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 */
53static unsigned short mbox_ready_timeout = 60;
54module_param(mbox_ready_timeout, ushort, 0644);
55MODULE_PARM_DESC(mbox_ready_timeout, "seconds to wait for mailbox ready");
56
57static 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 */
94struct cxl_dev_id {
95 struct cxl_dev_state *cxlds;
96};
97
98static 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, size: 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
114static bool cxl_mbox_background_complete(struct cxl_dev_state *cxlds)
115{
116 u64 reg;
117
118 reg = readq(addr: cxlds->regs.mbox + CXLDEV_MBOX_BG_CMD_STATUS_OFFSET);
119 return FIELD_GET(CXLDEV_MBOX_BG_CMD_COMMAND_PCT_MASK, reg) == 100;
120}
121
122static 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(addr: 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(wq: system_percpu_wq, dwork: &mds->security.poll_dwork, delay: 0);
140 mutex_unlock(lock: &cxl_mbox->mbox_mutex);
141 } else {
142 /* short-circuit the wait in __cxl_pci_mbox_send_cmd() */
143 rcuwait_wake_up(w: &cxl_mbox->mbox_wait);
144 }
145
146 return IRQ_HANDLED;
147}
148
149/*
150 * Sanitization operation polling mode.
151 */
152static 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(kn: 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(dwork: &mds->security.poll_dwork, delay: timeout * HZ);
172 }
173 mutex_unlock(lock: &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 */
198static 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(addr: 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(val: cmd_reg, addr: 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 addr: 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(addr: 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(addr: 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(dwork: &mds->security.poll_dwork,
312 delay: 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(addr: 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
351success:
352 /* #7 */
353 cmd_reg = readq(addr: 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
377static int cxl_pci_mbox_send(struct cxl_mailbox *cxl_mbox,
378 struct cxl_mbox_cmd *cmd)
379{
380 int rc;
381
382 mutex_lock(&cxl_mbox->mbox_mutex);
383 rc = __cxl_pci_mbox_send_cmd(cxl_mbox, mbox_cmd: cmd);
384 mutex_unlock(lock: &cxl_mbox->mbox_mutex);
385
386 return rc;
387}
388
389static 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(addr: 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(addr: cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
403 if (md_status & CXLMDEV_MBOX_IF_READY)
404 break;
405 if (msleep_interruptible(msecs: 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), nr: msgnum);
453 if (irq < 0)
454 return 0;
455
456 if (cxl_request_irq(cxlds, irq, thread_fn: 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(addr: cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
462 ctrl |= CXLDEV_MBOX_CTRL_BG_CMD_IRQ;
463 writel(val: ctrl, addr: 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 */
472static bool is_cxl_restricted(struct pci_dev *pdev)
473{
474 return pci_pcie_type(dev: pdev) == PCI_EXP_TYPE_RC_END;
475}
476
477static int cxl_rcrb_get_comp_regs(struct pci_dev *pdev,
478 struct cxl_register_map *map,
479 struct cxl_dport *dport)
480{
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: &dport);
490 if (!port)
491 return -EPROBE_DEFER;
492
493 component_reg_phys = cxl_rcd_component_reg_phys(dev: &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
