xref: /openbmc/linux/drivers/infiniband/hw/hfi1/pcie.c (revision 3ddc8b84)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2019 Intel Corporation.
4  */
5 
6 #include <linux/bitfield.h>
7 #include <linux/pci.h>
8 #include <linux/io.h>
9 #include <linux/delay.h>
10 #include <linux/vmalloc.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 	return 0;
69 
70 bail:
71 	hfi1_pcie_cleanup(pdev);
72 	return ret;
73 }
74 
75 /*
76  * Clean what was done in hfi1_pcie_init()
77  */
78 void hfi1_pcie_cleanup(struct pci_dev *pdev)
79 {
80 	pci_disable_device(pdev);
81 	/*
82 	 * Release regions should be called after the disable. OK to
83 	 * call if request regions has not been called or failed.
84 	 */
85 	pci_release_regions(pdev);
86 }
87 
88 /*
89  * Do remaining PCIe setup, once dd is allocated, and save away
90  * fields required to re-initialize after a chip reset, or for
91  * various other purposes
92  */
93 int hfi1_pcie_ddinit(struct hfi1_devdata *dd, struct pci_dev *pdev)
94 {
95 	unsigned long len;
96 	resource_size_t addr;
97 	int ret = 0;
98 	u32 rcv_array_count;
99 
100 	addr = pci_resource_start(pdev, 0);
101 	len = pci_resource_len(pdev, 0);
102 
103 	/*
104 	 * The TXE PIO buffers are at the tail end of the chip space.
105 	 * Cut them off and map them separately.
106 	 */
107 
108 	/* sanity check vs expectations */
109 	if (len != TXE_PIO_SEND + TXE_PIO_SIZE) {
110 		dd_dev_err(dd, "chip PIO range does not match\n");
111 		return -EINVAL;
112 	}
113 
114 	dd->kregbase1 = ioremap(addr, RCV_ARRAY);
115 	if (!dd->kregbase1) {
116 		dd_dev_err(dd, "UC mapping of kregbase1 failed\n");
117 		return -ENOMEM;
118 	}
119 	dd_dev_info(dd, "UC base1: %p for %x\n", dd->kregbase1, RCV_ARRAY);
120 
121 	/* verify that reads actually work, save revision for reset check */
122 	dd->revision = readq(dd->kregbase1 + CCE_REVISION);
123 	if (dd->revision == ~(u64)0) {
124 		dd_dev_err(dd, "Cannot read chip CSRs\n");
125 		goto nomem;
126 	}
127 
128 	rcv_array_count = readq(dd->kregbase1 + RCV_ARRAY_CNT);
129 	dd_dev_info(dd, "RcvArray count: %u\n", rcv_array_count);
130 	dd->base2_start  = RCV_ARRAY + rcv_array_count * 8;
131 
132 	dd->kregbase2 = ioremap(
133 		addr + dd->base2_start,
134 		TXE_PIO_SEND - dd->base2_start);
135 	if (!dd->kregbase2) {
136 		dd_dev_err(dd, "UC mapping of kregbase2 failed\n");
137 		goto nomem;
138 	}
139 	dd_dev_info(dd, "UC base2: %p for %x\n", dd->kregbase2,
140 		    TXE_PIO_SEND - dd->base2_start);
141 
142 	dd->piobase = ioremap_wc(addr + TXE_PIO_SEND, TXE_PIO_SIZE);
143 	if (!dd->piobase) {
144 		dd_dev_err(dd, "WC mapping of send buffers failed\n");
145 		goto nomem;
146 	}
147 	dd_dev_info(dd, "WC piobase: %p for %x\n", dd->piobase, TXE_PIO_SIZE);
148 
149 	dd->physaddr = addr;        /* used for io_remap, etc. */
150 
151 	/*
152 	 * Map the chip's RcvArray as write-combining to allow us
153 	 * to write an entire cacheline worth of entries in one shot.
154 	 */
155 	dd->rcvarray_wc = ioremap_wc(addr + RCV_ARRAY,
156 				     rcv_array_count * 8);
157 	if (!dd->rcvarray_wc) {
158 		dd_dev_err(dd, "WC mapping of receive array failed\n");
159 		goto nomem;
160 	}
161 	dd_dev_info(dd, "WC RcvArray: %p for %x\n",
162 		    dd->rcvarray_wc, rcv_array_count * 8);
163 
164 	dd->flags |= HFI1_PRESENT;	/* chip.c CSR routines now work */
165 	return 0;
166 nomem:
167 	ret = -ENOMEM;
168 	hfi1_pcie_ddcleanup(dd);
169 	return ret;
170 }
171 
172 /*
173  * Do PCIe cleanup related to dd, after chip-specific cleanup, etc.  Just prior
174  * to releasing the dd memory.
175  * Void because all of the core pcie cleanup functions are void.
176  */
177 void hfi1_pcie_ddcleanup(struct hfi1_devdata *dd)
178 {
179 	dd->flags &= ~HFI1_PRESENT;
180 	if (dd->kregbase1)
181 		iounmap(dd->kregbase1);
182 	dd->kregbase1 = NULL;
183 	if (dd->kregbase2)
184 		iounmap(dd->kregbase2);
185 	dd->kregbase2 = NULL;
186 	if (dd->rcvarray_wc)
187 		iounmap(dd->rcvarray_wc);
188 	dd->rcvarray_wc = NULL;
189 	if (dd->piobase)
190 		iounmap(dd->piobase);
191 	dd->piobase = NULL;
192 }
193 
194 /* return the PCIe link speed from the given link status */
195 static u32 extract_speed(u16 linkstat)
196 {
197 	u32 speed;
198 
199 	switch (linkstat & PCI_EXP_LNKSTA_CLS) {
200 	default: /* not defined, assume Gen1 */
201 	case PCI_EXP_LNKSTA_CLS_2_5GB:
202 		speed = 2500; /* Gen 1, 2.5GHz */
203 		break;
204 	case PCI_EXP_LNKSTA_CLS_5_0GB:
205 		speed = 5000; /* Gen 2, 5GHz */
206 		break;
207 	case PCI_EXP_LNKSTA_CLS_8_0GB:
208 		speed = 8000; /* Gen 3, 8GHz */
209 		break;
210 	}
211 	return speed;
212 }
213 
214 /* read the link status and set dd->{lbus_width,lbus_speed,lbus_info} */
215 static void update_lbus_info(struct hfi1_devdata *dd)
216 {
217 	u16 linkstat;
218 	int ret;
219 
220 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
221 	if (ret) {
222 		dd_dev_err(dd, "Unable to read from PCI config\n");
223 		return;
224 	}
225 
226 	dd->lbus_width = FIELD_GET(PCI_EXP_LNKSTA_NLW, linkstat);
227 	dd->lbus_speed = extract_speed(linkstat);
228 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
229 		 "PCIe,%uMHz,x%u", dd->lbus_speed, dd->lbus_width);
230 }
231 
232 /*
233  * Read in the current PCIe link width and speed.  Find if the link is
234  * Gen3 capable.
