4 * Copyright (C) 2009 Intel Corporation, Yu Zhao <yu.zhao@intel.com>
6 * PCI Express I/O Virtualization (IOV) support.
8 * Address Translation Service 1.0
11 #include <linux/pci.h>
12 #include <linux/slab.h>
13 #include <linux/mutex.h>
14 #include <linux/export.h>
15 #include <linux/string.h>
16 #include <linux/delay.h>
17 #include <linux/pci-ats.h>
20 #define VIRTFN_ID_LEN 16
22 static inline u8 virtfn_bus(struct pci_dev *dev, int id)
24 return dev->bus->number + ((dev->devfn + dev->sriov->offset +
25 dev->sriov->stride * id) >> 8);
28 static inline u8 virtfn_devfn(struct pci_dev *dev, int id)
30 return (dev->devfn + dev->sriov->offset +
31 dev->sriov->stride * id) & 0xff;
34 static struct pci_bus *virtfn_add_bus(struct pci_bus *bus, int busnr)
36 struct pci_bus *child;
38 if (bus->number == busnr)
41 child = pci_find_bus(pci_domain_nr(bus), busnr);
45 child = pci_add_new_bus(bus, NULL, busnr);
49 pci_bus_insert_busn_res(child, busnr, busnr);
54 static void virtfn_remove_bus(struct pci_bus *physbus, struct pci_bus *virtbus)
56 if (physbus != virtbus && list_empty(&virtbus->devices))
57 pci_remove_bus(virtbus);
60 static int virtfn_add(struct pci_dev *dev, int id, int reset)
65 char buf[VIRTFN_ID_LEN];
66 struct pci_dev *virtfn;
68 struct pci_sriov *iov = dev->sriov;
71 mutex_lock(&iov->dev->sriov->lock);
72 bus = virtfn_add_bus(dev->bus, virtfn_bus(dev, id));
76 virtfn = pci_alloc_dev(bus);
80 virtfn->devfn = virtfn_devfn(dev, id);
81 virtfn->vendor = dev->vendor;
82 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_DID, &virtfn->device);
83 pci_setup_device(virtfn);
84 virtfn->dev.parent = dev->dev.parent;
85 virtfn->physfn = pci_dev_get(dev);
86 virtfn->is_virtfn = 1;
87 virtfn->multifunction = 0;
89 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
90 res = dev->resource + PCI_IOV_RESOURCES + i;
93 virtfn->resource[i].name = pci_name(virtfn);
94 virtfn->resource[i].flags = res->flags;
95 size = resource_size(res);
96 do_div(size, iov->total_VFs);
97 virtfn->resource[i].start = res->start + size * id;
98 virtfn->resource[i].end = virtfn->resource[i].start + size - 1;
99 rc = request_resource(res, &virtfn->resource[i]);
104 __pci_reset_function(virtfn);
106 pci_device_add(virtfn, virtfn->bus);
107 mutex_unlock(&iov->dev->sriov->lock);
109 rc = pci_bus_add_device(virtfn);
110 sprintf(buf, "virtfn%u", id);
111 rc = sysfs_create_link(&dev->dev.kobj, &virtfn->dev.kobj, buf);
114 rc = sysfs_create_link(&virtfn->dev.kobj, &dev->dev.kobj, "physfn");
118 kobject_uevent(&virtfn->dev.kobj, KOBJ_CHANGE);
123 sysfs_remove_link(&dev->dev.kobj, buf);
126 mutex_lock(&iov->dev->sriov->lock);
127 pci_stop_and_remove_bus_device(virtfn);
129 virtfn_remove_bus(dev->bus, bus);
131 mutex_unlock(&iov->dev->sriov->lock);
136 static void virtfn_remove(struct pci_dev *dev, int id, int reset)
138 char buf[VIRTFN_ID_LEN];
139 struct pci_dev *virtfn;
140 struct pci_sriov *iov = dev->sriov;
142 virtfn = pci_get_domain_bus_and_slot(pci_domain_nr(dev->bus),
144 virtfn_devfn(dev, id));
149 device_release_driver(&virtfn->dev);
150 __pci_reset_function(virtfn);
153 sprintf(buf, "virtfn%u", id);
154 sysfs_remove_link(&dev->dev.kobj, buf);
156 * pci_stop_dev() could have been called for this virtfn already,
157 * so the directory for the virtfn may have been removed before.
