xref: /openbmc/linux/drivers/infiniband/hw/hfi1/pcie.c (revision c51d39010a1bccc9c1294e2d7c00005aefeb2b5c)
1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/pci.h>
49 #include <linux/io.h>
50 #include <linux/delay.h>
51 #include <linux/vmalloc.h>
52 #include <linux/aer.h>
53 #include <linux/module.h>
54 
55 #include "hfi.h"
56 #include "chip_registers.h"
57 #include "aspm.h"
58 
59 /* link speed vector for Gen3 speed - not in Linux headers */
60 #define GEN1_SPEED_VECTOR 0x1
61 #define GEN2_SPEED_VECTOR 0x2
62 #define GEN3_SPEED_VECTOR 0x3
63 
64 /*
65  * This file contains PCIe utility routines.
66  */
67 
68 /*
69  * Code to adjust PCIe capabilities.
70  */
71 static void tune_pcie_caps(struct hfi1_devdata *);
72 
73 /*
74  * Do all the common PCIe setup and initialization.
75  * devdata is not yet allocated, and is not allocated until after this
76  * routine returns success.  Therefore dd_dev_err() can't be used for error
77  * printing.
78  */
79 int hfi1_pcie_init(struct pci_dev *pdev, const struct pci_device_id *ent)
80 {
81 	int ret;
82 
83 	ret = pci_enable_device(pdev);
84 	if (ret) {
85 		/*
86 		 * This can happen (in theory) iff:
87 		 * We did a chip reset, and then failed to reprogram the
88 		 * BAR, or the chip reset due to an internal error.  We then
89 		 * unloaded the driver and reloaded it.
90 		 *
91 		 * Both reset cases set the BAR back to initial state.  For
92 		 * the latter case, the AER sticky error bit at offset 0x718
93 		 * should be set, but the Linux kernel doesn't yet know
94 		 * about that, it appears.  If the original BAR was retained
95 		 * in the kernel data structures, this may be OK.
96 		 */
97 		hfi1_early_err(&pdev->dev, "pci enable failed: error %d\n",
98 			       -ret);
99 		goto done;
100 	}
101 
102 	ret = pci_request_regions(pdev, DRIVER_NAME);
103 	if (ret) {
104 		hfi1_early_err(&pdev->dev,
105 			       "pci_request_regions fails: err %d\n", -ret);
106 		goto bail;
107 	}
108 
109 	ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
110 	if (ret) {
111 		/*
112 		 * If the 64 bit setup fails, try 32 bit.  Some systems
113 		 * do not setup 64 bit maps on systems with 2GB or less
114 		 * memory installed.
115 		 */
116 		ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
117 		if (ret) {
118 			hfi1_early_err(&pdev->dev,
119 				       "Unable to set DMA mask: %d\n", ret);
120 			goto bail;
121 		}
122 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
123 	} else {
124 		ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
125 	}
126 	if (ret) {
127 		hfi1_early_err(&pdev->dev,
128 			       "Unable to set DMA consistent mask: %d\n", ret);
129 		goto bail;
130 	}
131 
132 	pci_set_master(pdev);
133 	(void)pci_enable_pcie_error_reporting(pdev);
134 	goto done;
135 
136 bail:
137 	hfi1_pcie_cleanup(pdev);
138 done:
139 	return ret;
140 }
141 
142 /*
143  * Clean what was done in hfi1_pcie_init()
144  */
145 void hfi1_pcie_cleanup(struct pci_dev *pdev)
146 {
147 	pci_disable_device(pdev);
148 	/*
149 	 * Release regions should be called after the disable. OK to
150 	 * call if request regions has not been called or failed.
151 	 */
152 	pci_release_regions(pdev);
153 }
154 
155 /*
156  * Do remaining PCIe setup, once dd is allocated, and save away
157  * fields required to re-initialize after a chip reset, or for
158  * various other purposes
159  */
160 int hfi1_pcie_ddinit(struct hfi1_devdata *dd, struct pci_dev *pdev)
161 {
162 	unsigned long len;
163 	resource_size_t addr;
164 
165 	dd->pcidev = pdev;
166 	pci_set_drvdata(pdev, dd);
167 
168 	addr = pci_resource_start(pdev, 0);
169 	len = pci_resource_len(pdev, 0);
170 
171 	/*
172 	 * The TXE PIO buffers are at the tail end of the chip space.
173 	 * Cut them off and map them separately.
174 	 */
175 
176 	/* sanity check vs expectations */
177 	if (len != TXE_PIO_SEND + TXE_PIO_SIZE) {
178 		dd_dev_err(dd, "chip PIO range does not match\n");
179 		return -EINVAL;
180 	}
181 
182 	dd->kregbase = ioremap_nocache(addr, TXE_PIO_SEND);
183 	if (!dd->kregbase)
184 		return -ENOMEM;
185 
186 	dd->piobase = ioremap_wc(addr + TXE_PIO_SEND, TXE_PIO_SIZE);
187 	if (!dd->piobase) {
188 		iounmap(dd->kregbase);
189 		return -ENOMEM;
190 	}
191 
192 	dd->flags |= HFI1_PRESENT;	/* now register routines work */
193 
194 	dd->kregend = dd->kregbase + TXE_PIO_SEND;
195 	dd->physaddr = addr;        /* used for io_remap, etc. */
196 
197 	/*
198 	 * Re-map the chip's RcvArray as write-combining to allow us
199 	 * to write an entire cacheline worth of entries in one shot.
200 	 * If this re-map fails, just continue - the RcvArray programming
201 	 * function will handle both cases.
202 	 */
203 	dd->chip_rcv_array_count = read_csr(dd, RCV_ARRAY_CNT);
204 	dd->rcvarray_wc = ioremap_wc(addr + RCV_ARRAY,
205 				     dd->chip_rcv_array_count * 8);
206 	dd_dev_info(dd, "WC Remapped RcvArray: %p\n", dd->rcvarray_wc);
207 	/*
208 	 * Save BARs and command to rewrite after device reset.
209 	 */
210 	dd->pcibar0 = addr;
211 	dd->pcibar1 = addr >> 32;
212 	pci_read_config_dword(dd->pcidev, PCI_ROM_ADDRESS, &dd->pci_rom);
213 	pci_read_config_word(dd->pcidev, PCI_COMMAND, &dd->pci_command);
214 	pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &dd->pcie_devctl);
215 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL, &dd->pcie_lnkctl);
216 	pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL2,
217 				  &dd->pcie_devctl2);
218 	pci_read_config_dword(dd->pcidev, PCI_CFG_MSIX0, &dd->pci_msix0);
219 	pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE1, &dd->pci_lnkctl3);
220 	pci_read_config_dword(dd->pcidev, PCIE_CFG_TPH2, &dd->pci_tph2);
221 
222 	return 0;
223 }
224 
225 /*
226  * Do PCIe cleanup related to dd, after chip-specific cleanup, etc.  Just prior
227  * to releasing the dd memory.
