller must check return_code */
349 	}
350 
351 success:
352 	/* #7 */
353 	cmd_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
354 	out_len = FIELD_GET(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK, cmd_reg);
355 
356 	/* #8 */
357 	if (out_len && mbox_cmd->payload_out) {
358 		/*
359 		 * Sanitize the copy. If hardware misbehaves, out_len per the
360 		 * spec can actually be greater than the max allowed size (21
361 		 * bits available but spec defined 1M max). The caller also may
362 		 * have requested less data than the hardware supplied even
363 		 * within spec.
364 		 */
365 		size_t n;
366 
367 		n = min3(mbox_cmd->size_out, cxl_mbox->payload_size, out_len);
368 		memcpy_fromio(mbox_cmd->payload_out, payload, n);
369 		mbox_cmd->size_out = n;
370 	} else {
371 		mbox_cmd->size_out = 0;
372 	}
373 
374 	return 0;
375 }
376 
377 static int cxl_pci_mbox_send(struct cxl_mailbox *cxl_mbox,
378 			     struct cxl_mbox_cmd *cmd)
379 {
380 	int rc;
381 
382 	mutex_lock_io(&cxl_mbox->mbox_mutex);
383 	rc = __cxl_pci_mbox_send_cmd(cxl_mbox, cmd);
384 	mutex_unlock(&cxl_mbox->mbox_mutex);
385 
386 	return rc;
387 }
388 
389 static int cxl_pci_setup_mailbox(struct cxl_memdev_state *mds, bool irq_avail)
390 {
391 	struct cxl_dev_state *cxlds = &mds->cxlds;
392 	struct cxl_mailbox *cxl_mbox = &cxlds->cxl_mbox;
393 	const int cap = readl(cxlds->regs.mbox + CXLDEV_MBOX_CAPS_OFFSET);
394 	struct device *dev = cxlds->dev;
395 	unsigned long timeout;
396 	int irq, msgnum;
397 	u64 md_status;
398 	u32 ctrl;
399 
400 	timeout = jiffies + mbox_ready_timeout * HZ;
401 	do {
402 		md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
403 		if (md_status & CXLMDEV_MBOX_IF_READY)
404 			break;
405 		if (msleep_interruptible(100))
406 			break;
407 	} while (!time_after(jiffies, timeout));
408 
409 	if (!(md_status & CXLMDEV_MBOX_IF_READY)) {
410 		cxl_err(dev, md_status, "timeout awaiting mailbox ready");
411 		return -ETIMEDOUT;
412 	}
413 
414 	/*
415 	 * A command may be in flight from a previous driver instance,
416 	 * think kexec, do one doorbell wait so that
417 	 * __cxl_pci_mbox_send_cmd() can assume that it is the only
418 	 * source for future doorbell busy events.
419 	 */
420 	if (cxl_pci_mbox_wait_for_doorbell(cxlds) != 0) {
421 		cxl_err(dev, md_status, "timeout awaiting mailbox idle");
422 		return -ETIMEDOUT;
423 	}
424 
425 	cxl_mbox->mbox_send = cxl_pci_mbox_send;
426 	cxl_mbox->payload_size =
427 		1 << FIELD_GET(CXLDEV_MBOX_CAP_PAYLOAD_SIZE_MASK, cap);
428 
429 	/*
430 	 * CXL 2.0 8.2.8.4.3 Mailbox Capabilities Register
431 	 *
432 	 * If the size is too small, mandatory commands will not work and so
433 	 * there's no point in going forward. If the size is too large, there's
434 	 * no harm is soft limiting it.
