xref: /linux/drivers/infiniband/hw/hfi1/pcie.c (revision 5e3992fe72748ed3892be876f09d4d990548b7af)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2019 Intel Corporation.
4  */
5 
6 #include <linux/pci.h>
7 #include <linux/io.h>
8 #include <linux/delay.h>
9 #include <linux/vmalloc.h>
10 #include <linux/module.h>
11 
12 #include "hfi.h"
13 #include "chip_registers.h"
14 #include "aspm.h"
15 
16 /*
17  * This file contains PCIe utility routines.
18  */
19 
20 /*
21  * Do all the common PCIe setup and initialization.
22  */
23 int hfi1_pcie_init(struct hfi1_devdata *dd)
24 {
25 	int ret;
26 	struct pci_dev *pdev = dd->pcidev;
27 
28 	ret = pci_enable_device(pdev);
29 	if (ret) {
30 		/*
31 		 * This can happen (in theory) iff:
32 		 * We did a chip reset, and then failed to reprogram the
33 		 * BAR, or the chip reset due to an internal error.  We then
34 		 * unloaded the driver and reloaded it.
35 		 *
36 		 * Both reset cases set the BAR back to initial state.  For
37 		 * the latter case, the AER sticky error bit at offset 0x718
38 		 * should be set, but the Linux kernel doesn't yet know
39 		 * about that, it appears.  If the original BAR was retained
40 		 * in the kernel data structures, this may be OK.
41 		 */
42 		dd_dev_err(dd, "pci enable failed: error %d\n", -ret);
43 		return ret;
44 	}
45 
46 	ret = pci_request_regions(pdev, DRIVER_NAME);
47 	if (ret) {
48 		dd_dev_err(dd, "pci_request_regions fails: err %d\n", -ret);
49 		goto bail;
50 	}
51 
52 	ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
53 	if (ret) {
54 		/*
55 		 * If the 64 bit setup fails, try 32 bit.  Some systems
56 		 * do not setup 64 bit maps on systems with 2GB or less
57 		 * memory installed.
58 		 */
59 		ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
60 		if (ret) {
61 			dd_dev_err(dd, "Unable to set DMA mask: %d\n", ret);
62 			goto bail;
63 		}
64 	}
65 
66 	pci_set_master(pdev);
67 	return 0;
68 
69 bail:
70 	hfi1_pcie_cleanup(pdev);
71 	return ret;
72 }
73 
74 /*
75  * Clean what was done in hfi1_pcie_init()
76  */
77 void hfi1_pcie_cleanup(struct pci_dev *pdev)
78 {
79 	pci_disable_device(pdev);
80 	/*
81 	 * Release regions should be called after the disable. OK to
82 	 * call if request regions has not been called or failed.
83 	 */
84 	pci_release_regions(pdev);
85 }
86 
87 /*
88  * Do remaining PCIe setup, once dd is allocated, and save away
89  * fields required to re-initialize after a chip reset, or for
90  * various other purposes
91  */
92 int hfi1_pcie_ddinit(struct hfi1_devdata *dd, struct pci_dev *pdev)
93 {
94 	unsigned long len;
95 	resource_size_t addr;
96 	int ret = 0;
97 	u32 rcv_array_count;
98 
99 	addr = pci_resource_start(pdev, 0);
100 	len = pci_resource_len(pdev, 0);
101 
102 	/*
103 	 * The TXE PIO buffers are at the tail end of the chip space.
104 	 * Cut them off and map them separately.
105 	 */
106 
107 	/* sanity check vs expectations */
108 	if (len != TXE_PIO_SEND + TXE_PIO_SIZE) {
109 		dd_dev_err(dd, "chip PIO range does not match\n");
110 		return -EINVAL;
111 	}
112 
113 	dd->kregbase1 = ioremap(addr, RCV_ARRAY);
114 	if (!dd->kregbase1) {
115 		dd_dev_err(dd, "UC mapping of kregbase1 failed\n");
116 		return -ENOMEM;
117 	}
118 	dd_dev_info(dd, "UC base1: %p for %x\n", dd->kregbase1, RCV_ARRAY);
119 
120 	/* verify that reads actually work, save revision for reset check */
121 	dd->revision = readq(dd->kregbase1 + CCE_REVISION);
122 	if (dd->revision == ~(u64)0) {
123 		dd_dev_err(dd, "Cannot read chip CSRs\n");
124 		goto nomem;
125 	}
126 
127 	rcv_array_count = readq(dd->kregbase1 + RCV_ARRAY_CNT);
128 	dd_dev_info(dd, "RcvArray count: %u\n", rcv_array_count);
129 	dd->base2_start  = RCV_ARRAY + rcv_array_count * 8;
130 
131 	dd->kregbase2 = ioremap(
132 		addr + dd->base2_start,
133 		TXE_PIO_SEND - dd->base2_start);
134 	if (!dd->kregbase2) {
135 		dd_dev_err(dd, "UC mapping of kregbase2 failed\n");
136 		goto nomem;
137 	}
138 	dd_dev_info(dd, "UC base2: %p for %x\n", dd->kregbase2,
139 		    TXE_PIO_SEND - dd->base2_start);
140 
141 	dd->piobase = ioremap_wc(addr + TXE_PIO_SEND, TXE_PIO_SIZE);
142 	if (!dd->piobase) {
143 		dd_dev_err(dd, "WC mapping of send buffers failed\n");
144 		goto nomem;
145 	}
146 	dd_dev_info(dd, "WC piobase: %p for %x\n", dd->piobase, TXE_PIO_SIZE);
147 
148 	dd->physaddr = addr;        /* used for io_remap, etc. */
149 
150 	/*
151 	 * Map the chip's RcvArray as write-combining to allow us
152 	 * to write an entire cacheline worth of entries in one shot.
153 	 */
154 	dd->rcvarray_wc = ioremap_wc(addr + RCV_ARRAY,
155 				     rcv_array_count * 8);
156 	if (!dd->rcvarray_wc) {
157 		dd_dev_err(dd, "WC mapping of receive array failed\n");
158 		goto nomem;
159 	}
160 	dd_dev_info(dd, "WC RcvArray: %p for %x\n",
161 		    dd->rcvarray_wc, rcv_array_count * 8);
162 
163 	dd->flags |= HFI1_PRESENT;	/* chip.c CSR routines now work */
164 	return 0;
165 nomem:
166 	ret = -ENOMEM;
167 	hfi1_pcie_ddcleanup(dd);
168 	return ret;
169 }
170 
171 /*
172  * Do PCIe cleanup related to dd, after chip-specific cleanup, etc.  Just prior
173  * to releasing the dd memory.
174  * Void because all of the core pcie cleanup functions are void.
