• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *
3  * This file is provided under a dual BSD/GPLv2 license.  When using or
4  * redistributing this file, you may do so under either license.
5  *
6  * GPL LICENSE SUMMARY
7  *
8  * Copyright(c) 2015 Intel Corporation.
9  *
10  * This program is free software; you can redistribute it and/or modify
11  * it under the terms of version 2 of the GNU General Public License as
12  * published by the Free Software Foundation.
13  *
14  * This program is distributed in the hope that it will be useful, but
15  * WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17  * General Public License for more details.
18  *
19  * BSD LICENSE
20  *
21  * Copyright(c) 2015 Intel Corporation.
22  *
23  * Redistribution and use in source and binary forms, with or without
24  * modification, are permitted provided that the following conditions
25  * are met:
26  *
27  *  - Redistributions of source code must retain the above copyright
28  *    notice, this list of conditions and the following disclaimer.
29  *  - Redistributions in binary form must reproduce the above copyright
30  *    notice, this list of conditions and the following disclaimer in
31  *    the documentation and/or other materials provided with the
32  *    distribution.
33  *  - Neither the name of Intel Corporation nor the names of its
34  *    contributors may be used to endorse or promote products derived
35  *    from this software without specific prior written permission.
36  *
37  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
38  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
39  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
40  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
41  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
42  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
43  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
44  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
45  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
46  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
47  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48  *
49  */
50 
51 #include <linux/pci.h>
52 #include <linux/io.h>
53 #include <linux/delay.h>
54 #include <linux/vmalloc.h>
55 #include <linux/aer.h>
56 #include <linux/module.h>
57 
58 #include "hfi.h"
59 #include "chip_registers.h"
60 
61 /* link speed vector for Gen3 speed - not in Linux headers */
62 #define GEN1_SPEED_VECTOR 0x1
63 #define GEN2_SPEED_VECTOR 0x2
64 #define GEN3_SPEED_VECTOR 0x3
65 
66 /*
67  * This file contains PCIe utility routines.
68  */
69 
70 /*
71  * Code to adjust PCIe capabilities.
72  */
73 static void tune_pcie_caps(struct hfi1_devdata *);
74 
75 /*
76  * Do all the common PCIe setup and initialization.
77  * devdata is not yet allocated, and is not allocated until after this
78  * routine returns success.  Therefore dd_dev_err() can't be used for error
79  * printing.
80  */
hfi1_pcie_init(struct pci_dev * pdev,const struct pci_device_id * ent)81 int hfi1_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
82 {
83 	int ret;
84 
85 	ret = pci_enable_device(pdev);
86 	if (ret) {
87 		/*
88 		 * This can happen (in theory) iff:
89 		 * We did a chip reset, and then failed to reprogram the
90 		 * BAR, or the chip reset due to an internal error.  We then
91 		 * unloaded the driver and reloaded it.
92 		 *
93 		 * Both reset cases set the BAR back to initial state.  For
94 		 * the latter case, the AER sticky error bit at offset 0x718
95 		 * should be set, but the Linux kernel doesn't yet know
96 		 * about that, it appears.  If the original BAR was retained
97 		 * in the kernel data structures, this may be OK.
98 		 */
99 		hfi1_early_err(&pdev->dev, "pci enable failed: error %d\n",
100 			       -ret);
101 		goto done;
102 	}
103 
104 	ret = pci_request_regions(pdev, DRIVER_NAME);
105 	if (ret) {
106 		hfi1_early_err(&pdev->dev,
107 			       "pci_request_regions fails: err %d\n", -ret);
108 		goto bail;
109 	}
110 
111 	ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
112 	if (ret) {
113 		/*
114 		 * If the 64 bit setup fails, try 32 bit.  Some systems
115 		 * do not setup 64 bit maps on systems with 2GB or less
116 		 * memory installed.
117 		 */
118 		ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
119 		if (ret) {
120 			hfi1_early_err(&pdev->dev,
121 				       "Unable to set DMA mask: %d\n", ret);
122 			goto bail;
123 		}
124 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
125 	} else
126 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
127 	if (ret) {
128 		hfi1_early_err(&pdev->dev,
129 			       "Unable to set DMA consistent mask: %d\n", ret);
130 		goto bail;
131 	}
132 
133 	pci_set_master(pdev);
134 	ret = pci_enable_pcie_error_reporting(pdev);
135 	if (ret) {
136 		hfi1_early_err(&pdev->dev,
137 			       "Unable to enable pcie error reporting: %d\n",
138 			      ret);
139 		ret = 0;
140 	}
141 	goto done;
142 
143 bail:
144 	hfi1_pcie_cleanup(pdev);
145 done:
146 	return ret;
147 }
148 
149 /*
150  * Clean what was done in hfi1_pcie_init()
151  */
hfi1_pcie_cleanup(struct pci_dev * pdev)152 void hfi1_pcie_cleanup(struct pci_dev *pdev)
153 {
154 	pci_disable_device(pdev);
155 	/*
156 	 * Release regions should be called after the disable. OK to
157 	 * call if request regions has not been called or failed.
158 	 */
159 	pci_release_regions(pdev);
160 }
161 
162 /*
163  * Do remaining PCIe setup, once dd is allocated, and save away
164  * fields required to re-initialize after a chip reset, or for
165  * various other purposes
166  */
hfi1_pcie_ddinit(struct hfi1_devdata * dd,struct pci_dev * pdev,const struct pci_device_id * ent)167 int hfi1_pcie_ddinit(struct hfi1_devdata *dd, struct pci_dev *pdev,
168 		     const struct pci_device_id *ent)
169 {
170 	unsigned long len;
171 	resource_size_t addr;
172 
173 	dd->pcidev = pdev;
174 	pci_set_drvdata(pdev, dd);
175 
176 	addr = pci_resource_start(pdev, 0);
177 	len = pci_resource_len(pdev, 0);
178 
179 	/*
180 	 * The TXE PIO buffers are at the tail end of the chip space.
181 	 * Cut them off and map them separately.
