xref: /linux/drivers/infiniband/hw/qib/qib_pcie.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
1 /*
2  * Copyright (c) 2008, 2009 QLogic Corporation. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/pci.h>
34 #include <linux/io.h>
35 #include <linux/delay.h>
36 #include <linux/vmalloc.h>
37 #include <linux/aer.h>
38 #include <linux/module.h>
39 
40 #include "qib.h"
41 
42 /*
43  * This file contains PCIe utility routines that are common to the
44  * various QLogic InfiniPath adapters
45  */
46 
47 /*
48  * Code to adjust PCIe capabilities.
49  * To minimize the change footprint, we call it
50  * from qib_pcie_params, which every chip-specific
51  * file calls, even though this violates some
52  * expectations of harmlessness.
53  */
54 static void qib_tune_pcie_caps(struct qib_devdata *);
55 static void qib_tune_pcie_coalesce(struct qib_devdata *);
56 
57 /*
58  * Do all the common PCIe setup and initialization.
59  * devdata is not yet allocated, and is not allocated until after this
60  * routine returns success.  Therefore qib_dev_err() can't be used for error
61  * printing.
62  */
63 int qib_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
64 {
65 	int ret;
66 
67 	ret = pci_enable_device(pdev);
68 	if (ret) {
69 		/*
70 		 * This can happen (in theory) iff:
71 		 * We did a chip reset, and then failed to reprogram the
72 		 * BAR, or the chip reset due to an internal error.  We then
73 		 * unloaded the driver and reloaded it.
74 		 *
75 		 * Both reset cases set the BAR back to initial state.  For
76 		 * the latter case, the AER sticky error bit at offset 0x718
77 		 * should be set, but the Linux kernel doesn't yet know
78 		 * about that, it appears.  If the original BAR was retained
79 		 * in the kernel data structures, this may be OK.
80 		 */
81 		qib_early_err(&pdev->dev, "pci enable failed: error %d\n",
82 			      -ret);
83 		goto done;
84 	}
85 
86 	ret = pci_request_regions(pdev, QIB_DRV_NAME);
87 	if (ret) {
88 		qib_devinfo(pdev, "pci_request_regions fails: err %d\n", -ret);
89 		goto bail;
90 	}
91 
92 	ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
93 	if (ret) {
94 		/*
95 		 * If the 64 bit setup fails, try 32 bit.  Some systems
96 		 * do not setup 64 bit maps on systems with 2GB or less
97 		 * memory installed.
98 		 */
99 		ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
100 		if (ret) {
101 			qib_devinfo(pdev, "Unable to set DMA mask: %d\n", ret);
102 			goto bail;
103 		}
104 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
105 	} else
106 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
107 	if (ret) {
108 		qib_early_err(&pdev->dev,
109 			      "Unable to set DMA consistent mask: %d\n", ret);
110 		goto bail;
111 	}
112 
113 	pci_set_master(pdev);
114 	ret = pci_enable_pcie_error_reporting(pdev);
115 	if (ret) {
116 		qib_early_err(&pdev->dev,
117 			      "Unable to enable pcie error reporting: %d\n",
118 			      ret);
119 		ret = 0;
120 	}
121 	goto done;
122 
123 bail:
124 	pci_disable_device(pdev);
125 	pci_release_regions(pdev);
126 done:
127 	return ret;
128 }
129 
130 /*
131  * Do remaining PCIe setup, once dd is allocated, and save away
132  * fields required to re-initialize after a chip reset, or for
133  * various other purposes
134  */
135 int qib_pcie_ddinit(struct qib_devdata *dd, struct pci_dev *pdev,
136 		    const struct pci_device_id *ent)
137 {
138 	unsigned long len;
139 	resource_size_t addr;
140 
141 	dd->pcidev = pdev;
142 	pci_set_drvdata(pdev, dd);
143 
144 	addr = pci_resource_start(pdev, 0);
145 	len = pci_resource_len(pdev, 0);
146 
147 #if defined(__powerpc__)
148 	/* There isn't a generic way to specify writethrough mappings */
149 	dd->kregbase = __ioremap(addr, len, _PAGE_NO_CACHE | _PAGE_WRITETHRU);
150 #else
151 	dd->kregbase = ioremap_nocache(addr, len);
152 #endif
153 
154 	if (!dd->kregbase)
155 		return -ENOMEM;
156 
157 	dd->kregend = (u64 __iomem *)((void __iomem *) dd->kregbase + len);
158 	dd->physaddr = addr;        /* used for io_remap, etc. */
159 
160 	/*
161 	 * Save BARs to rewrite after device reset.  Save all 64 bits of
162 	 * BAR, just in case.