158 * Double check to avoid spurious sysfs warnings.
160 if (virtfn->dev.kobj.sd)
161 sysfs_remove_link(&virtfn->dev.kobj, "physfn");
163 mutex_lock(&iov->dev->sriov->lock);
164 pci_stop_and_remove_bus_device(virtfn);
165 virtfn_remove_bus(dev->bus, virtfn->bus);
166 mutex_unlock(&iov->dev->sriov->lock);
168 /* balance pci_get_domain_bus_and_slot() */
173 static int sriov_migration(struct pci_dev *dev)
176 struct pci_sriov *iov = dev->sriov;
181 if (!(iov->cap & PCI_SRIOV_CAP_VFM))
184 pci_read_config_word(dev, iov->pos + PCI_SRIOV_STATUS, &status);
185 if (!(status & PCI_SRIOV_STATUS_VFM))
188 schedule_work(&iov->mtask);
193 static void sriov_migration_task(struct work_struct *work)
198 struct pci_sriov *iov = container_of(work, struct pci_sriov, mtask);
200 for (i = iov->initial_VFs; i < iov->num_VFs; i++) {
201 state = readb(iov->mstate + i);
202 if (state == PCI_SRIOV_VFM_MI) {
203 writeb(PCI_SRIOV_VFM_AV, iov->mstate + i);
204 state = readb(iov->mstate + i);
205 if (state == PCI_SRIOV_VFM_AV)
206 virtfn_add(iov->self, i, 1);
207 } else if (state == PCI_SRIOV_VFM_MO) {
208 virtfn_remove(iov->self, i, 1);
209 writeb(PCI_SRIOV_VFM_UA, iov->mstate + i);
210 state = readb(iov->mstate + i);
211 if (state == PCI_SRIOV_VFM_AV)
212 virtfn_add(iov->self, i, 0);
216 pci_read_config_word(iov->self, iov->pos + PCI_SRIOV_STATUS, &status);
217 status &= ~PCI_SRIOV_STATUS_VFM;
218 pci_write_config_word(iov->self, iov->pos + PCI_SRIOV_STATUS, status);
221 static int sriov_enable_migration(struct pci_dev *dev, int nr_virtfn)
226 struct pci_sriov *iov = dev->sriov;
228 if (nr_virtfn <= iov->initial_VFs)
231 pci_read_config_dword(dev, iov->pos + PCI_SRIOV_VFM, &table);
232 bir = PCI_SRIOV_VFM_BIR(table);
233 if (bir > PCI_STD_RESOURCE_END)
236 table = PCI_SRIOV_VFM_OFFSET(table);
237 if (table + nr_virtfn > pci_resource_len(dev, bir))
240 pa = pci_resource_start(dev, bir) + table;
241 iov->mstate = ioremap(pa, nr_virtfn);
245 INIT_WORK(&iov->mtask, sriov_migration_task);
247 iov->ctrl |= PCI_SRIOV_CTRL_VFM | PCI_SRIOV_CTRL_INTR;
248 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
253 static void sriov_disable_migration(struct pci_dev *dev)
255 struct pci_sriov *iov = dev->sriov;
257 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFM | PCI_SRIOV_CTRL_INTR);
258 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
260 cancel_work_sync(&iov->mtask);
261 iounmap(iov->mstate);
264 static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
269 u16 offset, stride, initial;
270 struct resource *res;
271 struct pci_dev *pdev;
272 struct pci_sriov *iov = dev->sriov;
281 pci_read_config_word(dev, iov->pos + PCI_SRIOV_INITIAL_VF, &initial);
282 if (initial > iov->total_VFs ||
283 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (initial != iov->total_VFs)))
286 if (nr_virtfn < 0 || nr_virtfn > iov->total_VFs ||
287 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (nr_virtfn > initial)))
290 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_OFFSET, &offset);
291 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_STRIDE, &stride);
292 if (!offset || (nr_virtfn > 1 && !stride))
296 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
297 bars |= (1 << (i + PCI_IOV_RESOURCES));
298 res = dev->resource + PCI_IOV_RESOURCES + i;
302 if (nres != iov->nres) {
303 dev_err(&dev->dev, "not enough MMIO resources for SR-IOV\n");
307 iov->offset = offset;
308 iov->stride = stride;
