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>
38 #include <linux/module.h>
43 * This file contains PCIe utility routines that are common to the
44 * various QLogic InfiniPath adapters
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.
54 static int qib_tune_pcie_caps(struct qib_devdata *);
55 static int qib_tune_pcie_coalesce(struct qib_devdata *);
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
63 int qib_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
67 ret = pci_enable_device(pdev);
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.
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.
81 qib_early_err(&pdev->dev, "pci enable failed: error %d\n",
86 ret = pci_request_regions(pdev, QIB_DRV_NAME);
88 qib_devinfo(pdev, "pci_request_regions fails: err %d\n", -ret);
92 ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
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
99 ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
101 qib_devinfo(pdev, "Unable to set DMA mask: %d\n", ret);
104 ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
106 ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
108 qib_early_err(&pdev->dev,
109 "Unable to set DMA consistent mask: %d\n", ret);
113 pci_set_master(pdev);
114 ret = pci_enable_pcie_error_reporting(pdev);
116 qib_early_err(&pdev->dev,
117 "Unable to enable pcie error reporting: %d\n",
124 pci_disable_device(pdev);
125 pci_release_regions(pdev);
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
135 int qib_pcie_ddinit(struct qib_devdata *dd, struct pci_dev *pdev,
136 const struct pci_device_id *ent)
139 resource_size_t addr;
142 pci_set_drvdata(pdev, dd);
144 addr = pci_resource_start(pdev, 0);
145 len = pci_resource_len(pdev, 0);
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);
151 dd->kregbase = ioremap_nocache(addr, len);
157 dd->kregend = (u64 __iomem *)((void __iomem *) dd->kregbase + len);
158 dd->physaddr = addr; /* used for io_remap, etc. */
161 * Save BARs to rewrite after device reset. Save all 64 bits of
165 dd->pcibar1 = addr >> 32;
166 dd->deviceid = ent->device; /* save for later use */
167 dd->vendorid = ent->vendor;
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
177 void qib_pcie_ddcleanup(struct qib_devdata *dd)
179 u64 __iomem *base = (void __iomem *) dd->kregbase;
184 iounmap(dd->piobase);
186 iounmap(dd->userbase);
188 iounmap(dd->piovl15base);
190 pci_disable_device(dd->pcidev);
191 pci_release_regions(dd->pcidev);
193 pci_set_drvdata(dd->pcidev, NULL);
196 static void qib_msix_setup(struct qib_devdata *dd, int pos, u32 *msixcnt,
197 struct qib_msix_entry *qib_msix_entry)
202 struct msix_entry *msix_entry;
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);
213 for (i = 0; i < *msixcnt; i++)
214 msix_entry[i] = qib_msix_entry[i].msix;
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)
220 ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
223 ret = pci_enable_msix(dd->pcidev, msix_entry, tabsize);
228 "pci_enable_msix %d vectors failed: %d, falling back to INTx\n",
232 for (i = 0; i < tabsize; i++)
233 qib_msix_entry[i].msix = msix_entry[i];
238 qib_enable_intx(dd->pcidev);
243 * We save the msi lo and hi values, so we can restore them after
244 * chip reset (the kernel PCI infrastructure doesn't yet handle that
247 static int qib_msi_setup(struct qib_devdata *dd, int pos)
249 struct pci_dev *pdev = dd->pcidev;
253 ret = pci_enable_msi(pdev);
256 "pci_enable_msi failed: %d, interrupts may not work\n",
258 /* continue even if it fails, we may still be OK... */
260 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_LO,
262 pci_read_config_dword(pdev, pos + PCI_MSI_ADDRESS_HI,
264 pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &control);
265 /* now save the data (vector) info */
266 pci_read_config_word(pdev, pos + ((control & PCI_MSI_FLAGS_64BIT)
272 int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent,
273 struct qib_msix_entry *entry)
276 int pos = 0, pose, ret = 1;
278 pose = pci_pcie_cap(dd->pcidev);
280 qib_dev_err(dd, "Can't find PCI Express capability!\n");
281 /* set up something... */
283 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
287 pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSIX);
288 if (nent && *nent && pos) {
289 qib_msix_setup(dd, pos, nent, entry);
290 ret = 0; /* did it, either MSIx or INTx */
292 pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
294 ret = qib_msi_setup(dd, pos);
296 qib_dev_err(dd, "No PCI MSI or MSIx capability!\n");
299 qib_enable_intx(dd->pcidev);
301 pci_read_config_word(dd->pcidev, pose + PCI_EXP_LNKSTA, &linkstat);
303 * speed is bits 0-3, linkwidth is bits 4-8
304 * no defines for them in headers
306 speed = linkstat & 0xf;
309 dd->lbus_width = linkstat;
313 dd->lbus_speed = 2500; /* Gen1, 2.5GHz */
316 dd->lbus_speed = 5000; /* Gen1, 5GHz */
318 default: /* not defined, assume gen1 */
319 dd->lbus_speed = 2500;
324 * Check against expected pcie width and complain if "wrong"
325 * on first initialization, not afterwards (i.e., reset).