504static 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 struct cxl_dport *dport;
518 struct cxl_port *port __free(put_cxl_port) =
519 cxl_pci_find_port(pdev, dport: &dport);
520 if (!port)
521 return -EPROBE_DEFER;
522
523 rc = cxl_rcrb_get_comp_regs(pdev, map, dport);
524 if (rc)
525 return rc;
526
527 rc = cxl_dport_map_rcd_linkcap(pdev, dport);
528 if (rc)
529 return rc;
530
531 } else if (rc) {
532 return rc;
533 }
534
535 return cxl_setup_regs(map);
536}
537
538static int cxl_pci_ras_unmask(struct pci_dev *pdev)
539{
540 struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
541 void __iomem *addr;
542 u32 orig_val, val, mask;
543 u16 cap;
544 int rc;
545
546 if (!cxlds->regs.ras) {
547 dev_dbg(&pdev->dev, "No RAS registers.\n");
548 return 0;
549 }
550
551 /* BIOS has PCIe AER error control */
552 if (!pcie_aer_is_native(dev: pdev))
553 return 0;
554
555 rc = pcie_capability_read_word(dev: pdev, PCI_EXP_DEVCTL, val: &cap);
556 if (rc)
557 return rc;
558
559 if (cap & PCI_EXP_DEVCTL_URRE) {
560 addr = cxlds->regs.ras + CXL_RAS_UNCORRECTABLE_MASK_OFFSET;
561 orig_val = readl(addr);
562
563 mask = CXL_RAS_UNCORRECTABLE_MASK_MASK |
564 CXL_RAS_UNCORRECTABLE_MASK_F256B_MASK;
565 val = orig_val & ~mask;
566 writel(val, addr);
567 dev_dbg(&pdev->dev,
568 "Uncorrectable RAS Errors Mask: %#x -> %#x\n",
569 orig_val, val);
570 }
571
572 if (cap & PCI_EXP_DEVCTL_CERE) {
573 addr = cxlds->regs.ras + CXL_RAS_CORRECTABLE_MASK_OFFSET;
574 orig_val = readl(addr);
575 val = orig_val & ~CXL_RAS_CORRECTABLE_MASK_MASK;
576 writel(val, addr);
577 dev_dbg(&pdev->dev, "Correctable RAS Errors Mask: %#x -> %#x\n",
578 orig_val, val);
579 }
580
581 return 0;
582}
583
584static void free_event_buf(void *buf)
585{
586 kvfree(addr: buf);
587}
588
589/*
590 * There is a single buffer for reading event logs from the mailbox. All logs
591 * share this buffer protected by the mds->event_log_lock.
592 */
593static int cxl_mem_alloc_event_buf(struct cxl_memdev_state *mds)
594{
595 struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
596 struct cxl_get_event_payload *buf;
597
598 buf = kvmalloc(cxl_mbox->payload_size, GFP_KERNEL);
599 if (!buf)
600 return -ENOMEM;
601 mds->event.buf = buf;
602
603 return devm_add_action_or_reset(mds->cxlds.dev, free_event_buf, buf);
604}
605
606static bool cxl_alloc_irq_vectors(struct pci_dev *pdev)
607{
608 int nvecs;
609
610 /*
611 * Per CXL 3.0 3.1.1 CXL.io Endpoint a function on a CXL device must
612 * not generate INTx messages if that function participates in
613 * CXL.cache or CXL.mem.
614 *
615 * Additionally pci_alloc_irq_vectors() handles calling
616 * pci_free_irq_vectors() automatically despite not being called
617 * pcim_*. See pci_setup_msi_context().
618 */
619 nvecs = pci_alloc_irq_vectors(dev: pdev, min_vecs: 1, CXL_PCI_DEFAULT_MAX_VECTORS,
620 PCI_IRQ_MSIX | PCI_IRQ_MSI);
621 if (nvecs < 1) {
622 dev_dbg(&pdev->dev, "Failed to alloc irq vectors: %d\n", nvecs);
623 return false;
624 }
625 return true;
626}
627
628static irqreturn_t cxl_event_thread(int irq, void *id)
629{
630 struct cxl_dev_id *dev_id = id;
631 struct cxl_dev_state *cxlds = dev_id->cxlds;
632 struct cxl_memdev_state *mds = to_cxl_memdev_state(cxlds);
633 u32 status;
634
635 do {
636 /*
637 * CXL 3.0 8.2.8.3.1: The lower 32 bits are the status;
638 * ignore the reserved upper 32 bits
639 */
640 status = readl(addr: cxlds->regs.status + CXLDEV_DEV_EVENT_STATUS_OFFSET);
641 /* Ignore logs unknown to the driver */
642 status &= CXLDEV_EVENT_STATUS_ALL;
