4 * Copyright (C) 2009 Intel Corporation, Yu Zhao <yu.zhao@intel.com>
6 * PCI Express I/O Virtualization (IOV) support.
10 #include <linux/pci.h>
11 #include <linux/mutex.h>
12 #include <linux/string.h>
13 #include <linux/delay.h>
16 #define VIRTFN_ID_LEN 16
18 static inline u8 virtfn_bus(struct pci_dev *dev, int id)
20 return dev->bus->number + ((dev->devfn + dev->sriov->offset +
21 dev->sriov->stride * id) >> 8);
24 static inline u8 virtfn_devfn(struct pci_dev *dev, int id)
26 return (dev->devfn + dev->sriov->offset +
27 dev->sriov->stride * id) & 0xff;
30 static struct pci_bus *virtfn_add_bus(struct pci_bus *bus, int busnr)
33 struct pci_bus *child;
35 if (bus->number == busnr)
38 child = pci_find_bus(pci_domain_nr(bus), busnr);
42 child = pci_add_new_bus(bus, NULL, busnr);
46 child->subordinate = busnr;
47 child->dev.parent = bus->bridge;
48 rc = pci_bus_add_child(child);
50 pci_remove_bus(child);
57 static void virtfn_remove_bus(struct pci_bus *bus, int busnr)
59 struct pci_bus *child;
61 if (bus->number == busnr)
64 child = pci_find_bus(pci_domain_nr(bus), busnr);
67 if (list_empty(&child->devices))
68 pci_remove_bus(child);
71 static int virtfn_add(struct pci_dev *dev, int id, int reset)
76 char buf[VIRTFN_ID_LEN];
77 struct pci_dev *virtfn;
79 struct pci_sriov *iov = dev->sriov;
81 virtfn = alloc_pci_dev();
85 mutex_lock(&iov->dev->sriov->lock);
86 virtfn->bus = virtfn_add_bus(dev->bus, virtfn_bus(dev, id));
89 mutex_unlock(&iov->dev->sriov->lock);
92 virtfn->devfn = virtfn_devfn(dev, id);
93 virtfn->vendor = dev->vendor;
94 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_DID, &virtfn->device);
95 pci_setup_device(virtfn);
96 virtfn->dev.parent = dev->dev.parent;
98 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
99 res = dev->resource + PCI_IOV_RESOURCES + i;
102 virtfn->resource[i].name = pci_name(virtfn);
103 virtfn->resource[i].flags = res->flags;
104 size = resource_size(res);
105 do_div(size, iov->total);
106 virtfn->resource[i].start = res->start + size * id;
107 virtfn->resource[i].end = virtfn->resource[i].start + size - 1;
108 rc = request_resource(res, &virtfn->resource[i]);
113 pci_execute_reset_function(virtfn);
115 pci_device_add(virtfn, virtfn->bus);
116 mutex_unlock(&iov->dev->sriov->lock);
118 virtfn->physfn = pci_dev_get(dev);
119 virtfn->is_virtfn = 1;
121 rc = pci_bus_add_device(virtfn);
124 sprintf(buf, "virtfn%u", id);
125 rc = sysfs_create_link(&dev->dev.kobj, &virtfn->dev.kobj, buf);
128 rc = sysfs_create_link(&virtfn->dev.kobj, &dev->dev.kobj, "physfn");
132 kobject_uevent(&virtfn->dev.kobj, KOBJ_CHANGE);
137 sysfs_remove_link(&dev->dev.kobj, buf);
140 mutex_lock(&iov->dev->sriov->lock);
141 pci_remove_bus_device(virtfn);
142 virtfn_remove_bus(dev->bus, virtfn_bus(dev, id));
143 mutex_unlock(&iov->dev->sriov->lock);
148 static void virtfn_remove(struct pci_dev *dev, int id, int reset)
150 char buf[VIRTFN_ID_LEN];
152 struct pci_dev *virtfn;
153 struct pci_sriov *iov = dev->sriov;
155 bus = pci_find_bus(pci_domain_nr(dev->bus), virtfn_bus(dev, id));
159 virtfn = pci_get_slot(bus, virtfn_devfn(dev, id));
166 device_release_driver(&virtfn->dev);
167 pci_execute_reset_function(virtfn);
170 sprintf(buf, "virtfn%u", id);
171 sysfs_remove_link(&dev->dev.kobj, buf);
