xref: /linux/arch/powerpc/platforms/pseries/msi.c (revision d91517839e5d95adc0cf4b28caa7af62a71de526)
1 /*
2  * Copyright 2006 Jake Moilanen <moilanen@austin.ibm.com>, IBM Corp.
3  * Copyright 2006-2007 Michael Ellerman, IBM Corp.
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License
7  * as published by the Free Software Foundation; version 2 of the
8  * License.
9  *
10  */
11 
12 #include <linux/device.h>
13 #include <linux/irq.h>
14 #include <linux/msi.h>
15 
16 #include <asm/rtas.h>
17 #include <asm/hw_irq.h>
18 #include <asm/ppc-pci.h>
19 
20 static int query_token, change_token;
21 
22 #define RTAS_QUERY_FN		0
23 #define RTAS_CHANGE_FN		1
24 #define RTAS_RESET_FN		2
25 #define RTAS_CHANGE_MSI_FN	3
26 #define RTAS_CHANGE_MSIX_FN	4
27 #define RTAS_CHANGE_32MSI_FN	5
28 
29 /* RTAS Helpers */
30 
31 static int rtas_change_msi(struct pci_dn *pdn, u32 func, u32 num_irqs)
32 {
33 	u32 addr, seq_num, rtas_ret[3];
34 	unsigned long buid;
35 	int rc;
36 
37 	addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
38 	buid = pdn->phb->buid;
39 
40 	seq_num = 1;
41 	do {
42 		if (func == RTAS_CHANGE_MSI_FN || func == RTAS_CHANGE_MSIX_FN ||
43 		    func == RTAS_CHANGE_32MSI_FN)
44 			rc = rtas_call(change_token, 6, 4, rtas_ret, addr,
45 					BUID_HI(buid), BUID_LO(buid),
46 					func, num_irqs, seq_num);
47 		else
48 			rc = rtas_call(change_token, 6, 3, rtas_ret, addr,
49 					BUID_HI(buid), BUID_LO(buid),
50 					func, num_irqs, seq_num);
51 
52 		seq_num = rtas_ret[1];
53 	} while (rtas_busy_delay(rc));
54 
55 	/*
56 	 * If the RTAS call succeeded, return the number of irqs allocated.
57 	 * If not, make sure we return a negative error code.
58 	 */
59 	if (rc == 0)
60 		rc = rtas_ret[0];
61 	else if (rc > 0)
62 		rc = -rc;
63 
64 	pr_debug("rtas_msi: ibm,change_msi(func=%d,num=%d), got %d rc = %d\n",
65 		 func, num_irqs, rtas_ret[0], rc);
66 
67 	return rc;
68 }
69 
70 static void rtas_disable_msi(struct pci_dev *pdev)
71 {
72 	struct pci_dn *pdn;
73 
74 	pdn = pci_get_pdn(pdev);
75 	if (!pdn)
76 		return;
77 
78 	/*
79 	 * disabling MSI with the explicit interface also disables MSI-X
80 	 */
81 	if (rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, 0) != 0) {
82 		/*
83 		 * may have failed because explicit interface is not
84 		 * present
85 		 */
86 		if (rtas_change_msi(pdn, RTAS_CHANGE_FN, 0) != 0) {
87 			pr_debug("rtas_msi: Setting MSIs to 0 failed!\n");
88 		}
89 	}
90 }
91 
92 static int rtas_query_irq_number(struct pci_dn *pdn, int offset)
93 {
94 	u32 addr, rtas_ret[2];
95 	unsigned long buid;
96 	int rc;
97 
98 	addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
99 	buid = pdn->phb->buid;
100 
101 	do {
102 		rc = rtas_call(query_token, 4, 3, rtas_ret, addr,
103 			       BUID_HI(buid), BUID_LO(buid), offset);
104 	} while (rtas_busy_delay(rc));
105 
106 	if (rc) {
107 		pr_debug("rtas_msi: error (%d) querying source number\n", rc);
108 		return rc;
109 	}
110 
111 	return rtas_ret[0];
112 }
113 
114 static void rtas_teardown_msi_irqs(struct pci_dev *pdev)
115 {
116 	struct msi_desc *entry;
117 
118 	list_for_each_entry(entry, &pdev->msi_list, list) {
119 		if (entry->irq == NO_IRQ)
120 			continue;
121 
122 		irq_set_msi_desc(entry->irq, NULL);
123 		irq_dispose_mapping(entry->irq);
124 	}
125 
126 	rtas_disable_msi(pdev);
127 }
128 
129 static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
130 {
131 	struct device_node *dn;
132 	struct pci_dn *pdn;
133 	const __be32 *p;
134 	u32 req_msi;
135 
136 	pdn = pci_get_pdn(pdev);
137 	if (!pdn)
138 		return -ENODEV;
139 
140 	dn = pdn->node;
141 
142 	p = of_get_property(dn, prop_name, NULL);
