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
23static 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
34static 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
73static 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
95static 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
117static 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 == NO_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
132static int check_req(struct pci_dev *pdev, int nvec, char *prop_name)
133{
134	struct device_node *dn;
135	struct pci_dn *pdn;
136	const __be32 *p;
137	u32 req_msi;
138
139	pdn = pci_get_pdn(pdev);
140	if (!pdn)
141		return -ENODEV;
142
143	dn = pdn->node;
144
145	p = of_get_property(dn, prop_name, NULL);
146	if (!p) {
147		pr_debug("rtas_msi: No %s on %s\n", prop_name, dn->full_name);
148		return -ENOENT;
149	}
150
151	req_msi = be32_to_cpup(p);
152	if (req_msi < nvec) {
153		pr_debug("rtas_msi: %s requests < %d MSIs\n", prop_name, nvec);
154
155		if (req_msi == 0) /* Be paranoid */
156			return -ENOSPC;
157
158		return req_msi;
159	}
160
161	return 0;
162}
163
164static int check_req_msi(struct pci_dev *pdev, int nvec)
165{
166	return check_req(pdev, nvec, "ibm,req#msi");
167}
168
169static int check_req_msix(struct pci_dev *pdev, int nvec)
170{
171	return check_req(pdev, nvec, "ibm,req#msi-x");
172}
173
174/* Quota calculation */
175
176static struct device_node *find_pe_total_msi(struct pci_dev *dev, int *total)
177{
178	struct device_node *dn;
179	const __be32 *p;
180
181	dn = of_node_get(pci_device_to_OF_node(dev));
182	while (dn) {
183		p = of_get_property(dn, "ibm,pe-total-#msi", NULL);
184		if (p) {
185			pr_debug("rtas_msi: found prop on dn %s\n",
186				dn->full_name);
187			*total = be32_to_cpup(p);
188			return dn;
189		}
190
191		dn = of_get_next_parent(dn);
192	}
193
194	return NULL;
195}
196
197static struct device_node *find_pe_dn(struct pci_dev *dev, int *total)
198{
199	struct device_node *dn;
200	struct pci_dn *pdn;
201	struct eeh_dev *edev;
202
203	/* Found our PE and assume 8 at that point. */
204
205	dn = pci_device_to_OF_node(dev);
206	if (!dn)
207		return NULL;
208
209	/* Get the top level device in the PE */
210	edev = pdn_to_eeh_dev(PCI_DN(dn));
211	if (edev->pe)
212		edev = list_first_entry(&edev->pe->edevs, struct eeh_dev, list);
213	pdn = eeh_dev_to_pdn(edev);
214	dn = pdn ? pdn->node : NULL;
215	if (!dn)
216		return NULL;
217
218	/* We actually want the parent */
219	dn = of_get_parent(dn);
220	if (!dn)
221		return NULL;
222
223	/* Hardcode of 8 for old firmwares */
224	*total = 8;
225	pr_debug("rtas_msi: using PE dn %s\n", dn->full_name);
226
227	return dn;
228}
229
230struct msi_counts {
231	struct device_node *requestor;
232	int num_devices;
233	int request;
234	int quota;
235	int spare;
236	int over_quota;
237};
238
239static void *count_non_bridge_devices(struct device_node *dn, void *data)
240{
241	struct msi_counts *counts = data;
242	const __be32 *p;
243	u32 class;
244
245	pr_debug("rtas_msi: counting %s\n", dn->full_name);
246
247	p = of_get_property(dn, "class-code", NULL);
248	class = p ? be32_to_cpup(p) : 0;
249
250	if ((class >> 8) != PCI_CLASS_BRIDGE_PCI)
251		counts->num_devices++;
252
253	return NULL;
254}
255
256static void *count_spare_msis(struct device_node *dn, void *data)
257{
258	struct msi_counts *counts = data;
259	const __be32 *p;
260	int req;
261
262	if (dn == counts->requestor)
263		req = counts->request;
264	else {
265		/* We don't know if a driver will try to use MSI or MSI-X,
266		 * so we just have to punt and use the larger of the two. */
267		req = 0;
268		p = of_get_property(dn, "ibm,req#msi", NULL);
269		if (p)
270			req = be32_to_cpup(p);
271
272		p = of_get_property(dn, "ibm,req#msi-x", NULL);
273		if (p)
274			req = max(req, (int)be32_to_cpup(p));
275	}
276
277	if (req < counts->quota)
278		counts->spare += counts->quota - req;
279	else if (req > counts->quota)
280		counts->over_quota++;
281
282	return NULL;
283}
284
285static int msi_quota_for_device(struct pci_dev *dev, int request)
