xref: /openbmc/linux/drivers/infiniband/hw/hfi1/pcie.c (revision 99a15348)
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/aer.h>
11 #include <linux/module.h>
12 
13 #include "hfi.h"
14 #include "chip_registers.h"
15 #include "aspm.h"
16 
17 /*
18  * This file contains PCIe utility routines.
19  */
20 
21 /*
22  * Do all the common PCIe setup and initialization.
23  */
24 int hfi1_pcie_init(struct hfi1_devdata *dd)
25 {
26 	int ret;
27 	struct pci_dev *pdev = dd->pcidev;
28 
29 	ret = pci_enable_device(pdev);
30 	if (ret) {
31 		/*
32 		 * This can happen (in theory) iff:
33 		 * We did a chip reset, and then failed to reprogram the
34 		 * BAR, or the chip reset due to an internal error.  We then
35 		 * unloaded the driver and reloaded it.
36 		 *
37 		 * Both reset cases set the BAR back to initial state.  For
38 		 * the latter case, the AER sticky error bit at offset 0x718
39 		 * should be set, but the Linux kernel doesn't yet know
40 		 * about that, it appears.  If the original BAR was retained
41 		 * in the kernel data structures, this may be OK.
42 		 */
43 		dd_dev_err(dd, "pci enable failed: error %d\n", -ret);
44 		return ret;
45 	}
46 
47 	ret = pci_request_regions(pdev, DRIVER_NAME);
48 	if (ret) {
49 		dd_dev_err(dd, "pci_request_regions fails: err %d\n", -ret);
50 		goto bail;
51 	}
52 
53 	ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
54 	if (ret) {
55 		/*
56 		 * If the 64 bit setup fails, try 32 bit.  Some systems
57 		 * do not setup 64 bit maps on systems with 2GB or less
58 		 * memory installed.
59 		 */
60 		ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
61 		if (ret) {
62 			dd_dev_err(dd, "Unable to set DMA mask: %d\n", ret);
63 			goto bail;
64 		}
65 	}
66 
67 	pci_set_master(pdev);
68 	(void)pci_enable_pcie_error_reporting(pdev);
69 	return 0;
70 
71 bail:
72 	hfi1_pcie_cleanup(pdev);
73 	return ret;
74 }
75 
76 /*
77  * Clean what was done in hfi1_pcie_init()
78  */
79 void hfi1_pcie_cleanup(struct pci_dev *pdev)
80 {
81 	pci_disable_device(pdev);
82 	/*
83 	 * Release regions should be called after the disable. OK to
84 	 * call if request regions has not been called or failed.
85 	 */
86 	pci_release_regions(pdev);
87 }
88 
89 /*
90  * Do remaining PCIe setup, once dd is allocated, and save away
91  * fields required to re-initialize after a chip reset, or for
92  * various other purposes
93  */
94 int hfi1_pcie_ddinit(struct hfi1_devdata *dd, struct pci_dev *pdev)
95 {
96 	unsigned long len;
97 	resource_size_t addr;
98 	int ret = 0;
99 	u32 rcv_array_count;
100 
101 	addr = pci_resource_start(pdev, 0);
102 	len = pci_resource_len(pdev, 0);
103 
104 	/*
105 	 * The TXE PIO buffers are at the tail end of the chip space.
106 	 * Cut them off and map them separately.
107 	 */
108 
109 	/* sanity check vs expectations */
110 	if (len != TXE_PIO_SEND + TXE_PIO_SIZE) {
111 		dd_dev_err(dd, "chip PIO range does not match\n");
112 		return -EINVAL;
113 	}
114 
115 	dd->kregbase1 = ioremap(addr, RCV_ARRAY);
116 	if (!dd->kregbase1) {
117 		dd_dev_err(dd, "UC mapping of kregbase1 failed\n");
118 		return -ENOMEM;
119 	}
120 	dd_dev_info(dd, "UC base1: %p for %x\n", dd->kregbase1, RCV_ARRAY);
121 
122 	/* verify that reads actually work, save revision for reset check */
123 	dd->revision = readq(dd->kregbase1 + CCE_REVISION);
124 	if (dd->revision == ~(u64)0) {
125 		dd_dev_err(dd, "Cannot read chip CSRs\n");
126 		goto nomem;
127 	}
128 
129 	rcv_array_count = readq(dd->kregbase1 + RCV_ARRAY_CNT);
130 	dd_dev_info(dd, "RcvArray count: %u\n", rcv_array_count);
131 	dd->base2_start  = RCV_ARRAY + rcv_array_count * 8;
132 
133 	dd->kregbase2 = ioremap(
134 		addr + dd->base2_start,
135 		TXE_PIO_SEND - dd->base2_start);
136 	if (!dd->kregbase2) {
137 		dd_dev_err(dd, "UC mapping of kregbase2 failed\n");
138 		goto nomem;
139 	}
140 	dd_dev_info(dd, "UC base2: %p for %x\n", dd->kregbase2,
141 		    TXE_PIO_SEND - dd->base2_start);
142 
143 	dd->piobase = ioremap_wc(addr + TXE_PIO_SEND, TXE_PIO_SIZE);
144 	if (!dd->piobase) {
145 		dd_dev_err(dd, "WC mapping of send buffers failed\n");
146 		goto nomem;
147 	}
148 	dd_dev_info(dd, "WC piobase: %p for %x\n", dd->piobase, TXE_PIO_SIZE);
149 
150 	dd->physaddr = addr;        /* used for io_remap, etc. */
151 
152 	/*
153 	 * Map the chip's RcvArray as write-combining to allow us
154 	 * to write an entire cacheline worth of entries in one shot.
155 	 */
156 	dd->rcvarray_wc = ioremap_wc(addr + RCV_ARRAY,
157 				     rcv_array_count * 8);
158 	if (!dd->rcvarray_wc) {
159 		dd_dev_err(dd, "WC mapping of receive array failed\n");
160 		goto nomem;
161 	}
162 	dd_dev_info(dd, "WC RcvArray: %p for %x\n",
163 		    dd->rcvarray_wc, rcv_array_count * 8);
164 
165 	dd->flags |= HFI1_PRESENT;	/* chip.c CSR routines now work */
166 	return 0;
167 nomem:
168 	ret = -ENOMEM;
169 	hfi1_pcie_ddcleanup(dd);
170 	return ret;
171 }
172 
173 /*
174  * Do PCIe cleanup related to dd, after chip-specific cleanup, etc.  Just prior
175  * to releasing the dd memory.
176  * Void because all of the core pcie cleanup functions are void.