163 	 */
164 	dd->pcibar0 = addr;
165 	dd->pcibar1 = addr >> 32;
166 	dd->deviceid = ent->device; /* save for later use */
167 	dd->vendorid = ent->vendor;
168 
169 	return 0;
170 }
171 
172 /*
173  * Do PCIe cleanup, after chip-specific cleanup, etc.  Just prior
174  * to releasing the dd memory.
175  * void because none of the core pcie cleanup returns are void
176  */
177 void qib_pcie_ddcleanup(struct qib_devdata *dd)
178 {
179 	u64 __iomem *base = (void __iomem *) dd->kregbase;
180 
181 	dd->kregbase = NULL;
182 	iounmap(base);
183 	if (dd->piobase)
184 		iounmap(dd->piobase);
185 	if (dd->userbase)
186 		iounmap(dd->userbase);
187 	if (dd->piovl15base)
188 		iounmap(dd->piovl15base);
189 
190 	pci_disable_device(dd->pcidev);
191 	pci_release_regions(dd->pcidev);
192 
193 	pci_set_drvdata(dd->pcidev, NULL);
194 }
195 
196 static void qib_msix_setup(struct qib_devdata *dd, int pos, u32 *msixcnt,
197 			   struct qib_msix_entry *qib_msix_entry)
198 {
199 	int ret;
200 	int nvec = *msixcnt;
201 	struct msix_entry *msix_entry;
202 	int i;
203 
204 	ret = pci_msix_vec_count(dd->pcidev);
205 	if (ret < 0)
206 		goto do_intx;
207 
208 	nvec = min(nvec, ret);
209 
210 	/* We can't pass qib_msix_entry array to qib_msix_setup
211 	 * so use a dummy msix_entry array and copy the allocated
212 	 * irq back to the qib_msix_entry array. */
213 	msix_entry = kmalloc(nvec * sizeof(*msix_entry), GFP_KERNEL);
214 	if (!msix_entry)
215 		goto do_intx;
216 
217 	for (i = 0; i < nvec; i++)
218 		msix_entry[i] = qib_msix_entry[i].msix;
219 
220 	ret = pci_enable_msix_range(dd->pcidev, msix_entry, 1, nvec);
221 	if (ret < 0)
222 		goto free_msix_entry;
223 	else
224 		nvec = ret;
225 
226 	for (i = 0; i < nvec; i++)
227 		qib_msix_entry[i].msix = msix_entry[i];
228 
229 	kfree(msix_entry);
230 	*msixcnt = nvec;
231 	return;
232 
233 free_msix_entry:
234 	kfree(msix_entry);
235 
236 do_intx:
237 	qib_dev_err(dd, "pci_enable_msix_range %d vectors failed: %d, "
238 			"falling back to INTx\n", nvec, ret);
239 	*msixcnt = 0;
240 	qib_enable_intx(dd->pcidev);
241 }
242 
243 /**
244  * We save the msi lo and hi values, so we can restore them after
245  * chip reset (the kernel PCI infrastructure doesn't yet handle that
246  * correctly.
