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 #include <asm/machdep.h>
20 
21 #include "pseries.h"
22 
23 static int query_token, change_token;
24 
25 #define RTAS_QUERY_FN		0
26 #define RTAS_CHANGE_FN		1
27 #define RTAS_RESET_FN		2
28 #define RTAS_CHANGE_MSI_FN	3
29 #define RTAS_CHANGE_MSIX_FN	4
30 #define RTAS_CHANGE_32MSI_FN	5
31 
32 /* RTAS Helpers */
33 
34 static int rtas_change_msi(struct pci_dn *pdn, u32 func, u32 num_irqs)
35 {
36 	u32 addr, seq_num, rtas_ret[3];
37 	unsigned long buid;
38 	int rc;
39 
40 	addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
41 	buid = pdn->phb->buid;
42 
43 	seq_num = 1;
44 	do {
45 		if (func == RTAS_CHANGE_MSI_FN || func == RTAS_CHANGE_MSIX_FN ||
46 		    func == RTAS_CHANGE_32MSI_FN)
47 			rc = rtas_call(change_token, 6, 4, rtas_ret, addr,
48 					BUID_HI(buid), BUID_LO(buid),
49 					func, num_irqs, seq_num);
50 		else
51 			rc = rtas_call(change_token, 6, 3, rtas_ret, addr,
52 					BUID_HI(buid), BUID_LO(buid),
53 					func, num_irqs, seq_num);
54 
55 		seq_num = rtas_ret[1];
56 	} while (rtas_busy_delay(rc));
57 
58 	/*
59 	 * If the RTAS call succeeded, return the number of irqs allocated.
60 	 * If not, make sure we return a negative error code.
61 	 */
62 	if (rc == 0)
63 		rc = rtas_ret[0];
64 	else if (rc > 0)
65 		rc = -rc;
66 
67 	pr_debug("rtas_msi: ibm,change_msi(func=%d,num=%d), got %d rc = %d\n",
68 		 func, num_irqs, rtas_ret[0], rc);
69 
70 	return rc;
71 }
72 
73 static void rtas_disable_msi(struct pci_dev *pdev)
74 {
75 	struct pci_dn *pdn;
76 
77 	pdn = pci_get_pdn(pdev);
78 	if (!pdn)
79 		return;
80 
81 	/*
82 	 * disabling MSI with the explicit interface also disables MSI-X
83 	 */
84 	if (rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, 0) != 0) {
85 		/*
86 		 * may have failed because explicit interface is not
87 		 * present
88 		 */
89 		if (rtas_change_msi(pdn, RTAS_CHANGE_FN, 0) != 0) {
90 			pr_debug("rtas_msi: Setting MSIs to 0 failed!\n");
91 		}
92 	}
93 }
94 
95 static int rtas_query_irq_number(struct pci_dn *pdn, int offset)
96 {
97 	u32 addr, rtas_ret[2];
98 	unsigned long buid;
99 	int rc;
100 
101 	addr = rtas_config_addr(pdn->busno, pdn->devfn, 0);
102 	buid = pdn->phb->buid;
103 
104 	do {
105 		rc = rtas_call(query_token, 4, 3, rtas_ret, addr,
106 			       BUID_HI(buid), BUID_LO(buid), offset);
107 	} while (rtas_busy_delay(rc));
108 
109 	if (rc) {
110 		pr_debug("rtas_msi: error (%d) querying source number\n", rc);
111 		return rc;
112 	}
113 
114 	return rtas_ret[0];
115 }
116 
117 static void rtas_teardown_msi_irqs(struct pci_dev *pdev)
118 {
119 	struct msi_desc *entry;
120 
121 	for_each_pci_msi_entry(entry, pdev) {
122 		if (!entry->irq)
123 			continue;
124 
125 		irq_set_msi_desc(entry->irq, NULL);
126 		irq_dispose_mapping(entry->irq);
127 	}
128 
129 	rtas_disable_msi(pdev);
130 }
131 
132 static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
133 {
134 	struct device_node *dn;
135 	const __be32 *p;
136 	u32 req_msi;
137 
138 	dn = pci_device_to_OF_node(pdev);
139 
140 	p = of_get_property(dn, prop_name, NULL);
141 	if (!p) {
142 		pr_debug("rtas_msi: No %s on %pOF\n", prop_name, dn);
143 		return -ENOENT;
144 	}
145 
146 	req_msi = be32_to_cpup(p);
147 	if (req_msi < nvec) {
148 		pr_debug("rtas_msi: %s requests < %d MSIs\n", prop_name, nvec);
149 
150 		if (req_msi == 0) /* Be paranoid */
151 			return -ENOSPC;
