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 
39 #include "qib.h"
40 
41 /*
42  * This file contains PCIe utility routines that are common to the
43  * various QLogic InfiniPath adapters
44  */
45 
46 /*
47  * Code to adjust PCIe capabilities.
48  * To minimize the change footprint, we call it
49  * from qib_pcie_params, which every chip-specific
50  * file calls, even though this violates some
51  * expectations of harmlessness.
52  */
53 static int qib_tune_pcie_caps(struct qib_devdata *);
54 static int qib_tune_pcie_coalesce(struct qib_devdata *);
55 
56 /*
57  * Do all the common PCIe setup and initialization.
58  * devdata is not yet allocated, and is not allocated until after this
59  * routine returns success.  Therefore qib_dev_err() can't be used for error
60  * printing.
61  */
62 int qib_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
63 {
64 	int ret;
65 
66 	ret = pci_enable_device(pdev);
67 	if (ret) {
68 		/*
69 		 * This can happen (in theory) iff:
70 		 * We did a chip reset, and then failed to reprogram the
71 		 * BAR, or the chip reset due to an internal error.  We then
72 		 * unloaded the driver and reloaded it.
73 		 *
74 		 * Both reset cases set the BAR back to initial state.  For
75 		 * the latter case, the AER sticky error bit at offset 0x718
76 		 * should be set, but the Linux kernel doesn't yet know
77 		 * about that, it appears.  If the original BAR was retained
78 		 * in the kernel data structures, this may be OK.
79 		 */
80 		qib_early_err(&pdev->dev, "pci enable failed: error %d\n",
81 			      -ret);
82 		goto done;
83 	}
84 
85 	ret = pci_request_regions(pdev, QIB_DRV_NAME);
86 	if (ret) {
87 		qib_devinfo(pdev, "pci_request_regions fails: err %d\n", -ret);
88 		goto bail;
89 	}
90 
91 	ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
92 	if (ret) {
93 		/*
94 		 * If the 64 bit setup fails, try 32 bit.  Some systems
95 		 * do not setup 64 bit maps on systems with 2GB or less
96 		 * memory installed.
97 		 */
98 		ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
99 		if (ret) {
100 			qib_devinfo(pdev, "Unable to set DMA mask: %d\n", ret);
101 			goto bail;
102 		}
103 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
104 	} else
105 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
106 	if (ret) {
107 		qib_early_err(&pdev->dev,
108 			      "Unable to set DMA consistent mask: %d\n", ret);
109 		goto bail;
110 	}
111 
112 	pci_set_master(pdev);
113 	ret = pci_enable_pcie_error_reporting(pdev);
114 	if (ret) {
115 		qib_early_err(&pdev->dev,
116 			      "Unable to enable pcie error reporting: %d\n",
117 			      ret);
118 		ret = 0;
119 	}
120 	goto done;
121 
122 bail:
123 	pci_disable_device(pdev);
124 	pci_release_regions(pdev);
125 done:
126 	return ret;
127 }
128 
129 /*
130  * Do remaining PCIe setup, once dd is allocated, and save away
131  * fields required to re-initialize after a chip reset, or for
132  * various other purposes
133  */
134 int qib_pcie_ddinit(struct qib_devdata *dd, struct pci_dev *pdev,
135 		    const struct pci_device_id *ent)
136 {
137 	unsigned long len;
138 	resource_size_t addr;
139 
140 	dd->pcidev = pdev;
141 	pci_set_drvdata(pdev, dd);
142 
143 	addr = pci_resource_start(pdev, 0);
144 	len = pci_resource_len(pdev, 0);
145 
146 #if defined(__powerpc__)
147 	/* There isn't a generic way to specify writethrough mappings */
148 	dd->kregbase = __ioremap(addr, len, _PAGE_NO_CACHE | _PAGE_WRITETHRU);
149 #else
150 	dd->kregbase = ioremap_nocache(addr, len);
151 #endif
152 
153 	if (!dd->kregbase)
154 		return -ENOMEM;
155 
156 	dd->kregend = (u64 __iomem *)((void __iomem *) dd->kregbase + len);
157 	dd->physaddr = addr;        /* used for io_remap, etc. */
158 
159 	/*
160 	 * Save BARs to rewrite after device reset.  Save all 64 bits of
161 	 * BAR, just in case.
