1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright(c) 2020 Intel Corporation. All rights reserved. */
3 #include <linux/io-64-nonatomic-lo-hi.h>
4 #include <linux/moduleparam.h>
5 #include <linux/module.h>
6 #include <linux/delay.h>
7 #include <linux/sizes.h>
8 #include <linux/mutex.h>
9 #include <linux/list.h>
10 #include <linux/pci.h>
19 * This implements the PCI exclusive functionality for a CXL device as it is
20 * defined by the Compute Express Link specification. CXL devices may surface
21 * certain functionality even if it isn't CXL enabled. While this driver is
22 * focused around the PCI specific aspects of a CXL device, it binds to the
23 * specific CXL memory device class code, and therefore the implementation of
24 * cxl_pci is focused around CXL memory devices.
26 * The driver has several responsibilities, mainly:
27 * - Create the memX device and register on the CXL bus.
28 * - Enumerate device's register interface and map them.
29 * - Registers nvdimm bridge device with cxl_core.
30 * - Registers a CXL mailbox with cxl_core.
33 #define cxl_doorbell_busy(cxlds) \
34 (readl((cxlds)->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET) & \
35 CXLDEV_MBOX_CTRL_DOORBELL)
37 /* CXL 2.0 - 8.2.8.4 */
38 #define CXL_MAILBOX_TIMEOUT_MS (2 * HZ)
41 * CXL 2.0 ECN "Add Mailbox Ready Time" defines a capability field to
42 * dictate how long to wait for the mailbox to become ready. The new
43 * field allows the device to tell software the amount of time to wait
44 * before mailbox ready. This field per the spec theoretically allows
45 * for up to 255 seconds. 255 seconds is unreasonably long, its longer
46 * than the maximum SATA port link recovery wait. Default to 60 seconds
47 * until someone builds a CXL device that needs more time in practice.
49 static unsigned short mbox_ready_timeout = 60;
50 module_param(mbox_ready_timeout, ushort, 0644);
51 MODULE_PARM_DESC(mbox_ready_timeout,
52 "seconds to wait for mailbox ready / memory active status");
54 static int cxl_pci_mbox_wait_for_doorbell(struct cxl_dev_state *cxlds)
56 const unsigned long start = jiffies;
57 unsigned long end = start;
59 while (cxl_doorbell_busy(cxlds)) {
62 if (time_after(end, start + CXL_MAILBOX_TIMEOUT_MS)) {
63 /* Check again in case preempted before timeout test */
64 if (!cxl_doorbell_busy(cxlds))
71 dev_dbg(cxlds->dev, "Doorbell wait took %dms",
72 jiffies_to_msecs(end) - jiffies_to_msecs(start));
76 #define cxl_err(dev, status, msg) \
77 dev_err_ratelimited(dev, msg ", device state %s%s\n", \
78 status & CXLMDEV_DEV_FATAL ? " fatal" : "", \
79 status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
81 #define cxl_cmd_err(dev, cmd, status, msg) \
82 dev_err_ratelimited(dev, msg " (opcode: %#x), device state %s%s\n", \
84 status & CXLMDEV_DEV_FATAL ? " fatal" : "", \
85 status & CXLMDEV_FW_HALT ? " firmware-halt" : "")
88 * __cxl_pci_mbox_send_cmd() - Execute a mailbox command
89 * @cxlds: The device state to communicate with.
90 * @mbox_cmd: Command to send to the memory device.
92 * Context: Any context. Expects mbox_mutex to be held.
93 * Return: -ETIMEDOUT if timeout occurred waiting for completion. 0 on success.
94 * Caller should check the return code in @mbox_cmd to make sure it
97 * This is a generic form of the CXL mailbox send command thus only using the
98 * registers defined by the mailbox capability ID - CXL 2.0 8.2.8.4. Memory
99 * devices, and perhaps other types of CXL devices may have further information
100 * available upon error conditions. Driver facilities wishing to send mailbox
101 * commands should use the wrapper command.