235  */
236 int pcie_speeds(struct hfi1_devdata *dd)
237 {
238 	u32 linkcap;
239 	struct pci_dev *parent = dd->pcidev->bus->self;
240 	int ret;
241 
242 	if (!pci_is_pcie(dd->pcidev)) {
243 		dd_dev_err(dd, "Can't find PCI Express capability!\n");
244 		return -EINVAL;
245 	}
246 
247 	/* find if our max speed is Gen3 and parent supports Gen3 speeds */
248 	dd->link_gen3_capable = 1;
249 
250 	ret = pcie_capability_read_dword(dd->pcidev, PCI_EXP_LNKCAP, &linkcap);
251 	if (ret) {
252 		dd_dev_err(dd, "Unable to read from PCI config\n");
253 		return pcibios_err_to_errno(ret);
254 	}
255 
256 	if ((linkcap & PCI_EXP_LNKCAP_SLS) != PCI_EXP_LNKCAP_SLS_8_0GB) {
257 		dd_dev_info(dd,
258 			    "This HFI is not Gen3 capable, max speed 0x%x, need 0x3\n",
259 			    linkcap & PCI_EXP_LNKCAP_SLS);
260 		dd->link_gen3_capable = 0;
261 	}
262 
263 	/*
264 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
265 	 */
266 	if (parent &&
267 	    (dd->pcidev->bus->max_bus_speed == PCIE_SPEED_2_5GT ||
268 	     dd->pcidev->bus->max_bus_speed == PCIE_SPEED_5_0GT)) {
269 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
270 		dd->link_gen3_capable = 0;
271 	}
272 
273 	/* obtain the link width and current speed */
274 	update_lbus_info(dd);
275 
276 	dd_dev_info(dd, "%s\n", dd->lbus_info);
277 
278 	return 0;
279 }
280 
281 /*
282  * Restore command and BARs after a reset has wiped them out
283  *
284  * Returns 0 on success, otherwise a negative error value
285  */
286 int restore_pci_variables(struct hfi1_devdata *dd)
287 {
288 	int ret;
289 
290 	ret = pci_write_config_word(dd->pcidev, PCI_COMMAND, dd->pci_command);
291 	if (ret)
292 		goto error;
293 
294 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
295 				     dd->pcibar0);
296 	if (ret)
297 		goto error;
298 
299 	ret = pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
300 				     dd->pcibar1);
301 	if (ret)
302 		goto error;
303 
304 	ret = pci_write_config_dword(dd->pcidev, PCI_ROM_ADDRESS, dd->pci_rom);
305 	if (ret)
306 		goto error;
307 
308 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL,
309 					 dd->pcie_devctl);
310 	if (ret)
311 		goto error;
312 
313 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL,
314 					 dd->pcie_lnkctl);
315 	if (ret)
316 		goto error;
317 
318 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL2,
319 					 dd->pcie_devctl2);
320 	if (ret)
321 		goto error;
322 
323 	ret = pci_write_config_dword(dd->pcidev, PCI_CFG_MSIX0, dd->pci_msix0);
324 	if (ret)
325 		goto error;
326 
327 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
328 		ret = pci_write_config_dword(dd->pcidev, PCIE_CFG_TPH2,
329 					     dd->pci_tph2);
330 		if (ret)
331 			goto error;
332 	}
333 	return 0;
334 
335 error:
336 	dd_dev_err(dd, "Unable to write to PCI config\n");
337 	return pcibios_err_to_errno(ret);
338 }
339 
340 /*
341  * Save BARs and command to rewrite after device reset
342  *
343  * Returns 0 on success, otherwise a negative error value
344  */
345 int save_pci_variables(struct hfi1_devdata *dd)
346 {
347 	int ret;
348 
349 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
350 				    &dd->pcibar0);
351 	if (ret)
352 		goto error;
353 
354 	ret = pci_read_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
355 				    &dd->pcibar1);
356 	if (ret)
357 		goto error;
358 
359 	ret = pci_read_config_dword(dd->pcidev, PCI_ROM_ADDRESS, &dd->pci_rom);
360 	if (ret)
361 		goto error;
362 
363 	ret = pci_read_config_word(dd->pcidev, PCI_COMMAND, &dd->pci_command);
364 	if (ret)
365 		goto error;
366 
367 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL,
368 					&dd->pcie_devctl);
369 	if (ret)
370 		goto error;
371 
372 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL,
373 					&dd->pcie_lnkctl);
374 	if (ret)
375 		goto error;
376 
377 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL2,
378 					&dd->pcie_devctl2);
379 	if (ret)
380 		goto error;
381 
382 	ret = pci_read_config_dword(dd->pcidev, PCI_CFG_MSIX0, &dd->pci_msix0);
383 	if (ret)
384 		goto error;
385 
386 	if (pci_find_ext_capability(dd->pcidev, PCI_EXT_CAP_ID_TPH)) {
387 		ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_TPH2,
388 					    &dd->pci_tph2);
389 		if (ret)
390 			goto error;
391 	}
392 	return 0;
393 
394 error:
395 	dd_dev_err(dd, "Unable to read from PCI config\n");
396 	return pcibios_err_to_errno(ret);
397 }
398 
399 /*
400  * BIOS may not set PCIe bus-utilization parameters for best performance.
401  * Check and optionally adjust them to maximize our throughput.
402  */
403 static int hfi1_pcie_caps;
404 module_param_named(pcie_caps, hfi1_pcie_caps, int, 0444);
405 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
406 
407 /**
408  * tune_pcie_caps() - Code to adjust PCIe capabilities.
409  * @dd: Valid device data structure
410  *
411  */
412 void tune_pcie_caps(struct hfi1_devdata *dd)
413 {
414 	struct pci_dev *parent;
415 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
416 	u16 rc_mrrs, ep_mrrs, max_mrrs, ectl;
417 	int ret;
418 
419 	/*
420 	 * Turn on extended tags in DevCtl in case the BIOS has turned it off
421 	 * to improve WFR SDMA bandwidth
422 	 */
423 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &ectl);
424 	if ((!ret) && !(ectl & PCI_EXP_DEVCTL_EXT_TAG)) {
425 		dd_dev_info(dd, "Enabling PCIe extended tags\n");
426 		ectl |= PCI_EXP_DEVCTL_EXT_TAG;
427 		ret = pcie_capability_write_word(dd->pcidev,
428 						 PCI_EXP_DEVCTL, ectl);
429 		if (ret)
430 			dd_dev_info(dd, "Unable to write to PCI config\n");
431 	}
432 	/* Find out supported and configured values for parent (root) */
433 	parent = dd->pcidev->bus->self;
434 	/*
435 	 * The driver cannot perform the tuning if it does not have
436 	 * access to the upstream component.
