xref: /openbmc/linux/drivers/infiniband/hw/hfi1/pcie.c (revision 79e790ff)
1 /*
2  * Copyright(c) 2015 - 2019 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/pci.h>
49 #include <linux/io.h>
50 #include <linux/delay.h>
51 #include <linux/vmalloc.h>
52 #include <linux/aer.h>
53 #include <linux/module.h>
54 
55 #include "hfi.h"
56 #include "chip_registers.h"
57 #include "aspm.h"
58 
59 /*
60  * This file contains PCIe utility routines.
61  */
62 
63 /*
64  * Do all the common PCIe setup and initialization.
65  */
66 int hfi1_pcie_init(struct hfi1_devdata *dd)
67 {
68 	int ret;
69 	struct pci_dev *pdev = dd->pcidev;
70 
71 	ret = pci_enable_device(pdev);
72 	if (ret) {
73 		/*
74 		 * This can happen (in theory) iff:
75 		 * We did a chip reset, and then failed to reprogram the
76 		 * BAR, or the chip reset due to an internal error.  We then
77 		 * unloaded the driver and reloaded it.
78 		 *
79 		 * Both reset cases set the BAR back to initial state.  For
80 		 * the latter case, the AER sticky error bit at offset 0x718
81 		 * should be set, but the Linux kernel doesn't yet know
82 		 * about that, it appears.  If the original BAR was retained
83 		 * in the kernel data structures, this may be OK.
84 		 */
85 		dd_dev_err(dd, "pci enable failed: error %d\n", -ret);
86 		return ret;
87 	}
88 
89 	ret = pci_request_regions(pdev, DRIVER_NAME);
90 	if (ret) {
91 		dd_dev_err(dd, "pci_request_regions fails: err %d\n", -ret);
92 		goto bail;
93 	}
94 
95 	ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
96 	if (ret) {
97 		/*
98 		 * If the 64 bit setup fails, try 32 bit.  Some systems
99 		 * do not setup 64 bit maps on systems with 2GB or less
100 		 * memory installed.
101 		 */
102 		ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
103 		if (ret) {
104 			dd_dev_err(dd, "Unable to set DMA mask: %d\n", ret);
105 			goto bail;
106 		}
107 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
108 	} else {
109 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
110 	}
111 	if (ret) {
112 		dd_dev_err(dd, "Unable to set DMA consistent mask: %d\n", ret);
113 		goto bail;
114 	}
115 
116 	pci_set_master(pdev);
117 	(void)pci_enable_pcie_error_reporting(pdev);
118 	return 0;
119 
120 bail:
121 	hfi1_pcie_cleanup(pdev);
122 	return ret;
123 }
124 
125 /*
126  * Clean what was done in hfi1_pcie_init()
127  */
128 void hfi1_pcie_cleanup(struct pci_dev *pdev)
129 {
130 	pci_disable_device(pdev);
131 	/*
132 	 * Release regions should be called after the disable. OK to
133 	 * call if request regions has not been called or failed.
134 	 */
135 	pci_release_regions(pdev);
136 }
137 
138 /*
139  * Do remaining PCIe setup, once dd is allocated, and save away
140  * fields required to re-initialize after a chip reset, or for
141  * various other purposes
142  */
143 int hfi1_pcie_ddinit(struct hfi1_devdata *dd, struct pci_dev *pdev)
144 {
145 	unsigned long len;
146 	resource_size_t addr;
147 	int ret = 0;
148 	u32 rcv_array_count;
149 
150 	addr = pci_resource_start(pdev, 0);
151 	len = pci_resource_len(pdev, 0);
152 
153 	/*
154 	 * The TXE PIO buffers are at the tail end of the chip space.
155 	 * Cut them off and map them separately.
156 	 */
157 
158 	/* sanity check vs expectations */
159 	if (len != TXE_PIO_SEND + TXE_PIO_SIZE) {
160 		dd_dev_err(dd, "chip PIO range does not match\n");
161 		return -EINVAL;
162 	}
163 
164 	dd->kregbase1 = ioremap(addr, RCV_ARRAY);
165 	if (!dd->kregbase1) {
166 		dd_dev_err(dd, "UC mapping of kregbase1 failed\n");
167 		return -ENOMEM;
168 	}
169 	dd_dev_info(dd, "UC base1: %p for %x\n", dd->kregbase1, RCV_ARRAY);
170 
171 	/* verify that reads actually work, save revision for reset check */
172 	dd->revision = readq(dd->kregbase1 + CCE_REVISION);
173 	if (dd->revision == ~(u64)0) {
174 		dd_dev_err(dd, "Cannot read chip CSRs\n");
175 		goto nomem;
176 	}
177 
178 	rcv_array_count = readq(dd->kregbase1 + RCV_ARRAY_CNT);
179 	dd_dev_info(dd, "RcvArray count: %u\n", rcv_array_count);
180 	dd->base2_start  = RCV_ARRAY + rcv_array_count * 8;
181 
182 	dd->kregbase2 = ioremap(
183 		addr + dd->base2_start,
184 		TXE_PIO_SEND - dd->base2_start);
185 	if (!dd->kregbase2) {
186 		dd_dev_err(dd, "UC mapping of kregbase2 failed\n");
187 		goto nomem;
188 	}
189 	dd_dev_info(dd, "UC base2: %p for %x\n", dd->kregbase2,
190 		    TXE_PIO_SEND - dd->base2_start);
191 
192 	dd->piobase = ioremap_wc(addr + TXE_PIO_SEND, TXE_PIO_SIZE);
193 	if (!dd->piobase) {
194 		dd_dev_err(dd, "WC mapping of send buffers failed\n");
195 		goto nomem;
196 	}
197 	dd_dev_info(dd, "WC piobase: %p for %x\n", dd->piobase, TXE_PIO_SIZE);
198 
199 	dd->physaddr = addr;        /* used for io_remap, etc. */
200 
201 	/*
202 	 * Map the chip's RcvArray as write-combining to allow us
203 	 * to write an entire cacheline worth of entries in one shot.
204 	 */
205 	dd->rcvarray_wc = ioremap_wc(addr + RCV_ARRAY,
206 				     rcv_array_count * 8);
207 	if (!dd->rcvarray_wc) {
208 		dd_dev_err(dd, "WC mapping of receive array failed\n");
209 		goto nomem;
210 	}
211 	dd_dev_info(dd, "WC RcvArray: %p for %x\n",
212 		    dd->rcvarray_wc, rcv_array_count * 8);
213 
214 	dd->flags |= HFI1_PRESENT;	/* chip.c CSR routines now work */
215 	return 0;
216 nomem:
217 	ret = -ENOMEM;
218 	hfi1_pcie_ddcleanup(dd);
219 	return ret;
220 }
221 
222 /*
223  * Do PCIe cleanup related to dd, after chip-specific cleanup, etc.  Just prior
224  * to releasing the dd memory.
225  * Void because all of the core pcie cleanup functions are void.
