2 * Copyright (c) 2008, 2009 QLogic Corporation. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <linux/pci.h>
35 #include <linux/delay.h>
36 #include <linux/vmalloc.h>
37 #include <linux/aer.h>
42 * This file contains PCIe utility routines that are common to the
43 * various QLogic InfiniPath adapters
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.
53 static int qib_tune_pcie_caps(struct qib_devdata *);
54 static int qib_tune_pcie_coalesce(struct qib_devdata *);
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
62 int qib_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
66 ret = pci_enable_device(pdev);
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.
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.
80 qib_early_err(&pdev->dev, "pci enable failed: error %d\n",
85 ret = pci_request_regions(pdev, QIB_DRV_NAME);
87 qib_devinfo(pdev, "pci_request_regions fails: err %d\n", -ret);
91 ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
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
98 ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
100 qib_devinfo(pdev, "Unable to set DMA mask: %d\n", ret);
103 ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
105 ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
107 qib_early_err(&pdev->dev,
108 "Unable to set DMA consistent mask: %d\n", ret);
110 pci_set_master(pdev);
111 ret = pci_enable_pcie_error_reporting(pdev);
113 qib_early_err(&pdev->dev,
114 "Unable to enable pcie error reporting: %d\n",
119 pci_disable_device(pdev);
120 pci_release_regions(pdev);
126 * Do remaining PCIe setup, once dd is allocated, and save away
127 * fields required to re-initialize after a chip reset, or for
128 * various other purposes
130 int qib_pcie_ddinit(struct qib_devdata *dd, struct pci_dev *pdev,
131 const struct pci_device_id *ent)
134 resource_size_t addr;
137 pci_set_drvdata(pdev, dd);
139 addr = pci_resource_start(pdev, 0);
140 len = pci_resource_len(pdev, 0);
142 #if defined(__powerpc__)
143 /* There isn't a generic way to specify writethrough mappings */
144 dd->kregbase = __ioremap(addr, len, _PAGE_NO_CACHE | _PAGE_WRITETHRU);
146 dd->kregbase = ioremap_nocache(addr, len);
152 dd->kregend = (u64 __iomem *)((void __iomem *) dd->kregbase + len);
153 dd->physaddr = addr; /* used for io_remap, etc. */
156 * Save BARs to rewrite after device reset. Save all 64 bits of
160 dd->pcibar1 = addr >> 32;
161 dd->deviceid = ent->device; /* save for later use */
162 dd->vendorid = ent->vendor;
168 * Do PCIe cleanup, after chip-specific cleanup, etc. Just prior
169 * to releasing the dd memory.
170 * void because none of the core pcie cleanup returns are void
172 void qib_pcie_ddcleanup(struct qib_devdata *dd)
174 u64 __iomem *base = (void __iomem *) dd->kregbase;
179 iounmap(dd->piobase);
181 iounmap(dd->userbase);
183 iounmap(dd->piovl15base);
185 pci_disable_device(dd->pcidev);
186 pci_release_regions(dd->pcidev);
188 pci_set_drvdata(dd->pcidev, NULL);
191 static void qib_msix_setup(struct qib_devdata *dd, int pos, u32 *msixcnt,
192 struct msix_entry *msix_entry)
198 pci_read_config_word(dd->pcidev, pos + PCI_MSIX_FLAGS, &msix_flags);
199 tabsize = 1 + (msix_flags & PCI_MSIX_FLAGS_QSIZE);
200 if (tabsize > *msixcnt)
202 ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
205 ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
208 qib_dev_err(dd, "pci_enable_msix %d vectors failed: %d, "
209 "falling back to INTx\n", tabsize, ret);
215 qib_enable_intx(dd->pcidev);
220 * We save the msi lo and hi values, so we can restore them after
221 * chip reset (the kernel PCI infrastructure doesn't yet handle that
224 static int qib_msi_setup(struct qib_devdata *dd, int pos)
226 struct pci_dev *pdev = dd->pcidev;
230 ret = pci_enable_msi(pdev);
232 qib_dev_err(dd, "pci_enable_msi failed: %d, "
233 "interrupts may not work\n", ret);
234 /* continue even if it fails, we may still be OK... */
236 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO,
238 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI,
240 pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
241 /* now save the data (vector) info */