143 	if (!p) {
144 		pr_debug("rtas_msi: No %s on %s\n", prop_name, dn->full_name);
145 		return -ENOENT;
146 	}
147 
148 	req_msi = be32_to_cpup(p);
149 	if (req_msi < nvec) {
150 		pr_debug("rtas_msi: %s requests < %d MSIs\n", prop_name, nvec);
151 
152 		if (req_msi == 0) /* Be paranoid */
153 			return -ENOSPC;
154 
155 		return req_msi;
156 	}
157 
158 	return 0;
159 }
160 
161 static int check_req_msi(struct pci_dev *pdev, int nvec)
162 {
163 	return check_req(pdev, nvec, "ibm,req#msi");
164 }
165 
166 static int check_req_msix(struct pci_dev *pdev, int nvec)
167 {
168 	return check_req(pdev, nvec, "ibm,req#msi-x");
169 }
170 
171 /* Quota calculation */
172 
173 static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
174 {
175 	struct device_node *dn;
176 	const __be32 *p;
177 
178 	dn = of_node_get(pci_device_to_OF_node(dev));
179 	while (dn) {
180 		p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
181 		if (p) {
182 			pr_debug("rtas_msi: found prop on dn %s\n",
183 				dn->full_name);
184 			*total = be32_to_cpup(p);
185 			return dn;
186 		}
187 
188 		dn = of_get_next_parent(dn);
189 	}
190 
191 	return NULL;
192 }
193 
194 static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
195 {
196 	struct device_node *dn;
197 	struct eeh_dev *edev;
198 
199 	/* Found our PE and assume 8 at that point. */
200 
201 	dn = pci_device_to_OF_node(dev);
202 	if (!dn)
203 		return NULL;
204 
205 	/* Get the top level device in the PE */
206 	edev = of_node_to_eeh_dev(dn);
207 	if (edev->pe)
208 		edev = list_first_entry(&edev->pe->edevs, struct eeh_dev, list);
209 	dn = eeh_dev_to_of_node(edev);
210 	if (!dn)
211 		return NULL;
212 
213 	/* We actually want the parent */
214 	dn = of_get_parent(dn);
215 	if (!dn)
216 		return NULL;
217 
218 	/* Hardcode of 8 for old firmwares */
219 	*total = 8;
220 	pr_debug("rtas_msi: using PE dn %s\n", dn->full_name);
221 
222 	return dn;
223 }
224 
225 struct msi_counts {
226 	struct device_node *requestor;
227 	int num_devices;
228 	int request;
229 	int quota;
230 	int spare;
231 	int over_quota;
232 };
233 
234 static void *count_non_bridge_devices(struct device_node *dn, void *data)
235 {
236 	struct msi_counts *counts = data;
237 	const __be32 *p;
238 	u32 class;
239 
240 	pr_debug("rtas_msi: counting %s\n", dn->full_name);
241 
242 	p = of_get_property(dn, "class-code", NULL);
243 	class = p ? be32_to_cpup(p) : 0;
244 
245 	if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
246 		counts->num_devices++;
247 
248 	return NULL;
249 }
250 
251 static void *count_spare_msis(struct device_node *dn, void *data)
252 {
253 	struct msi_counts *counts = data;
254 	const __be32 *p;
255 	int req;
256 
257 	if (dn == counts->requestor)
258 		req = counts->request;
259 	else {
260 		/* We don't know if a driver will try to use MSI or MSI-X,
261 		 * so we just have to punt and use the larger of the two. */
262 		req = 0;
263 		p = of_get_property(dn, "ibm,req#msi", NULL);
264 		if (p)
265 			req = be32_to_cpup(p);
266 
267 		p = of_get_property(dn, "ibm,req#msi-x", NULL);
268 		if (p)
269 			req = max(req, (int)be32_to_cpup(p));
270 	}
271 
272 	if (req < counts->quota)
273 		counts->spare += counts->quota - req;
274 	else if (req > counts->quota)
275 		counts->over_quota++;
276 
277 	return NULL;
278 }
279 
280 static int msi_quota_for_device(struct pci_dev *dev, int request)
281 {
282 	struct device_node *pe_dn;
283 	struct msi_counts counts;
284 	int total;
285 
286 	pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
287 		  request);
288 
289 	pe_dn = find_pe_total_msi(dev, &total);
290 	if (!pe_dn)
291 		pe_dn = find_pe_dn(dev, &total);