286{
287	struct device_node *pe_dn;
288	struct msi_counts counts;
289	int total;
290
291	pr_debug("rtas_msi: calc quota for %s, request %d\n", pci_name(dev),
292		  request);
293
294	pe_dn = find_pe_total_msi(dev, &total);
295	if (!pe_dn)
296		pe_dn = find_pe_dn(dev, &total);
297
298	if (!pe_dn) {
299		pr_err("rtas_msi: couldn't find PE for %s\n", pci_name(dev));
300		goto out;
301	}
302
303	pr_debug("rtas_msi: found PE %s\n", pe_dn->full_name);
304
305	memset(&counts, 0, sizeof(struct msi_counts));
306
307	/* Work out how many devices we have below this PE */
308	traverse_pci_devices(pe_dn, count_non_bridge_devices, &counts);
309
310	if (counts.num_devices == 0) {
311		pr_err("rtas_msi: found 0 devices under PE for %s\n",
312			pci_name(dev));
313		goto out;
314	}
315
316	counts.quota = total / counts.num_devices;
317	if (request <= counts.quota)
318		goto out;
319
320	/* else, we have some more calculating to do */
321	counts.requestor = pci_device_to_OF_node(dev);
322	counts.request = request;
323	traverse_pci_devices(pe_dn, count_spare_msis, &counts);
324
325	/* If the quota isn't an integer multiple of the total, we can
326	 * use the remainder as spare MSIs for anyone that wants them. */
327	counts.spare += total % counts.num_devices;
328
329	/* Divide any spare by the number of over-quota requestors */
330	if (counts.over_quota)
331		counts.quota += counts.spare / counts.over_quota;
332
333	/* And finally clamp the request to the possibly adjusted quota */
334	request = min(counts.quota, request);
335
336	pr_debug("rtas_msi: request clamped to quota %d\n", request);
337out:
338	of_node_put(pe_dn);
339
340	return request;
341}
342
343static int check_msix_entries(struct pci_dev *pdev)
344{
345	struct msi_desc *entry;
346	int expected;
347
348	/* There's no way for us to express to firmware that we want
349	 * a discontiguous, or non-zero based, range of MSI-X entries.
350	 * So we must reject such requests. */
351
352	expected = 0;
353	for_each_pci_msi_entry(entry, pdev) {
354		if (entry->msi_attrib.entry_nr != expected) {
355			pr_debug("rtas_msi: bad MSI-X entries.\n");
356			return -EINVAL;
357		}
358		expected++;
359	}
360
361	return 0;
362}
363
364static void rtas_hack_32bit_msi_gen2(struct pci_dev *pdev)
365{
366	u32 addr_hi, addr_lo;
367
368	/*
369	 * We should only get in here for IODA1 configs. This is based on the
370	 * fact that we using RTAS for MSIs, we don't have the 32 bit MSI RTAS
371	 * support, and we are in a PCIe Gen2 slot.
372	 */
373	dev_info(&pdev->dev,
374		 "rtas_msi: No 32 bit MSI firmware support, forcing 32 bit MSI\n");
375	pci_read_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, &addr_hi);
376	addr_lo = 0xffff0000 | ((addr_hi >> (48 - 32)) << 4);
377	pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_LO, addr_lo);
378	pci_write_config_dword(pdev, pdev->msi_cap + PCI_MSI_ADDRESS_HI, 0);
379}
380
381static int rtas_setup_msi_irqs(struct pci_dev *pdev, int nvec_in, int type)
382{
383	struct pci_dn *pdn;
384	int hwirq, virq, i, quota, rc;
385	struct msi_desc *entry;
386	struct msi_msg msg;
387	int nvec = nvec_in;
388	int use_32bit_msi_hack = 0;
389
390	if (type == PCI_CAP_ID_MSIX)
391		rc = check_req_msix(pdev, nvec);
392	else
393		rc = check_req_msi(pdev, nvec);
394
395	if (rc)
396		return rc;
397
398	quota = msi_quota_for_device(pdev, nvec);
399
400	if (quota && quota < nvec)
401		return quota;
402
403	if (type == PCI_CAP_ID_MSIX && check_msix_entries(pdev))
404		return -EINVAL;
405
406	/*
407	 * Firmware currently refuse any non power of two allocation
408	 * so we round up if the quota will allow it.
409	 */
410	if (type == PCI_CAP_ID_MSIX) {
411		int m = roundup_pow_of_two(nvec);
412		quota = msi_quota_for_device(pdev, m);
413
414		if (quota >= m)
415			nvec = m;
416	}
417
418	pdn = pci_get_pdn(pdev);
419
420	/*
421	 * Try the new more explicit firmware interface, if that fails fall
422	 * back to the old interface. The old interface is known to never
423	 * return MSI-Xs.
