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 int qib_tune_pcie_caps(struct qib_devdata *);
55 static int 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 	u32 tabsize = 0;
201 	u16 msix_flags;
202 	struct msix_entry *msix_entry;
203 	int i;
204 
205 	/* We can't pass qib_msix_entry array to qib_msix_setup
206 	 * so use a dummy msix_entry array and copy the allocated
207 	 * irq back to the qib_msix_entry array. */
208 	msix_entry = kmalloc(*msixcnt * sizeof(*msix_entry), GFP_KERNEL);
209 	if (!msix_entry) {
210 		ret = -ENOMEM;
211 		goto do_intx;
212 	}
213 	for (i = 0; i < *msixcnt; i++)
214 		msix_entry[i] = qib_msix_entry[i].msix;
215 
216 	pci_read_config_word(dd->pcidev, pos + PCI_MSIX_FLAGS, &msix_flags);
217 	tabsize = 1 + (msix_flags & PCI_MSIX_FLAGS_QSIZE);
218 	if (tabsize > *msixcnt)
219 		tabsize = *msixcnt;
220 	ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
221 	if (ret > 0) {
222 		tabsize = ret;
223 		ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
224 	}
225 do_intx:
226 	if (ret) {
227 		qib_dev_err(dd, "pci_enable_msix %d vectors failed: %d, "
228 			    "falling back to INTx\n", tabsize, ret);
229 		tabsize = 0;
230 	}
231 	for (i = 0; i < tabsize; i++)
232 		qib_msix_entry[i].msix = msix_entry[i];
233 	kfree(msix_entry);
234 	*msixcnt = tabsize;
235 
236 	if (ret)
237 		qib_enable_intx(dd->pcidev);
238 
239 }
240 
241 /**
242  * We save the msi lo and hi values, so we can restore them after
243  * chip reset (the kernel PCI infrastructure doesn't yet handle that
244  * correctly.
245  */
246 static int qib_msi_setup(struct qib_devdata *dd, int pos)
247 {
248 	struct pci_dev *pdev = dd->pcidev;
249 	u16 control;
250 	int ret;
251 
252 	ret = pci_enable_msi(pdev);
253 	if (ret)
254 		qib_dev_err(dd, "pci_enable_msi failed: %d, "
255 			    "interrupts may not work\n", ret);
256 	/* continue even if it fails, we may still be OK... */
257 
258 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO,
259 			      &dd->msi_lo);
260 	pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI,
261 			      &dd->msi_hi);
262 	pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
263 	/* now save the data (vector) info */
264 	pci_read_config_word(pdev, pos + ((control & PCI_MSI_FLAGS_64BIT)
265 				    ? 12 : 8),
266 			     &dd->msi_data);
267 	return ret;
268 }
269 
270 int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent,
271 		    struct qib_msix_entry *entry)
272 {
273 	u16 linkstat, speed;
274 	int pos = 0, pose, ret = 1;
275 
276 	pose = pci_pcie_cap(dd->pcidev);
277 	if (!pose) {
278 		qib_dev_err(dd, "Can't find PCI Express capability!\n");
279 		/* set up something... */
280 		dd->lbus_width = 1;
281 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
282 		goto bail;
283 	}
284 
285 	pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSIX);
286 	if (nent && *nent && pos) {
287 		qib_msix_setup(dd, pos, nent, entry);
288 		ret = 0; /* did it, either MSIx or INTx */
289 	} else {
290 		pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
291 		if (pos)
292 			ret = qib_msi_setup(dd, pos);
293 		else
294 			qib_dev_err(dd, "No PCI MSI or MSIx capability!\n");
295 	}
296 	if (!pos)
297 		qib_enable_intx(dd->pcidev);
298 
299 	pci_read_config_word(dd->pcidev, pose + PCI_EXP_LNKSTA, &linkstat);
300 	/*
301 	 * speed is bits 0-3, linkwidth is bits 4-8
302 	 * no defines for them in headers
303 	 */
304 	speed = linkstat & 0xf;
305 	linkstat >>= 4;
306 	linkstat &= 0x1f;
307 	dd->lbus_width = linkstat;
308 
309 	switch (speed) {
310 	case 1:
311 		dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
312 		break;
313 	case 2:
314 		dd->lbus_speed = 5000; /* Gen1, 5GHz */
315 		break;
316 	default: /* not defined, assume gen1 */
317 		dd->lbus_speed = 2500;
318 		break;
319 	}
320 
321 	/*
322 	 * Check against expected pcie width and complain if "wrong"
323 	 * on first initialization, not afterwards (i.e., reset).