643 if (!status)
644 break;
645 cxl_mem_get_event_records(mds, status);
646 cond_resched();
647 } while (status);
648
649 return IRQ_HANDLED;
650}
651
652static int cxl_event_req_irq(struct cxl_dev_state *cxlds, u8 setting)
653{
654 struct pci_dev *pdev = to_pci_dev(cxlds->dev);
655 int irq;
656
657 if (FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting) != CXL_INT_MSI_MSIX)
658 return -ENXIO;
659
660 irq = pci_irq_vector(dev: pdev,
661 FIELD_GET(CXLDEV_EVENT_INT_MSGNUM_MASK, setting));
662 if (irq < 0)
663 return irq;
664
665 return cxl_request_irq(cxlds, irq, thread_fn: cxl_event_thread);
666}
667
668static int cxl_event_get_int_policy(struct cxl_memdev_state *mds,
669 struct cxl_event_interrupt_policy *policy)
670{
671 struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
672 struct cxl_mbox_cmd mbox_cmd = {
673 .opcode = CXL_MBOX_OP_GET_EVT_INT_POLICY,
674 .payload_out = policy,
675 .size_out = sizeof(*policy),
676 };
677 int rc;
678
679 rc = cxl_internal_send_cmd(cxl_mbox, cmd: &mbox_cmd);
680 if (rc < 0)
681 dev_err(mds->cxlds.dev,
682 "Failed to get event interrupt policy : %d", rc);
683
684 return rc;
685}
686
687static int cxl_event_config_msgnums(struct cxl_memdev_state *mds,
688 struct cxl_event_interrupt_policy *policy)
689{
690 struct cxl_mailbox *cxl_mbox = &mds->cxlds.cxl_mbox;
691 struct cxl_mbox_cmd mbox_cmd;
692 int rc;
693
694 *policy = (struct cxl_event_interrupt_policy) {
695 .info_settings = CXL_INT_MSI_MSIX,
696 .warn_settings = CXL_INT_MSI_MSIX,
697 .failure_settings = CXL_INT_MSI_MSIX,
698 .fatal_settings = CXL_INT_MSI_MSIX,
699 };
700
701 mbox_cmd = (struct cxl_mbox_cmd) {
702 .opcode = CXL_MBOX_OP_SET_EVT_INT_POLICY,
703 .payload_in = policy,
704 .size_in = sizeof(*policy),
705 };
706
707 rc = cxl_internal_send_cmd(cxl_mbox, cmd: &mbox_cmd);
708 if (rc < 0) {
709 dev_err(mds->cxlds.dev, "Failed to set event interrupt policy : %d",
710 rc);
711 return rc;
712 }
713
714 /* Retrieve final interrupt settings */
715 return cxl_event_get_int_policy(mds, policy);
716}
717
718static int cxl_event_irqsetup(struct cxl_memdev_state *mds)
719{
720 struct cxl_dev_state *cxlds = &mds->cxlds;
721 struct cxl_event_interrupt_policy policy;
722 int rc;
723
724 rc = cxl_event_config_msgnums(mds, policy: &policy);
725 if (rc)
726 return rc;
727
728 rc = cxl_event_req_irq(cxlds, setting: policy.info_settings);
729 if (rc) {
730 dev_err(cxlds->dev, "Failed to get interrupt for event Info log\n");
731 return rc;
732 }
733
734 rc = cxl_event_req_irq(cxlds, setting: policy.warn_settings);
735 if (rc) {
736 dev_err(cxlds->dev, "Failed to get interrupt for event Warn log\n");
737 return rc;
738 }
739
740 rc = cxl_event_req_irq(cxlds, setting: policy.failure_settings);
741 if (rc) {
742 dev_err(cxlds->dev, "Failed to get interrupt for event Failure log\n");
743 return rc;
744 }
745
746 rc = cxl_event_req_irq(cxlds, setting: policy.fatal_settings);
747 if (rc) {
748 dev_err(cxlds->dev, "Failed to get interrupt for event Fatal log\n");
749 return rc;
750 }
751
752 return 0;
753}
754
755static bool cxl_event_int_is_fw(u8 setting)
756{
757 u8 mode = FIELD_GET(CXLDEV_EVENT_INT_MODE_MASK, setting);
758
759 return mode == CXL_INT_FW;
760}
761
762static int cxl_event_config(struct pci_host_bridge *host_bridge,
763 struct cxl_memdev_state *mds, bool irq_avail)
764{
765 struct cxl_event_interrupt_policy policy;
766 int rc;
767
768 /*
769 * When BIOS maintains CXL error reporting control, it will process
770 * event records. Only one agent can do so.