162 	 */
163 	dd->pcibar0 = addr;
164 	dd->pcibar1 = addr >> 32;
165 	dd->deviceid = ent->device; /* save for later use */
166 	dd->vendorid = ent->vendor;
167 
168 	return 0;
169 }
170 
171 /*
172  * Do PCIe cleanup, after chip-specific cleanup, etc.  Just prior
173  * to releasing the dd memory.
174  * void because none of the core pcie cleanup returns are void
175  */
176 void qib_pcie_ddcleanup(struct qib_devdata *dd)
177 {
178 	u64 __iomem *base = (void __iomem *) dd->kregbase;
179 
180 	dd->kregbase = NULL;
181 	iounmap(base);
182 	if (dd->piobase)
183 		iounmap(dd->piobase);
184 	if (dd->userbase)
185 		iounmap(dd->userbase);
186 	if (dd->piovl15base)
187 		iounmap(dd->piovl15base);
188 
189 	pci_disable_device(dd->pcidev);
190 	pci_release_regions(dd->pcidev);
191 
192 	pci_set_drvdata(dd->pcidev, NULL);
193 }
194 
195 static void qib_msix_setup(struct qib_devdata *dd, int pos, u32 *msixcnt,
196 			   struct msix_entry *msix_entry)
197 {
198 	int ret;
199 	u32 tabsize = 0;
200 	u16 msix_flags;
201 
202 	pci_read_config_word(dd->pcidev, pos + PCI_MSIX_FLAGS, &msix_flags);
203 	tabsize = 1 + (msix_flags & PCI_MSIX_FLAGS_QSIZE);
204 	if (tabsize > *msixcnt)
205 		tabsize = *msixcnt;
206 	ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
207 	if (ret > 0) {
208 		tabsize = ret;
209 		ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
210 	}
211 	if (ret) {
212 		qib_dev_err(dd, "pci_enable_msix %d vectors failed: %d, "
213 			    "falling back to INTx\n", tabsize, ret);
214 		tabsize = 0;
215 	}
216 	*msixcnt = tabsize;
217 
218 	if (ret)
219 		qib_enable_intx(dd->pcidev);
220 
221 }
222 
223 /**
224  * We save the msi lo and hi values, so we can restore them after
225  * chip reset (the kernel PCI infrastructure doesn't yet handle that
226  * correctly.
227  */
228 static int qib_msi_setup(struct qib_devdata *dd, int pos)
229 {
230 	struct pci_dev *pdev = dd->pcidev;
231 	u16 control;
232 	int ret;
233 
234 	ret = pci_enable_msi(pdev);
235 	if (ret)
236 		qib_dev_err(dd, "pci_enable_msi failed: %d, "
237 			    "interrupts may not work\n", ret);
238 	/* continue even if it fails, we may still be OK... */
239 
240 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO,
241 			      &dd->msi_lo);
242 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI,
243 			      &dd->msi_hi);
244 	pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
245 	/* now save the data (vector) info */
246 	pci_read_config_word(pdev, pos + ((control & PCI_MSI_FLAGS_64BIT)
247 				    ? 12 : 8),
248 			     &dd->msi_data);
249 	return ret;
250 }
251 
252 int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent,
253 		    struct msix_entry *entry)
254 {
255 	u16 linkstat, speed;
256 	int pos = 0, pose, ret = 1;
257 
258 	pose = pci_find_capability(dd->pcidev, PCI_CAP_ID_EXP);
259 	if (!pose) {
260 		qib_dev_err(dd, "Can't find PCI Express capability!\n");
261 		/* set up something... */
262 		dd->lbus_width = 1;
263 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
264 		goto bail;
265 	}
266 
267 	pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSIX);
268 	if (nent && *nent && pos) {
269 		qib_msix_setup(dd, pos, nent, entry);
270 		ret = 0; /* did it, either MSIx or INTx */
271 	} else {
272 		pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
273 		if (pos)
274 			ret = qib_msi_setup(dd, pos);
275 		else
276 			qib_dev_err(dd, "No PCI MSI or MSIx capability!\n");
277 	}
278 	if (!pos)
279 		qib_enable_intx(dd->pcidev);
280 
281 	pci_read_config_word(dd->pcidev, pose + PCI_EXP_LNKSTA, &linkstat);
282 	/*
283 	 * speed is bits 0-3, linkwidth is bits 4-8
284 	 * no defines for them in headers
285 	 */
286 	speed = linkstat & 0xf;
287 	linkstat >>= 4;
288 	linkstat &= 0x1f;
289 	dd->lbus_width = linkstat;
290 
291 	switch (speed) {
292 	case 1:
293 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
294 		break;
295 	case 2:
296 		dd->lbus_speed = 5000; /* Gen1, 5GHz */
297 		break;
298 	default: /* not defined, assume gen1 */
299 		dd->lbus_speed = 2500;
300 		break;
301 	}
302 
303 	/*
304 	 * Check against expected pcie width and complain if "wrong"
305 	 * on first initialization, not afterwards (i.e., reset).