172 sysfs_remove_link(&virtfn->dev.kobj, "physfn");
174 mutex_lock(&iov->dev->sriov->lock);
175 pci_remove_bus_device(virtfn);
176 virtfn_remove_bus(dev->bus, virtfn_bus(dev, id));
177 mutex_unlock(&iov->dev->sriov->lock);
182 static int sriov_migration(struct pci_dev *dev)
185 struct pci_sriov *iov = dev->sriov;
190 if (!(iov->cap & PCI_SRIOV_CAP_VFM))
193 pci_read_config_word(dev, iov->pos + PCI_SRIOV_STATUS, &status);
194 if (!(status & PCI_SRIOV_STATUS_VFM))
197 schedule_work(&iov->mtask);
202 static void sriov_migration_task(struct work_struct *work)
207 struct pci_sriov *iov = container_of(work, struct pci_sriov, mtask);
209 for (i = iov->initial; i < iov->nr_virtfn; i++) {
210 state = readb(iov->mstate + i);
211 if (state == PCI_SRIOV_VFM_MI) {
212 writeb(PCI_SRIOV_VFM_AV, iov->mstate + i);
213 state = readb(iov->mstate + i);
214 if (state == PCI_SRIOV_VFM_AV)
215 virtfn_add(iov->self, i, 1);
216 } else if (state == PCI_SRIOV_VFM_MO) {
217 virtfn_remove(iov->self, i, 1);
218 writeb(PCI_SRIOV_VFM_UA, iov->mstate + i);
219 state = readb(iov->mstate + i);
220 if (state == PCI_SRIOV_VFM_AV)
221 virtfn_add(iov->self, i, 0);
225 pci_read_config_word(iov->self, iov->pos + PCI_SRIOV_STATUS, &status);
226 status &= ~PCI_SRIOV_STATUS_VFM;
227 pci_write_config_word(iov->self, iov->pos + PCI_SRIOV_STATUS, status);
230 static int sriov_enable_migration(struct pci_dev *dev, int nr_virtfn)
235 struct pci_sriov *iov = dev->sriov;
237 if (nr_virtfn <= iov->initial)
240 pci_read_config_dword(dev, iov->pos + PCI_SRIOV_VFM, &table);
241 bir = PCI_SRIOV_VFM_BIR(table);
242 if (bir > PCI_STD_RESOURCE_END)
245 table = PCI_SRIOV_VFM_OFFSET(table);
246 if (table + nr_virtfn > pci_resource_len(dev, bir))
249 pa = pci_resource_start(dev, bir) + table;
250 iov->mstate = ioremap(pa, nr_virtfn);
254 INIT_WORK(&iov->mtask, sriov_migration_task);
256 iov->ctrl |= PCI_SRIOV_CTRL_VFM | PCI_SRIOV_CTRL_INTR;
257 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
262 static void sriov_disable_migration(struct pci_dev *dev)
264 struct pci_sriov *iov = dev->sriov;
266 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFM | PCI_SRIOV_CTRL_INTR);
267 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
269 cancel_work_sync(&iov->mtask);
270 iounmap(iov->mstate);
273 static int sriov_enable(struct pci_dev *dev, int nr_virtfn)
278 u16 offset, stride, initial;
279 struct resource *res;
280 struct pci_dev *pdev;
281 struct pci_sriov *iov = dev->sriov;
289 pci_read_config_word(dev, iov->pos + PCI_SRIOV_INITIAL_VF, &initial);
290 if (initial > iov->total ||
291 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (initial != iov->total)))
294 if (nr_virtfn < 0 || nr_virtfn > iov->total ||
295 (!(iov->cap & PCI_SRIOV_CAP_VFM) && (nr_virtfn > initial)))
298 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, nr_virtfn);
299 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_OFFSET, &offset);
300 pci_read_config_word(dev, iov->pos + PCI_SRIOV_VF_STRIDE, &stride);
301 if (!offset || (nr_virtfn > 1 && !stride))
305 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
306 res = dev->resource + PCI_IOV_RESOURCES + i;
310 if (nres != iov->nres) {
311 dev_err(&dev->dev, "not enough MMIO resources for SR-IOV\n");