182 	 */
183 
184 	/* sanity check vs expectations */
185 	if (len != TXE_PIO_SEND + TXE_PIO_SIZE) {
186 		dd_dev_err(dd, "chip PIO range does not match\n");
187 		return -EINVAL;
188 	}
189 
190 	dd->kregbase = ioremap_nocache(addr, TXE_PIO_SEND);
191 	if (!dd->kregbase)
192 		return -ENOMEM;
193 
194 	dd->piobase = ioremap_wc(addr + TXE_PIO_SEND, TXE_PIO_SIZE);
195 	if (!dd->piobase) {
196 		iounmap(dd->kregbase);
197 		return -ENOMEM;
198 	}
199 
200 	dd->flags |= HFI1_PRESENT;	/* now register routines work */
201 
202 	dd->kregend = dd->kregbase + TXE_PIO_SEND;
203 	dd->physaddr = addr;        /* used for io_remap, etc. */
204 
205 	/*
206 	 * Re-map the chip's RcvArray as write-combining to allow us
207 	 * to write an entire cacheline worth of entries in one shot.
208 	 * If this re-map fails, just continue - the RcvArray programming
209 	 * function will handle both cases.
210 	 */
211 	dd->chip_rcv_array_count = read_csr(dd, RCV_ARRAY_CNT);
212 	dd->rcvarray_wc = ioremap_wc(addr + RCV_ARRAY,
213 				     dd->chip_rcv_array_count * 8);
214 	dd_dev_info(dd, "WC Remapped RcvArray: %p\n", dd->rcvarray_wc);
215 	/*
216 	 * Save BARs and command to rewrite after device reset.
217 	 */
218 	dd->pcibar0 = addr;
219 	dd->pcibar1 = addr >> 32;
220 	pci_read_config_dword(dd->pcidev, PCI_ROM_ADDRESS, &dd->pci_rom);
221 	pci_read_config_word(dd->pcidev, PCI_COMMAND, &dd->pci_command);
222 	pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &dd->pcie_devctl);
223 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL, &dd->pcie_lnkctl);
224 	pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL2,
225 							&dd->pcie_devctl2);
226 	pci_read_config_dword(dd->pcidev, PCI_CFG_MSIX0, &dd->pci_msix0);
227 	pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE1,
228 							&dd->pci_lnkctl3);
229 	pci_read_config_dword(dd->pcidev, PCIE_CFG_TPH2, &dd->pci_tph2);
230 
231 	return 0;
232 }
233 
234 /*
235  * Do PCIe cleanup related to dd, after chip-specific cleanup, etc.  Just prior
236  * to releasing the dd memory.
237  * Void because all of the core pcie cleanup functions are void.
238  */
hfi1_pcie_ddcleanup(struct hfi1_devdata * dd)239 void hfi1_pcie_ddcleanup(struct hfi1_devdata *dd)
240 {
241 	u64 __iomem *base = (void __iomem *) dd->kregbase;
242 
243 	dd->flags &= ~HFI1_PRESENT;
244 	dd->kregbase = NULL;
245 	iounmap(base);
246 	if (dd->rcvarray_wc)
247 		iounmap(dd->rcvarray_wc);
248 	if (dd->piobase)
249 		iounmap(dd->piobase);
250 
251 	pci_set_drvdata(dd->pcidev, NULL);
252 }
253 
254 /*
255  * Do a Function Level Reset (FLR) on the device.
256  * Based on static function drivers/pci/pci.c:pcie_flr().
257  */
hfi1_pcie_flr(struct hfi1_devdata * dd)258 void hfi1_pcie_flr(struct hfi1_devdata *dd)
259 {
260 	int i;
261 	u16 status;
262 
263 	/* no need to check for the capability - we know the device has it */
264 
265 	/* wait for Transaction Pending bit to clear, at most a few ms */
266 	for (i = 0; i < 4; i++) {
267 		if (i)
268 			msleep((1 << (i - 1)) * 100);
269 
270 		pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVSTA, &status);
271 		if (!(status & PCI_EXP_DEVSTA_TRPND))
272 			goto clear;
273 	}
274 
275 	dd_dev_err(dd, "Transaction Pending bit is not clearing, proceeding with reset anyway\n");
276 
277 clear:
278 	pcie_capability_set_word(dd->pcidev, PCI_EXP_DEVCTL,
279 						PCI_EXP_DEVCTL_BCR_FLR);
280 	/* PCIe spec requires the function to be back within 100ms */
281 	msleep(100);
282 }
283 
msix_setup(struct hfi1_devdata * dd,int pos,u32 * msixcnt,struct hfi1_msix_entry * hfi1_msix_entry)284 static void msix_setup(struct hfi1_devdata *dd, int pos, u32 *msixcnt,
285 		       struct hfi1_msix_entry *hfi1_msix_entry)
286 {
287 	int ret;
288 	int nvec = *msixcnt;
289 	struct msix_entry *msix_entry;
290 	int i;
291 
292 	/* We can't pass hfi1_msix_entry array to msix_setup
293 	 * so use a dummy msix_entry array and copy the allocated
294 	 * irq back to the hfi1_msix_entry array. */
295 	msix_entry = kmalloc_array(nvec, sizeof(*msix_entry), GFP_KERNEL);
296 	if (!msix_entry) {
297 		ret = -ENOMEM;
298 		goto do_intx;
299 	}
300 
301 	for (i = 0; i < nvec; i++)
302 		msix_entry[i] = hfi1_msix_entry[i].msix;
303 
304 	ret = pci_enable_msix_range(dd->pcidev, msix_entry, 1, nvec);
305 	if (ret < 0)
306 		goto free_msix_entry;
307 	nvec = ret;
308 
309 	for (i = 0; i < nvec; i++)
310 		hfi1_msix_entry[i].msix = msix_entry[i];
311 
312 	kfree(msix_entry);
313 	*msixcnt = nvec;
314 	return;
315 
316 free_msix_entry:
317 	kfree(msix_entry);
318 
319 do_intx:
320 	dd_dev_err(dd, "pci_enable_msix_range %d vectors failed: %d, falling back to INTx\n",
321 		   nvec, ret);
322 	*msixcnt = 0;
323 	hfi1_enable_intx(dd->pcidev);
324 
325 }
326 
327 /* return the PCIe link speed from the given link status */
extract_speed(u16 linkstat)328 static u32 extract_speed(u16 linkstat)
329 {
330 	u32 speed;
331 
332 	switch (linkstat & PCI_EXP_LNKSTA_CLS) {
333 	default: /* not defined, assume Gen1 */
334 	case PCI_EXP_LNKSTA_CLS_2_5GB:
335 		speed = 2500; /* Gen 1, 2.5GHz */
336 		break;
337 	case PCI_EXP_LNKSTA_CLS_5_0GB:
338 		speed = 5000; /* Gen 2, 5GHz */
339 		break;
340 	case GEN3_SPEED_VECTOR:
341 		speed = 8000; /* Gen 3, 8GHz */
342 		break;
343 	}
344 	return speed;
345 }
346 
347 /* return the PCIe link speed from the given link status */
extract_width(u16 linkstat)348 static u32 extract_width(u16 linkstat)
349 {
350 	return (linkstat & PCI_EXP_LNKSTA_NLW) >> PCI_EXP_LNKSTA_NLW_SHIFT;
351 }
352 
353 /* read the link status and set dd->{lbus_width,lbus_speed,lbus_info} */
update_lbus_info(struct hfi1_devdata * dd)354 static void update_lbus_info(struct hfi1_devdata *dd)
355 {
356 	u16 linkstat;
357 
358 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
359 	dd->lbus_width = extract_width(linkstat);
360 	dd->lbus_speed = extract_speed(linkstat);
361 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
362 		 "PCIe,%uMHz,x%u", dd->lbus_speed, dd->lbus_width);
363 }
364 
365 /*
366  * Read in the current PCIe link width and speed.  Find if the link is
367  * Gen3 capable.