228  * Void because all of the core pcie cleanup functions are void.
229  */
230 void hfi1_pcie_ddcleanup(struct hfi1_devdata *dd)
231 {
232 	u64 __iomem *base = (void __iomem *)dd->kregbase;
233 
234 	dd->flags &= ~HFI1_PRESENT;
235 	dd->kregbase = NULL;
236 	iounmap(base);
237 	if (dd->rcvarray_wc)
238 		iounmap(dd->rcvarray_wc);
239 	if (dd->piobase)
240 		iounmap(dd->piobase);
241 }
242 
243 /*
244  * Do a Function Level Reset (FLR) on the device.
245  * Based on static function drivers/pci/pci.c:pcie_flr().
246  */
247 void hfi1_pcie_flr(struct hfi1_devdata *dd)
248 {
249 	int i;
250 	u16 status;
251 
252 	/* no need to check for the capability - we know the device has it */
253 
254 	/* wait for Transaction Pending bit to clear, at most a few ms */
255 	for (i = 0; i < 4; i++) {
256 		if (i)
257 			msleep((1 << (i - 1)) * 100);
258 
259 		pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVSTA, &status);
260 		if (!(status & PCI_EXP_DEVSTA_TRPND))
261 			goto clear;
262 	}
263 
264 	dd_dev_err(dd, "Transaction Pending bit is not clearing, proceeding with reset anyway\n");
265 
266 clear:
267 	pcie_capability_set_word(dd->pcidev, PCI_EXP_DEVCTL,
268 				 PCI_EXP_DEVCTL_BCR_FLR);
269 	/* PCIe spec requires the function to be back within 100ms */
270 	msleep(100);
271 }
272 
273 static void msix_setup(struct hfi1_devdata *dd, int pos, u32 *msixcnt,
274 		       struct hfi1_msix_entry *hfi1_msix_entry)
275 {
276 	int ret;
277 	int nvec = *msixcnt;
278 	struct msix_entry *msix_entry;
279 	int i;
280 
281 	/*
282 	 * We can't pass hfi1_msix_entry array to msix_setup
283 	 * so use a dummy msix_entry array and copy the allocated
284 	 * irq back to the hfi1_msix_entry array.
285 	 */
286 	msix_entry = kmalloc_array(nvec, sizeof(*msix_entry), GFP_KERNEL);
287 	if (!msix_entry) {
288 		ret = -ENOMEM;
289 		goto do_intx;
290 	}
291 
292 	for (i = 0; i < nvec; i++)
293 		msix_entry[i] = hfi1_msix_entry[i].msix;
294 
295 	ret = pci_enable_msix_range(dd->pcidev, msix_entry, 1, nvec);
296 	if (ret < 0)
297 		goto free_msix_entry;
298 	nvec = ret;
299 
300 	for (i = 0; i < nvec; i++)
301 		hfi1_msix_entry[i].msix = msix_entry[i];
302 
303 	kfree(msix_entry);
304 	*msixcnt = nvec;
305 	return;
306 
307 free_msix_entry:
308 	kfree(msix_entry);
309 
310 do_intx:
311 	dd_dev_err(dd, "pci_enable_msix_range %d vectors failed: %d, falling back to INTx\n",
312 		   nvec, ret);
313 	*msixcnt = 0;
314 	hfi1_enable_intx(dd->pcidev);
315 }
316 
317 /* return the PCIe link speed from the given link status */
318 static u32 extract_speed(u16 linkstat)
319 {
320 	u32 speed;
321 
322 	switch (linkstat & PCI_EXP_LNKSTA_CLS) {
323 	default: /* not defined, assume Gen1 */
324 	case PCI_EXP_LNKSTA_CLS_2_5GB:
325 		speed = 2500; /* Gen 1, 2.5GHz */
326 		break;
327 	case PCI_EXP_LNKSTA_CLS_5_0GB:
328 		speed = 5000; /* Gen 2, 5GHz */
329 		break;
330 	case GEN3_SPEED_VECTOR:
331 		speed = 8000; /* Gen 3, 8GHz */
332 		break;
333 	}
334 	return speed;
335 }
336 
337 /* return the PCIe link speed from the given link status */
338 static u32 extract_width(u16 linkstat)
339 {
340 	return (linkstat & PCI_EXP_LNKSTA_NLW) >> PCI_EXP_LNKSTA_NLW_SHIFT;
341 }
342 
343 /* read the link status and set dd->{lbus_width,lbus_speed,lbus_info} */
344 static void update_lbus_info(struct hfi1_devdata *dd)
345 {
346 	u16 linkstat;
347 
348 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKSTA, &linkstat);
349 	dd->lbus_width = extract_width(linkstat);
350 	dd->lbus_speed = extract_speed(linkstat);
351 	snprintf(dd->lbus_info, sizeof(dd->lbus_info),
352 		 "PCIe,%uMHz,x%u", dd->lbus_speed, dd->lbus_width);
353 }
354 
355 /*
356  * Read in the current PCIe link width and speed.  Find if the link is
357  * Gen3 capable.
358  */
359 int pcie_speeds(struct hfi1_devdata *dd)
360 {
361 	u32 linkcap;
362 	struct pci_dev *parent = dd->pcidev->bus->self;
363 
364 	if (!pci_is_pcie(dd->pcidev)) {
365 		dd_dev_err(dd, "Can't find PCI Express capability!\n");
366 		return -EINVAL;
367 	}
368 
369 	/* find if our max speed is Gen3 and parent supports Gen3 speeds */
370 	dd->link_gen3_capable = 1;
371 
372 	pcie_capability_read_dword(dd->pcidev, PCI_EXP_LNKCAP, &linkcap);
373 	if ((linkcap & PCI_EXP_LNKCAP_SLS) != GEN3_SPEED_VECTOR) {
374 		dd_dev_info(dd,
375 			    "This HFI is not Gen3 capable, max speed 0x%x, need 0x3\n",
376 			    linkcap & PCI_EXP_LNKCAP_SLS);
377 		dd->link_gen3_capable = 0;
378 	}
379 
380 	/*
381 	 * bus->max_bus_speed is set from the bridge's linkcap Max Link Speed
382 	 */
383 	if (parent && dd->pcidev->bus->max_bus_speed != PCIE_SPEED_8_0GT) {
384 		dd_dev_info(dd, "Parent PCIe bridge does not support Gen3\n");
385 		dd->link_gen3_capable = 0;
386 	}
387 
388 	/* obtain the link width and current speed */
389 	update_lbus_info(dd);
390 
391 	dd_dev_info(dd, "%s\n", dd->lbus_info);
392 
393 	return 0;
394 }
395 
396 /*
397  * Returns in *nent:
398  *	- actual number of interrupts allocated
399  *	- 0 if fell back to INTx.
