1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright 2006-2007 Michael Ellerman, IBM Corp.
7 #include <linux/crash_dump.h>
8 #include <linux/device.h>
10 #include <linux/irqdomain.h>
11 #include <linux/msi.h>
14 #include <asm/hw_irq.h>
15 #include <asm/ppc-pci.h>
16 #include <asm/machdep.h>
21 static int query_token, change_token;
23 #define RTAS_QUERY_FN 0
24 #define RTAS_CHANGE_FN 1
25 #define RTAS_RESET_FN 2
26 #define RTAS_CHANGE_MSI_FN 3
27 #define RTAS_CHANGE_MSIX_FN 4
28 #define RTAS_CHANGE_32MSI_FN 5
29 #define RTAS_CHANGE_32MSIX_FN 6
33 static int rtas_change_msi(struct pci_dn *pdn, u32 func, u32 num_irqs)
35 u32 addr, seq_num, rtas_ret[3];
39 addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
40 buid = pdn->phb->buid;
44 if (func == RTAS_CHANGE_MSI_FN || func == RTAS_CHANGE_MSIX_FN ||
45 func == RTAS_CHANGE_32MSI_FN || func == RTAS_CHANGE_32MSIX_FN)
46 rc = rtas_call(change_token, 6, 4, rtas_ret, addr,
47 BUID_HI(buid), BUID_LO(buid),
48 func, num_irqs, seq_num);
50 rc = rtas_call(change_token, 6, 3, rtas_ret, addr,
51 BUID_HI(buid), BUID_LO(buid),
52 func, num_irqs, seq_num);
54 seq_num = rtas_ret[1];
55 } while (rtas_busy_delay(rc));
58 * If the RTAS call succeeded, return the number of irqs allocated.
59 * If not, make sure we return a negative error code.
66 pr_debug("rtas_msi: ibm,change_msi(func=%d,num=%d), got %d rc = %d\n",
67 func, num_irqs, rtas_ret[0], rc);
72 static void rtas_disable_msi(struct pci_dev *pdev)
76 pdn = pci_get_pdn(pdev);
81 * disabling MSI with the explicit interface also disables MSI-X
83 if (rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, 0) != 0) {
85 * may have failed because explicit interface is not
88 if (rtas_change_msi(pdn, RTAS_CHANGE_FN, 0) != 0) {
89 pr_debug("rtas_msi: Setting MSIs to 0 failed!\n");
94 static int rtas_query_irq_number(struct pci_dn *pdn, int offset)
96 u32 addr, rtas_ret[2];
100 addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
101 buid = pdn->phb->buid;
104 rc = rtas_call(query_token, 4, 3, rtas_ret, addr,
105 BUID_HI(buid), BUID_LO(buid), offset);
106 } while (rtas_busy_delay(rc));
109 pr_debug("rtas_msi: error (%d) querying source number\n", rc);
116 static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
118 struct device_node *dn;
122 dn = pci_device_to_OF_node(pdev);
124 p = of_get_property(dn, prop_name, NULL);
126 pr_debug("rtas_msi: No %s on %pOF\n", prop_name, dn);
130 req_msi = be32_to_cpup(p);
131 if (req_msi < nvec) {
132 pr_debug("rtas_msi: %s requests < %d MSIs\n", prop_name, nvec);
134 if (req_msi == 0) /* Be paranoid */
143 static int check_req_msi(struct pci_dev *pdev, int nvec)
145 return check_req(pdev, nvec, "ibm,req#msi");
148 static int check_req_msix(struct pci_dev *pdev, int nvec)
150 return check_req(pdev, nvec, "ibm,req#msi-x");
153 /* Quota calculation */
155 static struct device_node *__find_pe_total_msi(struct device_node *node, int *total)
157 struct device_node *dn;
160 dn = of_node_get(node);
162 p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
164 pr_debug("rtas_msi: found prop on dn %pOF\n",
166 *total = be32_to_cpup(p);
170 dn = of_get_next_parent(dn);
176 static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
178 return __find_pe_total_msi(pci_device_to_OF_node(dev), total);
181 static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
183 struct device_node *dn;
184 struct eeh_dev *edev;
186 /* Found our PE and assume 8 at that point. */
188 dn = pci_device_to_OF_node(dev);
192 /* Get the top level device in the PE */
193 edev = pdn_to_eeh_dev(PCI_DN(dn));
195 edev = list_first_entry(&edev->pe->edevs, struct eeh_dev,
197 dn = pci_device_to_OF_node(edev->pdev);
