1 // SPDX-License-Identifier: GPL-2.0
3 * PCI Express I/O Virtualization (IOV) support
5 * Address Translation Service 1.0
7 * Copyright (C) 2009 Intel Corporation, Yu Zhao <yu.zhao@intel.com>
10 #include <linux/pci.h>
11 #include <linux/slab.h>
12 #include <linux/mutex.h>
13 #include <linux/export.h>
14 #include <linux/string.h>
15 #include <linux/delay.h>
16 #include <linux/pci-ats.h>
19 #define VIRTFN_ID_LEN 16
21 int pci_iov_virtfn_bus(struct pci_dev *dev, int vf_id)
25 return dev->bus->number + ((dev->devfn + dev->sriov->offset +
26 dev->sriov->stride * vf_id) >> 8);
29 int pci_iov_virtfn_devfn(struct pci_dev *dev, int vf_id)
33 return (dev->devfn + dev->sriov->offset +
34 dev->sriov->stride * vf_id) & 0xff;
38 * Per SR-IOV spec sec 3.3.10 and 3.3.11, First VF Offset and VF Stride may
39 * change when NumVFs changes.
41 * Update iov->offset and iov->stride when NumVFs is written.
43 static inline void pci_iov_set_numvfs(struct pci_dev *dev, int nr_virtfn)
45 struct pci_sriov *iov = dev->sriov;
47 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, nr_virtfn);
48 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_OFFSET, &iov->offset);
49 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_STRIDE, &iov->stride);
53 * The PF consumes one bus number. NumVFs, First VF Offset, and VF Stride
54 * determine how many additional bus numbers will be consumed by VFs.
56 * Iterate over all valid NumVFs, validate offset and stride, and calculate
57 * the maximum number of bus numbers that could ever be required.
59 static int compute_max_vf_buses(struct pci_dev *dev)
61 struct pci_sriov *iov = dev->sriov;
62 int nr_virtfn, busnr, rc = 0;
64 for (nr_virtfn = iov->total_VFs; nr_virtfn; nr_virtfn--) {
65 pci_iov_set_numvfs(dev, nr_virtfn);
66 if (!iov->offset || (nr_virtfn > 1 && !iov->stride)) {
71 busnr = pci_iov_virtfn_bus(dev, nr_virtfn - 1);
72 if (busnr > iov->max_VF_buses)
73 iov->max_VF_buses = busnr;
77 pci_iov_set_numvfs(dev, 0);
81 static struct pci_bus *virtfn_add_bus(struct pci_bus *bus, int busnr)
83 struct pci_bus *child;
85 if (bus->number == busnr)
88 child = pci_find_bus(pci_domain_nr(bus), busnr);
92 child = pci_add_new_bus(bus, NULL, busnr);
96 pci_bus_insert_busn_res(child, busnr, busnr);
101 static void virtfn_remove_bus(struct pci_bus *physbus, struct pci_bus *virtbus)
103 if (physbus != virtbus && list_empty(&virtbus->devices))
104 pci_remove_bus(virtbus);
107 resource_size_t pci_iov_resource_size(struct pci_dev *dev, int resno)
112 return dev->sriov->barsz[resno - PCI_IOV_RESOURCES];
115 int pci_iov_add_virtfn(struct pci_dev *dev, int id)
120 char buf[VIRTFN_ID_LEN];
121 struct pci_dev *virtfn;
122 struct resource *res;
123 struct pci_sriov *iov = dev->sriov;
126 bus = virtfn_add_bus(dev->bus, pci_iov_virtfn_bus(dev, id));
130 virtfn = pci_alloc_dev(bus);
134 virtfn->devfn = pci_iov_virtfn_devfn(dev, id);
135 virtfn->vendor = dev->vendor;
136 virtfn->device = iov->vf_device;
137 rc = pci_setup_device(virtfn);
141 virtfn->dev.parent = dev->dev.parent;
142 virtfn->physfn = pci_dev_get(dev);
143 virtfn->is_virtfn = 1;
144 virtfn->multifunction = 0;
146 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
147 res = &dev->resource[i + PCI_IOV_RESOURCES];
150 virtfn->resource[i].name = pci_name(virtfn);
151 virtfn->resource[i].flags = res->flags;