368  */
pcie_speeds(struct hfi1_devdata * dd)369 int pcie_speeds(struct hfi1_devdata *dd)
370 {
371 	u32 linkcap;
372 
373 	if (!pci_is_pcie(dd->pcidev)) {
374 		dd_dev_err(dd, "Can't find PCI Express capability!\n");
375 		return -EINVAL;
376 	}
377 
378 	/* find if our max speed is Gen3 and parent supports Gen3 speeds */
379 	dd->link_gen3_capable = 1;
380 
381 	pcie_capability_read_dword(dd->pcidev, PCI_EXP_LNKCAP, &linkcap);
382 	if ((linkcap & PCI_EXP_LNKCAP_SLS) != GEN3_SPEED_VECTOR) {
383 		dd_dev_info(dd,
384 			"This HFI is not Gen3 capable, max speed 0x%x, need 0x3\n",
385 			linkcap & PCI_EXP_LNKCAP_SLS);
386 		dd->link_gen3_capable = 0;
387 	}
388 
389 	/*
390 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
391 	 */
392 	if (dd->pcidev->bus->max_bus_speed == PCIE_SPEED_2_5GT ||
393 	     dd->pcidev->bus->max_bus_speed == PCIE_SPEED_5_0GT) {
394 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
395 		dd->link_gen3_capable = 0;
396 	}
397 
398 	/* obtain the link width and current speed */
399 	update_lbus_info(dd);
400 
401 	/* check against expected pcie width and complain if "wrong" */
402 	if (dd->lbus_width < 16)
403 		dd_dev_err(dd, "PCIe width %u (x16 HFI)\n", dd->lbus_width);
404 
405 	return 0;
406 }
407 
408 /*
409  * Returns in *nent:
410  *	- actual number of interrupts allocated
411  *	- 0 if fell back to INTx.
412  */
request_msix(struct hfi1_devdata * dd,u32 * nent,struct hfi1_msix_entry * entry)413 void request_msix(struct hfi1_devdata *dd, u32 *nent,
414 		  struct hfi1_msix_entry *entry)
415 {
416 	int pos;
417 
418 	pos = dd->pcidev->msix_cap;
419 	if (*nent && pos) {
420 		msix_setup(dd, pos, nent, entry);
421 		/* did it, either MSI-X or INTx */
422 	} else {
423 		*nent = 0;
424 		hfi1_enable_intx(dd->pcidev);
425 	}
426 
427 	tune_pcie_caps(dd);
428 }
429 
430 /*
431  * Disable MSI-X.
432  */
hfi1_nomsix(struct hfi1_devdata * dd)433 void hfi1_nomsix(struct hfi1_devdata *dd)
434 {
435 	pci_disable_msix(dd->pcidev);
436 }
437 
hfi1_enable_intx(struct pci_dev * pdev)438 void hfi1_enable_intx(struct pci_dev *pdev)
439 {
440 	/* first, turn on INTx */
441 	pci_intx(pdev, 1);
442 	/* then turn off MSI-X */
443 	pci_disable_msix(pdev);
444 }
445 
446 /* restore command and BARs after a reset has wiped them out */
restore_pci_variables(struct hfi1_devdata * dd)447 void restore_pci_variables(struct hfi1_devdata *dd)
448 {
449 	pci_write_config_word(dd->pcidev, PCI_COMMAND, dd->pci_command);
450 	pci_write_config_dword(dd->pcidev,
451 				PCI_BASE_ADDRESS_0, dd->pcibar0);
452 	pci_write_config_dword(dd->pcidev,
453 				PCI_BASE_ADDRESS_1, dd->pcibar1);
454 	pci_write_config_dword(dd->pcidev,
455 				PCI_ROM_ADDRESS, dd->pci_rom);
456 	pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL, dd->pcie_devctl);
457 	pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL, dd->pcie_lnkctl);
458 	pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL2,
459 							dd->pcie_devctl2);
460 	pci_write_config_dword(dd->pcidev, PCI_CFG_MSIX0, dd->pci_msix0);
461 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE1,
462 							dd->pci_lnkctl3);
463 	pci_write_config_dword(dd->pcidev, PCIE_CFG_TPH2, dd->pci_tph2);
464 }
465 
466 
467 /*
468  * BIOS may not set PCIe bus-utilization parameters for best performance.
469  * Check and optionally adjust them to maximize our throughput.