226  */
227 void hfi1_pcie_ddcleanup(struct hfi1_devdata *dd)
228 {
229 	dd->flags &= ~HFI1_PRESENT;
230 	if (dd->kregbase1)
231 		iounmap(dd->kregbase1);
232 	dd->kregbase1 = NULL;
233 	if (dd->kregbase2)
234 		iounmap(dd->kregbase2);
235 	dd->kregbase2 = NULL;
236 	if (dd->rcvarray_wc)
237 		iounmap(dd->rcvarray_wc);
238 	dd->rcvarray_wc = NULL;
239 	if (dd->piobase)
240 		iounmap(dd->piobase);
241 	dd->piobase = NULL;
242 }
243 
244 /* return the PCIe link speed from the given link status */
245 static u32 extract_speed(u16 linkstat)
246 {
247 	u32 speed;
248 
249 	switch (linkstat & PCI_EXP_LNKSTA_CLS) {
250 	default: /* not defined, assume Gen1 */
251 	case PCI_EXP_LNKSTA_CLS_2_5GB:
252 		speed = 2500; /* Gen 1, 2.5GHz */
253 		break;
254 	case PCI_EXP_LNKSTA_CLS_5_0GB:
255 		speed = 5000; /* Gen 2, 5GHz */
256 		break;
257 	case PCI_EXP_LNKSTA_CLS_8_0GB:
258 		speed = 8000; /* Gen 3, 8GHz */
259 		break;
260 	}
261 	return speed;
262 }
263 
264 /* return the PCIe link speed from the given link status */
265 static u32 extract_width(u16 linkstat)
266 {
267 	return (linkstat & PCI_EXP_LNKSTA_NLW) >> PCI_EXP_LNKSTA_NLW_SHIFT;
268 }
269 
270 /* read the link status and set dd->{lbus_width,lbus_speed,lbus_info} */
271 static void update_lbus_info(struct hfi1_devdata *dd)
272 {
273 	u16 linkstat;
274 	int ret;
275 
276 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
277 	if (ret) {
278 		dd_dev_err(dd, "Unable to read from PCI config\n");
279 		return;
280 	}
281 
282 	dd->lbus_width = extract_width(linkstat);
283 	dd->lbus_speed = extract_speed(linkstat);
284 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
285 		 "PCIe,%uMHz,x%u", dd->lbus_speed, dd->lbus_width);
286 }
287 
288 /*
289  * Read in the current PCIe link width and speed.  Find if the link is
290  * Gen3 capable.
291  */
292 int pcie_speeds(struct hfi1_devdata *dd)
293 {
294 	u32 linkcap;
295 	struct pci_dev *parent = dd->pcidev->bus->self;
296 	int ret;
297 
298 	if (!pci_is_pcie(dd->pcidev)) {
299 		dd_dev_err(dd, "Can't find PCI Express capability!\n");
300 		return -EINVAL;
301 	}
302 
303 	/* find if our max speed is Gen3 and parent supports Gen3 speeds */
304 	dd->link_gen3_capable = 1;
305 
306 	ret = pcie_capability_read_dword(dd->pcidev, PCI_EXP_LNKCAP, &linkcap);
307 	if (ret) {
308 		dd_dev_err(dd, "Unable to read from PCI config\n");
309 		return pcibios_err_to_errno(ret);
310 	}
311 
312 	if ((linkcap & PCI_EXP_LNKCAP_SLS) != PCI_EXP_LNKCAP_SLS_8_0GB) {
313 		dd_dev_info(dd,
314 			    "This HFI is not Gen3 capable, max speed 0x%x, need 0x3\n",
315 			    linkcap & PCI_EXP_LNKCAP_SLS);
316 		dd->link_gen3_capable = 0;
317 	}
318 
319 	/*
320 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
321 	 */
322 	if (parent &&
323 	    (dd->pcidev->bus->max_bus_speed == PCIE_SPEED_2_5GT ||
324 	     dd->pcidev->bus->max_bus_speed == PCIE_SPEED_5_0GT)) {
325 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
326 		dd->link_gen3_capable = 0;
327 	}
328 
329 	/* obtain the link width and current speed */
330 	update_lbus_info(dd);
331 
332 	dd_dev_info(dd, "%s\n", dd->lbus_info);
333 
334 	return 0;
335 }
336 
337 /*
338  * Restore command and BARs after a reset has wiped them out
339  *
340  * Returns 0 on success, otherwise a negative error value
341  */
342 int restore_pci_variables(struct hfi1_devdata *dd)
343 {
344 	int ret;
345 
346 	ret = pci_write_config_word(dd->pcidev, PCI_COMMAND, dd->pci_command);
347 	if (ret)
348 		goto error;
349 
350 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
351 				     dd->pcibar0);
352 	if (ret)
353 		goto error;
354 
355 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
356 				     dd->pcibar1);
357 	if (ret)
358 		goto error;
359 
360 	ret = pci_write_config_dword(dd->pcidev, PCI_ROM_ADDRESS, dd->pci_rom);
361 	if (ret)
362 		goto error;
363 
364 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL,
365 					 dd->pcie_devctl);
366 	if (ret)
367 		goto error;
368 
369 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL,
370 					 dd->pcie_lnkctl);
371 	if (ret)
372 		goto error;
373 
374 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL2,
375 					 dd->pcie_devctl2);
376 	if (ret)
377 		goto error;
378 
379 	ret = pci_write_config_dword(dd->pcidev, PCI_CFG_MSIX0, dd->pci_msix0);
380 	if (ret)
381 		goto error;
382 
383 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
384 		ret = pci_write_config_dword(dd->pcidev, PCIE_CFG_TPH2,
385 					     dd->pci_tph2);
386 		if (ret)
387 			goto error;
388 	}
389 	return 0;
390 
391 error:
392 	dd_dev_err(dd, "Unable to write to PCI config\n");
393 	return pcibios_err_to_errno(ret);
394 }
395 
396 /*
397  * Save BARs and command to rewrite after device reset
398  *
399  * Returns 0 on success, otherwise a negative error value
400  */
401 int save_pci_variables(struct hfi1_devdata *dd)
402 {
403 	int ret;
404 
405 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
406 				    &dd->pcibar0);
407 	if (ret)
408 		goto error;
409 
410 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
411 				    &dd->pcibar1);
412 	if (ret)
413 		goto error;
414 
415 	ret = pci_read_config_dword(dd->pcidev, PCI_ROM_ADDRESS, &dd->pci_rom);
416 	if (ret)
417 		goto error;
418 
419 	ret = pci_read_config_word(dd->pcidev, PCI_COMMAND, &dd->pci_command);
420 	if (ret)
421 		goto error;
422 
423 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL,
424 					&dd->pcie_devctl);
425 	if (ret)
426 		goto error;
427 
428 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL,
429 					&dd->pcie_lnkctl);
430 	if (ret)
431 		goto error;
432 
433 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL2,
434 					&dd->pcie_devctl2);
435 	if (ret)
436 		goto error;
437 
438 	ret = pci_read_config_dword(dd->pcidev, PCI_CFG_MSIX0, &dd->pci_msix0);
439 	if (ret)
440 		goto error;
441 
442 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
443 		ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_TPH2,
444 					    &dd->pci_tph2);
445 		if (ret)
446 			goto error;
447 	}
448 	return 0;
449 
450 error:
451 	dd_dev_err(dd, "Unable to read from PCI config\n");
452 	return pcibios_err_to_errno(ret);
453 }
454 
455 /*
456  * BIOS may not set PCIe bus-utilization parameters for best performance.
457  * Check and optionally adjust them to maximize our throughput.