437 	 */
438 	if (!parent) {
439 		dd_dev_info(dd, "Parent not found\n");
440 		return;
441 	}
442 	if (!pci_is_root_bus(parent->bus)) {
443 		dd_dev_info(dd, "Parent not root\n");
444 		return;
445 	}
446 	if (!pci_is_pcie(parent)) {
447 		dd_dev_info(dd, "Parent is not PCI Express capable\n");
448 		return;
449 	}
450 	if (!pci_is_pcie(dd->pcidev)) {
451 		dd_dev_info(dd, "PCI device is not PCI Express capable\n");
452 		return;
453 	}
454 	rc_mpss = parent->pcie_mpss;
455 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
456 	/* Find out supported and configured values for endpoint (us) */
457 	ep_mpss = dd->pcidev->pcie_mpss;
458 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
459 
460 	/* Find max payload supported by root, endpoint */
461 	if (rc_mpss > ep_mpss)
462 		rc_mpss = ep_mpss;
463 
464 	/* If Supported greater than limit in module param, limit it */
465 	if (rc_mpss > (hfi1_pcie_caps & 7))
466 		rc_mpss = hfi1_pcie_caps & 7;
467 	/* If less than (allowed, supported), bump root payload */
468 	if (rc_mpss > rc_mps) {
469 		rc_mps = rc_mpss;
470 		pcie_set_mps(parent, 128 << rc_mps);
471 	}
472 	/* If less than (allowed, supported), bump endpoint payload */
473 	if (rc_mpss > ep_mps) {
474 		ep_mps = rc_mpss;
475 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
476 	}
477 
478 	/*
479 	 * Now the Read Request size.
480 	 * No field for max supported, but PCIe spec limits it to 4096,
481 	 * which is code '5' (log2(4096) - 7)
482 	 */
483 	max_mrrs = 5;
484 	if (max_mrrs > ((hfi1_pcie_caps >> 4) & 7))
485 		max_mrrs = (hfi1_pcie_caps >> 4) & 7;
486 
487 	max_mrrs = 128 << max_mrrs;
488 	rc_mrrs = pcie_get_readrq(parent);
489 	ep_mrrs = pcie_get_readrq(dd->pcidev);
490 
491 	if (max_mrrs > rc_mrrs) {
492 		rc_mrrs = max_mrrs;
493 		pcie_set_readrq(parent, rc_mrrs);
494 	}
495 	if (max_mrrs > ep_mrrs) {
496 		ep_mrrs = max_mrrs;
497 		pcie_set_readrq(dd->pcidev, ep_mrrs);
498 	}
499 }
500 
501 /* End of PCIe capability tuning */
502 
503 /*
504  * From here through hfi1_pci_err_handler definition is invoked via
505  * PCI error infrastructure, registered via pci
506  */
507 static pci_ers_result_t
508 pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
509 {
510 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
511 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
512 
513 	switch (state) {
514 	case pci_channel_io_normal:
515 		dd_dev_info(dd, "State Normal, ignoring\n");
516 		break;
517 
518 	case pci_channel_io_frozen:
519 		dd_dev_info(dd, "State Frozen, requesting reset\n");
520 		pci_disable_device(pdev);
521 		ret = PCI_ERS_RESULT_NEED_RESET;
522 		break;
523 
524 	case pci_channel_io_perm_failure:
525 		if (dd) {
526 			dd_dev_info(dd, "State Permanent Failure, disabling\n");
527 			/* no more register accesses! */
528 			dd->flags &= ~HFI1_PRESENT;
529 			hfi1_disable_after_error(dd);
530 		}
531 		 /* else early, or other problem */
532 		ret =  PCI_ERS_RESULT_DISCONNECT;
533 		break;
534 
535 	default: /* shouldn't happen */
536 		dd_dev_info(dd, "HFI1 PCI errors detected (state %d)\n",
537 			    state);
538 		break;
539 	}
540 	return ret;
541 }
542 
543 static pci_ers_result_t
544 pci_mmio_enabled(struct pci_dev *pdev)
545 {
546 	u64 words = 0U;
547 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
548 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
549 
550 	if (dd && dd->pport) {
551 		words = read_port_cntr(dd->pport, C_RX_WORDS, CNTR_INVALID_VL);
552 		if (words == ~0ULL)
553 			ret = PCI_ERS_RESULT_NEED_RESET;
554 		dd_dev_info(dd,
555 			    "HFI1 mmio_enabled function called, read wordscntr %llx, returning %d\n",
556 			    words, ret);
557 	}
558 	return  ret;
559 }
560 
561 static pci_ers_result_t
562 pci_slot_reset(struct pci_dev *pdev)
563 {
564 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
565 
566 	dd_dev_info(dd, "HFI1 slot_reset function called, ignored\n");
567 	return PCI_ERS_RESULT_CAN_RECOVER;
568 }
569 
570 static void
571 pci_resume(struct pci_dev *pdev)
572 {
573 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
574 
575 	dd_dev_info(dd, "HFI1 resume function called\n");
576 	/*
577 	 * Running jobs will fail, since it's asynchronous
578 	 * unlike sysfs-requested reset.   Better than
579 	 * doing nothing.
580 	 */
581 	hfi1_init(dd, 1); /* same as re-init after reset */
582 }
583 
584 const struct pci_error_handlers hfi1_pci_err_handler = {
585 	.error_detected = pci_error_detected,
586 	.mmio_enabled = pci_mmio_enabled,
587 	.slot_reset = pci_slot_reset,
588 	.resume = pci_resume,
589 };
590 
591 /*============================================================================*/
592 /* PCIe Gen3 support */
593 
594 /*
595  * This code is separated out because it is expected to be removed in the
596  * final shipping product.  If not, then it will be revisited and items
597  * will be moved to more standard locations.
