xref: /linux/drivers/infiniband/hw/hfi1/pcie.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
1 /*
2  * Copyright(c) 2015 - 2017 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  * Code to adjust PCIe capabilities.
65  */
66 static void tune_pcie_caps(struct hfi1_devdata *);
67 
68 /*
69  * Do all the common PCIe setup and initialization.
70  * devdata is not yet allocated, and is not allocated until after this
71  * routine returns success.  Therefore dd_dev_err() can't be used for error
72  * printing.
73  */
74 int hfi1_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
75 {
76 	int ret;
77 
78 	ret = pci_enable_device(pdev);
79 	if (ret) {
80 		/*
81 		 * This can happen (in theory) iff:
82 		 * We did a chip reset, and then failed to reprogram the
83 		 * BAR, or the chip reset due to an internal error.  We then
84 		 * unloaded the driver and reloaded it.
85 		 *
86 		 * Both reset cases set the BAR back to initial state.  For
87 		 * the latter case, the AER sticky error bit at offset 0x718
88 		 * should be set, but the Linux kernel doesn't yet know
89 		 * about that, it appears.  If the original BAR was retained
90 		 * in the kernel data structures, this may be OK.
91 		 */
92 		hfi1_early_err(&pdev->dev, "pci enable failed: error %d\n",
93 			       -ret);
94 		goto done;
95 	}
96 
97 	ret = pci_request_regions(pdev, DRIVER_NAME);
98 	if (ret) {
99 		hfi1_early_err(&pdev->dev,
100 			       "pci_request_regions fails: err %d\n", -ret);
101 		goto bail;
102 	}
103 
104 	ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
105 	if (ret) {
106 		/*
107 		 * If the 64 bit setup fails, try 32 bit.  Some systems
108 		 * do not setup 64 bit maps on systems with 2GB or less
109 		 * memory installed.
110 		 */
111 		ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
112 		if (ret) {
113 			hfi1_early_err(&pdev->dev,
114 				       "Unable to set DMA mask: %d\n", ret);
115 			goto bail;
116 		}
117 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
118 	} else {
119 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
120 	}
121 	if (ret) {
122 		hfi1_early_err(&pdev->dev,
123 			       "Unable to set DMA consistent mask: %d\n", ret);
124 		goto bail;
125 	}
126 
127 	pci_set_master(pdev);
128 	(void)pci_enable_pcie_error_reporting(pdev);
129 	goto done;
130 
131 bail:
132 	hfi1_pcie_cleanup(pdev);
133 done:
134 	return ret;
135 }
136 
137 /*
138  * Clean what was done in hfi1_pcie_init()
139  */
140 void hfi1_pcie_cleanup(struct pci_dev *pdev)
141 {
142 	pci_disable_device(pdev);
143 	/*
144 	 * Release regions should be called after the disable. OK to
145 	 * call if request regions has not been called or failed.
146 	 */
147 	pci_release_regions(pdev);
148 }
149 
150 /*
151  * Do remaining PCIe setup, once dd is allocated, and save away
152  * fields required to re-initialize after a chip reset, or for
153  * various other purposes
154  */
155 int hfi1_pcie_ddinit(struct hfi1_devdata *dd, struct pci_dev *pdev)
156 {
157 	unsigned long len;
158 	resource_size_t addr;
159 	int ret = 0;
160 	u32 rcv_array_count;
161 
162 	addr = pci_resource_start(pdev, 0);
163 	len = pci_resource_len(pdev, 0);
164 
165 	/*
166 	 * The TXE PIO buffers are at the tail end of the chip space.
167 	 * Cut them off and map them separately.
168 	 */
169 
170 	/* sanity check vs expectations */
171 	if (len != TXE_PIO_SEND + TXE_PIO_SIZE) {
172 		dd_dev_err(dd, "chip PIO range does not match\n");
173 		return -EINVAL;
174 	}
175 
176 	dd->kregbase1 = ioremap_nocache(addr, RCV_ARRAY);
177 	if (!dd->kregbase1) {
178 		dd_dev_err(dd, "UC mapping of kregbase1 failed\n");
179 		return -ENOMEM;
180 	}
181 	dd_dev_info(dd, "UC base1: %p for %x\n", dd->kregbase1, RCV_ARRAY);
182 
183 	/* verify that reads actually work, save revision for reset check */
184 	dd->revision = readq(dd->kregbase1 + CCE_REVISION);
185 	if (dd->revision == ~(u64)0) {
186 		dd_dev_err(dd, "Cannot read chip CSRs\n");
187 		goto nomem;
188 	}
189 
190 	rcv_array_count = readq(dd->kregbase1 + RCV_ARRAY_CNT);
191 	dd_dev_info(dd, "RcvArray count: %u\n", rcv_array_count);
192 	dd->base2_start  = RCV_ARRAY + rcv_array_count * 8;
193 
194 	dd->kregbase2 = ioremap_nocache(
195 		addr + dd->base2_start,
196 		TXE_PIO_SEND - dd->base2_start);
197 	if (!dd->kregbase2) {
198 		dd_dev_err(dd, "UC mapping of kregbase2 failed\n");
199 		goto nomem;
200 	}
201 	dd_dev_info(dd, "UC base2: %p for %x\n", dd->kregbase2,
202 		    TXE_PIO_SEND - dd->base2_start);
203 
204 	dd->piobase = ioremap_wc(addr + TXE_PIO_SEND, TXE_PIO_SIZE);
205 	if (!dd->piobase) {
206 		dd_dev_err(dd, "WC mapping of send buffers failed\n");
207 		goto nomem;
208 	}
209 	dd_dev_info(dd, "WC piobase: %p\n for %x", dd->piobase, TXE_PIO_SIZE);
210 
211 	dd->physaddr = addr;        /* used for io_remap, etc. */
212 
213 	/*
214 	 * Map the chip's RcvArray as write-combining to allow us
215 	 * to write an entire cacheline worth of entries in one shot.
216 	 */
217 	dd->rcvarray_wc = ioremap_wc(addr + RCV_ARRAY,
218 				     rcv_array_count * 8);
219 	if (!dd->rcvarray_wc) {
220 		dd_dev_err(dd, "WC mapping of receive array failed\n");
221 		goto nomem;
222 	}
223 	dd_dev_info(dd, "WC RcvArray: %p for %x\n",
224 		    dd->rcvarray_wc, rcv_array_count * 8);
225 
226 	dd->flags |= HFI1_PRESENT;	/* chip.c CSR routines now work */
227 	return 0;
228 nomem:
229 	ret = -ENOMEM;
230 	hfi1_pcie_ddcleanup(dd);
231 	return ret;
232 }
233 
234 /*
235  * Do PCIe cleanup related to dd, after chip-specific cleanup, etc.  Just prior
236  * to releasing the dd memory.
237  * Void because all of the core pcie cleanup functions are void.