315 iov->offset = offset;
316 iov->stride = stride;
318 if (virtfn_bus(dev, nr_virtfn - 1) > dev->bus->subordinate) {
319 dev_err(&dev->dev, "SR-IOV: bus number out of range\n");
323 if (iov->link != dev->devfn) {
324 pdev = pci_get_slot(dev->bus, iov->link);
330 if (!pdev->is_physfn)
333 rc = sysfs_create_link(&dev->dev.kobj,
334 &pdev->dev.kobj, "dep_link");
339 iov->ctrl |= PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE;
340 pci_block_user_cfg_access(dev);
341 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
343 pci_unblock_user_cfg_access(dev);
345 iov->initial = initial;
346 if (nr_virtfn < initial)
349 for (i = 0; i < initial; i++) {
350 rc = virtfn_add(dev, i, 0);
355 if (iov->cap & PCI_SRIOV_CAP_VFM) {
356 rc = sriov_enable_migration(dev, nr_virtfn);
361 kobject_uevent(&dev->dev.kobj, KOBJ_CHANGE);
362 iov->nr_virtfn = nr_virtfn;
367 for (j = 0; j < i; j++)
368 virtfn_remove(dev, j, 0);
370 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
371 pci_block_user_cfg_access(dev);
372 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
374 pci_unblock_user_cfg_access(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->nr_virtfn; i++)
394 virtfn_remove(dev, i, 0);
396 iov->ctrl &= ~(PCI_SRIOV_CTRL_VFE | PCI_SRIOV_CTRL_MSE);
397 pci_block_user_cfg_access(dev);
398 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
400 pci_unblock_user_cfg_access(dev);
402 if (iov->link != dev->devfn)
403 sysfs_remove_link(&dev->dev.kobj, "dep_link");
408 static int sriov_init(struct pci_dev *dev, int pos)
414 u16 ctrl, total, offset, stride;
415 struct pci_sriov *iov;
416 struct resource *res;
417 struct pci_dev *pdev;
419 if (dev->pcie_type != PCI_EXP_TYPE_RC_END &&
420 dev->pcie_type != PCI_EXP_TYPE_ENDPOINT)
423 pci_read_config_word(dev, pos + PCI_SRIOV_CTRL, &ctrl);
424 if (ctrl & PCI_SRIOV_CTRL_VFE) {
425 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, 0);
429 pci_read_config_word(dev, pos + PCI_SRIOV_TOTAL_VF, &total);
434 list_for_each_entry(pdev, &dev->bus->devices, bus_list)
439 if (pci_ari_enabled(dev->bus))
440 ctrl |= PCI_SRIOV_CTRL_ARI;
443 pci_write_config_word(dev, pos + PCI_SRIOV_CTRL, ctrl);
444 pci_write_config_word(dev, pos + PCI_SRIOV_NUM_VF, total);
445 pci_read_config_word(dev, pos + PCI_SRIOV_VF_OFFSET, &offset);
446 pci_read_config_word(dev, pos + PCI_SRIOV_VF_STRIDE, &stride);
447 if (!offset || (total > 1 && !stride))
450 pci_read_config_dword(dev, pos + PCI_SRIOV_SUP_PGSIZE, &pgsz);
451 i = PAGE_SHIFT > 12 ? PAGE_SHIFT - 12 : 0;
452 pgsz &= ~((1 << i) - 1);
457 pci_write_config_dword(dev, pos + PCI_SRIOV_SYS_PGSIZE, pgsz);
460 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
461 res = dev->resource + PCI_IOV_RESOURCES + i;
462 i += __pci_read_base(dev, pci_bar_unknown, res,
463 pos + PCI_SRIOV_BAR + i * 4);
466 if (resource_size(res) & (PAGE_SIZE - 1)) {
470 res->end = res->start + resource_size(res) * total - 1;
474 iov = kzalloc(sizeof(*iov), GFP_KERNEL);
484 iov->offset = offset;
485 iov->stride = stride;
488 pci_read_config_dword(dev, pos + PCI_SRIOV_CAP, &iov->cap);
489 pci_read_config_byte(dev, pos + PCI_SRIOV_FUNC_LINK, &iov->link);
492 iov->dev = pci_dev_get(pdev);
495 mutex_init(&iov->lock);
504 for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
