1 /*
2  * Copyright (c) 2010 - 2017 Intel Corporation.  All rights reserved.
3  * Copyright (c) 2008, 2009 QLogic Corporation. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/pci.h>
35 #include <linux/io.h>
36 #include <linux/delay.h>
37 #include <linux/vmalloc.h>
38 #include <linux/aer.h>
39 #include <linux/module.h>
40 
41 #include "qib.h"
42 
43 /*
44  * This file contains PCIe utility routines that are common to the
45  * various QLogic InfiniPath adapters
46  */
47 
48 /*
49  * Code to adjust PCIe capabilities.
50  * To minimize the change footprint, we call it
51  * from qib_pcie_params, which every chip-specific
52  * file calls, even though this violates some
53  * expectations of harmlessness.
54  */
55 static void qib_tune_pcie_caps(struct qib_devdata *);
56 static void qib_tune_pcie_coalesce(struct qib_devdata *);
57 
58 /*
59  * Do all the common PCIe setup and initialization.
60  * devdata is not yet allocated, and is not allocated until after this
61  * routine returns success.  Therefore qib_dev_err() can't be used for error
62  * printing.
63  */
64 int qib_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
65 {
66 	int ret;
67 
68 	ret = pci_enable_device(pdev);
69 	if (ret) {
70 		/*
71 		 * This can happen (in theory) iff:
72 		 * We did a chip reset, and then failed to reprogram the
73 		 * BAR, or the chip reset due to an internal error.  We then
74 		 * unloaded the driver and reloaded it.
75 		 *
76 		 * Both reset cases set the BAR back to initial state.  For
77 		 * the latter case, the AER sticky error bit at offset 0x718
78 		 * should be set, but the Linux kernel doesn't yet know
79 		 * about that, it appears.  If the original BAR was retained
80 		 * in the kernel data structures, this may be OK.
81 		 */
82 		qib_early_err(&pdev->dev, "pci enable failed: error %d\n",
83 			      -ret);
84 		goto done;
85 	}
86 
87 	ret = pci_request_regions(pdev, QIB_DRV_NAME);
88 	if (ret) {
89 		qib_devinfo(pdev, "pci_request_regions fails: err %d\n", -ret);
90 		goto bail;
91 	}
92 
93 	ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
94 	if (ret) {
95 		/*
96 		 * If the 64 bit setup fails, try 32 bit.  Some systems
97 		 * do not setup 64 bit maps on systems with 2GB or less
98 		 * memory installed.
99 		 */
100 		ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
101 		if (ret) {
102 			qib_devinfo(pdev, "Unable to set DMA mask: %d\n", ret);
103 			goto bail;
104 		}
105 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
106 	} else
107 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
108 	if (ret) {
109 		qib_early_err(&pdev->dev,
110 			      "Unable to set DMA consistent mask: %d\n", ret);
111 		goto bail;
112 	}
113 
114 	pci_set_master(pdev);
115 	ret = pci_enable_pcie_error_reporting(pdev);
116 	if (ret) {
117 		qib_early_err(&pdev->dev,
118 			      "Unable to enable pcie error reporting: %d\n",
119 			      ret);
120 		ret = 0;
121 	}
122 	goto done;
123 
124 bail:
125 	pci_disable_device(pdev);
126 	pci_release_regions(pdev);
127 done:
128 	return ret;
129 }
130 
131 /*
132  * Do remaining PCIe setup, once dd is allocated, and save away
133  * fields required to re-initialize after a chip reset, or for
134  * various other purposes
135  */
136 int qib_pcie_ddinit(struct qib_devdata *dd, struct pci_dev *pdev,
137 		    const struct pci_device_id *ent)
138 {
139 	unsigned long len;
140 	resource_size_t addr;
141 
142 	dd->pcidev = pdev;
143 	pci_set_drvdata(pdev, dd);
144 
145 	addr = pci_resource_start(pdev, 0);
146 	len = pci_resource_len(pdev, 0);
147 
148 	dd->kregbase = ioremap(addr, len);
149 	if (!dd->kregbase)
150 		return -ENOMEM;
151 
152 	dd->kregend = (u64 __iomem *)((void __iomem *) dd->kregbase + len);
153 	dd->physaddr = addr;        /* used for io_remap, etc. */
154 
155 	/*
156 	 * Save BARs to rewrite after device reset.  Save all 64 bits of
157 	 * BAR, just in case.