238  */
239 void hfi1_pcie_ddcleanup(struct hfi1_devdata *dd)
240 {
241 	dd->flags &= ~HFI1_PRESENT;
242 	if (dd->kregbase1)
243 		iounmap(dd->kregbase1);
244 	dd->kregbase1 = NULL;
245 	if (dd->kregbase2)
246 		iounmap(dd->kregbase2);
247 	dd->kregbase2 = NULL;
248 	if (dd->rcvarray_wc)
249 		iounmap(dd->rcvarray_wc);
250 	dd->rcvarray_wc = NULL;
251 	if (dd->piobase)
252 		iounmap(dd->piobase);
253 	dd->piobase = NULL;
254 }
255 
256 /* return the PCIe link speed from the given link status */
257 static u32 extract_speed(u16 linkstat)
258 {
259 	u32 speed;
260 
261 	switch (linkstat & PCI_EXP_LNKSTA_CLS) {
262 	default: /* not defined, assume Gen1 */
263 	case PCI_EXP_LNKSTA_CLS_2_5GB:
264 		speed = 2500; /* Gen 1, 2.5GHz */
265 		break;
266 	case PCI_EXP_LNKSTA_CLS_5_0GB:
267 		speed = 5000; /* Gen 2, 5GHz */
268 		break;
269 	case PCI_EXP_LNKSTA_CLS_8_0GB:
270 		speed = 8000; /* Gen 3, 8GHz */
271 		break;
272 	}
273 	return speed;
274 }
275 
276 /* return the PCIe link speed from the given link status */
277 static u32 extract_width(u16 linkstat)
278 {
279 	return (linkstat & PCI_EXP_LNKSTA_NLW) >> PCI_EXP_LNKSTA_NLW_SHIFT;
280 }
281 
282 /* read the link status and set dd->{lbus_width,lbus_speed,lbus_info} */
283 static void update_lbus_info(struct hfi1_devdata *dd)
284 {
285 	u16 linkstat;
286 	int ret;
287 
288 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
289 	if (ret) {
290 		dd_dev_err(dd, "Unable to read from PCI config\n");
291 		return;
292 	}
293 
294 	dd->lbus_width = extract_width(linkstat);
295 	dd->lbus_speed = extract_speed(linkstat);
296 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
297 		 "PCIe,%uMHz,x%u", dd->lbus_speed, dd->lbus_width);
298 }
299 
300 /*
301  * Read in the current PCIe link width and speed.  Find if the link is
302  * Gen3 capable.
303  */
304 int pcie_speeds(struct hfi1_devdata *dd)
305 {
306 	u32 linkcap;
307 	struct pci_dev *parent = dd->pcidev->bus->self;
308 	int ret;
309 
310 	if (!pci_is_pcie(dd->pcidev)) {
311 		dd_dev_err(dd, "Can't find PCI Express capability!\n");
312 		return -EINVAL;
313 	}
314 
315 	/* find if our max speed is Gen3 and parent supports Gen3 speeds */
316 	dd->link_gen3_capable = 1;
317 
318 	ret = pcie_capability_read_dword(dd->pcidev, PCI_EXP_LNKCAP, &linkcap);
319 	if (ret) {
320 		dd_dev_err(dd, "Unable to read from PCI config\n");
321 		return ret;
322 	}
323 
324 	if ((linkcap & PCI_EXP_LNKCAP_SLS) != PCI_EXP_LNKCAP_SLS_8_0GB) {
325 		dd_dev_info(dd,
326 			    "This HFI is not Gen3 capable, max speed 0x%x, need 0x3\n",
327 			    linkcap & PCI_EXP_LNKCAP_SLS);
328 		dd->link_gen3_capable = 0;
329 	}
330 
331 	/*
332 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
333 	 */
334 	if (parent && dd->pcidev->bus->max_bus_speed != PCIE_SPEED_8_0GT) {
335 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
336 		dd->link_gen3_capable = 0;
337 	}
338 
339 	/* obtain the link width and current speed */
340 	update_lbus_info(dd);
341 
342 	dd_dev_info(dd, "%s\n", dd->lbus_info);
343 
344 	return 0;
345 }
346 
347 /*
348  * Returns:
349  *	- actual number of interrupts allocated or
350  *      - error
351  */
352 int request_msix(struct hfi1_devdata *dd, u32 msireq)
353 {
354 	int nvec;
355 
356 	nvec = pci_alloc_irq_vectors(dd->pcidev, msireq, msireq, PCI_IRQ_MSIX);
357 	if (nvec < 0) {
358 		dd_dev_err(dd, "pci_alloc_irq_vectors() failed: %d\n", nvec);
359 		return nvec;
360 	}
361 
362 	tune_pcie_caps(dd);
363 
364 	return nvec;
365 }
366 
367 /* restore command and BARs after a reset has wiped them out */
368 int restore_pci_variables(struct hfi1_devdata *dd)
369 {
370 	int ret = 0;
371 
372 	ret = pci_write_config_word(dd->pcidev, PCI_COMMAND, dd->pci_command);
373 	if (ret)
374 		goto error;
375 
376 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
377 				     dd->pcibar0);
378 	if (ret)
379 		goto error;
380 
381 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
382 				     dd->pcibar1);
383 	if (ret)
384 		goto error;
385 
386 	ret = pci_write_config_dword(dd->pcidev, PCI_ROM_ADDRESS, dd->pci_rom);
387 	if (ret)
388 		goto error;
389 
390 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL,
391 					 dd->pcie_devctl);
392 	if (ret)
393 		goto error;
394 
395 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL,
396 					 dd->pcie_lnkctl);
397 	if (ret)
398 		goto error;
399 
400 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL2,
401 					 dd->pcie_devctl2);
402 	if (ret)
403 		goto error;
404 
405 	ret = pci_write_config_dword(dd->pcidev, PCI_CFG_MSIX0, dd->pci_msix0);
406 	if (ret)
407 		goto error;
408 
409 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
410 		ret = pci_write_config_dword(dd->pcidev, PCIE_CFG_TPH2,
411 					     dd->pci_tph2);
412 		if (ret)
413 			goto error;
414 	}
415 	return 0;
416 
417 error:
418 	dd_dev_err(dd, "Unable to write to PCI config\n");
419 	return ret;
420 }
421 
422 /* Save BARs and command to rewrite after device reset */
423 int save_pci_variables(struct hfi1_devdata *dd)
424 {
425 	int ret = 0;
426 
427 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
428 				    &dd->pcibar0);
429 	if (ret)
430 		goto error;
431 
432 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
433 				    &dd->pcibar1);
434 	if (ret)
435 		goto error;
436 
437 	ret = pci_read_config_dword(dd->pcidev, PCI_ROM_ADDRESS, &dd->pci_rom);
438 	if (ret)
439 		goto error;
440 
441 	ret = pci_read_config_word(dd->pcidev, PCI_COMMAND, &dd->pci_command);
442 	if (ret)
443 		goto error;
444 
445 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL,
446 					&dd->pcie_devctl);
447 	if (ret)
448 		goto error;
449 
450 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL,
451 					&dd->pcie_lnkctl);
452 	if (ret)
453 		goto error;
454 
455 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL2,
456 					&dd->pcie_devctl2);
457 	if (ret)
458 		goto error;
459 
460 	ret = pci_read_config_dword(dd->pcidev, PCI_CFG_MSIX0, &dd->pci_msix0);
461 	if (ret)
462 		goto error;
463 
464 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
465 		ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_TPH2,
466 					    &dd->pci_tph2);
467 		if (ret)
468 			goto error;
469 	}
470 	return 0;
471 
472 error:
473 	dd_dev_err(dd, "Unable to read from PCI config\n");
474 	return ret;
475 }
476 
477 /*
478  * BIOS may not set PCIe bus-utilization parameters for best performance.
479  * Check and optionally adjust them to maximize our throughput.