598  */
599 
600 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_STS field values */
601 #define DL_STATUS_HFI0 0x1	/* hfi0 firmware download complete */
602 #define DL_STATUS_HFI1 0x2	/* hfi1 firmware download complete */
603 #define DL_STATUS_BOTH 0x3	/* hfi0 and hfi1 firmware download complete */
604 
605 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_ERR field values */
606 #define DL_ERR_NONE		0x0	/* no error */
607 #define DL_ERR_SWAP_PARITY	0x1	/* parity error in SerDes interrupt */
608 					/*   or response data */
609 #define DL_ERR_DISABLED	0x2	/* hfi disabled */
610 #define DL_ERR_SECURITY	0x3	/* security check failed */
611 #define DL_ERR_SBUS		0x4	/* SBus status error */
612 #define DL_ERR_XFR_PARITY	0x5	/* parity error during ROM transfer*/
613 
614 /* gasket block secondary bus reset delay */
615 #define SBR_DELAY_US 200000	/* 200ms */
616 
617 static uint pcie_target = 3;
618 module_param(pcie_target, uint, S_IRUGO);
619 MODULE_PARM_DESC(pcie_target, "PCIe target speed (0 skip, 1-3 Gen1-3)");
620 
621 static uint pcie_force;
622 module_param(pcie_force, uint, S_IRUGO);
623 MODULE_PARM_DESC(pcie_force, "Force driver to do a PCIe firmware download even if already at target speed");
624 
625 static uint pcie_retry = 5;
626 module_param(pcie_retry, uint, S_IRUGO);
627 MODULE_PARM_DESC(pcie_retry, "Driver will try this many times to reach requested speed");
628 
629 #define UNSET_PSET 255
630 #define DEFAULT_DISCRETE_PSET 2	/* discrete HFI */
631 #define DEFAULT_MCP_PSET 6	/* MCP HFI */
632 static uint pcie_pset = UNSET_PSET;
633 module_param(pcie_pset, uint, S_IRUGO);
634 MODULE_PARM_DESC(pcie_pset, "PCIe Eq Pset value to use, range is 0-10");
635 
636 static uint pcie_ctle = 3; /* discrete on, integrated on */
637 module_param(pcie_ctle, uint, S_IRUGO);
638 MODULE_PARM_DESC(pcie_ctle, "PCIe static CTLE mode, bit 0 - discrete on/off, bit 1 - integrated on/off");
639 
640 /* equalization columns */
641 #define PREC 0
642 #define ATTN 1
643 #define POST 2
644 
645 /* discrete silicon preliminary equalization values */
646 static const u8 discrete_preliminary_eq[11][3] = {
647 	/* prec   attn   post */
648 	{  0x00,  0x00,  0x12 },	/* p0 */
649 	{  0x00,  0x00,  0x0c },	/* p1 */
650 	{  0x00,  0x00,  0x0f },	/* p2 */
651 	{  0x00,  0x00,  0x09 },	/* p3 */
652 	{  0x00,  0x00,  0x00 },	/* p4 */
653 	{  0x06,  0x00,  0x00 },	/* p5 */
654 	{  0x09,  0x00,  0x00 },	/* p6 */
655 	{  0x06,  0x00,  0x0f },	/* p7 */
656 	{  0x09,  0x00,  0x09 },	/* p8 */
657 	{  0x0c,  0x00,  0x00 },	/* p9 */
658 	{  0x00,  0x00,  0x18 },	/* p10 */
659 };
660 
661 /* integrated silicon preliminary equalization values */
662 static const u8 integrated_preliminary_eq[11][3] = {
663 	/* prec   attn   post */
664 	{  0x00,  0x1e,  0x07 },	/* p0 */
665 	{  0x00,  0x1e,  0x05 },	/* p1 */
666 	{  0x00,  0x1e,  0x06 },	/* p2 */
667 	{  0x00,  0x1e,  0x04 },	/* p3 */
668 	{  0x00,  0x1e,  0x00 },	/* p4 */
669 	{  0x03,  0x1e,  0x00 },	/* p5 */
670 	{  0x04,  0x1e,  0x00 },	/* p6 */
671 	{  0x03,  0x1e,  0x06 },	/* p7 */
672 	{  0x03,  0x1e,  0x04 },	/* p8 */
673 	{  0x05,  0x1e,  0x00 },	/* p9 */
674 	{  0x00,  0x1e,  0x0a },	/* p10 */
675 };
676 
677 static const u8 discrete_ctle_tunings[11][4] = {
678 	/* DC     LF     HF     BW */
679 	{  0x48,  0x0b,  0x04,  0x04 },	/* p0 */
680 	{  0x60,  0x05,  0x0f,  0x0a },	/* p1 */
681 	{  0x50,  0x09,  0x06,  0x06 },	/* p2 */
682 	{  0x68,  0x05,  0x0f,  0x0a },	/* p3 */
683 	{  0x80,  0x05,  0x0f,  0x0a },	/* p4 */
684 	{  0x70,  0x05,  0x0f,  0x0a },	/* p5 */
685 	{  0x68,  0x05,  0x0f,  0x0a },	/* p6 */
686 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
687 	{  0x48,  0x09,  0x06,  0x06 },	/* p8 */
688 	{  0x60,  0x05,  0x0f,  0x0a },	/* p9 */
689 	{  0x38,  0x0f,  0x00,  0x00 },	/* p10 */
690 };
691 
692 static const u8 integrated_ctle_tunings[11][4] = {
693 	/* DC     LF     HF     BW */
694 	{  0x38,  0x0f,  0x00,  0x00 },	/* p0 */
695 	{  0x38,  0x0f,  0x00,  0x00 },	/* p1 */
696 	{  0x38,  0x0f,  0x00,  0x00 },	/* p2 */
697 	{  0x38,  0x0f,  0x00,  0x00 },	/* p3 */
698 	{  0x58,  0x0a,  0x05,  0x05 },	/* p4 */
699 	{  0x48,  0x0a,  0x05,  0x05 },	/* p5 */
700 	{  0x40,  0x0a,  0x05,  0x05 },	/* p6 */
701 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
702 	{  0x38,  0x0f,  0x00,  0x00 },	/* p8 */
703 	{  0x38,  0x09,  0x06,  0x06 },	/* p9 */
704 	{  0x38,  0x0e,  0x01,  0x01 },	/* p10 */
705 };
706 
707 /* helper to format the value to write to hardware */
708 #define eq_value(pre, curr, post) \
709 	((((u32)(pre)) << \
710 			PCIE_CFG_REG_PL102_GEN3_EQ_PRE_CURSOR_PSET_SHIFT) \
711 	| (((u32)(curr)) << PCIE_CFG_REG_PL102_GEN3_EQ_CURSOR_PSET_SHIFT) \
712 	| (((u32)(post)) << \
713 		PCIE_CFG_REG_PL102_GEN3_EQ_POST_CURSOR_PSET_SHIFT))
714 
715 /*
716  * Load the given EQ preset table into the PCIe hardware.
717  */
718 static int load_eq_table(struct hfi1_devdata *dd, const u8 eq[11][3], u8 fs,
719 			 u8 div)
720 {
721 	struct pci_dev *pdev = dd->pcidev;
722 	u32 hit_error = 0;
723 	u32 violation;
724 	u32 i;
725 	u8 c_minus1, c0, c_plus1;
726 	int ret;
727 
728 	for (i = 0; i < 11; i++) {
729 		/* set index */
730 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL103, i);
731 		/* write the value */
732 		c_minus1 = eq[i][PREC] / div;
733 		c0 = fs - (eq[i][PREC] / div) - (eq[i][POST] / div);
734 		c_plus1 = eq[i][POST] / div;
735 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL102,
736 				       eq_value(c_minus1, c0, c_plus1));
737 		/* check if these coefficients violate EQ rules */
738 		ret = pci_read_config_dword(dd->pcidev,
739 					    PCIE_CFG_REG_PL105, &violation);
740 		if (ret) {
741 			dd_dev_err(dd, "Unable to read from PCI config\n");
742 			hit_error = 1;
743 			break;
744 		}
745 
746 		if (violation
747 		    & PCIE_CFG_REG_PL105_GEN3_EQ_VIOLATE_COEF_RULES_SMASK){
748 			if (hit_error == 0) {
749 				dd_dev_err(dd,
750 					   "Gen3 EQ Table Coefficient rule violations\n");
751 				dd_dev_err(dd, "         prec   attn   post\n");
752 			}
753 			dd_dev_err(dd, "   p%02d:   %02x     %02x     %02x\n",
754 				   i, (u32)eq[i][0], (u32)eq[i][1],
755 				   (u32)eq[i][2]);
756 			dd_dev_err(dd, "            %02x     %02x     %02x\n",
757 				   (u32)c_minus1, (u32)c0, (u32)c_plus1);
758 			hit_error = 1;
759 		}
760 	}
761 	if (hit_error)
762 		return -EINVAL;
763 	return 0;
764 }
765 
766 /*
767  * Steps to be done after the PCIe firmware is downloaded and
768  * before the SBR for the Pcie Gen3.