152 size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
153 virtfn->resource[i].start = res->start + size * id;
154 virtfn->resource[i].end = virtfn->resource[i].start + size - 1;
155 rc = request_resource(res, &virtfn->resource[i]);
159 pci_device_add(virtfn, virtfn->bus);
161 sprintf(buf, "virtfn%u", id);
162 rc = sysfs_create_link(&dev->dev.kobj, &virtfn->dev.kobj, buf);
165 rc = sysfs_create_link(&virtfn->dev.kobj, &dev->dev.kobj, "physfn");
169 kobject_uevent(&virtfn->dev.kobj, KOBJ_CHANGE);
171 pci_bus_add_device(virtfn);
176 sysfs_remove_link(&dev->dev.kobj, buf);
179 pci_stop_and_remove_bus_device(virtfn);
181 virtfn_remove_bus(dev->bus, bus);
187 void pci_iov_remove_virtfn(struct pci_dev *dev, int id)
189 char buf[VIRTFN_ID_LEN];
190 struct pci_dev *virtfn;
192 virtfn = pci_get_domain_bus_and_slot(pci_domain_nr(dev->bus),
193 pci_iov_virtfn_bus(dev, id),
194 pci_iov_virtfn_devfn(dev, id));
198 sprintf(buf, "virtfn%u", id);
199 sysfs_remove_link(&dev->dev.kobj, buf);
201 * pci_stop_dev() could have been called for this virtfn already,
202 * so the directory for the virtfn may have been removed before.
203 * Double check to avoid spurious sysfs warnings.
205 if (virtfn->dev.kobj.sd)
206 sysfs_remove_link(&virtfn->dev.kobj, "physfn");
208 pci_stop_and_remove_bus_device(virtfn);
209 virtfn_remove_bus(dev->bus, virtfn->bus);
211 /* balance pci_get_domain_bus_and_slot() */
216 int __weak pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
221 int __weak pcibios_sriov_disable(struct pci_dev *pdev)
226 static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
232 struct resource *res;
233 struct pci_dev *pdev;
234 struct pci_sriov *iov = dev->sriov;
244 pci_read_config_word(dev, iov->pos + PCI_SRIOV_INITIAL_VF, &initial);
245 if (initial > iov->total_VFs ||
246 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (initial != iov->total_VFs)))
249 if (nr_virtfn < 0 || nr_virtfn > iov->total_VFs ||
250 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (nr_virtfn > initial)))
254 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
255 bars |= (1 << (i + PCI_IOV_RESOURCES));
256 res = &dev->resource[i + PCI_IOV_RESOURCES];
260 if (nres != iov->nres) {
261 pci_err(dev, "not enough MMIO resources for SR-IOV\n");
265 bus = pci_iov_virtfn_bus(dev, nr_virtfn - 1);
266 if (bus > dev->bus->busn_res.end) {
267 pci_err(dev, "can't enable %d VFs (bus %02x out of range of %pR)\n",
268 nr_virtfn, bus, &dev->bus->busn_res);
272 if (pci_enable_resources(dev, bars)) {
273 pci_err(dev, "SR-IOV: IOV BARS not allocated\n");
277 if (iov->link != dev->devfn) {
278 pdev = pci_get_slot(dev->bus, iov->link);
282 if (!pdev->is_physfn) {
287 rc = sysfs_create_link(&dev->dev.kobj,
288 &pdev->dev.kobj, "dep_link");
294 iov->initial_VFs = initial;
295 if (nr_virtfn < initial)
298 rc = pcibios_sriov_enable(dev, initial);
300 pci_err(dev, "failure %d from pcibios_sriov_enable()\n", rc);
304 pci_iov_set_numvfs(dev, nr_virtfn);
305 iov->ctrl |= PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE;
306 pci_cfg_access_lock(dev);
307 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
309 pci_cfg_access_unlock(dev);
311 for (i = 0; i < initial; i++) {
312 rc = pci_iov_add_virtfn(dev, i);
317 kobject_uevent(&dev->dev.kobj, KOBJ_CHANGE);
318 iov->num_VFs = nr_virtfn;
324 pci_iov_remove_virtfn(dev, i);