480  */
481 static int hfi1_pcie_caps;
482 module_param_named(pcie_caps, hfi1_pcie_caps, int, S_IRUGO);
483 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
484 
485 uint aspm_mode = ASPM_MODE_DISABLED;
486 module_param_named(aspm, aspm_mode, uint, S_IRUGO);
487 MODULE_PARM_DESC(aspm, "PCIe ASPM: 0: disable, 1: enable, 2: dynamic");
488 
489 static void tune_pcie_caps(struct hfi1_devdata *dd)
490 {
491 	struct pci_dev *parent;
492 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
493 	u16 rc_mrrs, ep_mrrs, max_mrrs, ectl;
494 	int ret;
495 
496 	/*
497 	 * Turn on extended tags in DevCtl in case the BIOS has turned it off
498 	 * to improve WFR SDMA bandwidth
499 	 */
500 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &ectl);
501 	if ((!ret) && !(ectl & PCI_EXP_DEVCTL_EXT_TAG)) {
502 		dd_dev_info(dd, "Enabling PCIe extended tags\n");
503 		ectl |= PCI_EXP_DEVCTL_EXT_TAG;
504 		ret = pcie_capability_write_word(dd->pcidev,
505 						 PCI_EXP_DEVCTL, ectl);
506 		if (ret)
507 			dd_dev_info(dd, "Unable to write to PCI config\n");
508 	}
509 	/* Find out supported and configured values for parent (root) */
510 	parent = dd->pcidev->bus->self;
511 	/*
512 	 * The driver cannot perform the tuning if it does not have
513 	 * access to the upstream component.
514 	 */
515 	if (!parent) {
516 		dd_dev_info(dd, "Parent not found\n");
517 		return;
518 	}
519 	if (!pci_is_root_bus(parent->bus)) {
520 		dd_dev_info(dd, "Parent not root\n");
521 		return;
522 	}
523 	if (!pci_is_pcie(parent)) {
524 		dd_dev_info(dd, "Parent is not PCI Express capable\n");
525 		return;
526 	}
527 	if (!pci_is_pcie(dd->pcidev)) {
528 		dd_dev_info(dd, "PCI device is not PCI Express capable\n");
529 		return;
530 	}
531 	rc_mpss = parent->pcie_mpss;
532 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
533 	/* Find out supported and configured values for endpoint (us) */
534 	ep_mpss = dd->pcidev->pcie_mpss;
535 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
536 
537 	/* Find max payload supported by root, endpoint */
538 	if (rc_mpss > ep_mpss)
539 		rc_mpss = ep_mpss;
540 
541 	/* If Supported greater than limit in module param, limit it */
542 	if (rc_mpss > (hfi1_pcie_caps & 7))
543 		rc_mpss = hfi1_pcie_caps & 7;
544 	/* If less than (allowed, supported), bump root payload */
545 	if (rc_mpss > rc_mps) {
546 		rc_mps = rc_mpss;
547 		pcie_set_mps(parent, 128 << rc_mps);
548 	}
549 	/* If less than (allowed, supported), bump endpoint payload */
550 	if (rc_mpss > ep_mps) {
551 		ep_mps = rc_mpss;
552 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
553 	}
554 
555 	/*
556 	 * Now the Read Request size.
557 	 * No field for max supported, but PCIe spec limits it to 4096,
558 	 * which is code '5' (log2(4096) - 7)
559 	 */
560 	max_mrrs = 5;
561 	if (max_mrrs > ((hfi1_pcie_caps >> 4) & 7))
562 		max_mrrs = (hfi1_pcie_caps >> 4) & 7;
563 
564 	max_mrrs = 128 << max_mrrs;
565 	rc_mrrs = pcie_get_readrq(parent);
566 	ep_mrrs = pcie_get_readrq(dd->pcidev);
567 
568 	if (max_mrrs > rc_mrrs) {
569 		rc_mrrs = max_mrrs;
570 		pcie_set_readrq(parent, rc_mrrs);
571 	}
572 	if (max_mrrs > ep_mrrs) {
573 		ep_mrrs = max_mrrs;
574 		pcie_set_readrq(dd->pcidev, ep_mrrs);
575 	}
576 }
577 
578 /* End of PCIe capability tuning */
579 
580 /*
581  * From here through hfi1_pci_err_handler definition is invoked via
582  * PCI error infrastructure, registered via pci
583  */
584 static pci_ers_result_t
585 pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
586 {
587 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
588 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
589 
590 	switch (state) {
591 	case pci_channel_io_normal:
592 		dd_dev_info(dd, "State Normal, ignoring\n");
593 		break;
594 
595 	case pci_channel_io_frozen:
596 		dd_dev_info(dd, "State Frozen, requesting reset\n");
597 		pci_disable_device(pdev);
598 		ret = PCI_ERS_RESULT_NEED_RESET;
599 		break;
600 
601 	case pci_channel_io_perm_failure:
602 		if (dd) {
603 			dd_dev_info(dd, "State Permanent Failure, disabling\n");
604 			/* no more register accesses! */
605 			dd->flags &= ~HFI1_PRESENT;
606 			hfi1_disable_after_error(dd);
607 		}
608 		 /* else early, or other problem */
609 		ret =  PCI_ERS_RESULT_DISCONNECT;
610 		break;
611 
612 	default: /* shouldn't happen */
613 		dd_dev_info(dd, "HFI1 PCI errors detected (state %d)\n",
614 			    state);
615 		break;
616 	}
617 	return ret;
618 }
619 
620 static pci_ers_result_t
621 pci_mmio_enabled(struct pci_dev *pdev)
622 {
623 	u64 words = 0U;
624 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
625 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
626 
627 	if (dd && dd->pport) {
628 		words = read_port_cntr(dd->pport, C_RX_WORDS, CNTR_INVALID_VL);
629 		if (words == ~0ULL)
630 			ret = PCI_ERS_RESULT_NEED_RESET;
631 		dd_dev_info(dd,
632 			    "HFI1 mmio_enabled function called, read wordscntr %llx, returning %d\n",
633 			    words, ret);
634 	}
635 	return  ret;
636 }
637 
638 static pci_ers_result_t
639 pci_slot_reset(struct pci_dev *pdev)
640 {
641 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
642 
643 	dd_dev_info(dd, "HFI1 slot_reset function called, ignored\n");
644 	return PCI_ERS_RESULT_CAN_RECOVER;
645 }
646 
647 static void
648 pci_resume(struct pci_dev *pdev)
649 {
650 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
651 
652 	dd_dev_info(dd, "HFI1 resume function called\n");
653 	pci_cleanup_aer_uncorrect_error_status(pdev);
654 	/*
655 	 * Running jobs will fail, since it's asynchronous
656 	 * unlike sysfs-requested reset.   Better than
657 	 * doing nothing.
658 	 */
659 	hfi1_init(dd, 1); /* same as re-init after reset */
660 }
661 
662 const struct pci_error_handlers hfi1_pci_err_handler = {
663 	.error_detected = pci_error_detected,
664 	.mmio_enabled = pci_mmio_enabled,
665 	.slot_reset = pci_slot_reset,
666 	.resume = pci_resume,
667 };
668 
669 /*============================================================================*/
670 /* PCIe Gen3 support */
671 
672 /*
673  * This code is separated out because it is expected to be removed in the
674  * final shipping product.  If not, then it will be revisited and items
675  * will be moved to more standard locations.