470  */
471 static int hfi1_pcie_caps;
472 module_param_named(pcie_caps, hfi1_pcie_caps, int, S_IRUGO);
473 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
474 
tune_pcie_caps(struct hfi1_devdata * dd)475 static void tune_pcie_caps(struct hfi1_devdata *dd)
476 {
477 	struct pci_dev *parent;
478 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
479 	u16 rc_mrrs, ep_mrrs, max_mrrs;
480 
481 	/* Find out supported and configured values for parent (root) */
482 	parent = dd->pcidev->bus->self;
483 	if (!pci_is_root_bus(parent->bus)) {
484 		dd_dev_info(dd, "Parent not root\n");
485 		return;
486 	}
487 
488 	if (!pci_is_pcie(parent) || !pci_is_pcie(dd->pcidev))
489 		return;
490 	rc_mpss = parent->pcie_mpss;
491 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
492 	/* Find out supported and configured values for endpoint (us) */
493 	ep_mpss = dd->pcidev->pcie_mpss;
494 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
495 
496 	/* Find max payload supported by root, endpoint */
497 	if (rc_mpss > ep_mpss)
498 		rc_mpss = ep_mpss;
499 
500 	/* If Supported greater than limit in module param, limit it */
501 	if (rc_mpss > (hfi1_pcie_caps & 7))
502 		rc_mpss = hfi1_pcie_caps & 7;
503 	/* If less than (allowed, supported), bump root payload */
504 	if (rc_mpss > rc_mps) {
505 		rc_mps = rc_mpss;
506 		pcie_set_mps(parent, 128 << rc_mps);
507 	}
508 	/* If less than (allowed, supported), bump endpoint payload */
509 	if (rc_mpss > ep_mps) {
510 		ep_mps = rc_mpss;
511 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
512 	}
513 
514 	/*
515 	 * Now the Read Request size.
516 	 * No field for max supported, but PCIe spec limits it to 4096,
517 	 * which is code '5' (log2(4096) - 7)
518 	 */
519 	max_mrrs = 5;
520 	if (max_mrrs > ((hfi1_pcie_caps >> 4) & 7))
521 		max_mrrs = (hfi1_pcie_caps >> 4) & 7;
522 
523 	max_mrrs = 128 << max_mrrs;
524 	rc_mrrs = pcie_get_readrq(parent);
525 	ep_mrrs = pcie_get_readrq(dd->pcidev);
526 
527 	if (max_mrrs > rc_mrrs) {
528 		rc_mrrs = max_mrrs;
529 		pcie_set_readrq(parent, rc_mrrs);
530 	}
531 	if (max_mrrs > ep_mrrs) {
532 		ep_mrrs = max_mrrs;
533 		pcie_set_readrq(dd->pcidev, ep_mrrs);
534 	}
535 }
536 /* End of PCIe capability tuning */
537 
538 /*
539  * From here through hfi1_pci_err_handler definition is invoked via
540  * PCI error infrastructure, registered via pci
541  */
542 static pci_ers_result_t
pci_error_detected(struct pci_dev * pdev,pci_channel_state_t state)543 pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
544 {
545 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
546 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
547 
548 	switch (state) {
549 	case pci_channel_io_normal:
550 		dd_dev_info(dd, "State Normal, ignoring\n");
551 		break;
552 
553 	case pci_channel_io_frozen:
554 		dd_dev_info(dd, "State Frozen, requesting reset\n");
555 		pci_disable_device(pdev);
556 		ret = PCI_ERS_RESULT_NEED_RESET;
557 		break;
558 
559 	case pci_channel_io_perm_failure:
560 		if (dd) {
561 			dd_dev_info(dd, "State Permanent Failure, disabling\n");
562 			/* no more register accesses! */
563 			dd->flags &= ~HFI1_PRESENT;
564 			hfi1_disable_after_error(dd);
565 		}
566 		 /* else early, or other problem */
567 		ret =  PCI_ERS_RESULT_DISCONNECT;
568 		break;
569 
570 	default: /* shouldn't happen */
571 		dd_dev_info(dd, "HFI1 PCI errors detected (state %d)\n",
572 			    state);
573 		break;
574 	}
575 	return ret;
576 }
577 
578 static pci_ers_result_t
pci_mmio_enabled(struct pci_dev * pdev)579 pci_mmio_enabled(struct pci_dev *pdev)
580 {
581 	u64 words = 0U;
582 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
583 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
584 
585 	if (dd && dd->pport) {
586 		words = read_port_cntr(dd->pport, C_RX_WORDS, CNTR_INVALID_VL);
587 		if (words == ~0ULL)
588 			ret = PCI_ERS_RESULT_NEED_RESET;
589 		dd_dev_info(dd,
590 			    "HFI1 mmio_enabled function called, read wordscntr %Lx, returning %d\n",
591 			    words, ret);
592 	}
593 	return  ret;
594 }
595 
596 static pci_ers_result_t
pci_slot_reset(struct pci_dev * pdev)597 pci_slot_reset(struct pci_dev *pdev)
598 {
599 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
600 
601 	dd_dev_info(dd, "HFI1 slot_reset function called, ignored\n");
602 	return PCI_ERS_RESULT_CAN_RECOVER;
603 }
604 
605 static pci_ers_result_t
pci_link_reset(struct pci_dev * pdev)606 pci_link_reset(struct pci_dev *pdev)
607 {
608 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
609 
610 	dd_dev_info(dd, "HFI1 link_reset function called, ignored\n");
611 	return PCI_ERS_RESULT_CAN_RECOVER;
612 }
613 
614 static void
pci_resume(struct pci_dev * pdev)615 pci_resume(struct pci_dev *pdev)
616 {
617 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
618 
619 	dd_dev_info(dd, "HFI1 resume function called\n");
620 	pci_cleanup_aer_uncorrect_error_status(pdev);
621 	/*
622 	 * Running jobs will fail, since it's asynchronous
623 	 * unlike sysfs-requested reset.   Better than
624 	 * doing nothing.
625 	 */
626 	hfi1_init(dd, 1); /* same as re-init after reset */
627 }
628 
629 const struct pci_error_handlers hfi1_pci_err_handler = {
630 	.error_detected = pci_error_detected,
631 	.mmio_enabled = pci_mmio_enabled,
632 	.link_reset = pci_link_reset,
633 	.slot_reset = pci_slot_reset,
634 	.resume = pci_resume,
635 };
636 
637 /*============================================================================*/
638 /* PCIe Gen3 support */
639 
640 /*
641  * This code is separated out because it is expected to be removed in the
642  * final shipping product.  If not, then it will be revisited and items
643  * will be moved to more standard locations.