435 	 */
436 	cxl_mbox->payload_size = min_t(size_t, cxl_mbox->payload_size, SZ_1M);
437 	if (cxl_mbox->payload_size < 256) {
438 		dev_err(dev, "Mailbox is too small (%zub)",
439 			cxl_mbox->payload_size);
440 		return -ENXIO;
441 	}
442 
443 	dev_dbg(dev, "Mailbox payload sized %zu", cxl_mbox->payload_size);
444 
445 	INIT_DELAYED_WORK(&mds->security.poll_dwork, cxl_mbox_sanitize_work);
446 
447 	/* background command interrupts are optional */
448 	if (!(cap & CXLDEV_MBOX_CAP_BG_CMD_IRQ) || !irq_avail)
449 		return 0;
450 
451 	msgnum = FIELD_GET(CXLDEV_MBOX_CAP_IRQ_MSGNUM_MASK, cap);
452 	irq = pci_irq_vector(to_pci_dev(cxlds->dev), msgnum);
453 	if (irq < 0)
454 		return 0;
455 
456 	if (cxl_request_irq(cxlds, irq, cxl_pci_mbox_irq))
457 		return 0;
458 
459 	dev_dbg(cxlds->dev, "Mailbox interrupts enabled\n");
460 	/* enable background command mbox irq support */
461 	ctrl = readl(cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
462 	ctrl |= CXLDEV_MBOX_CTRL_BG_CMD_IRQ;
463 	writel(ctrl, cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
464 
465 	return 0;
466 }
467 
468 /*
469  * Assume that any RCIEP that emits the CXL memory expander class code
470  * is an RCD
471  */
472 static bool is_cxl_restricted(struct pci_dev *pdev)
473 {
474 	return pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END;
475 }
476 
477 static int cxl_rcrb_get_comp_regs(struct pci_dev *pdev,
478 				  struct cxl_register_map *map)
479 {
480 	struct cxl_dport *dport;
481 	resource_size_t component_reg_phys;
482 
483 	*map = (struct cxl_register_map) {
484 		.host = &pdev->dev,
485 		.resource = CXL_RESOURCE_NONE,
486 	};
487 
488 	struct cxl_port *port __free(put_cxl_port) =
489 		cxl_pci_find_port(pdev, &dport);
490 	if (!port)
491 		return -EPROBE_DEFER;
492 
493 	component_reg_phys = cxl_rcd_component_reg_phys(&pdev->dev, dport);
494 	if (component_reg_phys == CXL_RESOURCE_NONE)
495 		return -ENXIO;
496 
497 	map->resource = component_reg_phys;
498 	map->reg_type = CXL_REGLOC_RBI_COMPONENT;
499 	map->max_size = CXL_COMPONENT_REG_BLOCK_SIZE;
500 
501 	return 0;
502 }
503 
504 static int cxl_pci_setup_regs(struct pci_dev *pdev, enum cxl_regloc_type type,
505 			      struct cxl_register_map *map)
506 {
507 	int rc;
508 
509 	rc = cxl_find_regblock(pdev, type, map);
510 
511 	/*
512 	 * If the Register Locator DVSEC does not exist, check if it
513 	 * is an RCH and try to extract the Component Registers from
514 	 * an RCRB.
515 	 */
516 	if (rc && type == CXL_REGLOC_RBI_COMPONENT && is_cxl_restricted(pdev))
517 		rc = cxl_rcrb_get_comp_regs(pdev, map);
518 
519 	if (rc)
520 		return rc;
521 
522 	return cxl_setup_regs(map);
523 }
524 
525 static int cxl_pci_ras_unmask(struct pci_dev *pdev)
526 {
527 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
528 	void __iomem *addr;
529 	u32 orig_val, val, mask;
530 	u16 cap;
531 	int rc;
532 
533 	if (!cxlds->regs.ras) {
534 		dev_dbg(&pdev->dev, "No RAS registers.\n");
535 		return 0;
536 	}
537 
538 	/* BIOS has PCIe AER error control */
539 	if (!pcie_aer_is_native(pdev))
540 		return 0;
541 
542 	rc = pcie_capability_read_word(pdev, PCI_EXP_DEVCTL, &cap);
543 	if (rc)
544 		return rc;
545 
546 	if (cap & PCI_EXP_DEVCTL_URRE) {
547 		addr = cxlds->regs.ras + CXL_RAS_UNCORRECTABLE_MASK_OFFSET;
548 		orig_val = readl(addr);
549 
550 		mask = CXL_RAS_UNCORRECTABLE_MASK_MASK |
551 		       CXL_RAS_UNCORRECTABLE_MASK_F256B_MASK;
552 		val = orig_val & ~mask;
553 		writel(val, addr);
554 		dev_dbg(&pdev->dev,
555 			"Uncorrectable RAS Errors Mask: %#x -> %#x\n",
556 			orig_val, val);
557 	}
558 
559 	if (cap & PCI_EXP_DEVCTL_CERE) {
560 		addr = cxlds->regs.ras + CXL_RAS_CORRECTABLE_MASK_OFFSET;
561 		orig_val = readl(addr);
562 		val = orig_val & ~CXL_RAS_CORRECTABLE_MASK_MASK;
563 		writel(val, addr);
564 		dev_dbg(&pdev->dev, "Correctable RAS Errors Mask: %#x -> %#x\n",
565 			orig_val, val);
566 	}
567 
568 	return 0;
569 }
570 
571 static void free_event_buf(void *buf)
572 {
573 	kvfree(buf);
574 }
575 
576 /*
577  * There is a single buffer for reading event logs from the mailbox.  All logs
578  * share this buffer protected by the mds->event_log_lock.