177  */
178 void hfi1_pcie_ddcleanup(struct hfi1_devdata *dd)
179 {
180 	dd->flags &= ~HFI1_PRESENT;
181 	if (dd->kregbase1)
182 		iounmap(dd->kregbase1);
183 	dd->kregbase1 = NULL;
184 	if (dd->kregbase2)
185 		iounmap(dd->kregbase2);
186 	dd->kregbase2 = NULL;
187 	if (dd->rcvarray_wc)
188 		iounmap(dd->rcvarray_wc);
189 	dd->rcvarray_wc = NULL;
190 	if (dd->piobase)
191 		iounmap(dd->piobase);
192 	dd->piobase = NULL;
193 }
194 
195 /* return the PCIe link speed from the given link status */
196 static u32 extract_speed(u16 linkstat)
197 {
198 	u32 speed;
199 
200 	switch (linkstat & PCI_EXP_LNKSTA_CLS) {
201 	default: /* not defined, assume Gen1 */
202 	case PCI_EXP_LNKSTA_CLS_2_5GB:
203 		speed = 2500; /* Gen 1, 2.5GHz */
204 		break;
205 	case PCI_EXP_LNKSTA_CLS_5_0GB:
206 		speed = 5000; /* Gen 2, 5GHz */
207 		break;
208 	case PCI_EXP_LNKSTA_CLS_8_0GB:
209 		speed = 8000; /* Gen 3, 8GHz */
210 		break;
211 	}
212 	return speed;
213 }
214 
215 /* return the PCIe link speed from the given link status */
216 static u32 extract_width(u16 linkstat)
217 {
218 	return (linkstat & PCI_EXP_LNKSTA_NLW) >> PCI_EXP_LNKSTA_NLW_SHIFT;
219 }
220 
221 /* read the link status and set dd->{lbus_width,lbus_speed,lbus_info} */
222 static void update_lbus_info(struct hfi1_devdata *dd)
223 {
224 	u16 linkstat;
225 	int ret;
226 
227 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
228 	if (ret) {
229 		dd_dev_err(dd, "Unable to read from PCI config\n");
230 		return;
231 	}
232 
233 	dd->lbus_width = extract_width(linkstat);
234 	dd->lbus_speed = extract_speed(linkstat);
235 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
236 		 "PCIe,%uMHz,x%u", dd->lbus_speed, dd->lbus_width);
237 }
238 
239 /*
240  * Read in the current PCIe link width and speed.  Find if the link is
241  * Gen3 capable.
242  */
243 int pcie_speeds(struct hfi1_devdata *dd)
244 {
245 	u32 linkcap;
246 	struct pci_dev *parent = dd->pcidev->bus->self;
247 	int ret;
248 
249 	if (!pci_is_pcie(dd->pcidev)) {
250 		dd_dev_err(dd, "Can't find PCI Express capability!\n");
251 		return -EINVAL;
252 	}
253 
254 	/* find if our max speed is Gen3 and parent supports Gen3 speeds */
255 	dd->link_gen3_capable = 1;
256 
257 	ret = pcie_capability_read_dword(dd->pcidev, PCI_EXP_LNKCAP, &linkcap);
258 	if (ret) {
259 		dd_dev_err(dd, "Unable to read from PCI config\n");
260 		return pcibios_err_to_errno(ret);
261 	}
262 
263 	if ((linkcap & PCI_EXP_LNKCAP_SLS) != PCI_EXP_LNKCAP_SLS_8_0GB) {
264 		dd_dev_info(dd,
265 			    "This HFI is not Gen3 capable, max speed 0x%x, need 0x3\n",
266 			    linkcap & PCI_EXP_LNKCAP_SLS);
267 		dd->link_gen3_capable = 0;
268 	}
269 
270 	/*
271 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
272 	 */
273 	if (parent &&
274 	    (dd->pcidev->bus->max_bus_speed == PCIE_SPEED_2_5GT ||
275 	     dd->pcidev->bus->max_bus_speed == PCIE_SPEED_5_0GT)) {
276 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
277 		dd->link_gen3_capable = 0;
278 	}
279 
280 	/* obtain the link width and current speed */
281 	update_lbus_info(dd);
282 
283 	dd_dev_info(dd, "%s\n", dd->lbus_info);
284 
285 	return 0;
286 }
287 
288 /*
289  * Restore command and BARs after a reset has wiped them out
290  *
291  * Returns 0 on success, otherwise a negative error value
292  */
293 int restore_pci_variables(struct hfi1_devdata *dd)
294 {
295 	int ret;
296 
297 	ret = pci_write_config_word(dd->pcidev, PCI_COMMAND, dd->pci_command);
298 	if (ret)
299 		goto error;
300 
301 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
302 				     dd->pcibar0);
303 	if (ret)
304 		goto error;
305 
306 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
307 				     dd->pcibar1);
308 	if (ret)
309 		goto error;
310 
311 	ret = pci_write_config_dword(dd->pcidev, PCI_ROM_ADDRESS, dd->pci_rom);
312 	if (ret)
313 		goto error;
314 
315 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL,
316 					 dd->pcie_devctl);
317 	if (ret)
318 		goto error;
319 
320 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL,
321 					 dd->pcie_lnkctl);
322 	if (ret)
323 		goto error;
324 
325 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL2,
326 					 dd->pcie_devctl2);
327 	if (ret)
328 		goto error;
329 
330 	ret = pci_write_config_dword(dd->pcidev, PCI_CFG_MSIX0, dd->pci_msix0);
331 	if (ret)
332 		goto error;
333 
334 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
335 		ret = pci_write_config_dword(dd->pcidev, PCIE_CFG_TPH2,
336 					     dd->pci_tph2);
337 		if (ret)
338 			goto error;
339 	}
340 	return 0;
341 
342 error:
343 	dd_dev_err(dd, "Unable to write to PCI config\n");
344 	return pcibios_err_to_errno(ret);
345 }
346 
347 /*
348  * Save BARs and command to rewrite after device reset
349  *
350  * Returns 0 on success, otherwise a negative error value
351  */
352 int save_pci_variables(struct hfi1_devdata *dd)
353 {
354 	int ret;
355 
356 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
357 				    &dd->pcibar0);
358 	if (ret)
359 		goto error;
360 
361 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
362 				    &dd->pcibar1);
363 	if (ret)
364 		goto error;
365 
366 	ret = pci_read_config_dword(dd->pcidev, PCI_ROM_ADDRESS, &dd->pci_rom);
367 	if (ret)
368 		goto error;
369 
370 	ret = pci_read_config_word(dd->pcidev, PCI_COMMAND, &dd->pci_command);
371 	if (ret)
372 		goto error;
373 
374 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL,
375 					&dd->pcie_devctl);
376 	if (ret)
377 		goto error;
378 
379 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL,
380 					&dd->pcie_lnkctl);
381 	if (ret)
382 		goto error;
383 
384 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL2,
385 					&dd->pcie_devctl2);
386 	if (ret)
387 		goto error;
388 
389 	ret = pci_read_config_dword(dd->pcidev, PCI_CFG_MSIX0, &dd->pci_msix0);
390 	if (ret)
391 		goto error;
392 
393 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
394 		ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_TPH2,
395 					    &dd->pci_tph2);
396 		if (ret)
397 			goto error;
398 	}
399 	return 0;
400 
401 error:
402 	dd_dev_err(dd, "Unable to read from PCI config\n");
403 	return pcibios_err_to_errno(ret);
404 }
405 
406 /*
407  * BIOS may not set PCIe bus-utilization parameters for best performance.
408  * Check and optionally adjust them to maximize our throughput.
409  */
410 static int hfi1_pcie_caps;
411 module_param_named(pcie_caps, hfi1_pcie_caps, int, 0444);
412 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
413 
414 /**
415  * tune_pcie_caps() - Code to adjust PCIe capabilities.