771 */
772 if (!host_bridge->native_cxl_error)
773 return 0;
774
775 if (!irq_avail) {
776 dev_info(mds->cxlds.dev, "No interrupt support, disable event processing.\n");
777 return 0;
778 }
779
780 rc = cxl_event_get_int_policy(mds, policy: &policy);
781 if (rc)
782 return rc;
783
784 if (cxl_event_int_is_fw(setting: policy.info_settings) ||
785 cxl_event_int_is_fw(setting: policy.warn_settings) ||
786 cxl_event_int_is_fw(setting: policy.failure_settings) ||
787 cxl_event_int_is_fw(setting: policy.fatal_settings)) {
788 dev_err(mds->cxlds.dev,
789 "FW still in control of Event Logs despite _OSC settings\n");
790 return -EBUSY;
791 }
792
793 rc = cxl_mem_alloc_event_buf(mds);
794 if (rc)
795 return rc;
796
797 rc = cxl_event_irqsetup(mds);
798 if (rc)
799 return rc;
800
801 cxl_mem_get_event_records(mds, CXLDEV_EVENT_STATUS_ALL);
802
803 return 0;
804}
805
806static int cxl_pci_type3_init_mailbox(struct cxl_dev_state *cxlds)
807{
808 int rc;
809
810 /*
811 * Fail the init if there's no mailbox. For a type3 this is out of spec.
812 */
813 if (!cxlds->reg_map.device_map.mbox.valid)
814 return -ENODEV;
815
816 rc = cxl_mailbox_init(cxl_mbox: &cxlds->cxl_mbox, host: cxlds->dev);
817 if (rc)
818 return rc;
819
820 return 0;
821}
822
823static ssize_t rcd_pcie_cap_emit(struct device *dev, u16 offset, char *buf, size_t width)
824{
825 struct cxl_dev_state *cxlds = dev_get_drvdata(dev);
826 struct cxl_memdev *cxlmd = cxlds->cxlmd;
827 struct device *root_dev;
828 struct cxl_dport *dport;
829 struct cxl_port *root __free(put_cxl_port) =
830 cxl_mem_find_port(cxlmd, dport: &dport);
831
832 if (!root)
833 return -ENXIO;
834
835 root_dev = root->uport_dev;
836 if (!root_dev)
837 return -ENXIO;
838
839 if (!dport->regs.rcd_pcie_cap)
840 return -ENXIO;
841
842 guard(device)(T: root_dev);
843 if (!root_dev->driver)
844 return -ENXIO;
845
846 switch (width) {
847 case 2:
848 return sysfs_emit(buf, fmt: "%#x\n",
849 readw(addr: dport->regs.rcd_pcie_cap + offset));
850 case 4:
851 return sysfs_emit(buf, fmt: "%#x\n",
852 readl(addr: dport->regs.rcd_pcie_cap + offset));
853 default:
854 return -EINVAL;
855 }
856}
857
858static ssize_t rcd_link_cap_show(struct device *dev,
859 struct device_attribute *attr, char *buf)
860{
861 return rcd_pcie_cap_emit(dev, PCI_EXP_LNKCAP, buf, width: sizeof(u32));
862}
863static DEVICE_ATTR_RO(rcd_link_cap);
864
865static ssize_t rcd_link_ctrl_show(struct device *dev,
866 struct device_attribute *attr, char *buf)
867{
868 return rcd_pcie_cap_emit(dev, PCI_EXP_LNKCTL, buf, width: sizeof(u16));
869}
870static DEVICE_ATTR_RO(rcd_link_ctrl);
871
872static ssize_t rcd_link_status_show(struct device *dev,
873 struct device_attribute *attr, char *buf)
874{
875 return rcd_pcie_cap_emit(dev, PCI_EXP_LNKSTA, buf, width: sizeof(u16));
876}
877static DEVICE_ATTR_RO(rcd_link_status);
878
879static struct attribute *cxl_rcd_attrs[] = {
880 &dev_attr_rcd_link_cap.attr,
881 &dev_attr_rcd_link_ctrl.attr,
882 &dev_attr_rcd_link_status.attr,
883 NULL
884};
885
886static umode_t cxl_rcd_visible(struct kobject *kobj, struct attribute *a, int n)
887{
888 struct device *dev = kobj_to_dev(kobj);