175  */
176 void hfi1_pcie_ddcleanup(struct hfi1_devdata *dd)
177 {
178 	dd->flags &= ~HFI1_PRESENT;
179 	if (dd->kregbase1)
180 		iounmap(dd->kregbase1);
181 	dd->kregbase1 = NULL;
182 	if (dd->kregbase2)
183 		iounmap(dd->kregbase2);
184 	dd->kregbase2 = NULL;
185 	if (dd->rcvarray_wc)
186 		iounmap(dd->rcvarray_wc);
187 	dd->rcvarray_wc = NULL;
188 	if (dd->piobase)
189 		iounmap(dd->piobase);
190 	dd->piobase = NULL;
191 }
192 
193 /* return the PCIe link speed from the given link status */
194 static u32 extract_speed(u16 linkstat)
195 {
196 	u32 speed;
197 
198 	switch (linkstat & PCI_EXP_LNKSTA_CLS) {
199 	default: /* not defined, assume Gen1 */
200 	case PCI_EXP_LNKSTA_CLS_2_5GB:
201 		speed = 2500; /* Gen 1, 2.5GHz */
202 		break;
203 	case PCI_EXP_LNKSTA_CLS_5_0GB:
204 		speed = 5000; /* Gen 2, 5GHz */
205 		break;
206 	case PCI_EXP_LNKSTA_CLS_8_0GB:
207 		speed = 8000; /* Gen 3, 8GHz */
208 		break;
209 	}
210 	return speed;
211 }
212 
213 /* return the PCIe link speed from the given link status */
214 static u32 extract_width(u16 linkstat)
215 {
216 	return (linkstat & PCI_EXP_LNKSTA_NLW) >> PCI_EXP_LNKSTA_NLW_SHIFT;
217 }
218 
219 /* read the link status and set dd->{lbus_width,lbus_speed,lbus_info} */
220 static void update_lbus_info(struct hfi1_devdata *dd)
221 {
222 	u16 linkstat;
223 	int ret;
224 
225 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
226 	if (ret) {
227 		dd_dev_err(dd, "Unable to read from PCI config\n");
228 		return;
229 	}
230 
231 	dd->lbus_width = extract_width(linkstat);
232 	dd->lbus_speed = extract_speed(linkstat);
233 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
234 		 "PCIe,%uMHz,x%u", dd->lbus_speed, dd->lbus_width);
235 }
236 
237 /*
238  * Read in the current PCIe link width and speed.  Find if the link is
239  * Gen3 capable.
240  */
241 int pcie_speeds(struct hfi1_devdata *dd)
242 {
243 	u32 linkcap;
244 	struct pci_dev *parent = dd->pcidev->bus->self;
245 	int ret;
246 
247 	if (!pci_is_pcie(dd->pcidev)) {
248 		dd_dev_err(dd, "Can't find PCI Express capability!\n");
249 		return -EINVAL;
250 	}
251 
252 	/* find if our max speed is Gen3 and parent supports Gen3 speeds */
253 	dd->link_gen3_capable = 1;
254 
255 	ret = pcie_capability_read_dword(dd->pcidev, PCI_EXP_LNKCAP, &linkcap);
256 	if (ret) {
257 		dd_dev_err(dd, "Unable to read from PCI config\n");
258 		return pcibios_err_to_errno(ret);
259 	}
260 
261 	if ((linkcap & PCI_EXP_LNKCAP_SLS) != PCI_EXP_LNKCAP_SLS_8_0GB) {
262 		dd_dev_info(dd,
263 			    "This HFI is not Gen3 capable, max speed 0x%x, need 0x3\n",
264 			    linkcap & PCI_EXP_LNKCAP_SLS);
265 		dd->link_gen3_capable = 0;
266 	}
267 
268 	/*
269 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
270 	 */
271 	if (parent &&
272 	    (dd->pcidev->bus->max_bus_speed == PCIE_SPEED_2_5GT ||
273 	     dd->pcidev->bus->max_bus_speed == PCIE_SPEED_5_0GT)) {
274 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
275 		dd->link_gen3_capable = 0;
276 	}
277 
278 	/* obtain the link width and current speed */
279 	update_lbus_info(dd);
280 
281 	dd_dev_info(dd, "%s\n", dd->lbus_info);
282 
283 	return 0;
284 }
285 
286 /*
287  * Restore command and BARs after a reset has wiped them out
288  *
289  * Returns 0 on success, otherwise a negative error value
290  */
291 int restore_pci_variables(struct hfi1_devdata *dd)
292 {
293 	int ret;
294 
295 	ret = pci_write_config_word(dd->pcidev, PCI_COMMAND, dd->pci_command);
296 	if (ret)
297 		goto error;
298 
299 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
300 				     dd->pcibar0);
301 	if (ret)
302 		goto error;
303 
304 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
305 				     dd->pcibar1);
306 	if (ret)
307 		goto error;
308 
309 	ret = pci_write_config_dword(dd->pcidev, PCI_ROM_ADDRESS, dd->pci_rom);
310 	if (ret)
311 		goto error;
312 
313 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL,
314 					 dd->pcie_devctl);
315 	if (ret)
316 		goto error;
317 
318 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL,
319 					 dd->pcie_lnkctl);
320 	if (ret)
321 		goto error;
322 
323 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL2,
324 					 dd->pcie_devctl2);
325 	if (ret)
326 		goto error;
327 
328 	ret = pci_write_config_dword(dd->pcidev, PCI_CFG_MSIX0, dd->pci_msix0);
329 	if (ret)
330 		goto error;
331 
332 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
333 		ret = pci_write_config_dword(dd->pcidev, PCIE_CFG_TPH2,
334 					     dd->pci_tph2);
335 		if (ret)
336 			goto error;
337 	}
338 	return 0;
339 
340 error:
341 	dd_dev_err(dd, "Unable to write to PCI config\n");
342 	return pcibios_err_to_errno(ret);
343 }
344 
345 /*
346  * Save BARs and command to rewrite after device reset
347  *
348  * Returns 0 on success, otherwise a negative error value
349  */
350 int save_pci_variables(struct hfi1_devdata *dd)
351 {
352 	int ret;
353 
354 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
355 				    &dd->pcibar0);
356 	if (ret)
357 		goto error;
358 
359 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
360 				    &dd->pcibar1);
361 	if (ret)
362 		goto error;
363 
364 	ret = pci_read_config_dword(dd->pcidev, PCI_ROM_ADDRESS, &dd->pci_rom);
365 	if (ret)
366 		goto error;
367 
368 	ret = pci_read_config_word(dd->pcidev, PCI_COMMAND, &dd->pci_command);
369 	if (ret)
370 		goto error;
371 
372 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL,
373 					&dd->pcie_devctl);
374 	if (ret)
375 		goto error;
376 
377 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL,
378 					&dd->pcie_lnkctl);
379 	if (ret)
380 		goto error;
381 
382 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL2,
383 					&dd->pcie_devctl2);
384 	if (ret)
385 		goto error;
386 
387 	ret = pci_read_config_dword(dd->pcidev, PCI_CFG_MSIX0, &dd->pci_msix0);
388 	if (ret)
389 		goto error;
390 
391 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
392 		ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_TPH2,
393 					    &dd->pci_tph2);
394 		if (ret)
395 			goto error;
396 	}
397 	return 0;
398 
399 error:
400 	dd_dev_err(dd, "Unable to read from PCI config\n");
401 	return pcibios_err_to_errno(ret);
402 }
403 
404 /*
405  * BIOS may not set PCIe bus-utilization parameters for best performance.
406  * Check and optionally adjust them to maximize our throughput.
407  */
408 static int hfi1_pcie_caps;
409 module_param_named(pcie_caps, hfi1_pcie_caps, int, 0444);
410 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
411 
412 /**
413  * tune_pcie_caps() - Code to adjust PCIe capabilities.
