1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright 2016-2019 HabanaLabs, Ltd.
9 #define pr_fmt(fmt) "habanalabs: " fmt
11 #include "habanalabs.h"
13 #include <linux/pci.h>
14 #include <linux/aer.h>
15 #include <linux/module.h>
17 #define HL_DRIVER_AUTHOR "HabanaLabs Kernel Driver Team"
19 #define HL_DRIVER_DESC "Driver for HabanaLabs's AI Accelerators"
21 MODULE_AUTHOR(HL_DRIVER_AUTHOR);
22 MODULE_DESCRIPTION(HL_DRIVER_DESC);
23 MODULE_LICENSE("GPL v2");
26 static struct class *hl_class;
27 static DEFINE_IDR(hl_devs_idr);
28 static DEFINE_MUTEX(hl_devs_idr_lock);
30 static int timeout_locked = 30;
31 static int reset_on_lockup = 1;
32 static int memory_scrub = 1;
33 static ulong boot_error_status_mask = ULONG_MAX;
35 module_param(timeout_locked, int, 0444);
36 MODULE_PARM_DESC(timeout_locked,
37 "Device lockup timeout in seconds (0 = disabled, default 30s)");
39 module_param(reset_on_lockup, int, 0444);
40 MODULE_PARM_DESC(reset_on_lockup,
41 "Do device reset on lockup (0 = no, 1 = yes, default yes)");
43 module_param(memory_scrub, int, 0444);
44 MODULE_PARM_DESC(memory_scrub,
45 "Scrub device memory in various states (0 = no, 1 = yes, default yes)");
47 module_param(boot_error_status_mask, ulong, 0444);
48 MODULE_PARM_DESC(boot_error_status_mask,
49 "Mask of the error status during device CPU boot (If bitX is cleared then error X is masked. Default all 1's)");
51 #define PCI_VENDOR_ID_HABANALABS 0x1da3
53 #define PCI_IDS_GOYA 0x0001
54 #define PCI_IDS_GAUDI 0x1000
55 #define PCI_IDS_GAUDI_SEC 0x1010
57 static const struct pci_device_id ids[] = {
58 { PCI_DEVICE(PCI_VENDOR_ID_HABANALABS, PCI_IDS_GOYA), },
59 { PCI_DEVICE(PCI_VENDOR_ID_HABANALABS, PCI_IDS_GAUDI), },
60 { PCI_DEVICE(PCI_VENDOR_ID_HABANALABS, PCI_IDS_GAUDI_SEC), },
63 MODULE_DEVICE_TABLE(pci, ids);
66 * get_asic_type - translate device id to asic type
68 * @device: id of the PCI device
70 * Translate device id to asic type.
71 * In case of unidentified device, return -1
73 static enum hl_asic_type get_asic_type(u16 device)
75 enum hl_asic_type asic_type;
79 asic_type = ASIC_GOYA;
82 asic_type = ASIC_GAUDI;
84 case PCI_IDS_GAUDI_SEC:
85 asic_type = ASIC_GAUDI_SEC;
88 asic_type = ASIC_INVALID;
95 static bool is_asic_secured(enum hl_asic_type asic_type)
106 * hl_device_open - open function for habanalabs device
108 * @inode: pointer to inode structure
109 * @filp: pointer to file structure
111 * Called when process opens an habanalabs device.
113 int hl_device_open(struct inode *inode, struct file *filp)
115 enum hl_device_status status;
116 struct hl_device *hdev;
117 struct hl_fpriv *hpriv;
120 mutex_lock(&hl_devs_idr_lock);
121 hdev = idr_find(&hl_devs_idr, iminor(inode));
122 mutex_unlock(&hl_devs_idr_lock);
125 pr_err("Couldn't find device %d:%d\n",
126 imajor(inode), iminor(inode));
130 hpriv = kzalloc(sizeof(*hpriv), GFP_KERNEL);
135 filp->private_data = hpriv;
137 mutex_init(&hpriv->restore_phase_mutex);
138 kref_init(&hpriv->refcount);
139 nonseekable_open(inode, filp);
141 hl_cb_mgr_init(&hpriv->cb_mgr);
142 hl_ctx_mgr_init(&hpriv->ctx_mgr);
144 hpriv->taskpid = find_get_pid(current->pid);
146 mutex_lock(&hdev->fpriv_list_lock);
148 if (!hl_device_operational(hdev, &status)) {
149 dev_err_ratelimited(hdev->dev,
150 "Can't open %s because it is %s\n",
151 dev_name(hdev->dev), hdev->status[status]);
156 if (hdev->in_debug) {
157 dev_err_ratelimited(hdev->dev,
158 "Can't open %s because it is being debugged by another user\n",
159 dev_name(hdev->dev));
164 if (hdev->compute_ctx) {
165 dev_dbg_ratelimited(hdev->dev,
166 "Can't open %s because another user is working on it\n",
167 dev_name(hdev->dev));
172 rc = hl_ctx_create(hdev, hpriv);
174 dev_err(hdev->dev, "Failed to create context %d\n", rc);
178 /* Device is IDLE at this point so it is legal to change PLLs.