306 	 */
307 	if (minw && linkstat < minw)
308 		qib_dev_err(dd,
309 			    "PCIe width %u (x%u HCA), performance reduced\n",
310 			    linkstat, minw);
311 
312 	qib_tune_pcie_caps(dd);
313 
314 	qib_tune_pcie_coalesce(dd);
315 
316 bail:
317 	/* fill in string, even on errors */
318 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
319 		 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
320 	return ret;
321 }
322 
323 /*
324  * Setup pcie interrupt stuff again after a reset.  I'd like to just call
325  * pci_enable_msi() again for msi, but when I do that,
326  * the MSI enable bit doesn't get set in the command word, and
327  * we switch to to a different interrupt vector, which is confusing,
328  * so I instead just do it all inline.  Perhaps somehow can tie this
329  * into the PCIe hotplug support at some point
330  */
331 int qib_reinit_intr(struct qib_devdata *dd)
332 {
333 	int pos;
334 	u16 control;
335 	int ret = 0;
336 
337 	/* If we aren't using MSI, don't restore it */
338 	if (!dd->msi_lo)
339 		goto bail;
340 
341 	pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
342 	if (!pos) {
343 		qib_dev_err(dd, "Can't find MSI capability, "
344 			    "can't restore MSI settings\n");
345 		ret = 0;
346 		/* nothing special for MSIx, just MSI */
347 		goto bail;
348 	}
349 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
350 			       dd->msi_lo);
351 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
352 			       dd->msi_hi);
353 	pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
354 	if (!(control & PCI_MSI_FLAGS_ENABLE)) {
355 		control |= PCI_MSI_FLAGS_ENABLE;
356 		pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
357 				      control);
358 	}
359 	/* now rewrite the data (vector) info */
360 	pci_write_config_word(dd->pcidev, pos +
361 			      ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
362 			      dd->msi_data);
363 	ret = 1;
364 bail:
365 	if (!ret && (dd->flags & QIB_HAS_INTX)) {
366 		qib_enable_intx(dd->pcidev);
367 		ret = 1;
368 	}
369 
370 	/* and now set the pci master bit again */
371 	pci_set_master(dd->pcidev);
372 
373 	return ret;
374 }
375 
376 /*
377  * Disable msi interrupt if enabled, and clear msi_lo.
378  * This is used primarily for the fallback to INTx, but
379  * is also used in reinit after reset, and during cleanup.
380  */
381 void qib_nomsi(struct qib_devdata *dd)
382 {
383 	dd->msi_lo = 0;
384 	pci_disable_msi(dd->pcidev);
385 }
386 
387 /*
388  * Same as qib_nosmi, but for MSIx.
389  */
390 void qib_nomsix(struct qib_devdata *dd)
391 {
392 	pci_disable_msix(dd->pcidev);
393 }
394 
395 /*
396  * Similar to pci_intx(pdev, 1), except that we make sure
397  * msi(x) is off.