247  */
248 static int qib_msi_setup(struct qib_devdata *dd, int pos)
249 {
250 	struct pci_dev *pdev = dd->pcidev;
251 	u16 control;
252 	int ret;
253 
254 	ret = pci_enable_msi(pdev);
255 	if (ret)
256 		qib_dev_err(dd,
257 			"pci_enable_msi failed: %d, interrupts may not work\n",
258 			ret);
259 	/* continue even if it fails, we may still be OK... */
260 
261 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO,
262 			      &dd->msi_lo);
263 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI,
264 			      &dd->msi_hi);
265 	pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
266 	/* now save the data (vector) info */
267 	pci_read_config_word(pdev, pos + ((control & PCI_MSI_FLAGS_64BIT)
268 				    ? 12 : 8),
269 			     &dd->msi_data);
270 	return ret;
271 }
272 
273 int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent,
274 		    struct qib_msix_entry *entry)
275 {
276 	u16 linkstat, speed;
277 	int pos = 0, ret = 1;
278 
279 	if (!pci_is_pcie(dd->pcidev)) {
280 		qib_dev_err(dd, "Can't find PCI Express capability!\n");
281 		/* set up something... */
282 		dd->lbus_width = 1;
283 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
284 		goto bail;
285 	}
286 
287 	pos = dd->pcidev->msix_cap;
288 	if (nent && *nent && pos) {
289 		qib_msix_setup(dd, pos, nent, entry);
290 		ret = 0; /* did it, either MSIx or INTx */
291 	} else {
292 		pos = dd->pcidev->msi_cap;
293 		if (pos)
294 			ret = qib_msi_setup(dd, pos);
295 		else
296 			qib_dev_err(dd, "No PCI MSI or MSIx capability!\n");
297 	}
298 	if (!pos)
299 		qib_enable_intx(dd->pcidev);
300 
301 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
302 	/*
303 	 * speed is bits 0-3, linkwidth is bits 4-8
304 	 * no defines for them in headers
305 	 */
306 	speed = linkstat & 0xf;
307 	linkstat >>= 4;
308 	linkstat &= 0x1f;
309 	dd->lbus_width = linkstat;
310 
311 	switch (speed) {
312 	case 1:
313 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
314 		break;
315 	case 2:
316 		dd->lbus_speed = 5000; /* Gen1, 5GHz */
317 		break;
318 	default: /* not defined, assume gen1 */
319 		dd->lbus_speed = 2500;
320 		break;
321 	}
322 
323 	/*
324 	 * Check against expected pcie width and complain if "wrong"
325 	 * on first initialization, not afterwards (i.e., reset).
326 	 */
327 	if (minw && linkstat < minw)
328 		qib_dev_err(dd,
329 			    "PCIe width %u (x%u HCA), performance reduced\n",
330 			    linkstat, minw);
331 
332 	qib_tune_pcie_caps(dd);
333 
334 	qib_tune_pcie_coalesce(dd);
335 
336 bail:
337 	/* fill in string, even on errors */
338 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
339 		 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
340 	return ret;
341 }
342 
343 /*
344  * Setup pcie interrupt stuff again after a reset.  I'd like to just call
345  * pci_enable_msi() again for msi, but when I do that,
346  * the MSI enable bit doesn't get set in the command word, and
347  * we switch to to a different interrupt vector, which is confusing,
348  * so I instead just do it all inline.  Perhaps somehow can tie this
349  * into the PCIe hotplug support at some point
350  */
351 int qib_reinit_intr(struct qib_devdata *dd)
352 {
353 	int pos;
354 	u16 control;
355 	int ret = 0;
356 
357 	/* If we aren't using MSI, don't restore it */
358 	if (!dd->msi_lo)
359 		goto bail;
360 
361 	pos = dd->pcidev->msi_cap;
362 	if (!pos) {
363 		qib_dev_err(dd,
364 			"Can't find MSI capability, can't restore MSI settings\n");
365 		ret = 0;
366 		/* nothing special for MSIx, just MSI */
367 		goto bail;
368 	}
369 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
370 			       dd->msi_lo);
371 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
372 			       dd->msi_hi);
373 	pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
374 	if (!(control & PCI_MSI_FLAGS_ENABLE)) {
375 		control |= PCI_MSI_FLAGS_ENABLE;
376 		pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
377 				      control);
378 	}
379 	/* now rewrite the data (vector) info */
380 	pci_write_config_word(dd->pcidev, pos +
381 			      ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
382 			      dd->msi_data);
383 	ret = 1;
384 bail:
385 	if (!ret && (dd->flags & QIB_HAS_INTX)) {
386 		qib_enable_intx(dd->pcidev);
387 		ret = 1;
388 	}
389 
390 	/* and now set the pci master bit again */
391 	pci_set_master(dd->pcidev);
392 
393 	return ret;
394 }
395 
396 /*
397  * Disable msi interrupt if enabled, and clear msi_lo.