179 * There is no need to check anything because if the PLL is
180 * already HIGH, the set function will return without doing
183 hl_device_set_frequency(hdev, PLL_HIGH);
185 list_add(&hpriv->dev_node, &hdev->fpriv_list);
186 mutex_unlock(&hdev->fpriv_list_lock);
188 hl_debugfs_add_file(hpriv);
193 mutex_unlock(&hdev->fpriv_list_lock);
195 hl_cb_mgr_fini(hpriv->hdev, &hpriv->cb_mgr);
196 hl_ctx_mgr_fini(hpriv->hdev, &hpriv->ctx_mgr);
197 filp->private_data = NULL;
198 mutex_destroy(&hpriv->restore_phase_mutex);
199 put_pid(hpriv->taskpid);
206 int hl_device_open_ctrl(struct inode *inode, struct file *filp)
208 struct hl_device *hdev;
209 struct hl_fpriv *hpriv;
212 mutex_lock(&hl_devs_idr_lock);
213 hdev = idr_find(&hl_devs_idr, iminor(inode));
214 mutex_unlock(&hl_devs_idr_lock);
217 pr_err("Couldn't find device %d:%d\n",
218 imajor(inode), iminor(inode));
222 hpriv = kzalloc(sizeof(*hpriv), GFP_KERNEL);
226 mutex_lock(&hdev->fpriv_list_lock);
228 if (!hl_device_operational(hdev, NULL)) {
229 dev_err_ratelimited(hdev->dev_ctrl,
230 "Can't open %s because it is disabled or in reset\n",
231 dev_name(hdev->dev_ctrl));
236 list_add(&hpriv->dev_node, &hdev->fpriv_list);
237 mutex_unlock(&hdev->fpriv_list_lock);
240 filp->private_data = hpriv;
242 hpriv->is_control = true;
243 nonseekable_open(inode, filp);
245 hpriv->taskpid = find_get_pid(current->pid);
250 mutex_unlock(&hdev->fpriv_list_lock);
255 static void set_driver_behavior_per_device(struct hl_device *hdev)
257 hdev->fw_components = FW_TYPE_ALL_TYPES;
258 hdev->cpu_queues_enable = 1;
260 hdev->mmu_enable = 1;
261 hdev->clock_gating_mask = ULONG_MAX;
262 hdev->sram_scrambler_enable = 1;
263 hdev->dram_scrambler_enable = 1;
264 hdev->bmc_enable = 1;
265 hdev->hard_reset_on_fw_events = 1;
266 hdev->reset_on_preboot_fail = 1;
268 hdev->reset_pcilink = 0;
273 * create_hdev - create habanalabs device instance
275 * @dev: will hold the pointer to the new habanalabs device structure
276 * @pdev: pointer to the pci device
277 * @asic_type: in case of simulator device, which device is it
278 * @minor: in case of simulator device, the minor of the device
280 * Allocate memory for habanalabs device and initialize basic fields
281 * Identify the ASIC type
282 * Allocate ID (minor) for the device (only for real devices)
284 int create_hdev(struct hl_device **dev, struct pci_dev *pdev,
285 enum hl_asic_type asic_type, int minor)
287 struct hl_device *hdev;
288 int rc, main_id, ctrl_id = 0;
292 hdev = kzalloc(sizeof(*hdev), GFP_KERNEL);
296 /* First, we must find out which ASIC are we handling. This is needed
297 * to configure the behavior of the driver (kernel parameters)