458  */
459 static int hfi1_pcie_caps;
460 module_param_named(pcie_caps, hfi1_pcie_caps, int, 0444);
461 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
462 
463 /**
464  * tune_pcie_caps() - Code to adjust PCIe capabilities.
465  * @dd: Valid device data structure
466  *
467  */
468 void tune_pcie_caps(struct hfi1_devdata *dd)
469 {
470 	struct pci_dev *parent;
471 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
472 	u16 rc_mrrs, ep_mrrs, max_mrrs, ectl;
473 	int ret;
474 
475 	/*
476 	 * Turn on extended tags in DevCtl in case the BIOS has turned it off
477 	 * to improve WFR SDMA bandwidth
478 	 */
479 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &ectl);
480 	if ((!ret) && !(ectl & PCI_EXP_DEVCTL_EXT_TAG)) {
481 		dd_dev_info(dd, "Enabling PCIe extended tags\n");
482 		ectl |= PCI_EXP_DEVCTL_EXT_TAG;
483 		ret = pcie_capability_write_word(dd->pcidev,
484 						 PCI_EXP_DEVCTL, ectl);
485 		if (ret)
486 			dd_dev_info(dd, "Unable to write to PCI config\n");
487 	}
488 	/* Find out supported and configured values for parent (root) */
489 	parent = dd->pcidev->bus->self;
490 	/*
491 	 * The driver cannot perform the tuning if it does not have
492 	 * access to the upstream component.
493 	 */
494 	if (!parent) {
495 		dd_dev_info(dd, "Parent not found\n");
496 		return;
497 	}
498 	if (!pci_is_root_bus(parent->bus)) {
499 		dd_dev_info(dd, "Parent not root\n");
500 		return;
501 	}
502 	if (!pci_is_pcie(parent)) {
503 		dd_dev_info(dd, "Parent is not PCI Express capable\n");
504 		return;
505 	}
506 	if (!pci_is_pcie(dd->pcidev)) {
507 		dd_dev_info(dd, "PCI device is not PCI Express capable\n");
508 		return;
509 	}
510 	rc_mpss = parent->pcie_mpss;
511 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
512 	/* Find out supported and configured values for endpoint (us) */
513 	ep_mpss = dd->pcidev->pcie_mpss;
514 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
515 
516 	/* Find max payload supported by root, endpoint */
517 	if (rc_mpss > ep_mpss)
518 		rc_mpss = ep_mpss;
519 
520 	/* If Supported greater than limit in module param, limit it */
521 	if (rc_mpss > (hfi1_pcie_caps & 7))
522 		rc_mpss = hfi1_pcie_caps & 7;
523 	/* If less than (allowed, supported), bump root payload */
524 	if (rc_mpss > rc_mps) {
525 		rc_mps = rc_mpss;
526 		pcie_set_mps(parent, 128 << rc_mps);
527 	}
528 	/* If less than (allowed, supported), bump endpoint payload */
529 	if (rc_mpss > ep_mps) {
530 		ep_mps = rc_mpss;
531 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
532 	}
533 
534 	/*
535 	 * Now the Read Request size.
536 	 * No field for max supported, but PCIe spec limits it to 4096,
537 	 * which is code '5' (log2(4096) - 7)
538 	 */
539 	max_mrrs = 5;
540 	if (max_mrrs > ((hfi1_pcie_caps >> 4) & 7))
541 		max_mrrs = (hfi1_pcie_caps >> 4) & 7;
542 
543 	max_mrrs = 128 << max_mrrs;
544 	rc_mrrs = pcie_get_readrq(parent);
545 	ep_mrrs = pcie_get_readrq(dd->pcidev);
546 
547 	if (max_mrrs > rc_mrrs) {
548 		rc_mrrs = max_mrrs;
549 		pcie_set_readrq(parent, rc_mrrs);
550 	}
551 	if (max_mrrs > ep_mrrs) {
552 		ep_mrrs = max_mrrs;
553 		pcie_set_readrq(dd->pcidev, ep_mrrs);
554 	}
555 }
556 
557 /* End of PCIe capability tuning */
558 
559 /*
560  * From here through hfi1_pci_err_handler definition is invoked via
561  * PCI error infrastructure, registered via pci
562  */
563 static pci_ers_result_t
564 pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
565 {
566 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
567 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
568 
569 	switch (state) {
570 	case pci_channel_io_normal:
571 		dd_dev_info(dd, "State Normal, ignoring\n");
572 		break;
573 
574 	case pci_channel_io_frozen:
575 		dd_dev_info(dd, "State Frozen, requesting reset\n");
576 		pci_disable_device(pdev);
577 		ret = PCI_ERS_RESULT_NEED_RESET;
578 		break;
579 
580 	case pci_channel_io_perm_failure:
581 		if (dd) {
582 			dd_dev_info(dd, "State Permanent Failure, disabling\n");
583 			/* no more register accesses! */
584 			dd->flags &= ~HFI1_PRESENT;
585 			hfi1_disable_after_error(dd);
586 		}
587 		 /* else early, or other problem */
588 		ret =  PCI_ERS_RESULT_DISCONNECT;
589 		break;
590 
591 	default: /* shouldn't happen */
592 		dd_dev_info(dd, "HFI1 PCI errors detected (state %d)\n",
593 			    state);
594 		break;
595 	}
596 	return ret;
597 }
598 
599 static pci_ers_result_t
600 pci_mmio_enabled(struct pci_dev *pdev)
601 {
602 	u64 words = 0U;
603 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
604 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
605 
606 	if (dd && dd->pport) {
607 		words = read_port_cntr(dd->pport, C_RX_WORDS, CNTR_INVALID_VL);
608 		if (words == ~0ULL)
609 			ret = PCI_ERS_RESULT_NEED_RESET;
610 		dd_dev_info(dd,
611 			    "HFI1 mmio_enabled function called, read wordscntr %llx, returning %d\n",
612 			    words, ret);
613 	}
614 	return  ret;
615 }
616 
617 static pci_ers_result_t
618 pci_slot_reset(struct pci_dev *pdev)
619 {
620 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
621 
622 	dd_dev_info(dd, "HFI1 slot_reset function called, ignored\n");
623 	return PCI_ERS_RESULT_CAN_RECOVER;
624 }
625 
626 static void
627 pci_resume(struct pci_dev *pdev)
628 {
629 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
630 
631 	dd_dev_info(dd, "HFI1 resume function called\n");
632 	/*
633 	 * Running jobs will fail, since it's asynchronous
634 	 * unlike sysfs-requested reset.   Better than
635 	 * doing nothing.
636 	 */
637 	hfi1_init(dd, 1); /* same as re-init after reset */
638 }
639 
640 const struct pci_error_handlers hfi1_pci_err_handler = {
641 	.error_detected = pci_error_detected,
642 	.mmio_enabled = pci_mmio_enabled,
643 	.slot_reset = pci_slot_reset,
644 	.resume = pci_resume,
645 };
646 
647 /*============================================================================*/
648 /* PCIe Gen3 support */
649 
650 /*
651  * This code is separated out because it is expected to be removed in the
652  * final shipping product.  If not, then it will be revisited and items
653  * will be moved to more standard locations.