414  * @dd: Valid device data structure
415  *
416  */
417 void tune_pcie_caps(struct hfi1_devdata *dd)
418 {
419 	struct pci_dev *parent;
420 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
421 	u16 rc_mrrs, ep_mrrs, max_mrrs, ectl;
422 	int ret;
423 
424 	/*
425 	 * Turn on extended tags in DevCtl in case the BIOS has turned it off
426 	 * to improve WFR SDMA bandwidth
427 	 */
428 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &ectl);
429 	if ((!ret) && !(ectl & PCI_EXP_DEVCTL_EXT_TAG)) {
430 		dd_dev_info(dd, "Enabling PCIe extended tags\n");
431 		ectl |= PCI_EXP_DEVCTL_EXT_TAG;
432 		ret = pcie_capability_write_word(dd->pcidev,
433 						 PCI_EXP_DEVCTL, ectl);
434 		if (ret)
435 			dd_dev_info(dd, "Unable to write to PCI config\n");
436 	}
437 	/* Find out supported and configured values for parent (root) */
438 	parent = dd->pcidev->bus->self;
439 	/*
440 	 * The driver cannot perform the tuning if it does not have
441 	 * access to the upstream component.
442 	 */
443 	if (!parent) {
444 		dd_dev_info(dd, "Parent not found\n");
445 		return;
446 	}
447 	if (!pci_is_root_bus(parent->bus)) {
448 		dd_dev_info(dd, "Parent not root\n");
449 		return;
450 	}
451 	if (!pci_is_pcie(parent)) {
452 		dd_dev_info(dd, "Parent is not PCI Express capable\n");
453 		return;
454 	}
455 	if (!pci_is_pcie(dd->pcidev)) {
456 		dd_dev_info(dd, "PCI device is not PCI Express capable\n");
457 		return;
458 	}
459 	rc_mpss = parent->pcie_mpss;
460 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
461 	/* Find out supported and configured values for endpoint (us) */
462 	ep_mpss = dd->pcidev->pcie_mpss;
463 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
464 
465 	/* Find max payload supported by root, endpoint */
466 	if (rc_mpss > ep_mpss)
467 		rc_mpss = ep_mpss;
468 
469 	/* If Supported greater than limit in module param, limit it */
470 	if (rc_mpss > (hfi1_pcie_caps & 7))
471 		rc_mpss = hfi1_pcie_caps & 7;
472 	/* If less than (allowed, supported), bump root payload */
473 	if (rc_mpss > rc_mps) {
474 		rc_mps = rc_mpss;
475 		pcie_set_mps(parent, 128 << rc_mps);
476 	}
477 	/* If less than (allowed, supported), bump endpoint payload */
478 	if (rc_mpss > ep_mps) {
479 		ep_mps = rc_mpss;
480 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
481 	}
482 
483 	/*
484 	 * Now the Read Request size.
485 	 * No field for max supported, but PCIe spec limits it to 4096,
486 	 * which is code '5' (log2(4096) - 7)
487 	 */
488 	max_mrrs = 5;
489 	if (max_mrrs > ((hfi1_pcie_caps >> 4) & 7))
490 		max_mrrs = (hfi1_pcie_caps >> 4) & 7;
491 
492 	max_mrrs = 128 << max_mrrs;
493 	rc_mrrs = pcie_get_readrq(parent);
494 	ep_mrrs = pcie_get_readrq(dd->pcidev);
495 
496 	if (max_mrrs > rc_mrrs) {
497 		rc_mrrs = max_mrrs;
498 		pcie_set_readrq(parent, rc_mrrs);
499 	}
500 	if (max_mrrs > ep_mrrs) {
501 		ep_mrrs = max_mrrs;
502 		pcie_set_readrq(dd->pcidev, ep_mrrs);
503 	}
504 }
505 
506 /* End of PCIe capability tuning */
507 
508 /*
509  * From here through hfi1_pci_err_handler definition is invoked via
510  * PCI error infrastructure, registered via pci
511  */
512 static pci_ers_result_t
513 pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
514 {
515 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
516 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
517 
518 	switch (state) {
519 	case pci_channel_io_normal:
520 		dd_dev_info(dd, "State Normal, ignoring\n");
521 		break;
522 
523 	case pci_channel_io_frozen:
524 		dd_dev_info(dd, "State Frozen, requesting reset\n");
525 		pci_disable_device(pdev);
526 		ret = PCI_ERS_RESULT_NEED_RESET;
527 		break;
528 
529 	case pci_channel_io_perm_failure:
530 		if (dd) {
531 			dd_dev_info(dd, "State Permanent Failure, disabling\n");
532 			/* no more register accesses! */
533 			dd->flags &= ~HFI1_PRESENT;
534 			hfi1_disable_after_error(dd);
535 		}
536 		 /* else early, or other problem */
537 		ret =  PCI_ERS_RESULT_DISCONNECT;
538 		break;
539 
540 	default: /* shouldn't happen */
541 		dd_dev_info(dd, "HFI1 PCI errors detected (state %d)\n",
542 			    state);
543 		break;
544 	}
545 	return ret;
546 }
547 
548 static pci_ers_result_t
549 pci_mmio_enabled(struct pci_dev *pdev)
550 {
551 	u64 words = 0U;
552 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
553 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
554 
555 	if (dd && dd->pport) {
556 		words = read_port_cntr(dd->pport, C_RX_WORDS, CNTR_INVALID_VL);
557 		if (words == ~0ULL)
558 			ret = PCI_ERS_RESULT_NEED_RESET;
559 		dd_dev_info(dd,
560 			    "HFI1 mmio_enabled function called, read wordscntr %llx, returning %d\n",
561 			    words, ret);
562 	}
563 	return  ret;
564 }
565 
566 static pci_ers_result_t
567 pci_slot_reset(struct pci_dev *pdev)
568 {
569 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
570 
571 	dd_dev_info(dd, "HFI1 slot_reset function called, ignored\n");
572 	return PCI_ERS_RESULT_CAN_RECOVER;
573 }
574 
575 static void
576 pci_resume(struct pci_dev *pdev)
577 {
578 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
579 
580 	dd_dev_info(dd, "HFI1 resume function called\n");
581 	/*
582 	 * Running jobs will fail, since it's asynchronous
583 	 * unlike sysfs-requested reset.   Better than
584 	 * doing nothing.
585 	 */
586 	hfi1_init(dd, 1); /* same as re-init after reset */
587 }
588 
589 const struct pci_error_handlers hfi1_pci_err_handler = {
590 	.error_detected = pci_error_detected,
591 	.mmio_enabled = pci_mmio_enabled,
592 	.slot_reset = pci_slot_reset,
593 	.resume = pci_resume,
594 };
595 
596 /*============================================================================*/
597 /* PCIe Gen3 support */
598 
599 /*
600  * This code is separated out because it is expected to be removed in the
601  * final shipping product.  If not, then it will be revisited and items
602  * will be moved to more standard locations.