676  */
677 
678 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_STS field values */
679 #define DL_STATUS_HFI0 0x1	/* hfi0 firmware download complete */
680 #define DL_STATUS_HFI1 0x2	/* hfi1 firmware download complete */
681 #define DL_STATUS_BOTH 0x3	/* hfi0 and hfi1 firmware download complete */
682 
683 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_ERR field values */
684 #define DL_ERR_NONE		0x0	/* no error */
685 #define DL_ERR_SWAP_PARITY	0x1	/* parity error in SerDes interrupt */
686 					/*   or response data */
687 #define DL_ERR_DISABLED	0x2	/* hfi disabled */
688 #define DL_ERR_SECURITY	0x3	/* security check failed */
689 #define DL_ERR_SBUS		0x4	/* SBus status error */
690 #define DL_ERR_XFR_PARITY	0x5	/* parity error during ROM transfer*/
691 
692 /* gasket block secondary bus reset delay */
693 #define SBR_DELAY_US 200000	/* 200ms */
694 
695 static uint pcie_target = 3;
696 module_param(pcie_target, uint, S_IRUGO);
697 MODULE_PARM_DESC(pcie_target, "PCIe target speed (0 skip, 1-3 Gen1-3)");
698 
699 static uint pcie_force;
700 module_param(pcie_force, uint, S_IRUGO);
701 MODULE_PARM_DESC(pcie_force, "Force driver to do a PCIe firmware download even if already at target speed");
702 
703 static uint pcie_retry = 5;
704 module_param(pcie_retry, uint, S_IRUGO);
705 MODULE_PARM_DESC(pcie_retry, "Driver will try this many times to reach requested speed");
706 
707 #define UNSET_PSET 255
708 #define DEFAULT_DISCRETE_PSET 2	/* discrete HFI */
709 #define DEFAULT_MCP_PSET 6	/* MCP HFI */
710 static uint pcie_pset = UNSET_PSET;
711 module_param(pcie_pset, uint, S_IRUGO);
712 MODULE_PARM_DESC(pcie_pset, "PCIe Eq Pset value to use, range is 0-10");
713 
714 static uint pcie_ctle = 3; /* discrete on, integrated on */
715 module_param(pcie_ctle, uint, S_IRUGO);
716 MODULE_PARM_DESC(pcie_ctle, "PCIe static CTLE mode, bit 0 - discrete on/off, bit 1 - integrated on/off");
717 
718 /* equalization columns */
719 #define PREC 0
720 #define ATTN 1
721 #define POST 2
722 
723 /* discrete silicon preliminary equalization values */
724 static const u8 discrete_preliminary_eq[11][3] = {
725 	/* prec   attn   post */
726 	{  0x00,  0x00,  0x12 },	/* p0 */
727 	{  0x00,  0x00,  0x0c },	/* p1 */
728 	{  0x00,  0x00,  0x0f },	/* p2 */
729 	{  0x00,  0x00,  0x09 },	/* p3 */
730 	{  0x00,  0x00,  0x00 },	/* p4 */
731 	{  0x06,  0x00,  0x00 },	/* p5 */
732 	{  0x09,  0x00,  0x00 },	/* p6 */
733 	{  0x06,  0x00,  0x0f },	/* p7 */
734 	{  0x09,  0x00,  0x09 },	/* p8 */
735 	{  0x0c,  0x00,  0x00 },	/* p9 */
736 	{  0x00,  0x00,  0x18 },	/* p10 */
737 };
738 
739 /* integrated silicon preliminary equalization values */
740 static const u8 integrated_preliminary_eq[11][3] = {
741 	/* prec   attn   post */
742 	{  0x00,  0x1e,  0x07 },	/* p0 */
743 	{  0x00,  0x1e,  0x05 },	/* p1 */
744 	{  0x00,  0x1e,  0x06 },	/* p2 */
745 	{  0x00,  0x1e,  0x04 },	/* p3 */
746 	{  0x00,  0x1e,  0x00 },	/* p4 */
747 	{  0x03,  0x1e,  0x00 },	/* p5 */
748 	{  0x04,  0x1e,  0x00 },	/* p6 */
749 	{  0x03,  0x1e,  0x06 },	/* p7 */
750 	{  0x03,  0x1e,  0x04 },	/* p8 */
751 	{  0x05,  0x1e,  0x00 },	/* p9 */
752 	{  0x00,  0x1e,  0x0a },	/* p10 */
753 };
754 
755 static const u8 discrete_ctle_tunings[11][4] = {
756 	/* DC     LF     HF     BW */
757 	{  0x48,  0x0b,  0x04,  0x04 },	/* p0 */
758 	{  0x60,  0x05,  0x0f,  0x0a },	/* p1 */
759 	{  0x50,  0x09,  0x06,  0x06 },	/* p2 */
760 	{  0x68,  0x05,  0x0f,  0x0a },	/* p3 */
761 	{  0x80,  0x05,  0x0f,  0x0a },	/* p4 */
762 	{  0x70,  0x05,  0x0f,  0x0a },	/* p5 */
763 	{  0x68,  0x05,  0x0f,  0x0a },	/* p6 */
764 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
765 	{  0x48,  0x09,  0x06,  0x06 },	/* p8 */
766 	{  0x60,  0x05,  0x0f,  0x0a },	/* p9 */
767 	{  0x38,  0x0f,  0x00,  0x00 },	/* p10 */
768 };
769 
770 static const u8 integrated_ctle_tunings[11][4] = {
771 	/* DC     LF     HF     BW */
772 	{  0x38,  0x0f,  0x00,  0x00 },	/* p0 */
773 	{  0x38,  0x0f,  0x00,  0x00 },	/* p1 */
774 	{  0x38,  0x0f,  0x00,  0x00 },	/* p2 */
775 	{  0x38,  0x0f,  0x00,  0x00 },	/* p3 */
776 	{  0x58,  0x0a,  0x05,  0x05 },	/* p4 */
777 	{  0x48,  0x0a,  0x05,  0x05 },	/* p5 */
778 	{  0x40,  0x0a,  0x05,  0x05 },	/* p6 */
779 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
780 	{  0x38,  0x0f,  0x00,  0x00 },	/* p8 */
781 	{  0x38,  0x09,  0x06,  0x06 },	/* p9 */
782 	{  0x38,  0x0e,  0x01,  0x01 },	/* p10 */
783 };
784 
785 /* helper to format the value to write to hardware */
786 #define eq_value(pre, curr, post) \
787 	((((u32)(pre)) << \
788 			PCIE_CFG_REG_PL102_GEN3_EQ_PRE_CURSOR_PSET_SHIFT) \
789 	| (((u32)(curr)) << PCIE_CFG_REG_PL102_GEN3_EQ_CURSOR_PSET_SHIFT) \
790 	| (((u32)(post)) << \
791 		PCIE_CFG_REG_PL102_GEN3_EQ_POST_CURSOR_PSET_SHIFT))
792 
793 /*
794  * Load the given EQ preset table into the PCIe hardware.
