Loading...
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/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 */
54static void qib_tune_pcie_caps(struct qib_devdata *);
55static 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 */
63int 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 = dma_set_mask_and_coherent(&pdev->dev, 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 = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
100 if (ret) {
101 qib_devinfo(pdev, "Unable to set DMA mask: %d\n", ret);
102 goto bail;
103 }
104 }
105
106 pci_set_master(pdev);
107 goto done;
108
109bail:
110 pci_disable_device(pdev);
111 pci_release_regions(pdev);
112done:
113 return ret;
114}
115
116/*
117 * Do remaining PCIe setup, once dd is allocated, and save away
118 * fields required to re-initialize after a chip reset, or for
119 * various other purposes
120 */
121int qib_pcie_ddinit(struct qib_devdata *dd, struct pci_dev *pdev,
122 const struct pci_device_id *ent)
123{
124 unsigned long len;
125 resource_size_t addr;
126
127 dd->pcidev = pdev;
128 pci_set_drvdata(pdev, dd);
129
130 addr = pci_resource_start(pdev, 0);
131 len = pci_resource_len(pdev, 0);
132
133 dd->kregbase = ioremap(addr, len);
134 if (!dd->kregbase)
135 return -ENOMEM;
136
137 dd->kregend = (u64 __iomem *)((void __iomem *) dd->kregbase + len);
138 dd->physaddr = addr; /* used for io_remap, etc. */
139
140 /*
141 * Save BARs to rewrite after device reset. Save all 64 bits of
142 * BAR, just in case.
143 */
144 dd->pcibar0 = addr;
145 dd->pcibar1 = addr >> 32;
146 dd->deviceid = ent->device; /* save for later use */
147 dd->vendorid = ent->vendor;
148
149 return 0;
150}
151
152/*
153 * Do PCIe cleanup, after chip-specific cleanup, etc. Just prior
154 * to releasing the dd memory.
155 * void because none of the core pcie cleanup returns are void
156 */
157void qib_pcie_ddcleanup(struct qib_devdata *dd)
158{
159 u64 __iomem *base = (void __iomem *) dd->kregbase;
160
161 dd->kregbase = NULL;
162 iounmap(base);
163 if (dd->piobase)
164 iounmap(dd->piobase);
165 if (dd->userbase)
166 iounmap(dd->userbase);
167 if (dd->piovl15base)
168 iounmap(dd->piovl15base);
169
170 pci_disable_device(dd->pcidev);
171 pci_release_regions(dd->pcidev);
172
173 pci_set_drvdata(dd->pcidev, NULL);
174}
175
176/*
177 * We save the msi lo and hi values, so we can restore them after
178 * chip reset (the kernel PCI infrastructure doesn't yet handle that
179 * correctly.
180 */
181static void qib_cache_msi_info(struct qib_devdata *dd, int pos)
182{
183 struct pci_dev *pdev = dd->pcidev;
184 u16 control;
185
186 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO, &dd->msi_lo);
187 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI, &dd->msi_hi);
188 pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
189
190 /* now save the data (vector) info */
191 pci_read_config_word(pdev,
192 pos + ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
193 &dd->msi_data);
194}
195
196int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent)
197{
198 u16 linkstat, speed;
199 int nvec;
200 int maxvec;
201 unsigned int flags = PCI_IRQ_MSIX | PCI_IRQ_MSI;
202
203 if (!pci_is_pcie(dd->pcidev)) {
204 qib_dev_err(dd, "Can't find PCI Express capability!\n");
205 /* set up something... */
206 dd->lbus_width = 1;
207 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
208 nvec = -1;
209 goto bail;
210 }
211
212 if (dd->flags & QIB_HAS_INTX)
213 flags |= PCI_IRQ_LEGACY;
214 maxvec = (nent && *nent) ? *nent : 1;
215 nvec = pci_alloc_irq_vectors(dd->pcidev, 1, maxvec, flags);
216 if (nvec < 0)
217 goto bail;
218
219 /*
220 * If nent exists, make sure to record how many vectors were allocated.
221 * If msix_enabled is false, return 0 so the fallback code works
222 * correctly.