324 	 */
325 	if (minw && linkstat < minw)
326 		qib_dev_err(dd,
327 			    "PCIe width %u (x%u HCA), performance reduced\n",
328 			    linkstat, minw);
329 
330 	qib_tune_pcie_caps(dd);
331 
332 	qib_tune_pcie_coalesce(dd);
333 
334 bail:
335 	/* fill in string, even on errors */
336 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
337 		 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
338 	return ret;
339 }
340 
341 /*
342  * Setup pcie interrupt stuff again after a reset.  I'd like to just call
343  * pci_enable_msi() again for msi, but when I do that,
344  * the MSI enable bit doesn't get set in the command word, and
345  * we switch to to a different interrupt vector, which is confusing,
346  * so I instead just do it all inline.  Perhaps somehow can tie this
347  * into the PCIe hotplug support at some point
348  */
349 int qib_reinit_intr(struct qib_devdata *dd)
350 {
351 	int pos;
352 	u16 control;
353 	int ret = 0;
354 
355 	/* If we aren't using MSI, don't restore it */
356 	if (!dd->msi_lo)
357 		goto bail;
358 
359 	pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
360 	if (!pos) {
361 		qib_dev_err(dd, "Can't find MSI capability, "
362 			    "can't restore MSI settings\n");
363 		ret = 0;
364 		/* nothing special for MSIx, just MSI */
365 		goto bail;
366 	}
367 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
368 			       dd->msi_lo);
369 	pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
370 			       dd->msi_hi);
371 	pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
372 	if (!(control & PCI_MSI_FLAGS_ENABLE)) {
373 		control |= PCI_MSI_FLAGS_ENABLE;
374 		pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
375 				      control);
376 	}
377 	/* now rewrite the data (vector) info */
378 	pci_write_config_word(dd->pcidev, pos +
379 			      ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
380 			      dd->msi_data);
381 	ret = 1;
382 bail:
383 	if (!ret && (dd->flags & QIB_HAS_INTX)) {
384 		qib_enable_intx(dd->pcidev);
385 		ret = 1;
386 	}
387 
388 	/* and now set the pci master bit again */
389 	pci_set_master(dd->pcidev);
390 
391 	return ret;
392 }
393 
394 /*
395  * Disable msi interrupt if enabled, and clear msi_lo.
396  * This is used primarily for the fallback to INTx, but
397  * is also used in reinit after reset, and during cleanup.
398  */
399 void qib_nomsi(struct qib_devdata *dd)
400 {
401 	dd->msi_lo = 0;
402 	pci_disable_msi(dd->pcidev);
403 }
404 
405 /*
406  * Same as qib_nosmi, but for MSIx.
407  */
408 void qib_nomsix(struct qib_devdata *dd)
409 {
410 	pci_disable_msix(dd->pcidev);
411 }
412 
413 /*
414  * Similar to pci_intx(pdev, 1), except that we make sure
415  * msi(x) is off.