795  */
796 static int load_eq_table(struct hfi1_devdata *dd, const u8 eq[11][3], u8 fs,
797 			 u8 div)
798 {
799 	struct pci_dev *pdev = dd->pcidev;
800 	u32 hit_error = 0;
801 	u32 violation;
802 	u32 i;
803 	u8 c_minus1, c0, c_plus1;
804 	int ret;
805 
806 	for (i = 0; i < 11; i++) {
807 		/* set index */
808 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL103, i);
809 		/* write the value */
810 		c_minus1 = eq[i][PREC] / div;
811 		c0 = fs - (eq[i][PREC] / div) - (eq[i][POST] / div);
812 		c_plus1 = eq[i][POST] / div;
813 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL102,
814 				       eq_value(c_minus1, c0, c_plus1));
815 		/* check if these coefficients violate EQ rules */
816 		ret = pci_read_config_dword(dd->pcidev,
817 					    PCIE_CFG_REG_PL105, &violation);
818 		if (ret) {
819 			dd_dev_err(dd, "Unable to read from PCI config\n");
820 			hit_error = 1;
821 			break;
822 		}
823 
824 		if (violation
825 		    & PCIE_CFG_REG_PL105_GEN3_EQ_VIOLATE_COEF_RULES_SMASK){
826 			if (hit_error == 0) {
827 				dd_dev_err(dd,
828 					   "Gen3 EQ Table Coefficient rule violations\n");
829 				dd_dev_err(dd, "         prec   attn   post\n");
830 			}
831 			dd_dev_err(dd, "   p%02d:   %02x     %02x     %02x\n",
832 				   i, (u32)eq[i][0], (u32)eq[i][1],
833 				   (u32)eq[i][2]);
834 			dd_dev_err(dd, "            %02x     %02x     %02x\n",
835 				   (u32)c_minus1, (u32)c0, (u32)c_plus1);
836 			hit_error = 1;
837 		}
838 	}
839 	if (hit_error)
840 		return -EINVAL;
841 	return 0;
842 }
843 
844 /*
845  * Steps to be done after the PCIe firmware is downloaded and
846  * before the SBR for the Pcie Gen3.
847  * The SBus resource is already being held.
848  */
849 static void pcie_post_steps(struct hfi1_devdata *dd)
850 {
851 	int i;
852 
853 	set_sbus_fast_mode(dd);
854 	/*
855 	 * Write to the PCIe PCSes to set the G3_LOCKED_NEXT bits to 1.
856 	 * This avoids a spurious framing error that can otherwise be
857 	 * generated by the MAC layer.
858 	 *
859 	 * Use individual addresses since no broadcast is set up.
860 	 */
861 	for (i = 0; i < NUM_PCIE_SERDES; i++) {
862 		sbus_request(dd, pcie_pcs_addrs[dd->hfi1_id][i],
863 			     0x03, WRITE_SBUS_RECEIVER, 0x00022132);
864 	}
865 
866 	clear_sbus_fast_mode(dd);
867 }
868 
869 /*
870  * Trigger a secondary bus reset (SBR) on ourselves using our parent.
871  *
872  * Based on pci_parent_bus_reset() which is not exported by the
873  * kernel core.
874  */
875 static int trigger_sbr(struct hfi1_devdata *dd)
876 {
877 	struct pci_dev *dev = dd->pcidev;
878 	struct pci_dev *pdev;
879 
880 	/* need a parent */
881 	if (!dev->bus->self) {
882 		dd_dev_err(dd, "%s: no parent device\n", __func__);
883 		return -ENOTTY;
884 	}
885 
886 	/* should not be anyone else on the bus */
887 	list_for_each_entry(pdev, &dev->bus->devices, bus_list)
888 		if (pdev != dev) {
889 			dd_dev_err(dd,
890 				   "%s: another device is on the same bus\n",
891 				   __func__);
892 			return -ENOTTY;
893 		}
894 
895 	/*
896 	 * A secondary bus reset (SBR) issues a hot reset to our device.
897 	 * The following routine does a 1s wait after the reset is dropped
898 	 * per PCI Trhfa (recovery time).  PCIe 3.0 section 6.6.1 -
899 	 * Conventional Reset, paragraph 3, line 35 also says that a 1s
900 	 * delay after a reset is required.  Per spec requirements,
901 	 * the link is either working or not after that point.
902 	 */
903 	return pci_reset_bus(dev);
904 }
905 
906 /*
907  * Write the given gasket interrupt register.
908  */
909 static void write_gasket_interrupt(struct hfi1_devdata *dd, int index,
910 				   u16 code, u16 data)
911 {
912 	write_csr(dd, ASIC_PCIE_SD_INTRPT_LIST + (index * 8),
913 		  (((u64)code << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_CODE_SHIFT) |
914 		   ((u64)data << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_DATA_SHIFT)));
915 }
916 
917 /*
918  * Tell the gasket logic how to react to the reset.
919  */
920 static void arm_gasket_logic(struct hfi1_devdata *dd)
921 {
922 	u64 reg;
923 
924 	reg = (((u64)1 << dd->hfi1_id) <<
925 	       ASIC_PCIE_SD_HOST_CMD_INTRPT_CMD_SHIFT) |
926 	      ((u64)pcie_serdes_broadcast[dd->hfi1_id] <<
927 	       ASIC_PCIE_SD_HOST_CMD_SBUS_RCVR_ADDR_SHIFT |
928 	       ASIC_PCIE_SD_HOST_CMD_SBR_MODE_SMASK |
929 	       ((u64)SBR_DELAY_US & ASIC_PCIE_SD_HOST_CMD_TIMER_MASK) <<
930 	       ASIC_PCIE_SD_HOST_CMD_TIMER_SHIFT);
931 	write_csr(dd, ASIC_PCIE_SD_HOST_CMD, reg);
932 	/* read back to push the write */
933 	read_csr(dd, ASIC_PCIE_SD_HOST_CMD);
934 }
935 
936 /*
937  * CCE_PCIE_CTRL long name helpers
938  * We redefine these shorter macros to use in the code while leaving
939  * chip_registers.h to be autogenerated from the hardware spec.
940  */
941 #define LANE_BUNDLE_MASK              CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_MASK
942 #define LANE_BUNDLE_SHIFT             CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_SHIFT
943 #define LANE_DELAY_MASK               CCE_PCIE_CTRL_PCIE_LANE_DELAY_MASK
944 #define LANE_DELAY_SHIFT              CCE_PCIE_CTRL_PCIE_LANE_DELAY_SHIFT
945 #define MARGIN_OVERWRITE_ENABLE_SHIFT CCE_PCIE_CTRL_XMT_MARGIN_OVERWRITE_ENABLE_SHIFT
946 #define MARGIN_SHIFT                  CCE_PCIE_CTRL_XMT_MARGIN_SHIFT
947 #define MARGIN_G1_G2_OVERWRITE_MASK   CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_MASK
948 #define MARGIN_G1_G2_OVERWRITE_SHIFT  CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_SHIFT
949 #define MARGIN_GEN1_GEN2_MASK         CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_MASK
950 #define MARGIN_GEN1_GEN2_SHIFT        CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_SHIFT
951 
952  /*
953   * Write xmt_margin for full-swing (WFR-B) or half-swing (WFR-C).
954   */
955 static void write_xmt_margin(struct hfi1_devdata *dd, const char *fname)
956 {
957 	u64 pcie_ctrl;
958 	u64 xmt_margin;
959 	u64 xmt_margin_oe;
960 	u64 lane_delay;
961 	u64 lane_bundle;
962 
963 	pcie_ctrl = read_csr(dd, CCE_PCIE_CTRL);
964 
965 	/*
966 	 * For Discrete, use full-swing.