300 hdev->asic_type = get_asic_type(pdev->device);
301 if (hdev->asic_type == ASIC_INVALID) {
302 dev_err(&pdev->dev, "Unsupported ASIC\n");
307 hdev->asic_type = asic_type;
311 hdev->asic_prop.fw_security_disabled =
312 !is_asic_secured(pdev->device);
314 hdev->asic_prop.fw_security_disabled = true;
316 /* Assign status description string */
317 strncpy(hdev->status[HL_DEVICE_STATUS_MALFUNCTION],
318 "disabled", HL_STR_MAX);
319 strncpy(hdev->status[HL_DEVICE_STATUS_IN_RESET],
320 "in reset", HL_STR_MAX);
321 strncpy(hdev->status[HL_DEVICE_STATUS_NEEDS_RESET],
322 "needs reset", HL_STR_MAX);
324 hdev->major = hl_major;
325 hdev->reset_on_lockup = reset_on_lockup;
326 hdev->memory_scrub = memory_scrub;
327 hdev->boot_error_status_mask = boot_error_status_mask;
331 set_driver_behavior_per_device(hdev);
334 hdev->timeout_jiffies = msecs_to_jiffies(timeout_locked * 1000);
336 hdev->timeout_jiffies = MAX_SCHEDULE_TIMEOUT;
338 hdev->disabled = true;
339 hdev->pdev = pdev; /* can be NULL in case of simulator device */
341 /* Set default DMA mask to 32 bits */
344 mutex_lock(&hl_devs_idr_lock);
346 /* Always save 2 numbers, 1 for main device and 1 for control.
347 * They must be consecutive
349 main_id = idr_alloc(&hl_devs_idr, hdev, 0, HL_MAX_MINORS,
353 ctrl_id = idr_alloc(&hl_devs_idr, hdev, main_id + 1,
354 main_id + 2, GFP_KERNEL);
356 mutex_unlock(&hl_devs_idr_lock);
358 if ((main_id < 0) || (ctrl_id < 0)) {
359 if ((main_id == -ENOSPC) || (ctrl_id == -ENOSPC))
360 pr_err("too many devices in the system\n");
363 mutex_lock(&hl_devs_idr_lock);
364 idr_remove(&hl_devs_idr, main_id);
365 mutex_unlock(&hl_devs_idr_lock);
373 hdev->id_control = ctrl_id;
385 * destroy_hdev - destroy habanalabs device instance
387 * @dev: pointer to the habanalabs device structure
390 void destroy_hdev(struct hl_device *hdev)
392 /* Remove device from the device list */
393 mutex_lock(&hl_devs_idr_lock);
394 idr_remove(&hl_devs_idr, hdev->id);
395 idr_remove(&hl_devs_idr, hdev->id_control);
396 mutex_unlock(&hl_devs_idr_lock);
401 static int hl_pmops_suspend(struct device *dev)
403 struct hl_device *hdev = dev_get_drvdata(dev);
405 pr_debug("Going to suspend PCI device\n");
408 pr_err("device pointer is NULL in suspend\n");
412 return hl_device_suspend(hdev);
415 static int hl_pmops_resume(struct device *dev)
417 struct hl_device *hdev = dev_get_drvdata(dev);
419 pr_debug("Going to resume PCI device\n");
422 pr_err("device pointer is NULL in resume\n");
426 return hl_device_resume(hdev);
430 * hl_pci_probe - probe PCI habanalabs devices
432 * @pdev: pointer to pci device
433 * @id: pointer to pci device id structure
435 * Standard PCI probe function for habanalabs device.