327 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
328 pci_cfg_access_lock(dev);
329 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
331 pci_cfg_access_unlock(dev);
333 pcibios_sriov_disable(dev);
335 if (iov->link != dev->devfn)
336 sysfs_remove_link(&dev->dev.kobj, "dep_link");
338 pci_iov_set_numvfs(dev, 0);
342 static void sriov_disable(struct pci_dev *dev)
345 struct pci_sriov *iov = dev->sriov;
350 for (i = 0; i < iov->num_VFs; i++)
351 pci_iov_remove_virtfn(dev, i);
353 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
354 pci_cfg_access_lock(dev);
355 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
357 pci_cfg_access_unlock(dev);
359 pcibios_sriov_disable(dev);
361 if (iov->link != dev->devfn)
362 sysfs_remove_link(&dev->dev.kobj, "dep_link");
365 pci_iov_set_numvfs(dev, 0);
368 static int sriov_init(struct pci_dev *dev, int pos)
375 struct pci_sriov *iov;
376 struct resource *res;
377 struct pci_dev *pdev;
379 pci_read_config_word(dev, pos + PCI_SRIOV_CTRL, &ctrl);
380 if (ctrl & PCI_SRIOV_CTRL_VFE) {
381 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, 0);
386 list_for_each_entry(pdev, &dev->bus->devices, bus_list)
391 if (pci_ari_enabled(dev->bus))
392 ctrl |= PCI_SRIOV_CTRL_ARI;
395 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, ctrl);
397 pci_read_config_word(dev, pos + PCI_SRIOV_TOTAL_VF, &total);
401 pci_read_config_dword(dev, pos + PCI_SRIOV_SUP_PGSIZE, &pgsz);
402 i = PAGE_SHIFT > 12 ? PAGE_SHIFT - 12 : 0;
403 pgsz &= ~((1 << i) - 1);
408 pci_write_config_dword(dev, pos + PCI_SRIOV_SYS_PGSIZE, pgsz);
410 iov = kzalloc(sizeof(*iov), GFP_KERNEL);
415 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
416 res = &dev->resource[i + PCI_IOV_RESOURCES];
418 * If it is already FIXED, don't change it, something
419 * (perhaps EA or header fixups) wants it this way.
421 if (res->flags & IORESOURCE_PCI_FIXED)
422 bar64 = (res->flags & IORESOURCE_MEM_64) ? 1 : 0;
424 bar64 = __pci_read_base(dev, pci_bar_unknown, res,
425 pos + PCI_SRIOV_BAR + i * 4);
428 if (resource_size(res) & (PAGE_SIZE - 1)) {
432 iov->barsz[i] = resource_size(res);
433 res->end = res->start + resource_size(res) * total - 1;
434 pci_info(dev, "VF(n) BAR%d space: %pR (contains BAR%d for %d VFs)\n",
443 iov->total_VFs = total;
444 pci_read_config_word(dev, pos + PCI_SRIOV_VF_DID, &iov->vf_device);
447 iov->drivers_autoprobe = true;
448 pci_read_config_dword(dev, pos + PCI_SRIOV_CAP, &iov->cap);
449 pci_read_config_byte(dev, pos + PCI_SRIOV_FUNC_LINK, &iov->link);
450 if (pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END)
451 iov->link = PCI_DEVFN(PCI_SLOT(dev->devfn), iov->link);
454 iov->dev = pci_dev_get(pdev);
460 rc = compute_max_vf_buses(dev);
470 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
471 res = &dev->resource[i + PCI_IOV_RESOURCES];
479 static void sriov_release(struct pci_dev *dev)
481 BUG_ON(dev->sriov->num_VFs);
483 if (dev != dev->sriov->dev)
484 pci_dev_put(dev->sriov->dev);
490 static void sriov_restore_state(struct pci_dev *dev)
494 struct pci_sriov *iov = dev->sriov;
496 pci_read_config_word(dev, iov->pos + PCI_SRIOV_CTRL, &ctrl);
497 if (ctrl & PCI_SRIOV_CTRL_VFE)
501 * Restore PCI_SRIOV_CTRL_ARI before pci_iov_set_numvfs() because
502 * it reads offset & stride, which depend on PCI_SRIOV_CTRL_ARI.