579  */
580 static int cxl_mem_alloc_event_buf(struct cxl_memdev_state *mds)
581 {
582 	struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
583 	struct cxl_get_event_payload *buf;
584 
585 	buf = kvmalloc(cxl_mbox->payload_size, GFP_KERNEL);
586 	if (!buf)
587 		return -ENOMEM;
588 	mds->event.buf = buf;
589 
590 	return devm_add_action_or_reset(mds->cxlds.dev, free_event_buf, buf);
591 }
592 
593 static bool cxl_alloc_irq_vectors(struct pci_dev *pdev)
594 {
595 	int nvecs;
596 
597 	/*
598 	 * Per CXL 3.0 3.1.1 CXL.io Endpoint a function on a CXL device must
599 	 * not generate INTx messages if that function participates in
600 	 * CXL.cache or CXL.mem.
601 	 *
602 	 * Additionally pci_alloc_irq_vectors() handles calling
603 	 * pci_free_irq_vectors() automatically despite not being called
604 	 * pcim_*.  See pci_setup_msi_context().
605 	 */
606 	nvecs = pci_alloc_irq_vectors(pdev, 1, CXL_PCI_DEFAULT_MAX_VECTORS,
607 				      PCI_IRQ_MSIX | PCI_IRQ_MSI);
608 	if (nvecs < 1) {
609 		dev_dbg(&pdev->dev, "Failed to alloc irq vectors: %d\n", nvecs);
610 		return false;
611 	}
612 	return true;
613 }
614 
615 static irqreturn_t cxl_event_thread(int irq, void *id)
616 {
617 	struct cxl_dev_id *dev_id = id;
618 	struct cxl_dev_state *cxlds = dev_id->cxlds;
619 	struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
620 	u32 status;
621 
622 	do {
623 		/*
624 		 * CXL 3.0 8.2.8.3.1: The lower 32 bits are the status;
625 		 * ignore the reserved upper 32 bits
626 		 */
627 		status = readl(cxlds->regs.status + CXLDEV_DEV_EVENT_STATUS_OFFSET);
628 		/* Ignore logs unknown to the driver */
629 		status &= CXLDEV_EVENT_STATUS_ALL;
630 		if (!status)
631 			break;
632 		cxl_mem_get_event_records(mds, status);
633 		cond_resched();
634 	} while (status);
635 
636 	return IRQ_HANDLED;
637 }
638 
639 static int cxl_event_req_irq(struct cxl_dev_state *cxlds, u8 setting)
640 {
641 	struct pci_dev *pdev = to_pci_dev(cxlds->dev);
642 	int irq;
643 
644 	if (FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting) != CXL_INT_MSI_MSIX)
645 		return -ENXIO;
646 
647 	irq =  pci_irq_vector(pdev,
648 			      FIELD_GET(CXLDEV_EVENT_INT_MSGNUM_MASK, setting));
649 	if (irq < 0)
650 		return irq;
651 
652 	return cxl_request_irq(cxlds, irq, cxl_event_thread);
653 }
654 
655 static int cxl_event_get_int_policy(struct cxl_memdev_state *mds,
656 				    struct cxl_event_interrupt_policy *policy)
657 {
658 	struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
659 	struct cxl_mbox_cmd mbox_cmd = {
660 		.opcode = CXL_MBOX_OP_GET_EVT_INT_POLICY,
661 		.payload_out = policy,
662 		.size_out = sizeof(*policy),
663 	};
664 	int rc;
665 
666 	rc = cxl_internal_send_cmd(cxl_mbox, &mbox_cmd);
667 	if (rc < 0)
668 		dev_err(mds->cxlds.dev,
669 			"Failed to get event interrupt policy : %d", rc);
670 
671 	return rc;
672 }
673 
674 static int cxl_event_config_msgnums(struct cxl_memdev_state *mds,
675 				    struct cxl_event_interrupt_policy *policy)
676 {
677 	struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
678 	struct cxl_mbox_cmd mbox_cmd;
679 	int rc;
680 
681 	*policy = (struct cxl_event_interrupt_policy) {
682 		.info_settings = CXL_INT_MSI_MSIX,