644  */
645 
646 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_STS field values */
647 #define DL_STATUS_HFI0 0x1	/* hfi0 firmware download complete */
648 #define DL_STATUS_HFI1 0x2	/* hfi1 firmware download complete */
649 #define DL_STATUS_BOTH 0x3	/* hfi0 and hfi1 firmware download complete */
650 
651 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_ERR field values */
652 #define DL_ERR_NONE		0x0	/* no error */
653 #define DL_ERR_SWAP_PARITY	0x1	/* parity error in SerDes interrupt */
654 					/*   or response data */
655 #define DL_ERR_DISABLED	0x2	/* hfi disabled */
656 #define DL_ERR_SECURITY	0x3	/* security check failed */
657 #define DL_ERR_SBUS		0x4	/* SBus status error */
658 #define DL_ERR_XFR_PARITY	0x5	/* parity error during ROM transfer*/
659 
660 /* gasket block secondary bus reset delay */
661 #define SBR_DELAY_US 200000	/* 200ms */
662 
663 /* mask for PCIe capability register lnkctl2 target link speed */
664 #define LNKCTL2_TARGET_LINK_SPEED_MASK 0xf
665 
666 static uint pcie_target = 3;
667 module_param(pcie_target, uint, S_IRUGO);
668 MODULE_PARM_DESC(pcie_target, "PCIe target speed (0 skip, 1-3 Gen1-3)");
669 
670 static uint pcie_force;
671 module_param(pcie_force, uint, S_IRUGO);
672 MODULE_PARM_DESC(pcie_force, "Force driver to do a PCIe firmware download even if already at target speed");
673 
674 static uint pcie_retry = 5;
675 module_param(pcie_retry, uint, S_IRUGO);
676 MODULE_PARM_DESC(pcie_retry, "Driver will try this many times to reach requested speed");
677 
678 #define UNSET_PSET 255
679 #define DEFAULT_DISCRETE_PSET 2	/* discrete HFI */
680 #define DEFAULT_MCP_PSET 4	/* MCP HFI */
681 static uint pcie_pset = UNSET_PSET;
682 module_param(pcie_pset, uint, S_IRUGO);
683 MODULE_PARM_DESC(pcie_pset, "PCIe Eq Pset value to use, range is 0-10");
684 
685 /* equalization columns */
686 #define PREC 0
687 #define ATTN 1
688 #define POST 2
689 
690 /* discrete silicon preliminary equalization values */
691 static const u8 discrete_preliminary_eq[11][3] = {
692 	/* prec   attn   post */
693 	{  0x00,  0x00,  0x12 },	/* p0 */
694 	{  0x00,  0x00,  0x0c },	/* p1 */
695 	{  0x00,  0x00,  0x0f },	/* p2 */
696 	{  0x00,  0x00,  0x09 },	/* p3 */
697 	{  0x00,  0x00,  0x00 },	/* p4 */
698 	{  0x06,  0x00,  0x00 },	/* p5 */
699 	{  0x09,  0x00,  0x00 },	/* p6 */
700 	{  0x06,  0x00,  0x0f },	/* p7 */
701 	{  0x09,  0x00,  0x09 },	/* p8 */
702 	{  0x0c,  0x00,  0x00 },	/* p9 */
703 	{  0x00,  0x00,  0x18 },	/* p10 */
704 };
705 
706 /* integrated silicon preliminary equalization values */
707 static const u8 integrated_preliminary_eq[11][3] = {
708 	/* prec   attn   post */
709 	{  0x00,  0x1e,  0x07 },	/* p0 */
710 	{  0x00,  0x1e,  0x05 },	/* p1 */
711 	{  0x00,  0x1e,  0x06 },	/* p2 */
712 	{  0x00,  0x1e,  0x04 },	/* p3 */
713 	{  0x00,  0x1e,  0x00 },	/* p4 */
714 	{  0x03,  0x1e,  0x00 },	/* p5 */
715 	{  0x04,  0x1e,  0x00 },	/* p6 */
716 	{  0x03,  0x1e,  0x06 },	/* p7 */
717 	{  0x03,  0x1e,  0x04 },	/* p8 */
718 	{  0x05,  0x1e,  0x00 },	/* p9 */
719 	{  0x00,  0x1e,  0x0a },	/* p10 */
720 };
721 
722 /* helper to format the value to write to hardware */
723 #define eq_value(pre, curr, post) \
724 	((((u32)(pre)) << \
725 			PCIE_CFG_REG_PL102_GEN3_EQ_PRE_CURSOR_PSET_SHIFT) \
726 	| (((u32)(curr)) << PCIE_CFG_REG_PL102_GEN3_EQ_CURSOR_PSET_SHIFT) \
727 	| (((u32)(post)) << \
728 		PCIE_CFG_REG_PL102_GEN3_EQ_POST_CURSOR_PSET_SHIFT))
729 
730 /*
731  * Load the given EQ preset table into the PCIe hardware.
732  */
load_eq_table(struct hfi1_devdata * dd,const u8 eq[11][3],u8 fs,u8 div)733 static int load_eq_table(struct hfi1_devdata *dd, const u8 eq[11][3], u8 fs,
734 			 u8 div)
735 {
736 	struct pci_dev *pdev = dd->pcidev;
737 	u32 hit_error = 0;
738 	u32 violation;
739 	u32 i;
740 	u8 c_minus1, c0, c_plus1;
741 
742 	for (i = 0; i < 11; i++) {
743 		/* set index */
744 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL103, i);
745 		/* write the value */
746 		c_minus1 = eq[i][PREC] / div;
747 		c0 = fs - (eq[i][PREC] / div) - (eq[i][POST] / div);
748 		c_plus1 = eq[i][POST] / div;
749 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL102,
750 			eq_value(c_minus1, c0, c_plus1));
751 		/* check if these coefficients violate EQ rules */
752 		pci_read_config_dword(dd->pcidev, PCIE_CFG_REG_PL105,
753 								&violation);
754 		if (violation
755 		    & PCIE_CFG_REG_PL105_GEN3_EQ_VIOLATE_COEF_RULES_SMASK){
756 			if (hit_error == 0) {
757 				dd_dev_err(dd,
758 					"Gen3 EQ Table Coefficient rule violations\n");
759 				dd_dev_err(dd, "         prec   attn   post\n");
760 			}
761 			dd_dev_err(dd, "   p%02d:   %02x     %02x     %02x\n",
762 				i, (u32)eq[i][0], (u32)eq[i][1], (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 hardware mutex is already being held.
777  */
pcie_post_steps(struct hfi1_devdata * dd)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  */
trigger_sbr(struct hfi1_devdata * dd)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 	 * A secondary bus reset (SBR) issues a hot reset to our device.
826 	 * The following routine does a 1s wait after the reset is dropped
827 	 * per PCI Trhfa (recovery time).  PCIe 3.0 section 6.6.1 -
828 	 * Conventional Reset, paragraph 3, line 35 also says that a 1s
829 	 * delay after a reset is required.  Per spec requirements,
830 	 * the link is either working or not after that point.