504 ctrl &= ~PCI_SRIOV_CTRL_ARI;
505 ctrl |= iov->ctrl & PCI_SRIOV_CTRL_ARI;
506 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, ctrl);
508 for (i = PCI_IOV_RESOURCES; i <= PCI_IOV_RESOURCE_END; i++)
509 pci_update_resource(dev, i);
511 pci_write_config_dword(dev, iov->pos + PCI_SRIOV_SYS_PGSIZE, iov->pgsz);
512 pci_iov_set_numvfs(dev, iov->num_VFs);
513 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
514 if (iov->ctrl & PCI_SRIOV_CTRL_VFE)
519 * pci_iov_init - initialize the IOV capability
520 * @dev: the PCI device
522 * Returns 0 on success, or negative on failure.
524 int pci_iov_init(struct pci_dev *dev)
528 if (!pci_is_pcie(dev))
531 pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_SRIOV);
533 return sriov_init(dev, pos);
539 * pci_iov_release - release resources used by the IOV capability
540 * @dev: the PCI device
542 void pci_iov_release(struct pci_dev *dev)
549 * pci_iov_update_resource - update a VF BAR
550 * @dev: the PCI device
551 * @resno: the resource number
553 * Update a VF BAR in the SR-IOV capability of a PF.
555 void pci_iov_update_resource(struct pci_dev *dev, int resno)
557 struct pci_sriov *iov = dev->is_physfn ? dev->sriov : NULL;
558 struct resource *res = dev->resource + resno;
559 int vf_bar = resno - PCI_IOV_RESOURCES;
560 struct pci_bus_region region;
566 * The generic pci_restore_bars() path calls this for all devices,
567 * including VFs and non-SR-IOV devices. If this is not a PF, we
568 * have nothing to do.
573 pci_read_config_word(dev, iov->pos + PCI_SRIOV_CTRL, &cmd);
574 if ((cmd & PCI_SRIOV_CTRL_VFE) && (cmd & PCI_SRIOV_CTRL_MSE)) {
575 dev_WARN(&dev->dev, "can't update enabled VF BAR%d %pR\n",
581 * Ignore unimplemented BARs, unused resource slots for 64-bit
582 * BARs, and non-movable resources, e.g., those described via
583 * Enhanced Allocation.
588 if (res->flags & IORESOURCE_UNSET)
591 if (res->flags & IORESOURCE_PCI_FIXED)
594 pcibios_resource_to_bus(dev->bus, ®ion, res);
596 new |= res->flags & ~PCI_BASE_ADDRESS_MEM_MASK;
598 reg = iov->pos + PCI_SRIOV_BAR + 4 * vf_bar;
599 pci_write_config_dword(dev, reg, new);
600 if (res->flags & IORESOURCE_MEM_64) {
601 new = region.start >> 16 >> 16;
602 pci_write_config_dword(dev, reg + 4, new);
606 resource_size_t __weak pcibios_iov_resource_alignment(struct pci_dev *dev,
609 return pci_iov_resource_size(dev, resno);
613 * pci_sriov_resource_alignment - get resource alignment for VF BAR
614 * @dev: the PCI device
615 * @resno: the resource number
617 * Returns the alignment of the VF BAR found in the SR-IOV capability.
618 * This is not the same as the resource size which is defined as
619 * the VF BAR size multiplied by the number of VFs. The alignment
620 * is just the VF BAR size.