223 */
224 if (nent)
225 *nent = !dd->pcidev->msix_enabled ? 0 : nvec;
226
227 if (dd->pcidev->msi_enabled)
228 qib_cache_msi_info(dd, dd->pcidev->msi_cap);
229
230 pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
231 /*
232 * speed is bits 0-3, linkwidth is bits 4-8
233 * no defines for them in headers
234 */
235 speed = linkstat & 0xf;
236 linkstat >>= 4;
237 linkstat &= 0x1f;
238 dd->lbus_width = linkstat;
239
240 switch (speed) {
241 case 1:
242 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
243 break;
244 case 2:
245 dd->lbus_speed = 5000; /* Gen1, 5GHz */
246 break;
247 default: /* not defined, assume gen1 */
248 dd->lbus_speed = 2500;
249 break;
250 }
251
252 /*
253 * Check against expected pcie width and complain if "wrong"
254 * on first initialization, not afterwards (i.e., reset).
255 */
256 if (minw && linkstat < minw)
257 qib_dev_err(dd,
258 "PCIe width %u (x%u HCA), performance reduced\n",
259 linkstat, minw);
260
261 qib_tune_pcie_caps(dd);
262
263 qib_tune_pcie_coalesce(dd);
264
265bail:
266 /* fill in string, even on errors */
267 snprintf(dd->lbus_info, sizeof(dd->lbus_info),
268 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
269 return nvec < 0 ? nvec : 0;
270}
271
272/**
273 * qib_free_irq - Cleanup INTx and MSI interrupts
274 * @dd: valid pointer to qib dev data
275 *
276 * Since cleanup for INTx and MSI interrupts is trivial, have a common
277 * routine.
278 *
279 */
280void qib_free_irq(struct qib_devdata *dd)
281{
282 pci_free_irq(dd->pcidev, 0, dd);
283 pci_free_irq_vectors(dd->pcidev);
284}
285
286/*
287 * Setup pcie interrupt stuff again after a reset. I'd like to just call
288 * pci_enable_msi() again for msi, but when I do that,
289 * the MSI enable bit doesn't get set in the command word, and
290 * we switch to a different interrupt vector, which is confusing,
291 * so I instead just do it all inline. Perhaps somehow can tie this
292 * into the PCIe hotplug support at some point
293 */
294int qib_reinit_intr(struct qib_devdata *dd)
295{
296 int pos;
297 u16 control;
298 int ret = 0;
299
300 /* If we aren't using MSI, don't restore it */
301 if (!dd->msi_lo)
302 goto bail;
303
304 pos = dd->pcidev->msi_cap;
305 if (!pos) {
306 qib_dev_err(dd,
307 "Can't find MSI capability, can't restore MSI settings\n");
308 ret = 0;
309 /* nothing special for MSIx, just MSI */
310 goto bail;
311 }
312 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
313 dd->msi_lo);
314 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
315 dd->msi_hi);
316 pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
317 if (!(control & PCI_MSI_FLAGS_ENABLE)) {
318 control |= PCI_MSI_FLAGS_ENABLE;
319 pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
320 control);
321 }
322 /* now rewrite the data (vector) info */
323 pci_write_config_word(dd->pcidev, pos +
324 ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
325 dd->msi_data);
326 ret = 1;
327bail:
328 qib_free_irq(dd);
329
330 if (!ret && (dd->flags & QIB_HAS_INTX))
331 ret = 1;
332
333 /* and now set the pci master bit again */
334 pci_set_master(dd->pcidev);
335
336 return ret;
337}
338
339/*
340 * These two routines are helper routines for the device reset code
341 * to move all the pcie code out of the chip-specific driver code.
342 */
343void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
344{
345 pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
346 pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
347 pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
348}
349
350void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
351{
352 int r;
353
354 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
355 dd->pcibar0);
356 if (r)
357 qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
358 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
359 dd->pcibar1);
360 if (r)
361 qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
362 /* now re-enable memory access, and restore cosmetic settings */
363 pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
364 pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
365 pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
366 r = pci_enable_device(dd->pcidev);
367 if (r)
368 qib_dev_err(dd,
369 "pci_enable_device failed after reset: %d\n", r);
370}
371
372
373static int qib_pcie_coalesce;
374module_param_named(pcie_coalesce, qib_pcie_coalesce, int, S_IRUGO);
375MODULE_PARM_DESC(pcie_coalesce, "tune PCIe coalescing on some Intel chipsets");
376
377/*
378 * Enable PCIe completion and data coalescing, on Intel 5x00 and 7300
379 * chipsets. This is known to be unsafe for some revisions of some
380 * of these chipsets, with some BIOS settings, and enabling it on those
381 * systems may result in the system crashing, and/or data corruption.