242 pci_read_config_word(pdev, pos + ((control & PCI_MSI_FLAGS_64BIT)
248 int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent,
249 struct msix_entry *entry)
252 int pos = 0, pose, ret = 1;
254 pose = pci_find_capability(dd->pcidev, PCI_CAP_ID_EXP);
256 qib_dev_err(dd, "Can't find PCI Express capability!\n");
257 /* set up something... */
259 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
263 pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSIX);
264 if (nent && *nent && pos) {
265 qib_msix_setup(dd, pos, nent, entry);
266 ret = 0; /* did it, either MSIx or INTx */
268 pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
270 ret = qib_msi_setup(dd, pos);
272 qib_dev_err(dd, "No PCI MSI or MSIx capability!\n");
275 qib_enable_intx(dd->pcidev);
277 pci_read_config_word(dd->pcidev, pose + PCI_EXP_LNKSTA, &linkstat);
279 * speed is bits 0-3, linkwidth is bits 4-8
280 * no defines for them in headers
282 speed = linkstat & 0xf;
285 dd->lbus_width = linkstat;
289 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
292 dd->lbus_speed = 5000; /* Gen1, 5GHz */
294 default: /* not defined, assume gen1 */
295 dd->lbus_speed = 2500;
300 * Check against expected pcie width and complain if "wrong"
301 * on first initialization, not afterwards (i.e., reset).
303 if (minw && linkstat < minw)
305 "PCIe width %u (x%u HCA), performance reduced\n",
308 qib_tune_pcie_caps(dd);
310 qib_tune_pcie_coalesce(dd);
313 /* fill in string, even on errors */
314 snprintf(dd->lbus_info, sizeof(dd->lbus_info),
315 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
320 * Setup pcie interrupt stuff again after a reset. I'd like to just call
321 * pci_enable_msi() again for msi, but when I do that,
322 * the MSI enable bit doesn't get set in the command word, and
323 * we switch to to a different interrupt vector, which is confusing,
324 * so I instead just do it all inline. Perhaps somehow can tie this
325 * into the PCIe hotplug support at some point
327 int qib_reinit_intr(struct qib_devdata *dd)
333 /* If we aren't using MSI, don't restore it */
337 pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
339 qib_dev_err(dd, "Can't find MSI capability, "
340 "can't restore MSI settings\n");
342 /* nothing special for MSIx, just MSI */
345 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
347 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
349 pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
350 if (!(control & PCI_MSI_FLAGS_ENABLE)) {
351 control |= PCI_MSI_FLAGS_ENABLE;
352 pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
355 /* now rewrite the data (vector) info */
356 pci_write_config_word(dd->pcidev, pos +
357 ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
361 if (!ret && (dd->flags & QIB_HAS_INTX)) {
362 qib_enable_intx(dd->pcidev);
366 /* and now set the pci master bit again */
367 pci_set_master(dd->pcidev);
373 * Disable msi interrupt if enabled, and clear msi_lo.
374 * This is used primarily for the fallback to INTx, but
375 * is also used in reinit after reset, and during cleanup.
377 void qib_nomsi(struct qib_devdata *dd)
380 pci_disable_msi(dd->pcidev);
384 * Same as qib_nosmi, but for MSIx.
386 void qib_nomsix(struct qib_devdata *dd)
388 pci_disable_msix(dd->pcidev);
392 * Similar to pci_intx(pdev, 1), except that we make sure
395 void qib_enable_intx(struct pci_dev *pdev)
400 /* first, turn on INTx */
401 pci_read_config_word(pdev, PCI_COMMAND, &cw);
402 new = cw & ~PCI_COMMAND_INTX_DISABLE;
404 pci_write_config_word(pdev, PCI_COMMAND, new);
406 pos = pci_find_capability(pdev, PCI_CAP_ID_MSI);
408 /* then turn off MSI */
409 pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &cw);
410 new = cw & ~PCI_MSI_FLAGS_ENABLE;
412 pci_write_config_word(pdev, pos + PCI_MSI_FLAGS, new);
414 pos = pci_find_capability(pdev, PCI_CAP_ID_MSIX);
416 /* then turn off MSIx */
417 pci_read_config_word(pdev, pos + PCI_MSIX_FLAGS, &cw);
418 new = cw & ~PCI_MSIX_FLAGS_ENABLE;
420 pci_write_config_word(pdev, pos + PCI_MSIX_FLAGS, new);
425 * These two routines are helper routines for the device reset code
426 * to move all the pcie code out of the chip-specific driver code.