310 if (virtfn_bus(dev, nr_virtfn - 1) > dev->bus->busn_res.end) {
311 dev_err(&dev->dev, "SR-IOV: bus number out of range\n");
315 if (pci_enable_resources(dev, bars)) {
316 dev_err(&dev->dev, "SR-IOV: IOV BARS not allocated\n");
320 if (iov->link != dev->devfn) {
321 pdev = pci_get_slot(dev->bus, iov->link);
325 if (!pdev->is_physfn) {
330 rc = sysfs_create_link(&dev->dev.kobj,
331 &pdev->dev.kobj, "dep_link");
337 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, nr_virtfn);
338 iov->ctrl |= PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE;
339 pci_cfg_access_lock(dev);
340 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
342 pci_cfg_access_unlock(dev);
344 iov->initial_VFs = initial;
345 if (nr_virtfn < initial)
348 for (i = 0; i < initial; i++) {
349 rc = virtfn_add(dev, i, 0);
354 if (iov->cap & PCI_SRIOV_CAP_VFM) {
355 rc = sriov_enable_migration(dev, nr_virtfn);
360 kobject_uevent(&dev->dev.kobj, KOBJ_CHANGE);
361 iov->num_VFs = nr_virtfn;
366 for (j = 0; j < i; j++)
367 virtfn_remove(dev, j, 0);
369 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
370 pci_cfg_access_lock(dev);
371 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
372 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, 0);
374 pci_cfg_access_unlock(dev);
376 if (iov->link != dev->devfn)
377 sysfs_remove_link(&dev->dev.kobj, "dep_link");
382 static void sriov_disable(struct pci_dev *dev)
385 struct pci_sriov *iov = dev->sriov;
390 if (iov->cap & PCI_SRIOV_CAP_VFM)
391 sriov_disable_migration(dev);
393 for (i = 0; i < iov->num_VFs; i++)
394 virtfn_remove(dev, i, 0);
396 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
397 pci_cfg_access_lock(dev);
398 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
400 pci_cfg_access_unlock(dev);
402 if (iov->link != dev->devfn)
403 sysfs_remove_link(&dev->dev.kobj, "dep_link");
406 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, 0);
409 static int sriov_init(struct pci_dev *dev, int pos)
415 u16 ctrl, total, offset, stride;
416 struct pci_sriov *iov;
417 struct resource *res;
418 struct pci_dev *pdev;
420 if (pci_pcie_type(dev) != PCI_EXP_TYPE_RC_END &&
421 pci_pcie_type(dev) != PCI_EXP_TYPE_ENDPOINT)
424 pci_read_config_word(dev, pos + PCI_SRIOV_CTRL, &ctrl);
425 if (ctrl & PCI_SRIOV_CTRL_VFE) {
426 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, 0);
430 pci_read_config_word(dev, pos + PCI_SRIOV_TOTAL_VF, &total);
435 list_for_each_entry(pdev, &dev->bus->devices, bus_list)
440 if (pci_ari_enabled(dev->bus))
441 ctrl |= PCI_SRIOV_CTRL_ARI;
444 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, ctrl);
445 pci_write_config_word(dev, pos + PCI_SRIOV_NUM_VF, 0);
446 pci_read_config_word(dev, pos + PCI_SRIOV_VF_OFFSET, &offset);
447 pci_read_config_word(dev, pos + PCI_SRIOV_VF_STRIDE, &stride);
448 if (!offset || (total > 1 && !stride))
451 pci_read_config_dword(dev, pos + PCI_SRIOV_SUP_PGSIZE, &pgsz);
452 i = PAGE_SHIFT > 12 ? PAGE_SHIFT - 12 : 0;
453 pgsz &= ~((1 << i) - 1);
458 pci_write_config_dword(dev, pos + PCI_SRIOV_SYS_PGSIZE, pgsz);
461 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
462 res = dev->resource + PCI_IOV_RESOURCES + i;
463 i += __pci_read_base(dev, pci_bar_unknown, res,
464 pos + PCI_SRIOV_BAR + i * 4);
467 if (resource_size(res) & (PAGE_SIZE - 1)) {
471 res->end = res->start + resource_size(res) * total - 1;
475 iov = kzalloc(sizeof(*iov), GFP_KERNEL);
484 iov->total_VFs = total;
485 iov->offset = offset;
486 iov->stride = stride;