382 */
383static void qib_tune_pcie_coalesce(struct qib_devdata *dd)
384{
385 struct pci_dev *parent;
386 u16 devid;
387 u32 mask, bits, val;
388
389 if (!qib_pcie_coalesce)
390 return;
391
392 /* Find out supported and configured values for parent (root) */
393 parent = dd->pcidev->bus->self;
394 if (parent->bus->parent) {
395 qib_devinfo(dd->pcidev, "Parent not root\n");
396 return;
397 }
398 if (!pci_is_pcie(parent))
399 return;
400 if (parent->vendor != 0x8086)
401 return;
402
403 /*
404 * - bit 12: Max_rdcmp_Imt_EN: need to set to 1
405 * - bit 11: COALESCE_FORCE: need to set to 0
406 * - bit 10: COALESCE_EN: need to set to 1
407 * (but limitations on some on some chipsets)
408 *
409 * On the Intel 5000, 5100, and 7300 chipsets, there is
410 * also: - bit 25:24: COALESCE_MODE, need to set to 0
411 */
412 devid = parent->device;
413 if (devid >= 0x25e2 && devid <= 0x25fa) {
414 /* 5000 P/V/X/Z */
415 if (parent->revision <= 0xb2)
416 bits = 1U << 10;
417 else
418 bits = 7U << 10;
419 mask = (3U << 24) | (7U << 10);
420 } else if (devid >= 0x65e2 && devid <= 0x65fa) {
421 /* 5100 */
422 bits = 1U << 10;
423 mask = (3U << 24) | (7U << 10);
424 } else if (devid >= 0x4021 && devid <= 0x402e) {
425 /* 5400 */
426 bits = 7U << 10;
427 mask = 7U << 10;
428 } else if (devid >= 0x3604 && devid <= 0x360a) {
429 /* 7300 */
430 bits = 7U << 10;
431 mask = (3U << 24) | (7U << 10);
432 } else {
433 /* not one of the chipsets that we know about */
434 return;
435 }
436 pci_read_config_dword(parent, 0x48, &val);
437 val &= ~mask;
438 val |= bits;
439 pci_write_config_dword(parent, 0x48, val);
440}
441
442/*
443 * BIOS may not set PCIe bus-utilization parameters for best performance.
444 * Check and optionally adjust them to maximize our throughput.
445 */
446static int qib_pcie_caps;
447module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
448MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
449
450static void qib_tune_pcie_caps(struct qib_devdata *dd)
451{
452 struct pci_dev *parent;
453 u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
454 u16 rc_mrrs, ep_mrrs, max_mrrs;
455
456 /* Find out supported and configured values for parent (root) */
457 parent = dd->pcidev->bus->self;
458 if (!pci_is_root_bus(parent->bus)) {
459 qib_devinfo(dd->pcidev, "Parent not root\n");
460 return;
461 }
462
463 if (!pci_is_pcie(parent) || !pci_is_pcie(dd->pcidev))
464 return;
465
466 rc_mpss = parent->pcie_mpss;
467 rc_mps = ffs(pcie_get_mps(parent)) - 8;
468 /* Find out supported and configured values for endpoint (us) */
469 ep_mpss = dd->pcidev->pcie_mpss;
470 ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
471
472 /* Find max payload supported by root, endpoint */
473 if (rc_mpss > ep_mpss)
474 rc_mpss = ep_mpss;
475
476 /* If Supported greater than limit in module param, limit it */
477 if (rc_mpss > (qib_pcie_caps & 7))
478 rc_mpss = qib_pcie_caps & 7;
479 /* If less than (allowed, supported), bump root payload */
480 if (rc_mpss > rc_mps) {
481 rc_mps = rc_mpss;
482 pcie_set_mps(parent, 128 << rc_mps);
483 }
484 /* If less than (allowed, supported), bump endpoint payload */
485 if (rc_mpss > ep_mps) {
486 ep_mps = rc_mpss;
487 pcie_set_mps(dd->pcidev, 128 << ep_mps);
488 }
489
490 /*
491 * Now the Read Request size.