201 /* We actually want the parent */
202 dn = of_get_parent(dn);
206 /* Hardcode of 8 for old firmwares */
208 pr_debug("rtas_msi: using PE dn %pOF\n", dn);
214 struct device_node *requestor;
222 static void *count_non_bridge_devices(struct device_node *dn, void *data)
224 struct msi_counts *counts = data;
228 pr_debug("rtas_msi: counting %pOF\n", dn);
230 p = of_get_property(dn, "class-code", NULL);
231 class = p ? be32_to_cpup(p) : 0;
233 if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
234 counts->num_devices++;
239 static void *count_spare_msis(struct device_node *dn, void *data)
241 struct msi_counts *counts = data;
245 if (dn == counts->requestor)
246 req = counts->request;
248 /* We don't know if a driver will try to use MSI or MSI-X,
249 * so we just have to punt and use the larger of the two. */
251 p = of_get_property(dn, "ibm,req#msi", NULL);
253 req = be32_to_cpup(p);
255 p = of_get_property(dn, "ibm,req#msi-x", NULL);
257 req = max(req, (int)be32_to_cpup(p));
260 if (req < counts->quota)
261 counts->spare += counts->quota - req;
262 else if (req > counts->quota)
263 counts->over_quota++;
268 static int msi_quota_for_device(struct pci_dev *dev, int request)
270 struct device_node *pe_dn;
271 struct msi_counts counts;
274 pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
277 pe_dn = find_pe_total_msi(dev, &total);
279 pe_dn = find_pe_dn(dev, &total);
282 pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
286 pr_debug("rtas_msi: found PE %pOF\n", pe_dn);
288 memset(&counts, 0, sizeof(struct msi_counts));
290 /* Work out how many devices we have below this PE */
291 pci_traverse_device_nodes(pe_dn, count_non_bridge_devices, &counts);
293 if (counts.num_devices == 0) {
294 pr_err("rtas_msi: found 0 devices under PE for %s\n",
299 counts.quota = total / counts.num_devices;
300 if (request <= counts.quota)
303 /* else, we have some more calculating to do */
304 counts.requestor = pci_device_to_OF_node(dev);
305 counts.request = request;
306 pci_traverse_device_nodes(pe_dn, count_spare_msis, &counts);
308 /* If the quota isn't an integer multiple of the total, we can
309 * use the remainder as spare MSIs for anyone that wants them. */
310 counts.spare += total % counts.num_devices;
312 /* Divide any spare by the number of over-quota requestors */
313 if (counts.over_quota)
314 counts.quota += counts.spare / counts.over_quota;
316 /* And finally clamp the request to the possibly adjusted quota */
317 request = min(counts.quota, request);
319 pr_debug("rtas_msi: request clamped to quota %d\n", request);
326 static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
328 u32 addr_hi, addr_lo;
331 * We should only get in here for IODA1 configs. This is based on the
332 * fact that we using RTAS for MSIs, we don't have the 32 bit MSI RTAS
333 * support, and we are in a PCIe Gen2 slot.
336 "rtas_msi: No 32 bit MSI firmware support, forcing 32 bit MSI\n");
337 pci_read_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, &addr_hi);
338 addr_lo = 0xffff0000 | ((addr_hi >> (48 - 32)) << 4);
339 pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_LO, addr_lo);
340 pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, 0);
343 static int rtas_prepare_msi_irqs(struct pci_dev *pdev, int nvec_in, int type,
344 msi_alloc_info_t *arg)
349 int use_32bit_msi_hack = 0;
351 if (type == PCI_CAP_ID_MSIX)
352 rc = check_req_msix(pdev, nvec);
354 rc = check_req_msi(pdev, nvec);
359 quota = msi_quota_for_device(pdev, nvec);
361 if (quota && quota < nvec)
365 * Firmware currently refuse any non power of two allocation
366 * so we round up if the quota will allow it.