152 
153 		return req_msi;
154 	}
155 
156 	return 0;
157 }
158 
159 static int check_req_msi(struct pci_dev *pdev, int nvec)
160 {
161 	return check_req(pdev, nvec, "ibm,req#msi");
162 }
163 
164 static int check_req_msix(struct pci_dev *pdev, int nvec)
165 {
166 	return check_req(pdev, nvec, "ibm,req#msi-x");
167 }
168 
169 /* Quota calculation */
170 
171 static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
172 {
173 	struct device_node *dn;
174 	const __be32 *p;
175 
176 	dn = of_node_get(pci_device_to_OF_node(dev));
177 	while (dn) {
178 		p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
179 		if (p) {
180 			pr_debug("rtas_msi: found prop on dn %pOF\n",
181 				dn);
182 			*total = be32_to_cpup(p);
183 			return dn;
184 		}
185 
186 		dn = of_get_next_parent(dn);
187 	}
188 
189 	return NULL;
190 }
191 
192 static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
193 {
194 	struct device_node *dn;
195 	struct eeh_dev *edev;
196 
197 	/* Found our PE and assume 8 at that point. */
198 
199 	dn = pci_device_to_OF_node(dev);
200 	if (!dn)
201 		return NULL;
202 
203 	/* Get the top level device in the PE */
204 	edev = pdn_to_eeh_dev(PCI_DN(dn));
205 	if (edev->pe)
206 		edev = list_first_entry(&edev->pe->edevs, struct eeh_dev,
207 					entry);
208 	dn = pci_device_to_OF_node(edev->pdev);
209 	if (!dn)
210 		return NULL;
211 
212 	/* We actually want the parent */
213 	dn = of_get_parent(dn);
214 	if (!dn)
215 		return NULL;
216 
217 	/* Hardcode of 8 for old firmwares */
218 	*total = 8;
219 	pr_debug("rtas_msi: using PE dn %pOF\n", dn);
220 
221 	return dn;
222 }
223 
224 struct msi_counts {
225 	struct device_node *requestor;
226 	int num_devices;
227 	int request;
228 	int quota;
229 	int spare;
230 	int over_quota;
231 };
232 
233 static void *count_non_bridge_devices(struct device_node *dn, void *data)
234 {
235 	struct msi_counts *counts = data;
236 	const __be32 *p;
237 	u32 class;
238 
239 	pr_debug("rtas_msi: counting %pOF\n", dn);
240 
241 	p = of_get_property(dn, "class-code", NULL);
242 	class = p ? be32_to_cpup(p) : 0;
243 
244 	if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
245 		counts->num_devices++;
246 
247 	return NULL;
248 }
249 
250 static void *count_spare_msis(struct device_node *dn, void *data)
251 {
252 	struct msi_counts *counts = data;
253 	const __be32 *p;
254 	int req;
255 
256 	if (dn == counts->requestor)
257 		req = counts->request;
258 	else {
259 		/* We don't know if a driver will try to use MSI or MSI-X,
260 		 * so we just have to punt and use the larger of the two. */
261 		req = 0;
262 		p = of_get_property(dn, "ibm,req#msi", NULL);
263 		if (p)
264 			req = be32_to_cpup(p);
265 
266 		p = of_get_property(dn, "ibm,req#msi-x", NULL);
267 		if (p)
268 			req = max(req, (int)be32_to_cpup(p));
269 	}
270 
271 	if (req < counts->quota)
272 		counts->spare += counts->quota - req;
273 	else if (req > counts->quota)
274 		counts->over_quota++;
275 
276 	return NULL;
277 }
278 
279 static int msi_quota_for_device(struct pci_dev *dev, int request)
280 {
281 	struct device_node *pe_dn;
282 	struct msi_counts counts;
283 	int total;
284 
285 	pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
286 		  request);
287 
288 	pe_dn = find_pe_total_msi(dev, &total);
289 	if (!pe_dn)
290 		pe_dn = find_pe_dn(dev, &total);
291 
292 	if (!pe_dn) {