398  */
399 void qib_enable_intx(struct pci_dev *pdev)
400 {
401 	u16 cw, new;
402 	int pos;
403 
404 	/* first, turn on INTx */
405 	pci_read_config_word(pdev, PCI_COMMAND, &cw);
406 	new = cw & ~PCI_COMMAND_INTX_DISABLE;
407 	if (new != cw)
408 		pci_write_config_word(pdev, PCI_COMMAND, new);
409 
410 	pos = pci_find_capability(pdev, PCI_CAP_ID_MSI);
411 	if (pos) {
412 		/* then turn off MSI */
413 		pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &cw);
414 		new = cw & ~PCI_MSI_FLAGS_ENABLE;
415 		if (new != cw)
416 			pci_write_config_word(pdev, pos + PCI_MSI_FLAGS, new);
417 	}
418 	pos = pci_find_capability(pdev, PCI_CAP_ID_MSIX);
419 	if (pos) {
420 		/* then turn off MSIx */
421 		pci_read_config_word(pdev, pos + PCI_MSIX_FLAGS, &cw);
422 		new = cw & ~PCI_MSIX_FLAGS_ENABLE;
423 		if (new != cw)
424 			pci_write_config_word(pdev, pos + PCI_MSIX_FLAGS, new);
425 	}
426 }
427 
428 /*
429  * These two routines are helper routines for the device reset code
430  * to move all the pcie code out of the chip-specific driver code.
431  */
432 void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
433 {
434 	pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
435 	pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
436 	pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
437 }
438 
439 void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
440 {
441 	int r;
442 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
443 				   dd->pcibar0);
444 	if (r)
445 		qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
446 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
447 				   dd->pcibar1);
448 	if (r)
449 		qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
450 	/* now re-enable memory access, and restore cosmetic settings */
451 	pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
452 	pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
453 	pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
454 	r = pci_enable_device(dd->pcidev);
455 	if (r)
456 		qib_dev_err(dd, "pci_enable_device failed after "
457 			    "reset: %d\n", r);
458 }
459 
460 /* code to adjust PCIe capabilities. */
461 
462 static int fld2val(int wd, int mask)
463 {
464 	int lsbmask;
465 
466 	if (!mask)
467 		return 0;
468 	wd &= mask;
469 	lsbmask = mask ^ (mask & (mask - 1));
470 	wd /= lsbmask;
471 	return wd;
472 }
473 
474 static int val2fld(int wd, int mask)
475 {
476 	int lsbmask;
477 
478 	if (!mask)
479 		return 0;
480 	lsbmask = mask ^ (mask & (mask - 1));
481 	wd *= lsbmask;
482 	return wd;
483 }
484 
485 static int qib_pcie_coalesce;
486 module_param_named(pcie_coalesce, qib_pcie_coalesce, int, S_IRUGO);
487 MODULE_PARM_DESC(pcie_coalesce, "tune PCIe colescing on some Intel chipsets");
488 
489 /*
490  * Enable PCIe completion and data coalescing, on Intel 5x00 and 7300
491  * chipsets.   This is known to be unsafe for some revisions of some
492  * of these chipsets, with some BIOS settings, and enabling it on those
493  * systems may result in the system crashing, and/or data corruption.
494  */
495 static int qib_tune_pcie_coalesce(struct qib_devdata *dd)
496 {
497 	int r;
498 	struct pci_dev *parent;
499 	int ppos;
500 	u16 devid;
501 	u32 mask, bits, val;
502 
503 	if (!qib_pcie_coalesce)
504 		return 0;
505 
506 	/* Find out supported and configured values for parent (root) */
507 	parent = dd->pcidev->bus->self;
508 	if (parent->bus->parent) {
509 		qib_devinfo(dd->pcidev, "Parent not root\n");
510 		return 1;
511 	}
512 	ppos = pci_find_capability(parent, PCI_CAP_ID_EXP);
513 	if (!ppos)
514 		return 1;
515 	if (parent->vendor != 0x8086)
516 		return 1;
517 
518 	/*
519 	 *  - bit 12: Max_rdcmp_Imt_EN: need to set to 1
520 	 *  - bit 11: COALESCE_FORCE: need to set to 0
521 	 *  - bit 10: COALESCE_EN: need to set to 1
522 	 *  (but limitations on some on some chipsets)
523 	 *
524 	 *  On the Intel 5000, 5100, and 7300 chipsets, there is
525 	 *  also: - bit 25:24: COALESCE_MODE, need to set to 0
526 	 */
527 	devid = parent->device;
528 	if (devid >= 0x25e2 && devid <= 0x25fa) {
529 		/* 5000 P/V/X/Z */
530 		if (parent->revision <= 0xb2)
531 			bits = 1U << 10;
532 		else
533 			bits = 7U << 10;
534 		mask = (3U << 24) | (7U << 10);
535 	} else if (devid >= 0x65e2 && devid <= 0x65fa) {
536 		/* 5100 */
537 		bits = 1U << 10;
538 		mask = (3U << 24) | (7U << 10);
539 	} else if (devid >= 0x4021 && devid <= 0x402e) {
540 		/* 5400 */
541 		bits = 7U << 10;
542 		mask = 7U << 10;
543 	} else if (devid >= 0x3604 && devid <= 0x360a) {
544 		/* 7300 */
545 		bits = 7U << 10;
546 		mask = (3U << 24) | (7U << 10);
547 	} else {
548 		/* not one of the chipsets that we know about */
549 		return 1;
550 	}
551 	pci_read_config_dword(parent, 0x48, &val);
552 	val &= ~mask;
553 	val |= bits;
554 	r = pci_write_config_dword(parent, 0x48, val);
555 	return 0;
556 }
557 
558 /*
559  * BIOS may not set PCIe bus-utilization parameters for best performance.