158 	 */
159 	dd->pcibar0 = addr;
160 	dd->pcibar1 = addr >> 32;
161 	dd->deviceid = ent->device; /* save for later use */
162 	dd->vendorid = ent->vendor;
163 
164 	return 0;
165 }
166 
167 /*
168  * Do PCIe cleanup, after chip-specific cleanup, etc.  Just prior
169  * to releasing the dd memory.
170  * void because none of the core pcie cleanup returns are void
171  */
172 void qib_pcie_ddcleanup(struct qib_devdata *dd)
173 {
174 	u64 __iomem *base = (void __iomem *) dd->kregbase;
175 
176 	dd->kregbase = NULL;
177 	iounmap(base);
178 	if (dd->piobase)
179 		iounmap(dd->piobase);
180 	if (dd->userbase)
181 		iounmap(dd->userbase);
182 	if (dd->piovl15base)
183 		iounmap(dd->piovl15base);
184 
185 	pci_disable_device(dd->pcidev);
186 	pci_release_regions(dd->pcidev);
187 
188 	pci_set_drvdata(dd->pcidev, NULL);
189 }
190 
191 /**
192  * We save the msi lo and hi values, so we can restore them after
193  * chip reset (the kernel PCI infrastructure doesn't yet handle that
194  * correctly.
195  */
196 static void qib_cache_msi_info(struct qib_devdata *dd, int pos)
197 {
198 	struct pci_dev *pdev = dd->pcidev;
199 	u16 control;
200 
201 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO, &dd->msi_lo);
202 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI, &dd->msi_hi);
203 	pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
204 
205 	/* now save the data (vector) info */
206 	pci_read_config_word(pdev,
207 			     pos + ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
208 			     &dd->msi_data);
209 }
210 
211 int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent)
212 {
213 	u16 linkstat, speed;
214 	int nvec;
215 	int maxvec;
216 	unsigned int flags = PCI_IRQ_MSIX | PCI_IRQ_MSI;
217 
218 	if (!pci_is_pcie(dd->pcidev)) {
219 		qib_dev_err(dd, "Can't find PCI Express capability!\n");
220 		/* set up something... */
221 		dd->lbus_width = 1;
222 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
223 		nvec = -1;
224 		goto bail;
225 	}
226 
227 	if (dd->flags & QIB_HAS_INTX)
228 		flags |= PCI_IRQ_LEGACY;
229 	maxvec = (nent && *nent) ? *nent : 1;
230 	nvec = pci_alloc_irq_vectors(dd->pcidev, 1, maxvec, flags);
231 	if (nvec < 0)
232 		goto bail;
233 
234 	/*
235 	 * If nent exists, make sure to record how many vectors were allocated.
236 	 * If msix_enabled is false, return 0 so the fallback code works
237 	 * correctly.
238 	 */
239 	if (nent)
240 		*nent = !dd->pcidev->msix_enabled ? 0 : nvec;
241 
242 	if (dd->pcidev->msi_enabled)
243 		qib_cache_msi_info(dd, dd->pcidev->msi_cap);
244 
245 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
246 	/*
247 	 * speed is bits 0-3, linkwidth is bits 4-8
248 	 * no defines for them in headers
249 	 */
250 	speed = linkstat & 0xf;
251 	linkstat >>= 4;
252 	linkstat &= 0x1f;
253 	dd->lbus_width = linkstat;
254 
255 	switch (speed) {
256 	case 1:
257 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
258 		break;
259 	case 2:
260 		dd->lbus_speed = 5000; /* Gen1, 5GHz */
261 		break;
262 	default: /* not defined, assume gen1 */
263 		dd->lbus_speed = 2500;
264 		break;
265 	}
266 
267 	/*
268 	 * Check against expected pcie width and complain if "wrong"
269 	 * on first initialization, not afterwards (i.e., reset).