400  */
401 void request_msix(struct hfi1_devdata *dd, u32 *nent,
402 		  struct hfi1_msix_entry *entry)
403 {
404 	int pos;
405 
406 	pos = dd->pcidev->msix_cap;
407 	if (*nent && pos) {
408 		msix_setup(dd, pos, nent, entry);
409 		/* did it, either MSI-X or INTx */
410 	} else {
411 		*nent = 0;
412 		hfi1_enable_intx(dd->pcidev);
413 	}
414 
415 	tune_pcie_caps(dd);
416 }
417 
418 void hfi1_enable_intx(struct pci_dev *pdev)
419 {
420 	/* first, turn on INTx */
421 	pci_intx(pdev, 1);
422 	/* then turn off MSI-X */
423 	pci_disable_msix(pdev);
424 }
425 
426 /* restore command and BARs after a reset has wiped them out */
427 void restore_pci_variables(struct hfi1_devdata *dd)
428 {
429 	pci_write_config_word(dd->pcidev, PCI_COMMAND, dd->pci_command);
430 	pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0, dd->pcibar0);
431 	pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1, dd->pcibar1);
432 	pci_write_config_dword(dd->pcidev, PCI_ROM_ADDRESS, dd->pci_rom);
433 	pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL, dd->pcie_devctl);
434 	pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL, dd->pcie_lnkctl);
435 	pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL2,
436 				   dd->pcie_devctl2);
437 	pci_write_config_dword(dd->pcidev, PCI_CFG_MSIX0, dd->pci_msix0);
438 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE1, dd->pci_lnkctl3);
439 	pci_write_config_dword(dd->pcidev, PCIE_CFG_TPH2, dd->pci_tph2);
440 }
441 
442 /*
443  * BIOS may not set PCIe bus-utilization parameters for best performance.
444  * Check and optionally adjust them to maximize our throughput.
445  */
446 static int hfi1_pcie_caps;
447 module_param_named(pcie_caps, hfi1_pcie_caps, int, S_IRUGO);
448 MODULE_PARM_DESC(pcie_caps, "Max PCIe tuning: Payload (0..3), ReadReq (4..7)");
449 
450 uint aspm_mode = ASPM_MODE_DISABLED;
451 module_param_named(aspm, aspm_mode, uint, S_IRUGO);
452 MODULE_PARM_DESC(aspm, "PCIe ASPM: 0: disable, 1: enable, 2: dynamic");
453 
454 static void tune_pcie_caps(struct hfi1_devdata *dd)
455 {
456 	struct pci_dev *parent;
457 	u16 rc_mpss, rc_mps, ep_mpss, ep_mps;
458 	u16 rc_mrrs, ep_mrrs, max_mrrs, ectl;
459 
460 	/*
461 	 * Turn on extended tags in DevCtl in case the BIOS has turned it off
462 	 * to improve WFR SDMA bandwidth
463 	 */
464 	pcie_capability_read_word(dd->pcidev, PCI_EXP_DEVCTL, &ectl);
465 	if (!(ectl & PCI_EXP_DEVCTL_EXT_TAG)) {
466 		dd_dev_info(dd, "Enabling PCIe extended tags\n");
467 		ectl |= PCI_EXP_DEVCTL_EXT_TAG;
468 		pcie_capability_write_word(dd->pcidev, PCI_EXP_DEVCTL, ectl);
469 	}
470 	/* Find out supported and configured values for parent (root) */
471 	parent = dd->pcidev->bus->self;
472 	/*
473 	 * The driver cannot perform the tuning if it does not have
474 	 * access to the upstream component.
475 	 */
476 	if (!parent)
477 		return;
478 	if (!pci_is_root_bus(parent->bus)) {
479 		dd_dev_info(dd, "Parent not root\n");
480 		return;
481 	}
482 
483 	if (!pci_is_pcie(parent) || !pci_is_pcie(dd->pcidev))
484 		return;
485 	rc_mpss = parent->pcie_mpss;
486 	rc_mps = ffs(pcie_get_mps(parent)) - 8;
487 	/* Find out supported and configured values for endpoint (us) */
488 	ep_mpss = dd->pcidev->pcie_mpss;
489 	ep_mps = ffs(pcie_get_mps(dd->pcidev)) - 8;
490 
491 	/* Find max payload supported by root, endpoint */
492 	if (rc_mpss > ep_mpss)
493 		rc_mpss = ep_mpss;
494 
495 	/* If Supported greater than limit in module param, limit it */
496 	if (rc_mpss > (hfi1_pcie_caps & 7))
497 		rc_mpss = hfi1_pcie_caps & 7;
498 	/* If less than (allowed, supported), bump root payload */
499 	if (rc_mpss > rc_mps) {
500 		rc_mps = rc_mpss;
501 		pcie_set_mps(parent, 128 << rc_mps);
502 	}
503 	/* If less than (allowed, supported), bump endpoint payload */
504 	if (rc_mpss > ep_mps) {
505 		ep_mps = rc_mpss;
506 		pcie_set_mps(dd->pcidev, 128 << ep_mps);
507 	}
508 
509 	/*
510 	 * Now the Read Request size.