492 * No field for max supported, but PCIe spec limits it to 4096,
493 * which is code '5' (log2(4096) - 7)
494 */
495 max_mrrs = 5;
496 if (max_mrrs > ((qib_pcie_caps >> 4) & 7))
497 max_mrrs = (qib_pcie_caps >> 4) & 7;
498
499 max_mrrs = 128 << max_mrrs;
500 rc_mrrs = pcie_get_readrq(parent);
501 ep_mrrs = pcie_get_readrq(dd->pcidev);
502
503 if (max_mrrs > rc_mrrs) {
504 rc_mrrs = max_mrrs;
505 pcie_set_readrq(parent, rc_mrrs);
506 }
507 if (max_mrrs > ep_mrrs) {
508 ep_mrrs = max_mrrs;
509 pcie_set_readrq(dd->pcidev, ep_mrrs);
510 }
511}
512/* End of PCIe capability tuning */
513
514/*
515 * From here through qib_pci_err_handler definition is invoked via
516 * PCI error infrastructure, registered via pci
517 */
518static pci_ers_result_t
519qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
520{
521 struct qib_devdata *dd = pci_get_drvdata(pdev);
522 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
523
524 switch (state) {
525 case pci_channel_io_normal:
526 qib_devinfo(pdev, "State Normal, ignoring\n");
527 break;
528
529 case pci_channel_io_frozen:
530 qib_devinfo(pdev, "State Frozen, requesting reset\n");
531 pci_disable_device(pdev);
532 ret = PCI_ERS_RESULT_NEED_RESET;
533 break;
534
535 case pci_channel_io_perm_failure:
536 qib_devinfo(pdev, "State Permanent Failure, disabling\n");
537 if (dd) {
538 /* no more register accesses! */
539 dd->flags &= ~QIB_PRESENT;
540 qib_disable_after_error(dd);
541 }
542 /* else early, or other problem */
543 ret = PCI_ERS_RESULT_DISCONNECT;
544 break;
545
546 default: /* shouldn't happen */
547 qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
548 state);
549 break;
550 }
551 return ret;
552}
553
554static pci_ers_result_t
555qib_pci_mmio_enabled(struct pci_dev *pdev)
556{
557 u64 words = 0U;
558 struct qib_devdata *dd = pci_get_drvdata(pdev);
559 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
560
561 if (dd && dd->pport) {
562 words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
563 if (words == ~0ULL)
564 ret = PCI_ERS_RESULT_NEED_RESET;
565 }
566 qib_devinfo(pdev,
567 "QIB mmio_enabled function called, read wordscntr %Lx, returning %d\n",
568 words, ret);
569 return ret;
570}
571
572static pci_ers_result_t
573qib_pci_slot_reset(struct pci_dev *pdev)
574{
575 qib_devinfo(pdev, "QIB slot_reset function called, ignored\n");
576 return PCI_ERS_RESULT_CAN_RECOVER;
577}
578
579static void
580qib_pci_resume(struct pci_dev *pdev)
581{
582 struct qib_devdata *dd = pci_get_drvdata(pdev);
583
584 qib_devinfo(pdev, "QIB resume function called\n");
585 /*
586 * Running jobs will fail, since it's asynchronous
587 * unlike sysfs-requested reset. Better than
588 * doing nothing.