398  * This is used primarily for the fallback to INTx, but
399  * is also used in reinit after reset, and during cleanup.
400  */
401 void qib_nomsi(struct qib_devdata *dd)
402 {
403 	dd->msi_lo = 0;
404 	pci_disable_msi(dd->pcidev);
405 }
406 
407 /*
408  * Same as qib_nosmi, but for MSIx.
409  */
410 void qib_nomsix(struct qib_devdata *dd)
411 {
412 	pci_disable_msix(dd->pcidev);
413 }
414 
415 /*
416  * Similar to pci_intx(pdev, 1), except that we make sure
417  * msi(x) is off.
418  */
419 void qib_enable_intx(struct pci_dev *pdev)
420 {
421 	u16 cw, new;
422 	int pos;
423 
424 	/* first, turn on INTx */
425 	pci_read_config_word(pdev, PCI_COMMAND, &cw);
426 	new = cw & ~PCI_COMMAND_INTX_DISABLE;
427 	if (new != cw)
428 		pci_write_config_word(pdev, PCI_COMMAND, new);
429 
430 	pos = pdev->msi_cap;
431 	if (pos) {
432 		/* then turn off MSI */
433 		pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &cw);
434 		new = cw & ~PCI_MSI_FLAGS_ENABLE;
435 		if (new != cw)
436 			pci_write_config_word(pdev, pos + PCI_MSI_FLAGS, new);
437 	}
438 	pos = pdev->msix_cap;
439 	if (pos) {
440 		/* then turn off MSIx */
441 		pci_read_config_word(pdev, pos + PCI_MSIX_FLAGS, &cw);
442 		new = cw & ~PCI_MSIX_FLAGS_ENABLE;
443 		if (new != cw)
444 			pci_write_config_word(pdev, pos + PCI_MSIX_FLAGS, new);
445 	}
446 }
447 
448 /*
449  * These two routines are helper routines for the device reset code
450  * to move all the pcie code out of the chip-specific driver code.
451  */
452 void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
453 {
454 	pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
455 	pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
456 	pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
457 }
458 
459 void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
460 {
461 	int r;
462 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
463 				   dd->pcibar0);
464 	if (r)
465 		qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
466 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
467 				   dd->pcibar1);
468 	if (r)
469 		qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
470 	/* now re-enable memory access, and restore cosmetic settings */
471 	pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
472 	pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
473 	pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
474 	r = pci_enable_device(dd->pcidev);
475 	if (r)
476 		qib_dev_err(dd,
477 			"pci_enable_device failed after reset: %d\n", r);
478 }
479 
480 
481 static int qib_pcie_coalesce;
482 module_param_named(pcie_coalesce, qib_pcie_coalesce, int, S_IRUGO);
483 MODULE_PARM_DESC(pcie_coalesce, "tune PCIe colescing on some Intel chipsets");
484 
485 /*
486  * Enable PCIe completion and data coalescing, on Intel 5x00 and 7300
487  * chipsets.   This is known to be unsafe for some revisions of some
488  * of these chipsets, with some BIOS settings, and enabling it on those
489  * systems may result in the system crashing, and/or data corruption.