831 	 */
832 	pci_reset_bridge_secondary_bus(dev->bus->self);
833 
834 	return 0;
835 }
836 
837 /*
838  * Write the given gasket interrupt register.
839  */
write_gasket_interrupt(struct hfi1_devdata * dd,int index,u16 code,u16 data)840 static void write_gasket_interrupt(struct hfi1_devdata *dd, int index,
841 				   u16 code, u16 data)
842 {
843 	write_csr(dd, ASIC_PCIE_SD_INTRPT_LIST + (index * 8),
844 	    (((u64)code << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_CODE_SHIFT)
845 	    |((u64)data << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_DATA_SHIFT)));
846 }
847 
848 /*
849  * Tell the gasket logic how to react to the reset.
850  */
arm_gasket_logic(struct hfi1_devdata * dd)851 static void arm_gasket_logic(struct hfi1_devdata *dd)
852 {
853 	u64 reg;
854 
855 	reg = (((u64)1 << dd->hfi1_id)
856 			<< ASIC_PCIE_SD_HOST_CMD_INTRPT_CMD_SHIFT)
857 		| ((u64)pcie_serdes_broadcast[dd->hfi1_id]
858 			<< ASIC_PCIE_SD_HOST_CMD_SBUS_RCVR_ADDR_SHIFT
859 		| ASIC_PCIE_SD_HOST_CMD_SBR_MODE_SMASK
860 		| ((u64)SBR_DELAY_US & ASIC_PCIE_SD_HOST_CMD_TIMER_MASK)
861 			<< ASIC_PCIE_SD_HOST_CMD_TIMER_SHIFT
862 		);
863 	write_csr(dd, ASIC_PCIE_SD_HOST_CMD, reg);
864 	/* read back to push the write */
865 	read_csr(dd, ASIC_PCIE_SD_HOST_CMD);
866 }
867 
868 /*
869  * Do all the steps needed to transition the PCIe link to Gen3 speed.
870  */
do_pcie_gen3_transition(struct hfi1_devdata * dd)871 int do_pcie_gen3_transition(struct hfi1_devdata *dd)
872 {
873 	struct pci_dev *parent;
874 	u64 fw_ctrl;
875 	u64 reg, therm;
876 	u32 reg32, fs, lf;
877 	u32 status, err;
878 	int ret;
879 	int do_retry, retry_count = 0;
880 	uint default_pset;
881 	u16 target_vector, target_speed;
882 	u16 lnkctl, lnkctl2, vendor;
883 	u8 nsbr = 1;
884 	u8 div;
885 	const u8 (*eq)[3];
886 	int return_error = 0;
887 
888 	/* PCIe Gen3 is for the ASIC only */
889 	if (dd->icode != ICODE_RTL_SILICON)
890 		return 0;
891 
892 	if (pcie_target == 1) {			/* target Gen1 */
893 		target_vector = GEN1_SPEED_VECTOR;
894 		target_speed = 2500;
895 	} else if (pcie_target == 2) {		/* target Gen2 */
896 		target_vector = GEN2_SPEED_VECTOR;
897 		target_speed = 5000;
898 	} else if (pcie_target == 3) {		/* target Gen3 */
899 		target_vector = GEN3_SPEED_VECTOR;
900 		target_speed = 8000;
901 	} else {
902 		/* off or invalid target - skip */
903 		dd_dev_info(dd, "%s: Skipping PCIe transition\n", __func__);
904 		return 0;
905 	}
906 
907 	/* if already at target speed, done (unless forced) */
908 	if (dd->lbus_speed == target_speed) {
909 		dd_dev_info(dd, "%s: PCIe already at gen%d, %s\n", __func__,
910 			pcie_target,
911 			pcie_force ? "re-doing anyway" : "skipping");
912 		if (!pcie_force)
913 			return 0;
914 	}
915 
916 	/*
917 	 * A0 needs an additional SBR
918 	 */
919 	if (is_a0(dd))
920 		nsbr++;
921 
922 	/*
923 	 * Do the Gen3 transition.  Steps are those of the PCIe Gen3
924 	 * recipe.
925 	 */
926 
927 	/* step 1: pcie link working in gen1/gen2 */
928 
929 	/* step 2: if either side is not capable of Gen3, done */
930 	if (pcie_target == 3 && !dd->link_gen3_capable) {
931 		dd_dev_err(dd, "The PCIe link is not Gen3 capable\n");
932 		ret = -ENOSYS;
933 		goto done_no_mutex;
934 	}
935 
936 	/* hold the HW mutex across the firmware download and SBR */
937 	ret = acquire_hw_mutex(dd);
938 	if (ret)
939 		return ret;
940 
941 	/* make sure thermal polling is not causing interrupts */
942 	therm = read_csr(dd, ASIC_CFG_THERM_POLL_EN);
943 	if (therm) {
944 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x0);
945 		msleep(100);
946 		dd_dev_info(dd, "%s: Disabled therm polling\n",
947 			    __func__);
948 	}
949 
950 retry:
951 
952 	if (therm) {
953 		/*
954 		 * toggle SPICO_ENABLE to get back to the state
955 		 * just after the firmware load
956 		 */
957 		sbus_request(dd, SBUS_MASTER_BROADCAST, 0x01,
958 			WRITE_SBUS_RECEIVER, 0x00000040);
959 		sbus_request(dd, SBUS_MASTER_BROADCAST, 0x01,
960 			WRITE_SBUS_RECEIVER, 0x00000140);
961 	}
962 
963 	/* step 3: download SBus Master firmware */
964 	/* step 4: download PCIe Gen3 SerDes firmware */
965 	dd_dev_info(dd, "%s: downloading firmware\n", __func__);
966 	ret = load_pcie_firmware(dd);
967 	if (ret)
968 		goto done;
969 
970 	/* step 5: set up device parameter settings */
971 	dd_dev_info(dd, "%s: setting PCIe registers\n", __func__);
972 
973 	/*
974 	 * PcieCfgSpcie1 - Link Control 3
975 	 * Leave at reset value.  No need to set PerfEq - link equalization
976 	 * will be performed automatically after the SBR when the target
977 	 * speed is 8GT/s.
978 	 */
979 
980 	/* clear all 16 per-lane error bits (PCIe: Lane Error Status) */
981 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, 0xffff);
982 
983 	/* step 5a: Set Synopsys Port Logic registers */
984 
985 	/*
986 	 * PcieCfgRegPl2 - Port Force Link
987 	 *
988 	 * Set the low power field to 0x10 to avoid unnecessary power
989 	 * management messages.  All other fields are zero.