428 void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
430 pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
431 pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
432 pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
435 void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
438 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
441 qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
442 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
445 qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
446 /* now re-enable memory access, and restore cosmetic settings */
447 pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
448 pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
449 pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
450 r = pci_enable_device(dd->pcidev);
452 qib_dev_err(dd, "pci_enable_device failed after "
456 /* code to adjust PCIe capabilities. */
458 static int fld2val(int wd, int mask)
465 lsbmask = mask ^ (mask & (mask - 1));
470 static int val2fld(int wd, int mask)
476 lsbmask = mask ^ (mask & (mask - 1));
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");
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.
491 static int qib_tune_pcie_coalesce(struct qib_devdata *dd)
494 struct pci_dev *parent;
499 if (!qib_pcie_coalesce)
502 /* Find out supported and configured values for parent (root) */
503 parent = dd->pcidev->bus->self;
504 if (parent->bus->parent) {
505 qib_devinfo(dd->pcidev, "Parent not root\n");
508 ppos = pci_find_capability(parent, PCI_CAP_ID_EXP);
511 if (parent->vendor != 0x8086)
515 * - bit 12: Max_rdcmp_Imt_EN: need to set to 1
516 * - bit 11: COALESCE_FORCE: need to set to 0
517 * - bit 10: COALESCE_EN: need to set to 1
518 * (but limitations on some on some chipsets)
520 * On the Intel 5000, 5100, and 7300 chipsets, there is
521 * also: - bit 25:24: COALESCE_MODE, need to set to 0
523 devid = parent->device;
524 if (devid >= 0x25e2 && devid <= 0x25fa) {
528 pci_read_config_byte(parent, PCI_REVISION_ID, &rev);
533 mask = (3U << 24) | (7U << 10);
534 } else if (devid >= 0x65e2 && devid <= 0x65fa) {
537 mask = (3U << 24) | (7U << 10);
538 } else if (devid >= 0x4021 && devid <= 0x402e) {
542 } else if (devid >= 0x3604 && devid <= 0x360a) {
545 mask = (3U << 24) | (7U << 10);
547 /* not one of the chipsets that we know about */
550 pci_read_config_dword(parent, 0x48, &val);
553 r = pci_write_config_dword(parent, 0x48, val);
558 * BIOS may not set PCIe bus-utilization parameters for best performance.
559 * Check and optionally adjust them to maximize our throughput.
561 static int qib_pcie_caps;
562 module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
563 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (4lsb), ReadReq (D4..7)");
565 static int qib_tune_pcie_caps(struct qib_devdata *dd)
567 int ret = 1; /* Assume the worst */
568 struct pci_dev *parent;
570 u16 pcaps, pctl, ecaps, ectl;
574 /* Find out supported and configured values for parent (root) */
575 parent = dd->pcidev->bus->self;
576 if (parent->bus->parent) {
577 qib_devinfo(dd->pcidev, "Parent not root\n");
580 ppos = pci_find_capability(parent, PCI_CAP_ID_EXP);
582 pci_read_config_word(parent, ppos + PCI_EXP_DEVCAP, &pcaps);
583 pci_read_config_word(parent, ppos + PCI_EXP_DEVCTL, &pctl);
586 /* Find out supported and configured values for endpoint (us) */
587 epos = pci_find_capability(dd->pcidev, PCI_CAP_ID_EXP);
589 pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCAP, &ecaps);
590 pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, &ectl);
594 /* Find max payload supported by root, endpoint */
595 rc_sup = fld2val(pcaps, PCI_EXP_DEVCAP_PAYLOAD);
596 ep_sup = fld2val(ecaps, PCI_EXP_DEVCAP_PAYLOAD);
600 rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_PAYLOAD);
601 ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_PAYLOAD);
603 /* If Supported greater than limit in module param, limit it */
604 if (rc_sup > (qib_pcie_caps & 7))
605 rc_sup = qib_pcie_caps & 7;
606 /* If less than (allowed, supported), bump root payload */
607 if (rc_sup > rc_cur) {
609 pctl = (pctl & ~PCI_EXP_DEVCTL_PAYLOAD) |
610 val2fld(rc_cur, PCI_EXP_DEVCTL_PAYLOAD);
611 pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
613 /* If less than (allowed, supported), bump endpoint payload */
614 if (rc_sup > ep_cur) {
616 ectl = (ectl & ~PCI_EXP_DEVCTL_PAYLOAD) |
617 val2fld(ep_cur, PCI_EXP_DEVCTL_PAYLOAD);
618 pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
622 * Now the Read Request size.