270 	 */
271 	if (minw && linkstat < minw)
272 		qib_dev_err(dd,
273 			    "PCIe width %u (x%u HCA), performance reduced\n",
274 			    linkstat, minw);
275 
276 	qib_tune_pcie_caps(dd);
277 
278 	qib_tune_pcie_coalesce(dd);
279 
280 bail:
281 	/* fill in string, even on errors */
282 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
283 		 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
284 	return nvec < 0 ? nvec : 0;
285 }
286 
287 /**
288  * qib_free_irq - Cleanup INTx and MSI interrupts
289  * @dd: valid pointer to qib dev data
290  *
291  * Since cleanup for INTx and MSI interrupts is trivial, have a common
292  * routine.
293  *
294  */
295 void qib_free_irq(struct qib_devdata *dd)
296 {
297 	pci_free_irq(dd->pcidev, 0, dd);
298 	pci_free_irq_vectors(dd->pcidev);
299 }
300 
301 /*
302  * Setup pcie interrupt stuff again after a reset.  I'd like to just call
303  * pci_enable_msi() again for msi, but when I do that,
304  * the MSI enable bit doesn't get set in the command word, and
305  * we switch to to a different interrupt vector, which is confusing,
306  * so I instead just do it all inline.  Perhaps somehow can tie this
307  * into the PCIe hotplug support at some point
308  */
309 int qib_reinit_intr(struct qib_devdata *dd)
310 {
311 	int pos;
312 	u16 control;
313 	int ret = 0;
314 
315 	/* If we aren't using MSI, don't restore it */
316 	if (!dd->msi_lo)
317 		goto bail;
318 
319 	pos = dd->pcidev->msi_cap;
320 	if (!pos) {
321 		qib_dev_err(dd,
322 			"Can't find MSI capability, can't restore MSI settings\n");
323 		ret = 0;
324 		/* nothing special for MSIx, just MSI */
325 		goto bail;
326 	}
327 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
328 			       dd->msi_lo);
329 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
330 			       dd->msi_hi);
331 	pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
332 	if (!(control & PCI_MSI_FLAGS_ENABLE)) {
333 		control |= PCI_MSI_FLAGS_ENABLE;
334 		pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
335 				      control);
336 	}
337 	/* now rewrite the data (vector) info */
338 	pci_write_config_word(dd->pcidev, pos +
339 			      ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
340 			      dd->msi_data);
341 	ret = 1;
342 bail:
343 	qib_free_irq(dd);
344 
345 	if (!ret && (dd->flags & QIB_HAS_INTX))
346 		ret = 1;
347 
348 	/* and now set the pci master bit again */
349 	pci_set_master(dd->pcidev);
350 
351 	return ret;
352 }
353 
354 /*
355  * These two routines are helper routines for the device reset code
356  * to move all the pcie code out of the chip-specific driver code.
357  */
358 void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
359 {
360 	pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
361 	pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
362 	pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
363 }
364 
365 void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
366 {
367 	int r;
368 
369 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
370 				   dd->pcibar0);
371 	if (r)
372 		qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
373 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
374 				   dd->pcibar1);
375 	if (r)
376 		qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
377 	/* now re-enable memory access, and restore cosmetic settings */
378 	pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
379 	pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
380 	pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
381 	r = pci_enable_device(dd->pcidev);
382 	if (r)
383 		qib_dev_err(dd,
384 			"pci_enable_device failed after reset: %d\n", r);
385 }
386 
387 
388 static int qib_pcie_coalesce;
389 module_param_named(pcie_coalesce, qib_pcie_coalesce, int, S_IRUGO);
390 MODULE_PARM_DESC(pcie_coalesce, "tune PCIe coalescing on some Intel chipsets");
391 
392 /*
393  * Enable PCIe completion and data coalescing, on Intel 5x00 and 7300
394  * chipsets.   This is known to be unsafe for some revisions of some
395  * of these chipsets, with some BIOS settings, and enabling it on those
396  * systems may result in the system crashing, and/or data corruption.