416  */
417 void qib_enable_intx(struct pci_dev *pdev)
418 {
419 	u16 cw, new;
420 	int pos;
421 
422 	/* first, turn on INTx */
423 	pci_read_config_word(pdev, PCI_COMMAND, &cw);
424 	new = cw & ~PCI_COMMAND_INTX_DISABLE;
425 	if (new != cw)
426 		pci_write_config_word(pdev, PCI_COMMAND, new);
427 
428 	pos = pci_find_capability(pdev, PCI_CAP_ID_MSI);
429 	if (pos) {
430 		/* then turn off MSI */
431 		pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &cw);
432 		new = cw & ~PCI_MSI_FLAGS_ENABLE;
433 		if (new != cw)
434 			pci_write_config_word(pdev, pos + PCI_MSI_FLAGS, new);
435 	}
436 	pos = pci_find_capability(pdev, PCI_CAP_ID_MSIX);
437 	if (pos) {
438 		/* then turn off MSIx */
439 		pci_read_config_word(pdev, pos + PCI_MSIX_FLAGS, &cw);
440 		new = cw & ~PCI_MSIX_FLAGS_ENABLE;
441 		if (new != cw)
442 			pci_write_config_word(pdev, pos + PCI_MSIX_FLAGS, new);
443 	}
444 }
445 
446 /*
447  * These two routines are helper routines for the device reset code
448  * to move all the pcie code out of the chip-specific driver code.
449  */
450 void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
451 {
452 	pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
453 	pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
454 	pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
455 }
456 
457 void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
458 {
459 	int r;
460 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
461 				   dd->pcibar0);
462 	if (r)
463 		qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
464 	r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
465 				   dd->pcibar1);
466 	if (r)
467 		qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
468 	/* now re-enable memory access, and restore cosmetic settings */
469 	pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
470 	pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
471 	pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
472 	r = pci_enable_device(dd->pcidev);
473 	if (r)
474 		qib_dev_err(dd, "pci_enable_device failed after "
475 			    "reset: %d\n", r);
476 }
477 
478 /* code to adjust PCIe capabilities. */
479 
480 static int fld2val(int wd, int mask)
481 {
482 	int lsbmask;
483 
484 	if (!mask)
485 		return 0;
486 	wd &= mask;
487 	lsbmask = mask ^ (mask & (mask - 1));
488 	wd /= lsbmask;
489 	return wd;
490 }
491 
492 static int val2fld(int wd, int mask)
493 {
494 	int lsbmask;
495 
496 	if (!mask)
497 		return 0;
498 	lsbmask = mask ^ (mask & (mask - 1));
499 	wd *= lsbmask;
500 	return wd;
501 }
502 
503 static int qib_pcie_coalesce;
504 module_param_named(pcie_coalesce, qib_pcie_coalesce, int, S_IRUGO);
505 MODULE_PARM_DESC(pcie_coalesce, "tune PCIe colescing on some Intel chipsets");
506 
507 /*
508  * Enable PCIe completion and data coalescing, on Intel 5x00 and 7300
509  * chipsets.   This is known to be unsafe for some revisions of some
510  * of these chipsets, with some BIOS settings, and enabling it on those
511  * systems may result in the system crashing, and/or data corruption.
512  */
513 static int qib_tune_pcie_coalesce(struct qib_devdata *dd)
514 {
515 	int r;
516 	struct pci_dev *parent;
517 	int ppos;
518 	u16 devid;
519 	u32 mask, bits, val;
520 
521 	if (!qib_pcie_coalesce)
522 		return 0;
523 
524 	/* Find out supported and configured values for parent (root) */
525 	parent = dd->pcidev->bus->self;
526 	if (parent->bus->parent) {
527 		qib_devinfo(dd->pcidev, "Parent not root\n");
528 		return 1;
529 	}
530 	ppos = pci_pcie_cap(parent);
531 	if (!ppos)
532 		return 1;
533 	if (parent->vendor != 0x8086)
534 		return 1;
535 
536 	/*
537 	 *  - bit 12: Max_rdcmp_Imt_EN: need to set to 1
538 	 *  - bit 11: COALESCE_FORCE: need to set to 0
539 	 *  - bit 10: COALESCE_EN: need to set to 1
540 	 *  (but limitations on some on some chipsets)
541 	 *
542 	 *  On the Intel 5000, 5100, and 7300 chipsets, there is
543 	 *  also: - bit 25:24: COALESCE_MODE, need to set to 0
544 	 */
545 	devid = parent->device;
546 	if (devid >= 0x25e2 && devid <= 0x25fa) {
547 		/* 5000 P/V/X/Z */
548 		if (parent->revision <= 0xb2)
549 			bits = 1U << 10;
550 		else
551 			bits = 7U << 10;
552 		mask = (3U << 24) | (7U << 10);
553 	} else if (devid >= 0x65e2 && devid <= 0x65fa) {
554 		/* 5100 */
555 		bits = 1U << 10;
556 		mask = (3U << 24) | (7U << 10);
557 	} else if (devid >= 0x4021 && devid <= 0x402e) {
558 		/* 5400 */
559 		bits = 7U << 10;
560 		mask = 7U << 10;
561 	} else if (devid >= 0x3604 && devid <= 0x360a) {
562 		/* 7300 */
563 		bits = 7U << 10;
564 		mask = (3U << 24) | (7U << 10);
565 	} else {
566 		/* not one of the chipsets that we know about */
567 		return 1;
568 	}
569 	pci_read_config_dword(parent, 0x48, &val);
570 	val &= ~mask;
571 	val |= bits;
572 	r = pci_write_config_dword(parent, 0x48, val);
573 	return 0;
574 }
575 
576 /*
577  * BIOS may not set PCIe bus-utilization parameters for best performance.