293 		pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
294 		goto out;
295 	}
296 
297 	pr_debug("rtas_msi: found PE %pOF\n", pe_dn);
298 
299 	memset(&counts, 0, sizeof(struct msi_counts));
300 
301 	/* Work out how many devices we have below this PE */
302 	pci_traverse_device_nodes(pe_dn, count_non_bridge_devices, &counts);
303 
304 	if (counts.num_devices == 0) {
305 		pr_err("rtas_msi: found 0 devices under PE for %s\n",
306 			pci_name(dev));
307 		goto out;
308 	}
309 
310 	counts.quota = total / counts.num_devices;
311 	if (request <= counts.quota)
312 		goto out;
313 
314 	/* else, we have some more calculating to do */
315 	counts.requestor = pci_device_to_OF_node(dev);
316 	counts.request = request;
317 	pci_traverse_device_nodes(pe_dn, count_spare_msis, &counts);
318 
319 	/* If the quota isn't an integer multiple of the total, we can
320 	 * use the remainder as spare MSIs for anyone that wants them. */
321 	counts.spare += total % counts.num_devices;
322 
323 	/* Divide any spare by the number of over-quota requestors */
324 	if (counts.over_quota)
325 		counts.quota += counts.spare / counts.over_quota;
326 
327 	/* And finally clamp the request to the possibly adjusted quota */
328 	request = min(counts.quota, request);
329 
330 	pr_debug("rtas_msi: request clamped to quota %d\n", request);
331 out:
332 	of_node_put(pe_dn);
333 
334 	return request;
335 }
336 
337 static int check_msix_entries(struct pci_dev *pdev)
338 {
339 	struct msi_desc *entry;
340 	int expected;
341 
342 	/* There's no way for us to express to firmware that we want
343 	 * a discontiguous, or non-zero based, range of MSI-X entries.
344 	 * So we must reject such requests. */
345 
346 	expected = 0;
347 	for_each_pci_msi_entry(entry, pdev) {
348 		if (entry->msi_attrib.entry_nr != expected) {
349 			pr_debug("rtas_msi: bad MSI-X entries.\n");
350 			return -EINVAL;
351 		}
352 		expected++;
353 	}
354 
355 	return 0;
356 }
357 
358 static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
359 {
360 	u32 addr_hi, addr_lo;
361 
362 	/*
363 	 * We should only get in here for IODA1 configs. This is based on the
364 	 * fact that we using RTAS for MSIs, we don't have the 32 bit MSI RTAS
365 	 * support, and we are in a PCIe Gen2 slot.
366 	 */
367 	dev_info(&pdev->dev,
368 		 "rtas_msi: No 32 bit MSI firmware support, forcing 32 bit MSI\n");
369 	pci_read_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, &addr_hi);
370 	addr_lo = 0xffff0000 | ((addr_hi >> (48 - 32)) << 4);
371 	pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_LO, addr_lo);
372 	pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, 0);
373 }
374 
375 static int rtas_setup_msi_irqs(struct pci_dev *pdev, int nvec_in, int type)
376 {
377 	struct pci_dn *pdn;
378 	int hwirq, virq, i, quota, rc;
379 	struct msi_desc *entry;
380 	struct msi_msg msg;
381 	int nvec = nvec_in;
382 	int use_32bit_msi_hack = 0;
383 
384 	if (type == PCI_CAP_ID_MSIX)
385 		rc = check_req_msix(pdev, nvec);
386 	else
387 		rc = check_req_msi(pdev, nvec);
388 
389 	if (rc)
390 		return rc;
391 
392 	quota = msi_quota_for_device(pdev, nvec);
393 
394 	if (quota && quota < nvec)
395 		return quota;
396 
397 	if (type == PCI_CAP_ID_MSIX && check_msix_entries(pdev))
398 		return -EINVAL;
399 
400 	/*
401 	 * Firmware currently refuse any non power of two allocation
402 	 * so we round up if the quota will allow it.