416  * @dd: Valid device data structure
417  *
418  */
419 void tune_pcie_caps(struct hfi1_devdata *dd)
420 {
421 	struct pci_dev *parent;
422 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
423 	u16 rc_mrrs, ep_mrrs, max_mrrs, ectl;
424 	int ret;
425 
426 	/*
427 	 * Turn on extended tags in DevCtl in case the BIOS has turned it off
428 	 * to improve WFR SDMA bandwidth
429 	 */
430 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &ectl);
431 	if ((!ret) && !(ectl & PCI_EXP_DEVCTL_EXT_TAG)) {
432 		dd_dev_info(dd, "Enabling PCIe extended tags\n");
433 		ectl |= PCI_EXP_DEVCTL_EXT_TAG;
434 		ret = pcie_capability_write_word(dd->pcidev,
435 						 PCI_EXP_DEVCTL, ectl);
436 		if (ret)
437 			dd_dev_info(dd, "Unable to write to PCI config\n");
438 	}
439 	/* Find out supported and configured values for parent (root) */
440 	parent = dd->pcidev->bus->self;
441 	/*
442 	 * The driver cannot perform the tuning if it does not have
443 	 * access to the upstream component.
444 	 */
445 	if (!parent) {
446 		dd_dev_info(dd, "Parent not found\n");
447 		return;
448 	}
449 	if (!pci_is_root_bus(parent->bus)) {
450 		dd_dev_info(dd, "Parent not root\n");
451 		return;
452 	}
453 	if (!pci_is_pcie(parent)) {
454 		dd_dev_info(dd, "Parent is not PCI Express capable\n");
455 		return;
456 	}
457 	if (!pci_is_pcie(dd->pcidev)) {
458 		dd_dev_info(dd, "PCI device is not PCI Express capable\n");
459 		return;
460 	}
461 	rc_mpss = parent->pcie_mpss;
462 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
463 	/* Find out supported and configured values for endpoint (us) */
464 	ep_mpss = dd->pcidev->pcie_mpss;
465 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
466 
467 	/* Find max payload supported by root, endpoint */
468 	if (rc_mpss > ep_mpss)
469 		rc_mpss = ep_mpss;
470 
471 	/* If Supported greater than limit in module param, limit it */
472 	if (rc_mpss > (hfi1_pcie_caps & 7))
473 		rc_mpss = hfi1_pcie_caps & 7;
474 	/* If less than (allowed, supported), bump root payload */
475 	if (rc_mpss > rc_mps) {
476 		rc_mps = rc_mpss;
477 		pcie_set_mps(parent, 128 << rc_mps);
478 	}
479 	/* If less than (allowed, supported), bump endpoint payload */
480 	if (rc_mpss > ep_mps) {
481 		ep_mps = rc_mpss;
482 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
483 	}
484 
485 	/*
486 	 * Now the Read Request size.
487 	 * No field for max supported, but PCIe spec limits it to 4096,
488 	 * which is code '5' (log2(4096) - 7)
489 	 */
490 	max_mrrs = 5;
491 	if (max_mrrs > ((hfi1_pcie_caps >> 4) & 7))
492 		max_mrrs = (hfi1_pcie_caps >> 4) & 7;
493 
494 	max_mrrs = 128 << max_mrrs;
495 	rc_mrrs = pcie_get_readrq(parent);
496 	ep_mrrs = pcie_get_readrq(dd->pcidev);
497 
498 	if (max_mrrs > rc_mrrs) {
499 		rc_mrrs = max_mrrs;
500 		pcie_set_readrq(parent, rc_mrrs);
501 	}
502 	if (max_mrrs > ep_mrrs) {
503 		ep_mrrs = max_mrrs;
504 		pcie_set_readrq(dd->pcidev, ep_mrrs);
505 	}
506 }
507 
508 /* End of PCIe capability tuning */
509 
510 /*
511  * From here through hfi1_pci_err_handler definition is invoked via
512  * PCI error infrastructure, registered via pci
513  */
514 static pci_ers_result_t
515 pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
516 {
517 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
518 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
519 
520 	switch (state) {
521 	case pci_channel_io_normal:
522 		dd_dev_info(dd, "State Normal, ignoring\n");
523 		break;
524 
525 	case pci_channel_io_frozen:
526 		dd_dev_info(dd, "State Frozen, requesting reset\n");
527 		pci_disable_device(pdev);
528 		ret = PCI_ERS_RESULT_NEED_RESET;
529 		break;
530 
531 	case pci_channel_io_perm_failure:
532 		if (dd) {
533 			dd_dev_info(dd, "State Permanent Failure, disabling\n");
534 			/* no more register accesses! */
535 			dd->flags &= ~HFI1_PRESENT;
536 			hfi1_disable_after_error(dd);
537 		}
538 		 /* else early, or other problem */
539 		ret =  PCI_ERS_RESULT_DISCONNECT;
540 		break;
541 
542 	default: /* shouldn't happen */
543 		dd_dev_info(dd, "HFI1 PCI errors detected (state %d)\n",
544 			    state);
545 		break;
546 	}
547 	return ret;
548 }
549 
550 static pci_ers_result_t
551 pci_mmio_enabled(struct pci_dev *pdev)
552 {
553 	u64 words = 0U;
554 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
555 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
556 
557 	if (dd && dd->pport) {
558 		words = read_port_cntr(dd->pport, C_RX_WORDS, CNTR_INVALID_VL);
559 		if (words == ~0ULL)
560 			ret = PCI_ERS_RESULT_NEED_RESET;
561 		dd_dev_info(dd,
562 			    "HFI1 mmio_enabled function called, read wordscntr %llx, returning %d\n",
563 			    words, ret);
564 	}
565 	return  ret;
566 }
567 
568 static pci_ers_result_t
569 pci_slot_reset(struct pci_dev *pdev)
570 {
571 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
572 
573 	dd_dev_info(dd, "HFI1 slot_reset function called, ignored\n");
574 	return PCI_ERS_RESULT_CAN_RECOVER;
575 }
576 
577 static void
578 pci_resume(struct pci_dev *pdev)
579 {
580 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
581 
582 	dd_dev_info(dd, "HFI1 resume function called\n");
583 	/*
584 	 * Running jobs will fail, since it's asynchronous
585 	 * unlike sysfs-requested reset.   Better than
586 	 * doing nothing.
587 	 */
588 	hfi1_init(dd, 1); /* same as re-init after reset */
589 }
590 
591 const struct pci_error_handlers hfi1_pci_err_handler = {
592 	.error_detected = pci_error_detected,
593 	.mmio_enabled = pci_mmio_enabled,
594 	.slot_reset = pci_slot_reset,
595 	.resume = pci_resume,
596 };
597 
598 /*============================================================================*/
599 /* PCIe Gen3 support */
600 
601 /*
602  * This code is separated out because it is expected to be removed in the
603  * final shipping product.  If not, then it will be revisited and items
604  * will be moved to more standard locations.