654  */
655 
656 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_STS field values */
657 #define DL_STATUS_HFI0 0x1	/* hfi0 firmware download complete */
658 #define DL_STATUS_HFI1 0x2	/* hfi1 firmware download complete */
659 #define DL_STATUS_BOTH 0x3	/* hfi0 and hfi1 firmware download complete */
660 
661 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_ERR field values */
662 #define DL_ERR_NONE		0x0	/* no error */
663 #define DL_ERR_SWAP_PARITY	0x1	/* parity error in SerDes interrupt */
664 					/*   or response data */
665 #define DL_ERR_DISABLED	0x2	/* hfi disabled */
666 #define DL_ERR_SECURITY	0x3	/* security check failed */
667 #define DL_ERR_SBUS		0x4	/* SBus status error */
668 #define DL_ERR_XFR_PARITY	0x5	/* parity error during ROM transfer*/
669 
670 /* gasket block secondary bus reset delay */
671 #define SBR_DELAY_US 200000	/* 200ms */
672 
673 static uint pcie_target = 3;
674 module_param(pcie_target, uint, S_IRUGO);
675 MODULE_PARM_DESC(pcie_target, "PCIe target speed (0 skip, 1-3 Gen1-3)");
676 
677 static uint pcie_force;
678 module_param(pcie_force, uint, S_IRUGO);
679 MODULE_PARM_DESC(pcie_force, "Force driver to do a PCIe firmware download even if already at target speed");
680 
681 static uint pcie_retry = 5;
682 module_param(pcie_retry, uint, S_IRUGO);
683 MODULE_PARM_DESC(pcie_retry, "Driver will try this many times to reach requested speed");
684 
685 #define UNSET_PSET 255
686 #define DEFAULT_DISCRETE_PSET 2	/* discrete HFI */
687 #define DEFAULT_MCP_PSET 6	/* MCP HFI */
688 static uint pcie_pset = UNSET_PSET;
689 module_param(pcie_pset, uint, S_IRUGO);
690 MODULE_PARM_DESC(pcie_pset, "PCIe Eq Pset value to use, range is 0-10");
691 
692 static uint pcie_ctle = 3; /* discrete on, integrated on */
693 module_param(pcie_ctle, uint, S_IRUGO);
694 MODULE_PARM_DESC(pcie_ctle, "PCIe static CTLE mode, bit 0 - discrete on/off, bit 1 - integrated on/off");
695 
696 /* equalization columns */
697 #define PREC 0
698 #define ATTN 1
699 #define POST 2
700 
701 /* discrete silicon preliminary equalization values */
702 static const u8 discrete_preliminary_eq[11][3] = {
703 	/* prec   attn   post */
704 	{  0x00,  0x00,  0x12 },	/* p0 */
705 	{  0x00,  0x00,  0x0c },	/* p1 */
706 	{  0x00,  0x00,  0x0f },	/* p2 */
707 	{  0x00,  0x00,  0x09 },	/* p3 */
708 	{  0x00,  0x00,  0x00 },	/* p4 */
709 	{  0x06,  0x00,  0x00 },	/* p5 */
710 	{  0x09,  0x00,  0x00 },	/* p6 */
711 	{  0x06,  0x00,  0x0f },	/* p7 */
712 	{  0x09,  0x00,  0x09 },	/* p8 */
713 	{  0x0c,  0x00,  0x00 },	/* p9 */
714 	{  0x00,  0x00,  0x18 },	/* p10 */
715 };
716 
717 /* integrated silicon preliminary equalization values */
718 static const u8 integrated_preliminary_eq[11][3] = {
719 	/* prec   attn   post */
720 	{  0x00,  0x1e,  0x07 },	/* p0 */
721 	{  0x00,  0x1e,  0x05 },	/* p1 */
722 	{  0x00,  0x1e,  0x06 },	/* p2 */
723 	{  0x00,  0x1e,  0x04 },	/* p3 */
724 	{  0x00,  0x1e,  0x00 },	/* p4 */
725 	{  0x03,  0x1e,  0x00 },	/* p5 */
726 	{  0x04,  0x1e,  0x00 },	/* p6 */
727 	{  0x03,  0x1e,  0x06 },	/* p7 */
728 	{  0x03,  0x1e,  0x04 },	/* p8 */
729 	{  0x05,  0x1e,  0x00 },	/* p9 */
730 	{  0x00,  0x1e,  0x0a },	/* p10 */
731 };
732 
733 static const u8 discrete_ctle_tunings[11][4] = {
734 	/* DC     LF     HF     BW */
735 	{  0x48,  0x0b,  0x04,  0x04 },	/* p0 */
736 	{  0x60,  0x05,  0x0f,  0x0a },	/* p1 */
737 	{  0x50,  0x09,  0x06,  0x06 },	/* p2 */
738 	{  0x68,  0x05,  0x0f,  0x0a },	/* p3 */
739 	{  0x80,  0x05,  0x0f,  0x0a },	/* p4 */
740 	{  0x70,  0x05,  0x0f,  0x0a },	/* p5 */
741 	{  0x68,  0x05,  0x0f,  0x0a },	/* p6 */
742 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
743 	{  0x48,  0x09,  0x06,  0x06 },	/* p8 */
744 	{  0x60,  0x05,  0x0f,  0x0a },	/* p9 */
745 	{  0x38,  0x0f,  0x00,  0x00 },	/* p10 */
746 };
747 
748 static const u8 integrated_ctle_tunings[11][4] = {
749 	/* DC     LF     HF     BW */
750 	{  0x38,  0x0f,  0x00,  0x00 },	/* p0 */
751 	{  0x38,  0x0f,  0x00,  0x00 },	/* p1 */
752 	{  0x38,  0x0f,  0x00,  0x00 },	/* p2 */
753 	{  0x38,  0x0f,  0x00,  0x00 },	/* p3 */
754 	{  0x58,  0x0a,  0x05,  0x05 },	/* p4 */
755 	{  0x48,  0x0a,  0x05,  0x05 },	/* p5 */
756 	{  0x40,  0x0a,  0x05,  0x05 },	/* p6 */
757 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
758 	{  0x38,  0x0f,  0x00,  0x00 },	/* p8 */
759 	{  0x38,  0x09,  0x06,  0x06 },	/* p9 */
760 	{  0x38,  0x0e,  0x01,  0x01 },	/* p10 */
761 };
762 
763 /* helper to format the value to write to hardware */
764 #define eq_value(pre, curr, post) \
765 	((((u32)(pre)) << \
766 			PCIE_CFG_REG_PL102_GEN3_EQ_PRE_CURSOR_PSET_SHIFT) \
767 	| (((u32)(curr)) << PCIE_CFG_REG_PL102_GEN3_EQ_CURSOR_PSET_SHIFT) \
768 	| (((u32)(post)) << \
769 		PCIE_CFG_REG_PL102_GEN3_EQ_POST_CURSOR_PSET_SHIFT))
770 
771 /*
772  * Load the given EQ preset table into the PCIe hardware.