103 * The CXL spec allows for up to two mailboxes. The intention is for the primary
104 * mailbox to be OS controlled and the secondary mailbox to be used by system
105 * firmware. This allows the OS and firmware to communicate with the device and
106 * not need to coordinate with each other. The driver only uses the primary
109 static int __cxl_pci_mbox_send_cmd(struct cxl_dev_state *cxlds,
110 struct cxl_mbox_cmd *mbox_cmd)
112 void __iomem *payload = cxlds->regs.mbox + CXLDEV_MBOX_PAYLOAD_OFFSET;
113 struct device *dev = cxlds->dev;
114 u64 cmd_reg, status_reg;
118 lockdep_assert_held(&cxlds->mbox_mutex);
121 * Here are the steps from 8.2.8.4 of the CXL 2.0 spec.
122 * 1. Caller reads MB Control Register to verify doorbell is clear
123 * 2. Caller writes Command Register
124 * 3. Caller writes Command Payload Registers if input payload is non-empty
125 * 4. Caller writes MB Control Register to set doorbell
126 * 5. Caller either polls for doorbell to be clear or waits for interrupt if configured
127 * 6. Caller reads MB Status Register to fetch Return code
128 * 7. If command successful, Caller reads Command Register to get Payload Length
129 * 8. If output payload is non-empty, host reads Command Payload Registers
131 * Hardware is free to do whatever it wants before the doorbell is rung,
132 * and isn't allowed to change anything after it clears the doorbell. As
133 * such, steps 2 and 3 can happen in any order, and steps 6, 7, 8 can
134 * also happen in any order (though some orders might not make sense).
138 if (cxl_doorbell_busy(cxlds)) {
140 readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
142 cxl_cmd_err(cxlds->dev, mbox_cmd, md_status,
143 "mailbox queue busy");
147 cmd_reg = FIELD_PREP(CXLDEV_MBOX_CMD_COMMAND_OPCODE_MASK,
149 if (mbox_cmd->size_in) {
150 if (WARN_ON(!mbox_cmd->payload_in))
153 cmd_reg |= FIELD_PREP(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK,
155 memcpy_toio(payload, mbox_cmd->payload_in, mbox_cmd->size_in);
159 writeq(cmd_reg, cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
162 dev_dbg(dev, "Sending command\n");
163 writel(CXLDEV_MBOX_CTRL_DOORBELL,
164 cxlds->regs.mbox + CXLDEV_MBOX_CTRL_OFFSET);
167 rc = cxl_pci_mbox_wait_for_doorbell(cxlds);
168 if (rc == -ETIMEDOUT) {
169 u64 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
171 cxl_cmd_err(cxlds->dev, mbox_cmd, md_status, "mailbox timeout");
176 status_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_STATUS_OFFSET);
177 mbox_cmd->return_code =
178 FIELD_GET(CXLDEV_MBOX_STATUS_RET_CODE_MASK, status_reg);
180 if (mbox_cmd->return_code != 0) {
181 dev_dbg(dev, "Mailbox operation had an error\n");
186 cmd_reg = readq(cxlds->regs.mbox + CXLDEV_MBOX_CMD_OFFSET);
187 out_len = FIELD_GET(CXLDEV_MBOX_CMD_PAYLOAD_LENGTH_MASK, cmd_reg);
190 if (out_len && mbox_cmd->payload_out) {
192 * Sanitize the copy. If hardware misbehaves, out_len per the
193 * spec can actually be greater than the max allowed size (21
194 * bits available but spec defined 1M max). The caller also may
195 * have requested less data than the hardware supplied even
198 size_t n = min3(mbox_cmd->size_out, cxlds->payload_size, out_len);
200 memcpy_fromio(mbox_cmd->payload_out, payload, n);
201 mbox_cmd->size_out = n;
203 mbox_cmd->size_out = 0;
209 static int cxl_pci_mbox_send(struct cxl_dev_state *cxlds, struct cxl_mbox_cmd *cmd)
213 mutex_lock_io(&cxlds->mbox_mutex);
214 rc = __cxl_pci_mbox_send_cmd(cxlds, cmd);
215 mutex_unlock(&cxlds->mbox_mutex);
220 static int cxl_pci_setup_mailbox(struct cxl_dev_state *cxlds)
222 const int cap = readl(cxlds->regs.mbox + CXLDEV_MBOX_CAPS_OFFSET);
223 unsigned long timeout;
226 timeout = jiffies + mbox_ready_timeout * HZ;
228 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
229 if (md_status & CXLMDEV_MBOX_IF_READY)
231 if (msleep_interruptible(100))
233 } while (!time_after(jiffies, timeout));
235 if (!(md_status & CXLMDEV_MBOX_IF_READY)) {
236 cxl_err(cxlds->dev, md_status,
237 "timeout awaiting mailbox ready");
242 * A command may be in flight from a previous driver instance,
243 * think kexec, do one doorbell wait so that
244 * __cxl_pci_mbox_send_cmd() can assume that it is the only
245 * source for future doorbell busy events.