603  */
604 
605 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_STS field values */
606 #define DL_STATUS_HFI0 0x1	/* hfi0 firmware download complete */
607 #define DL_STATUS_HFI1 0x2	/* hfi1 firmware download complete */
608 #define DL_STATUS_BOTH 0x3	/* hfi0 and hfi1 firmware download complete */
609 
610 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_ERR field values */
611 #define DL_ERR_NONE		0x0	/* no error */
612 #define DL_ERR_SWAP_PARITY	0x1	/* parity error in SerDes interrupt */
613 					/*   or response data */
614 #define DL_ERR_DISABLED	0x2	/* hfi disabled */
615 #define DL_ERR_SECURITY	0x3	/* security check failed */
616 #define DL_ERR_SBUS		0x4	/* SBus status error */
617 #define DL_ERR_XFR_PARITY	0x5	/* parity error during ROM transfer*/
618 
619 /* gasket block secondary bus reset delay */
620 #define SBR_DELAY_US 200000	/* 200ms */
621 
622 static uint pcie_target = 3;
623 module_param(pcie_target, uint, S_IRUGO);
624 MODULE_PARM_DESC(pcie_target, "PCIe target speed (0 skip, 1-3 Gen1-3)");
625 
626 static uint pcie_force;
627 module_param(pcie_force, uint, S_IRUGO);
628 MODULE_PARM_DESC(pcie_force, "Force driver to do a PCIe firmware download even if already at target speed");
629 
630 static uint pcie_retry = 5;
631 module_param(pcie_retry, uint, S_IRUGO);
632 MODULE_PARM_DESC(pcie_retry, "Driver will try this many times to reach requested speed");
633 
634 #define UNSET_PSET 255
635 #define DEFAULT_DISCRETE_PSET 2	/* discrete HFI */
636 #define DEFAULT_MCP_PSET 6	/* MCP HFI */
637 static uint pcie_pset = UNSET_PSET;
638 module_param(pcie_pset, uint, S_IRUGO);
639 MODULE_PARM_DESC(pcie_pset, "PCIe Eq Pset value to use, range is 0-10");
640 
641 static uint pcie_ctle = 3; /* discrete on, integrated on */
642 module_param(pcie_ctle, uint, S_IRUGO);
643 MODULE_PARM_DESC(pcie_ctle, "PCIe static CTLE mode, bit 0 - discrete on/off, bit 1 - integrated on/off");
644 
645 /* equalization columns */
646 #define PREC 0
647 #define ATTN 1
648 #define POST 2
649 
650 /* discrete silicon preliminary equalization values */
651 static const u8 discrete_preliminary_eq[11][3] = {
652 	/* prec   attn   post */
653 	{  0x00,  0x00,  0x12 },	/* p0 */
654 	{  0x00,  0x00,  0x0c },	/* p1 */
655 	{  0x00,  0x00,  0x0f },	/* p2 */
656 	{  0x00,  0x00,  0x09 },	/* p3 */
657 	{  0x00,  0x00,  0x00 },	/* p4 */
658 	{  0x06,  0x00,  0x00 },	/* p5 */
659 	{  0x09,  0x00,  0x00 },	/* p6 */
660 	{  0x06,  0x00,  0x0f },	/* p7 */
661 	{  0x09,  0x00,  0x09 },	/* p8 */
662 	{  0x0c,  0x00,  0x00 },	/* p9 */
663 	{  0x00,  0x00,  0x18 },	/* p10 */
664 };
665 
666 /* integrated silicon preliminary equalization values */
667 static const u8 integrated_preliminary_eq[11][3] = {
668 	/* prec   attn   post */
669 	{  0x00,  0x1e,  0x07 },	/* p0 */
670 	{  0x00,  0x1e,  0x05 },	/* p1 */
671 	{  0x00,  0x1e,  0x06 },	/* p2 */
672 	{  0x00,  0x1e,  0x04 },	/* p3 */
673 	{  0x00,  0x1e,  0x00 },	/* p4 */
674 	{  0x03,  0x1e,  0x00 },	/* p5 */
675 	{  0x04,  0x1e,  0x00 },	/* p6 */
676 	{  0x03,  0x1e,  0x06 },	/* p7 */
677 	{  0x03,  0x1e,  0x04 },	/* p8 */
678 	{  0x05,  0x1e,  0x00 },	/* p9 */
679 	{  0x00,  0x1e,  0x0a },	/* p10 */
680 };
681 
682 static const u8 discrete_ctle_tunings[11][4] = {
683 	/* DC     LF     HF     BW */
684 	{  0x48,  0x0b,  0x04,  0x04 },	/* p0 */
685 	{  0x60,  0x05,  0x0f,  0x0a },	/* p1 */
686 	{  0x50,  0x09,  0x06,  0x06 },	/* p2 */
687 	{  0x68,  0x05,  0x0f,  0x0a },	/* p3 */
688 	{  0x80,  0x05,  0x0f,  0x0a },	/* p4 */
689 	{  0x70,  0x05,  0x0f,  0x0a },	/* p5 */
690 	{  0x68,  0x05,  0x0f,  0x0a },	/* p6 */
691 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
692 	{  0x48,  0x09,  0x06,  0x06 },	/* p8 */
693 	{  0x60,  0x05,  0x0f,  0x0a },	/* p9 */
694 	{  0x38,  0x0f,  0x00,  0x00 },	/* p10 */
695 };
696 
697 static const u8 integrated_ctle_tunings[11][4] = {
698 	/* DC     LF     HF     BW */
699 	{  0x38,  0x0f,  0x00,  0x00 },	/* p0 */
700 	{  0x38,  0x0f,  0x00,  0x00 },	/* p1 */
701 	{  0x38,  0x0f,  0x00,  0x00 },	/* p2 */
702 	{  0x38,  0x0f,  0x00,  0x00 },	/* p3 */
703 	{  0x58,  0x0a,  0x05,  0x05 },	/* p4 */
704 	{  0x48,  0x0a,  0x05,  0x05 },	/* p5 */
705 	{  0x40,  0x0a,  0x05,  0x05 },	/* p6 */
706 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
707 	{  0x38,  0x0f,  0x00,  0x00 },	/* p8 */
708 	{  0x38,  0x09,  0x06,  0x06 },	/* p9 */
709 	{  0x38,  0x0e,  0x01,  0x01 },	/* p10 */
710 };
711 
712 /* helper to format the value to write to hardware */
713 #define eq_value(pre, curr, post) \
714 	((((u32)(pre)) << \
715 			PCIE_CFG_REG_PL102_GEN3_EQ_PRE_CURSOR_PSET_SHIFT) \
716 	| (((u32)(curr)) << PCIE_CFG_REG_PL102_GEN3_EQ_CURSOR_PSET_SHIFT) \
717 	| (((u32)(post)) << \
718 		PCIE_CFG_REG_PL102_GEN3_EQ_POST_CURSOR_PSET_SHIFT))
719 
720 /*
721  * Load the given EQ preset table into the PCIe hardware.