773  */
774 static int load_eq_table(struct hfi1_devdata *dd, const u8 eq[11][3], u8 fs,
775 			 u8 div)
776 {
777 	struct pci_dev *pdev = dd->pcidev;
778 	u32 hit_error = 0;
779 	u32 violation;
780 	u32 i;
781 	u8 c_minus1, c0, c_plus1;
782 	int ret;
783 
784 	for (i = 0; i < 11; i++) {
785 		/* set index */
786 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL103, i);
787 		/* write the value */
788 		c_minus1 = eq[i][PREC] / div;
789 		c0 = fs - (eq[i][PREC] / div) - (eq[i][POST] / div);
790 		c_plus1 = eq[i][POST] / div;
791 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL102,
792 				       eq_value(c_minus1, c0, c_plus1));
793 		/* check if these coefficients violate EQ rules */
794 		ret = pci_read_config_dword(dd->pcidev,
795 					    PCIE_CFG_REG_PL105, &violation);
796 		if (ret) {
797 			dd_dev_err(dd, "Unable to read from PCI config\n");
798 			hit_error = 1;
799 			break;
800 		}
801 
802 		if (violation
803 		    & PCIE_CFG_REG_PL105_GEN3_EQ_VIOLATE_COEF_RULES_SMASK){
804 			if (hit_error == 0) {
805 				dd_dev_err(dd,
806 					   "Gen3 EQ Table Coefficient rule violations\n");
807 				dd_dev_err(dd, "         prec   attn   post\n");
808 			}
809 			dd_dev_err(dd, "   p%02d:   %02x     %02x     %02x\n",
810 				   i, (u32)eq[i][0], (u32)eq[i][1],
811 				   (u32)eq[i][2]);
812 			dd_dev_err(dd, "            %02x     %02x     %02x\n",
813 				   (u32)c_minus1, (u32)c0, (u32)c_plus1);
814 			hit_error = 1;
815 		}
816 	}
817 	if (hit_error)
818 		return -EINVAL;
819 	return 0;
820 }
821 
822 /*
823  * Steps to be done after the PCIe firmware is downloaded and
824  * before the SBR for the Pcie Gen3.
825  * The SBus resource is already being held.
826  */
827 static void pcie_post_steps(struct hfi1_devdata *dd)
828 {
829 	int i;
830 
831 	set_sbus_fast_mode(dd);
832 	/*
833 	 * Write to the PCIe PCSes to set the G3_LOCKED_NEXT bits to 1.
834 	 * This avoids a spurious framing error that can otherwise be
835 	 * generated by the MAC layer.
836 	 *
837 	 * Use individual addresses since no broadcast is set up.
838 	 */
839 	for (i = 0; i < NUM_PCIE_SERDES; i++) {
840 		sbus_request(dd, pcie_pcs_addrs[dd->hfi1_id][i],
841 			     0x03, WRITE_SBUS_RECEIVER, 0x00022132);
842 	}
843 
844 	clear_sbus_fast_mode(dd);
845 }
846 
847 /*
848  * Trigger a secondary bus reset (SBR) on ourselves using our parent.
849  *
850  * Based on pci_parent_bus_reset() which is not exported by the
851  * kernel core.
852  */
853 static int trigger_sbr(struct hfi1_devdata *dd)
854 {
855 	struct pci_dev *dev = dd->pcidev;
856 	struct pci_dev *pdev;
857 
858 	/* need a parent */
859 	if (!dev->bus->self) {
860 		dd_dev_err(dd, "%s: no parent device\n", __func__);
861 		return -ENOTTY;
862 	}
863 
864 	/* should not be anyone else on the bus */
865 	list_for_each_entry(pdev, &dev->bus->devices, bus_list)
866 		if (pdev != dev) {
867 			dd_dev_err(dd,
868 				   "%s: another device is on the same bus\n",
869 				   __func__);
870 			return -ENOTTY;
871 		}
872 
873 	/*
874 	 * This is an end around to do an SBR during probe time. A new API needs
875 	 * to be implemented to have cleaner interface but this fixes the
876 	 * current brokenness
877 	 */
878 	return pci_bridge_secondary_bus_reset(dev->bus->self);
879 }
880 
881 /*
882  * Write the given gasket interrupt register.
883  */
884 static void write_gasket_interrupt(struct hfi1_devdata *dd, int index,
885 				   u16 code, u16 data)
886 {
887 	write_csr(dd, ASIC_PCIE_SD_INTRPT_LIST + (index * 8),
888 		  (((u64)code << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_CODE_SHIFT) |
889 		   ((u64)data << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_DATA_SHIFT)));
890 }
891 
892 /*
893  * Tell the gasket logic how to react to the reset.
894  */
895 static void arm_gasket_logic(struct hfi1_devdata *dd)
896 {
897 	u64 reg;
898 
899 	reg = (((u64)1 << dd->hfi1_id) <<
900 	       ASIC_PCIE_SD_HOST_CMD_INTRPT_CMD_SHIFT) |
901 	      ((u64)pcie_serdes_broadcast[dd->hfi1_id] <<
902 	       ASIC_PCIE_SD_HOST_CMD_SBUS_RCVR_ADDR_SHIFT |
903 	       ASIC_PCIE_SD_HOST_CMD_SBR_MODE_SMASK |
904 	       ((u64)SBR_DELAY_US & ASIC_PCIE_SD_HOST_CMD_TIMER_MASK) <<
905 	       ASIC_PCIE_SD_HOST_CMD_TIMER_SHIFT);
906 	write_csr(dd, ASIC_PCIE_SD_HOST_CMD, reg);
907 	/* read back to push the write */
908 	read_csr(dd, ASIC_PCIE_SD_HOST_CMD);
909 }
910 
911 /*
912  * CCE_PCIE_CTRL long name helpers
913  * We redefine these shorter macros to use in the code while leaving
914  * chip_registers.h to be autogenerated from the hardware spec.
915  */
916 #define LANE_BUNDLE_MASK              CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_MASK
917 #define LANE_BUNDLE_SHIFT             CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_SHIFT
918 #define LANE_DELAY_MASK               CCE_PCIE_CTRL_PCIE_LANE_DELAY_MASK
919 #define LANE_DELAY_SHIFT              CCE_PCIE_CTRL_PCIE_LANE_DELAY_SHIFT
920 #define MARGIN_OVERWRITE_ENABLE_SHIFT CCE_PCIE_CTRL_XMT_MARGIN_OVERWRITE_ENABLE_SHIFT
921 #define MARGIN_SHIFT                  CCE_PCIE_CTRL_XMT_MARGIN_SHIFT
922 #define MARGIN_G1_G2_OVERWRITE_MASK   CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_MASK
923 #define MARGIN_G1_G2_OVERWRITE_SHIFT  CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_SHIFT
924 #define MARGIN_GEN1_GEN2_MASK         CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_MASK
925 #define MARGIN_GEN1_GEN2_SHIFT        CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_SHIFT
926 
927  /*
928   * Write xmt_margin for full-swing (WFR-B) or half-swing (WFR-C).
929   */
930 static void write_xmt_margin(struct hfi1_devdata *dd, const char *fname)
931 {
932 	u64 pcie_ctrl;
933 	u64 xmt_margin;
934 	u64 xmt_margin_oe;
935 	u64 lane_delay;
936 	u64 lane_bundle;
937 
938 	pcie_ctrl = read_csr(dd, CCE_PCIE_CTRL);
939 
940 	/*
941 	 * For Discrete, use full-swing.
942 	 *  - PCIe TX defaults to full-swing.
943 	 *    Leave this register as default.
944 	 * For Integrated, use half-swing
945 	 *  - Copy xmt_margin and xmt_margin_oe
946 	 *    from Gen1/Gen2 to Gen3.