605  */
606 
607 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_STS field values */
608 #define DL_STATUS_HFI0 0x1	/* hfi0 firmware download complete */
609 #define DL_STATUS_HFI1 0x2	/* hfi1 firmware download complete */
610 #define DL_STATUS_BOTH 0x3	/* hfi0 and hfi1 firmware download complete */
611 
612 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_ERR field values */
613 #define DL_ERR_NONE		0x0	/* no error */
614 #define DL_ERR_SWAP_PARITY	0x1	/* parity error in SerDes interrupt */
615 					/*   or response data */
616 #define DL_ERR_DISABLED	0x2	/* hfi disabled */
617 #define DL_ERR_SECURITY	0x3	/* security check failed */
618 #define DL_ERR_SBUS		0x4	/* SBus status error */
619 #define DL_ERR_XFR_PARITY	0x5	/* parity error during ROM transfer*/
620 
621 /* gasket block secondary bus reset delay */
622 #define SBR_DELAY_US 200000	/* 200ms */
623 
624 static uint pcie_target = 3;
625 module_param(pcie_target, uint, S_IRUGO);
626 MODULE_PARM_DESC(pcie_target, "PCIe target speed (0 skip, 1-3 Gen1-3)");
627 
628 static uint pcie_force;
629 module_param(pcie_force, uint, S_IRUGO);
630 MODULE_PARM_DESC(pcie_force, "Force driver to do a PCIe firmware download even if already at target speed");
631 
632 static uint pcie_retry = 5;
633 module_param(pcie_retry, uint, S_IRUGO);
634 MODULE_PARM_DESC(pcie_retry, "Driver will try this many times to reach requested speed");
635 
636 #define UNSET_PSET 255
637 #define DEFAULT_DISCRETE_PSET 2	/* discrete HFI */
638 #define DEFAULT_MCP_PSET 6	/* MCP HFI */
639 static uint pcie_pset = UNSET_PSET;
640 module_param(pcie_pset, uint, S_IRUGO);
641 MODULE_PARM_DESC(pcie_pset, "PCIe Eq Pset value to use, range is 0-10");
642 
643 static uint pcie_ctle = 3; /* discrete on, integrated on */
644 module_param(pcie_ctle, uint, S_IRUGO);
645 MODULE_PARM_DESC(pcie_ctle, "PCIe static CTLE mode, bit 0 - discrete on/off, bit 1 - integrated on/off");
646 
647 /* equalization columns */
648 #define PREC 0
649 #define ATTN 1
650 #define POST 2
651 
652 /* discrete silicon preliminary equalization values */
653 static const u8 discrete_preliminary_eq[11][3] = {
654 	/* prec   attn   post */
655 	{  0x00,  0x00,  0x12 },	/* p0 */
656 	{  0x00,  0x00,  0x0c },	/* p1 */
657 	{  0x00,  0x00,  0x0f },	/* p2 */
658 	{  0x00,  0x00,  0x09 },	/* p3 */
659 	{  0x00,  0x00,  0x00 },	/* p4 */
660 	{  0x06,  0x00,  0x00 },	/* p5 */
661 	{  0x09,  0x00,  0x00 },	/* p6 */
662 	{  0x06,  0x00,  0x0f },	/* p7 */
663 	{  0x09,  0x00,  0x09 },	/* p8 */
664 	{  0x0c,  0x00,  0x00 },	/* p9 */
665 	{  0x00,  0x00,  0x18 },	/* p10 */
666 };
667 
668 /* integrated silicon preliminary equalization values */
669 static const u8 integrated_preliminary_eq[11][3] = {
670 	/* prec   attn   post */
671 	{  0x00,  0x1e,  0x07 },	/* p0 */
672 	{  0x00,  0x1e,  0x05 },	/* p1 */
673 	{  0x00,  0x1e,  0x06 },	/* p2 */
674 	{  0x00,  0x1e,  0x04 },	/* p3 */
675 	{  0x00,  0x1e,  0x00 },	/* p4 */
676 	{  0x03,  0x1e,  0x00 },	/* p5 */
677 	{  0x04,  0x1e,  0x00 },	/* p6 */
678 	{  0x03,  0x1e,  0x06 },	/* p7 */
679 	{  0x03,  0x1e,  0x04 },	/* p8 */
680 	{  0x05,  0x1e,  0x00 },	/* p9 */
681 	{  0x00,  0x1e,  0x0a },	/* p10 */
682 };
683 
684 static const u8 discrete_ctle_tunings[11][4] = {
685 	/* DC     LF     HF     BW */
686 	{  0x48,  0x0b,  0x04,  0x04 },	/* p0 */
687 	{  0x60,  0x05,  0x0f,  0x0a },	/* p1 */
688 	{  0x50,  0x09,  0x06,  0x06 },	/* p2 */
689 	{  0x68,  0x05,  0x0f,  0x0a },	/* p3 */
690 	{  0x80,  0x05,  0x0f,  0x0a },	/* p4 */
691 	{  0x70,  0x05,  0x0f,  0x0a },	/* p5 */
692 	{  0x68,  0x05,  0x0f,  0x0a },	/* p6 */
693 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
694 	{  0x48,  0x09,  0x06,  0x06 },	/* p8 */
695 	{  0x60,  0x05,  0x0f,  0x0a },	/* p9 */
696 	{  0x38,  0x0f,  0x00,  0x00 },	/* p10 */
697 };
698 
699 static const u8 integrated_ctle_tunings[11][4] = {
700 	/* DC     LF     HF     BW */
701 	{  0x38,  0x0f,  0x00,  0x00 },	/* p0 */
702 	{  0x38,  0x0f,  0x00,  0x00 },	/* p1 */
703 	{  0x38,  0x0f,  0x00,  0x00 },	/* p2 */
704 	{  0x38,  0x0f,  0x00,  0x00 },	/* p3 */
705 	{  0x58,  0x0a,  0x05,  0x05 },	/* p4 */
706 	{  0x48,  0x0a,  0x05,  0x05 },	/* p5 */
707 	{  0x40,  0x0a,  0x05,  0x05 },	/* p6 */
708 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
709 	{  0x38,  0x0f,  0x00,  0x00 },	/* p8 */
710 	{  0x38,  0x09,  0x06,  0x06 },	/* p9 */
711 	{  0x38,  0x0e,  0x01,  0x01 },	/* p10 */
712 };
713 
714 /* helper to format the value to write to hardware */
715 #define eq_value(pre, curr, post) \
716 	((((u32)(pre)) << \
717 			PCIE_CFG_REG_PL102_GEN3_EQ_PRE_CURSOR_PSET_SHIFT) \
718 	| (((u32)(curr)) << PCIE_CFG_REG_PL102_GEN3_EQ_CURSOR_PSET_SHIFT) \
719 	| (((u32)(post)) << \
720 		PCIE_CFG_REG_PL102_GEN3_EQ_POST_CURSOR_PSET_SHIFT))
721 
722 /*
723  * Load the given EQ preset table into the PCIe hardware.