511 	 * No field for max supported, but PCIe spec limits it to 4096,
512 	 * which is code '5' (log2(4096) - 7)
513 	 */
514 	max_mrrs = 5;
515 	if (max_mrrs > ((hfi1_pcie_caps >> 4) & 7))
516 		max_mrrs = (hfi1_pcie_caps >> 4) & 7;
517 
518 	max_mrrs = 128 << max_mrrs;
519 	rc_mrrs = pcie_get_readrq(parent);
520 	ep_mrrs = pcie_get_readrq(dd->pcidev);
521 
522 	if (max_mrrs > rc_mrrs) {
523 		rc_mrrs = max_mrrs;
524 		pcie_set_readrq(parent, rc_mrrs);
525 	}
526 	if (max_mrrs > ep_mrrs) {
527 		ep_mrrs = max_mrrs;
528 		pcie_set_readrq(dd->pcidev, ep_mrrs);
529 	}
530 }
531 
532 /* End of PCIe capability tuning */
533 
534 /*
535  * From here through hfi1_pci_err_handler definition is invoked via
536  * PCI error infrastructure, registered via pci
537  */
538 static pci_ers_result_t
539 pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
540 {
541 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
542 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
543 
544 	switch (state) {
545 	case pci_channel_io_normal:
546 		dd_dev_info(dd, "State Normal, ignoring\n");
547 		break;
548 
549 	case pci_channel_io_frozen:
550 		dd_dev_info(dd, "State Frozen, requesting reset\n");
551 		pci_disable_device(pdev);
552 		ret = PCI_ERS_RESULT_NEED_RESET;
553 		break;
554 
555 	case pci_channel_io_perm_failure:
556 		if (dd) {
557 			dd_dev_info(dd, "State Permanent Failure, disabling\n");
558 			/* no more register accesses! */
559 			dd->flags &= ~HFI1_PRESENT;
560 			hfi1_disable_after_error(dd);
561 		}
562 		 /* else early, or other problem */
563 		ret =  PCI_ERS_RESULT_DISCONNECT;
564 		break;
565 
566 	default: /* shouldn't happen */
567 		dd_dev_info(dd, "HFI1 PCI errors detected (state %d)\n",
568 			    state);
569 		break;
570 	}
571 	return ret;
572 }
573 
574 static pci_ers_result_t
575 pci_mmio_enabled(struct pci_dev *pdev)
576 {
577 	u64 words = 0U;
578 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
579 	pci_ers_result_t ret = PCI_ERS_RESULT_RECOVERED;
580 
581 	if (dd && dd->pport) {
582 		words = read_port_cntr(dd->pport, C_RX_WORDS, CNTR_INVALID_VL);
583 		if (words == ~0ULL)
584 			ret = PCI_ERS_RESULT_NEED_RESET;
585 		dd_dev_info(dd,
586 			    "HFI1 mmio_enabled function called, read wordscntr %Lx, returning %d\n",
587 			    words, ret);
588 	}
589 	return  ret;
590 }
591 
592 static pci_ers_result_t
593 pci_slot_reset(struct pci_dev *pdev)
594 {
595 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
596 
597 	dd_dev_info(dd, "HFI1 slot_reset function called, ignored\n");
598 	return PCI_ERS_RESULT_CAN_RECOVER;
599 }
600 
601 static pci_ers_result_t
602 pci_link_reset(struct pci_dev *pdev)
603 {
604 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
605 
606 	dd_dev_info(dd, "HFI1 link_reset function called, ignored\n");
607 	return PCI_ERS_RESULT_CAN_RECOVER;
608 }
609 
610 static void
611 pci_resume(struct pci_dev *pdev)
612 {
613 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
614 
615 	dd_dev_info(dd, "HFI1 resume function called\n");
616 	pci_cleanup_aer_uncorrect_error_status(pdev);
617 	/*
618 	 * Running jobs will fail, since it's asynchronous
619 	 * unlike sysfs-requested reset.   Better than
620 	 * doing nothing.
621 	 */
622 	hfi1_init(dd, 1); /* same as re-init after reset */
623 }
624 
625 const struct pci_error_handlers hfi1_pci_err_handler = {
626 	.error_detected = pci_error_detected,
627 	.mmio_enabled = pci_mmio_enabled,
628 	.link_reset = pci_link_reset,
629 	.slot_reset = pci_slot_reset,
630 	.resume = pci_resume,
631 };
632 
633 /*============================================================================*/
634 /* PCIe Gen3 support */
635 
636 /*
637  * This code is separated out because it is expected to be removed in the
638  * final shipping product.  If not, then it will be revisited and items
639  * will be moved to more standard locations.
640  */
641 
642 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_STS field values */
643 #define DL_STATUS_HFI0 0x1	/* hfi0 firmware download complete */
644 #define DL_STATUS_HFI1 0x2	/* hfi1 firmware download complete */
645 #define DL_STATUS_BOTH 0x3	/* hfi0 and hfi1 firmware download complete */
646 
647 /* ASIC_PCI_SD_HOST_STATUS.FW_DNLD_ERR field values */
648 #define DL_ERR_NONE		0x0	/* no error */
649 #define DL_ERR_SWAP_PARITY	0x1	/* parity error in SerDes interrupt */
650 					/*   or response data */
651 #define DL_ERR_DISABLED	0x2	/* hfi disabled */
652 #define DL_ERR_SECURITY	0x3	/* security check failed */
653 #define DL_ERR_SBUS		0x4	/* SBus status error */
654 #define DL_ERR_XFR_PARITY	0x5	/* parity error during ROM transfer*/
655 
656 /* gasket block secondary bus reset delay */
657 #define SBR_DELAY_US 200000	/* 200ms */
658 
659 /* mask for PCIe capability register lnkctl2 target link speed */
660 #define LNKCTL2_TARGET_LINK_SPEED_MASK 0xf
661 
662 static uint pcie_target = 3;
663 module_param(pcie_target, uint, S_IRUGO);
664 MODULE_PARM_DESC(pcie_target, "PCIe target speed (0 skip, 1-3 Gen1-3)");
665 
666 static uint pcie_force;
667 module_param(pcie_force, uint, S_IRUGO);
668 MODULE_PARM_DESC(pcie_force, "Force driver to do a PCIe firmware download even if already at target speed");
669 
670 static uint pcie_retry = 5;
671 module_param(pcie_retry, uint, S_IRUGO);
672 MODULE_PARM_DESC(pcie_retry, "Driver will try this many times to reach requested speed");
673 
674 #define UNSET_PSET 255
675 #define DEFAULT_DISCRETE_PSET 2	/* discrete HFI */
676 #define DEFAULT_MCP_PSET 4	/* MCP HFI */
677 static uint pcie_pset = UNSET_PSET;
678 module_param(pcie_pset, uint, S_IRUGO);
679 MODULE_PARM_DESC(pcie_pset, "PCIe Eq Pset value to use, range is 0-10");
680 
681 static uint pcie_ctle = 1; /* discrete on, integrated off */
682 module_param(pcie_ctle, uint, S_IRUGO);
683 MODULE_PARM_DESC(pcie_ctle, "PCIe static CTLE mode, bit 0 - discrete on/off, bit 1 - integrated on/off");
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 static const u8 discrete_ctle_tunings[11][4] = {
723 	/* DC     LF     HF     BW */
724 	{  0x48,  0x0b,  0x04,  0x04 },	/* p0 */
725 	{  0x60,  0x05,  0x0f,  0x0a },	/* p1 */
726 	{  0x50,  0x09,  0x06,  0x06 },	/* p2 */
727 	{  0x68,  0x05,  0x0f,  0x0a },	/* p3 */
728 	{  0x80,  0x05,  0x0f,  0x0a },	/* p4 */
729 	{  0x70,  0x05,  0x0f,  0x0a },	/* p5 */
730 	{  0x68,  0x05,  0x0f,  0x0a },	/* p6 */
731 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
732 	{  0x48,  0x09,  0x06,  0x06 },	/* p8 */
733 	{  0x60,  0x05,  0x0f,  0x0a },	/* p9 */
734 	{  0x38,  0x0f,  0x00,  0x00 },	/* p10 */
735 };
736 
737 static const u8 integrated_ctle_tunings[11][4] = {
738 	/* DC     LF     HF     BW */
739 	{  0x38,  0x0f,  0x00,  0x00 },	/* p0 */
740 	{  0x38,  0x0f,  0x00,  0x00 },	/* p1 */
741 	{  0x38,  0x0f,  0x00,  0x00 },	/* p2 */
742 	{  0x38,  0x0f,  0x00,  0x00 },	/* p3 */
743 	{  0x58,  0x0a,  0x05,  0x05 },	/* p4 */
744 	{  0x48,  0x0a,  0x05,  0x05 },	/* p5 */
745 	{  0x40,  0x0a,  0x05,  0x05 },	/* p6 */
746 	{  0x38,  0x0f,  0x00,  0x00 },	/* p7 */
747 	{  0x38,  0x0f,  0x00,  0x00 },	/* p8 */
748 	{  0x38,  0x09,  0x06,  0x06 },	/* p9 */
749 	{  0x38,  0x0e,  0x01,  0x01 },	/* p10 */
750 };
751 
752 /* helper to format the value to write to hardware */
753 #define eq_value(pre, curr, post) \
754 	((((u32)(pre)) << \
755 			PCIE_CFG_REG_PL102_GEN3_EQ_PRE_CURSOR_PSET_SHIFT) \
756 	| (((u32)(curr)) << PCIE_CFG_REG_PL102_GEN3_EQ_CURSOR_PSET_SHIFT) \
757 	| (((u32)(post)) << \
758 		PCIE_CFG_REG_PL102_GEN3_EQ_POST_CURSOR_PSET_SHIFT))
759 
760 /*
761  * Load the given EQ preset table into the PCIe hardware.