947 	 */
948 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL1) { /* integrated */
949 		/* extract initial fields */
950 		xmt_margin = (pcie_ctrl >> MARGIN_GEN1_GEN2_SHIFT)
951 			      & MARGIN_GEN1_GEN2_MASK;
952 		xmt_margin_oe = (pcie_ctrl >> MARGIN_G1_G2_OVERWRITE_SHIFT)
953 				 & MARGIN_G1_G2_OVERWRITE_MASK;
954 		lane_delay = (pcie_ctrl >> LANE_DELAY_SHIFT) & LANE_DELAY_MASK;
955 		lane_bundle = (pcie_ctrl >> LANE_BUNDLE_SHIFT)
956 			       & LANE_BUNDLE_MASK;
957 
958 		/*
959 		 * For A0, EFUSE values are not set.  Override with the
960 		 * correct values.
961 		 */
962 		if (is_ax(dd)) {
963 			/*
964 			 * xmt_margin and OverwiteEnabel should be the
965 			 * same for Gen1/Gen2 and Gen3
966 			 */
967 			xmt_margin = 0x5;
968 			xmt_margin_oe = 0x1;
969 			lane_delay = 0xF; /* Delay 240ns. */
970 			lane_bundle = 0x0; /* Set to 1 lane. */
971 		}
972 
973 		/* overwrite existing values */
974 		pcie_ctrl = (xmt_margin << MARGIN_GEN1_GEN2_SHIFT)
975 			| (xmt_margin_oe << MARGIN_G1_G2_OVERWRITE_SHIFT)
976 			| (xmt_margin << MARGIN_SHIFT)
977 			| (xmt_margin_oe << MARGIN_OVERWRITE_ENABLE_SHIFT)
978 			| (lane_delay << LANE_DELAY_SHIFT)
979 			| (lane_bundle << LANE_BUNDLE_SHIFT);
980 
981 		write_csr(dd, CCE_PCIE_CTRL, pcie_ctrl);
982 	}
983 
984 	dd_dev_dbg(dd, "%s: program XMT margin, CcePcieCtrl 0x%llx\n",
985 		   fname, pcie_ctrl);
986 }
987 
988 /*
989  * Do all the steps needed to transition the PCIe link to Gen3 speed.
990  */
991 int do_pcie_gen3_transition(struct hfi1_devdata *dd)
992 {
993 	struct pci_dev *parent = dd->pcidev->bus->self;
994 	u64 fw_ctrl;
995 	u64 reg, therm;
996 	u32 reg32, fs, lf;
997 	u32 status, err;
998 	int ret;
999 	int do_retry, retry_count = 0;
1000 	int intnum = 0;
1001 	uint default_pset;
1002 	uint pset = pcie_pset;
1003 	u16 target_vector, target_speed;
1004 	u16 lnkctl2, vendor;
1005 	u8 div;
1006 	const u8 (*eq)[3];
1007 	const u8 (*ctle_tunings)[4];
1008 	uint static_ctle_mode;
1009 	int return_error = 0;
1010 	u32 target_width;
1011 
1012 	/* PCIe Gen3 is for the ASIC only */
1013 	if (dd->icode != ICODE_RTL_SILICON)
1014 		return 0;
1015 
1016 	if (pcie_target == 1) {			/* target Gen1 */
1017 		target_vector = PCI_EXP_LNKCTL2_TLS_2_5GT;
1018 		target_speed = 2500;
1019 	} else if (pcie_target == 2) {		/* target Gen2 */
1020 		target_vector = PCI_EXP_LNKCTL2_TLS_5_0GT;
1021 		target_speed = 5000;
1022 	} else if (pcie_target == 3) {		/* target Gen3 */
1023 		target_vector = PCI_EXP_LNKCTL2_TLS_8_0GT;
1024 		target_speed = 8000;
1025 	} else {
1026 		/* off or invalid target - skip */
1027 		dd_dev_info(dd, "%s: Skipping PCIe transition\n", __func__);
1028 		return 0;
1029 	}
1030 
1031 	/* if already at target speed, done (unless forced) */
1032 	if (dd->lbus_speed == target_speed) {
1033 		dd_dev_info(dd, "%s: PCIe already at gen%d, %s\n", __func__,
1034 			    pcie_target,
1035 			    pcie_force ? "re-doing anyway" : "skipping");
1036 		if (!pcie_force)
1037 			return 0;
1038 	}
1039 
1040 	/*
1041 	 * The driver cannot do the transition if it has no access to the
1042 	 * upstream component
1043 	 */
1044 	if (!parent) {
1045 		dd_dev_info(dd, "%s: No upstream, Can't do gen3 transition\n",
1046 			    __func__);
1047 		return 0;
1048 	}
1049 
1050 	/* Previous Gen1/Gen2 bus width */
1051 	target_width = dd->lbus_width;
1052 
1053 	/*
1054 	 * Do the Gen3 transition.  Steps are those of the PCIe Gen3
1055 	 * recipe.
1056 	 */
1057 
1058 	/* step 1: pcie link working in gen1/gen2 */
1059 
1060 	/* step 2: if either side is not capable of Gen3, done */
1061 	if (pcie_target == 3 && !dd->link_gen3_capable) {
1062 		dd_dev_err(dd, "The PCIe link is not Gen3 capable\n");
1063 		ret = -ENOSYS;
1064 		goto done_no_mutex;
1065 	}
1066 
1067 	/* hold the SBus resource across the firmware download and SBR */
1068 	ret = acquire_chip_resource(dd, CR_SBUS, SBUS_TIMEOUT);
1069 	if (ret) {
1070 		dd_dev_err(dd, "%s: unable to acquire SBus resource\n",
1071 			   __func__);
1072 		return ret;
1073 	}
1074 
1075 	/* make sure thermal polling is not causing interrupts */
1076 	therm = read_csr(dd, ASIC_CFG_THERM_POLL_EN);
1077 	if (therm) {
1078 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x0);
1079 		msleep(100);
1080 		dd_dev_info(dd, "%s: Disabled therm polling\n",
1081 			    __func__);
1082 	}
1083 
1084 retry:
1085 	/* the SBus download will reset the spico for thermal */
1086 
1087 	/* step 3: download SBus Master firmware */
1088 	/* step 4: download PCIe Gen3 SerDes firmware */
1089 	dd_dev_info(dd, "%s: downloading firmware\n", __func__);
1090 	ret = load_pcie_firmware(dd);
1091 	if (ret) {
1092 		/* do not proceed if the firmware cannot be downloaded */
1093 		return_error = 1;
1094 		goto done;
1095 	}
1096 
1097 	/* step 5: set up device parameter settings */
1098 	dd_dev_info(dd, "%s: setting PCIe registers\n", __func__);
1099 
1100 	/*
1101 	 * PcieCfgSpcie1 - Link Control 3
1102 	 * Leave at reset value.  No need to set PerfEq - link equalization
1103 	 * will be performed automatically after the SBR when the target
1104 	 * speed is 8GT/s.
1105 	 */
1106 
1107 	/* clear all 16 per-lane error bits (PCIe: Lane Error Status) */
1108 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, 0xffff);
1109 
1110 	/* step 5a: Set Synopsys Port Logic registers */
1111 
1112 	/*
1113 	 * PcieCfgRegPl2 - Port Force Link
1114 	 *
1115 	 * Set the low power field to 0x10 to avoid unnecessary power
1116 	 * management messages.  All other fields are zero.