560  * Check and optionally adjust them to maximize our throughput.
561  */
562 static int qib_pcie_caps;
563 module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
564 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (4lsb), ReadReq (D4..7)");
565 
566 static int qib_tune_pcie_caps(struct qib_devdata *dd)
567 {
568 	int ret = 1; /* Assume the worst */
569 	struct pci_dev *parent;
570 	int ppos, epos;
571 	u16 pcaps, pctl, ecaps, ectl;
572 	int rc_sup, ep_sup;
573 	int rc_cur, ep_cur;
574 
575 	/* Find out supported and configured values for parent (root) */
576 	parent = dd->pcidev->bus->self;
577 	if (parent->bus->parent) {
578 		qib_devinfo(dd->pcidev, "Parent not root\n");
579 		goto bail;
580 	}
581 	ppos = pci_find_capability(parent, PCI_CAP_ID_EXP);
582 	if (ppos) {
583 		pci_read_config_word(parent, ppos + PCI_EXP_DEVCAP, &pcaps);
584 		pci_read_config_word(parent, ppos + PCI_EXP_DEVCTL, &pctl);
585 	} else
586 		goto bail;
587 	/* Find out supported and configured values for endpoint (us) */
588 	epos = pci_find_capability(dd->pcidev, PCI_CAP_ID_EXP);
589 	if (epos) {
590 		pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCAP, &ecaps);
591 		pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, &ectl);
592 	} else
593 		goto bail;
594 	ret = 0;
595 	/* Find max payload supported by root, endpoint */
596 	rc_sup = fld2val(pcaps, PCI_EXP_DEVCAP_PAYLOAD);
597 	ep_sup = fld2val(ecaps, PCI_EXP_DEVCAP_PAYLOAD);
598 	if (rc_sup > ep_sup)
599 		rc_sup = ep_sup;
600 
601 	rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_PAYLOAD);
602 	ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_PAYLOAD);
603 
604 	/* If Supported greater than limit in module param, limit it */
605 	if (rc_sup > (qib_pcie_caps & 7))
606 		rc_sup = qib_pcie_caps & 7;
607 	/* If less than (allowed, supported), bump root payload */
608 	if (rc_sup > rc_cur) {
609 		rc_cur = rc_sup;
610 		pctl = (pctl & ~PCI_EXP_DEVCTL_PAYLOAD) |
611 			val2fld(rc_cur, PCI_EXP_DEVCTL_PAYLOAD);
612 		pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
613 	}
614 	/* If less than (allowed, supported), bump endpoint payload */
615 	if (rc_sup > ep_cur) {
616 		ep_cur = rc_sup;
617 		ectl = (ectl & ~PCI_EXP_DEVCTL_PAYLOAD) |
618 			val2fld(ep_cur, PCI_EXP_DEVCTL_PAYLOAD);
619 		pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
620 	}
621 
622 	/*
623 	 * Now the Read Request size.