247 if (cxl_pci_mbox_wait_for_doorbell(cxlds) != 0) {
248 cxl_err(cxlds->dev, md_status, "timeout awaiting mailbox idle");
252 cxlds->mbox_send = cxl_pci_mbox_send;
253 cxlds->payload_size =
254 1 << FIELD_GET(CXLDEV_MBOX_CAP_PAYLOAD_SIZE_MASK, cap);
257 * CXL 2.0 8.2.8.4.3 Mailbox Capabilities Register
259 * If the size is too small, mandatory commands will not work and so
260 * there's no point in going forward. If the size is too large, there's
261 * no harm is soft limiting it.
263 cxlds->payload_size = min_t(size_t, cxlds->payload_size, SZ_1M);
264 if (cxlds->payload_size < 256) {
265 dev_err(cxlds->dev, "Mailbox is too small (%zub)",
266 cxlds->payload_size);
270 dev_dbg(cxlds->dev, "Mailbox payload sized %zu",
271 cxlds->payload_size);
276 static int cxl_map_regblock(struct pci_dev *pdev, struct cxl_register_map *map)
279 int bar = map->barno;
280 struct device *dev = &pdev->dev;
281 resource_size_t offset = map->block_offset;
283 /* Basic sanity check that BAR is big enough */
284 if (pci_resource_len(pdev, bar) < offset) {
285 dev_err(dev, "BAR%d: %pr: too small (offset: %pa)\n", bar,
286 &pdev->resource[bar], &offset);
290 addr = pci_iomap(pdev, bar, 0);
292 dev_err(dev, "failed to map registers\n");
296 dev_dbg(dev, "Mapped CXL Memory Device resource bar %u @ %pa\n",
299 map->base = addr + map->block_offset;
303 static void cxl_unmap_regblock(struct pci_dev *pdev,
304 struct cxl_register_map *map)
306 pci_iounmap(pdev, map->base - map->block_offset);
310 static int cxl_probe_regs(struct pci_dev *pdev, struct cxl_register_map *map)
312 struct cxl_component_reg_map *comp_map;
313 struct cxl_device_reg_map *dev_map;
314 struct device *dev = &pdev->dev;
315 void __iomem *base = map->base;
317 switch (map->reg_type) {
318 case CXL_REGLOC_RBI_COMPONENT:
319 comp_map = &map->component_map;
320 cxl_probe_component_regs(dev, base, comp_map);
321 if (!comp_map->hdm_decoder.valid) {
322 dev_err(dev, "HDM decoder registers not found\n");
326 dev_dbg(dev, "Set up component registers\n");
328 case CXL_REGLOC_RBI_MEMDEV:
329 dev_map = &map->device_map;
330 cxl_probe_device_regs(dev, base, dev_map);
331 if (!dev_map->status.valid || !dev_map->mbox.valid ||
332 !dev_map->memdev.valid) {
333 dev_err(dev, "registers not found: %s%s%s\n",
334 !dev_map->status.valid ? "status " : "",
335 !dev_map->mbox.valid ? "mbox " : "",
336 !dev_map->memdev.valid ? "memdev " : "");
340 dev_dbg(dev, "Probing device registers...\n");
349 static int cxl_map_regs(struct cxl_dev_state *cxlds, struct cxl_register_map *map)
351 struct device *dev = cxlds->dev;
352 struct pci_dev *pdev = to_pci_dev(dev);