762  */
763 static int load_eq_table(struct hfi1_devdata *dd, const u8 eq[11][3], u8 fs,
764 			 u8 div)
765 {
766 	struct pci_dev *pdev = dd->pcidev;
767 	u32 hit_error = 0;
768 	u32 violation;
769 	u32 i;
770 	u8 c_minus1, c0, c_plus1;
771 
772 	for (i = 0; i < 11; i++) {
773 		/* set index */
774 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL103, i);
775 		/* write the value */
776 		c_minus1 = eq[i][PREC] / div;
777 		c0 = fs - (eq[i][PREC] / div) - (eq[i][POST] / div);
778 		c_plus1 = eq[i][POST] / div;
779 		pci_write_config_dword(pdev, PCIE_CFG_REG_PL102,
780 				       eq_value(c_minus1, c0, c_plus1));
781 		/* check if these coefficients violate EQ rules */
782 		pci_read_config_dword(dd->pcidev, PCIE_CFG_REG_PL105,
783 				      &violation);
784 		if (violation
785 		    & PCIE_CFG_REG_PL105_GEN3_EQ_VIOLATE_COEF_RULES_SMASK){
786 			if (hit_error == 0) {
787 				dd_dev_err(dd,
788 					   "Gen3 EQ Table Coefficient rule violations\n");
789 				dd_dev_err(dd, "         prec   attn   post\n");
790 			}
791 			dd_dev_err(dd, "   p%02d:   %02x     %02x     %02x\n",
792 				   i, (u32)eq[i][0], (u32)eq[i][1],
793 				   (u32)eq[i][2]);
794 			dd_dev_err(dd, "            %02x     %02x     %02x\n",
795 				   (u32)c_minus1, (u32)c0, (u32)c_plus1);
796 			hit_error = 1;
797 		}
798 	}
799 	if (hit_error)
800 		return -EINVAL;
801 	return 0;
802 }
803 
804 /*
805  * Steps to be done after the PCIe firmware is downloaded and
806  * before the SBR for the Pcie Gen3.
807  * The SBus resource is already being held.
808  */
809 static void pcie_post_steps(struct hfi1_devdata *dd)
810 {
811 	int i;
812 
813 	set_sbus_fast_mode(dd);
814 	/*
815 	 * Write to the PCIe PCSes to set the G3_LOCKED_NEXT bits to 1.
816 	 * This avoids a spurious framing error that can otherwise be
817 	 * generated by the MAC layer.
818 	 *
819 	 * Use individual addresses since no broadcast is set up.
820 	 */
821 	for (i = 0; i < NUM_PCIE_SERDES; i++) {
822 		sbus_request(dd, pcie_pcs_addrs[dd->hfi1_id][i],
823 			     0x03, WRITE_SBUS_RECEIVER, 0x00022132);
824 	}
825 
826 	clear_sbus_fast_mode(dd);
827 }
828 
829 /*
830  * Trigger a secondary bus reset (SBR) on ourselves using our parent.
831  *
832  * Based on pci_parent_bus_reset() which is not exported by the
833  * kernel core.
834  */
835 static int trigger_sbr(struct hfi1_devdata *dd)
836 {
837 	struct pci_dev *dev = dd->pcidev;
838 	struct pci_dev *pdev;
839 
840 	/* need a parent */
841 	if (!dev->bus->self) {
842 		dd_dev_err(dd, "%s: no parent device\n", __func__);
843 		return -ENOTTY;
844 	}
845 
846 	/* should not be anyone else on the bus */
847 	list_for_each_entry(pdev, &dev->bus->devices, bus_list)
848 		if (pdev != dev) {
849 			dd_dev_err(dd,
850 				   "%s: another device is on the same bus\n",
851 				   __func__);
852 			return -ENOTTY;
853 		}
854 
855 	/*
856 	 * A secondary bus reset (SBR) issues a hot reset to our device.
857 	 * The following routine does a 1s wait after the reset is dropped
858 	 * per PCI Trhfa (recovery time).  PCIe 3.0 section 6.6.1 -
859 	 * Conventional Reset, paragraph 3, line 35 also says that a 1s
860 	 * delay after a reset is required.  Per spec requirements,
861 	 * the link is either working or not after that point.
862 	 */
863 	pci_reset_bridge_secondary_bus(dev->bus->self);
864 
865 	return 0;
866 }
867 
868 /*
869  * Write the given gasket interrupt register.
870  */
871 static void write_gasket_interrupt(struct hfi1_devdata *dd, int index,
872 				   u16 code, u16 data)
873 {
874 	write_csr(dd, ASIC_PCIE_SD_INTRPT_LIST + (index * 8),
875 		  (((u64)code << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_CODE_SHIFT) |
876 		   ((u64)data << ASIC_PCIE_SD_INTRPT_LIST_INTRPT_DATA_SHIFT)));
877 }
878 
879 /*
880  * Tell the gasket logic how to react to the reset.
881  */
882 static void arm_gasket_logic(struct hfi1_devdata *dd)
883 {
884 	u64 reg;
885 
886 	reg = (((u64)1 << dd->hfi1_id) <<
887 	       ASIC_PCIE_SD_HOST_CMD_INTRPT_CMD_SHIFT) |
888 	      ((u64)pcie_serdes_broadcast[dd->hfi1_id] <<
889 	       ASIC_PCIE_SD_HOST_CMD_SBUS_RCVR_ADDR_SHIFT |
890 	       ASIC_PCIE_SD_HOST_CMD_SBR_MODE_SMASK |
891 	       ((u64)SBR_DELAY_US & ASIC_PCIE_SD_HOST_CMD_TIMER_MASK) <<
892 	       ASIC_PCIE_SD_HOST_CMD_TIMER_SHIFT);
893 	write_csr(dd, ASIC_PCIE_SD_HOST_CMD, reg);
894 	/* read back to push the write */
895 	read_csr(dd, ASIC_PCIE_SD_HOST_CMD);
896 }
897 
898 /*
899  * CCE_PCIE_CTRL long name helpers
900  * We redefine these shorter macros to use in the code while leaving
901  * chip_registers.h to be autogenerated from the hardware spec.