578  * Check and optionally adjust them to maximize our throughput.
579  */
580 static int qib_pcie_caps;
581 module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
582 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
583 
584 static int qib_tune_pcie_caps(struct qib_devdata *dd)
585 {
586 	int ret = 1; /* Assume the worst */
587 	struct pci_dev *parent;
588 	int ppos, epos;
589 	u16 pcaps, pctl, ecaps, ectl;
590 	int rc_sup, ep_sup;
591 	int rc_cur, ep_cur;
592 
593 	/* Find out supported and configured values for parent (root) */
594 	parent = dd->pcidev->bus->self;
595 	if (parent->bus->parent) {
596 		qib_devinfo(dd->pcidev, "Parent not root\n");
597 		goto bail;
598 	}
599 	ppos = pci_pcie_cap(parent);
600 	if (ppos) {
601 		pci_read_config_word(parent, ppos + PCI_EXP_DEVCAP, &pcaps);
602 		pci_read_config_word(parent, ppos + PCI_EXP_DEVCTL, &pctl);
603 	} else
604 		goto bail;
605 	/* Find out supported and configured values for endpoint (us) */
606 	epos = pci_pcie_cap(dd->pcidev);
607 	if (epos) {
608 		pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCAP, &ecaps);
609 		pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, &ectl);
610 	} else
611 		goto bail;
612 	ret = 0;
613 	/* Find max payload supported by root, endpoint */
614 	rc_sup = fld2val(pcaps, PCI_EXP_DEVCAP_PAYLOAD);
615 	ep_sup = fld2val(ecaps, PCI_EXP_DEVCAP_PAYLOAD);
616 	if (rc_sup > ep_sup)
617 		rc_sup = ep_sup;
618 
619 	rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_PAYLOAD);
620 	ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_PAYLOAD);
621 
622 	/* If Supported greater than limit in module param, limit it */
623 	if (rc_sup > (qib_pcie_caps & 7))
624 		rc_sup = qib_pcie_caps & 7;
625 	/* If less than (allowed, supported), bump root payload */
626 	if (rc_sup > rc_cur) {
627 		rc_cur = rc_sup;
628 		pctl = (pctl & ~PCI_EXP_DEVCTL_PAYLOAD) |
629 			val2fld(rc_cur, PCI_EXP_DEVCTL_PAYLOAD);
630 		pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
631 	}
632 	/* If less than (allowed, supported), bump endpoint payload */
633 	if (rc_sup > ep_cur) {
634 		ep_cur = rc_sup;
635 		ectl = (ectl & ~PCI_EXP_DEVCTL_PAYLOAD) |
636 			val2fld(ep_cur, PCI_EXP_DEVCTL_PAYLOAD);
637 		pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
638 	}
639 
640 	/*
641 	 * Now the Read Request size.