589 */
590 qib_init(dd, 1); /* same as re-init after reset */
591}
592
593const struct pci_error_handlers qib_pci_err_handler = {
594 .error_detected = qib_pci_error_detected,
595 .mmio_enabled = qib_pci_mmio_enabled,
596 .slot_reset = qib_pci_slot_reset,
597 .resume = qib_pci_resume,
598};
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 */
55static void qib_tune_pcie_caps(struct qib_devdata *);
56static 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 */
64int 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 = dma_set_mask_and_coherent(&pdev->dev, 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 = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
101 if (ret) {
102 qib_devinfo(pdev, "Unable to set DMA mask: %d\n", ret);
103 goto bail;
104 }
105 }
106
107 pci_set_master(pdev);
108 ret = pci_enable_pcie_error_reporting(pdev);
109 if (ret) {
110 qib_early_err(&pdev->dev,
111 "Unable to enable pcie error reporting: %d\n",
112 ret);
113 ret = 0;
114 }
115 goto done;
116
117bail:
118 pci_disable_device(pdev);
119 pci_release_regions(pdev);
120done:
121 return ret;
122}
123
124/*
125 * Do remaining PCIe setup, once dd is allocated, and save away
126 * fields required to re-initialize after a chip reset, or for
127 * various other purposes
128 */
129int qib_pcie_ddinit(struct qib_devdata *dd, struct pci_dev *pdev,
130 const struct pci_device_id *ent)
131{
132 unsigned long len;
133 resource_size_t addr;
134
135 dd->pcidev = pdev;
136 pci_set_drvdata(pdev, dd);
137
138 addr = pci_resource_start(pdev, 0);
139 len = pci_resource_len(pdev, 0);
140
141 dd->kregbase = ioremap(addr, len);
142 if (!dd->kregbase)
143 return -ENOMEM;
144
145 dd->kregend = (u64 __iomem *)((void __iomem *) dd->kregbase + len);
146 dd->physaddr = addr; /* used for io_remap, etc. */
147
148 /*
149 * Save BARs to rewrite after device reset. Save all 64 bits of
150 * BAR, just in case.
151 */
152 dd->pcibar0 = addr;
153 dd->pcibar1 = addr >> 32;
154 dd->deviceid = ent->device; /* save for later use */
155 dd->vendorid = ent->vendor;
156
157 return 0;
158}
159
160/*
161 * Do PCIe cleanup, after chip-specific cleanup, etc. Just prior
162 * to releasing the dd memory.
163 * void because none of the core pcie cleanup returns are void
164 */
165void qib_pcie_ddcleanup(struct qib_devdata *dd)
166{
167 u64 __iomem *base = (void __iomem *) dd->kregbase;
168
169 dd->kregbase = NULL;
170 iounmap(base);
171 if (dd->piobase)
172 iounmap(dd->piobase);
173 if (dd->userbase)
174 iounmap(dd->userbase);
175 if (dd->piovl15base)
176 iounmap(dd->piovl15base);
177
178 pci_disable_device(dd->pcidev);
179 pci_release_regions(dd->pcidev);
180
181 pci_set_drvdata(dd->pcidev, NULL);
182}
183
184/*
185 * We save the msi lo and hi values, so we can restore them after
186 * chip reset (the kernel PCI infrastructure doesn't yet handle that
187 * correctly.
188 */
189static void qib_cache_msi_info(struct qib_devdata *dd, int pos)
190{
191 struct pci_dev *pdev = dd->pcidev;
192 u16 control;
193
194 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO, &dd->msi_lo);
195 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI, &dd->msi_hi);
196 pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
197
198 /* now save the data (vector) info */
199 pci_read_config_word(pdev,
200 pos + ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
201 &dd->msi_data);
202}
203
204int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent)
205{
206 u16 linkstat, speed;
207 int nvec;
208 int maxvec;
209 unsigned int flags = PCI_IRQ_MSIX | PCI_IRQ_MSI;
210
211 if (!pci_is_pcie(dd->pcidev)) {
212 qib_dev_err(dd, "Can't find PCI Express capability!\n");
213 /* set up something... */
214 dd->lbus_width = 1;
215 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
216 nvec = -1;
217 goto bail;
218 }
219
220 if (dd->flags & QIB_HAS_INTX)
221 flags |= PCI_IRQ_LEGACY;
222 maxvec = (nent && *nent) ? *nent : 1;
223 nvec = pci_alloc_irq_vectors(dd->pcidev, 1, maxvec, flags);
224 if (nvec < 0)
225 goto bail;
226
227 /*
228 * If nent exists, make sure to record how many vectors were allocated.