354 switch (map->reg_type) {
355 case CXL_REGLOC_RBI_COMPONENT:
356 cxl_map_component_regs(pdev, &cxlds->regs.component, map);
357 dev_dbg(dev, "Mapping component registers...\n");
359 case CXL_REGLOC_RBI_MEMDEV:
360 cxl_map_device_regs(pdev, &cxlds->regs.device_regs, map);
361 dev_dbg(dev, "Probing device registers...\n");
370 static int cxl_setup_regs(struct pci_dev *pdev, enum cxl_regloc_type type,
371 struct cxl_register_map *map)
375 rc = cxl_find_regblock(pdev, type, map);
379 rc = cxl_map_regblock(pdev, map);
383 rc = cxl_probe_regs(pdev, map);
384 cxl_unmap_regblock(pdev, map);
389 static int wait_for_valid(struct cxl_dev_state *cxlds)
391 struct pci_dev *pdev = to_pci_dev(cxlds->dev);
392 int d = cxlds->cxl_dvsec, rc;
396 * Memory_Info_Valid: When set, indicates that the CXL Range 1 Size high
397 * and Size Low registers are valid. Must be set within 1 second of
398 * deassertion of reset to CXL device. Likely it is already set by the
399 * time this runs, but otherwise give a 1.5 second timeout in case of
402 rc = pci_read_config_dword(pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(0), &val);
406 if (val & CXL_DVSEC_MEM_INFO_VALID)
411 rc = pci_read_config_dword(pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(0), &val);
415 if (val & CXL_DVSEC_MEM_INFO_VALID)
422 * Wait up to @mbox_ready_timeout for the device to report memory
425 static int wait_for_media_ready(struct cxl_dev_state *cxlds)
427 struct pci_dev *pdev = to_pci_dev(cxlds->dev);
428 int d = cxlds->cxl_dvsec;
433 rc = wait_for_valid(cxlds);
437 for (i = mbox_ready_timeout; i; i--) {
441 rc = pci_read_config_dword(
442 pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(0), &temp);
446 active = FIELD_GET(CXL_DVSEC_MEM_ACTIVE, temp);
454 "timeout awaiting memory active after %d seconds\n",
459 md_status = readq(cxlds->regs.memdev + CXLMDEV_STATUS_OFFSET);
460 if (!CXLMDEV_READY(md_status))
466 static int cxl_dvsec_ranges(struct cxl_dev_state *cxlds)
468 struct cxl_endpoint_dvsec_info *info = &cxlds->info;
469 struct pci_dev *pdev = to_pci_dev(cxlds->dev);
470 int d = cxlds->cxl_dvsec;
471 int hdm_count, rc, i;
477 rc = pci_read_config_word(pdev, d + CXL_DVSEC_CAP_OFFSET, &cap);
481 rc = pci_read_config_word(pdev, d + CXL_DVSEC_CTRL_OFFSET, &ctrl);
485 if (!(cap & CXL_DVSEC_MEM_CAPABLE))
489 * It is not allowed by spec for MEM.capable to be set and have 0 legacy
490 * HDM decoders (values > 2 are also undefined as of CXL 2.0). As this
491 * driver is for a spec defined class code which must be CXL.mem
492 * capable, there is no point in continuing to enable CXL.mem.
