1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright 2006 Jake Moilanen <moilanen@austin.ibm.com>, IBM Corp.
4 * Copyright 2006-2007 Michael Ellerman, IBM Corp.
7 #include <linux/crash_dump.h>
8 #include <linux/device.h>
10 #include <linux/msi.h>
13 #include <asm/hw_irq.h>
14 #include <asm/ppc-pci.h>
15 #include <asm/machdep.h>
19 static int query_token, change_token;
21 #define RTAS_QUERY_FN 0
22 #define RTAS_CHANGE_FN 1
23 #define RTAS_RESET_FN 2
24 #define RTAS_CHANGE_MSI_FN 3
25 #define RTAS_CHANGE_MSIX_FN 4
26 #define RTAS_CHANGE_32MSI_FN 5
30 static int rtas_change_msi(struct pci_dn *pdn, u32 func, u32 num_irqs)
32 u32 addr, seq_num, rtas_ret[3];
36 addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
37 buid = pdn->phb->buid;
41 if (func == RTAS_CHANGE_MSI_FN || func == RTAS_CHANGE_MSIX_FN ||
42 func == RTAS_CHANGE_32MSI_FN)
43 rc = rtas_call(change_token, 6, 4, rtas_ret, addr,
44 BUID_HI(buid), BUID_LO(buid),
45 func, num_irqs, seq_num);
47 rc = rtas_call(change_token, 6, 3, rtas_ret, addr,
48 BUID_HI(buid), BUID_LO(buid),
49 func, num_irqs, seq_num);
51 seq_num = rtas_ret[1];
52 } while (rtas_busy_delay(rc));
55 * If the RTAS call succeeded, return the number of irqs allocated.
56 * If not, make sure we return a negative error code.
63 pr_debug("rtas_msi: ibm,change_msi(func=%d,num=%d), got %d rc = %d\n",
64 func, num_irqs, rtas_ret[0], rc);
69 static void rtas_disable_msi(struct pci_dev *pdev)
73 pdn = pci_get_pdn(pdev);
78 * disabling MSI with the explicit interface also disables MSI-X
80 if (rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, 0) != 0) {
82 * may have failed because explicit interface is not
85 if (rtas_change_msi(pdn, RTAS_CHANGE_FN, 0) != 0) {
86 pr_debug("rtas_msi: Setting MSIs to 0 failed!\n");
91 static int rtas_query_irq_number(struct pci_dn *pdn, int offset)
93 u32 addr, rtas_ret[2];
97 addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
98 buid = pdn->phb->buid;
101 rc = rtas_call(query_token, 4, 3, rtas_ret, addr,
102 BUID_HI(buid), BUID_LO(buid), offset);
103 } while (rtas_busy_delay(rc));
106 pr_debug("rtas_msi: error (%d) querying source number\n", rc);
113 static void rtas_teardown_msi_irqs(struct pci_dev *pdev)
115 struct msi_desc *entry;
117 for_each_pci_msi_entry(entry, pdev) {
121 irq_set_msi_desc(entry->irq, NULL);
122 irq_dispose_mapping(entry->irq);
125 rtas_disable_msi(pdev);
128 static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
130 struct device_node *dn;
134 dn = pci_device_to_OF_node(pdev);
136 p = of_get_property(dn, prop_name, NULL);
138 pr_debug("rtas_msi: No %s on %pOF\n", prop_name, dn);
142 req_msi = be32_to_cpup(p);
143 if (req_msi < nvec) {
144 pr_debug("rtas_msi: %s requests < %d MSIs\n", prop_name, nvec);
146 if (req_msi == 0) /* Be paranoid */
155 static int check_req_msi(struct pci_dev *pdev, int nvec)
157 return check_req(pdev, nvec, "ibm,req#msi");
160 static int check_req_msix(struct pci_dev *pdev, int nvec)
162 return check_req(pdev, nvec, "ibm,req#msi-x");
165 /* Quota calculation */
167 static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
169 struct device_node *dn;
172 dn = of_node_get(pci_device_to_OF_node(dev));
174 p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
176 pr_debug("rtas_msi: found prop on dn %pOF\n",
178 *total = be32_to_cpup(p);
182 dn = of_get_next_parent(dn);
188 static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
190 struct device_node *dn;
191 struct eeh_dev *edev;
193 /* Found our PE and assume 8 at that point. */
195 dn = pci_device_to_OF_node(dev);
199 /* Get the top level device in the PE */
200 edev = pdn_to_eeh_dev(PCI_DN(dn));
202 edev = list_first_entry(&edev->pe->edevs, struct eeh_dev,
204 dn = pci_device_to_OF_node(edev->pdev);
208 /* We actually want the parent */
209 dn = of_get_parent(dn);
213 /* Hardcode of 8 for old firmwares */
215 pr_debug("rtas_msi: using PE dn %pOF\n", dn);
221 struct device_node *requestor;