889 struct pci_dev *pdev = to_pci_dev(dev);
890
891 if (is_cxl_restricted(pdev))
892 return a->mode;
893
894 return 0;
895}
896
897static struct attribute_group cxl_rcd_group = {
898 .attrs = cxl_rcd_attrs,
899 .is_visible = cxl_rcd_visible,
900};
901__ATTRIBUTE_GROUPS(cxl_rcd);
902
903static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
904{
905 struct pci_host_bridge *host_bridge = pci_find_host_bridge(bus: pdev->bus);
906 struct cxl_dpa_info range_info = { 0 };
907 struct cxl_memdev_state *mds;
908 struct cxl_dev_state *cxlds;
909 struct cxl_register_map map;
910 struct cxl_memdev *cxlmd;
911 int rc, pmu_count;
912 unsigned int i;
913 bool irq_avail;
914
915 /*
916 * Double check the anonymous union trickery in struct cxl_regs
917 * FIXME switch to struct_group()
918 */
919 BUILD_BUG_ON(offsetof(struct cxl_regs, memdev) !=
920 offsetof(struct cxl_regs, device_regs.memdev));
921
922 rc = pcim_enable_device(pdev);
923 if (rc)
924 return rc;
925 pci_set_master(dev: pdev);
926
927 mds = cxl_memdev_state_create(dev: &pdev->dev);
928 if (IS_ERR(ptr: mds))
929 return PTR_ERR(ptr: mds);
930 cxlds = &mds->cxlds;
931 pci_set_drvdata(pdev, data: cxlds);
932
933 cxlds->rcd = is_cxl_restricted(pdev);
934 cxlds->serial = pci_get_dsn(dev: pdev);
935 cxlds->cxl_dvsec = pci_find_dvsec_capability(
936 dev: pdev, PCI_VENDOR_ID_CXL, CXL_DVSEC_PCIE_DEVICE);
937 if (!cxlds->cxl_dvsec)
938 dev_warn(&pdev->dev,
939 "Device DVSEC not present, skip CXL.mem init\n");
940
941 rc = cxl_pci_setup_regs(pdev, type: CXL_REGLOC_RBI_MEMDEV, map: &map);
942 if (rc)
943 return rc;
944
945 rc = cxl_map_device_regs(map: &map, regs: &cxlds->regs.device_regs);
946 if (rc)
947 return rc;
948
949 /*
950 * If the component registers can't be found, the cxl_pci driver may
951 * still be useful for management functions so don't return an error.
952 */
953 rc = cxl_pci_setup_regs(pdev, type: CXL_REGLOC_RBI_COMPONENT,
954 map: &cxlds->reg_map);
955 if (rc)
956 dev_warn(&pdev->dev, "No component registers (%d)\n", rc);
957 else if (!cxlds->reg_map.component_map.ras.valid)
958 dev_dbg(&pdev->dev, "RAS registers not found\n");
959
960 rc = cxl_map_component_regs(map: &cxlds->reg_map, regs: &cxlds->regs.component,
961 BIT(CXL_CM_CAP_CAP_ID_RAS));
962 if (rc)
963 dev_dbg(&pdev->dev, "Failed to map RAS capability.\n");
964
965 rc = cxl_pci_type3_init_mailbox(cxlds);
966 if (rc)
967 return rc;
968
969 rc = cxl_await_media_ready(cxlds);
970 if (rc == 0)
971 cxlds->media_ready = true;
972 else
973 dev_warn(&pdev->dev, "Media not active (%d)\n", rc);
974
975 irq_avail = cxl_alloc_irq_vectors(pdev);
976
977 rc = cxl_pci_setup_mailbox(mds, irq_avail);
978 if (rc)
979 return rc;
980
981 rc = cxl_enumerate_cmds(mds);
982 if (rc)
983 return rc;
984
985 rc = cxl_set_timestamp(mds);
986 if (rc)
987 return rc;
988
989 rc = cxl_poison_state_init(mds);
990 if (rc)
991 return rc;
992
993 rc = cxl_dev_state_identify(mds);
994 if (rc)
995 return rc;
996
997 rc = cxl_mem_dpa_fetch(mds, info: &range_info);
998 if (rc)
999 return rc;
1000
1001 rc = cxl_dpa_setup(cxlds, info: &range_info);
1002 if (rc)
1003 return rc;