292 
293 	if (!pe_dn) {
294 		pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
295 		goto out;
296 	}
297 
298 	pr_debug("rtas_msi: found PE %s\n", pe_dn->full_name);
299 
300 	memset(&counts, 0, sizeof(struct msi_counts));
301 
302 	/* Work out how many devices we have below this PE */
303 	traverse_pci_devices(pe_dn, count_non_bridge_devices, &counts);
304 
305 	if (counts.num_devices == 0) {
306 		pr_err("rtas_msi: found 0 devices under PE for %s\n",
307 			pci_name(dev));
308 		goto out;
309 	}
310 
311 	counts.quota = total / counts.num_devices;
312 	if (request <= counts.quota)
313 		goto out;
314 
315 	/* else, we have some more calculating to do */
316 	counts.requestor = pci_device_to_OF_node(dev);
317 	counts.request = request;
318 	traverse_pci_devices(pe_dn, count_spare_msis, &counts);
319 
320 	/* If the quota isn't an integer multiple of the total, we can
321 	 * use the remainder as spare MSIs for anyone that wants them. */
322 	counts.spare += total % counts.num_devices;
323 
324 	/* Divide any spare by the number of over-quota requestors */
325 	if (counts.over_quota)
326 		counts.quota += counts.spare / counts.over_quota;
327 
328 	/* And finally clamp the request to the possibly adjusted quota */
329 	request = min(counts.quota, request);
330 
331 	pr_debug("rtas_msi: request clamped to quota %d\n", request);
332 out:
333 	of_node_put(pe_dn);
334 
335 	return request;
336 }
337 
338 static int rtas_msi_check_device(struct pci_dev *pdev, int nvec, int type)
339 {
340 	int quota, rc;
341 
342 	if (type == PCI_CAP_ID_MSIX)
343 		rc = check_req_msix(pdev, nvec);
344 	else
345 		rc = check_req_msi(pdev, nvec);
346 
347 	if (rc)
348 		return rc;
349 
350 	quota = msi_quota_for_device(pdev, nvec);
351 
352 	if (quota && quota < nvec)
353 		return quota;
354 
355 	return 0;
356 }
357 
358 static int check_msix_entries(struct pci_dev *pdev)
359 {
360 	struct msi_desc *entry;
361 	int expected;
362 
363 	/* There's no way for us to express to firmware that we want
364 	 * a discontiguous, or non-zero based, range of MSI-X entries.
365 	 * So we must reject such requests. */
366 
367 	expected = 0;
368 	list_for_each_entry(entry, &pdev->msi_list, list) {
369 		if (entry->msi_attrib.entry_nr != expected) {
370 			pr_debug("rtas_msi: bad MSI-X entries.\n");
371 			return -EINVAL;
372 		}
373 		expected++;
374 	}
375 
376 	return 0;
377 }
378 
379 static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
380 {
381 	u32 addr_hi, addr_lo;
382 
383 	/*
384 	 * We should only get in here for IODA1 configs. This is based on the
385 	 * fact that we using RTAS for MSIs, we don't have the 32 bit MSI RTAS
386 	 * support, and we are in a PCIe Gen2 slot.
387 	 */
388 	dev_info(&pdev->dev,
389 		 "rtas_msi: No 32 bit MSI firmware support, forcing 32 bit MSI\n");
390 	pci_read_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, &addr_hi);
391 	addr_lo = 0xffff0000 | ((addr_hi >> (48 - 32)) << 4);
392 	pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_LO, addr_lo);
393 	pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, 0);
394 }
395 
396 static int rtas_setup_msi_irqs(struct pci_dev *pdev, int nvec_in, int type)
397 {
398 	struct pci_dn *pdn;
399 	int hwirq, virq, i, rc;
400 	struct msi_desc *entry;
401 	struct msi_msg msg;
402 	int nvec = nvec_in;
403 	int use_32bit_msi_hack = 0;
404 
405 	pdn = pci_get_pdn(pdev);
406 	if (!pdn)
407 		return -ENODEV;
408 
409 	if (type == PCI_CAP_ID_MSIX && check_msix_entries(pdev))
410 		return -EINVAL;
411 
412 	/*
413 	 * Firmware currently refuse any non power of two allocation
414 	 * so we round up if the quota will allow it.