1117 	 */
1118 	reg32 = 0x10ul << PCIE_CFG_REG_PL2_LOW_PWR_ENT_CNT_SHIFT;
1119 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL2, reg32);
1120 
1121 	/*
1122 	 * PcieCfgRegPl100 - Gen3 Control
1123 	 *
1124 	 * turn off PcieCfgRegPl100.Gen3ZRxDcNonCompl
1125 	 * turn on PcieCfgRegPl100.EqEieosCnt
1126 	 * Everything else zero.
1127 	 */
1128 	reg32 = PCIE_CFG_REG_PL100_EQ_EIEOS_CNT_SMASK;
1129 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL100, reg32);
1130 
1131 	/*
1132 	 * PcieCfgRegPl101 - Gen3 EQ FS and LF
1133 	 * PcieCfgRegPl102 - Gen3 EQ Presets to Coefficients Mapping
1134 	 * PcieCfgRegPl103 - Gen3 EQ Preset Index
1135 	 * PcieCfgRegPl105 - Gen3 EQ Status
1136 	 *
1137 	 * Give initial EQ settings.
1138 	 */
1139 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL0) { /* discrete */
1140 		/* 1000mV, FS=24, LF = 8 */
1141 		fs = 24;
1142 		lf = 8;
1143 		div = 3;
1144 		eq = discrete_preliminary_eq;
1145 		default_pset = DEFAULT_DISCRETE_PSET;
1146 		ctle_tunings = discrete_ctle_tunings;
1147 		/* bit 0 - discrete on/off */
1148 		static_ctle_mode = pcie_ctle & 0x1;
1149 	} else {
1150 		/* 400mV, FS=29, LF = 9 */
1151 		fs = 29;
1152 		lf = 9;
1153 		div = 1;
1154 		eq = integrated_preliminary_eq;
1155 		default_pset = DEFAULT_MCP_PSET;
1156 		ctle_tunings = integrated_ctle_tunings;
1157 		/* bit 1 - integrated on/off */
1158 		static_ctle_mode = (pcie_ctle >> 1) & 0x1;
1159 	}
1160 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL101,
1161 			       (fs <<
1162 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_FS_SHIFT) |
1163 			       (lf <<
1164 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_LF_SHIFT));
1165 	ret = load_eq_table(dd, eq, fs, div);
1166 	if (ret)
1167 		goto done;
1168 
1169 	/*
1170 	 * PcieCfgRegPl106 - Gen3 EQ Control
1171 	 *
1172 	 * Set Gen3EqPsetReqVec, leave other fields 0.
1173 	 */
1174 	if (pset == UNSET_PSET)
1175 		pset = default_pset;
1176 	if (pset > 10) {	/* valid range is 0-10, inclusive */
1177 		dd_dev_err(dd, "%s: Invalid Eq Pset %u, setting to %d\n",
1178 			   __func__, pset, default_pset);
1179 		pset = default_pset;
1180 	}
1181 	dd_dev_info(dd, "%s: using EQ Pset %u\n", __func__, pset);
1182 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL106,
1183 			       ((1 << pset) <<
1184 			PCIE_CFG_REG_PL106_GEN3_EQ_PSET_REQ_VEC_SHIFT) |
1185 			PCIE_CFG_REG_PL106_GEN3_EQ_EVAL2MS_DISABLE_SMASK |
1186 			PCIE_CFG_REG_PL106_GEN3_EQ_PHASE23_EXIT_MODE_SMASK);
1187 
1188 	/*
1189 	 * step 5b: Do post firmware download steps via SBus
1190 	 */
1191 	dd_dev_info(dd, "%s: doing pcie post steps\n", __func__);
1192 	pcie_post_steps(dd);
1193 
1194 	/*
1195 	 * step 5c: Program gasket interrupts
1196 	 */
1197 	/* set the Rx Bit Rate to REFCLK ratio */
1198 	write_gasket_interrupt(dd, intnum++, 0x0006, 0x0050);
1199 	/* disable pCal for PCIe Gen3 RX equalization */
1200 	/* select adaptive or static CTLE */
1201 	write_gasket_interrupt(dd, intnum++, 0x0026,
1202 			       0x5b01 | (static_ctle_mode << 3));
1203 	/*
1204 	 * Enable iCal for PCIe Gen3 RX equalization, and set which
1205 	 * evaluation of RX_EQ_EVAL will launch the iCal procedure.
1206 	 */
1207 	write_gasket_interrupt(dd, intnum++, 0x0026, 0x5202);
1208 
1209 	if (static_ctle_mode) {
1210 		/* apply static CTLE tunings */
1211 		u8 pcie_dc, pcie_lf, pcie_hf, pcie_bw;
1212 
1213 		pcie_dc = ctle_tunings[pset][0];
1214 		pcie_lf = ctle_tunings[pset][1];
1215 		pcie_hf = ctle_tunings[pset][2];
1216 		pcie_bw = ctle_tunings[pset][3];
1217 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0200 | pcie_dc);
1218 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0100 | pcie_lf);
1219 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0000 | pcie_hf);
1220 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x5500 | pcie_bw);
1221 	}
1222 
1223 	/* terminate list */
1224 	write_gasket_interrupt(dd, intnum++, 0x0000, 0x0000);
1225 
1226 	/*
1227 	 * step 5d: program XMT margin
1228 	 */
1229 	write_xmt_margin(dd, __func__);
1230 
1231 	/*
1232 	 * step 5e: disable active state power management (ASPM). It
1233 	 * will be enabled if required later
1234 	 */
1235 	dd_dev_info(dd, "%s: clearing ASPM\n", __func__);
1236 	aspm_hw_disable_l1(dd);
1237 
1238 	/*
1239 	 * step 5f: clear DirectSpeedChange
1240 	 * PcieCfgRegPl67.DirectSpeedChange must be zero to prevent the
1241 	 * change in the speed target from starting before we are ready.
1242 	 * This field defaults to 0 and we are not changing it, so nothing
1243 	 * needs to be done.
1244 	 */
1245 
1246 	/* step 5g: Set target link speed */
1247 	/*
1248 	 * Set target link speed to be target on both device and parent.
1249 	 * On setting the parent: Some system BIOSs "helpfully" set the
1250 	 * parent target speed to Gen2 to match the ASIC's initial speed.
1251 	 * We can set the target Gen3 because we have already checked
1252 	 * that it is Gen3 capable earlier.