1004
1005 rc = devm_cxl_setup_features(cxlds);
1006 if (rc)
1007 dev_dbg(&pdev->dev, "No CXL Features discovered\n");
1008
1009 cxlmd = devm_cxl_add_memdev(host: &pdev->dev, cxlds);
1010 if (IS_ERR(ptr: cxlmd))
1011 return PTR_ERR(ptr: cxlmd);
1012
1013 rc = devm_cxl_setup_fw_upload(host: &pdev->dev, mds);
1014 if (rc)
1015 return rc;
1016
1017 rc = devm_cxl_sanitize_setup_notifier(host: &pdev->dev, cxlmd);
1018 if (rc)
1019 return rc;
1020
1021 rc = devm_cxl_setup_fwctl(host: &pdev->dev, cxlmd);
1022 if (rc)
1023 dev_dbg(&pdev->dev, "No CXL FWCTL setup\n");
1024
1025 pmu_count = cxl_count_regblock(pdev, type: CXL_REGLOC_RBI_PMU);
1026 if (pmu_count < 0)
1027 return pmu_count;
1028
1029 for (i = 0; i < pmu_count; i++) {
1030 struct cxl_pmu_regs pmu_regs;
1031
1032 rc = cxl_find_regblock_instance(pdev, type: CXL_REGLOC_RBI_PMU, map: &map, index: i);
1033 if (rc) {
1034 dev_dbg(&pdev->dev, "Could not find PMU regblock\n");
1035 break;
1036 }
1037
1038 rc = cxl_map_pmu_regs(map: &map, regs: &pmu_regs);
1039 if (rc) {
1040 dev_dbg(&pdev->dev, "Could not map PMU regs\n");
1041 break;
1042 }
1043
1044 rc = devm_cxl_pmu_add(parent: cxlds->dev, regs: &pmu_regs, assoc_id: cxlmd->id, idx: i, type: CXL_PMU_MEMDEV);
1045 if (rc) {
1046 dev_dbg(&pdev->dev, "Could not add PMU instance\n");
1047 break;
1048 }
1049 }
1050
1051 rc = cxl_event_config(host_bridge, mds, irq_avail);
1052 if (rc)
1053 return rc;
1054
1055 if (cxl_pci_ras_unmask(pdev))
1056 dev_dbg(&pdev->dev, "No RAS reporting unmasked\n");
1057
1058 pci_save_state(dev: pdev);
1059
1060 return rc;
1061}
1062
1063static const struct pci_device_id cxl_mem_pci_tbl[] = {
1064 /* PCI class code for CXL.mem Type-3 Devices */
1065 { PCI_DEVICE_CLASS((PCI_CLASS_MEMORY_CXL << 8 | CXL_MEMORY_PROGIF), ~0)},
1066 { /* terminate list */ },
1067};
1068MODULE_DEVICE_TABLE(pci, cxl_mem_pci_tbl);
1069
1070static pci_ers_result_t cxl_slot_reset(struct pci_dev *pdev)
1071{
1072 struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
1073 struct cxl_memdev *cxlmd = cxlds->cxlmd;
1074 struct device *dev = &cxlmd->dev;
1075
1076 dev_info(&pdev->dev, "%s: restart CXL.mem after slot reset\n",
1077 dev_name(dev));
1078 pci_restore_state(dev: pdev);
1079 if (device_attach(dev) <= 0)
1080 return PCI_ERS_RESULT_DISCONNECT;
1081 return PCI_ERS_RESULT_RECOVERED;
1082}
1083
1084static void cxl_error_resume(struct pci_dev *pdev)
1085{
1086 struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
1087 struct cxl_memdev *cxlmd = cxlds->cxlmd;
1088 struct device *dev = &cxlmd->dev;
1089
1090 dev_info(&pdev->dev, "%s: error resume %s\n", dev_name(dev),
1091 dev->driver ? "successful" : "failed");
1092}
1093
1094static void cxl_reset_done(struct pci_dev *pdev)
1095{
1096 struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
1097 struct cxl_memdev *cxlmd = cxlds->cxlmd;
1098 struct device *dev = &pdev->dev;
1099
1100 /*
1101 * FLR does not expect to touch the HDM decoders and related
1102 * registers. SBR, however, will wipe all device configurations.
1103 * Issue a warning if there was an active decoder before the reset
1104 * that no longer exists.
