• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright 2006 Jake Moilanen <moilanen@austin.ibm.com>, IBM Corp.
4  * Copyright 2006-2007 Michael Ellerman, IBM Corp.
5  */
6 
7 #include <linux/crash_dump.h>
8 #include <linux/device.h>
9 #include <linux/irq.h>
10 #include <linux/msi.h>
11 
12 #include <asm/rtas.h>
13 #include <asm/hw_irq.h>
14 #include <asm/ppc-pci.h>
15 #include <asm/machdep.h>
16 
17 #include "pseries.h"
18 
19 static int query_token, change_token;
20 
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
27 
28 /* RTAS Helpers */
29 
rtas_change_msi(struct pci_dn * pdn,u32 func,u32 num_irqs)30 static int rtas_change_msi(struct pci_dn *pdn, u32 func, u32 num_irqs)
31 {
32 	u32 addr, seq_num, rtas_ret[3];
33 	unsigned long buid;
34 	int rc;
35 
36 	addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
37 	buid = pdn->phb->buid;
38 
39 	seq_num = 1;
40 	do {
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);
46 		else
47 			rc = rtas_call(change_token, 6, 3, rtas_ret, addr,
48 					BUID_HI(buid), BUID_LO(buid),
49 					func, num_irqs, seq_num);
50 
51 		seq_num = rtas_ret[1];
52 	} while (rtas_busy_delay(rc));
53 
54 	/*
55 	 * If the RTAS call succeeded, return the number of irqs allocated.
56 	 * If not, make sure we return a negative error code.
57 	 */
58 	if (rc == 0)
59 		rc = rtas_ret[0];
60 	else if (rc > 0)
61 		rc = -rc;
62 
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);
65 
66 	return rc;
67 }
68 
rtas_disable_msi(struct pci_dev * pdev)69 static void rtas_disable_msi(struct pci_dev *pdev)
70 {
71 	struct pci_dn *pdn;
72 
73 	pdn = pci_get_pdn(pdev);
74 	if (!pdn)
75 		return;
76 
77 	/*
78 	 * disabling MSI with the explicit interface also disables MSI-X
79 	 */
80 	if (rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, 0) != 0) {
81 		/*
82 		 * may have failed because explicit interface is not
83 		 * present
84 		 */
85 		if (rtas_change_msi(pdn, RTAS_CHANGE_FN, 0) != 0) {
86 			pr_debug("rtas_msi: Setting MSIs to 0 failed!\n");
87 		}
88 	}
89 }
90 
rtas_query_irq_number(struct pci_dn * pdn,int offset)91 static int rtas_query_irq_number(struct pci_dn *pdn, int offset)
92 {
93 	u32 addr, rtas_ret[2];
94 	unsigned long buid;
95 	int rc;
96 
97 	addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
98 	buid = pdn->phb->buid;
99 
100 	do {
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));
104 
105 	if (rc) {
106 		pr_debug("rtas_msi: error (%d) querying source number\n", rc);
107 		return rc;
108 	}
109 
110 	return rtas_ret[0];
111 }
112 
rtas_teardown_msi_irqs(struct pci_dev * pdev)113 static void rtas_teardown_msi_irqs(struct pci_dev *pdev)
114 {
115 	struct msi_desc *entry;
116 
117 	for_each_pci_msi_entry(entry, pdev) {
118 		if (!entry->irq)
119 			continue;
120 
121 		irq_set_msi_desc(entry->irq, NULL);
122 		irq_dispose_mapping(entry->irq);
123 	}
124 
125 	rtas_disable_msi(pdev);
126 }
127 
check_req(struct pci_dev * pdev,int nvec,char * prop_name)128 static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
129 {
130 	struct device_node *dn;
131 	const __be32 *p;
132 	u32 req_msi;
133 
134 	dn = pci_device_to_OF_node(pdev);
135 
136 	p = of_get_property(dn, prop_name, NULL);
137 	if (!p) {
138 		pr_debug("rtas_msi: No %s on %pOF\n", prop_name, dn);
139 		return -ENOENT;
140 	}
141 
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);
145 
146 		if (req_msi == 0) /* Be paranoid */
147 			return -ENOSPC;
148 
149 		return req_msi;
150 	}
151 
152 	return 0;
153 }
154 
check_req_msi(struct pci_dev * pdev,int nvec)155 static int check_req_msi(struct pci_dev *pdev, int nvec)
156 {
157 	return check_req(pdev, nvec, "ibm,req#msi");
158 }
159 
check_req_msix(struct pci_dev * pdev,int nvec)160 static int check_req_msix(struct pci_dev *pdev, int nvec)
161 {
162 	return check_req(pdev, nvec, "ibm,req#msi-x");
163 }
164 
165 /* Quota calculation */
166 
find_pe_total_msi(struct pci_dev * dev,int * total)167 static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
168 {
169 	struct device_node *dn;
170 	const __be32 *p;
171 
172 	dn = of_node_get(pci_device_to_OF_node(dev));
173 	while (dn) {
174 		p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
175 		if (p) {
176 			pr_debug("rtas_msi: found prop on dn %pOF\n",
177 				dn);
178 			*total = be32_to_cpup(p);
179 			return dn;
180 		}
181 
182 		dn = of_get_next_parent(dn);
183 	}
184 
185 	return NULL;