724  */
725 static int load_eq_table(struct hfi1_devdata *dd, const u8 eq[11][3], u8 fs,
726 			 u8 div)
727 {
728 	struct pci_dev *pdev = dd->pcidev;
729 	u32 hit_error = 0;
730 	u32 violation;
731 	u32 i;
732 	u8 c_minus1, c0, c_plus1;
733 	int ret;
734 
735 	for (i = 0; i < 11; i++) {
736 		/* set index */
737 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL103, i);
738 		/* write the value */
739 		c_minus1 = eq[i][PREC] / div;
740 		c0 = fs - (eq[i][PREC] / div) - (eq[i][POST] / div);
741 		c_plus1 = eq[i][POST] / div;
742 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL102,
743 				       eq_value(c_minus1, c0, c_plus1));
744 		/* check if these coefficients violate EQ rules */
745 		ret = pci_read_config_dword(dd->pcidev,
746 					    PCIE_CFG_REG_PL105, &violation);
747 		if (ret) {
748 			dd_dev_err(dd, "Unable to read from PCI config\n");
749 			hit_error = 1;
750 			break;
751 		}
752 
753 		if (violation
754 		    & PCIE_CFG_REG_PL105_GEN3_EQ_VIOLATE_COEF_RULES_SMASK){
755 			if (hit_error == 0) {
756 				dd_dev_err(dd,
757 					   "Gen3 EQ Table Coefficient rule violations\n");
758 				dd_dev_err(dd, "         prec   attn   post\n");
759 			}
760 			dd_dev_err(dd, "   p%02d:   %02x     %02x     %02x\n",
761 				   i, (u32)eq[i][0], (u32)eq[i][1],
762 				   (u32)eq[i][2]);
763 			dd_dev_err(dd, "            %02x     %02x     %02x\n",
764 				   (u32)c_minus1, (u32)c0, (u32)c_plus1);
765 			hit_error = 1;
766 		}
767 	}
768 	if (hit_error)
769 		return -EINVAL;
770 	return 0;
771 }
772 
773 /*
774  * Steps to be done after the PCIe firmware is downloaded and
775  * before the SBR for the Pcie Gen3.
776  * The SBus resource is already being held.
777  */
778 static void pcie_post_steps(struct hfi1_devdata *dd)
779 {
780 	int i;
781 
782 	set_sbus_fast_mode(dd);
783 	/*
784 	 * Write to the PCIe PCSes to set the G3_LOCKED_NEXT bits to 1.
785 	 * This avoids a spurious framing error that can otherwise be
786 	 * generated by the MAC layer.
787 	 *
788 	 * Use individual addresses since no broadcast is set up.
789 	 */
790 	for (i = 0; i < NUM_PCIE_SERDES; i++) {
791 		sbus_request(dd, pcie_pcs_addrs[dd->hfi1_id][i],
792 			     0x03, WRITE_SBUS_RECEIVER, 0x00022132);
793 	}
794 
795 	clear_sbus_fast_mode(dd);
796 }
797 
798 /*
799  * Trigger a secondary bus reset (SBR) on ourselves using our parent.
800  *
801  * Based on pci_parent_bus_reset() which is not exported by the
802  * kernel core.
803  */
804 static int trigger_sbr(struct hfi1_devdata *dd)
805 {
806 	struct pci_dev *dev = dd->pcidev;
807 	struct pci_dev *pdev;
808 
809 	/* need a parent */
810 	if (!dev->bus->self) {
811 		dd_dev_err(dd, "%s: no parent device\n", __func__);
812 		return -ENOTTY;
813 	}
814 
815 	/* should not be anyone else on the bus */
816 	list_for_each_entry(pdev, &dev->bus->devices, bus_list)
817 		if (pdev != dev) {
818 			dd_dev_err(dd,
819 				   "%s: another device is on the same bus\n",
820 				   __func__);
821 			return -ENOTTY;
822 		}
823 
824 	/*
825 	 * This is an end around to do an SBR during probe time. A new API needs
826 	 * to be implemented to have cleaner interface but this fixes the
827 	 * current brokenness
828 	 */
829 	return pci_bridge_secondary_bus_reset(dev->bus->self);
830 }
831 
832 /*
833  * Write the given gasket interrupt register.
834  */
835 static void write_gasket_interrupt(struct hfi1_devdata *dd, int index,
836 				   u16 code, u16 data)
837 {
838 	write_csr(dd, ASIC_PCIE_SD_INTRPT_LIST + (index * 8),
839 		  (((u64)code << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_CODE_SHIFT) |
840 		   ((u64)data << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_DATA_SHIFT)));
841 }
842 
843 /*
844  * Tell the gasket logic how to react to the reset.
845  */
846 static void arm_gasket_logic(struct hfi1_devdata *dd)
847 {
848 	u64 reg;
849 
850 	reg = (((u64)1 << dd->hfi1_id) <<
851 	       ASIC_PCIE_SD_HOST_CMD_INTRPT_CMD_SHIFT) |
852 	      ((u64)pcie_serdes_broadcast[dd->hfi1_id] <<
853 	       ASIC_PCIE_SD_HOST_CMD_SBUS_RCVR_ADDR_SHIFT |
854 	       ASIC_PCIE_SD_HOST_CMD_SBR_MODE_SMASK |
855 	       ((u64)SBR_DELAY_US & ASIC_PCIE_SD_HOST_CMD_TIMER_MASK) <<
856 	       ASIC_PCIE_SD_HOST_CMD_TIMER_SHIFT);
857 	write_csr(dd, ASIC_PCIE_SD_HOST_CMD, reg);
858 	/* read back to push the write */
859 	read_csr(dd, ASIC_PCIE_SD_HOST_CMD);
860 }
861 
862 /*
863  * CCE_PCIE_CTRL long name helpers
864  * We redefine these shorter macros to use in the code while leaving
865  * chip_registers.h to be autogenerated from the hardware spec.
866  */
867 #define LANE_BUNDLE_MASK              CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_MASK
868 #define LANE_BUNDLE_SHIFT             CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_SHIFT
869 #define LANE_DELAY_MASK               CCE_PCIE_CTRL_PCIE_LANE_DELAY_MASK
870 #define LANE_DELAY_SHIFT              CCE_PCIE_CTRL_PCIE_LANE_DELAY_SHIFT
871 #define MARGIN_OVERWRITE_ENABLE_SHIFT CCE_PCIE_CTRL_XMT_MARGIN_OVERWRITE_ENABLE_SHIFT
872 #define MARGIN_SHIFT                  CCE_PCIE_CTRL_XMT_MARGIN_SHIFT
873 #define MARGIN_G1_G2_OVERWRITE_MASK   CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_MASK
874 #define MARGIN_G1_G2_OVERWRITE_SHIFT  CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_SHIFT
875 #define MARGIN_GEN1_GEN2_MASK         CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_MASK
876 #define MARGIN_GEN1_GEN2_SHIFT        CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_SHIFT
877 
878  /*
879   * Write xmt_margin for full-swing (WFR-B) or half-swing (WFR-C).
880   */
881 static void write_xmt_margin(struct hfi1_devdata *dd, const char *fname)
882 {
883 	u64 pcie_ctrl;
884 	u64 xmt_margin;
885 	u64 xmt_margin_oe;
886 	u64 lane_delay;
887 	u64 lane_bundle;
888 
889 	pcie_ctrl = read_csr(dd, CCE_PCIE_CTRL);
890 
891 	/*
892 	 * For Discrete, use full-swing.
893 	 *  - PCIe TX defaults to full-swing.
894 	 *    Leave this register as default.
895 	 * For Integrated, use half-swing
896 	 *  - Copy xmt_margin and xmt_margin_oe
897 	 *    from Gen1/Gen2 to Gen3.