368 if (type == PCI_CAP_ID_MSIX) {
369 int m = roundup_pow_of_two(nvec);
370 quota = msi_quota_for_device(pdev, m);
376 pdn = pci_get_pdn(pdev);
379 * Try the new more explicit firmware interface, if that fails fall
380 * back to the old interface. The old interface is known to never
384 if (type == PCI_CAP_ID_MSI) {
385 if (pdev->no_64bit_msi) {
386 rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
389 * We only want to run the 32 bit MSI hack below if
390 * the max bus speed is Gen2 speed
392 if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
395 use_32bit_msi_hack = 1;
401 rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
404 pr_debug("rtas_msi: trying the old firmware call.\n");
405 rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
408 if (use_32bit_msi_hack && rc > 0)
409 rtas_hack_32bit_msi_gen2(pdev);
411 if (pdev->no_64bit_msi)
412 rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSIX_FN, nvec);
414 rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
418 if (nvec != nvec_in) {
422 pr_debug("rtas_msi: rtas_change_msi() failed\n");
429 static int pseries_msi_ops_prepare(struct irq_domain *domain, struct device *dev,
430 int nvec, msi_alloc_info_t *arg)
432 struct pci_dev *pdev = to_pci_dev(dev);
433 int type = pdev->msix_enabled ? PCI_CAP_ID_MSIX : PCI_CAP_ID_MSI;
435 return rtas_prepare_msi_irqs(pdev, nvec, type, arg);
439 * ->msi_free() is called before irq_domain_free_irqs_top() when the
440 * handler data is still available. Use that to clear the XIVE
443 static void pseries_msi_ops_msi_free(struct irq_domain *domain,
444 struct msi_domain_info *info,
448 xive_irq_free_data(irq);
452 * RTAS can not disable one MSI at a time. It's all or nothing. Do it
453 * at the end after all IRQs have been freed.
455 static void pseries_msi_post_free(struct irq_domain *domain, struct device *dev)
457 if (WARN_ON_ONCE(!dev_is_pci(dev)))
460 rtas_disable_msi(to_pci_dev(dev));
463 static struct msi_domain_ops pseries_pci_msi_domain_ops = {
464 .msi_prepare = pseries_msi_ops_prepare,
465 .msi_free = pseries_msi_ops_msi_free,
466 .msi_post_free = pseries_msi_post_free,
469 static void pseries_msi_shutdown(struct irq_data *d)
472 if (d->chip->irq_shutdown)
473 d->chip->irq_shutdown(d);
476 static void pseries_msi_mask(struct irq_data *d)
479 irq_chip_mask_parent(d);
482 static void pseries_msi_unmask(struct irq_data *d)
484 pci_msi_unmask_irq(d);
485 irq_chip_unmask_parent(d);
488 static void pseries_msi_write_msg(struct irq_data *data, struct msi_msg *msg)
490 struct msi_desc *entry = irq_data_get_msi_desc(data);
493 * Do not update the MSIx vector table. It's not strictly necessary
494 * because the table is initialized by the underlying hypervisor, PowerVM
495 * or QEMU/KVM. However, if the MSIx vector entry is cleared, any further
496 * activation will fail. This can happen in some drivers (eg. IPR) which
497 * deactivate an IRQ used for testing MSI support.
502 static struct irq_chip pseries_pci_msi_irq_chip = {
503 .name = "pSeries-PCI-MSI",
504 .irq_shutdown = pseries_msi_shutdown,
505 .irq_mask = pseries_msi_mask,
506 .irq_unmask = pseries_msi_unmask,
507 .irq_eoi = irq_chip_eoi_parent,
508 .irq_write_msi_msg = pseries_msi_write_msg,
513 * Set MSI_FLAG_MSIX_CONTIGUOUS as there is no way to express to
514 * firmware to request a discontiguous or non-zero based range of
515 * MSI-X entries. Core code will reject such setup attempts.
517 static struct msi_domain_info pseries_msi_domain_info = {
518 .flags = (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
519 MSI_FLAG_MULTI_PCI_MSI | MSI_FLAG_PCI_MSIX |
520 MSI_FLAG_MSIX_CONTIGUOUS),
521 .ops = &pseries_pci_msi_domain_ops,
522 .chip = &pseries_pci_msi_irq_chip,
525 static void pseries_msi_compose_msg(struct irq_data *data, struct msi_msg *msg)
527 __pci_read_msi_msg(irq_data_get_msi_desc(data), msg);
530 static struct irq_chip pseries_msi_irq_chip = {
531 .name = "pSeries-MSI",
532 .irq_shutdown = pseries_msi_shutdown,
533 .irq_mask = irq_chip_mask_parent,
534 .irq_unmask = irq_chip_unmask_parent,
535 .irq_eoi = irq_chip_eoi_parent,
536 .irq_set_affinity = irq_chip_set_affinity_parent,