186 }
187 
find_pe_dn(struct pci_dev * dev,int * total)188 static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
189 {
190 	struct device_node *dn;
191 	struct eeh_dev *edev;
192 
193 	/* Found our PE and assume 8 at that point. */
194 
195 	dn = pci_device_to_OF_node(dev);
196 	if (!dn)
197 		return NULL;
198 
199 	/* Get the top level device in the PE */
200 	edev = pdn_to_eeh_dev(PCI_DN(dn));
201 	if (edev->pe)
202 		edev = list_first_entry(&edev->pe->edevs, struct eeh_dev,
203 					entry);
204 	dn = pci_device_to_OF_node(edev->pdev);
205 	if (!dn)
206 		return NULL;
207 
208 	/* We actually want the parent */
209 	dn = of_get_parent(dn);
210 	if (!dn)
211 		return NULL;
212 
213 	/* Hardcode of 8 for old firmwares */
214 	*total = 8;
215 	pr_debug("rtas_msi: using PE dn %pOF\n", dn);
216 
217 	return dn;
218 }
219 
220 struct msi_counts {
221 	struct device_node *requestor;
222 	int num_devices;
223 	int request;
224 	int quota;
225 	int spare;
226 	int over_quota;
227 };
228 
count_non_bridge_devices(struct device_node * dn,void * data)229 static void *count_non_bridge_devices(struct device_node *dn, void *data)
230 {
231 	struct msi_counts *counts = data;
232 	const __be32 *p;
233 	u32 class;
234 
235 	pr_debug("rtas_msi: counting %pOF\n", dn);
236 
237 	p = of_get_property(dn, "class-code", NULL);
238 	class = p ? be32_to_cpup(p) : 0;
239 
240 	if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
241 		counts->num_devices++;
242 
243 	return NULL;
244 }
245 
count_spare_msis(struct device_node * dn,void * data)246 static void *count_spare_msis(struct device_node *dn, void *data)
247 {
248 	struct msi_counts *counts = data;
249 	const __be32 *p;
250 	int req;
251 
252 	if (dn == counts->requestor)
253 		req = counts->request;
254 	else {
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. */
257 		req = 0;
258 		p = of_get_property(dn, "ibm,req#msi", NULL);
259 		if (p)
260 			req = be32_to_cpup(p);
261 
262 		p = of_get_property(dn, "ibm,req#msi-x", NULL);
263 		if (p)
264 			req = max(req, (int)be32_to_cpup(p));
265 	}
266 
267 	if (req < counts->quota)
268 		counts->spare += counts->quota - req;
269 	else if (req > counts->quota)
270 		counts->over_quota++;
271 
272 	return NULL;
273 }
274 
msi_quota_for_device(struct pci_dev * dev,int request)275 static int msi_quota_for_device(struct pci_dev *dev, int request)
276 {
277 	struct device_node *pe_dn;
278 	struct msi_counts counts;
279 	int total;
280 
281 	pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
282 		  request);
283 
284 	pe_dn = find_pe_total_msi(dev, &total);
285 	if (!pe_dn)
286 		pe_dn = find_pe_dn(dev, &total);
287 
288 	if (!pe_dn) {
289 		pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
290 		goto out;
291 	}
292 
293 	pr_debug("rtas_msi: found PE %pOF\n", pe_dn);
294 
295 	memset(&counts, 0, sizeof(struct msi_counts));
296 
297 	/* Work out how many devices we have below this PE */
298 	pci_traverse_device_nodes(pe_dn, count_non_bridge_devices, &counts);
299 
300 	if (counts.num_devices == 0) {
301 		pr_err("rtas_msi: found 0 devices under PE for %s\n",
302 			pci_name(dev));
303 		goto out;
304 	}
305 
306 	counts.quota = total / counts.num_devices;
307 	if (request <= counts.quota)
308 		goto out;
309 
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);
314 
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;
318 
319 	/* Divide any spare by the number of over-quota requestors */
320 	if (counts.over_quota)
321 		counts.quota += counts.spare / counts.over_quota;
322 
323 	/* And finally clamp the request to the possibly adjusted quota */
324 	request = min(counts.quota, request);
325 
326 	pr_debug("rtas_msi: request clamped to quota %d\n", request);
327 out:
328 	of_node_put(pe_dn);
329 
330 	return request;
331 }
332 
check_msix_entries(struct pci_dev * pdev)333 static int check_msix_entries(struct pci_dev *pdev)
334 {
335 	struct msi_desc *entry;
336 	int expected;
337 
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. */
341 
342 	expected = 0;
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");
346 			return -EINVAL;
347 		}
348 		expected++;
349 	}
350 
351 	return 0;
352 }
353 
rtas_hack_32bit_msi_gen2(struct pci_dev * pdev)354 static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
355 {
356 	u32 addr_hi, addr_lo;
357 
358 	/*
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.