902  */
903 #define LANE_BUNDLE_MASK              CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_MASK
904 #define LANE_BUNDLE_SHIFT             CCE_PCIE_CTRL_PCIE_LANE_BUNDLE_SHIFT
905 #define LANE_DELAY_MASK               CCE_PCIE_CTRL_PCIE_LANE_DELAY_MASK
906 #define LANE_DELAY_SHIFT              CCE_PCIE_CTRL_PCIE_LANE_DELAY_SHIFT
907 #define MARGIN_OVERWRITE_ENABLE_SHIFT CCE_PCIE_CTRL_XMT_MARGIN_OVERWRITE_ENABLE_SHIFT
908 #define MARGIN_SHIFT                  CCE_PCIE_CTRL_XMT_MARGIN_SHIFT
909 #define MARGIN_G1_G2_OVERWRITE_MASK   CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_MASK
910 #define MARGIN_G1_G2_OVERWRITE_SHIFT  CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_OVERWRITE_ENABLE_SHIFT
911 #define MARGIN_GEN1_GEN2_MASK         CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_MASK
912 #define MARGIN_GEN1_GEN2_SHIFT        CCE_PCIE_CTRL_XMT_MARGIN_GEN1_GEN2_SHIFT
913 
914  /*
915   * Write xmt_margin for full-swing (WFR-B) or half-swing (WFR-C).
916   */
917 static void write_xmt_margin(struct hfi1_devdata *dd, const char *fname)
918 {
919 	u64 pcie_ctrl;
920 	u64 xmt_margin;
921 	u64 xmt_margin_oe;
922 	u64 lane_delay;
923 	u64 lane_bundle;
924 
925 	pcie_ctrl = read_csr(dd, CCE_PCIE_CTRL);
926 
927 	/*
928 	 * For Discrete, use full-swing.
929 	 *  - PCIe TX defaults to full-swing.
930 	 *    Leave this register as default.
931 	 * For Integrated, use half-swing
932 	 *  - Copy xmt_margin and xmt_margin_oe
933 	 *    from Gen1/Gen2 to Gen3.
934 	 */
935 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL1) { /* integrated */
936 		/* extract initial fields */
937 		xmt_margin = (pcie_ctrl >> MARGIN_GEN1_GEN2_SHIFT)
938 			      & MARGIN_GEN1_GEN2_MASK;
939 		xmt_margin_oe = (pcie_ctrl >> MARGIN_G1_G2_OVERWRITE_SHIFT)
940 				 & MARGIN_G1_G2_OVERWRITE_MASK;
941 		lane_delay = (pcie_ctrl >> LANE_DELAY_SHIFT) & LANE_DELAY_MASK;
942 		lane_bundle = (pcie_ctrl >> LANE_BUNDLE_SHIFT)
943 			       & LANE_BUNDLE_MASK;
944 
945 		/*
946 		 * For A0, EFUSE values are not set.  Override with the
947 		 * correct values.
948 		 */
949 		if (is_ax(dd)) {
950 			/*
951 			 * xmt_margin and OverwiteEnabel should be the
952 			 * same for Gen1/Gen2 and Gen3
953 			 */
954 			xmt_margin = 0x5;
955 			xmt_margin_oe = 0x1;
956 			lane_delay = 0xF; /* Delay 240ns. */
957 			lane_bundle = 0x0; /* Set to 1 lane. */
958 		}
959 
960 		/* overwrite existing values */
961 		pcie_ctrl = (xmt_margin << MARGIN_GEN1_GEN2_SHIFT)
962 			| (xmt_margin_oe << MARGIN_G1_G2_OVERWRITE_SHIFT)
963 			| (xmt_margin << MARGIN_SHIFT)
964 			| (xmt_margin_oe << MARGIN_OVERWRITE_ENABLE_SHIFT)
965 			| (lane_delay << LANE_DELAY_SHIFT)
966 			| (lane_bundle << LANE_BUNDLE_SHIFT);
967 
968 		write_csr(dd, CCE_PCIE_CTRL, pcie_ctrl);
969 	}
970 
971 	dd_dev_dbg(dd, "%s: program XMT margin, CcePcieCtrl 0x%llx\n",
972 		   fname, pcie_ctrl);
973 }
974 
975 /*
976  * Do all the steps needed to transition the PCIe link to Gen3 speed.
977  */
978 int do_pcie_gen3_transition(struct hfi1_devdata *dd)
979 {
980 	struct pci_dev *parent = dd->pcidev->bus->self;
981 	u64 fw_ctrl;
982 	u64 reg, therm;
983 	u32 reg32, fs, lf;
984 	u32 status, err;
985 	int ret;
986 	int do_retry, retry_count = 0;
987 	int intnum = 0;
988 	uint default_pset;
989 	u16 target_vector, target_speed;
990 	u16 lnkctl2, vendor;
991 	u8 div;
992 	const u8 (*eq)[3];
993 	const u8 (*ctle_tunings)[4];
994 	uint static_ctle_mode;
995 	int return_error = 0;
996 
997 	/* PCIe Gen3 is for the ASIC only */
998 	if (dd->icode != ICODE_RTL_SILICON)
999 		return 0;
1000 
1001 	if (pcie_target == 1) {			/* target Gen1 */
1002 		target_vector = GEN1_SPEED_VECTOR;
1003 		target_speed = 2500;
1004 	} else if (pcie_target == 2) {		/* target Gen2 */
1005 		target_vector = GEN2_SPEED_VECTOR;
1006 		target_speed = 5000;
1007 	} else if (pcie_target == 3) {		/* target Gen3 */
1008 		target_vector = GEN3_SPEED_VECTOR;
1009 		target_speed = 8000;
1010 	} else {
1011 		/* off or invalid target - skip */
1012 		dd_dev_info(dd, "%s: Skipping PCIe transition\n", __func__);
1013 		return 0;
1014 	}
1015 
1016 	/* if already at target speed, done (unless forced) */
1017 	if (dd->lbus_speed == target_speed) {
1018 		dd_dev_info(dd, "%s: PCIe already at gen%d, %s\n", __func__,
1019 			    pcie_target,
1020 			    pcie_force ? "re-doing anyway" : "skipping");
1021 		if (!pcie_force)
1022 			return 0;
1023 	}
1024 
1025 	/*
1026 	 * The driver cannot do the transition if it has no access to the
1027 	 * upstream component
1028 	 */
1029 	if (!parent) {
1030 		dd_dev_info(dd, "%s: No upstream, Can't do gen3 transition\n",
1031 			    __func__);
1032 		return 0;
1033 	}
1034 
1035 	/*
1036 	 * Do the Gen3 transition.  Steps are those of the PCIe Gen3
1037 	 * recipe.