683 		.warn_settings = CXL_INT_MSI_MSIX,
684 		.failure_settings = CXL_INT_MSI_MSIX,
685 		.fatal_settings = CXL_INT_MSI_MSIX,
686 	};
687 
688 	mbox_cmd = (struct cxl_mbox_cmd) {
689 		.opcode = CXL_MBOX_OP_SET_EVT_INT_POLICY,
690 		.payload_in = policy,
691 		.size_in = sizeof(*policy),
692 	};
693 
694 	rc = cxl_internal_send_cmd(cxl_mbox, &mbox_cmd);
695 	if (rc < 0) {
696 		dev_err(mds->cxlds.dev, "Failed to set event interrupt policy : %d",
697 			rc);
698 		return rc;
699 	}
700 
701 	/* Retrieve final interrupt settings */
702 	return cxl_event_get_int_policy(mds, policy);
703 }
704 
705 static int cxl_event_irqsetup(struct cxl_memdev_state *mds)
706 {
707 	struct cxl_dev_state *cxlds = &mds->cxlds;
708 	struct cxl_event_interrupt_policy policy;
709 	int rc;
710 
711 	rc = cxl_event_config_msgnums(mds, &policy);
712 	if (rc)
713 		return rc;
714 
715 	rc = cxl_event_req_irq(cxlds, policy.info_settings);
716 	if (rc) {
717 		dev_err(cxlds->dev, "Failed to get interrupt for event Info log\n");
718 		return rc;
719 	}
720 
721 	rc = cxl_event_req_irq(cxlds, policy.warn_settings);
722 	if (rc) {
723 		dev_err(cxlds->dev, "Failed to get interrupt for event Warn log\n");
724 		return rc;
725 	}
726 
727 	rc = cxl_event_req_irq(cxlds, policy.failure_settings);
728 	if (rc) {
729 		dev_err(cxlds->dev, "Failed to get interrupt for event Failure log\n");
730 		return rc;
731 	}
732 
733 	rc = cxl_event_req_irq(cxlds, policy.fatal_settings);
734 	if (rc) {
735 		dev_err(cxlds->dev, "Failed to get interrupt for event Fatal log\n");
736 		return rc;
737 	}
738 
739 	return 0;
740 }
741 
742 static bool cxl_event_int_is_fw(u8 setting)
743 {
744 	u8 mode = FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting);
745 
746 	return mode == CXL_INT_FW;
747 }
748 
749 static int cxl_event_config(struct pci_host_bridge *host_bridge,
750 			    struct cxl_memdev_state *mds, bool irq_avail)
751 {
752 	struct cxl_event_interrupt_policy policy;
753 	int rc;
754 
755 	/*
756 	 * When BIOS maintains CXL error reporting control, it will process
757 	 * event records.  Only one agent can do so.
758 	 */
759 	if (!host_bridge->native_cxl_error)
760 		return 0;
761 
762 	if (!irq_avail) {
763 		dev_info(mds->cxlds.dev, "No interrupt support, disable event processing.\n");
764 		return 0;
765 	}
766 
767 	rc = cxl_mem_alloc_event_buf(mds);
768 	if (rc)
769 		return rc;
770 
771 	rc = cxl_event_get_int_policy(mds, &policy);
772 	if (rc)
773 		return rc;
774 
775 	if (cxl_event_int_is_fw(policy.info_settings) ||
776 	    cxl_event_int_is_fw(policy.warn_settings) ||
777 	    cxl_event_int_is_fw(policy.failure_settings) ||
778 	    cxl_event_int_is_fw(policy.fatal_settings)) {
779 		dev_err(mds->cxlds.dev,
780 			"FW still in control of Event Logs despite _OSC settings\n");
781 		return -EBUSY;
782 	}
783 
784 	rc = cxl_event_irqsetup(mds);
785 	if (rc)
786 		return rc;
787 
788 	cxl_mem_get_event_records(mds, CXLDEV_EVENT_STATUS_ALL);
789 
790 	return 0;
791 }
792 
793 static int cxl_pci_type3_init_mailbox(struct cxl_dev_state *cxlds)
794 {
795 	int rc;
796 
797 	/*
798 	 * Fail the init if there's no mailbox. For a type3 this is out of spec.