494 hdm_count = FIELD_GET(CXL_DVSEC_HDM_COUNT_MASK, cap);
495 if (!hdm_count || hdm_count > 2)
498 rc = wait_for_valid(cxlds);
502 info->mem_enabled = FIELD_GET(CXL_DVSEC_MEM_ENABLE, ctrl);
504 for (i = 0; i < hdm_count; i++) {
508 rc = pci_read_config_dword(
509 pdev, d + CXL_DVSEC_RANGE_SIZE_HIGH(i), &temp);
513 size = (u64)temp << 32;
515 rc = pci_read_config_dword(
516 pdev, d + CXL_DVSEC_RANGE_SIZE_LOW(i), &temp);
520 size |= temp & CXL_DVSEC_MEM_SIZE_LOW_MASK;
522 rc = pci_read_config_dword(
523 pdev, d + CXL_DVSEC_RANGE_BASE_HIGH(i), &temp);
527 base = (u64)temp << 32;
529 rc = pci_read_config_dword(
530 pdev, d + CXL_DVSEC_RANGE_BASE_LOW(i), &temp);
534 base |= temp & CXL_DVSEC_MEM_BASE_LOW_MASK;
536 info->dvsec_range[i] = (struct range) {
538 .end = base + size - 1
548 static int cxl_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
550 struct cxl_register_map map;
551 struct cxl_memdev *cxlmd;
552 struct cxl_dev_state *cxlds;
556 * Double check the anonymous union trickery in struct cxl_regs
557 * FIXME switch to struct_group()
559 BUILD_BUG_ON(offsetof(struct cxl_regs, memdev) !=
560 offsetof(struct cxl_regs, device_regs.memdev));
562 rc = pcim_enable_device(pdev);
566 cxlds = cxl_dev_state_create(&pdev->dev);
568 return PTR_ERR(cxlds);
570 cxlds->serial = pci_get_dsn(pdev);
571 cxlds->cxl_dvsec = pci_find_dvsec_capability(
572 pdev, PCI_DVSEC_VENDOR_ID_CXL, CXL_DVSEC_PCIE_DEVICE);
573 if (!cxlds->cxl_dvsec)
575 "Device DVSEC not present, skip CXL.mem init\n");
577 cxlds->wait_media_ready = wait_for_media_ready;
579 rc = cxl_setup_regs(pdev, CXL_REGLOC_RBI_MEMDEV, &map);
583 rc = cxl_map_regs(cxlds, &map);
588 * If the component registers can't be found, the cxl_pci driver may
589 * still be useful for management functions so don't return an error.
591 cxlds->component_reg_phys = CXL_RESOURCE_NONE;
592 rc = cxl_setup_regs(pdev, CXL_REGLOC_RBI_COMPONENT, &map);
594 dev_warn(&pdev->dev, "No component registers (%d)\n", rc);
596 cxlds->component_reg_phys = cxl_regmap_to_base(pdev, &map);
598 rc = cxl_pci_setup_mailbox(cxlds);
602 rc = cxl_enumerate_cmds(cxlds);
606 rc = cxl_dev_state_identify(cxlds);
610 rc = cxl_mem_create_range_info(cxlds);
614 rc = cxl_dvsec_ranges(cxlds);
617 "Failed to get DVSEC range information (%d)\n", rc);
619 cxlmd = devm_cxl_add_memdev(cxlds);
621 return PTR_ERR(cxlmd);
623 if (range_len(&cxlds->pmem_range) && IS_ENABLED(CONFIG_CXL_PMEM))
624 rc = devm_cxl_add_nvdimm(&pdev->dev, cxlmd);
629 static const struct pci_device_id cxl_mem_pci_tbl[] = {
630 /* PCI class code for CXL.mem Type-3 Devices */
631 { PCI_DEVICE_CLASS((PCI_CLASS_MEMORY_CXL << 8 | CXL_MEMORY_PROGIF), ~0)},
632 { /* terminate list */ },
634 MODULE_DEVICE_TABLE(pci, cxl_mem_pci_tbl);
636 static struct pci_driver cxl_pci_driver = {
637 .name = KBUILD_MODNAME,
638 .id_table = cxl_mem_pci_tbl,
639 .probe = cxl_pci_probe,
641 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
645 MODULE_LICENSE("GPL v2");
646 module_pci_driver(cxl_pci_driver);
647 MODULE_IMPORT_NS(CXL);