1038 	 */
1039 
1040 	/* step 1: pcie link working in gen1/gen2 */
1041 
1042 	/* step 2: if either side is not capable of Gen3, done */
1043 	if (pcie_target == 3 && !dd->link_gen3_capable) {
1044 		dd_dev_err(dd, "The PCIe link is not Gen3 capable\n");
1045 		ret = -ENOSYS;
1046 		goto done_no_mutex;
1047 	}
1048 
1049 	/* hold the SBus resource across the firmware download and SBR */
1050 	ret = acquire_chip_resource(dd, CR_SBUS, SBUS_TIMEOUT);
1051 	if (ret) {
1052 		dd_dev_err(dd, "%s: unable to acquire SBus resource\n",
1053 			   __func__);
1054 		return ret;
1055 	}
1056 
1057 	/* make sure thermal polling is not causing interrupts */
1058 	therm = read_csr(dd, ASIC_CFG_THERM_POLL_EN);
1059 	if (therm) {
1060 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x0);
1061 		msleep(100);
1062 		dd_dev_info(dd, "%s: Disabled therm polling\n",
1063 			    __func__);
1064 	}
1065 
1066 retry:
1067 	/* the SBus download will reset the spico for thermal */
1068 
1069 	/* step 3: download SBus Master firmware */
1070 	/* step 4: download PCIe Gen3 SerDes firmware */
1071 	dd_dev_info(dd, "%s: downloading firmware\n", __func__);
1072 	ret = load_pcie_firmware(dd);
1073 	if (ret) {
1074 		/* do not proceed if the firmware cannot be downloaded */
1075 		return_error = 1;
1076 		goto done;
1077 	}
1078 
1079 	/* step 5: set up device parameter settings */
1080 	dd_dev_info(dd, "%s: setting PCIe registers\n", __func__);
1081 
1082 	/*
1083 	 * PcieCfgSpcie1 - Link Control 3
1084 	 * Leave at reset value.  No need to set PerfEq - link equalization
1085 	 * will be performed automatically after the SBR when the target
1086 	 * speed is 8GT/s.
1087 	 */
1088 
1089 	/* clear all 16 per-lane error bits (PCIe: Lane Error Status) */
1090 	pci_write_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, 0xffff);
1091 
1092 	/* step 5a: Set Synopsys Port Logic registers */
1093 
1094 	/*
1095 	 * PcieCfgRegPl2 - Port Force Link
1096 	 *
1097 	 * Set the low power field to 0x10 to avoid unnecessary power
1098 	 * management messages.  All other fields are zero.
1099 	 */
1100 	reg32 = 0x10ul << PCIE_CFG_REG_PL2_LOW_PWR_ENT_CNT_SHIFT;
1101 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL2, reg32);
1102 
1103 	/*
1104 	 * PcieCfgRegPl100 - Gen3 Control
1105 	 *
1106 	 * turn off PcieCfgRegPl100.Gen3ZRxDcNonCompl
1107 	 * turn on PcieCfgRegPl100.EqEieosCnt
1108 	 * Everything else zero.
1109 	 */
1110 	reg32 = PCIE_CFG_REG_PL100_EQ_EIEOS_CNT_SMASK;
1111 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL100, reg32);
1112 
1113 	/*
1114 	 * PcieCfgRegPl101 - Gen3 EQ FS and LF
1115 	 * PcieCfgRegPl102 - Gen3 EQ Presets to Coefficients Mapping
1116 	 * PcieCfgRegPl103 - Gen3 EQ Preset Index
1117 	 * PcieCfgRegPl105 - Gen3 EQ Status
1118 	 *
1119 	 * Give initial EQ settings.
1120 	 */
1121 	if (dd->pcidev->device == PCI_DEVICE_ID_INTEL0) { /* discrete */
1122 		/* 1000mV, FS=24, LF = 8 */
1123 		fs = 24;
1124 		lf = 8;
1125 		div = 3;
1126 		eq = discrete_preliminary_eq;
1127 		default_pset = DEFAULT_DISCRETE_PSET;
1128 		ctle_tunings = discrete_ctle_tunings;
1129 		/* bit 0 - discrete on/off */
1130 		static_ctle_mode = pcie_ctle & 0x1;
1131 	} else {
1132 		/* 400mV, FS=29, LF = 9 */
1133 		fs = 29;
1134 		lf = 9;
1135 		div = 1;
1136 		eq = integrated_preliminary_eq;
1137 		default_pset = DEFAULT_MCP_PSET;
1138 		ctle_tunings = integrated_ctle_tunings;
1139 		/* bit 1 - integrated on/off */
1140 		static_ctle_mode = (pcie_ctle >> 1) & 0x1;
1141 	}
1142 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL101,
1143 			       (fs <<
1144 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_FS_SHIFT) |
1145 			       (lf <<
1146 				PCIE_CFG_REG_PL101_GEN3_EQ_LOCAL_LF_SHIFT));
1147 	ret = load_eq_table(dd, eq, fs, div);
1148 	if (ret)
1149 		goto done;
1150 
1151 	/*
1152 	 * PcieCfgRegPl106 - Gen3 EQ Control
1153 	 *
1154 	 * Set Gen3EqPsetReqVec, leave other fields 0.
1155 	 */
1156 	if (pcie_pset == UNSET_PSET)
1157 		pcie_pset = default_pset;
1158 	if (pcie_pset > 10) {	/* valid range is 0-10, inclusive */
1159 		dd_dev_err(dd, "%s: Invalid Eq Pset %u, setting to %d\n",
1160 			   __func__, pcie_pset, default_pset);
1161 		pcie_pset = default_pset;
1162 	}
1163 	dd_dev_info(dd, "%s: using EQ Pset %u\n", __func__, pcie_pset);
1164 	pci_write_config_dword(dd->pcidev, PCIE_CFG_REG_PL106,
1165 			       ((1 << pcie_pset) <<
1166 			PCIE_CFG_REG_PL106_GEN3_EQ_PSET_REQ_VEC_SHIFT) |
1167 			PCIE_CFG_REG_PL106_GEN3_EQ_EVAL2MS_DISABLE_SMASK |
1168 			PCIE_CFG_REG_PL106_GEN3_EQ_PHASE23_EXIT_MODE_SMASK);
1169 
1170 	/*
1171 	 * step 5b: Do post firmware download steps via SBus
1172 	 */
1173 	dd_dev_info(dd, "%s: doing pcie post steps\n", __func__);
1174 	pcie_post_steps(dd);
1175 
1176 	/*
1177 	 * step 5c: Program gasket interrupts
1178 	 */
1179 	/* set the Rx Bit Rate to REFCLK ratio */
1180 	write_gasket_interrupt(dd, intnum++, 0x0006, 0x0050);
1181 	/* disable pCal for PCIe Gen3 RX equalization */
1182 	/* select adaptive or static CTLE */
1183 	write_gasket_interrupt(dd, intnum++, 0x0026,
1184 			       0x5b01 | (static_ctle_mode << 3));
1185 	/*
1186 	 * Enable iCal for PCIe Gen3 RX equalization, and set which
1187 	 * evaluation of RX_EQ_EVAL will launch the iCal procedure.