799 	 */
800 	if (!cxlds->reg_map.device_map.mbox.valid)
801 		return -ENODEV;
802 
803 	rc = cxl_mailbox_init(&cxlds->cxl_mbox, cxlds->dev);
804 	if (rc)
805 		return rc;
806 
807 	return 0;
808 }
809 
810 static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
811 {
812 	struct pci_host_bridge *host_bridge = pci_find_host_bridge(pdev->bus);
813 	struct cxl_memdev_state *mds;
814 	struct cxl_dev_state *cxlds;
815 	struct cxl_register_map map;
816 	struct cxl_memdev *cxlmd;
817 	int i, rc, pmu_count;
818 	bool irq_avail;
819 
820 	/*
821 	 * Double check the anonymous union trickery in struct cxl_regs
822 	 * FIXME switch to struct_group()
823 	 */
824 	BUILD_BUG_ON(offsetof(struct cxl_regs, memdev) !=
825 		     offsetof(struct cxl_regs, device_regs.memdev));
826 
827 	rc = pcim_enable_device(pdev);
828 	if (rc)
829 		return rc;
830 	pci_set_master(pdev);
831 
832 	mds = cxl_memdev_state_create(&pdev->dev);
833 	if (IS_ERR(mds))
834 		return PTR_ERR(mds);
835 	cxlds = &mds->cxlds;
836 	pci_set_drvdata(pdev, cxlds);
837 
838 	cxlds->rcd = is_cxl_restricted(pdev);
839 	cxlds->serial = pci_get_dsn(pdev);
840 	cxlds->cxl_dvsec = pci_find_dvsec_capability(
841 		pdev, PCI_VENDOR_ID_CXL, CXL_DVSEC_PCIE_DEVICE);
842 	if (!cxlds->cxl_dvsec)
843 		dev_warn(&pdev->dev,
844 			 "Device DVSEC not present, skip CXL.mem init\n");
845 
846 	rc = cxl_pci_setup_regs(pdev, CXL_REGLOC_RBI_MEMDEV, &map);
847 	if (rc)
848 		return rc;
849 
850 	rc = cxl_map_device_regs(&map, &cxlds->regs.device_regs);
851 	if (rc)
852 		return rc;
853 
854 	/*
855 	 * If the component registers can't be found, the cxl_pci driver may
856 	 * still be useful for management functions so don't return an error.
857 	 */
858 	rc = cxl_pci_setup_regs(pdev, CXL_REGLOC_RBI_COMPONENT,
859 				&cxlds->reg_map);
860 	if (rc)
861 		dev_warn(&pdev->dev, "No component registers (%d)\n", rc);
862 	else if (!cxlds->reg_map.component_map.ras.valid)
863 		dev_dbg(&pdev->dev, "RAS registers not found\n");
864 
865 	rc = cxl_map_component_regs(&cxlds->reg_map, &cxlds->regs.component,
866 				    BIT(CXL_CM_CAP_CAP_ID_RAS));
867 	if (rc)
868 		dev_dbg(&pdev->dev, "Failed to map RAS capability.\n");
869 
870 	rc = cxl_pci_type3_init_mailbox(cxlds);
871 	if (rc)
872 		return rc;
873 
874 	rc = cxl_await_media_ready(cxlds);
875 	if (rc == 0)
876 		cxlds->media_ready = true;
877 	else
878 		dev_warn(&pdev->dev, "Media not active (%d)\n", rc);
879 
880 	irq_avail = cxl_alloc_irq_vectors(pdev);
881 
882 	rc = cxl_pci_setup_mailbox(mds, irq_avail);
883 	if (rc)
884 		return rc;
885 
886 	rc = cxl_enumerate_cmds(mds);
887 	if (rc)
888 		return rc;
889 
890 	rc = cxl_set_timestamp(mds);
891 	if (rc)
892 		return rc;
893 
894 	rc = cxl_poison_state_init(mds);
895 	if (rc)
896 		return rc;
897 
898 	rc = cxl_dev_state_identify(mds);
899 	if (rc)
900 		return rc;
901 