722  */
723 static int load_eq_table(struct hfi1_devdata *dd, const u8 eq[11][3], u8 fs,
724 			 u8 div)
725 {
726 	struct pci_dev *pdev = dd->pcidev;
727 	u32 hit_error = 0;
728 	u32 violation;
729 	u32 i;
730 	u8 c_minus1, c0, c_plus1;
731 	int ret;
732 
733 	for (i = 0; i < 11; i++) {
734 		/* set index */
735 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL103, i);
736 		/* write the value */
737 		c_minus1 = eq[i][PREC] / div;
738 		c0 = fs - (eq[i][PREC] / div) - (eq[i][POST] / div);
739 		c_plus1 = eq[i][POST] / div;
740 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL102,
741 				       eq_value(c_minus1, c0, c_plus1));
742 		/* check if these coefficients violate EQ rules */
743 		ret = pci_read_config_dword(dd->pcidev,
744 					    PCIE_CFG_REG_PL105, &violation);
745 		if (ret) {
746 			dd_dev_err(dd, "Unable to read from PCI config\n");
747 			hit_error = 1;
748 			break;
749 		}
750 
751 		if (violation
752 		    & PCIE_CFG_REG_PL105_GEN3_EQ_VIOLATE_COEF_RULES_SMASK){
753 			if (hit_error == 0) {
754 				dd_dev_err(dd,
755 					   "Gen3 EQ Table Coefficient rule violations\n");
756 				dd_dev_err(dd, "         prec   attn   post\n");
757 			}
758 			dd_dev_err(dd, "   p%02d:   %02x     %02x     %02x\n",
759 				   i, (u32)eq[i][0], (u32)eq[i][1],
760 				   (u32)eq[i][2]);
761 			dd_dev_err(dd, "            %02x     %02x     %02x\n",
762 				   (u32)c_minus1, (u32)c0, (u32)c_plus1);
763 			hit_error = 1;
764 		}
765 	}
766 	if (hit_error)
767 		return -EINVAL;
768 	return 0;
769 }
770 
771 /*
772  * Steps to be done after the PCIe firmware is downloaded and
773  * before the SBR for the Pcie Gen3.
774  * The SBus resource is already being held.
775  */
776 static void pcie_post_steps(struct hfi1_devdata *dd)
777 {
778 	int i;
779 
780 	set_sbus_fast_mode(dd);
781 	/*
782 	 * Write to the PCIe PCSes to set the G3_LOCKED_NEXT bits to 1.
783 	 * This avoids a spurious framing error that can otherwise be
784 	 * generated by the MAC layer.
785 	 *
786 	 * Use individual addresses since no broadcast is set up.
787 	 */
788 	for (i = 0; i < NUM_PCIE_SERDES; i++) {
789 		sbus_request(dd, pcie_pcs_addrs[dd->hfi1_id][i],
790 			     0x03, WRITE_SBUS_RECEIVER, 0x00022132);
791 	}
792 
793 	clear_sbus_fast_mode(dd);
794 }
795 
796 /*
797  * Trigger a secondary bus reset (SBR) on ourselves using our parent.
798  *
799  * Based on pci_parent_bus_reset() which is not exported by the
800  * kernel core.
801  */
802 static int trigger_sbr(struct hfi1_devdata *dd)
803 {
804 	struct pci_dev *dev = dd->pcidev;
805 	struct pci_dev *pdev;
806 
807 	/* need a parent */
808 	if (!dev->bus->self) {
809 		dd_dev_err(dd, "%s: no parent device\n", __func__);
810 		return -ENOTTY;
811 	}
812 
813 	/* should not be anyone else on the bus */
814 	list_for_each_entry(pdev, &dev->bus->devices, bus_list)
815 		if (pdev != dev) {
816 			dd_dev_err(dd,
817 				   "%s: another device is on the same bus\n",
818 				   __func__);
819 			return -ENOTTY;
820 		}
821 
822 	/*
823 	 * This is an end around to do an SBR during probe time. A new API needs
824 	 * to be implemented to have cleaner interface but this fixes the
825 	 * current brokenness
826 	 */
827 	return pci_bridge_secondary_bus_reset(dev->bus->self);
828 }
829 
830 /*
831  * Write the given gasket interrupt register.
832  */
833 static void write_gasket_interrupt(struct hfi1_devdata *dd, int index,
834 				   u16 code, u16 data)
835 {
836 	write_csr(dd, ASIC_PCIE_SD_INTRPT_LIST + (index * 8),
837 		  (((u64)code << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_CODE_SHIFT) |
838 		   ((u64)data << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_DATA_SHIFT)));
839 }
840 
841 /*
842  * Tell the gasket logic how to react to the reset.
843  */
844 static void arm_gasket_logic(struct hfi1_devdata *dd)
845 {
846 	u64 reg;
847 
848 	reg = (((u64)1 << dd->hfi1_id) <<
849 	       ASIC_PCIE_SD_HOST_CMD_INTRPT_CMD_SHIFT) |
850 	      ((u64)pcie_serdes_broadcast[dd->hfi1_id] <<
851 	       ASIC_PCIE_SD_HOST_CMD_SBUS_RCVR_ADDR_SHIFT |
852 	       ASIC_PCIE_SD_HOST_CMD_SBR_MODE_SMASK |
853 	       ((u64)SBR_DELAY_US & ASIC_PCIE_SD_HOST_CMD_TIMER_MASK) <<
854 	       ASIC_PCIE_SD_HOST_CMD_TIMER_SHIFT);
855 	write_csr(dd, ASIC_PCIE_SD_HOST_CMD, reg);
856 	/* read back to push the write */
857 	read_csr(dd, ASIC_PCIE_SD_HOST_CMD);
858 }
859 
860 /*
861  * CCE_PCIE_CTRL long name helpers
862  * We redefine these shorter macros to use in the code while leaving
863  * chip_registers.h to be autogenerated from the hardware spec.
864  */
865 #define LANE_BUNDLE_MASK              CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_MASK
866 #define LANE_BUNDLE_SHIFT             CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_SHIFT
867 #define LANE_DELAY_MASK               CCE_PCIE_CTRL_PCIE_LANE_DELAY_MASK
868 #define LANE_DELAY_SHIFT              CCE_PCIE_CTRL_PCIE_LANE_DELAY_SHIFT
869 #define MARGIN_OVERWRITE_ENABLE_SHIFT CCE_PCIE_CTRL_XMT_MARGIN_OVERWRITE_ENABLE_SHIFT
870 #define MARGIN_SHIFT                  CCE_PCIE_CTRL_XMT_MARGIN_SHIFT
871 #define MARGIN_G1_G2_OVERWRITE_MASK   CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_MASK
872 #define MARGIN_G1_G2_OVERWRITE_SHIFT  CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_SHIFT
873 #define MARGIN_GEN1_GEN2_MASK         CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_MASK
874 #define MARGIN_GEN1_GEN2_SHIFT        CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_SHIFT
875 
876  /*
877   * Write xmt_margin for full-swing (WFR-B) or half-swing (WFR-C).
878   */
879 static void write_xmt_margin(struct hfi1_devdata *dd, const char *fname)
880 {
881 	u64 pcie_ctrl;
882 	u64 xmt_margin;
883 	u64 xmt_margin_oe;
884 	u64 lane_delay;
885 	u64 lane_bundle;
886 
887 	pcie_ctrl = read_csr(dd, CCE_PCIE_CTRL);
888 
889 	/*
890 	 * For Discrete, use full-swing.
891 	 *  - PCIe TX defaults to full-swing.
892 	 *    Leave this register as default.
893 	 * For Integrated, use half-swing
894 	 *  - Copy xmt_margin and xmt_margin_oe
895 	 *    from Gen1/Gen2 to Gen3.