990 	 */
991 	reg32 = 0x10ul << PCIE_CFG_REG_PL2_LOW_PWR_ENT_CNT_SHIFT;
992 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL2, reg32);
993 
994 	/*
995 	 * PcieCfgRegPl100 - Gen3 Control
996 	 *
997 	 * turn off PcieCfgRegPl100.Gen3ZRxDcNonCompl
998 	 * turn on PcieCfgRegPl100.EqEieosCnt (erratum)
999 	 * Everything else zero.
1000 	 */
1001 	reg32 = PCIE_CFG_REG_PL100_EQ_EIEOS_CNT_SMASK;
1002 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL100, reg32);
1003 
1004 	/*
1005 	 * PcieCfgRegPl101 - Gen3 EQ FS and LF
1006 	 * PcieCfgRegPl102 - Gen3 EQ Presets to Coefficients Mapping
1007 	 * PcieCfgRegPl103 - Gen3 EQ Preset Index
1008 	 * PcieCfgRegPl105 - Gen3 EQ Status
1009 	 *
1010 	 * Give initial EQ settings.
1011 	 */
1012 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL0) { /* discrete */
1013 		/* 1000mV, FS=24, LF = 8 */
1014 		fs = 24;
1015 		lf = 8;
1016 		div = 3;
1017 		eq = discrete_preliminary_eq;
1018 		default_pset = DEFAULT_DISCRETE_PSET;
1019 	} else {
1020 		/* 400mV, FS=29, LF = 9 */
1021 		fs = 29;
1022 		lf = 9;
1023 		div = 1;
1024 		eq = integrated_preliminary_eq;
1025 		default_pset = DEFAULT_MCP_PSET;
1026 	}
1027 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL101,
1028 		(fs << PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_FS_SHIFT)
1029 		| (lf << PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_LF_SHIFT));
1030 	ret = load_eq_table(dd, eq, fs, div);
1031 	if (ret)
1032 		goto done;
1033 
1034 	/*
1035 	 * PcieCfgRegPl106 - Gen3 EQ Control
1036 	 *
1037 	 * Set Gen3EqPsetReqVec, leave other fields 0.
1038 	 */
1039 	if (pcie_pset == UNSET_PSET)
1040 		pcie_pset = default_pset;
1041 	if (pcie_pset > 10) {	/* valid range is 0-10, inclusive */
1042 		dd_dev_err(dd, "%s: Invalid Eq Pset %u, setting to %d\n",
1043 			__func__, pcie_pset, default_pset);
1044 		pcie_pset = default_pset;
1045 	}
1046 	dd_dev_info(dd, "%s: using EQ Pset %u\n", __func__, pcie_pset);
1047 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL106,
1048 		((1 << pcie_pset)
1049 			<< PCIE_CFG_REG_PL106_GEN3_EQ_PSET_REQ_VEC_SHIFT)
1050 		| PCIE_CFG_REG_PL106_GEN3_EQ_EVAL2MS_DISABLE_SMASK
1051 		| PCIE_CFG_REG_PL106_GEN3_EQ_PHASE23_EXIT_MODE_SMASK);
1052 
1053 	/*
1054 	 * step 5b: Do post firmware download steps via SBus
1055 	 */
1056 	dd_dev_info(dd, "%s: doing pcie post steps\n", __func__);
1057 	pcie_post_steps(dd);
1058 
1059 	/*
1060 	 * step 5c: Program gasket interrupts
1061 	 */
1062 	/* set the Rx Bit Rate to REFCLK ratio */
1063 	write_gasket_interrupt(dd, 0, 0x0006, 0x0050);
1064 	/* disable pCal for PCIe Gen3 RX equalization */
1065 	write_gasket_interrupt(dd, 1, 0x0026, 0x5b01);
1066 	/*
1067 	 * Enable iCal for PCIe Gen3 RX equalization, and set which
1068 	 * evaluation of RX_EQ_EVAL will launch the iCal procedure.
1069 	 */
1070 	write_gasket_interrupt(dd, 2, 0x0026, 0x5202);
1071 	/* terminate list */
1072 	write_gasket_interrupt(dd, 3, 0x0000, 0x0000);
1073 
1074 	/*
1075 	 * step 5d: program XMT margin
1076 	 * Right now, leave the default alone.  To change, do a
1077 	 * read-modify-write of:
1078 	 *	CcePcieCtrl.XmtMargin
1079 	 *	CcePcieCtrl.XmitMarginOverwriteEnable
1080 	 */
1081 
1082 	/* step 5e: disable active state power management (ASPM) */
1083 	dd_dev_info(dd, "%s: clearing ASPM\n", __func__);
1084 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL, &lnkctl);
1085 	lnkctl &= ~PCI_EXP_LNKCTL_ASPMC;
1086 	pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL, lnkctl);
1087 
1088 	/*
1089 	 * step 5f: clear DirectSpeedChange
1090 	 * PcieCfgRegPl67.DirectSpeedChange must be zero to prevent the
1091 	 * change in the speed target from starting before we are ready.
1092 	 * This field defaults to 0 and we are not changing it, so nothing
1093 	 * needs to be done.
1094 	 */
1095 
1096 	/* step 5g: Set target link speed */
1097 	/*
1098 	 * Set target link speed to be target on both device and parent.
1099 	 * On setting the parent: Some system BIOSs "helpfully" set the
1100 	 * parent target speed to Gen2 to match the ASIC's initial speed.
1101 	 * We can set the target Gen3 because we have already checked
1102 	 * that it is Gen3 capable earlier.