623 * No field for max supported, but PCIe spec limits it to 4096,
624 * which is code '5' (log2(4096) - 7)
627 if (rc_sup > ((qib_pcie_caps >> 4) & 7))
628 rc_sup = (qib_pcie_caps >> 4) & 7;
629 rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_READRQ);
630 ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_READRQ);
632 if (rc_sup > rc_cur) {
634 pctl = (pctl & ~PCI_EXP_DEVCTL_READRQ) |
635 val2fld(rc_cur, PCI_EXP_DEVCTL_READRQ);
636 pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
638 if (rc_sup > ep_cur) {
640 ectl = (ectl & ~PCI_EXP_DEVCTL_READRQ) |
641 val2fld(ep_cur, PCI_EXP_DEVCTL_READRQ);
642 pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
647 /* End of PCIe capability tuning */
650 * From here through qib_pci_err_handler definition is invoked via
651 * PCI error infrastructure, registered via pci
653 static pci_ers_result_t
654 qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
656 struct qib_devdata *dd = pci_get_drvdata(pdev);
657 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
660 case pci_channel_io_normal:
661 qib_devinfo(pdev, "State Normal, ignoring\n");
664 case pci_channel_io_frozen:
665 qib_devinfo(pdev, "State Frozen, requesting reset\n");
666 pci_disable_device(pdev);
667 ret = PCI_ERS_RESULT_NEED_RESET;
670 case pci_channel_io_perm_failure:
671 qib_devinfo(pdev, "State Permanent Failure, disabling\n");
673 /* no more register accesses! */
674 dd->flags &= ~QIB_PRESENT;
675 qib_disable_after_error(dd);
677 /* else early, or other problem */
678 ret = PCI_ERS_RESULT_DISCONNECT;
681 default: /* shouldn't happen */
682 qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
689 static pci_ers_result_t
690 qib_pci_mmio_enabled(struct pci_dev *pdev)
693 struct qib_devdata *dd = pci_get_drvdata(pdev);
694 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
696 if (dd && dd->pport) {
697 words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
699 ret = PCI_ERS_RESULT_NEED_RESET;
701 qib_devinfo(pdev, "QIB mmio_enabled function called, "
702 "read wordscntr %Lx, returning %d\n", words, ret);
706 static pci_ers_result_t
707 qib_pci_slot_reset(struct pci_dev *pdev)
709 qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
710 return PCI_ERS_RESULT_CAN_RECOVER;
713 static pci_ers_result_t
714 qib_pci_link_reset(struct pci_dev *pdev)
716 qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
717 return PCI_ERS_RESULT_CAN_RECOVER;
721 qib_pci_resume(struct pci_dev *pdev)
723 struct qib_devdata *dd = pci_get_drvdata(pdev);
724 qib_devinfo(pdev, "QIB resume function called\n");
725 pci_cleanup_aer_uncorrect_error_status(pdev);
727 * Running jobs will fail, since it's asynchronous
728 * unlike sysfs-requested reset. Better than
731 qib_init(dd, 1); /* same as re-init after reset */
734 struct pci_error_handlers qib_pci_err_handler = {
735 .error_detected = qib_pci_error_detected,
736 .mmio_enabled = qib_pci_mmio_enabled,
737 .link_reset = qib_pci_link_reset,
738 .slot_reset = qib_pci_slot_reset,
739 .resume = qib_pci_resume,