902 	rc = cxl_mem_create_range_info(mds);
903 	if (rc)
904 		return rc;
905 
906 	cxlmd = devm_cxl_add_memdev(&pdev->dev, cxlds);
907 	if (IS_ERR(cxlmd))
908 		return PTR_ERR(cxlmd);
909 
910 	rc = devm_cxl_setup_fw_upload(&pdev->dev, mds);
911 	if (rc)
912 		return rc;
913 
914 	rc = devm_cxl_sanitize_setup_notifier(&pdev->dev, cxlmd);
915 	if (rc)
916 		return rc;
917 
918 	pmu_count = cxl_count_regblock(pdev, CXL_REGLOC_RBI_PMU);
919 	for (i = 0; i < pmu_count; i++) {
920 		struct cxl_pmu_regs pmu_regs;
921 
922 		rc = cxl_find_regblock_instance(pdev, CXL_REGLOC_RBI_PMU, &map, i);
923 		if (rc) {
924 			dev_dbg(&pdev->dev, "Could not find PMU regblock\n");
925 			break;
926 		}
927 
928 		rc = cxl_map_pmu_regs(&map, &pmu_regs);
929 		if (rc) {
930 			dev_dbg(&pdev->dev, "Could not map PMU regs\n");
931 			break;
932 		}
933 
934 		rc = devm_cxl_pmu_add(cxlds->dev, &pmu_regs, cxlmd->id, i, CXL_PMU_MEMDEV);
935 		if (rc) {
936 			dev_dbg(&pdev->dev, "Could not add PMU instance\n");
937 			break;
938 		}
939 	}
940 
941 	rc = cxl_event_config(host_bridge, mds, irq_avail);
942 	if (rc)
943 		return rc;
944 
945 	rc = cxl_pci_ras_unmask(pdev);
946 	if (rc)
947 		dev_dbg(&pdev->dev, "No RAS reporting unmasked\n");
948 
949 	pci_save_state(pdev);
950 
951 	return rc;
952 }
953 
954 static const struct pci_device_id cxl_mem_pci_tbl[] = {
955 	/* PCI class code for CXL.mem Type-3 Devices */
956 	{ PCI_DEVICE_CLASS((PCI_CLASS_MEMORY_CXL << 8 | CXL_MEMORY_PROGIF), ~0)},
957 	{ /* terminate list */ },
958 };
959 MODULE_DEVICE_TABLE(pci, cxl_mem_pci_tbl);
960 
961 static pci_ers_result_t cxl_slot_reset(struct pci_dev *pdev)
962 {
963 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
964 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
965 	struct device *dev = &cxlmd->dev;
966 
967 	dev_info(&pdev->dev, "%s: restart CXL.mem after slot reset\n",
968 		 dev_name(dev));
969 	pci_restore_state(pdev);
970 	if (device_attach(dev) <= 0)
971 		return PCI_ERS_RESULT_DISCONNECT;
972 	return PCI_ERS_RESULT_RECOVERED;
973 }
974 
975 static void cxl_error_resume(struct pci_dev *pdev)
976 {
977 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
978 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
979 	struct device *dev = &cxlmd->dev;
980 
981 	dev_info(&pdev->dev, "%s: error resume %s\n", dev_name(dev),
982 		 dev->driver ? "successful" : "failed");
983 }
984 
985 static void cxl_reset_done(struct pci_dev *pdev)
986 {
987 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
988 	struct cxl_memdev *cxlmd = cxlds->cxlmd;
989 	struct device *dev = &pdev->dev;
990 
991 	/*
992 	 * FLR does not expect to touch the HDM decoders and related
993 	 * registers.  SBR, however, will wipe all device configurations.
994 	 * Issue a warning if there was an active decoder before the reset
995 	 * that no longer exists.