898 	 */
899 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL1) { /* integrated */
900 		/* extract initial fields */
901 		xmt_margin = (pcie_ctrl >> MARGIN_GEN1_GEN2_SHIFT)
902 			      & MARGIN_GEN1_GEN2_MASK;
903 		xmt_margin_oe = (pcie_ctrl >> MARGIN_G1_G2_OVERWRITE_SHIFT)
904 				 & MARGIN_G1_G2_OVERWRITE_MASK;
905 		lane_delay = (pcie_ctrl >> LANE_DELAY_SHIFT) & LANE_DELAY_MASK;
906 		lane_bundle = (pcie_ctrl >> LANE_BUNDLE_SHIFT)
907 			       & LANE_BUNDLE_MASK;
908 
909 		/*
910 		 * For A0, EFUSE values are not set.  Override with the
911 		 * correct values.
912 		 */
913 		if (is_ax(dd)) {
914 			/*
915 			 * xmt_margin and OverwiteEnabel should be the
916 			 * same for Gen1/Gen2 and Gen3
917 			 */
918 			xmt_margin = 0x5;
919 			xmt_margin_oe = 0x1;
920 			lane_delay = 0xF; /* Delay 240ns. */
921 			lane_bundle = 0x0; /* Set to 1 lane. */
922 		}
923 
924 		/* overwrite existing values */
925 		pcie_ctrl = (xmt_margin << MARGIN_GEN1_GEN2_SHIFT)
926 			| (xmt_margin_oe << MARGIN_G1_G2_OVERWRITE_SHIFT)
927 			| (xmt_margin << MARGIN_SHIFT)
928 			| (xmt_margin_oe << MARGIN_OVERWRITE_ENABLE_SHIFT)
929 			| (lane_delay << LANE_DELAY_SHIFT)
930 			| (lane_bundle << LANE_BUNDLE_SHIFT);
931 
932 		write_csr(dd, CCE_PCIE_CTRL, pcie_ctrl);
933 	}
934 
935 	dd_dev_dbg(dd, "%s: program XMT margin, CcePcieCtrl 0x%llx\n",
936 		   fname, pcie_ctrl);
937 }
938 
939 /*
940  * Do all the steps needed to transition the PCIe link to Gen3 speed.
941  */
942 int do_pcie_gen3_transition(struct hfi1_devdata *dd)
943 {
944 	struct pci_dev *parent = dd->pcidev->bus->self;
945 	u64 fw_ctrl;
946 	u64 reg, therm;
947 	u32 reg32, fs, lf;
948 	u32 status, err;
949 	int ret;
950 	int do_retry, retry_count = 0;
951 	int intnum = 0;
952 	uint default_pset;
953 	uint pset = pcie_pset;
954 	u16 target_vector, target_speed;
955 	u16 lnkctl2, vendor;
956 	u8 div;
957 	const u8 (*eq)[3];
958 	const u8 (*ctle_tunings)[4];
959 	uint static_ctle_mode;
960 	int return_error = 0;
961 	u32 target_width;
962 
963 	/* PCIe Gen3 is for the ASIC only */
964 	if (dd->icode != ICODE_RTL_SILICON)
965 		return 0;
966 
967 	if (pcie_target == 1) {			/* target Gen1 */
968 		target_vector = PCI_EXP_LNKCTL2_TLS_2_5GT;
969 		target_speed = 2500;
970 	} else if (pcie_target == 2) {		/* target Gen2 */
971 		target_vector = PCI_EXP_LNKCTL2_TLS_5_0GT;
972 		target_speed = 5000;
973 	} else if (pcie_target == 3) {		/* target Gen3 */
974 		target_vector = PCI_EXP_LNKCTL2_TLS_8_0GT;
975 		target_speed = 8000;
976 	} else {
977 		/* off or invalid target - skip */
978 		dd_dev_info(dd, "%s: Skipping PCIe transition\n", __func__);
979 		return 0;
980 	}
981 
982 	/* if already at target speed, done (unless forced) */
983 	if (dd->lbus_speed == target_speed) {
984 		dd_dev_info(dd, "%s: PCIe already at gen%d, %s\n", __func__,
985 			    pcie_target,
986 			    pcie_force ? "re-doing anyway" : "skipping");
987 		if (!pcie_force)
988 			return 0;
989 	}
990 
991 	/*
992 	 * The driver cannot do the transition if it has no access to the
993 	 * upstream component
994 	 */
995 	if (!parent) {
996 		dd_dev_info(dd, "%s: No upstream, Can't do gen3 transition\n",
997 			    __func__);
998 		return 0;
999 	}
1000 
1001 	/* Previous Gen1/Gen2 bus width */
1002 	target_width = dd->lbus_width;
1003 
1004 	/*
1005 	 * Do the Gen3 transition.  Steps are those of the PCIe Gen3
1006 	 * recipe.
1007 	 */
1008 
1009 	/* step 1: pcie link working in gen1/gen2 */
1010 
1011 	/* step 2: if either side is not capable of Gen3, done */
1012 	if (pcie_target == 3 && !dd->link_gen3_capable) {
1013 		dd_dev_err(dd, "The PCIe link is not Gen3 capable\n");
1014 		ret = -ENOSYS;
1015 		goto done_no_mutex;
1016 	}
1017 
1018 	/* hold the SBus resource across the firmware download and SBR */
1019 	ret = acquire_chip_resource(dd, CR_SBUS, SBUS_TIMEOUT);
1020 	if (ret) {
1021 		dd_dev_err(dd, "%s: unable to acquire SBus resource\n",
1022 			   __func__);
1023 		return ret;
1024 	}
1025 
1026 	/* make sure thermal polling is not causing interrupts */
1027 	therm = read_csr(dd, ASIC_CFG_THERM_POLL_EN);
1028 	if (therm) {
1029 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x0);
1030 		msleep(100);
1031 		dd_dev_info(dd, "%s: Disabled therm polling\n",
1032 			    __func__);
1033 	}
1034 
1035 retry:
1036 	/* the SBus download will reset the spico for thermal */
1037 
1038 	/* step 3: download SBus Master firmware */
1039 	/* step 4: download PCIe Gen3 SerDes firmware */
1040 	dd_dev_info(dd, "%s: downloading firmware\n", __func__);
1041 	ret = load_pcie_firmware(dd);
1042 	if (ret) {
1043 		/* do not proceed if the firmware cannot be downloaded */
1044 		return_error = 1;
1045 		goto done;
1046 	}
1047 
1048 	/* step 5: set up device parameter settings */
1049 	dd_dev_info(dd, "%s: setting PCIe registers\n", __func__);
1050 
1051 	/*
1052 	 * PcieCfgSpcie1 - Link Control 3
1053 	 * Leave at reset value.  No need to set PerfEq - link equalization
1054 	 * will be performed automatically after the SBR when the target
1055 	 * speed is 8GT/s.
1056 	 */
1057 
1058 	/* clear all 16 per-lane error bits (PCIe: Lane Error Status) */
1059 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, 0xffff);
1060 
1061 	/* step 5a: Set Synopsys Port Logic registers */
1062 
1063 	/*
1064 	 * PcieCfgRegPl2 - Port Force Link
1065 	 *
1066 	 * Set the low power field to 0x10 to avoid unnecessary power
1067 	 * management messages.  All other fields are zero.