489 pci_read_config_dword(dev, pos + PCI_SRIOV_CAP, &iov->cap);
490 pci_read_config_byte(dev, pos + PCI_SRIOV_FUNC_LINK, &iov->link);
491 if (pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END)
492 iov->link = PCI_DEVFN(PCI_SLOT(dev->devfn), iov->link);
495 iov->dev = pci_dev_get(pdev);
499 mutex_init(&iov->lock);
507 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
508 res = dev->resource + PCI_IOV_RESOURCES + i;
515 static void sriov_release(struct pci_dev *dev)
517 BUG_ON(dev->sriov->num_VFs);
519 if (dev != dev->sriov->dev)
520 pci_dev_put(dev->sriov->dev);
522 mutex_destroy(&dev->sriov->lock);
528 static void sriov_restore_state(struct pci_dev *dev)
532 struct pci_sriov *iov = dev->sriov;
534 pci_read_config_word(dev, iov->pos + PCI_SRIOV_CTRL, &ctrl);
535 if (ctrl & PCI_SRIOV_CTRL_VFE)
538 for (i = PCI_IOV_RESOURCES; i <= PCI_IOV_RESOURCE_END; i++)
539 pci_update_resource(dev, i);
541 pci_write_config_dword(dev, iov->pos + PCI_SRIOV_SYS_PGSIZE, iov->pgsz);
542 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, iov->num_VFs);
543 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
544 if (iov->ctrl & PCI_SRIOV_CTRL_VFE)
549 * pci_iov_init - initialize the IOV capability
550 * @dev: the PCI device
552 * Returns 0 on success, or negative on failure.
554 int pci_iov_init(struct pci_dev *dev)
558 if (!pci_is_pcie(dev))
561 pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_SRIOV);
563 return sriov_init(dev, pos);
569 * pci_iov_release - release resources used by the IOV capability
570 * @dev: the PCI device
572 void pci_iov_release(struct pci_dev *dev)
579 * pci_iov_resource_bar - get position of the SR-IOV BAR
580 * @dev: the PCI device
581 * @resno: the resource number
582 * @type: the BAR type to be filled in
584 * Returns position of the BAR encapsulated in the SR-IOV capability.
586 int pci_iov_resource_bar(struct pci_dev *dev, int resno,
587 enum pci_bar_type *type)
589 if (resno < PCI_IOV_RESOURCES || resno > PCI_IOV_RESOURCE_END)
592 BUG_ON(!dev->is_physfn);
594 *type = pci_bar_unknown;
596 return dev->sriov->pos + PCI_SRIOV_BAR +
597 4 * (resno - PCI_IOV_RESOURCES);
601 * pci_sriov_resource_alignment - get resource alignment for VF BAR
602 * @dev: the PCI device
603 * @resno: the resource number
605 * Returns the alignment of the VF BAR found in the SR-IOV capability.
606 * This is not the same as the resource size which is defined as
607 * the VF BAR size multiplied by the number of VFs. The alignment
608 * is just the VF BAR size.
610 resource_size_t pci_sriov_resource_alignment(struct pci_dev *dev, int resno)
613 enum pci_bar_type type;
614 int reg = pci_iov_resource_bar(dev, resno, &type);
619 __pci_read_base(dev, type, &tmp, reg);
620 return resource_alignment(&tmp);
624 * pci_restore_iov_state - restore the state of the IOV capability
625 * @dev: the PCI device
627 void pci_restore_iov_state(struct pci_dev *dev)
630 sriov_restore_state(dev);
634 * pci_iov_bus_range - find bus range used by Virtual Function
637 * Returns max number of buses (exclude current one) used by Virtual
640 int pci_iov_bus_range(struct pci_bus *bus)
646 list_for_each_entry(dev, &bus->devices, bus_list) {
649 busnr = virtfn_bus(dev, dev->sriov->total_VFs - 1);
654 return max ? max - bus->number : 0;
658 * pci_enable_sriov - enable the SR-IOV capability
659 * @dev: the PCI device
660 * @nr_virtfn: number of virtual functions to enable
662 * Returns 0 on success, or negative on failure.