327 if (minw && linkstat < minw)
329 "PCIe width %u (x%u HCA), performance reduced\n",
332 qib_tune_pcie_caps(dd);
334 qib_tune_pcie_coalesce(dd);
337 /* fill in string, even on errors */
338 snprintf(dd->lbus_info, sizeof(dd->lbus_info),
339 "PCIe,%uMHz,x%u\n", dd->lbus_speed, dd->lbus_width);
344 * Setup pcie interrupt stuff again after a reset. I'd like to just call
345 * pci_enable_msi() again for msi, but when I do that,
346 * the MSI enable bit doesn't get set in the command word, and
347 * we switch to to a different interrupt vector, which is confusing,
348 * so I instead just do it all inline. Perhaps somehow can tie this
349 * into the PCIe hotplug support at some point
351 int qib_reinit_intr(struct qib_devdata *dd)
357 /* If we aren't using MSI, don't restore it */
361 pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI);
364 "Can't find MSI capability, can't restore MSI settings\n");
366 /* nothing special for MSIx, just MSI */
369 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
371 pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
373 pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
374 if (!(control & PCI_MSI_FLAGS_ENABLE)) {
375 control |= PCI_MSI_FLAGS_ENABLE;
376 pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
379 /* now rewrite the data (vector) info */
380 pci_write_config_word(dd->pcidev, pos +
381 ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
385 if (!ret && (dd->flags & QIB_HAS_INTX)) {
386 qib_enable_intx(dd->pcidev);
390 /* and now set the pci master bit again */
391 pci_set_master(dd->pcidev);
397 * Disable msi interrupt if enabled, and clear msi_lo.
398 * This is used primarily for the fallback to INTx, but
399 * is also used in reinit after reset, and during cleanup.
401 void qib_nomsi(struct qib_devdata *dd)
404 pci_disable_msi(dd->pcidev);
408 * Same as qib_nosmi, but for MSIx.
410 void qib_nomsix(struct qib_devdata *dd)
412 pci_disable_msix(dd->pcidev);
416 * Similar to pci_intx(pdev, 1), except that we make sure
419 void qib_enable_intx(struct pci_dev *pdev)
424 /* first, turn on INTx */
425 pci_read_config_word(pdev, PCI_COMMAND, &cw);
426 new = cw & ~PCI_COMMAND_INTX_DISABLE;
428 pci_write_config_word(pdev, PCI_COMMAND, new);
430 pos = pci_find_capability(pdev, PCI_CAP_ID_MSI);
432 /* then turn off MSI */
433 pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &cw);
434 new = cw & ~PCI_MSI_FLAGS_ENABLE;
436 pci_write_config_word(pdev, pos + PCI_MSI_FLAGS, new);
438 pos = pci_find_capability(pdev, PCI_CAP_ID_MSIX);
440 /* then turn off MSIx */
441 pci_read_config_word(pdev, pos + PCI_MSIX_FLAGS, &cw);
442 new = cw & ~PCI_MSIX_FLAGS_ENABLE;
444 pci_write_config_word(pdev, pos + PCI_MSIX_FLAGS, new);
449 * These two routines are helper routines for the device reset code
450 * to move all the pcie code out of the chip-specific driver code.