1253 	 */
1254 	dd_dev_info(dd, "%s: setting parent target link speed\n", __func__);
1255 	ret = pcie_capability_read_word(parent, PCI_EXP_LNKCTL2, &lnkctl2);
1256 	if (ret) {
1257 		dd_dev_err(dd, "Unable to read from PCI config\n");
1258 		return_error = 1;
1259 		goto done;
1260 	}
1261 
1262 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1263 		    (u32)lnkctl2);
1264 	/* only write to parent if target is not as high as ours */
1265 	if ((lnkctl2 & PCI_EXP_LNKCTL2_TLS) < target_vector) {
1266 		lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1267 		lnkctl2 |= target_vector;
1268 		dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1269 			    (u32)lnkctl2);
1270 		ret = pcie_capability_write_word(parent,
1271 						 PCI_EXP_LNKCTL2, lnkctl2);
1272 		if (ret) {
1273 			dd_dev_err(dd, "Unable to write to PCI config\n");
1274 			return_error = 1;
1275 			goto done;
1276 		}
1277 	} else {
1278 		dd_dev_info(dd, "%s: ..target speed is OK\n", __func__);
1279 	}
1280 
1281 	dd_dev_info(dd, "%s: setting target link speed\n", __func__);
1282 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL2, &lnkctl2);
1283 	if (ret) {
1284 		dd_dev_err(dd, "Unable to read from PCI config\n");
1285 		return_error = 1;
1286 		goto done;
1287 	}
1288 
1289 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1290 		    (u32)lnkctl2);
1291 	lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1292 	lnkctl2 |= target_vector;
1293 	dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1294 		    (u32)lnkctl2);
1295 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL2, lnkctl2);
1296 	if (ret) {
1297 		dd_dev_err(dd, "Unable to write to PCI config\n");
1298 		return_error = 1;
1299 		goto done;
1300 	}
1301 
1302 	/* step 5h: arm gasket logic */
1303 	/* hold DC in reset across the SBR */
1304 	write_csr(dd, CCE_DC_CTRL, CCE_DC_CTRL_DC_RESET_SMASK);
1305 	(void)read_csr(dd, CCE_DC_CTRL); /* DC reset hold */
1306 	/* save firmware control across the SBR */
1307 	fw_ctrl = read_csr(dd, MISC_CFG_FW_CTRL);
1308 
1309 	dd_dev_info(dd, "%s: arming gasket logic\n", __func__);
1310 	arm_gasket_logic(dd);
1311 
1312 	/*
1313 	 * step 6: quiesce PCIe link
1314 	 * The chip has already been reset, so there will be no traffic
1315 	 * from the chip.  Linux has no easy way to enforce that it will
1316 	 * not try to access the device, so we just need to hope it doesn't
1317 	 * do it while we are doing the reset.
1318 	 */
1319 
1320 	/*
1321 	 * step 7: initiate the secondary bus reset (SBR)
1322 	 * step 8: hardware brings the links back up
1323 	 * step 9: wait for link speed transition to be complete
1324 	 */
1325 	dd_dev_info(dd, "%s: calling trigger_sbr\n", __func__);
1326 	ret = trigger_sbr(dd);
1327 	if (ret)
1328 		goto done;
1329 
1330 	/* step 10: decide what to do next */
1331 
1332 	/* check if we can read PCI space */
1333 	ret = pci_read_config_word(dd->pcidev, PCI_VENDOR_ID, &vendor);
1334 	if (ret) {
1335 		dd_dev_info(dd,
1336 			    "%s: read of VendorID failed after SBR, err %d\n",
1337 			    __func__, ret);
1338 		return_error = 1;
1339 		goto done;
1340 	}
1341 	if (vendor == 0xffff) {
1342 		dd_dev_info(dd, "%s: VendorID is all 1s after SBR\n", __func__);
1343 		return_error = 1;
1344 		ret = -EIO;
1345 		goto done;
1346 	}
1347 
1348 	/* restore PCI space registers we know were reset */
1349 	dd_dev_info(dd, "%s: calling restore_pci_variables\n", __func__);
1350 	ret = restore_pci_variables(dd);
1351 	if (ret) {
1352 		dd_dev_err(dd, "%s: Could not restore PCI variables\n",
1353 			   __func__);
1354 		return_error = 1;
1355 		goto done;
1356 	}
1357 
1358 	/* restore firmware control */
1359 	write_csr(dd, MISC_CFG_FW_CTRL, fw_ctrl);
1360 
1361 	/*
1362 	 * Check the gasket block status.
1363 	 *
1364 	 * This is the first CSR read after the SBR.  If the read returns
1365 	 * all 1s (fails), the link did not make it back.
1366 	 *
1367 	 * Once we're sure we can read and write, clear the DC reset after
1368 	 * the SBR.  Then check for any per-lane errors. Then look over
1369 	 * the status.
1370 	 */
1371 	reg = read_csr(dd, ASIC_PCIE_SD_HOST_STATUS);
1372 	dd_dev_info(dd, "%s: gasket block status: 0x%llx\n", __func__, reg);
1373 	if (reg == ~0ull) {	/* PCIe read failed/timeout */
1374 		dd_dev_err(dd, "SBR failed - unable to read from device\n");
1375 		return_error = 1;
1376 		ret = -ENOSYS;
1377 		goto done;
1378 	}
1379 
1380 	/* clear the DC reset */
1381 	write_csr(dd, CCE_DC_CTRL, 0);
1382 
1383 	/* Set the LED off */
1384 	setextled(dd, 0);
1385 
1386 	/* check for any per-lane errors */
1387 	ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, &reg32);
1388 	if (ret) {
1389 		dd_dev_err(dd, "Unable to read from PCI config\n");
1390 		return_error = 1;
1391 		goto done;
1392 	}
1393 
1394 	dd_dev_info(dd, "%s: per-lane errors: 0x%x\n", __func__, reg32);
1395 
1396 	/* extract status, look for our HFI */
1397 	status = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_SHIFT)
1398 			& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_MASK;
1399 	if ((status & (1 << dd->hfi1_id)) == 0) {
1400 		dd_dev_err(dd,
1401 			   "%s: gasket status 0x%x, expecting 0x%x\n",
1402 			   __func__, status, 1 << dd->hfi1_id);
1403 		ret = -EIO;
1404 		goto done;
1405 	}
1406 
1407 	/* extract error */
1408 	err = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_SHIFT)
1409 		& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_MASK;
1410 	if (err) {
1411 		dd_dev_err(dd, "%s: gasket error %d\n", __func__, err);
1412 		ret = -EIO;
1413 		goto done;
1414 	}
1415 
1416 	/* update our link information cache */
1417 	update_lbus_info(dd);
1418 	dd_dev_info(dd, "%s: new speed and width: %s\n", __func__,
1419 		    dd->lbus_info);
1420 
1421 	if (dd->lbus_speed != target_speed ||
1422 	    dd->lbus_width < target_width) { /* not target */
1423 		/* maybe retry */
1424 		do_retry = retry_count < pcie_retry;
1425 		dd_dev_err(dd, "PCIe link speed or width did not match target%s\n",
1426 			   do_retry ? ", retrying" : "");
1427 		retry_count++;
1428 		if (do_retry) {
1429 			msleep(100); /* allow time to settle */
1430 			goto retry;
1431 		}
1432 		ret = -EIO;
1433 	}
1434 
1435 done:
1436 	if (therm) {
1437 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x1);
1438 		msleep(100);
1439 		dd_dev_info(dd, "%s: Re-enable therm polling\n",
1440 			    __func__);
1441 	}
1442 	release_chip_resource(dd, CR_SBUS);
1443 done_no_mutex:
1444 	/* return no error if it is OK to be at current speed */
1445 	if (ret && !return_error) {
1446 		dd_dev_err(dd, "Proceeding at current speed PCIe speed\n");
1447 		ret = 0;
1448 	}
1449 
1450 	dd_dev_info(dd, "%s: done\n", __func__);
1451 	return ret;
1452 }
1453