436 * Create a new habanalabs device and initialize it according to the
439 static int hl_pci_probe(struct pci_dev *pdev,
440 const struct pci_device_id *id)
442 struct hl_device *hdev;
445 dev_info(&pdev->dev, HL_NAME
446 " device found [%04x:%04x] (rev %x)\n",
447 (int)pdev->vendor, (int)pdev->device, (int)pdev->revision);
449 rc = create_hdev(&hdev, pdev, ASIC_INVALID, -1);
453 pci_set_drvdata(pdev, hdev);
455 pci_enable_pcie_error_reporting(pdev);
457 rc = hl_device_init(hdev, hl_class);
459 dev_err(&pdev->dev, "Fatal error during habanalabs device init\n");
467 pci_set_drvdata(pdev, NULL);
474 * hl_pci_remove - remove PCI habanalabs devices
476 * @pdev: pointer to pci device
478 * Standard PCI remove function for habanalabs device
480 static void hl_pci_remove(struct pci_dev *pdev)
482 struct hl_device *hdev;
484 hdev = pci_get_drvdata(pdev);
488 hl_device_fini(hdev);
489 pci_disable_pcie_error_reporting(pdev);
490 pci_set_drvdata(pdev, NULL);
495 * hl_pci_err_detected - a PCI bus error detected on this device
497 * @pdev: pointer to pci device
498 * @state: PCI error type
500 * Called by the PCI subsystem whenever a non-correctable
501 * PCI bus error is detected
503 static pci_ers_result_t
504 hl_pci_err_detected(struct pci_dev *pdev, pci_channel_state_t state)
506 struct hl_device *hdev = pci_get_drvdata(pdev);
507 enum pci_ers_result result;
510 case pci_channel_io_normal:
511 return PCI_ERS_RESULT_CAN_RECOVER;
513 case pci_channel_io_frozen:
514 dev_warn(hdev->dev, "frozen state error detected\n");
515 result = PCI_ERS_RESULT_NEED_RESET;
518 case pci_channel_io_perm_failure:
519 dev_warn(hdev->dev, "failure state error detected\n");
520 result = PCI_ERS_RESULT_DISCONNECT;
524 result = PCI_ERS_RESULT_NONE;
527 hdev->asic_funcs->halt_engines(hdev, true);
533 * hl_pci_err_resume - resume after a PCI slot reset
535 * @pdev: pointer to pci device
538 static void hl_pci_err_resume(struct pci_dev *pdev)
540 struct hl_device *hdev = pci_get_drvdata(pdev);
542 dev_warn(hdev->dev, "Resuming device after PCI slot reset\n");
543 hl_device_resume(hdev);
547 * hl_pci_err_slot_reset - a PCI slot reset has just happened
549 * @pdev: pointer to pci device
551 * Determine if the driver can recover from the PCI slot reset
553 static pci_ers_result_t hl_pci_err_slot_reset(struct pci_dev *pdev)
555 return PCI_ERS_RESULT_RECOVERED;
558 static const struct dev_pm_ops hl_pm_ops = {
559 .suspend = hl_pmops_suspend,
560 .resume = hl_pmops_resume,
563 static const struct pci_error_handlers hl_pci_err_handler = {
564 .error_detected = hl_pci_err_detected,
565 .slot_reset = hl_pci_err_slot_reset,
566 .resume = hl_pci_err_resume,
569 static struct pci_driver hl_pci_driver = {
572 .probe = hl_pci_probe,
573 .remove = hl_pci_remove,
574 .shutdown = hl_pci_remove,
575 .driver.pm = &hl_pm_ops,
576 .err_handler = &hl_pci_err_handler,
580 * hl_init - Initialize the habanalabs kernel driver
582 static int __init hl_init(void)
587 pr_info("loading driver\n");
589 rc = alloc_chrdev_region(&dev, 0, HL_MAX_MINORS, HL_NAME);
591 pr_err("unable to get major\n");
595 hl_major = MAJOR(dev);
597 hl_class = class_create(THIS_MODULE, HL_NAME);
598 if (IS_ERR(hl_class)) {
599 pr_err("failed to allocate class\n");
600 rc = PTR_ERR(hl_class);
606 rc = pci_register_driver(&hl_pci_driver);
608 pr_err("failed to register pci device\n");
612 pr_debug("driver loaded\n");
618 class_destroy(hl_class);
620 unregister_chrdev_region(MKDEV(hl_major, 0), HL_MAX_MINORS);
625 * hl_exit - Release all resources of the habanalabs kernel driver
627 static void __exit hl_exit(void)
629 pci_unregister_driver(&hl_pci_driver);
632 * Removing debugfs must be after all devices or simulator devices
633 * have been removed because otherwise we get a bug in the
634 * debugfs module for referencing NULL objects
638 class_destroy(hl_class);
639 unregister_chrdev_region(MKDEV(hl_major, 0), HL_MAX_MINORS);
641 idr_destroy(&hl_devs_idr);
643 pr_debug("driver removed\n");
646 module_init(hl_init);
647 module_exit(hl_exit);