996 	 */
997 	guard(device)(&cxlmd->dev);
998 	if (cxlmd->endpoint &&
999 	    cxl_endpoint_decoder_reset_detected(cxlmd->endpoint)) {
1000 		dev_crit(dev, "SBR happened without memory regions removal.\n");
1001 		dev_crit(dev, "System may be unstable if regions hosted system memory.\n");
1002 		add_taint(TAINT_USER, LOCKDEP_STILL_OK);
1003 	}
1004 }
1005 
1006 static const struct pci_error_handlers cxl_error_handlers = {
1007 	.error_detected	= cxl_error_detected,
1008 	.slot_reset	= cxl_slot_reset,
1009 	.resume		= cxl_error_resume,
1010 	.cor_error_detected	= cxl_cor_error_detected,
1011 	.reset_done	= cxl_reset_done,
1012 };
1013 
1014 static struct pci_driver cxl_pci_driver = {
1015 	.name			= KBUILD_MODNAME,
1016 	.id_table		= cxl_mem_pci_tbl,
1017 	.probe			= cxl_pci_probe,
1018 	.err_handler		= &cxl_error_handlers,
1019 	.driver	= {
1020 		.probe_type	= PROBE_PREFER_ASYNCHRONOUS,
1021 	},
1022 };
1023 
1024 #define CXL_EVENT_HDR_FLAGS_REC_SEVERITY GENMASK(1, 0)
1025 static void cxl_handle_cper_event(enum cxl_event_type ev_type,
1026 				  struct cxl_cper_event_rec *rec)
1027 {
1028 	struct cper_cxl_event_devid *device_id = &rec->hdr.device_id;
1029 	struct pci_dev *pdev __free(pci_dev_put) = NULL;
1030 	enum cxl_event_log_type log_type;
1031 	struct cxl_dev_state *cxlds;
1032 	unsigned int devfn;
1033 	u32 hdr_flags;
1034 
1035 	pr_debug("CPER event %d for device %u:%u:%u.%u\n", ev_type,
1036 		 device_id->segment_num, device_id->bus_num,
1037 		 device_id->device_num, device_id->func_num);
1038 
1039 	devfn = PCI_DEVFN(device_id->device_num, device_id->func_num);
1040 	pdev = pci_get_domain_bus_and_slot(device_id->segment_num,
1041 					   device_id->bus_num, devfn);
1042 	if (!pdev)
1043 		return;
1044 
1045 	guard(device)(&pdev->dev);
1046 	if (pdev->driver != &cxl_pci_driver)
1047 		return;
1048 
1049 	cxlds = pci_get_drvdata(pdev);
1050 	if (!cxlds)
1051 		return;
1052 
1053 	/* Fabricate a log type */
1054 	hdr_flags = get_unaligned_le24(rec->event.generic.hdr.flags);
1055 	log_type = FIELD_GET(CXL_EVENT_HDR_FLAGS_REC_SEVERITY, hdr_flags);
1056 
1057 	cxl_event_trace_record(cxlds->cxlmd, log_type, ev_type,
1058 			       &uuid_null, &rec->event);
1059 }
1060 
1061 static void cxl_cper_work_fn(struct work_struct *work)
1062 {
1063 	struct cxl_cper_work_data wd;
1064 
1065 	while (cxl_cper_kfifo_get(&wd))
1066 		cxl_handle_cper_event(wd.event_type, &wd.rec);
1067 }
1068 static DECLARE_WORK(cxl_cper_work, cxl_cper_work_fn);
1069 
1070 static int __init cxl_pci_driver_init(void)
1071 {
1072 	int rc;
1073 
1074 	rc = pci_register_driver(&cxl_pci_driver);
1075 	if (rc)
1076 		return rc;
1077 
1078 	rc = cxl_cper_register_work(&cxl_cper_work);
1079 	if (rc)
1080 		pci_unregister_driver(&cxl_pci_driver);
1081 
1082 	return rc;
1083 }
1084 
1085 static void __exit cxl_pci_driver_exit(void)
1086 {
1087 	cxl_cper_unregister_work(&cxl_cper_work);
1088 	cancel_work_sync(&cxl_cper_work);
1089 	pci_unregister_driver(&cxl_pci_driver);
1090 }
1091 
1092 module_init(cxl_pci_driver_init);
1093 module_exit(cxl_pci_driver_exit);
1094 MODULE_DESCRIPTION("CXL: PCI manageability");
1095 MODULE_LICENSE("GPL v2");
1096 MODULE_IMPORT_NS(CXL);
1097