229 static void *count_non_bridge_devices(struct device_node *dn, void *data)
231 struct msi_counts *counts = data;
235 pr_debug("rtas_msi: counting %pOF\n", dn);
237 p = of_get_property(dn, "class-code", NULL);
238 class = p ? be32_to_cpup(p) : 0;
240 if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
241 counts->num_devices++;
246 static void *count_spare_msis(struct device_node *dn, void *data)
248 struct msi_counts *counts = data;
252 if (dn == counts->requestor)
253 req = counts->request;
255 /* We don't know if a driver will try to use MSI or MSI-X,
256 * so we just have to punt and use the larger of the two. */
258 p = of_get_property(dn, "ibm,req#msi", NULL);
260 req = be32_to_cpup(p);
262 p = of_get_property(dn, "ibm,req#msi-x", NULL);
264 req = max(req, (int)be32_to_cpup(p));
267 if (req < counts->quota)
268 counts->spare += counts->quota - req;
269 else if (req > counts->quota)
270 counts->over_quota++;
275 static int msi_quota_for_device(struct pci_dev *dev, int request)
277 struct device_node *pe_dn;
278 struct msi_counts counts;
281 pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
284 pe_dn = find_pe_total_msi(dev, &total);
286 pe_dn = find_pe_dn(dev, &total);
289 pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
293 pr_debug("rtas_msi: found PE %pOF\n", pe_dn);
295 memset(&counts, 0, sizeof(struct msi_counts));
297 /* Work out how many devices we have below this PE */
298 pci_traverse_device_nodes(pe_dn, count_non_bridge_devices, &counts);
300 if (counts.num_devices == 0) {
301 pr_err("rtas_msi: found 0 devices under PE for %s\n",
306 counts.quota = total / counts.num_devices;
307 if (request <= counts.quota)
310 /* else, we have some more calculating to do */
311 counts.requestor = pci_device_to_OF_node(dev);
312 counts.request = request;
313 pci_traverse_device_nodes(pe_dn, count_spare_msis, &counts);
315 /* If the quota isn't an integer multiple of the total, we can
316 * use the remainder as spare MSIs for anyone that wants them. */
317 counts.spare += total % counts.num_devices;
319 /* Divide any spare by the number of over-quota requestors */
320 if (counts.over_quota)
321 counts.quota += counts.spare / counts.over_quota;
323 /* And finally clamp the request to the possibly adjusted quota */
324 request = min(counts.quota, request);
326 pr_debug("rtas_msi: request clamped to quota %d\n", request);
333 static int check_msix_entries(struct pci_dev *pdev)
335 struct msi_desc *entry;
338 /* There's no way for us to express to firmware that we want
339 * a discontiguous, or non-zero based, range of MSI-X entries.
340 * So we must reject such requests. */
343 for_each_pci_msi_entry(entry, pdev) {
344 if (entry->msi_attrib.entry_nr != expected) {
345 pr_debug("rtas_msi: bad MSI-X entries.\n");
354 static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
356 u32 addr_hi, addr_lo;
359 * We should only get in here for IODA1 configs. This is based on the
360 * fact that we using RTAS for MSIs, we don't have the 32 bit MSI RTAS
361 * support, and we are in a PCIe Gen2 slot.
364 "rtas_msi: No 32 bit MSI firmware support, forcing 32 bit MSI\n");
365 pci_read_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, &addr_hi);
366 addr_lo = 0xffff0000 | ((addr_hi >> (48 - 32)) << 4);
367 pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_LO, addr_lo);
368 pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, 0);
371 static int rtas_setup_msi_irqs(struct pci_dev *pdev, int nvec_in, int type)
374 int hwirq, virq, i, quota, rc;
375 struct msi_desc *entry;
378 int use_32bit_msi_hack = 0;
380 if (type == PCI_CAP_ID_MSIX)
381 rc = check_req_msix(pdev, nvec);
383 rc = check_req_msi(pdev, nvec);
388 quota = msi_quota_for_device(pdev, nvec);
390 if (quota && quota < nvec)
393 if (type == PCI_CAP_ID_MSIX && check_msix_entries(pdev))
397 * Firmware currently refuse any non power of two allocation
398 * so we round up if the quota will allow it.