490  */
491 static void qib_tune_pcie_coalesce(struct qib_devdata *dd)
492 {
493 	int r;
494 	struct pci_dev *parent;
495 	u16 devid;
496 	u32 mask, bits, val;
497 
498 	if (!qib_pcie_coalesce)
499 		return;
500 
501 	/* Find out supported and configured values for parent (root) */
502 	parent = dd->pcidev->bus->self;
503 	if (parent->bus->parent) {
504 		qib_devinfo(dd->pcidev, "Parent not root\n");
505 		return;
506 	}
507 	if (!pci_is_pcie(parent))
508 		return;
509 	if (parent->vendor != 0x8086)
510 		return;
511 
512 	/*
513 	 *  - bit 12: Max_rdcmp_Imt_EN: need to set to 1
514 	 *  - bit 11: COALESCE_FORCE: need to set to 0
515 	 *  - bit 10: COALESCE_EN: need to set to 1
516 	 *  (but limitations on some on some chipsets)
517 	 *
518 	 *  On the Intel 5000, 5100, and 7300 chipsets, there is
519 	 *  also: - bit 25:24: COALESCE_MODE, need to set to 0
520 	 */
521 	devid = parent->device;
522 	if (devid >= 0x25e2 && devid <= 0x25fa) {
523 		/* 5000 P/V/X/Z */
524 		if (parent->revision <= 0xb2)
525 			bits = 1U << 10;
526 		else
527 			bits = 7U << 10;
528 		mask = (3U << 24) | (7U << 10);
529 	} else if (devid >= 0x65e2 && devid <= 0x65fa) {
530 		/* 5100 */
531 		bits = 1U << 10;
532 		mask = (3U << 24) | (7U << 10);
533 	} else if (devid >= 0x4021 && devid <= 0x402e) {
534 		/* 5400 */
535 		bits = 7U << 10;
536 		mask = 7U << 10;
537 	} else if (devid >= 0x3604 && devid <= 0x360a) {
538 		/* 7300 */
539 		bits = 7U << 10;
540 		mask = (3U << 24) | (7U << 10);
541 	} else {
542 		/* not one of the chipsets that we know about */
543 		return;
544 	}
545 	pci_read_config_dword(parent, 0x48, &val);
546 	val &= ~mask;
547 	val |= bits;
548 	r = pci_write_config_dword(parent, 0x48, val);
549 }
550 
551 /*
552  * BIOS may not set PCIe bus-utilization parameters for best performance.
553  * Check and optionally adjust them to maximize our throughput.
554  */
555 static int qib_pcie_caps;
556 module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
557 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
558 
559 static void qib_tune_pcie_caps(struct qib_devdata *dd)
560 {
561 	struct pci_dev *parent;
562 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
563 	u16 rc_mrrs, ep_mrrs, max_mrrs;
564 
565 	/* Find out supported and configured values for parent (root) */
566 	parent = dd->pcidev->bus->self;
567 	if (!pci_is_root_bus(parent->bus)) {
568 		qib_devinfo(dd->pcidev, "Parent not root\n");
569 		return;
570 	}
571 
572 	if (!pci_is_pcie(parent) || !pci_is_pcie(dd->pcidev))
573 		return;
574 
575 	rc_mpss = parent->pcie_mpss;
576 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
577 	/* Find out supported and configured values for endpoint (us) */
578 	ep_mpss = dd->pcidev->pcie_mpss;
579 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
580 
581 	/* Find max payload supported by root, endpoint */
582 	if (rc_mpss > ep_mpss)
583 		rc_mpss = ep_mpss;
584 
585 	/* If Supported greater than limit in module param, limit it */
586 	if (rc_mpss > (qib_pcie_caps & 7))
587 		rc_mpss = qib_pcie_caps & 7;
588 	/* If less than (allowed, supported), bump root payload */
589 	if (rc_mpss > rc_mps) {
590 		rc_mps = rc_mpss;
591 		pcie_set_mps(parent, 128 << rc_mps);
592 	}
593 	/* If less than (allowed, supported), bump endpoint payload */
594 	if (rc_mpss > ep_mps) {
595 		ep_mps = rc_mpss;
596 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
597 	}
598 
599 	/*
600 	 * Now the Read Request size.