229 * If msix_enabled is false, return 0 so the fallback code works
230 * correctly.
231 */
232 if (nent)
233 *nent = !dd->pcidev->msix_enabled ? 0 : nvec;
234
235 if (dd->pcidev->msi_enabled)
236 qib_cache_msi_info(dd, dd->pcidev->msi_cap);
237
238 pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
239 /*
240 * speed is bits 0-3, linkwidth is bits 4-8
241 * no defines for them in headers
242 */
243 speed = linkstat & 0xf;
244 linkstat >>= 4;
245 linkstat &= 0x1f;
246 dd->lbus_width = linkstat;
247
248 switch (speed) {
249 case 1:
250 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
251 break;
252 case 2:
253 dd->lbus_speed = 5000; /* Gen1, 5GHz */
254 break;
255 default: /* not defined, assume gen1 */
256 dd->lbus_speed = 2500;
257 break;
258 }
259
260 /*
261 * Check against expected pcie width and complain if "wrong"
262 * on first initialization, not afterwards (i.e., reset).
263 */
264 if (minw && linkstat < minw)
265 qib_dev_err(dd,
266 "PCIe width %u (x%u HCA), performance reduced\n",
267 linkstat, minw);
268
269 qib_tune_pcie_caps(dd);
270
271 qib_tune_pcie_coalesce(dd);
272
273bail:
274 /* fill in string, even on errors */
275 snprintf(dd->lbus_info, sizeof(dd->lbus_info),
276 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
277 return nvec < 0 ? nvec : 0;
278}
279
280/**
281 * qib_free_irq - Cleanup INTx and MSI interrupts
282 * @dd: valid pointer to qib dev data
283 *
284 * Since cleanup for INTx and MSI interrupts is trivial, have a common
285 * routine.
286 *
287 */
288void qib_free_irq(struct qib_devdata *dd)
289{
290 pci_free_irq(dd->pcidev, 0, dd);
291 pci_free_irq_vectors(dd->pcidev);
292}
293
294/*
295 * Setup pcie interrupt stuff again after a reset. I'd like to just call
296 * pci_enable_msi() again for msi, but when I do that,
297 * the MSI enable bit doesn't get set in the command word, and
298 * we switch to a different interrupt vector, which is confusing,
299 * so I instead just do it all inline. Perhaps somehow can tie this
300 * into the PCIe hotplug support at some point
301 */
302int qib_reinit_intr(struct qib_devdata *dd)
303{
304 int pos;
305 u16 control;
306 int ret = 0;
307
308 /* If we aren't using MSI, don't restore it */
309 if (!dd->msi_lo)
310 goto bail;
311
312 pos = dd->pcidev->msi_cap;
313 if (!pos) {
314 qib_dev_err(dd,
315 "Can't find MSI capability, can't restore MSI settings\n");
316 ret = 0;
317 /* nothing special for MSIx, just MSI */
318 goto bail;
319 }
320 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
321 dd->msi_lo);
322 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
323 dd->msi_hi);
324 pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
325 if (!(control & PCI_MSI_FLAGS_ENABLE)) {
326 control |= PCI_MSI_FLAGS_ENABLE;
327 pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
328 control);
329 }
330 /* now rewrite the data (vector) info */
331 pci_write_config_word(dd->pcidev, pos +
332 ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
333 dd->msi_data);
334 ret = 1;
335bail:
336 qib_free_irq(dd);
337
338 if (!ret && (dd->flags & QIB_HAS_INTX))
339 ret = 1;
340
341 /* and now set the pci master bit again */
342 pci_set_master(dd->pcidev);
343
344 return ret;
345}
346
347/*
348 * These two routines are helper routines for the device reset code
349 * to move all the pcie code out of the chip-specific driver code.