1103 	 */
1104 	dd_dev_info(dd, "%s: setting parent target link speed\n", __func__);
1105 	parent = dd->pcidev->bus->self;
1106 	pcie_capability_read_word(parent, PCI_EXP_LNKCTL2, &lnkctl2);
1107 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1108 		(u32)lnkctl2);
1109 	/* only write to parent if target is not as high as ours */
1110 	if ((lnkctl2 & LNKCTL2_TARGET_LINK_SPEED_MASK) < target_vector) {
1111 		lnkctl2 &= ~LNKCTL2_TARGET_LINK_SPEED_MASK;
1112 		lnkctl2 |= target_vector;
1113 		dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1114 			(u32)lnkctl2);
1115 		pcie_capability_write_word(parent, PCI_EXP_LNKCTL2, lnkctl2);
1116 	} else {
1117 		dd_dev_info(dd, "%s: ..target speed is OK\n", __func__);
1118 	}
1119 
1120 	dd_dev_info(dd, "%s: setting target link speed\n", __func__);
1121 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL2, &lnkctl2);
1122 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1123 		(u32)lnkctl2);
1124 	lnkctl2 &= ~LNKCTL2_TARGET_LINK_SPEED_MASK;
1125 	lnkctl2 |= target_vector;
1126 	dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1127 		(u32)lnkctl2);
1128 	pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL2, lnkctl2);
1129 
1130 	/* step 5h: arm gasket logic */
1131 	/* hold DC in reset across the SBR */
1132 	write_csr(dd, CCE_DC_CTRL, CCE_DC_CTRL_DC_RESET_SMASK);
1133 	(void) read_csr(dd, CCE_DC_CTRL); /* DC reset hold */
1134 	/* save firmware control across the SBR */
1135 	fw_ctrl = read_csr(dd, MISC_CFG_FW_CTRL);
1136 
1137 	dd_dev_info(dd, "%s: arming gasket logic\n", __func__);
1138 	arm_gasket_logic(dd);
1139 
1140 	/*
1141 	 * step 6: quiesce PCIe link
1142 	 * The chip has already been reset, so there will be no traffic
1143 	 * from the chip.  Linux has no easy way to enforce that it will
1144 	 * not try to access the device, so we just need to hope it doesn't
1145 	 * do it while we are doing the reset.
1146 	 */
1147 
1148 	/*
1149 	 * step 7: initiate the secondary bus reset (SBR)
1150 	 * step 8: hardware brings the links back up
1151 	 * step 9: wait for link speed transition to be complete
1152 	 */
1153 	dd_dev_info(dd, "%s: calling trigger_sbr\n", __func__);
1154 	ret = trigger_sbr(dd);
1155 	if (ret)
1156 		goto done;
1157 
1158 	/* step 10: decide what to do next */
1159 
1160 	/* check if we can read PCI space */
1161 	ret = pci_read_config_word(dd->pcidev, PCI_VENDOR_ID, &vendor);
1162 	if (ret) {
1163 		dd_dev_info(dd,
1164 			"%s: read of VendorID failed after SBR, err %d\n",
1165 			__func__, ret);
1166 		return_error = 1;
1167 		goto done;
1168 	}
1169 	if (vendor == 0xffff) {
1170 		dd_dev_info(dd, "%s: VendorID is all 1s after SBR\n", __func__);
1171 		return_error = 1;
1172 		ret = -EIO;
1173 		goto done;
1174 	}
1175 
1176 	/* restore PCI space registers we know were reset */
1177 	dd_dev_info(dd, "%s: calling restore_pci_variables\n", __func__);
1178 	restore_pci_variables(dd);
1179 	/* restore firmware control */
1180 	write_csr(dd, MISC_CFG_FW_CTRL, fw_ctrl);
1181 
1182 	/*
1183 	 * Check the gasket block status.
1184 	 *
1185 	 * This is the first CSR read after the SBR.  If the read returns
1186 	 * all 1s (fails), the link did not make it back.
1187 	 *
1188 	 * Once we're sure we can read and write, clear the DC reset after
1189 	 * the SBR.  Then check for any per-lane errors. Then look over
1190 	 * the status.
1191 	 */
1192 	reg = read_csr(dd, ASIC_PCIE_SD_HOST_STATUS);
1193 	dd_dev_info(dd, "%s: gasket block status: 0x%llx\n", __func__, reg);
1194 	if (reg == ~0ull) {	/* PCIe read failed/timeout */
1195 		dd_dev_err(dd, "SBR failed - unable to read from device\n");
1196 		return_error = 1;
1197 		ret = -ENOSYS;
1198 		goto done;
1199 	}
1200 
1201 	/* clear the DC reset */
1202 	write_csr(dd, CCE_DC_CTRL, 0);
1203 
1204 	/* Set the LED off */
1205 	if (is_a0(dd))
1206 		setextled(dd, 0);
1207 
1208 	/* check for any per-lane errors */
1209 	pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, &reg32);
1210 	dd_dev_info(dd, "%s: per-lane errors: 0x%x\n", __func__, reg32);
1211 
1212 	/* extract status, look for our HFI */
1213 	status = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_SHIFT)
1214 			& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_MASK;
1215 	if ((status & (1 << dd->hfi1_id)) == 0) {
1216 		dd_dev_err(dd,
1217 			"%s: gasket status 0x%x, expecting 0x%x\n",
1218 			__func__, status, 1 << dd->hfi1_id);
1219 		ret = -EIO;
1220 		goto done;
1221 	}
1222 
1223 	/* extract error */
1224 	err = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_SHIFT)
1225 		& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_MASK;
1226 	if (err) {
1227 		dd_dev_err(dd, "%s: gasket error %d\n", __func__, err);
1228 		ret = -EIO;
1229 		goto done;
1230 	}
1231 
1232 	/* update our link information cache */
1233 	update_lbus_info(dd);
1234 	dd_dev_info(dd, "%s: new speed and width: %s\n", __func__,
1235 		dd->lbus_info);
1236 
1237 	if (dd->lbus_speed != target_speed) { /* not target */
1238 		/* maybe retry */
1239 		do_retry = retry_count < pcie_retry;
1240 		dd_dev_err(dd, "PCIe link speed did not switch to Gen%d%s\n",
1241 			pcie_target, do_retry ? ", retrying" : "");
1242 		retry_count++;
1243 		if (do_retry) {
1244 			msleep(100); /* allow time to settle */
1245 			goto retry;
1246 		}
1247 		ret = -EIO;
1248 	}
1249 
1250 done:
1251 	if (therm) {
1252 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x1);
1253 		msleep(100);
1254 		dd_dev_info(dd, "%s: Re-enable therm polling\n",
1255 			    __func__);
1256 	}
1257 	release_hw_mutex(dd);
1258 done_no_mutex:
1259 	/* return no error if it is OK to be at current speed */
1260 	if (ret && !return_error) {
1261 		dd_dev_err(dd, "Proceeding at current speed PCIe speed\n");
1262 		ret = 0;
1263 	}
1264 
1265 	dd_dev_info(dd, "%s: done\n", __func__);
1266 	return ret;
1267 }
1268