642 	 * No field for max supported, but PCIe spec limits it to 4096,
643 	 * which is code '5' (log2(4096) - 7)
644 	 */
645 	rc_sup = 5;
646 	if (rc_sup > ((qib_pcie_caps >> 4) & 7))
647 		rc_sup = (qib_pcie_caps >> 4) & 7;
648 	rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_READRQ);
649 	ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_READRQ);
650 
651 	if (rc_sup > rc_cur) {
652 		rc_cur = rc_sup;
653 		pctl = (pctl & ~PCI_EXP_DEVCTL_READRQ) |
654 			val2fld(rc_cur, PCI_EXP_DEVCTL_READRQ);
655 		pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
656 	}
657 	if (rc_sup > ep_cur) {
658 		ep_cur = rc_sup;
659 		ectl = (ectl & ~PCI_EXP_DEVCTL_READRQ) |
660 			val2fld(ep_cur, PCI_EXP_DEVCTL_READRQ);
661 		pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
662 	}
663 bail:
664 	return ret;
665 }
666 /* End of PCIe capability tuning */
667 
668 /*
669  * From here through qib_pci_err_handler definition is invoked via
670  * PCI error infrastructure, registered via pci
671  */
672 static pci_ers_result_t
673 qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
674 {
675 	struct qib_devdata *dd = pci_get_drvdata(pdev);
676 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
677 
678 	switch (state) {
679 	case pci_channel_io_normal:
680 		qib_devinfo(pdev, "State Normal, ignoring\n");
681 		break;
682 
683 	case pci_channel_io_frozen:
684 		qib_devinfo(pdev, "State Frozen, requesting reset\n");
685 		pci_disable_device(pdev);
686 		ret = PCI_ERS_RESULT_NEED_RESET;
687 		break;
688 
689 	case pci_channel_io_perm_failure:
690 		qib_devinfo(pdev, "State Permanent Failure, disabling\n");
691 		if (dd) {
692 			/* no more register accesses! */
693 			dd->flags &= ~QIB_PRESENT;
694 			qib_disable_after_error(dd);
695 		}
696 		 /* else early, or other problem */
697 		ret =  PCI_ERS_RESULT_DISCONNECT;
698 		break;
699 
700 	default: /* shouldn't happen */
701 		qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
702 			state);
703 		break;
704 	}
705 	return ret;
706 }
707 
708 static pci_ers_result_t
709 qib_pci_mmio_enabled(struct pci_dev *pdev)
710 {
711 	u64 words = 0U;
712 	struct qib_devdata *dd = pci_get_drvdata(pdev);
713 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
714 
715 	if (dd && dd->pport) {
716 		words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
717 		if (words == ~0ULL)
718 			ret = PCI_ERS_RESULT_NEED_RESET;
719 	}
720 	qib_devinfo(pdev, "QIB mmio_enabled function called, "
721 		 "read wordscntr %Lx, returning %d\n", words, ret);
722 	return  ret;
723 }
724 
725 static pci_ers_result_t
726 qib_pci_slot_reset(struct pci_dev *pdev)
727 {
728 	qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
729 	return PCI_ERS_RESULT_CAN_RECOVER;
730 }
731 
732 static pci_ers_result_t
733 qib_pci_link_reset(struct pci_dev *pdev)
734 {
735 	qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
736 	return PCI_ERS_RESULT_CAN_RECOVER;
737 }
738 
739 static void
740 qib_pci_resume(struct pci_dev *pdev)
741 {
742 	struct qib_devdata *dd = pci_get_drvdata(pdev);
743 	qib_devinfo(pdev, "QIB resume function called\n");
744 	pci_cleanup_aer_uncorrect_error_status(pdev);
745 	/*
746 	 * Running jobs will fail, since it's asynchronous
747 	 * unlike sysfs-requested reset.   Better than
748 	 * doing nothing.
749 	 */
750 	qib_init(dd, 1); /* same as re-init after reset */
751 }
752 
753 struct pci_error_handlers qib_pci_err_handler = {
754 	.error_detected = qib_pci_error_detected,
755 	.mmio_enabled = qib_pci_mmio_enabled,
756 	.link_reset = qib_pci_link_reset,
757 	.slot_reset = qib_pci_slot_reset,
758 	.resume = qib_pci_resume,
759 };
760