769  * The SBus resource is already being held.
770  */
771 static void pcie_post_steps(struct hfi1_devdata *dd)
772 {
773 	int i;
774 
775 	set_sbus_fast_mode(dd);
776 	/*
777 	 * Write to the PCIe PCSes to set the G3_LOCKED_NEXT bits to 1.
778 	 * This avoids a spurious framing error that can otherwise be
779 	 * generated by the MAC layer.
780 	 *
781 	 * Use individual addresses since no broadcast is set up.
782 	 */
783 	for (i = 0; i < NUM_PCIE_SERDES; i++) {
784 		sbus_request(dd, pcie_pcs_addrs[dd->hfi1_id][i],
785 			     0x03, WRITE_SBUS_RECEIVER, 0x00022132);
786 	}
787 
788 	clear_sbus_fast_mode(dd);
789 }
790 
791 /*
792  * Trigger a secondary bus reset (SBR) on ourselves using our parent.
793  *
794  * Based on pci_parent_bus_reset() which is not exported by the
795  * kernel core.
796  */
797 static int trigger_sbr(struct hfi1_devdata *dd)
798 {
799 	struct pci_dev *dev = dd->pcidev;
800 	struct pci_dev *pdev;
801 
802 	/* need a parent */
803 	if (!dev->bus->self) {
804 		dd_dev_err(dd, "%s: no parent device\n", __func__);
805 		return -ENOTTY;
806 	}
807 
808 	/* should not be anyone else on the bus */
809 	list_for_each_entry(pdev, &dev->bus->devices, bus_list)
810 		if (pdev != dev) {
811 			dd_dev_err(dd,
812 				   "%s: another device is on the same bus\n",
813 				   __func__);
814 			return -ENOTTY;
815 		}
816 
817 	/*
818 	 * This is an end around to do an SBR during probe time. A new API needs
819 	 * to be implemented to have cleaner interface but this fixes the
820 	 * current brokenness
821 	 */
822 	return pci_bridge_secondary_bus_reset(dev->bus->self);
823 }
824 
825 /*
826  * Write the given gasket interrupt register.
827  */
828 static void write_gasket_interrupt(struct hfi1_devdata *dd, int index,
829 				   u16 code, u16 data)
830 {
831 	write_csr(dd, ASIC_PCIE_SD_INTRPT_LIST + (index * 8),
832 		  (((u64)code << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_CODE_SHIFT) |
833 		   ((u64)data << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_DATA_SHIFT)));
834 }
835 
836 /*
837  * Tell the gasket logic how to react to the reset.
838  */
839 static void arm_gasket_logic(struct hfi1_devdata *dd)
840 {
841 	u64 reg;
842 
843 	reg = (((u64)1 << dd->hfi1_id) <<
844 	       ASIC_PCIE_SD_HOST_CMD_INTRPT_CMD_SHIFT) |
845 	      ((u64)pcie_serdes_broadcast[dd->hfi1_id] <<
846 	       ASIC_PCIE_SD_HOST_CMD_SBUS_RCVR_ADDR_SHIFT |
847 	       ASIC_PCIE_SD_HOST_CMD_SBR_MODE_SMASK |
848 	       ((u64)SBR_DELAY_US & ASIC_PCIE_SD_HOST_CMD_TIMER_MASK) <<
849 	       ASIC_PCIE_SD_HOST_CMD_TIMER_SHIFT);
850 	write_csr(dd, ASIC_PCIE_SD_HOST_CMD, reg);
851 	/* read back to push the write */
852 	read_csr(dd, ASIC_PCIE_SD_HOST_CMD);
853 }
854 
855 /*
856  * CCE_PCIE_CTRL long name helpers
857  * We redefine these shorter macros to use in the code while leaving
858  * chip_registers.h to be autogenerated from the hardware spec.
859  */
860 #define LANE_BUNDLE_MASK              CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_MASK
861 #define LANE_BUNDLE_SHIFT             CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_SHIFT
862 #define LANE_DELAY_MASK               CCE_PCIE_CTRL_PCIE_LANE_DELAY_MASK
863 #define LANE_DELAY_SHIFT              CCE_PCIE_CTRL_PCIE_LANE_DELAY_SHIFT
864 #define MARGIN_OVERWRITE_ENABLE_SHIFT CCE_PCIE_CTRL_XMT_MARGIN_OVERWRITE_ENABLE_SHIFT
865 #define MARGIN_SHIFT                  CCE_PCIE_CTRL_XMT_MARGIN_SHIFT
866 #define MARGIN_G1_G2_OVERWRITE_MASK   CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_MASK
867 #define MARGIN_G1_G2_OVERWRITE_SHIFT  CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_SHIFT
868 #define MARGIN_GEN1_GEN2_MASK         CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_MASK
869 #define MARGIN_GEN1_GEN2_SHIFT        CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_SHIFT
870 
871  /*
872   * Write xmt_margin for full-swing (WFR-B) or half-swing (WFR-C).
873   */
874 static void write_xmt_margin(struct hfi1_devdata *dd, const char *fname)
875 {
876 	u64 pcie_ctrl;
877 	u64 xmt_margin;
878 	u64 xmt_margin_oe;
879 	u64 lane_delay;
880 	u64 lane_bundle;
881 
882 	pcie_ctrl = read_csr(dd, CCE_PCIE_CTRL);
883 
884 	/*
885 	 * For Discrete, use full-swing.
886 	 *  - PCIe TX defaults to full-swing.
887 	 *    Leave this register as default.
888 	 * For Integrated, use half-swing
889 	 *  - Copy xmt_margin and xmt_margin_oe
890 	 *    from Gen1/Gen2 to Gen3.
891 	 */
892 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL1) { /* integrated */
893 		/* extract initial fields */
894 		xmt_margin = (pcie_ctrl >> MARGIN_GEN1_GEN2_SHIFT)
895 			      & MARGIN_GEN1_GEN2_MASK;
896 		xmt_margin_oe = (pcie_ctrl >> MARGIN_G1_G2_OVERWRITE_SHIFT)
897 				 & MARGIN_G1_G2_OVERWRITE_MASK;
898 		lane_delay = (pcie_ctrl >> LANE_DELAY_SHIFT) & LANE_DELAY_MASK;
899 		lane_bundle = (pcie_ctrl >> LANE_BUNDLE_SHIFT)
900 			       & LANE_BUNDLE_MASK;
901 
902 		/*
903 		 * For A0, EFUSE values are not set.  Override with the
904 		 * correct values.