896 	 */
897 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL1) { /* integrated */
898 		/* extract initial fields */
899 		xmt_margin = (pcie_ctrl >> MARGIN_GEN1_GEN2_SHIFT)
900 			      & MARGIN_GEN1_GEN2_MASK;
901 		xmt_margin_oe = (pcie_ctrl >> MARGIN_G1_G2_OVERWRITE_SHIFT)
902 				 & MARGIN_G1_G2_OVERWRITE_MASK;
903 		lane_delay = (pcie_ctrl >> LANE_DELAY_SHIFT) & LANE_DELAY_MASK;
904 		lane_bundle = (pcie_ctrl >> LANE_BUNDLE_SHIFT)
905 			       & LANE_BUNDLE_MASK;
906 
907 		/*
908 		 * For A0, EFUSE values are not set.  Override with the
909 		 * correct values.
910 		 */
911 		if (is_ax(dd)) {
912 			/*
913 			 * xmt_margin and OverwiteEnabel should be the
914 			 * same for Gen1/Gen2 and Gen3
915 			 */
916 			xmt_margin = 0x5;
917 			xmt_margin_oe = 0x1;
918 			lane_delay = 0xF; /* Delay 240ns. */
919 			lane_bundle = 0x0; /* Set to 1 lane. */
920 		}
921 
922 		/* overwrite existing values */
923 		pcie_ctrl = (xmt_margin << MARGIN_GEN1_GEN2_SHIFT)
924 			| (xmt_margin_oe << MARGIN_G1_G2_OVERWRITE_SHIFT)
925 			| (xmt_margin << MARGIN_SHIFT)
926 			| (xmt_margin_oe << MARGIN_OVERWRITE_ENABLE_SHIFT)
927 			| (lane_delay << LANE_DELAY_SHIFT)
928 			| (lane_bundle << LANE_BUNDLE_SHIFT);
929 
930 		write_csr(dd, CCE_PCIE_CTRL, pcie_ctrl);
931 	}
932 
933 	dd_dev_dbg(dd, "%s: program XMT margin, CcePcieCtrl 0x%llx\n",
934 		   fname, pcie_ctrl);
935 }
936 
937 /*
938  * Do all the steps needed to transition the PCIe link to Gen3 speed.
939  */
940 int do_pcie_gen3_transition(struct hfi1_devdata *dd)
941 {
942 	struct pci_dev *parent = dd->pcidev->bus->self;
943 	u64 fw_ctrl;
944 	u64 reg, therm;
945 	u32 reg32, fs, lf;
946 	u32 status, err;
947 	int ret;
948 	int do_retry, retry_count = 0;
949 	int intnum = 0;
950 	uint default_pset;
951 	uint pset = pcie_pset;
952 	u16 target_vector, target_speed;
953 	u16 lnkctl2, vendor;
954 	u8 div;
955 	const u8 (*eq)[3];
956 	const u8 (*ctle_tunings)[4];
957 	uint static_ctle_mode;
958 	int return_error = 0;
959 	u32 target_width;
960 
961 	/* PCIe Gen3 is for the ASIC only */
962 	if (dd->icode != ICODE_RTL_SILICON)
963 		return 0;
964 
965 	if (pcie_target == 1) {			/* target Gen1 */
966 		target_vector = PCI_EXP_LNKCTL2_TLS_2_5GT;
967 		target_speed = 2500;
968 	} else if (pcie_target == 2) {		/* target Gen2 */
969 		target_vector = PCI_EXP_LNKCTL2_TLS_5_0GT;
970 		target_speed = 5000;
971 	} else if (pcie_target == 3) {		/* target Gen3 */
972 		target_vector = PCI_EXP_LNKCTL2_TLS_8_0GT;
973 		target_speed = 8000;
974 	} else {
975 		/* off or invalid target - skip */
976 		dd_dev_info(dd, "%s: Skipping PCIe transition\n", __func__);
977 		return 0;
978 	}
979 
980 	/* if already at target speed, done (unless forced) */
981 	if (dd->lbus_speed == target_speed) {
982 		dd_dev_info(dd, "%s: PCIe already at gen%d, %s\n", __func__,
983 			    pcie_target,
984 			    pcie_force ? "re-doing anyway" : "skipping");
985 		if (!pcie_force)
986 			return 0;
987 	}
988 
989 	/*
990 	 * The driver cannot do the transition if it has no access to the
991 	 * upstream component
992 	 */
993 	if (!parent) {
994 		dd_dev_info(dd, "%s: No upstream, Can't do gen3 transition\n",
995 			    __func__);
996 		return 0;
997 	}
998 
999 	/* Previous Gen1/Gen2 bus width */
1000 	target_width = dd->lbus_width;
1001 
1002 	/*
1003 	 * Do the Gen3 transition.  Steps are those of the PCIe Gen3
1004 	 * recipe.
1005 	 */
1006 
1007 	/* step 1: pcie link working in gen1/gen2 */
1008 
1009 	/* step 2: if either side is not capable of Gen3, done */
1010 	if (pcie_target == 3 && !dd->link_gen3_capable) {
1011 		dd_dev_err(dd, "The PCIe link is not Gen3 capable\n");
1012 		ret = -ENOSYS;
1013 		goto done_no_mutex;
1014 	}
1015 
1016 	/* hold the SBus resource across the firmware download and SBR */
1017 	ret = acquire_chip_resource(dd, CR_SBUS, SBUS_TIMEOUT);
1018 	if (ret) {
1019 		dd_dev_err(dd, "%s: unable to acquire SBus resource\n",
1020 			   __func__);
1021 		return ret;
1022 	}
1023 
1024 	/* make sure thermal polling is not causing interrupts */
1025 	therm = read_csr(dd, ASIC_CFG_THERM_POLL_EN);
1026 	if (therm) {
1027 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x0);
1028 		msleep(100);
1029 		dd_dev_info(dd, "%s: Disabled therm polling\n",
1030 			    __func__);
1031 	}
1032 
1033 retry:
1034 	/* the SBus download will reset the spico for thermal */
1035 
1036 	/* step 3: download SBus Master firmware */
1037 	/* step 4: download PCIe Gen3 SerDes firmware */
1038 	dd_dev_info(dd, "%s: downloading firmware\n", __func__);
1039 	ret = load_pcie_firmware(dd);
1040 	if (ret) {
1041 		/* do not proceed if the firmware cannot be downloaded */
1042 		return_error = 1;
1043 		goto done;
1044 	}
1045 
1046 	/* step 5: set up device parameter settings */
1047 	dd_dev_info(dd, "%s: setting PCIe registers\n", __func__);
1048 
1049 	/*
1050 	 * PcieCfgSpcie1 - Link Control 3
1051 	 * Leave at reset value.  No need to set PerfEq - link equalization
1052 	 * will be performed automatically after the SBR when the target
1053 	 * speed is 8GT/s.
1054 	 */
1055 
1056 	/* clear all 16 per-lane error bits (PCIe: Lane Error Status) */
1057 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, 0xffff);
1058 
1059 	/* step 5a: Set Synopsys Port Logic registers */
1060 
1061 	/*
1062 	 * PcieCfgRegPl2 - Port Force Link
1063 	 *
1064 	 * Set the low power field to 0x10 to avoid unnecessary power
1065 	 * management messages.  All other fields are zero.