1068 	 */
1069 	reg32 = 0x10ul << PCIE_CFG_REG_PL2_LOW_PWR_ENT_CNT_SHIFT;
1070 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL2, reg32);
1071 
1072 	/*
1073 	 * PcieCfgRegPl100 - Gen3 Control
1074 	 *
1075 	 * turn off PcieCfgRegPl100.Gen3ZRxDcNonCompl
1076 	 * turn on PcieCfgRegPl100.EqEieosCnt
1077 	 * Everything else zero.
1078 	 */
1079 	reg32 = PCIE_CFG_REG_PL100_EQ_EIEOS_CNT_SMASK;
1080 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL100, reg32);
1081 
1082 	/*
1083 	 * PcieCfgRegPl101 - Gen3 EQ FS and LF
1084 	 * PcieCfgRegPl102 - Gen3 EQ Presets to Coefficients Mapping
1085 	 * PcieCfgRegPl103 - Gen3 EQ Preset Index
1086 	 * PcieCfgRegPl105 - Gen3 EQ Status
1087 	 *
1088 	 * Give initial EQ settings.
1089 	 */
1090 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL0) { /* discrete */
1091 		/* 1000mV, FS=24, LF = 8 */
1092 		fs = 24;
1093 		lf = 8;
1094 		div = 3;
1095 		eq = discrete_preliminary_eq;
1096 		default_pset = DEFAULT_DISCRETE_PSET;
1097 		ctle_tunings = discrete_ctle_tunings;
1098 		/* bit 0 - discrete on/off */
1099 		static_ctle_mode = pcie_ctle & 0x1;
1100 	} else {
1101 		/* 400mV, FS=29, LF = 9 */
1102 		fs = 29;
1103 		lf = 9;
1104 		div = 1;
1105 		eq = integrated_preliminary_eq;
1106 		default_pset = DEFAULT_MCP_PSET;
1107 		ctle_tunings = integrated_ctle_tunings;
1108 		/* bit 1 - integrated on/off */
1109 		static_ctle_mode = (pcie_ctle >> 1) & 0x1;
1110 	}
1111 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL101,
1112 			       (fs <<
1113 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_FS_SHIFT) |
1114 			       (lf <<
1115 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_LF_SHIFT));
1116 	ret = load_eq_table(dd, eq, fs, div);
1117 	if (ret)
1118 		goto done;
1119 
1120 	/*
1121 	 * PcieCfgRegPl106 - Gen3 EQ Control
1122 	 *
1123 	 * Set Gen3EqPsetReqVec, leave other fields 0.
1124 	 */
1125 	if (pset == UNSET_PSET)
1126 		pset = default_pset;
1127 	if (pset > 10) {	/* valid range is 0-10, inclusive */
1128 		dd_dev_err(dd, "%s: Invalid Eq Pset %u, setting to %d\n",
1129 			   __func__, pset, default_pset);
1130 		pset = default_pset;
1131 	}
1132 	dd_dev_info(dd, "%s: using EQ Pset %u\n", __func__, pset);
1133 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL106,
1134 			       ((1 << pset) <<
1135 			PCIE_CFG_REG_PL106_GEN3_EQ_PSET_REQ_VEC_SHIFT) |
1136 			PCIE_CFG_REG_PL106_GEN3_EQ_EVAL2MS_DISABLE_SMASK |
1137 			PCIE_CFG_REG_PL106_GEN3_EQ_PHASE23_EXIT_MODE_SMASK);
1138 
1139 	/*
1140 	 * step 5b: Do post firmware download steps via SBus
1141 	 */
1142 	dd_dev_info(dd, "%s: doing pcie post steps\n", __func__);
1143 	pcie_post_steps(dd);
1144 
1145 	/*
1146 	 * step 5c: Program gasket interrupts
1147 	 */
1148 	/* set the Rx Bit Rate to REFCLK ratio */
1149 	write_gasket_interrupt(dd, intnum++, 0x0006, 0x0050);
1150 	/* disable pCal for PCIe Gen3 RX equalization */
1151 	/* select adaptive or static CTLE */
1152 	write_gasket_interrupt(dd, intnum++, 0x0026,
1153 			       0x5b01 | (static_ctle_mode << 3));
1154 	/*
1155 	 * Enable iCal for PCIe Gen3 RX equalization, and set which
1156 	 * evaluation of RX_EQ_EVAL will launch the iCal procedure.
1157 	 */
1158 	write_gasket_interrupt(dd, intnum++, 0x0026, 0x5202);
1159 
1160 	if (static_ctle_mode) {
1161 		/* apply static CTLE tunings */
1162 		u8 pcie_dc, pcie_lf, pcie_hf, pcie_bw;
1163 
1164 		pcie_dc = ctle_tunings[pset][0];
1165 		pcie_lf = ctle_tunings[pset][1];
1166 		pcie_hf = ctle_tunings[pset][2];
1167 		pcie_bw = ctle_tunings[pset][3];
1168 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0200 | pcie_dc);
1169 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0100 | pcie_lf);
1170 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0000 | pcie_hf);
1171 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x5500 | pcie_bw);
1172 	}
1173 
1174 	/* terminate list */
1175 	write_gasket_interrupt(dd, intnum++, 0x0000, 0x0000);
1176 
1177 	/*
1178 	 * step 5d: program XMT margin
1179 	 */
1180 	write_xmt_margin(dd, __func__);
1181 
1182 	/*
1183 	 * step 5e: disable active state power management (ASPM). It
1184 	 * will be enabled if required later
1185 	 */
1186 	dd_dev_info(dd, "%s: clearing ASPM\n", __func__);
1187 	aspm_hw_disable_l1(dd);
1188 
1189 	/*
1190 	 * step 5f: clear DirectSpeedChange
1191 	 * PcieCfgRegPl67.DirectSpeedChange must be zero to prevent the
1192 	 * change in the speed target from starting before we are ready.
1193 	 * This field defaults to 0 and we are not changing it, so nothing
1194 	 * needs to be done.
1195 	 */
1196 
1197 	/* step 5g: Set target link speed */
1198 	/*
1199 	 * Set target link speed to be target on both device and parent.
1200 	 * On setting the parent: Some system BIOSs "helpfully" set the
1201 	 * parent target speed to Gen2 to match the ASIC's initial speed.
1202 	 * We can set the target Gen3 because we have already checked
1203 	 * that it is Gen3 capable earlier.