905 		 */
906 		if (is_ax(dd)) {
907 			/*
908 			 * xmt_margin and OverwiteEnabel should be the
909 			 * same for Gen1/Gen2 and Gen3
910 			 */
911 			xmt_margin = 0x5;
912 			xmt_margin_oe = 0x1;
913 			lane_delay = 0xF; /* Delay 240ns. */
914 			lane_bundle = 0x0; /* Set to 1 lane. */
915 		}
916 
917 		/* overwrite existing values */
918 		pcie_ctrl = (xmt_margin << MARGIN_GEN1_GEN2_SHIFT)
919 			| (xmt_margin_oe << MARGIN_G1_G2_OVERWRITE_SHIFT)
920 			| (xmt_margin << MARGIN_SHIFT)
921 			| (xmt_margin_oe << MARGIN_OVERWRITE_ENABLE_SHIFT)
922 			| (lane_delay << LANE_DELAY_SHIFT)
923 			| (lane_bundle << LANE_BUNDLE_SHIFT);
924 
925 		write_csr(dd, CCE_PCIE_CTRL, pcie_ctrl);
926 	}
927 
928 	dd_dev_dbg(dd, "%s: program XMT margin, CcePcieCtrl 0x%llx\n",
929 		   fname, pcie_ctrl);
930 }
931 
932 /*
933  * Do all the steps needed to transition the PCIe link to Gen3 speed.
934  */
935 int do_pcie_gen3_transition(struct hfi1_devdata *dd)
936 {
937 	struct pci_dev *parent = dd->pcidev->bus->self;
938 	u64 fw_ctrl;
939 	u64 reg, therm;
940 	u32 reg32, fs, lf;
941 	u32 status, err;
942 	int ret;
943 	int do_retry, retry_count = 0;
944 	int intnum = 0;
945 	uint default_pset;
946 	uint pset = pcie_pset;
947 	u16 target_vector, target_speed;
948 	u16 lnkctl2, vendor;
949 	u8 div;
950 	const u8 (*eq)[3];
951 	const u8 (*ctle_tunings)[4];
952 	uint static_ctle_mode;
953 	int return_error = 0;
954 	u32 target_width;
955 
956 	/* PCIe Gen3 is for the ASIC only */
957 	if (dd->icode != ICODE_RTL_SILICON)
958 		return 0;
959 
960 	if (pcie_target == 1) {			/* target Gen1 */
961 		target_vector = PCI_EXP_LNKCTL2_TLS_2_5GT;
962 		target_speed = 2500;
963 	} else if (pcie_target == 2) {		/* target Gen2 */
964 		target_vector = PCI_EXP_LNKCTL2_TLS_5_0GT;
965 		target_speed = 5000;
966 	} else if (pcie_target == 3) {		/* target Gen3 */
967 		target_vector = PCI_EXP_LNKCTL2_TLS_8_0GT;
968 		target_speed = 8000;
969 	} else {
970 		/* off or invalid target - skip */
971 		dd_dev_info(dd, "%s: Skipping PCIe transition\n", __func__);
972 		return 0;
973 	}
974 
975 	/* if already at target speed, done (unless forced) */
976 	if (dd->lbus_speed == target_speed) {
977 		dd_dev_info(dd, "%s: PCIe already at gen%d, %s\n", __func__,
978 			    pcie_target,
979 			    pcie_force ? "re-doing anyway" : "skipping");
980 		if (!pcie_force)
981 			return 0;
982 	}
983 
984 	/*
985 	 * The driver cannot do the transition if it has no access to the
986 	 * upstream component
987 	 */
988 	if (!parent) {
989 		dd_dev_info(dd, "%s: No upstream, Can't do gen3 transition\n",
990 			    __func__);
991 		return 0;
992 	}
993 
994 	/* Previous Gen1/Gen2 bus width */
995 	target_width = dd->lbus_width;
996 
997 	/*
998 	 * Do the Gen3 transition.  Steps are those of the PCIe Gen3
999 	 * recipe.
1000 	 */
1001 
1002 	/* step 1: pcie link working in gen1/gen2 */
1003 
1004 	/* step 2: if either side is not capable of Gen3, done */
1005 	if (pcie_target == 3 && !dd->link_gen3_capable) {
1006 		dd_dev_err(dd, "The PCIe link is not Gen3 capable\n");
1007 		ret = -ENOSYS;
1008 		goto done_no_mutex;
1009 	}
1010 
1011 	/* hold the SBus resource across the firmware download and SBR */
1012 	ret = acquire_chip_resource(dd, CR_SBUS, SBUS_TIMEOUT);
1013 	if (ret) {
1014 		dd_dev_err(dd, "%s: unable to acquire SBus resource\n",
1015 			   __func__);
1016 		return ret;
1017 	}
1018 
1019 	/* make sure thermal polling is not causing interrupts */
1020 	therm = read_csr(dd, ASIC_CFG_THERM_POLL_EN);
1021 	if (therm) {
1022 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x0);
1023 		msleep(100);
1024 		dd_dev_info(dd, "%s: Disabled therm polling\n",
1025 			    __func__);
1026 	}
1027 
1028 retry:
1029 	/* the SBus download will reset the spico for thermal */
1030 
1031 	/* step 3: download SBus Master firmware */
1032 	/* step 4: download PCIe Gen3 SerDes firmware */
1033 	dd_dev_info(dd, "%s: downloading firmware\n", __func__);
1034 	ret = load_pcie_firmware(dd);
1035 	if (ret) {
1036 		/* do not proceed if the firmware cannot be downloaded */
1037 		return_error = 1;
1038 		goto done;
1039 	}
1040 
1041 	/* step 5: set up device parameter settings */
1042 	dd_dev_info(dd, "%s: setting PCIe registers\n", __func__);
1043 
1044 	/*
1045 	 * PcieCfgSpcie1 - Link Control 3
1046 	 * Leave at reset value.  No need to set PerfEq - link equalization
1047 	 * will be performed automatically after the SBR when the target
1048 	 * speed is 8GT/s.
1049 	 */
1050 
1051 	/* clear all 16 per-lane error bits (PCIe: Lane Error Status) */
1052 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, 0xffff);
1053 
1054 	/* step 5a: Set Synopsys Port Logic registers */
1055 
1056 	/*
1057 	 * PcieCfgRegPl2 - Port Force Link
1058 	 *
1059 	 * Set the low power field to 0x10 to avoid unnecessary power
1060 	 * management messages.  All other fields are zero.
1061 	 */
1062 	reg32 = 0x10ul << PCIE_CFG_REG_PL2_LOW_PWR_ENT_CNT_SHIFT;
1063 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL2, reg32);
1064 
1065 	/*
1066 	 * PcieCfgRegPl100 - Gen3 Control
1067 	 *
1068 	 * turn off PcieCfgRegPl100.Gen3ZRxDcNonCompl
1069 	 * turn on PcieCfgRegPl100.EqEieosCnt
1070 	 * Everything else zero.