1066 	 */
1067 	reg32 = 0x10ul << PCIE_CFG_REG_PL2_LOW_PWR_ENT_CNT_SHIFT;
1068 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL2, reg32);
1069 
1070 	/*
1071 	 * PcieCfgRegPl100 - Gen3 Control
1072 	 *
1073 	 * turn off PcieCfgRegPl100.Gen3ZRxDcNonCompl
1074 	 * turn on PcieCfgRegPl100.EqEieosCnt
1075 	 * Everything else zero.
1076 	 */
1077 	reg32 = PCIE_CFG_REG_PL100_EQ_EIEOS_CNT_SMASK;
1078 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL100, reg32);
1079 
1080 	/*
1081 	 * PcieCfgRegPl101 - Gen3 EQ FS and LF
1082 	 * PcieCfgRegPl102 - Gen3 EQ Presets to Coefficients Mapping
1083 	 * PcieCfgRegPl103 - Gen3 EQ Preset Index
1084 	 * PcieCfgRegPl105 - Gen3 EQ Status
1085 	 *
1086 	 * Give initial EQ settings.
1087 	 */
1088 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL0) { /* discrete */
1089 		/* 1000mV, FS=24, LF = 8 */
1090 		fs = 24;
1091 		lf = 8;
1092 		div = 3;
1093 		eq = discrete_preliminary_eq;
1094 		default_pset = DEFAULT_DISCRETE_PSET;
1095 		ctle_tunings = discrete_ctle_tunings;
1096 		/* bit 0 - discrete on/off */
1097 		static_ctle_mode = pcie_ctle & 0x1;
1098 	} else {
1099 		/* 400mV, FS=29, LF = 9 */
1100 		fs = 29;
1101 		lf = 9;
1102 		div = 1;
1103 		eq = integrated_preliminary_eq;
1104 		default_pset = DEFAULT_MCP_PSET;
1105 		ctle_tunings = integrated_ctle_tunings;
1106 		/* bit 1 - integrated on/off */
1107 		static_ctle_mode = (pcie_ctle >> 1) & 0x1;
1108 	}
1109 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL101,
1110 			       (fs <<
1111 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_FS_SHIFT) |
1112 			       (lf <<
1113 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_LF_SHIFT));
1114 	ret = load_eq_table(dd, eq, fs, div);
1115 	if (ret)
1116 		goto done;
1117 
1118 	/*
1119 	 * PcieCfgRegPl106 - Gen3 EQ Control
1120 	 *
1121 	 * Set Gen3EqPsetReqVec, leave other fields 0.
1122 	 */
1123 	if (pset == UNSET_PSET)
1124 		pset = default_pset;
1125 	if (pset > 10) {	/* valid range is 0-10, inclusive */
1126 		dd_dev_err(dd, "%s: Invalid Eq Pset %u, setting to %d\n",
1127 			   __func__, pset, default_pset);
1128 		pset = default_pset;
1129 	}
1130 	dd_dev_info(dd, "%s: using EQ Pset %u\n", __func__, pset);
1131 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL106,
1132 			       ((1 << pset) <<
1133 			PCIE_CFG_REG_PL106_GEN3_EQ_PSET_REQ_VEC_SHIFT) |
1134 			PCIE_CFG_REG_PL106_GEN3_EQ_EVAL2MS_DISABLE_SMASK |
1135 			PCIE_CFG_REG_PL106_GEN3_EQ_PHASE23_EXIT_MODE_SMASK);
1136 
1137 	/*
1138 	 * step 5b: Do post firmware download steps via SBus
1139 	 */
1140 	dd_dev_info(dd, "%s: doing pcie post steps\n", __func__);
1141 	pcie_post_steps(dd);
1142 
1143 	/*
1144 	 * step 5c: Program gasket interrupts
1145 	 */
1146 	/* set the Rx Bit Rate to REFCLK ratio */
1147 	write_gasket_interrupt(dd, intnum++, 0x0006, 0x0050);
1148 	/* disable pCal for PCIe Gen3 RX equalization */
1149 	/* select adaptive or static CTLE */
1150 	write_gasket_interrupt(dd, intnum++, 0x0026,
1151 			       0x5b01 | (static_ctle_mode << 3));
1152 	/*
1153 	 * Enable iCal for PCIe Gen3 RX equalization, and set which
1154 	 * evaluation of RX_EQ_EVAL will launch the iCal procedure.
1155 	 */
1156 	write_gasket_interrupt(dd, intnum++, 0x0026, 0x5202);
1157 
1158 	if (static_ctle_mode) {
1159 		/* apply static CTLE tunings */
1160 		u8 pcie_dc, pcie_lf, pcie_hf, pcie_bw;
1161 
1162 		pcie_dc = ctle_tunings[pset][0];
1163 		pcie_lf = ctle_tunings[pset][1];
1164 		pcie_hf = ctle_tunings[pset][2];
1165 		pcie_bw = ctle_tunings[pset][3];
1166 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0200 | pcie_dc);
1167 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0100 | pcie_lf);
1168 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0000 | pcie_hf);
1169 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x5500 | pcie_bw);
1170 	}
1171 
1172 	/* terminate list */
1173 	write_gasket_interrupt(dd, intnum++, 0x0000, 0x0000);
1174 
1175 	/*
1176 	 * step 5d: program XMT margin
1177 	 */
1178 	write_xmt_margin(dd, __func__);
1179 
1180 	/*
1181 	 * step 5e: disable active state power management (ASPM). It
1182 	 * will be enabled if required later
1183 	 */
1184 	dd_dev_info(dd, "%s: clearing ASPM\n", __func__);
1185 	aspm_hw_disable_l1(dd);
1186 
1187 	/*
1188 	 * step 5f: clear DirectSpeedChange
1189 	 * PcieCfgRegPl67.DirectSpeedChange must be zero to prevent the
1190 	 * change in the speed target from starting before we are ready.
1191 	 * This field defaults to 0 and we are not changing it, so nothing
1192 	 * needs to be done.
1193 	 */
1194 
1195 	/* step 5g: Set target link speed */
1196 	/*
1197 	 * Set target link speed to be target on both device and parent.
1198 	 * On setting the parent: Some system BIOSs "helpfully" set the
1199 	 * parent target speed to Gen2 to match the ASIC's initial speed.
1200 	 * We can set the target Gen3 because we have already checked
1201 	 * that it is Gen3 capable earlier.