397  */
398 static void qib_tune_pcie_coalesce(struct qib_devdata *dd)
399 {
400 	struct pci_dev *parent;
401 	u16 devid;
402 	u32 mask, bits, val;
403 
404 	if (!qib_pcie_coalesce)
405 		return;
406 
407 	/* Find out supported and configured values for parent (root) */
408 	parent = dd->pcidev->bus->self;
409 	if (parent->bus->parent) {
410 		qib_devinfo(dd->pcidev, "Parent not root\n");
411 		return;
412 	}
413 	if (!pci_is_pcie(parent))
414 		return;
415 	if (parent->vendor != 0x8086)
416 		return;
417 
418 	/*
419 	 *  - bit 12: Max_rdcmp_Imt_EN: need to set to 1
420 	 *  - bit 11: COALESCE_FORCE: need to set to 0
421 	 *  - bit 10: COALESCE_EN: need to set to 1
422 	 *  (but limitations on some on some chipsets)
423 	 *
424 	 *  On the Intel 5000, 5100, and 7300 chipsets, there is
425 	 *  also: - bit 25:24: COALESCE_MODE, need to set to 0
426 	 */
427 	devid = parent->device;
428 	if (devid >= 0x25e2 && devid <= 0x25fa) {
429 		/* 5000 P/V/X/Z */
430 		if (parent->revision <= 0xb2)
431 			bits = 1U << 10;
432 		else
433 			bits = 7U << 10;
434 		mask = (3U << 24) | (7U << 10);
435 	} else if (devid >= 0x65e2 && devid <= 0x65fa) {
436 		/* 5100 */
437 		bits = 1U << 10;
438 		mask = (3U << 24) | (7U << 10);
439 	} else if (devid >= 0x4021 && devid <= 0x402e) {
440 		/* 5400 */
441 		bits = 7U << 10;
442 		mask = 7U << 10;
443 	} else if (devid >= 0x3604 && devid <= 0x360a) {
444 		/* 7300 */
445 		bits = 7U << 10;
446 		mask = (3U << 24) | (7U << 10);
447 	} else {
448 		/* not one of the chipsets that we know about */
449 		return;
450 	}
451 	pci_read_config_dword(parent, 0x48, &val);
452 	val &= ~mask;
453 	val |= bits;
454 	pci_write_config_dword(parent, 0x48, val);
455 }
456 
457 /*
458  * BIOS may not set PCIe bus-utilization parameters for best performance.
459  * Check and optionally adjust them to maximize our throughput.
460  */
461 static int qib_pcie_caps;
462 module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
463 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
464 
465 static void qib_tune_pcie_caps(struct qib_devdata *dd)
466 {
467 	struct pci_dev *parent;
468 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
469 	u16 rc_mrrs, ep_mrrs, max_mrrs;
470 
471 	/* Find out supported and configured values for parent (root) */
472 	parent = dd->pcidev->bus->self;
473 	if (!pci_is_root_bus(parent->bus)) {
474 		qib_devinfo(dd->pcidev, "Parent not root\n");
475 		return;
476 	}
477 
478 	if (!pci_is_pcie(parent) || !pci_is_pcie(dd->pcidev))
479 		return;
480 
481 	rc_mpss = parent->pcie_mpss;
482 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
483 	/* Find out supported and configured values for endpoint (us) */
484 	ep_mpss = dd->pcidev->pcie_mpss;
485 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
486 
487 	/* Find max payload supported by root, endpoint */
488 	if (rc_mpss > ep_mpss)
489 		rc_mpss = ep_mpss;
490 
491 	/* If Supported greater than limit in module param, limit it */
492 	if (rc_mpss > (qib_pcie_caps & 7))
493 		rc_mpss = qib_pcie_caps & 7;
494 	/* If less than (allowed, supported), bump root payload */
495 	if (rc_mpss > rc_mps) {
496 		rc_mps = rc_mpss;
497 		pcie_set_mps(parent, 128 << rc_mps);
498 	}
499 	/* If less than (allowed, supported), bump endpoint payload */
500 	if (rc_mpss > ep_mps) {
501 		ep_mps = rc_mpss;
502 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
503 	}
504 
505 	/*
506 	 * Now the Read Request size.