400 if (type == PCI_CAP_ID_MSIX) {
401 int m = roundup_pow_of_two(nvec);
402 quota = msi_quota_for_device(pdev, m);
408 pdn = pci_get_pdn(pdev);
411 * Try the new more explicit firmware interface, if that fails fall
412 * back to the old interface. The old interface is known to never
416 if (type == PCI_CAP_ID_MSI) {
417 if (pdev->no_64bit_msi) {
418 rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
421 * We only want to run the 32 bit MSI hack below if
422 * the max bus speed is Gen2 speed
424 if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
427 use_32bit_msi_hack = 1;
433 rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
436 pr_debug("rtas_msi: trying the old firmware call.\n");
437 rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
440 if (use_32bit_msi_hack && rc > 0)
441 rtas_hack_32bit_msi_gen2(pdev);
443 rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
446 if (nvec != nvec_in) {
450 pr_debug("rtas_msi: rtas_change_msi() failed\n");
455 for_each_pci_msi_entry(entry, pdev) {
456 hwirq = rtas_query_irq_number(pdn, i++);
458 pr_debug("rtas_msi: error (%d) getting hwirq\n", rc);
463 * Depending on the number of online CPUs in the original
464 * kernel, it is likely for CPU #0 to be offline in a kdump
465 * kernel. The associated IRQs in the affinity mappings
466 * provided by irq_create_affinity_masks() are thus not
467 * started by irq_startup(), as per-design for managed IRQs.
468 * This can be a problem with multi-queue block devices driven
469 * by blk-mq : such a non-started IRQ is very likely paired
470 * with the single queue enforced by blk-mq during kdump (see
471 * blk_mq_alloc_tag_set()). This causes the device to remain
472 * silent and likely hangs the guest at some point.
474 * We don't really care for fine-grained affinity when doing
475 * kdump actually : simply ignore the pre-computed affinity
476 * masks in this case and let the default mask with all CPUs
477 * be used when creating the IRQ mappings.
479 if (is_kdump_kernel())
480 virq = irq_create_mapping(NULL, hwirq);
482 virq = irq_create_mapping_affinity(NULL, hwirq,
486 pr_debug("rtas_msi: Failed mapping hwirq %d\n", hwirq);
490 dev_dbg(&pdev->dev, "rtas_msi: allocated virq %d\n", virq);
491 irq_set_msi_desc(virq, entry);
493 /* Read config space back so we can restore after reset */
494 __pci_read_msi_msg(entry, &msg);
501 static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
503 /* No LSI -> leave MSIs (if any) configured */
505 dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
509 /* No MSI -> MSIs can't have been assigned by fw, leave LSI */
510 if (check_req_msi(pdev, 1) && check_req_msix(pdev, 1)) {
511 dev_dbg(&pdev->dev, "rtas_msi: no req#msi/x, nothing to do.\n");
515 dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
516 rtas_disable_msi(pdev);
519 static int rtas_msi_init(void)
521 struct pci_controller *phb;
523 query_token = rtas_token("ibm,query-interrupt-source-number");
524 change_token = rtas_token("ibm,change-msi");
526 if ((query_token == RTAS_UNKNOWN_SERVICE) ||
527 (change_token == RTAS_UNKNOWN_SERVICE)) {
528 pr_debug("rtas_msi: no RTAS tokens, no MSI support.\n");
532 pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
534 WARN_ON(pseries_pci_controller_ops.setup_msi_irqs);
535 pseries_pci_controller_ops.setup_msi_irqs = rtas_setup_msi_irqs;
536 pseries_pci_controller_ops.teardown_msi_irqs = rtas_teardown_msi_irqs;
538 list_for_each_entry(phb, &hose_list, list_node) {
539 WARN_ON(phb->controller_ops.setup_msi_irqs);
540 phb->controller_ops.setup_msi_irqs = rtas_setup_msi_irqs;
541 phb->controller_ops.teardown_msi_irqs = rtas_teardown_msi_irqs;
544 WARN_ON(ppc_md.pci_irq_fixup);
545 ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
549 machine_arch_initcall(pseries, rtas_msi_init);