967 	 *  - PCIe TX defaults to full-swing.
968 	 *    Leave this register as default.
969 	 * For Integrated, use half-swing
970 	 *  - Copy xmt_margin and xmt_margin_oe
971 	 *    from Gen1/Gen2 to Gen3.
972 	 */
973 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL1) { /* integrated */
974 		/* extract initial fields */
975 		xmt_margin = (pcie_ctrl >> MARGIN_GEN1_GEN2_SHIFT)
976 			      & MARGIN_GEN1_GEN2_MASK;
977 		xmt_margin_oe = (pcie_ctrl >> MARGIN_G1_G2_OVERWRITE_SHIFT)
978 				 & MARGIN_G1_G2_OVERWRITE_MASK;
979 		lane_delay = (pcie_ctrl >> LANE_DELAY_SHIFT) & LANE_DELAY_MASK;
980 		lane_bundle = (pcie_ctrl >> LANE_BUNDLE_SHIFT)
981 			       & LANE_BUNDLE_MASK;
982 
983 		/*
984 		 * For A0, EFUSE values are not set.  Override with the
985 		 * correct values.
986 		 */
987 		if (is_ax(dd)) {
988 			/*
989 			 * xmt_margin and OverwiteEnabel should be the
990 			 * same for Gen1/Gen2 and Gen3
991 			 */
992 			xmt_margin = 0x5;
993 			xmt_margin_oe = 0x1;
994 			lane_delay = 0xF; /* Delay 240ns. */
995 			lane_bundle = 0x0; /* Set to 1 lane. */
996 		}
997 
998 		/* overwrite existing values */
999 		pcie_ctrl = (xmt_margin << MARGIN_GEN1_GEN2_SHIFT)
1000 			| (xmt_margin_oe << MARGIN_G1_G2_OVERWRITE_SHIFT)
1001 			| (xmt_margin << MARGIN_SHIFT)
1002 			| (xmt_margin_oe << MARGIN_OVERWRITE_ENABLE_SHIFT)
1003 			| (lane_delay << LANE_DELAY_SHIFT)
1004 			| (lane_bundle << LANE_BUNDLE_SHIFT);
1005 
1006 		write_csr(dd, CCE_PCIE_CTRL, pcie_ctrl);
1007 	}
1008 
1009 	dd_dev_dbg(dd, "%s: program XMT margin, CcePcieCtrl 0x%llx\n",
1010 		   fname, pcie_ctrl);
1011 }
1012 
1013 /*
1014  * Do all the steps needed to transition the PCIe link to Gen3 speed.
1015  */
1016 int do_pcie_gen3_transition(struct hfi1_devdata *dd)
1017 {
1018 	struct pci_dev *parent = dd->pcidev->bus->self;
1019 	u64 fw_ctrl;
1020 	u64 reg, therm;
1021 	u32 reg32, fs, lf;
1022 	u32 status, err;
1023 	int ret;
1024 	int do_retry, retry_count = 0;
1025 	int intnum = 0;
1026 	uint default_pset;
1027 	uint pset = pcie_pset;
1028 	u16 target_vector, target_speed;
1029 	u16 lnkctl2, vendor;
1030 	u8 div;
1031 	const u8 (*eq)[3];
1032 	const u8 (*ctle_tunings)[4];
1033 	uint static_ctle_mode;
1034 	int return_error = 0;
1035 
1036 	/* PCIe Gen3 is for the ASIC only */
1037 	if (dd->icode != ICODE_RTL_SILICON)
1038 		return 0;
1039 
1040 	if (pcie_target == 1) {			/* target Gen1 */
1041 		target_vector = PCI_EXP_LNKCTL2_TLS_2_5GT;
1042 		target_speed = 2500;
1043 	} else if (pcie_target == 2) {		/* target Gen2 */
1044 		target_vector = PCI_EXP_LNKCTL2_TLS_5_0GT;
1045 		target_speed = 5000;
1046 	} else if (pcie_target == 3) {		/* target Gen3 */
1047 		target_vector = PCI_EXP_LNKCTL2_TLS_8_0GT;
1048 		target_speed = 8000;
1049 	} else {
1050 		/* off or invalid target - skip */
1051 		dd_dev_info(dd, "%s: Skipping PCIe transition\n", __func__);
1052 		return 0;
1053 	}
1054 
1055 	/* if already at target speed, done (unless forced) */
1056 	if (dd->lbus_speed == target_speed) {
1057 		dd_dev_info(dd, "%s: PCIe already at gen%d, %s\n", __func__,
1058 			    pcie_target,
1059 			    pcie_force ? "re-doing anyway" : "skipping");
1060 		if (!pcie_force)
1061 			return 0;
1062 	}
1063 
1064 	/*
1065 	 * The driver cannot do the transition if it has no access to the
1066 	 * upstream component
1067 	 */
1068 	if (!parent) {
1069 		dd_dev_info(dd, "%s: No upstream, Can't do gen3 transition\n",
1070 			    __func__);
1071 		return 0;
1072 	}
1073 
1074 	/*
1075 	 * Do the Gen3 transition.  Steps are those of the PCIe Gen3
1076 	 * recipe.
1077 	 */
1078 
1079 	/* step 1: pcie link working in gen1/gen2 */
1080 
1081 	/* step 2: if either side is not capable of Gen3, done */
1082 	if (pcie_target == 3 && !dd->link_gen3_capable) {
1083 		dd_dev_err(dd, "The PCIe link is not Gen3 capable\n");
1084 		ret = -ENOSYS;
1085 		goto done_no_mutex;
1086 	}
1087 
1088 	/* hold the SBus resource across the firmware download and SBR */
1089 	ret = acquire_chip_resource(dd, CR_SBUS, SBUS_TIMEOUT);
1090 	if (ret) {
1091 		dd_dev_err(dd, "%s: unable to acquire SBus resource\n",
1092 			   __func__);
1093 		return ret;
1094 	}
1095 
1096 	/* make sure thermal polling is not causing interrupts */
1097 	therm = read_csr(dd, ASIC_CFG_THERM_POLL_EN);
1098 	if (therm) {
1099 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x0);
1100 		msleep(100);
1101 		dd_dev_info(dd, "%s: Disabled therm polling\n",
1102 			    __func__);
1103 	}
1104 
1105 retry:
1106 	/* the SBus download will reset the spico for thermal */
1107 
1108 	/* step 3: download SBus Master firmware */
1109 	/* step 4: download PCIe Gen3 SerDes firmware */
1110 	dd_dev_info(dd, "%s: downloading firmware\n", __func__);
1111 	ret = load_pcie_firmware(dd);
1112 	if (ret) {
1113 		/* do not proceed if the firmware cannot be downloaded */
1114 		return_error = 1;
1115 		goto done;
1116 	}
1117 
1118 	/* step 5: set up device parameter settings */
1119 	dd_dev_info(dd, "%s: setting PCIe registers\n", __func__);
1120 
1121 	/*
1122 	 * PcieCfgSpcie1 - Link Control 3
1123 	 * Leave at reset value.  No need to set PerfEq - link equalization
1124 	 * will be performed automatically after the SBR when the target
1125 	 * speed is 8GT/s.