362 	 */
363 	dev_info(&pdev->dev,
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);
369 }
370 
rtas_setup_msi_irqs(struct pci_dev * pdev,int nvec_in,int type)371 static int rtas_setup_msi_irqs(struct pci_dev *pdev, int nvec_in, int type)
372 {
373 	struct pci_dn *pdn;
374 	int hwirq, virq, i, quota, rc;
375 	struct msi_desc *entry;
376 	struct msi_msg msg;
377 	int nvec = nvec_in;
378 	int use_32bit_msi_hack = 0;
379 
380 	if (type == PCI_CAP_ID_MSIX)
381 		rc = check_req_msix(pdev, nvec);
382 	else
383 		rc = check_req_msi(pdev, nvec);
384 
385 	if (rc)
386 		return rc;
387 
388 	quota = msi_quota_for_device(pdev, nvec);
389 
390 	if (quota && quota < nvec)
391 		return quota;
392 
393 	if (type == PCI_CAP_ID_MSIX && check_msix_entries(pdev))
394 		return -EINVAL;
395 
396 	/*
397 	 * Firmware currently refuse any non power of two allocation
398 	 * so we round up if the quota will allow it.
399 	 */
400 	if (type == PCI_CAP_ID_MSIX) {
401 		int m = roundup_pow_of_two(nvec);
402 		quota = msi_quota_for_device(pdev, m);
403 
404 		if (quota >= m)
405 			nvec = m;
406 	}
407 
408 	pdn = pci_get_pdn(pdev);
409 
410 	/*
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
413 	 * return MSI-Xs.
414 	 */
415 again:
416 	if (type == PCI_CAP_ID_MSI) {
417 		if (pdev->no_64bit_msi) {
418 			rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
419 			if (rc < 0) {
420 				/*
421 				 * We only want to run the 32 bit MSI hack below if
422 				 * the max bus speed is Gen2 speed
423 				 */
424 				if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
425 					return rc;
426 
427 				use_32bit_msi_hack = 1;
428 			}
429 		} else
430 			rc = -1;
431 
432 		if (rc < 0)
433 			rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
434 
435 		if (rc < 0) {
436 			pr_debug("rtas_msi: trying the old firmware call.\n");
437 			rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
438 		}
439 
440 		if (use_32bit_msi_hack && rc > 0)
441 			rtas_hack_32bit_msi_gen2(pdev);
442 	} else
443 		rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
444 
445 	if (rc != nvec) {
446 		if (nvec != nvec_in) {
447 			nvec = nvec_in;
448 			goto again;
449 		}
450 		pr_debug("rtas_msi: rtas_change_msi() failed\n");
451 		return rc;
452 	}
453 
454 	i = 0;
455 	for_each_pci_msi_entry(entry, pdev) {
456 		hwirq = rtas_query_irq_number(pdn, i++);
457 		if (hwirq < 0) {
458 			pr_debug("rtas_msi: error (%d) getting hwirq\n", rc);
459 			return hwirq;
460 		}
461 
462 		/*
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.
473 		 *
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.
478 		 */
479 		if (is_kdump_kernel())
480 			virq = irq_create_mapping(NULL, hwirq);
481 		else
482 			virq = irq_create_mapping_affinity(NULL, hwirq,
483 							   entry->affinity);
484 
485 		if (!virq) {
486 			pr_debug("rtas_msi: Failed mapping hwirq %d\n", hwirq);
487 			return -ENOSPC;
488 		}
489 
490 		dev_dbg(&pdev->dev, "rtas_msi: allocated virq %d\n", virq);
491 		irq_set_msi_desc(virq, entry);
492 
493 		/* Read config space back so we can restore after reset */
494 		__pci_read_msi_msg(entry, &msg);
495 		entry->msg = msg;
496 	}
497 
498 	return 0;
499 }
500 
rtas_msi_pci_irq_fixup(struct pci_dev * pdev)501 static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
502 {
503 	/* No LSI -> leave MSIs (if any) configured */
504 	if (!pdev->irq) {
505 		dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
506 		return;
507 	}
508 
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");
512 		return;
513 	}
514 
515 	dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
516 	rtas_disable_msi(pdev);
517 }
518 
rtas_msi_init(void)519 static int rtas_msi_init(void)
520 {
521 	struct pci_controller *phb;
522 
523 	query_token  = rtas_token("ibm,query-interrupt-source-number");
524 	change_token = rtas_token("ibm,change-msi");
525 
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");
529 		return -1;
530 	}
531 
532 	pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
533 
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;
537 
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;
542 	}
543 
544 	WARN_ON(ppc_md.pci_irq_fixup);
545 	ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
546 
547 	return 0;
548 }
549 machine_arch_initcall(pseries, rtas_msi_init);
550