624 	 * No field for max supported, but PCIe spec limits it to 4096,
625 	 * which is code '5' (log2(4096) - 7)
626 	 */
627 	rc_sup = 5;
628 	if (rc_sup > ((qib_pcie_caps >> 4) & 7))
629 		rc_sup = (qib_pcie_caps >> 4) & 7;
630 	rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_READRQ);
631 	ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_READRQ);
632 
633 	if (rc_sup > rc_cur) {
634 		rc_cur = rc_sup;
635 		pctl = (pctl & ~PCI_EXP_DEVCTL_READRQ) |
636 			val2fld(rc_cur, PCI_EXP_DEVCTL_READRQ);
637 		pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
638 	}
639 	if (rc_sup > ep_cur) {
640 		ep_cur = rc_sup;
641 		ectl = (ectl & ~PCI_EXP_DEVCTL_READRQ) |
642 			val2fld(ep_cur, PCI_EXP_DEVCTL_READRQ);
643 		pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
644 	}
645 bail:
646 	return ret;
647 }
648 /* End of PCIe capability tuning */
649 
650 /*
651  * From here through qib_pci_err_handler definition is invoked via
652  * PCI error infrastructure, registered via pci
653  */
654 static pci_ers_result_t
655 qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
656 {
657 	struct qib_devdata *dd = pci_get_drvdata(pdev);
658 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
659 
660 	switch (state) {
661 	case pci_channel_io_normal:
662 		qib_devinfo(pdev, "State Normal, ignoring\n");
663 		break;
664 
665 	case pci_channel_io_frozen:
666 		qib_devinfo(pdev, "State Frozen, requesting reset\n");
667 		pci_disable_device(pdev);
668 		ret = PCI_ERS_RESULT_NEED_RESET;
669 		break;
670 
671 	case pci_channel_io_perm_failure:
672 		qib_devinfo(pdev, "State Permanent Failure, disabling\n");
673 		if (dd) {
674 			/* no more register accesses! */
675 			dd->flags &= ~QIB_PRESENT;
676 			qib_disable_after_error(dd);
677 		}
678 		 /* else early, or other problem */
679 		ret =  PCI_ERS_RESULT_DISCONNECT;
680 		break;
681 
682 	default: /* shouldn't happen */
683 		qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
684 			state);
685 		break;
686 	}
687 	return ret;
688 }
689 
690 static pci_ers_result_t
691 qib_pci_mmio_enabled(struct pci_dev *pdev)
692 {
693 	u64 words = 0U;
694 	struct qib_devdata *dd = pci_get_drvdata(pdev);
695 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
696 
697 	if (dd && dd->pport) {
698 		words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
699 		if (words == ~0ULL)
700 			ret = PCI_ERS_RESULT_NEED_RESET;
701 	}
702 	qib_devinfo(pdev, "QIB mmio_enabled function called, "
703 		 "read wordscntr %Lx, returning %d\n", words, ret);
704 	return  ret;
705 }
706 
707 static pci_ers_result_t
708 qib_pci_slot_reset(struct pci_dev *pdev)
709 {
710 	qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
711 	return PCI_ERS_RESULT_CAN_RECOVER;
712 }
713 
714 static pci_ers_result_t
715 qib_pci_link_reset(struct pci_dev *pdev)
716 {
717 	qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
718 	return PCI_ERS_RESULT_CAN_RECOVER;
719 }
720 
721 static void
722 qib_pci_resume(struct pci_dev *pdev)
723 {
724 	struct qib_devdata *dd = pci_get_drvdata(pdev);
725 	qib_devinfo(pdev, "QIB resume function called\n");
726 	pci_cleanup_aer_uncorrect_error_status(pdev);
727 	/*
728 	 * Running jobs will fail, since it's asynchronous
729 	 * unlike sysfs-requested reset.   Better than
730 	 * doing nothing.
731 	 */
732 	qib_init(dd, 1); /* same as re-init after reset */
733 }
734 
735 struct pci_error_handlers qib_pci_err_handler = {
736 	.error_detected = qib_pci_error_detected,
737 	.mmio_enabled = qib_pci_mmio_enabled,
738 	.link_reset = qib_pci_link_reset,
739 	.slot_reset = qib_pci_slot_reset,
740 	.resume = qib_pci_resume,
741 };
742