1126 	 */
1127 
1128 	/* clear all 16 per-lane error bits (PCIe: Lane Error Status) */
1129 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, 0xffff);
1130 
1131 	/* step 5a: Set Synopsys Port Logic registers */
1132 
1133 	/*
1134 	 * PcieCfgRegPl2 - Port Force Link
1135 	 *
1136 	 * Set the low power field to 0x10 to avoid unnecessary power
1137 	 * management messages.  All other fields are zero.
1138 	 */
1139 	reg32 = 0x10ul << PCIE_CFG_REG_PL2_LOW_PWR_ENT_CNT_SHIFT;
1140 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL2, reg32);
1141 
1142 	/*
1143 	 * PcieCfgRegPl100 - Gen3 Control
1144 	 *
1145 	 * turn off PcieCfgRegPl100.Gen3ZRxDcNonCompl
1146 	 * turn on PcieCfgRegPl100.EqEieosCnt
1147 	 * Everything else zero.
1148 	 */
1149 	reg32 = PCIE_CFG_REG_PL100_EQ_EIEOS_CNT_SMASK;
1150 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL100, reg32);
1151 
1152 	/*
1153 	 * PcieCfgRegPl101 - Gen3 EQ FS and LF
1154 	 * PcieCfgRegPl102 - Gen3 EQ Presets to Coefficients Mapping
1155 	 * PcieCfgRegPl103 - Gen3 EQ Preset Index
1156 	 * PcieCfgRegPl105 - Gen3 EQ Status
1157 	 *
1158 	 * Give initial EQ settings.
1159 	 */
1160 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL0) { /* discrete */
1161 		/* 1000mV, FS=24, LF = 8 */
1162 		fs = 24;
1163 		lf = 8;
1164 		div = 3;
1165 		eq = discrete_preliminary_eq;
1166 		default_pset = DEFAULT_DISCRETE_PSET;
1167 		ctle_tunings = discrete_ctle_tunings;
1168 		/* bit 0 - discrete on/off */
1169 		static_ctle_mode = pcie_ctle & 0x1;
1170 	} else {
1171 		/* 400mV, FS=29, LF = 9 */
1172 		fs = 29;
1173 		lf = 9;
1174 		div = 1;
1175 		eq = integrated_preliminary_eq;
1176 		default_pset = DEFAULT_MCP_PSET;
1177 		ctle_tunings = integrated_ctle_tunings;
1178 		/* bit 1 - integrated on/off */
1179 		static_ctle_mode = (pcie_ctle >> 1) & 0x1;
1180 	}
1181 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL101,
1182 			       (fs <<
1183 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_FS_SHIFT) |
1184 			       (lf <<
1185 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_LF_SHIFT));
1186 	ret = load_eq_table(dd, eq, fs, div);
1187 	if (ret)
1188 		goto done;
1189 
1190 	/*
1191 	 * PcieCfgRegPl106 - Gen3 EQ Control
1192 	 *
1193 	 * Set Gen3EqPsetReqVec, leave other fields 0.
1194 	 */
1195 	if (pset == UNSET_PSET)
1196 		pset = default_pset;
1197 	if (pset > 10) {	/* valid range is 0-10, inclusive */
1198 		dd_dev_err(dd, "%s: Invalid Eq Pset %u, setting to %d\n",
1199 			   __func__, pset, default_pset);
1200 		pset = default_pset;
1201 	}
1202 	dd_dev_info(dd, "%s: using EQ Pset %u\n", __func__, pset);
1203 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL106,
1204 			       ((1 << pset) <<
1205 			PCIE_CFG_REG_PL106_GEN3_EQ_PSET_REQ_VEC_SHIFT) |
1206 			PCIE_CFG_REG_PL106_GEN3_EQ_EVAL2MS_DISABLE_SMASK |
1207 			PCIE_CFG_REG_PL106_GEN3_EQ_PHASE23_EXIT_MODE_SMASK);
1208 
1209 	/*
1210 	 * step 5b: Do post firmware download steps via SBus
1211 	 */
1212 	dd_dev_info(dd, "%s: doing pcie post steps\n", __func__);
1213 	pcie_post_steps(dd);
1214 
1215 	/*
1216 	 * step 5c: Program gasket interrupts
1217 	 */
1218 	/* set the Rx Bit Rate to REFCLK ratio */
1219 	write_gasket_interrupt(dd, intnum++, 0x0006, 0x0050);
1220 	/* disable pCal for PCIe Gen3 RX equalization */
1221 	/* select adaptive or static CTLE */
1222 	write_gasket_interrupt(dd, intnum++, 0x0026,
1223 			       0x5b01 | (static_ctle_mode << 3));
1224 	/*
1225 	 * Enable iCal for PCIe Gen3 RX equalization, and set which
1226 	 * evaluation of RX_EQ_EVAL will launch the iCal procedure.
1227 	 */
1228 	write_gasket_interrupt(dd, intnum++, 0x0026, 0x5202);
1229 
1230 	if (static_ctle_mode) {
1231 		/* apply static CTLE tunings */
1232 		u8 pcie_dc, pcie_lf, pcie_hf, pcie_bw;
1233 
1234 		pcie_dc = ctle_tunings[pset][0];
1235 		pcie_lf = ctle_tunings[pset][1];
1236 		pcie_hf = ctle_tunings[pset][2];
1237 		pcie_bw = ctle_tunings[pset][3];
1238 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0200 | pcie_dc);
1239 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0100 | pcie_lf);
1240 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0000 | pcie_hf);
1241 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x5500 | pcie_bw);
1242 	}
1243 
1244 	/* terminate list */
1245 	write_gasket_interrupt(dd, intnum++, 0x0000, 0x0000);
1246 
1247 	/*
1248 	 * step 5d: program XMT margin
1249 	 */
1250 	write_xmt_margin(dd, __func__);
1251 
1252 	/*
1253 	 * step 5e: disable active state power management (ASPM). It
1254 	 * will be enabled if required later
1255 	 */
1256 	dd_dev_info(dd, "%s: clearing ASPM\n", __func__);
1257 	aspm_hw_disable_l1(dd);
1258 
1259 	/*
1260 	 * step 5f: clear DirectSpeedChange
1261 	 * PcieCfgRegPl67.DirectSpeedChange must be zero to prevent the
1262 	 * change in the speed target from starting before we are ready.
1263 	 * This field defaults to 0 and we are not changing it, so nothing
1264 	 * needs to be done.
1265 	 */
1266 
1267 	/* step 5g: Set target link speed */
1268 	/*
1269 	 * Set target link speed to be target on both device and parent.
1270 	 * On setting the parent: Some system BIOSs "helpfully" set the
1271 	 * parent target speed to Gen2 to match the ASIC's initial speed.
1272 	 * We can set the target Gen3 because we have already checked
1273 	 * that it is Gen3 capable earlier.