452 void qib_pcie_getcmd(struct qib_devdata *dd, u16 *cmd, u8 *iline, u8 *cline)
454 pci_read_config_word(dd->pcidev, PCI_COMMAND, cmd);
455 pci_read_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
456 pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
459 void qib_pcie_reenable(struct qib_devdata *dd, u16 cmd, u8 iline, u8 cline)
462 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
465 qib_dev_err(dd, "rewrite of BAR0 failed: %d\n", r);
466 r = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
469 qib_dev_err(dd, "rewrite of BAR1 failed: %d\n", r);
470 /* now re-enable memory access, and restore cosmetic settings */
471 pci_write_config_word(dd->pcidev, PCI_COMMAND, cmd);
472 pci_write_config_byte(dd->pcidev, PCI_INTERRUPT_LINE, iline);
473 pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE, cline);
474 r = pci_enable_device(dd->pcidev);
477 "pci_enable_device failed after reset: %d\n", r);
480 /* code to adjust PCIe capabilities. */
482 static int fld2val(int wd, int mask)
489 lsbmask = mask ^ (mask & (mask - 1));
494 static int val2fld(int wd, int mask)
500 lsbmask = mask ^ (mask & (mask - 1));
505 static int qib_pcie_coalesce;
506 module_param_named(pcie_coalesce, qib_pcie_coalesce, int, S_IRUGO);
507 MODULE_PARM_DESC(pcie_coalesce, "tune PCIe colescing on some Intel chipsets");
510 * Enable PCIe completion and data coalescing, on Intel 5x00 and 7300
511 * chipsets. This is known to be unsafe for some revisions of some
512 * of these chipsets, with some BIOS settings, and enabling it on those
513 * systems may result in the system crashing, and/or data corruption.
515 static int qib_tune_pcie_coalesce(struct qib_devdata *dd)
518 struct pci_dev *parent;
523 if (!qib_pcie_coalesce)
526 /* Find out supported and configured values for parent (root) */
527 parent = dd->pcidev->bus->self;
528 if (parent->bus->parent) {
529 qib_devinfo(dd->pcidev, "Parent not root\n");
532 ppos = pci_pcie_cap(parent);
535 if (parent->vendor != 0x8086)
539 * - bit 12: Max_rdcmp_Imt_EN: need to set to 1
540 * - bit 11: COALESCE_FORCE: need to set to 0
541 * - bit 10: COALESCE_EN: need to set to 1
542 * (but limitations on some on some chipsets)
544 * On the Intel 5000, 5100, and 7300 chipsets, there is
545 * also: - bit 25:24: COALESCE_MODE, need to set to 0
547 devid = parent->device;
548 if (devid >= 0x25e2 && devid <= 0x25fa) {
550 if (parent->revision <= 0xb2)
554 mask = (3U << 24) | (7U << 10);
555 } else if (devid >= 0x65e2 && devid <= 0x65fa) {
558 mask = (3U << 24) | (7U << 10);
559 } else if (devid >= 0x4021 && devid <= 0x402e) {
563 } else if (devid >= 0x3604 && devid <= 0x360a) {
566 mask = (3U << 24) | (7U << 10);
568 /* not one of the chipsets that we know about */
571 pci_read_config_dword(parent, 0x48, &val);
574 r = pci_write_config_dword(parent, 0x48, val);
579 * BIOS may not set PCIe bus-utilization parameters for best performance.
580 * Check and optionally adjust them to maximize our throughput.
582 static int qib_pcie_caps;
583 module_param_named(pcie_caps, qib_pcie_caps, int, S_IRUGO);
584 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
586 static int qib_tune_pcie_caps(struct qib_devdata *dd)
588 int ret = 1; /* Assume the worst */
589 struct pci_dev *parent;
591 u16 pcaps, pctl, ecaps, ectl;
595 /* Find out supported and configured values for parent (root) */
596 parent = dd->pcidev->bus->self;
597 if (parent->bus->parent) {
598 qib_devinfo(dd->pcidev, "Parent not root\n");
601 ppos = pci_pcie_cap(parent);
603 pci_read_config_word(parent, ppos + PCI_EXP_DEVCAP, &pcaps);
604 pci_read_config_word(parent, ppos + PCI_EXP_DEVCTL, &pctl);
607 /* Find out supported and configured values for endpoint (us) */
608 epos = pci_pcie_cap(dd->pcidev);
610 pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCAP, &ecaps);
611 pci_read_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, &ectl);
615 /* Find max payload supported by root, endpoint */
616 rc_sup = fld2val(pcaps, PCI_EXP_DEVCAP_PAYLOAD);
617 ep_sup = fld2val(ecaps, PCI_EXP_DEVCAP_PAYLOAD);
621 rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_PAYLOAD);
622 ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_PAYLOAD);
624 /* If Supported greater than limit in module param, limit it */
625 if (rc_sup > (qib_pcie_caps & 7))
626 rc_sup = qib_pcie_caps & 7;
627 /* If less than (allowed, supported), bump root payload */
628 if (rc_sup > rc_cur) {
630 pctl = (pctl & ~PCI_EXP_DEVCTL_PAYLOAD) |
631 val2fld(rc_cur, PCI_EXP_DEVCTL_PAYLOAD);
632 pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
634 /* If less than (allowed, supported), bump endpoint payload */
635 if (rc_sup > ep_cur) {
637 ectl = (ectl & ~PCI_EXP_DEVCTL_PAYLOAD) |
638 val2fld(ep_cur, PCI_EXP_DEVCTL_PAYLOAD);
639 pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
643 * Now the Read Request size.