507 	 * No field for max supported, but PCIe spec limits it to 4096,
508 	 * which is code '5' (log2(4096) - 7)
509 	 */
510 	max_mrrs = 5;
511 	if (max_mrrs > ((qib_pcie_caps >> 4) & 7))
512 		max_mrrs = (qib_pcie_caps >> 4) & 7;
513 
514 	max_mrrs = 128 << max_mrrs;
515 	rc_mrrs = pcie_get_readrq(parent);
516 	ep_mrrs = pcie_get_readrq(dd->pcidev);
517 
518 	if (max_mrrs > rc_mrrs) {
519 		rc_mrrs = max_mrrs;
520 		pcie_set_readrq(parent, rc_mrrs);
521 	}
522 	if (max_mrrs > ep_mrrs) {
523 		ep_mrrs = max_mrrs;
524 		pcie_set_readrq(dd->pcidev, ep_mrrs);
525 	}
526 }
527 /* End of PCIe capability tuning */
528 
529 /*
530  * From here through qib_pci_err_handler definition is invoked via
531  * PCI error infrastructure, registered via pci
532  */
533 static pci_ers_result_t
534 qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
535 {
536 	struct qib_devdata *dd = pci_get_drvdata(pdev);
537 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
538 
539 	switch (state) {
540 	case pci_channel_io_normal:
541 		qib_devinfo(pdev, "State Normal, ignoring\n");
542 		break;
543 
544 	case pci_channel_io_frozen:
545 		qib_devinfo(pdev, "State Frozen, requesting reset\n");
546 		pci_disable_device(pdev);
547 		ret = PCI_ERS_RESULT_NEED_RESET;
548 		break;
549 
550 	case pci_channel_io_perm_failure:
551 		qib_devinfo(pdev, "State Permanent Failure, disabling\n");
552 		if (dd) {
553 			/* no more register accesses! */
554 			dd->flags &= ~QIB_PRESENT;
555 			qib_disable_after_error(dd);
556 		}
557 		 /* else early, or other problem */
558 		ret =  PCI_ERS_RESULT_DISCONNECT;
559 		break;
560 
561 	default: /* shouldn't happen */
562 		qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
563 			state);
564 		break;
565 	}
566 	return ret;
567 }
568 
569 static pci_ers_result_t
570 qib_pci_mmio_enabled(struct pci_dev *pdev)
571 {
572 	u64 words = 0U;
573 	struct qib_devdata *dd = pci_get_drvdata(pdev);
574 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
575 
576 	if (dd && dd->pport) {
577 		words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
578 		if (words == ~0ULL)
579 			ret = PCI_ERS_RESULT_NEED_RESET;
580 	}
581 	qib_devinfo(pdev,
582 		"QIB mmio_enabled function called, read wordscntr %Lx, returning %d\n",
583 		words, ret);
584 	return  ret;
585 }
586 
587 static pci_ers_result_t
588 qib_pci_slot_reset(struct pci_dev *pdev)
589 {
590 	qib_devinfo(pdev, "QIB slot_reset function called, ignored\n");
591 	return PCI_ERS_RESULT_CAN_RECOVER;
592 }
593 
594 static void
595 qib_pci_resume(struct pci_dev *pdev)
596 {
597 	struct qib_devdata *dd = pci_get_drvdata(pdev);
598 
599 	qib_devinfo(pdev, "QIB resume function called\n");
600 	/*
601 	 * Running jobs will fail, since it's asynchronous
602 	 * unlike sysfs-requested reset.   Better than
603 	 * doing nothing.
604 	 */
605 	qib_init(dd, 1); /* same as re-init after reset */
606 }
607 
608 const struct pci_error_handlers qib_pci_err_handler = {
609 	.error_detected = qib_pci_error_detected,
610 	.mmio_enabled = qib_pci_mmio_enabled,
611 	.slot_reset = qib_pci_slot_reset,
612 	.resume = qib_pci_resume,
613 };
614