415 	 */
416 	if (type == PCI_CAP_ID_MSIX) {
417 		int m = roundup_pow_of_two(nvec);
418 		int quota = msi_quota_for_device(pdev, m);
419 
420 		if (quota >= m)
421 			nvec = m;
422 	}
423 
424 	/*
425 	 * Try the new more explicit firmware interface, if that fails fall
426 	 * back to the old interface. The old interface is known to never
427 	 * return MSI-Xs.
428 	 */
429 again:
430 	if (type == PCI_CAP_ID_MSI) {
431 		if (pdn->force_32bit_msi) {
432 			rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
433 			if (rc < 0) {
434 				/*
435 				 * We only want to run the 32 bit MSI hack below if
436 				 * the max bus speed is Gen2 speed
437 				 */
438 				if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
439 					return rc;
440 
441 				use_32bit_msi_hack = 1;
442 			}
443 		} else
444 			rc = -1;
445 
446 		if (rc < 0)
447 			rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
448 
449 		if (rc < 0) {
450 			pr_debug("rtas_msi: trying the old firmware call.\n");
451 			rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
452 		}
453 
454 		if (use_32bit_msi_hack && rc > 0)
455 			rtas_hack_32bit_msi_gen2(pdev);
456 	} else
457 		rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
458 
459 	if (rc != nvec) {
460 		if (nvec != nvec_in) {
461 			nvec = nvec_in;
462 			goto again;
463 		}
464 		pr_debug("rtas_msi: rtas_change_msi() failed\n");
465 		return rc;
466 	}
467 
468 	i = 0;
469 	list_for_each_entry(entry, &pdev->msi_list, list) {
470 		hwirq = rtas_query_irq_number(pdn, i++);
471 		if (hwirq < 0) {
472 			pr_debug("rtas_msi: error (%d) getting hwirq\n", rc);
473 			return hwirq;
474 		}
475 
476 		virq = irq_create_mapping(NULL, hwirq);
477 
478 		if (virq == NO_IRQ) {
479 			pr_debug("rtas_msi: Failed mapping hwirq %d\n", hwirq);
480 			return -ENOSPC;
481 		}
482 
483 		dev_dbg(&pdev->dev, "rtas_msi: allocated virq %d\n", virq);
484 		irq_set_msi_desc(virq, entry);
485 
486 		/* Read config space back so we can restore after reset */
487 		read_msi_msg(virq, &msg);
488 		entry->msg = msg;
489 	}
490 
491 	return 0;
492 }
493 
494 static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
495 {
496 	/* No LSI -> leave MSIs (if any) configured */
497 	if (pdev->irq == NO_IRQ) {
498 		dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
499 		return;
500 	}
501 
502 	/* No MSI -> MSIs can't have been assigned by fw, leave LSI */
503 	if (check_req_msi(pdev, 1) && check_req_msix(pdev, 1)) {
504 		dev_dbg(&pdev->dev, "rtas_msi: no req#msi/x, nothing to do.\n");
505 		return;
506 	}
507 
508 	dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
509 	rtas_disable_msi(pdev);
510 }
511 
512 static int rtas_msi_init(void)
513 {
514 	query_token  = rtas_token("ibm,query-interrupt-source-number");
515 	change_token = rtas_token("ibm,change-msi");
516 
517 	if ((query_token == RTAS_UNKNOWN_SERVICE) ||
518 			(change_token == RTAS_UNKNOWN_SERVICE)) {
519 		pr_debug("rtas_msi: no RTAS tokens, no MSI support.\n");
520 		return -1;
521 	}
522 
523 	pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
524 
525 	WARN_ON(ppc_md.setup_msi_irqs);
526 	ppc_md.setup_msi_irqs = rtas_setup_msi_irqs;
527 	ppc_md.teardown_msi_irqs = rtas_teardown_msi_irqs;
528 	ppc_md.msi_check_device = rtas_msi_check_device;
529 
530 	WARN_ON(ppc_md.pci_irq_fixup);
531 	ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
532 
533 	return 0;
534 }
535 arch_initcall(rtas_msi_init);
536 
537