505 res = dev->resource + PCI_IOV_RESOURCES + i;
512 static void sriov_release(struct pci_dev *dev)
514 BUG_ON(dev->sriov->nr_virtfn);
516 if (dev == dev->sriov->dev)
517 mutex_destroy(&dev->sriov->lock);
519 pci_dev_put(dev->sriov->dev);
525 static void sriov_restore_state(struct pci_dev *dev)
529 struct pci_sriov *iov = dev->sriov;
531 pci_read_config_word(dev, iov->pos + PCI_SRIOV_CTRL, &ctrl);
532 if (ctrl & PCI_SRIOV_CTRL_VFE)
535 for (i = PCI_IOV_RESOURCES; i <= PCI_IOV_RESOURCE_END; i++)
536 pci_update_resource(dev, i);
538 pci_write_config_dword(dev, iov->pos + PCI_SRIOV_SYS_PGSIZE, iov->pgsz);
539 pci_write_config_word(dev, iov->pos + PCI_SRIOV_NUM_VF, iov->nr_virtfn);
540 pci_write_config_word(dev, iov->pos + PCI_SRIOV_CTRL, iov->ctrl);
541 if (iov->ctrl & PCI_SRIOV_CTRL_VFE)
546 * pci_iov_init - initialize the IOV capability
547 * @dev: the PCI device
549 * Returns 0 on success, or negative on failure.
551 int pci_iov_init(struct pci_dev *dev)
558 pos = pci_find_ext_capability(dev, PCI_EXT_CAP_ID_SRIOV);
560 return sriov_init(dev, pos);
566 * pci_iov_release - release resources used by the IOV capability
567 * @dev: the PCI device
569 void pci_iov_release(struct pci_dev *dev)
576 * pci_iov_resource_bar - get position of the SR-IOV BAR
577 * @dev: the PCI device
578 * @resno: the resource number
579 * @type: the BAR type to be filled in
581 * Returns position of the BAR encapsulated in the SR-IOV capability.
583 int pci_iov_resource_bar(struct pci_dev *dev, int resno,
584 enum pci_bar_type *type)
586 if (resno < PCI_IOV_RESOURCES || resno > PCI_IOV_RESOURCE_END)
589 BUG_ON(!dev->is_physfn);
591 *type = pci_bar_unknown;
593 return dev->sriov->pos + PCI_SRIOV_BAR +
594 4 * (resno - PCI_IOV_RESOURCES);
598 * pci_restore_iov_state - restore the state of the IOV capability
599 * @dev: the PCI device
601 void pci_restore_iov_state(struct pci_dev *dev)
604 sriov_restore_state(dev);
608 * pci_iov_bus_range - find bus range used by Virtual Function
611 * Returns max number of buses (exclude current one) used by Virtual
614 int pci_iov_bus_range(struct pci_bus *bus)
620 list_for_each_entry(dev, &bus->devices, bus_list) {
623 busnr = virtfn_bus(dev, dev->sriov->total - 1);
628 return max ? max - bus->number : 0;
632 * pci_enable_sriov - enable the SR-IOV capability
633 * @dev: the PCI device
635 * Returns 0 on success, or negative on failure.
637 int pci_enable_sriov(struct pci_dev *dev, int nr_virtfn)
644 return sriov_enable(dev, nr_virtfn);
646 EXPORT_SYMBOL_GPL(pci_enable_sriov);
649 * pci_disable_sriov - disable the SR-IOV capability
650 * @dev: the PCI device
652 void pci_disable_sriov(struct pci_dev *dev)
661 EXPORT_SYMBOL_GPL(pci_disable_sriov);
664 * pci_sriov_migration - notify SR-IOV core of Virtual Function Migration
665 * @dev: the PCI device
667 * Returns IRQ_HANDLED if the IRQ is handled, or IRQ_NONE if not.
669 * Physical Function driver is responsible to register IRQ handler using
670 * VF Migration Interrupt Message Number, and call this function when the
671 * interrupt is generated by the hardware.
673 irqreturn_t pci_sriov_migration(struct pci_dev *dev)
678 return sriov_migration(dev) ? IRQ_HANDLED : IRQ_NONE;
680 EXPORT_SYMBOL_GPL(pci_sriov_migration);