1071 	 */
1072 	reg32 = PCIE_CFG_REG_PL100_EQ_EIEOS_CNT_SMASK;
1073 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL100, reg32);
1074 
1075 	/*
1076 	 * PcieCfgRegPl101 - Gen3 EQ FS and LF
1077 	 * PcieCfgRegPl102 - Gen3 EQ Presets to Coefficients Mapping
1078 	 * PcieCfgRegPl103 - Gen3 EQ Preset Index
1079 	 * PcieCfgRegPl105 - Gen3 EQ Status
1080 	 *
1081 	 * Give initial EQ settings.
1082 	 */
1083 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL0) { /* discrete */
1084 		/* 1000mV, FS=24, LF = 8 */
1085 		fs = 24;
1086 		lf = 8;
1087 		div = 3;
1088 		eq = discrete_preliminary_eq;
1089 		default_pset = DEFAULT_DISCRETE_PSET;
1090 		ctle_tunings = discrete_ctle_tunings;
1091 		/* bit 0 - discrete on/off */
1092 		static_ctle_mode = pcie_ctle & 0x1;
1093 	} else {
1094 		/* 400mV, FS=29, LF = 9 */
1095 		fs = 29;
1096 		lf = 9;
1097 		div = 1;
1098 		eq = integrated_preliminary_eq;
1099 		default_pset = DEFAULT_MCP_PSET;
1100 		ctle_tunings = integrated_ctle_tunings;
1101 		/* bit 1 - integrated on/off */
1102 		static_ctle_mode = (pcie_ctle >> 1) & 0x1;
1103 	}
1104 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL101,
1105 			       (fs <<
1106 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_FS_SHIFT) |
1107 			       (lf <<
1108 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_LF_SHIFT));
1109 	ret = load_eq_table(dd, eq, fs, div);
1110 	if (ret)
1111 		goto done;
1112 
1113 	/*
1114 	 * PcieCfgRegPl106 - Gen3 EQ Control
1115 	 *
1116 	 * Set Gen3EqPsetReqVec, leave other fields 0.
1117 	 */
1118 	if (pset == UNSET_PSET)
1119 		pset = default_pset;
1120 	if (pset > 10) {	/* valid range is 0-10, inclusive */
1121 		dd_dev_err(dd, "%s: Invalid Eq Pset %u, setting to %d\n",
1122 			   __func__, pset, default_pset);
1123 		pset = default_pset;
1124 	}
1125 	dd_dev_info(dd, "%s: using EQ Pset %u\n", __func__, pset);
1126 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL106,
1127 			       ((1 << pset) <<
1128 			PCIE_CFG_REG_PL106_GEN3_EQ_PSET_REQ_VEC_SHIFT) |
1129 			PCIE_CFG_REG_PL106_GEN3_EQ_EVAL2MS_DISABLE_SMASK |
1130 			PCIE_CFG_REG_PL106_GEN3_EQ_PHASE23_EXIT_MODE_SMASK);
1131 
1132 	/*
1133 	 * step 5b: Do post firmware download steps via SBus
1134 	 */
1135 	dd_dev_info(dd, "%s: doing pcie post steps\n", __func__);
1136 	pcie_post_steps(dd);
1137 
1138 	/*
1139 	 * step 5c: Program gasket interrupts
1140 	 */
1141 	/* set the Rx Bit Rate to REFCLK ratio */
1142 	write_gasket_interrupt(dd, intnum++, 0x0006, 0x0050);
1143 	/* disable pCal for PCIe Gen3 RX equalization */
1144 	/* select adaptive or static CTLE */
1145 	write_gasket_interrupt(dd, intnum++, 0x0026,
1146 			       0x5b01 | (static_ctle_mode << 3));
1147 	/*
1148 	 * Enable iCal for PCIe Gen3 RX equalization, and set which
1149 	 * evaluation of RX_EQ_EVAL will launch the iCal procedure.
1150 	 */
1151 	write_gasket_interrupt(dd, intnum++, 0x0026, 0x5202);
1152 
1153 	if (static_ctle_mode) {
1154 		/* apply static CTLE tunings */
1155 		u8 pcie_dc, pcie_lf, pcie_hf, pcie_bw;
1156 
1157 		pcie_dc = ctle_tunings[pset][0];
1158 		pcie_lf = ctle_tunings[pset][1];
1159 		pcie_hf = ctle_tunings[pset][2];
1160 		pcie_bw = ctle_tunings[pset][3];
1161 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0200 | pcie_dc);
1162 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0100 | pcie_lf);
1163 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0000 | pcie_hf);
1164 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x5500 | pcie_bw);
1165 	}
1166 
1167 	/* terminate list */
1168 	write_gasket_interrupt(dd, intnum++, 0x0000, 0x0000);
1169 
1170 	/*
1171 	 * step 5d: program XMT margin
1172 	 */
1173 	write_xmt_margin(dd, __func__);
1174 
1175 	/*
1176 	 * step 5e: disable active state power management (ASPM). It
1177 	 * will be enabled if required later
1178 	 */
1179 	dd_dev_info(dd, "%s: clearing ASPM\n", __func__);
1180 	aspm_hw_disable_l1(dd);
1181 
1182 	/*
1183 	 * step 5f: clear DirectSpeedChange
1184 	 * PcieCfgRegPl67.DirectSpeedChange must be zero to prevent the
1185 	 * change in the speed target from starting before we are ready.
1186 	 * This field defaults to 0 and we are not changing it, so nothing
1187 	 * needs to be done.
1188 	 */
1189 
1190 	/* step 5g: Set target link speed */
1191 	/*
1192 	 * Set target link speed to be target on both device and parent.
1193 	 * On setting the parent: Some system BIOSs "helpfully" set the
1194 	 * parent target speed to Gen2 to match the ASIC's initial speed.
1195 	 * We can set the target Gen3 because we have already checked
1196 	 * that it is Gen3 capable earlier.