350 */
351void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
352{
353 pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
354 pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
355 pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
356}
357
358void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
359{
360 int r;
361
362 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
363 dd->pcibar0);
364 if (r)
365 qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
366 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
367 dd->pcibar1);
368 if (r)
369 qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
370 /* now re-enable memory access, and restore cosmetic settings */
371 pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
372 pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
373 pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
374 r = pci_enable_device(dd->pcidev);
375 if (r)
376 qib_dev_err(dd,
377 "pci_enable_device failed after reset: %d\n", r);
378}
379
380
381static int qib_pcie_coalesce;
382module_param_named(pcie_coalesce, qib_pcie_coalesce, int, S_IRUGO);
383MODULE_PARM_DESC(pcie_coalesce, "tune PCIe coalescing on some Intel chipsets");
384
385/*
386 * Enable PCIe completion and data coalescing, on Intel 5x00 and 7300
387 * chipsets. This is known to be unsafe for some revisions of some
388 * of these chipsets, with some BIOS settings, and enabling it on those
389 * systems may result in the system crashing, and/or data corruption.
390 */
391static void qib_tune_pcie_coalesce(struct qib_devdata *dd)
392{
393 struct pci_dev *parent;
394 u16 devid;
395 u32 mask, bits, val;
396
397 if (!qib_pcie_coalesce)
398 return;
399
400 /* Find out supported and configured values for parent (root) */
401 parent = dd->pcidev->bus->self;
402 if (parent->bus->parent) {
403 qib_devinfo(dd->pcidev, "Parent not root\n");
404 return;
405 }
406 if (!pci_is_pcie(parent))
407 return;
408 if (parent->vendor != 0x8086)
409 return;
410
411 /*
412 * - bit 12: Max_rdcmp_Imt_EN: need to set to 1
413 * - bit 11: COALESCE_FORCE: need to set to 0
414 * - bit 10: COALESCE_EN: need to set to 1
415 * (but limitations on some on some chipsets)
416 *
417 * On the Intel 5000, 5100, and 7300 chipsets, there is
418 * also: - bit 25:24: COALESCE_MODE, need to set to 0
419 */
420 devid = parent->device;
421 if (devid >= 0x25e2 && devid <= 0x25fa) {
422 /* 5000 P/V/X/Z */
423 if (parent->revision <= 0xb2)
424 bits = 1U << 10;
425 else
426 bits = 7U << 10;
427 mask = (3U << 24) | (7U << 10);
428 } else if (devid >= 0x65e2 && devid <= 0x65fa) {
429 /* 5100 */
430 bits = 1U << 10;
431 mask = (3U << 24) | (7U << 10);
432 } else if (devid >= 0x4021 && devid <= 0x402e) {
433 /* 5400 */
434 bits = 7U << 10;
435 mask = 7U << 10;
436 } else if (devid >= 0x3604 && devid <= 0x360a) {
437 /* 7300 */
438 bits = 7U << 10;
439 mask = (3U << 24) | (7U << 10);
440 } else {
441 /* not one of the chipsets that we know about */
442 return;
443 }
444 pci_read_config_dword(parent, 0x48, &val);
445 val &= ~mask;
446 val |= bits;
447 pci_write_config_dword(parent, 0x48, val);
448}
449
450/*
451 * BIOS may not set PCIe bus-utilization parameters for best performance.
452 * Check and optionally adjust them to maximize our throughput.
453 */
454static int qib_pcie_caps;
455module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
456MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
457
458static void qib_tune_pcie_caps(struct qib_devdata *dd)
459{
460 struct pci_dev *parent;
461 u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
462 u16 rc_mrrs, ep_mrrs, max_mrrs;
463
464 /* Find out supported and configured values for parent (root) */
465 parent = dd->pcidev->bus->self;
466 if (!pci_is_root_bus(parent->bus)) {
467 qib_devinfo(dd->pcidev, "Parent not root\n");
468 return;
469 }
470
471 if (!pci_is_pcie(parent) || !pci_is_pcie(dd->pcidev))
472 return;
473
474 rc_mpss = parent->pcie_mpss;
475 rc_mps = ffs(pcie_get_mps(parent)) - 8;
476 /* Find out supported and configured values for endpoint (us) */
477 ep_mpss = dd->pcidev->pcie_mpss;
478 ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
479
480 /* Find max payload supported by root, endpoint */
481 if (rc_mpss > ep_mpss)
482 rc_mpss = ep_mpss;
483
484 /* If Supported greater than limit in module param, limit it */
485 if (rc_mpss > (qib_pcie_caps & 7))
486 rc_mpss = qib_pcie_caps & 7;
487 /* If less than (allowed, supported), bump root payload */
488 if (rc_mpss > rc_mps) {
489 rc_mps = rc_mpss;
490 pcie_set_mps(parent, 128 << rc_mps);
491 }
492 /* If less than (allowed, supported), bump endpoint payload */
493 if (rc_mpss > ep_mps) {
494 ep_mps = rc_mpss;
495 pcie_set_mps(dd->pcidev, 128 << ep_mps);
496 }
497
498 /*
499 * Now the Read Request size.