622 resource_size_t pci_sriov_resource_alignment(struct pci_dev *dev, int resno)
624 return pcibios_iov_resource_alignment(dev, resno);
628 * pci_restore_iov_state - restore the state of the IOV capability
629 * @dev: the PCI device
631 void pci_restore_iov_state(struct pci_dev *dev)
634 sriov_restore_state(dev);
638 * pci_vf_drivers_autoprobe - set PF property drivers_autoprobe for VFs
639 * @dev: the PCI device
640 * @auto_probe: set VF drivers auto probe flag
642 void pci_vf_drivers_autoprobe(struct pci_dev *dev, bool auto_probe)
645 dev->sriov->drivers_autoprobe = auto_probe;
649 * pci_iov_bus_range - find bus range used by Virtual Function
652 * Returns max number of buses (exclude current one) used by Virtual
655 int pci_iov_bus_range(struct pci_bus *bus)
660 list_for_each_entry(dev, &bus->devices, bus_list) {
663 if (dev->sriov->max_VF_buses > max)
664 max = dev->sriov->max_VF_buses;
667 return max ? max - bus->number : 0;
671 * pci_enable_sriov - enable the SR-IOV capability
672 * @dev: the PCI device
673 * @nr_virtfn: number of virtual functions to enable
675 * Returns 0 on success, or negative on failure.
677 int pci_enable_sriov(struct pci_dev *dev, int nr_virtfn)
684 return sriov_enable(dev, nr_virtfn);
686 EXPORT_SYMBOL_GPL(pci_enable_sriov);
689 * pci_disable_sriov - disable the SR-IOV capability
690 * @dev: the PCI device
692 void pci_disable_sriov(struct pci_dev *dev)
701 EXPORT_SYMBOL_GPL(pci_disable_sriov);
704 * pci_num_vf - return number of VFs associated with a PF device_release_driver
705 * @dev: the PCI device
707 * Returns number of VFs, or 0 if SR-IOV is not enabled.
709 int pci_num_vf(struct pci_dev *dev)
714 return dev->sriov->num_VFs;
716 EXPORT_SYMBOL_GPL(pci_num_vf);
719 * pci_vfs_assigned - returns number of VFs are assigned to a guest
720 * @dev: the PCI device
722 * Returns number of VFs belonging to this device that are assigned to a guest.
723 * If device is not a physical function returns 0.
725 int pci_vfs_assigned(struct pci_dev *dev)
727 struct pci_dev *vfdev;
728 unsigned int vfs_assigned = 0;
729 unsigned short dev_id;
731 /* only search if we are a PF */
736 * determine the device ID for the VFs, the vendor ID will be the
737 * same as the PF so there is no need to check for that one
739 dev_id = dev->sriov->vf_device;
741 /* loop through all the VFs to see if we own any that are assigned */
742 vfdev = pci_get_device(dev->vendor, dev_id, NULL);
745 * It is considered assigned if it is a virtual function with
746 * our dev as the physical function and the assigned bit is set
748 if (vfdev->is_virtfn && (vfdev->physfn == dev) &&
749 pci_is_dev_assigned(vfdev))
752 vfdev = pci_get_device(dev->vendor, dev_id, vfdev);
757 EXPORT_SYMBOL_GPL(pci_vfs_assigned);
760 * pci_sriov_set_totalvfs -- reduce the TotalVFs available
761 * @dev: the PCI PF device
762 * @numvfs: number that should be used for TotalVFs supported
764 * Should be called from PF driver's probe routine with
765 * device's mutex held.
767 * Returns 0 if PF is an SRIOV-capable device and
768 * value of numvfs valid. If not a PF return -ENOSYS;
769 * if numvfs is invalid return -EINVAL;
770 * if VFs already enabled, return -EBUSY.
772 int pci_sriov_set_totalvfs(struct pci_dev *dev, u16 numvfs)
776 if (numvfs > dev->sriov->total_VFs)
779 /* Shouldn't change if VFs already enabled */
780 if (dev->sriov->ctrl & PCI_SRIOV_CTRL_VFE)
783 dev->sriov->driver_max_VFs = numvfs;
787 EXPORT_SYMBOL_GPL(pci_sriov_set_totalvfs);
790 * pci_sriov_get_totalvfs -- get total VFs supported on this device
791 * @dev: the PCI PF device
793 * For a PCIe device with SRIOV support, return the PCIe
794 * SRIOV capability value of TotalVFs or the value of driver_max_VFs
795 * if the driver reduced it. Otherwise 0.
797 int pci_sriov_get_totalvfs(struct pci_dev *dev)
802 if (dev->sriov->driver_max_VFs)
803 return dev->sriov->driver_max_VFs;
805 return dev->sriov->total_VFs;
807 EXPORT_SYMBOL_GPL(pci_sriov_get_totalvfs);