601 	 * No field for max supported, but PCIe spec limits it to 4096,
602 	 * which is code '5' (log2(4096) - 7)
603 	 */
604 	max_mrrs = 5;
605 	if (max_mrrs > ((qib_pcie_caps >> 4) & 7))
606 		max_mrrs = (qib_pcie_caps >> 4) & 7;
607 
608 	max_mrrs = 128 << max_mrrs;
609 	rc_mrrs = pcie_get_readrq(parent);
610 	ep_mrrs = pcie_get_readrq(dd->pcidev);
611 
612 	if (max_mrrs > rc_mrrs) {
613 		rc_mrrs = max_mrrs;
614 		pcie_set_readrq(parent, rc_mrrs);
615 	}
616 	if (max_mrrs > ep_mrrs) {
617 		ep_mrrs = max_mrrs;
618 		pcie_set_readrq(dd->pcidev, ep_mrrs);
619 	}
620 }
621 /* End of PCIe capability tuning */
622 
623 /*
624  * From here through qib_pci_err_handler definition is invoked via
625  * PCI error infrastructure, registered via pci
626  */
627 static pci_ers_result_t
628 qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
629 {
630 	struct qib_devdata *dd = pci_get_drvdata(pdev);
631 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
632 
633 	switch (state) {
634 	case pci_channel_io_normal:
635 		qib_devinfo(pdev, "State Normal, ignoring\n");
636 		break;
637 
638 	case pci_channel_io_frozen:
639 		qib_devinfo(pdev, "State Frozen, requesting reset\n");
640 		pci_disable_device(pdev);
641 		ret = PCI_ERS_RESULT_NEED_RESET;
642 		break;
643 
644 	case pci_channel_io_perm_failure:
645 		qib_devinfo(pdev, "State Permanent Failure, disabling\n");
646 		if (dd) {
647 			/* no more register accesses! */
648 			dd->flags &= ~QIB_PRESENT;
649 			qib_disable_after_error(dd);
650 		}
651 		 /* else early, or other problem */
652 		ret =  PCI_ERS_RESULT_DISCONNECT;
653 		break;
654 
655 	default: /* shouldn't happen */
656 		qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
657 			state);
658 		break;
659 	}
660 	return ret;
661 }
662 
663 static pci_ers_result_t
664 qib_pci_mmio_enabled(struct pci_dev *pdev)
665 {
666 	u64 words = 0U;
667 	struct qib_devdata *dd = pci_get_drvdata(pdev);
668 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
669 
670 	if (dd && dd->pport) {
671 		words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
672 		if (words == ~0ULL)
673 			ret = PCI_ERS_RESULT_NEED_RESET;
674 	}
675 	qib_devinfo(pdev,
676 		"QIB mmio_enabled function called, read wordscntr %Lx, returning %d\n",
677 		words, ret);
678 	return  ret;
679 }
680 
681 static pci_ers_result_t
682 qib_pci_slot_reset(struct pci_dev *pdev)
683 {
684 	qib_devinfo(pdev, "QIB slot_reset function called, ignored\n");
685 	return PCI_ERS_RESULT_CAN_RECOVER;
686 }
687 
688 static pci_ers_result_t
689 qib_pci_link_reset(struct pci_dev *pdev)
690 {
691 	qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
692 	return PCI_ERS_RESULT_CAN_RECOVER;
693 }
694 
695 static void
696 qib_pci_resume(struct pci_dev *pdev)
697 {
698 	struct qib_devdata *dd = pci_get_drvdata(pdev);
699 	qib_devinfo(pdev, "QIB resume function called\n");
700 	pci_cleanup_aer_uncorrect_error_status(pdev);
701 	/*
702 	 * Running jobs will fail, since it's asynchronous
703 	 * unlike sysfs-requested reset.   Better than
704 	 * doing nothing.
705 	 */
706 	qib_init(dd, 1); /* same as re-init after reset */
707 }
708 
709 const struct pci_error_handlers qib_pci_err_handler = {
710 	.error_detected = qib_pci_error_detected,
711 	.mmio_enabled = qib_pci_mmio_enabled,
712 	.link_reset = qib_pci_link_reset,
713 	.slot_reset = qib_pci_slot_reset,
714 	.resume = qib_pci_resume,
715 };
716