1274 	 */
1275 	dd_dev_info(dd, "%s: setting parent target link speed\n", __func__);
1276 	ret = pcie_capability_read_word(parent, PCI_EXP_LNKCTL2, &lnkctl2);
1277 	if (ret) {
1278 		dd_dev_err(dd, "Unable to read from PCI config\n");
1279 		return_error = 1;
1280 		goto done;
1281 	}
1282 
1283 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1284 		    (u32)lnkctl2);
1285 	/* only write to parent if target is not as high as ours */
1286 	if ((lnkctl2 & PCI_EXP_LNKCTL2_TLS) < target_vector) {
1287 		lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1288 		lnkctl2 |= target_vector;
1289 		dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1290 			    (u32)lnkctl2);
1291 		ret = pcie_capability_write_word(parent,
1292 						 PCI_EXP_LNKCTL2, lnkctl2);
1293 		if (ret) {
1294 			dd_dev_err(dd, "Unable to write to PCI config\n");
1295 			return_error = 1;
1296 			goto done;
1297 		}
1298 	} else {
1299 		dd_dev_info(dd, "%s: ..target speed is OK\n", __func__);
1300 	}
1301 
1302 	dd_dev_info(dd, "%s: setting target link speed\n", __func__);
1303 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL2, &lnkctl2);
1304 	if (ret) {
1305 		dd_dev_err(dd, "Unable to read from PCI config\n");
1306 		return_error = 1;
1307 		goto done;
1308 	}
1309 
1310 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1311 		    (u32)lnkctl2);
1312 	lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1313 	lnkctl2 |= target_vector;
1314 	dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1315 		    (u32)lnkctl2);
1316 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL2, lnkctl2);
1317 	if (ret) {
1318 		dd_dev_err(dd, "Unable to write to PCI config\n");
1319 		return_error = 1;
1320 		goto done;
1321 	}
1322 
1323 	/* step 5h: arm gasket logic */
1324 	/* hold DC in reset across the SBR */
1325 	write_csr(dd, CCE_DC_CTRL, CCE_DC_CTRL_DC_RESET_SMASK);
1326 	(void)read_csr(dd, CCE_DC_CTRL); /* DC reset hold */
1327 	/* save firmware control across the SBR */
1328 	fw_ctrl = read_csr(dd, MISC_CFG_FW_CTRL);
1329 
1330 	dd_dev_info(dd, "%s: arming gasket logic\n", __func__);
1331 	arm_gasket_logic(dd);
1332 
1333 	/*
1334 	 * step 6: quiesce PCIe link
1335 	 * The chip has already been reset, so there will be no traffic
1336 	 * from the chip.  Linux has no easy way to enforce that it will
1337 	 * not try to access the device, so we just need to hope it doesn't
1338 	 * do it while we are doing the reset.
1339 	 */
1340 
1341 	/*
1342 	 * step 7: initiate the secondary bus reset (SBR)
1343 	 * step 8: hardware brings the links back up
1344 	 * step 9: wait for link speed transition to be complete
1345 	 */
1346 	dd_dev_info(dd, "%s: calling trigger_sbr\n", __func__);
1347 	ret = trigger_sbr(dd);
1348 	if (ret)
1349 		goto done;
1350 
1351 	/* step 10: decide what to do next */
1352 
1353 	/* check if we can read PCI space */
1354 	ret = pci_read_config_word(dd->pcidev, PCI_VENDOR_ID, &vendor);
1355 	if (ret) {
1356 		dd_dev_info(dd,
1357 			    "%s: read of VendorID failed after SBR, err %d\n",
1358 			    __func__, ret);
1359 		return_error = 1;
1360 		goto done;
1361 	}
1362 	if (vendor == 0xffff) {
1363 		dd_dev_info(dd, "%s: VendorID is all 1s after SBR\n", __func__);
1364 		return_error = 1;
1365 		ret = -EIO;
1366 		goto done;
1367 	}
1368 
1369 	/* restore PCI space registers we know were reset */
1370 	dd_dev_info(dd, "%s: calling restore_pci_variables\n", __func__);
1371 	ret = restore_pci_variables(dd);
1372 	if (ret) {
1373 		dd_dev_err(dd, "%s: Could not restore PCI variables\n",
1374 			   __func__);
1375 		return_error = 1;
1376 		goto done;
1377 	}
1378 
1379 	/* restore firmware control */
1380 	write_csr(dd, MISC_CFG_FW_CTRL, fw_ctrl);
1381 
1382 	/*
1383 	 * Check the gasket block status.
1384 	 *
1385 	 * This is the first CSR read after the SBR.  If the read returns
1386 	 * all 1s (fails), the link did not make it back.
1387 	 *
1388 	 * Once we're sure we can read and write, clear the DC reset after
1389 	 * the SBR.  Then check for any per-lane errors. Then look over
1390 	 * the status.
1391 	 */
1392 	reg = read_csr(dd, ASIC_PCIE_SD_HOST_STATUS);
1393 	dd_dev_info(dd, "%s: gasket block status: 0x%llx\n", __func__, reg);
1394 	if (reg == ~0ull) {	/* PCIe read failed/timeout */
1395 		dd_dev_err(dd, "SBR failed - unable to read from device\n");
1396 		return_error = 1;
1397 		ret = -ENOSYS;
1398 		goto done;
1399 	}
1400 
1401 	/* clear the DC reset */
1402 	write_csr(dd, CCE_DC_CTRL, 0);
1403 
1404 	/* Set the LED off */
1405 	setextled(dd, 0);
1406 
1407 	/* check for any per-lane errors */
1408 	ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, &reg32);
1409 	if (ret) {
1410 		dd_dev_err(dd, "Unable to read from PCI config\n");
1411 		return_error = 1;
1412 		goto done;
1413 	}
1414 
1415 	dd_dev_info(dd, "%s: per-lane errors: 0x%x\n", __func__, reg32);
1416 
1417 	/* extract status, look for our HFI */
1418 	status = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_SHIFT)
1419 			& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_MASK;
1420 	if ((status & (1 << dd->hfi1_id)) == 0) {
1421 		dd_dev_err(dd,
1422 			   "%s: gasket status 0x%x, expecting 0x%x\n",
1423 			   __func__, status, 1 << dd->hfi1_id);
1424 		ret = -EIO;
1425 		goto done;
1426 	}
1427 
1428 	/* extract error */
1429 	err = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_SHIFT)
1430 		& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_MASK;
1431 	if (err) {
1432 		dd_dev_err(dd, "%s: gasket error %d\n", __func__, err);
1433 		ret = -EIO;
1434 		goto done;
1435 	}
1436 
1437 	/* update our link information cache */
1438 	update_lbus_info(dd);
1439 	dd_dev_info(dd, "%s: new speed and width: %s\n", __func__,
1440 		    dd->lbus_info);
1441 
1442 	if (dd->lbus_speed != target_speed) { /* not target */
1443 		/* maybe retry */
1444 		do_retry = retry_count < pcie_retry;
1445 		dd_dev_err(dd, "PCIe link speed did not switch to Gen%d%s\n",
1446 			   pcie_target, do_retry ? ", retrying" : "");
1447 		retry_count++;
1448 		if (do_retry) {
1449 			msleep(100); /* allow time to settle */
1450 			goto retry;
1451 		}
1452 		ret = -EIO;
1453 	}
1454 
1455 done:
1456 	if (therm) {
1457 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x1);
1458 		msleep(100);
1459 		dd_dev_info(dd, "%s: Re-enable therm polling\n",
1460 			    __func__);
1461 	}
1462 	release_chip_resource(dd, CR_SBUS);
1463 done_no_mutex:
1464 	/* return no error if it is OK to be at current speed */
1465 	if (ret && !return_error) {
1466 		dd_dev_err(dd, "Proceeding at current speed PCIe speed\n");
1467 		ret = 0;
1468 	}
1469 
1470 	dd_dev_info(dd, "%s: done\n", __func__);
1471 	return ret;
1472 }
1473