664 int pci_enable_sriov(struct pci_dev *dev, int nr_virtfn)
671 return sriov_enable(dev, nr_virtfn);
673 EXPORT_SYMBOL_GPL(pci_enable_sriov);
676 * pci_disable_sriov - disable the SR-IOV capability
677 * @dev: the PCI device
679 void pci_disable_sriov(struct pci_dev *dev)
688 EXPORT_SYMBOL_GPL(pci_disable_sriov);
691 * pci_sriov_migration - notify SR-IOV core of Virtual Function Migration
692 * @dev: the PCI device
694 * Returns IRQ_HANDLED if the IRQ is handled, or IRQ_NONE if not.
696 * Physical Function driver is responsible to register IRQ handler using
697 * VF Migration Interrupt Message Number, and call this function when the
698 * interrupt is generated by the hardware.
700 irqreturn_t pci_sriov_migration(struct pci_dev *dev)
705 return sriov_migration(dev) ? IRQ_HANDLED : IRQ_NONE;
707 EXPORT_SYMBOL_GPL(pci_sriov_migration);
710 * pci_num_vf - return number of VFs associated with a PF device_release_driver
711 * @dev: the PCI device
713 * Returns number of VFs, or 0 if SR-IOV is not enabled.
715 int pci_num_vf(struct pci_dev *dev)
720 return dev->sriov->num_VFs;
722 EXPORT_SYMBOL_GPL(pci_num_vf);
725 * pci_vfs_assigned - returns number of VFs are assigned to a guest
726 * @dev: the PCI device
728 * Returns number of VFs belonging to this device that are assigned to a guest.
729 * If device is not a physical function returns 0.
731 int pci_vfs_assigned(struct pci_dev *dev)
733 struct pci_dev *vfdev;
734 unsigned int vfs_assigned = 0;
735 unsigned short dev_id;
737 /* only search if we are a PF */
742 * determine the device ID for the VFs, the vendor ID will be the
743 * same as the PF so there is no need to check for that one
745 pci_read_config_word(dev, dev->sriov->pos + PCI_SRIOV_VF_DID, &dev_id);
747 /* loop through all the VFs to see if we own any that are assigned */
748 vfdev = pci_get_device(dev->vendor, dev_id, NULL);
751 * It is considered assigned if it is a virtual function with
752 * our dev as the physical function and the assigned bit is set
754 if (vfdev->is_virtfn && (vfdev->physfn == dev) &&
755 (vfdev->dev_flags & PCI_DEV_FLAGS_ASSIGNED))
758 vfdev = pci_get_device(dev->vendor, dev_id, vfdev);
763 EXPORT_SYMBOL_GPL(pci_vfs_assigned);
766 * pci_sriov_set_totalvfs -- reduce the TotalVFs available
767 * @dev: the PCI PF device
768 * @numvfs: number that should be used for TotalVFs supported
770 * Should be called from PF driver's probe routine with
771 * device's mutex held.
773 * Returns 0 if PF is an SRIOV-capable device and
774 * value of numvfs valid. If not a PF return -ENOSYS;
775 * if numvfs is invalid return -EINVAL;
776 * if VFs already enabled, return -EBUSY.
778 int pci_sriov_set_totalvfs(struct pci_dev *dev, u16 numvfs)
782 if (numvfs > dev->sriov->total_VFs)
785 /* Shouldn't change if VFs already enabled */
786 if (dev->sriov->ctrl & PCI_SRIOV_CTRL_VFE)
789 dev->sriov->driver_max_VFs = numvfs;
793 EXPORT_SYMBOL_GPL(pci_sriov_set_totalvfs);
796 * pci_sriov_get_totalvfs -- get total VFs supported on this device
797 * @dev: the PCI PF device
799 * For a PCIe device with SRIOV support, return the PCIe
800 * SRIOV capability value of TotalVFs or the value of driver_max_VFs
801 * if the driver reduced it. Otherwise 0.
803 int pci_sriov_get_totalvfs(struct pci_dev *dev)
808 if (dev->sriov->driver_max_VFs)
809 return dev->sriov->driver_max_VFs;
811 return dev->sriov->total_VFs;
813 EXPORT_SYMBOL_GPL(pci_sriov_get_totalvfs);