1202 	 */
1203 	dd_dev_info(dd, "%s: setting parent target link speed\n", __func__);
1204 	ret = pcie_capability_read_word(parent, PCI_EXP_LNKCTL2, &lnkctl2);
1205 	if (ret) {
1206 		dd_dev_err(dd, "Unable to read from PCI config\n");
1207 		return_error = 1;
1208 		goto done;
1209 	}
1210 
1211 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1212 		    (u32)lnkctl2);
1213 	/* only write to parent if target is not as high as ours */
1214 	if ((lnkctl2 & PCI_EXP_LNKCTL2_TLS) < target_vector) {
1215 		lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1216 		lnkctl2 |= target_vector;
1217 		dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1218 			    (u32)lnkctl2);
1219 		ret = pcie_capability_write_word(parent,
1220 						 PCI_EXP_LNKCTL2, lnkctl2);
1221 		if (ret) {
1222 			dd_dev_err(dd, "Unable to write to PCI config\n");
1223 			return_error = 1;
1224 			goto done;
1225 		}
1226 	} else {
1227 		dd_dev_info(dd, "%s: ..target speed is OK\n", __func__);
1228 	}
1229 
1230 	dd_dev_info(dd, "%s: setting target link speed\n", __func__);
1231 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL2, &lnkctl2);
1232 	if (ret) {
1233 		dd_dev_err(dd, "Unable to read from PCI config\n");
1234 		return_error = 1;
1235 		goto done;
1236 	}
1237 
1238 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1239 		    (u32)lnkctl2);
1240 	lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1241 	lnkctl2 |= target_vector;
1242 	dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1243 		    (u32)lnkctl2);
1244 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL2, lnkctl2);
1245 	if (ret) {
1246 		dd_dev_err(dd, "Unable to write to PCI config\n");
1247 		return_error = 1;
1248 		goto done;
1249 	}
1250 
1251 	/* step 5h: arm gasket logic */
1252 	/* hold DC in reset across the SBR */
1253 	write_csr(dd, CCE_DC_CTRL, CCE_DC_CTRL_DC_RESET_SMASK);
1254 	(void)read_csr(dd, CCE_DC_CTRL); /* DC reset hold */
1255 	/* save firmware control across the SBR */
1256 	fw_ctrl = read_csr(dd, MISC_CFG_FW_CTRL);
1257 
1258 	dd_dev_info(dd, "%s: arming gasket logic\n", __func__);
1259 	arm_gasket_logic(dd);
1260 
1261 	/*
1262 	 * step 6: quiesce PCIe link
1263 	 * The chip has already been reset, so there will be no traffic
1264 	 * from the chip.  Linux has no easy way to enforce that it will
1265 	 * not try to access the device, so we just need to hope it doesn't
1266 	 * do it while we are doing the reset.
1267 	 */
1268 
1269 	/*
1270 	 * step 7: initiate the secondary bus reset (SBR)
1271 	 * step 8: hardware brings the links back up
1272 	 * step 9: wait for link speed transition to be complete
1273 	 */
1274 	dd_dev_info(dd, "%s: calling trigger_sbr\n", __func__);
1275 	ret = trigger_sbr(dd);
1276 	if (ret)
1277 		goto done;
1278 
1279 	/* step 10: decide what to do next */
1280 
1281 	/* check if we can read PCI space */
1282 	ret = pci_read_config_word(dd->pcidev, PCI_VENDOR_ID, &vendor);
1283 	if (ret) {
1284 		dd_dev_info(dd,
1285 			    "%s: read of VendorID failed after SBR, err %d\n",
1286 			    __func__, ret);
1287 		return_error = 1;
1288 		goto done;
1289 	}
1290 	if (vendor == 0xffff) {
1291 		dd_dev_info(dd, "%s: VendorID is all 1s after SBR\n", __func__);
1292 		return_error = 1;
1293 		ret = -EIO;
1294 		goto done;
1295 	}
1296 
1297 	/* restore PCI space registers we know were reset */
1298 	dd_dev_info(dd, "%s: calling restore_pci_variables\n", __func__);
1299 	ret = restore_pci_variables(dd);
1300 	if (ret) {
1301 		dd_dev_err(dd, "%s: Could not restore PCI variables\n",
1302 			   __func__);
1303 		return_error = 1;
1304 		goto done;
1305 	}
1306 
1307 	/* restore firmware control */
1308 	write_csr(dd, MISC_CFG_FW_CTRL, fw_ctrl);
1309 
1310 	/*
1311 	 * Check the gasket block status.
1312 	 *
1313 	 * This is the first CSR read after the SBR.  If the read returns
1314 	 * all 1s (fails), the link did not make it back.
1315 	 *
1316 	 * Once we're sure we can read and write, clear the DC reset after
1317 	 * the SBR.  Then check for any per-lane errors. Then look over
1318 	 * the status.
1319 	 */
1320 	reg = read_csr(dd, ASIC_PCIE_SD_HOST_STATUS);
1321 	dd_dev_info(dd, "%s: gasket block status: 0x%llx\n", __func__, reg);
1322 	if (reg == ~0ull) {	/* PCIe read failed/timeout */
1323 		dd_dev_err(dd, "SBR failed - unable to read from device\n");
1324 		return_error = 1;
1325 		ret = -ENOSYS;
1326 		goto done;
1327 	}
1328 
1329 	/* clear the DC reset */
1330 	write_csr(dd, CCE_DC_CTRL, 0);
1331 
1332 	/* Set the LED off */
1333 	setextled(dd, 0);
1334 
1335 	/* check for any per-lane errors */
1336 	ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, &reg32);
1337 	if (ret) {
1338 		dd_dev_err(dd, "Unable to read from PCI config\n");
1339 		return_error = 1;
1340 		goto done;
1341 	}
1342 
1343 	dd_dev_info(dd, "%s: per-lane errors: 0x%x\n", __func__, reg32);
1344 
1345 	/* extract status, look for our HFI */
1346 	status = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_SHIFT)
1347 			& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_MASK;
1348 	if ((status & (1 << dd->hfi1_id)) == 0) {
1349 		dd_dev_err(dd,
1350 			   "%s: gasket status 0x%x, expecting 0x%x\n",
1351 			   __func__, status, 1 << dd->hfi1_id);
1352 		ret = -EIO;
1353 		goto done;
1354 	}
1355 
1356 	/* extract error */
1357 	err = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_SHIFT)
1358 		& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_MASK;
1359 	if (err) {
1360 		dd_dev_err(dd, "%s: gasket error %d\n", __func__, err);
1361 		ret = -EIO;
1362 		goto done;
1363 	}
1364 
1365 	/* update our link information cache */
1366 	update_lbus_info(dd);
1367 	dd_dev_info(dd, "%s: new speed and width: %s\n", __func__,
1368 		    dd->lbus_info);
1369 
1370 	if (dd->lbus_speed != target_speed ||
1371 	    dd->lbus_width < target_width) { /* not target */
1372 		/* maybe retry */
1373 		do_retry = retry_count < pcie_retry;
1374 		dd_dev_err(dd, "PCIe link speed or width did not match target%s\n",
1375 			   do_retry ? ", retrying" : "");
1376 		retry_count++;
1377 		if (do_retry) {
1378 			msleep(100); /* allow time to settle */
1379 			goto retry;
1380 		}
1381 		ret = -EIO;
1382 	}
1383 
1384 done:
1385 	if (therm) {
1386 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x1);
1387 		msleep(100);
1388 		dd_dev_info(dd, "%s: Re-enable therm polling\n",
1389 			    __func__);
1390 	}
1391 	release_chip_resource(dd, CR_SBUS);
1392 done_no_mutex:
1393 	/* return no error if it is OK to be at current speed */
1394 	if (ret && !return_error) {
1395 		dd_dev_err(dd, "Proceeding at current speed PCIe speed\n");
1396 		ret = 0;
1397 	}
1398 
1399 	dd_dev_info(dd, "%s: done\n", __func__);
1400 	return ret;
1401 }
1402