1105 */
1106 guard(device)(T: &cxlmd->dev);
1107 if (cxlmd->endpoint &&
1108 cxl_endpoint_decoder_reset_detected(port: cxlmd->endpoint)) {
1109 dev_crit(dev, "SBR happened without memory regions removal.\n");
1110 dev_crit(dev, "System may be unstable if regions hosted system memory.\n");
1111 add_taint(TAINT_USER, LOCKDEP_STILL_OK);
1112 }
1113}
1114
1115static const struct pci_error_handlers cxl_error_handlers = {
1116 .error_detected = cxl_error_detected,
1117 .slot_reset = cxl_slot_reset,
1118 .resume = cxl_error_resume,
1119 .cor_error_detected = cxl_cor_error_detected,
1120 .reset_done = cxl_reset_done,
1121};
1122
1123static struct pci_driver cxl_pci_driver = {
1124 .name = KBUILD_MODNAME,
1125 .id_table = cxl_mem_pci_tbl,
1126 .probe = cxl_pci_probe,
1127 .err_handler = &cxl_error_handlers,
1128 .dev_groups = cxl_rcd_groups,
1129 .driver = {
1130 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
1131 },
1132};
1133
1134#define CXL_EVENT_HDR_FLAGS_REC_SEVERITY GENMASK(1, 0)
1135static void cxl_handle_cper_event(enum cxl_event_type ev_type,
1136 struct cxl_cper_event_rec *rec)
1137{
1138 struct cper_cxl_event_devid *device_id = &rec->hdr.device_id;
1139 struct pci_dev *pdev __free(pci_dev_put) = NULL;
1140 enum cxl_event_log_type log_type;
1141 struct cxl_dev_state *cxlds;
1142 unsigned int devfn;
1143 u32 hdr_flags;
1144
1145 pr_debug("CPER event %d for device %u:%u:%u.%u\n", ev_type,
1146 device_id->segment_num, device_id->bus_num,
1147 device_id->device_num, device_id->func_num);
1148
1149 devfn = PCI_DEVFN(device_id->device_num, device_id->func_num);
1150 pdev = pci_get_domain_bus_and_slot(domain: device_id->segment_num,
1151 bus: device_id->bus_num, devfn);
1152 if (!pdev)
1153 return;
1154
1155 guard(device)(T: &pdev->dev);
1156 if (pdev->driver != &cxl_pci_driver)
1157 return;
1158
1159 cxlds = pci_get_drvdata(pdev);
1160 if (!cxlds)
1161 return;
1162
1163 /* Fabricate a log type */
1164 hdr_flags = get_unaligned_le24(p: rec->event.generic.hdr.flags);
1165 log_type = FIELD_GET(CXL_EVENT_HDR_FLAGS_REC_SEVERITY, hdr_flags);
1166
1167 cxl_event_trace_record(cxlmd: cxlds->cxlmd, type: log_type, event_type: ev_type,
1168 uuid: &uuid_null, evt: &rec->event);
1169}
1170
1171static void cxl_cper_work_fn(struct work_struct *work)
1172{
1173 struct cxl_cper_work_data wd;
1174
1175 while (cxl_cper_kfifo_get(wd: &wd))
1176 cxl_handle_cper_event(ev_type: wd.event_type, rec: &wd.rec);
1177}
1178static DECLARE_WORK(cxl_cper_work, cxl_cper_work_fn);
1179
1180static int __init cxl_pci_driver_init(void)
1181{
1182 int rc;
1183
1184 rc = pci_register_driver(&cxl_pci_driver);
1185 if (rc)
1186 return rc;
1187
1188 rc = cxl_cper_register_work(work: &cxl_cper_work);
1189 if (rc)
1190 pci_unregister_driver(dev: &cxl_pci_driver);
1191
1192 return rc;
1193}
1194
1195static void __exit cxl_pci_driver_exit(void)
1196{
1197 cxl_cper_unregister_work(work: &cxl_cper_work);
1198 cancel_work_sync(work: &cxl_cper_work);
1199 pci_unregister_driver(dev: &cxl_pci_driver);
1200}
1201
1202module_init(cxl_pci_driver_init);
1203module_exit(cxl_pci_driver_exit);
1204MODULE_DESCRIPTION("CXL: PCI manageability");
1205MODULE_LICENSE("GPL v2");
1206MODULE_IMPORT_NS("CXL");
1207

source code of linux/drivers/cxl/pci.c