537 .irq_compose_msi_msg = pseries_msi_compose_msg,
540 static int pseries_irq_parent_domain_alloc(struct irq_domain *domain, unsigned int virq,
541 irq_hw_number_t hwirq)
543 struct irq_fwspec parent_fwspec;
546 parent_fwspec.fwnode = domain->parent->fwnode;
547 parent_fwspec.param_count = 2;
548 parent_fwspec.param[0] = hwirq;
549 parent_fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
551 ret = irq_domain_alloc_irqs_parent(domain, virq, 1, &parent_fwspec);
558 static int pseries_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
559 unsigned int nr_irqs, void *arg)
561 struct pci_controller *phb = domain->host_data;
562 msi_alloc_info_t *info = arg;
563 struct msi_desc *desc = info->desc;
564 struct pci_dev *pdev = msi_desc_to_pci_dev(desc);
568 hwirq = rtas_query_irq_number(pci_get_pdn(pdev), desc->msi_index);
570 dev_err(&pdev->dev, "Failed to query HW IRQ: %d\n", hwirq);
574 dev_dbg(&pdev->dev, "%s bridge %pOF %d/%x #%d\n", __func__,
575 phb->dn, virq, hwirq, nr_irqs);
577 for (i = 0; i < nr_irqs; i++) {
578 ret = pseries_irq_parent_domain_alloc(domain, virq + i, hwirq + i);
582 irq_domain_set_hwirq_and_chip(domain, virq + i, hwirq + i,
583 &pseries_msi_irq_chip, domain->host_data);
589 /* TODO: handle RTAS cleanup in ->msi_finish() ? */
590 irq_domain_free_irqs_parent(domain, virq, i - 1);
594 static void pseries_irq_domain_free(struct irq_domain *domain, unsigned int virq,
595 unsigned int nr_irqs)
597 struct irq_data *d = irq_domain_get_irq_data(domain, virq);
598 struct pci_controller *phb = irq_data_get_irq_chip_data(d);
600 pr_debug("%s bridge %pOF %d #%d\n", __func__, phb->dn, virq, nr_irqs);
602 /* XIVE domain data is cleared through ->msi_free() */
605 static const struct irq_domain_ops pseries_irq_domain_ops = {
606 .alloc = pseries_irq_domain_alloc,
607 .free = pseries_irq_domain_free,
610 static int __pseries_msi_allocate_domains(struct pci_controller *phb,
613 struct irq_domain *parent = irq_get_default_host();
615 phb->fwnode = irq_domain_alloc_named_id_fwnode("pSeries-MSI",
620 phb->dev_domain = irq_domain_create_hierarchy(parent, 0, count,
622 &pseries_irq_domain_ops, phb);
623 if (!phb->dev_domain) {
624 pr_err("PCI: failed to create IRQ domain bridge %pOF (domain %d)\n",
625 phb->dn, phb->global_number);
626 irq_domain_free_fwnode(phb->fwnode);
630 phb->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(phb->dn),
631 &pseries_msi_domain_info,
633 if (!phb->msi_domain) {
634 pr_err("PCI: failed to create MSI IRQ domain bridge %pOF (domain %d)\n",
635 phb->dn, phb->global_number);
636 irq_domain_free_fwnode(phb->fwnode);
637 irq_domain_remove(phb->dev_domain);
644 int pseries_msi_allocate_domains(struct pci_controller *phb)
648 if (!__find_pe_total_msi(phb->dn, &count)) {
649 pr_err("PCI: failed to find MSIs for bridge %pOF (domain %d)\n",
650 phb->dn, phb->global_number);
654 return __pseries_msi_allocate_domains(phb, count);
657 void pseries_msi_free_domains(struct pci_controller *phb)
660 irq_domain_remove(phb->msi_domain);
662 irq_domain_remove(phb->dev_domain);
664 irq_domain_free_fwnode(phb->fwnode);
667 static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
669 /* No LSI -> leave MSIs (if any) configured */
671 dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
675 /* No MSI -> MSIs can't have been assigned by fw, leave LSI */
676 if (check_req_msi(pdev, 1) && check_req_msix(pdev, 1)) {
677 dev_dbg(&pdev->dev, "rtas_msi: no req#msi/x, nothing to do.\n");
681 dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
682 rtas_disable_msi(pdev);
685 static int rtas_msi_init(void)
687 query_token = rtas_function_token(RTAS_FN_IBM_QUERY_INTERRUPT_SOURCE_NUMBER);
688 change_token = rtas_function_token(RTAS_FN_IBM_CHANGE_MSI);
690 if ((query_token == RTAS_UNKNOWN_SERVICE) ||
691 (change_token == RTAS_UNKNOWN_SERVICE)) {
692 pr_debug("rtas_msi: no RTAS tokens, no MSI support.\n");
696 pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
698 WARN_ON(ppc_md.pci_irq_fixup);
699 ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
703 machine_arch_initcall(pseries, rtas_msi_init);