644 * No field for max supported, but PCIe spec limits it to 4096,
645 * which is code '5' (log2(4096) - 7)
648 if (rc_sup > ((qib_pcie_caps >> 4) & 7))
649 rc_sup = (qib_pcie_caps >> 4) & 7;
650 rc_cur = fld2val(pctl, PCI_EXP_DEVCTL_READRQ);
651 ep_cur = fld2val(ectl, PCI_EXP_DEVCTL_READRQ);
653 if (rc_sup > rc_cur) {
655 pctl = (pctl & ~PCI_EXP_DEVCTL_READRQ) |
656 val2fld(rc_cur, PCI_EXP_DEVCTL_READRQ);
657 pci_write_config_word(parent, ppos + PCI_EXP_DEVCTL, pctl);
659 if (rc_sup > ep_cur) {
661 ectl = (ectl & ~PCI_EXP_DEVCTL_READRQ) |
662 val2fld(ep_cur, PCI_EXP_DEVCTL_READRQ);
663 pci_write_config_word(dd->pcidev, epos + PCI_EXP_DEVCTL, ectl);
668 /* End of PCIe capability tuning */
671 * From here through qib_pci_err_handler definition is invoked via
672 * PCI error infrastructure, registered via pci
674 static pci_ers_result_t
675 qib_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
677 struct qib_devdata *dd = pci_get_drvdata(pdev);
678 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
681 case pci_channel_io_normal:
682 qib_devinfo(pdev, "State Normal, ignoring\n");
685 case pci_channel_io_frozen:
686 qib_devinfo(pdev, "State Frozen, requesting reset\n");
687 pci_disable_device(pdev);
688 ret = PCI_ERS_RESULT_NEED_RESET;
691 case pci_channel_io_perm_failure:
692 qib_devinfo(pdev, "State Permanent Failure, disabling\n");
694 /* no more register accesses! */
695 dd->flags &= ~QIB_PRESENT;
696 qib_disable_after_error(dd);
698 /* else early, or other problem */
699 ret = PCI_ERS_RESULT_DISCONNECT;
702 default: /* shouldn't happen */
703 qib_devinfo(pdev, "QIB PCI errors detected (state %d)\n",
710 static pci_ers_result_t
711 qib_pci_mmio_enabled(struct pci_dev *pdev)
714 struct qib_devdata *dd = pci_get_drvdata(pdev);
715 pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
717 if (dd && dd->pport) {
718 words = dd->f_portcntr(dd->pport, QIBPORTCNTR_WORDRCV);
720 ret = PCI_ERS_RESULT_NEED_RESET;
723 "QIB mmio_enabled function called, read wordscntr %Lx, returning %d\n",
728 static pci_ers_result_t
729 qib_pci_slot_reset(struct pci_dev *pdev)
731 qib_devinfo(pdev, "QIB slot_reset function called, ignored\n");
732 return PCI_ERS_RESULT_CAN_RECOVER;
735 static pci_ers_result_t
736 qib_pci_link_reset(struct pci_dev *pdev)
738 qib_devinfo(pdev, "QIB link_reset function called, ignored\n");
739 return PCI_ERS_RESULT_CAN_RECOVER;
743 qib_pci_resume(struct pci_dev *pdev)
745 struct qib_devdata *dd = pci_get_drvdata(pdev);
746 qib_devinfo(pdev, "QIB resume function called\n");
747 pci_cleanup_aer_uncorrect_error_status(pdev);
749 * Running jobs will fail, since it's asynchronous
750 * unlike sysfs-requested reset. Better than
753 qib_init(dd, 1); /* same as re-init after reset */
756 struct pci_error_handlers qib_pci_err_handler = {
757 .error_detected = qib_pci_error_detected,
758 .mmio_enabled = qib_pci_mmio_enabled,
759 .link_reset = qib_pci_link_reset,
760 .slot_reset = qib_pci_slot_reset,
761 .resume = qib_pci_resume,