424	 */
425again:
426	if (type == PCI_CAP_ID_MSI) {
427		if (pdev->no_64bit_msi) {
428			rc = rtas_change_msi(pdn, RTAS_CHANGE_32MSI_FN, nvec);
429			if (rc < 0) {
430				/*
431				 * We only want to run the 32 bit MSI hack below if
432				 * the max bus speed is Gen2 speed
433				 */
434				if (pdev->bus->max_bus_speed != PCIE_SPEED_5_0GT)
435					return rc;
436
437				use_32bit_msi_hack = 1;
438			}
439		} else
440			rc = -1;
441
442		if (rc < 0)
443			rc = rtas_change_msi(pdn, RTAS_CHANGE_MSI_FN, nvec);
444
445		if (rc < 0) {
446			pr_debug("rtas_msi: trying the old firmware call.\n");
447			rc = rtas_change_msi(pdn, RTAS_CHANGE_FN, nvec);
448		}
449
450		if (use_32bit_msi_hack && rc > 0)
451			rtas_hack_32bit_msi_gen2(pdev);
452	} else
453		rc = rtas_change_msi(pdn, RTAS_CHANGE_MSIX_FN, nvec);
454
455	if (rc != nvec) {
456		if (nvec != nvec_in) {
457			nvec = nvec_in;
458			goto again;
459		}
460		pr_debug("rtas_msi: rtas_change_msi() failed\n");
461		return rc;
462	}
463
464	i = 0;
465	for_each_pci_msi_entry(entry, pdev) {
466		hwirq = rtas_query_irq_number(pdn, i++);
467		if (hwirq < 0) {
468			pr_debug("rtas_msi: error (%d) getting hwirq\n", rc);
469			return hwirq;
470		}
471
472		virq = irq_create_mapping(NULL, hwirq);
473
474		if (virq == NO_IRQ) {
475			pr_debug("rtas_msi: Failed mapping hwirq %d\n", hwirq);
476			return -ENOSPC;
477		}
478
479		dev_dbg(&pdev->dev, "rtas_msi: allocated virq %d\n", virq);
480		irq_set_msi_desc(virq, entry);
481
482		/* Read config space back so we can restore after reset */
483		__pci_read_msi_msg(entry, &msg);
484		entry->msg = msg;
485	}
486
487	return 0;
488}
489
490static void rtas_msi_pci_irq_fixup(struct pci_dev *pdev)
491{
492	/* No LSI -> leave MSIs (if any) configured */
493	if (pdev->irq == NO_IRQ) {
494		dev_dbg(&pdev->dev, "rtas_msi: no LSI, nothing to do.\n");
495		return;
496	}
497
498	/* No MSI -> MSIs can't have been assigned by fw, leave LSI */
499	if (check_req_msi(pdev, 1) && check_req_msix(pdev, 1)) {
500		dev_dbg(&pdev->dev, "rtas_msi: no req#msi/x, nothing to do.\n");
501		return;
502	}
503
504	dev_dbg(&pdev->dev, "rtas_msi: disabling existing MSI.\n");
505	rtas_disable_msi(pdev);
506}
507
508static int rtas_msi_init(void)
509{
510	struct pci_controller *phb;
511
512	query_token  = rtas_token("ibm,query-interrupt-source-number");
513	change_token = rtas_token("ibm,change-msi");
514
515	if ((query_token == RTAS_UNKNOWN_SERVICE) ||
516			(change_token == RTAS_UNKNOWN_SERVICE)) {
517		pr_debug("rtas_msi: no RTAS tokens, no MSI support.\n");
518		return -1;
519	}
520
521	pr_debug("rtas_msi: Registering RTAS MSI callbacks.\n");
522
523	WARN_ON(pseries_pci_controller_ops.setup_msi_irqs);
524	pseries_pci_controller_ops.setup_msi_irqs = rtas_setup_msi_irqs;
525	pseries_pci_controller_ops.teardown_msi_irqs = rtas_teardown_msi_irqs;
526
527	list_for_each_entry(phb, &hose_list, list_node) {
528		WARN_ON(phb->controller_ops.setup_msi_irqs);
529		phb->controller_ops.setup_msi_irqs = rtas_setup_msi_irqs;
530		phb->controller_ops.teardown_msi_irqs = rtas_teardown_msi_irqs;
531	}
532
533	WARN_ON(ppc_md.pci_irq_fixup);
534	ppc_md.pci_irq_fixup = rtas_msi_pci_irq_fixup;
535
536	return 0;
537}
538machine_arch_initcall(pseries, rtas_msi_init);
539