1197 	 */
1198 	dd_dev_info(dd, "%s: setting parent target link speed\n", __func__);
1199 	ret = pcie_capability_read_word(parent, PCI_EXP_LNKCTL2, &lnkctl2);
1200 	if (ret) {
1201 		dd_dev_err(dd, "Unable to read from PCI config\n");
1202 		return_error = 1;
1203 		goto done;
1204 	}
1205 
1206 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1207 		    (u32)lnkctl2);
1208 	/* only write to parent if target is not as high as ours */
1209 	if ((lnkctl2 & PCI_EXP_LNKCTL2_TLS) < target_vector) {
1210 		lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1211 		lnkctl2 |= target_vector;
1212 		dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1213 			    (u32)lnkctl2);
1214 		ret = pcie_capability_write_word(parent,
1215 						 PCI_EXP_LNKCTL2, lnkctl2);
1216 		if (ret) {
1217 			dd_dev_err(dd, "Unable to write to PCI config\n");
1218 			return_error = 1;
1219 			goto done;
1220 		}
1221 	} else {
1222 		dd_dev_info(dd, "%s: ..target speed is OK\n", __func__);
1223 	}
1224 
1225 	dd_dev_info(dd, "%s: setting target link speed\n", __func__);
1226 	ret = pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL2, &lnkctl2);
1227 	if (ret) {
1228 		dd_dev_err(dd, "Unable to read from PCI config\n");
1229 		return_error = 1;
1230 		goto done;
1231 	}
1232 
1233 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1234 		    (u32)lnkctl2);
1235 	lnkctl2 &= ~PCI_EXP_LNKCTL2_TLS;
1236 	lnkctl2 |= target_vector;
1237 	dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1238 		    (u32)lnkctl2);
1239 	ret = pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL2, lnkctl2);
1240 	if (ret) {
1241 		dd_dev_err(dd, "Unable to write to PCI config\n");
1242 		return_error = 1;
1243 		goto done;
1244 	}
1245 
1246 	/* step 5h: arm gasket logic */
1247 	/* hold DC in reset across the SBR */
1248 	write_csr(dd, CCE_DC_CTRL, CCE_DC_CTRL_DC_RESET_SMASK);
1249 	(void)read_csr(dd, CCE_DC_CTRL); /* DC reset hold */
1250 	/* save firmware control across the SBR */
1251 	fw_ctrl = read_csr(dd, MISC_CFG_FW_CTRL);
1252 
1253 	dd_dev_info(dd, "%s: arming gasket logic\n", __func__);
1254 	arm_gasket_logic(dd);
1255 
1256 	/*
1257 	 * step 6: quiesce PCIe link
1258 	 * The chip has already been reset, so there will be no traffic
1259 	 * from the chip.  Linux has no easy way to enforce that it will
1260 	 * not try to access the device, so we just need to hope it doesn't
1261 	 * do it while we are doing the reset.
1262 	 */
1263 
1264 	/*
1265 	 * step 7: initiate the secondary bus reset (SBR)
1266 	 * step 8: hardware brings the links back up
1267 	 * step 9: wait for link speed transition to be complete
1268 	 */
1269 	dd_dev_info(dd, "%s: calling trigger_sbr\n", __func__);
1270 	ret = trigger_sbr(dd);
1271 	if (ret)
1272 		goto done;
1273 
1274 	/* step 10: decide what to do next */
1275 
1276 	/* check if we can read PCI space */
1277 	ret = pci_read_config_word(dd->pcidev, PCI_VENDOR_ID, &vendor);
1278 	if (ret) {
1279 		dd_dev_info(dd,
1280 			    "%s: read of VendorID failed after SBR, err %d\n",
1281 			    __func__, ret);
1282 		return_error = 1;
1283 		goto done;
1284 	}
1285 	if (vendor == 0xffff) {
1286 		dd_dev_info(dd, "%s: VendorID is all 1s after SBR\n", __func__);
1287 		return_error = 1;
1288 		ret = -EIO;
1289 		goto done;
1290 	}
1291 
1292 	/* restore PCI space registers we know were reset */
1293 	dd_dev_info(dd, "%s: calling restore_pci_variables\n", __func__);
1294 	ret = restore_pci_variables(dd);
1295 	if (ret) {
1296 		dd_dev_err(dd, "%s: Could not restore PCI variables\n",
1297 			   __func__);
1298 		return_error = 1;
1299 		goto done;
1300 	}
1301 
1302 	/* restore firmware control */
1303 	write_csr(dd, MISC_CFG_FW_CTRL, fw_ctrl);
1304 
1305 	/*
1306 	 * Check the gasket block status.
1307 	 *
1308 	 * This is the first CSR read after the SBR.  If the read returns
1309 	 * all 1s (fails), the link did not make it back.
1310 	 *
1311 	 * Once we're sure we can read and write, clear the DC reset after
1312 	 * the SBR.  Then check for any per-lane errors. Then look over
1313 	 * the status.
1314 	 */
1315 	reg = read_csr(dd, ASIC_PCIE_SD_HOST_STATUS);
1316 	dd_dev_info(dd, "%s: gasket block status: 0x%llx\n", __func__, reg);
1317 	if (reg == ~0ull) {	/* PCIe read failed/timeout */
1318 		dd_dev_err(dd, "SBR failed - unable to read from device\n");
1319 		return_error = 1;
1320 		ret = -ENOSYS;
1321 		goto done;
1322 	}
1323 
1324 	/* clear the DC reset */
1325 	write_csr(dd, CCE_DC_CTRL, 0);
1326 
1327 	/* Set the LED off */
1328 	setextled(dd, 0);
1329 
1330 	/* check for any per-lane errors */
1331 	ret = pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, &reg32);
1332 	if (ret) {
1333 		dd_dev_err(dd, "Unable to read from PCI config\n");
1334 		return_error = 1;
1335 		goto done;
1336 	}
1337 
1338 	dd_dev_info(dd, "%s: per-lane errors: 0x%x\n", __func__, reg32);
1339 
1340 	/* extract status, look for our HFI */
1341 	status = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_SHIFT)
1342 			& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_MASK;
1343 	if ((status & (1 << dd->hfi1_id)) == 0) {
1344 		dd_dev_err(dd,
1345 			   "%s: gasket status 0x%x, expecting 0x%x\n",
1346 			   __func__, status, 1 << dd->hfi1_id);
1347 		ret = -EIO;
1348 		goto done;
1349 	}
1350 
1351 	/* extract error */
1352 	err = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_SHIFT)
1353 		& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_MASK;
1354 	if (err) {
1355 		dd_dev_err(dd, "%s: gasket error %d\n", __func__, err);
1356 		ret = -EIO;
1357 		goto done;
1358 	}
1359 
1360 	/* update our link information cache */
1361 	update_lbus_info(dd);
1362 	dd_dev_info(dd, "%s: new speed and width: %s\n", __func__,
1363 		    dd->lbus_info);
1364 
1365 	if (dd->lbus_speed != target_speed ||
1366 	    dd->lbus_width < target_width) { /* not target */
1367 		/* maybe retry */
1368 		do_retry = retry_count < pcie_retry;
1369 		dd_dev_err(dd, "PCIe link speed or width did not match target%s\n",
1370 			   do_retry ? ", retrying" : "");
1371 		retry_count++;
1372 		if (do_retry) {
1373 			msleep(100); /* allow time to settle */
1374 			goto retry;
1375 		}
1376 		ret = -EIO;
1377 	}
1378 
1379 done:
1380 	if (therm) {
1381 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x1);
1382 		msleep(100);
1383 		dd_dev_info(dd, "%s: Re-enable therm polling\n",
1384 			    __func__);
1385 	}
1386 	release_chip_resource(dd, CR_SBUS);
1387 done_no_mutex:
1388 	/* return no error if it is OK to be at current speed */
1389 	if (ret && !return_error) {
1390 		dd_dev_err(dd, "Proceeding at current speed PCIe speed\n");
1391 		ret = 0;
1392 	}
1393 
1394 	dd_dev_info(dd, "%s: done\n", __func__);
1395 	return ret;
1396 }
1397