403 	 */
404 	if (type == PCI_CAP_ID_MSIX) {
405 		int m = roundup_pow_of_two(nvec);
406 		quota = msi_quota_for_device(pdev, m);
407 
408 		if (quota >= m)
409 			nvec = m;
410 	}
411 
412 	pdn = pci_get_pdn(pdev);
413 
414 	/*
415 	 * Try the new more explicit firmware interface, if that fails fall
416 	 * back to the old interface. The old interface is known to never
417 	 * return MSI-Xs.
418 	 */
419 again:
420 	if (type == PCI_CAP_ID_MSI) {
421 		if (pdev->no_64bit_msi) {
422 			rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
423 			if (rc < 0) {
424 				/*
425 				 * We only want to run the 32 bit MSI hack below if
426 				 * the max bus speed is Gen2 speed
427 				 */
428 				if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
429 					return rc;
430 
431 				use_32bit_msi_hack = 1;
432 			}
433 		} else
434 			rc = -1;
435 
436 		if (rc < 0)
437 			rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
438 
439 		if (rc < 0) {
440 			pr_debug("rtas_msi: trying the old firmware call.\n");
441 			rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
442 		}
443 
444 		if (use_32bit_msi_hack && rc > 0)
445 			rtas_hack_32bit_msi_gen2(pdev);
446 	} else
447 		rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
448 
449 	if (rc != nvec) {
450 		if (nvec != nvec_in) {
451 			nvec = nvec_in;
452 			goto again;
453 		}
454 		pr_debug("rtas_msi: rtas_change_msi() failed\n");
455 		return rc;
456 	}
457 
458 	i = 0;
459 	for_each_pci_msi_entry(entry, pdev) {
460 		hwirq = rtas_query_irq_number(pdn, i++);
461 		if (hwirq < 0) {
462 			pr_debug("rtas_msi: error (%d) getting hwirq\n", rc);
463 			return hwirq;
464 		}
465 
466 		virq = irq_create_mapping(NULL, hwirq);
467 
468 		if (!virq) {
469 			pr_debug("rtas_msi: Failed mapping hwirq %d\n", hwirq);
470 			return -ENOSPC;
471 		}
472 
473 		dev_dbg(&pdev->dev, "rtas_msi: allocated virq %d\n", virq);
474 		irq_set_msi_desc(virq, entry);
475 
476 		/* Read config space back so we can restore after reset */
477 		__pci_read_msi_msg(entry, &msg);
478 		entry->msg = msg;
479 	}
480 
481 	return 0;
482 }
483 
484 static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
485 {
486 	/* No LSI -> leave MSIs (if any) configured */
487 	if (!pdev->irq) {
488 		dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
489 		return;
490 	}
491 
492 	/* No MSI -> MSIs can't have been assigned by fw, leave LSI */
493 	if (check_req_msi(pdev, 1) && check_req_msix(pdev, 1)) {
494 		dev_dbg(&pdev->dev, "rtas_msi: no req#msi/x, nothing to do.\n");
495 		return;
496 	}
497 
498 	dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
499 	rtas_disable_msi(pdev);
500 }
501 
502 static int rtas_msi_init(void)
503 {
504 	struct pci_controller *phb;
505 
506 	query_token  = rtas_token("ibm,query-interrupt-source-number");
507 	change_token = rtas_token("ibm,change-msi");
508 
509 	if ((query_token == RTAS_UNKNOWN_SERVICE) ||
510 			(change_token == RTAS_UNKNOWN_SERVICE)) {
511 		pr_debug("rtas_msi: no RTAS tokens, no MSI support.\n");
512 		return -1;
513 	}
514 
515 	pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
516 
517 	WARN_ON(pseries_pci_controller_ops.setup_msi_irqs);
518 	pseries_pci_controller_ops.setup_msi_irqs = rtas_setup_msi_irqs;
519 	pseries_pci_controller_ops.teardown_msi_irqs = rtas_teardown_msi_irqs;
520 
521 	list_for_each_entry(phb, &hose_list, list_node) {
522 		WARN_ON(phb->controller_ops.setup_msi_irqs);
523 		phb->controller_ops.setup_msi_irqs = rtas_setup_msi_irqs;
524 		phb->controller_ops.teardown_msi_irqs = rtas_teardown_msi_irqs;
525 	}
526 
527 	WARN_ON(ppc_md.pci_irq_fixup);
528 	ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
529 
530 	return 0;
531 }
532 machine_arch_initcall(pseries, rtas_msi_init);
533