500 * No field for max supported, but PCIe spec limits it to 4096,
501 * which is code '5' (log2(4096) - 7)
502 */
503 max_mrrs = 5;
504 if (max_mrrs > ((qib_pcie_caps >> 4) & 7))
505 max_mrrs = (qib_pcie_caps >> 4) & 7;
506
507 max_mrrs = 128 << max_mrrs;
508 rc_mrrs = pcie_get_readrq(parent);
509 ep_mrrs = pcie_get_readrq(dd->pcidev);
510
511 if (max_mrrs > rc_mrrs) {
512 rc_mrrs = max_mrrs;
513 pcie_set_readrq(parent, rc_mrrs);
514 }
515 if (max_mrrs > ep_mrrs) {
516 ep_mrrs = max_mrrs;
517 pcie_set_readrq(dd->pcidev, ep_mrrs);
518 }
519}
520/* End of PCIe capability tuning */
521
522/*
523 * From here through qib_pci_err_handler definition is invoked via
524 * PCI error infrastructure, registered via pci
525 */
526static pci_ers_result_t
527qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
528{
529 struct qib_devdata *dd = pci_get_drvdata(pdev);
530 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
531
532 switch (state) {
533 case pci_channel_io_normal:
534 qib_devinfo(pdev, "State Normal, ignoring\n");
535 break;
536
537 case pci_channel_io_frozen:
538 qib_devinfo(pdev, "State Frozen, requesting reset\n");
539 pci_disable_device(pdev);
540 ret = PCI_ERS_RESULT_NEED_RESET;
541 break;
542
543 case pci_channel_io_perm_failure:
544 qib_devinfo(pdev, "State Permanent Failure, disabling\n");
545 if (dd) {
546 /* no more register accesses! */
547 dd->flags &= ~QIB_PRESENT;
548 qib_disable_after_error(dd);
549 }
550 /* else early, or other problem */
551 ret = PCI_ERS_RESULT_DISCONNECT;
552 break;
553
554 default: /* shouldn't happen */
555 qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
556 state);
557 break;
558 }
559 return ret;
560}
561
562static pci_ers_result_t
563qib_pci_mmio_enabled(struct pci_dev *pdev)
564{
565 u64 words = 0U;
566 struct qib_devdata *dd = pci_get_drvdata(pdev);
567 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
568
569 if (dd && dd->pport) {
570 words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
571 if (words == ~0ULL)
572 ret = PCI_ERS_RESULT_NEED_RESET;
573 }
574 qib_devinfo(pdev,
575 "QIB mmio_enabled function called, read wordscntr %Lx, returning %d\n",
576 words, ret);
577 return ret;
578}
579
580static pci_ers_result_t
581qib_pci_slot_reset(struct pci_dev *pdev)
582{
583 qib_devinfo(pdev, "QIB slot_reset function called, ignored\n");
584 return PCI_ERS_RESULT_CAN_RECOVER;
585}
586
587static void
588qib_pci_resume(struct pci_dev *pdev)
589{
590 struct qib_devdata *dd = pci_get_drvdata(pdev);
591
592 qib_devinfo(pdev, "QIB resume function called\n");
593 /*
594 * Running jobs will fail, since it's asynchronous
595 * unlike sysfs-requested reset. Better than
596 * doing nothing.
597 */
598 qib_init(dd, 1); /* same as re-init after reset */
599}
600
601const struct pci_error_handlers qib_pci_err_handler = {
602 .error_detected = qib_pci_error_detected,
603 .mmio_enabled = qib_pci_mmio_enabled,
604 .slot_reset = qib_pci_slot_reset,
605 .resume = qib_pci_resume,
606};