1204 	 */
1205 	dd_dev_info(dd, "%s: setting parent target link speed\n", __func__);
1206 	ret = pcie_capability_read_word(parent, PCI_EXP_LNKCTL2, &lnkctl2);
1207 	if (ret) {
1208 		dd_dev_err(dd, "Unable to read from PCI config\n");
1209 		return_error = 1;
1210 		goto done;
1211 	}
1212 
1213 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1214 		    (u32)lnkctl2);
1215 	/* only write to parent if target is not as high as ours */
1216 	if ((lnkctl2 & PCI_EXP_LNKCTL2_TLS) < target_vector) {
1217 		lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1218 		lnkctl2 |= target_vector;
1219 		dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1220 			    (u32)lnkctl2);
1221 		ret = pcie_capability_write_word(parent,
1222 						 PCI_EXP_LNKCTL2, lnkctl2);
1223 		if (ret) {
1224 			dd_dev_err(dd, "Unable to write to PCI config\n");
1225 			return_error = 1;
1226 			goto done;
1227 		}
1228 	} else {
1229 		dd_dev_info(dd, "%s: ..target speed is OK\n", __func__);
1230 	}
1231 
1232 	dd_dev_info(dd, "%s: setting target link speed\n", __func__);
1233 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL2, &lnkctl2);
1234 	if (ret) {
1235 		dd_dev_err(dd, "Unable to read from PCI config\n");
1236 		return_error = 1;
1237 		goto done;
1238 	}
1239 
1240 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1241 		    (u32)lnkctl2);
1242 	lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1243 	lnkctl2 |= target_vector;
1244 	dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1245 		    (u32)lnkctl2);
1246 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL2, lnkctl2);
1247 	if (ret) {
1248 		dd_dev_err(dd, "Unable to write to PCI config\n");
1249 		return_error = 1;
1250 		goto done;
1251 	}
1252 
1253 	/* step 5h: arm gasket logic */
1254 	/* hold DC in reset across the SBR */
1255 	write_csr(dd, CCE_DC_CTRL, CCE_DC_CTRL_DC_RESET_SMASK);
1256 	(void)read_csr(dd, CCE_DC_CTRL); /* DC reset hold */
1257 	/* save firmware control across the SBR */
1258 	fw_ctrl = read_csr(dd, MISC_CFG_FW_CTRL);
1259 
1260 	dd_dev_info(dd, "%s: arming gasket logic\n", __func__);
1261 	arm_gasket_logic(dd);
1262 
1263 	/*
1264 	 * step 6: quiesce PCIe link
1265 	 * The chip has already been reset, so there will be no traffic
1266 	 * from the chip.  Linux has no easy way to enforce that it will
1267 	 * not try to access the device, so we just need to hope it doesn't
1268 	 * do it while we are doing the reset.
1269 	 */
1270 
1271 	/*
1272 	 * step 7: initiate the secondary bus reset (SBR)
1273 	 * step 8: hardware brings the links back up
1274 	 * step 9: wait for link speed transition to be complete
1275 	 */
1276 	dd_dev_info(dd, "%s: calling trigger_sbr\n", __func__);
1277 	ret = trigger_sbr(dd);
1278 	if (ret)
1279 		goto done;
1280 
1281 	/* step 10: decide what to do next */
1282 
1283 	/* check if we can read PCI space */
1284 	ret = pci_read_config_word(dd->pcidev, PCI_VENDOR_ID, &vendor);
1285 	if (ret) {
1286 		dd_dev_info(dd,
1287 			    "%s: read of VendorID failed after SBR, err %d\n",
1288 			    __func__, ret);
1289 		return_error = 1;
1290 		goto done;
1291 	}
1292 	if (vendor == 0xffff) {
1293 		dd_dev_info(dd, "%s: VendorID is all 1s after SBR\n", __func__);
1294 		return_error = 1;
1295 		ret = -EIO;
1296 		goto done;
1297 	}
1298 
1299 	/* restore PCI space registers we know were reset */
1300 	dd_dev_info(dd, "%s: calling restore_pci_variables\n", __func__);
1301 	ret = restore_pci_variables(dd);
1302 	if (ret) {
1303 		dd_dev_err(dd, "%s: Could not restore PCI variables\n",
1304 			   __func__);
1305 		return_error = 1;
1306 		goto done;
1307 	}
1308 
1309 	/* restore firmware control */
1310 	write_csr(dd, MISC_CFG_FW_CTRL, fw_ctrl);
1311 
1312 	/*
1313 	 * Check the gasket block status.
1314 	 *
1315 	 * This is the first CSR read after the SBR.  If the read returns
1316 	 * all 1s (fails), the link did not make it back.
1317 	 *
1318 	 * Once we're sure we can read and write, clear the DC reset after
1319 	 * the SBR.  Then check for any per-lane errors. Then look over
1320 	 * the status.
1321 	 */
1322 	reg = read_csr(dd, ASIC_PCIE_SD_HOST_STATUS);
1323 	dd_dev_info(dd, "%s: gasket block status: 0x%llx\n", __func__, reg);
1324 	if (reg == ~0ull) {	/* PCIe read failed/timeout */
1325 		dd_dev_err(dd, "SBR failed - unable to read from device\n");
1326 		return_error = 1;
1327 		ret = -ENOSYS;
1328 		goto done;
1329 	}
1330 
1331 	/* clear the DC reset */
1332 	write_csr(dd, CCE_DC_CTRL, 0);
1333 
1334 	/* Set the LED off */
1335 	setextled(dd, 0);
1336 
1337 	/* check for any per-lane errors */
1338 	ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, &reg32);
1339 	if (ret) {
1340 		dd_dev_err(dd, "Unable to read from PCI config\n");
1341 		return_error = 1;
1342 		goto done;
1343 	}
1344 
1345 	dd_dev_info(dd, "%s: per-lane errors: 0x%x\n", __func__, reg32);
1346 
1347 	/* extract status, look for our HFI */
1348 	status = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_SHIFT)
1349 			& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_MASK;
1350 	if ((status & (1 << dd->hfi1_id)) == 0) {
1351 		dd_dev_err(dd,
1352 			   "%s: gasket status 0x%x, expecting 0x%x\n",
1353 			   __func__, status, 1 << dd->hfi1_id);
1354 		ret = -EIO;
1355 		goto done;
1356 	}
1357 
1358 	/* extract error */
1359 	err = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_SHIFT)
1360 		& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_MASK;
1361 	if (err) {
1362 		dd_dev_err(dd, "%s: gasket error %d\n", __func__, err);
1363 		ret = -EIO;
1364 		goto done;
1365 	}
1366 
1367 	/* update our link information cache */
1368 	update_lbus_info(dd);
1369 	dd_dev_info(dd, "%s: new speed and width: %s\n", __func__,
1370 		    dd->lbus_info);
1371 
1372 	if (dd->lbus_speed != target_speed ||
1373 	    dd->lbus_width < target_width) { /* not target */
1374 		/* maybe retry */
1375 		do_retry = retry_count < pcie_retry;
1376 		dd_dev_err(dd, "PCIe link speed or width did not match target%s\n",
1377 			   do_retry ? ", retrying" : "");
1378 		retry_count++;
1379 		if (do_retry) {
1380 			msleep(100); /* allow time to settle */
1381 			goto retry;
1382 		}
1383 		ret = -EIO;
1384 	}
1385 
1386 done:
1387 	if (therm) {
1388 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x1);
1389 		msleep(100);
1390 		dd_dev_info(dd, "%s: Re-enable therm polling\n",
1391 			    __func__);
1392 	}
1393 	release_chip_resource(dd, CR_SBUS);
1394 done_no_mutex:
1395 	/* return no error if it is OK to be at current speed */
1396 	if (ret && !return_error) {
1397 		dd_dev_err(dd, "Proceeding at current speed PCIe speed\n");
1398 		ret = 0;
1399 	}
1400 
1401 	dd_dev_info(dd, "%s: done\n", __func__);
1402 	return ret;
1403 }
1404