1188 	 */
1189 	write_gasket_interrupt(dd, intnum++, 0x0026, 0x5202);
1190 
1191 	if (static_ctle_mode) {
1192 		/* apply static CTLE tunings */
1193 		u8 pcie_dc, pcie_lf, pcie_hf, pcie_bw;
1194 
1195 		pcie_dc = ctle_tunings[pcie_pset][0];
1196 		pcie_lf = ctle_tunings[pcie_pset][1];
1197 		pcie_hf = ctle_tunings[pcie_pset][2];
1198 		pcie_bw = ctle_tunings[pcie_pset][3];
1199 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0200 | pcie_dc);
1200 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0100 | pcie_lf);
1201 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x0000 | pcie_hf);
1202 		write_gasket_interrupt(dd, intnum++, 0x0026, 0x5500 | pcie_bw);
1203 	}
1204 
1205 	/* terminate list */
1206 	write_gasket_interrupt(dd, intnum++, 0x0000, 0x0000);
1207 
1208 	/*
1209 	 * step 5d: program XMT margin
1210 	 */
1211 	write_xmt_margin(dd, __func__);
1212 
1213 	/*
1214 	 * step 5e: disable active state power management (ASPM). It
1215 	 * will be enabled if required later
1216 	 */
1217 	dd_dev_info(dd, "%s: clearing ASPM\n", __func__);
1218 	aspm_hw_disable_l1(dd);
1219 
1220 	/*
1221 	 * step 5f: clear DirectSpeedChange
1222 	 * PcieCfgRegPl67.DirectSpeedChange must be zero to prevent the
1223 	 * change in the speed target from starting before we are ready.
1224 	 * This field defaults to 0 and we are not changing it, so nothing
1225 	 * needs to be done.
1226 	 */
1227 
1228 	/* step 5g: Set target link speed */
1229 	/*
1230 	 * Set target link speed to be target on both device and parent.
1231 	 * On setting the parent: Some system BIOSs "helpfully" set the
1232 	 * parent target speed to Gen2 to match the ASIC's initial speed.
1233 	 * We can set the target Gen3 because we have already checked
1234 	 * that it is Gen3 capable earlier.
1235 	 */
1236 	dd_dev_info(dd, "%s: setting parent target link speed\n", __func__);
1237 	pcie_capability_read_word(parent, PCI_EXP_LNKCTL2, &lnkctl2);
1238 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1239 		    (u32)lnkctl2);
1240 	/* only write to parent if target is not as high as ours */
1241 	if ((lnkctl2 & LNKCTL2_TARGET_LINK_SPEED_MASK) < target_vector) {
1242 		lnkctl2 &= ~LNKCTL2_TARGET_LINK_SPEED_MASK;
1243 		lnkctl2 |= target_vector;
1244 		dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1245 			    (u32)lnkctl2);
1246 		pcie_capability_write_word(parent, PCI_EXP_LNKCTL2, lnkctl2);
1247 	} else {
1248 		dd_dev_info(dd, "%s: ..target speed is OK\n", __func__);
1249 	}
1250 
1251 	dd_dev_info(dd, "%s: setting target link speed\n", __func__);
1252 	pcie_capability_read_word(dd->pcidev, PCI_EXP_LNKCTL2, &lnkctl2);
1253 	dd_dev_info(dd, "%s: ..old link control2: 0x%x\n", __func__,
1254 		    (u32)lnkctl2);
1255 	lnkctl2 &= ~LNKCTL2_TARGET_LINK_SPEED_MASK;
1256 	lnkctl2 |= target_vector;
1257 	dd_dev_info(dd, "%s: ..new link control2: 0x%x\n", __func__,
1258 		    (u32)lnkctl2);
1259 	pcie_capability_write_word(dd->pcidev, PCI_EXP_LNKCTL2, lnkctl2);
1260 
1261 	/* step 5h: arm gasket logic */
1262 	/* hold DC in reset across the SBR */
1263 	write_csr(dd, CCE_DC_CTRL, CCE_DC_CTRL_DC_RESET_SMASK);
1264 	(void)read_csr(dd, CCE_DC_CTRL); /* DC reset hold */
1265 	/* save firmware control across the SBR */
1266 	fw_ctrl = read_csr(dd, MISC_CFG_FW_CTRL);
1267 
1268 	dd_dev_info(dd, "%s: arming gasket logic\n", __func__);
1269 	arm_gasket_logic(dd);
1270 
1271 	/*
1272 	 * step 6: quiesce PCIe link
1273 	 * The chip has already been reset, so there will be no traffic
1274 	 * from the chip.  Linux has no easy way to enforce that it will
1275 	 * not try to access the device, so we just need to hope it doesn't
1276 	 * do it while we are doing the reset.
1277 	 */
1278 
1279 	/*
1280 	 * step 7: initiate the secondary bus reset (SBR)
1281 	 * step 8: hardware brings the links back up
1282 	 * step 9: wait for link speed transition to be complete
1283 	 */
1284 	dd_dev_info(dd, "%s: calling trigger_sbr\n", __func__);
1285 	ret = trigger_sbr(dd);
1286 	if (ret)
1287 		goto done;
1288 
1289 	/* step 10: decide what to do next */
1290 
1291 	/* check if we can read PCI space */
1292 	ret = pci_read_config_word(dd->pcidev, PCI_VENDOR_ID, &vendor);
1293 	if (ret) {
1294 		dd_dev_info(dd,
1295 			    "%s: read of VendorID failed after SBR, err %d\n",
1296 			    __func__, ret);
1297 		return_error = 1;
1298 		goto done;
1299 	}
1300 	if (vendor == 0xffff) {
1301 		dd_dev_info(dd, "%s: VendorID is all 1s after SBR\n", __func__);
1302 		return_error = 1;
1303 		ret = -EIO;
1304 		goto done;
1305 	}
1306 
1307 	/* restore PCI space registers we know were reset */
1308 	dd_dev_info(dd, "%s: calling restore_pci_variables\n", __func__);
1309 	restore_pci_variables(dd);
1310 	/* restore firmware control */
1311 	write_csr(dd, MISC_CFG_FW_CTRL, fw_ctrl);
1312 
1313 	/*
1314 	 * Check the gasket block status.
1315 	 *
1316 	 * This is the first CSR read after the SBR.  If the read returns
1317 	 * all 1s (fails), the link did not make it back.
1318 	 *
1319 	 * Once we're sure we can read and write, clear the DC reset after
1320 	 * the SBR.  Then check for any per-lane errors. Then look over
1321 	 * the status.
1322 	 */
1323 	reg = read_csr(dd, ASIC_PCIE_SD_HOST_STATUS);
1324 	dd_dev_info(dd, "%s: gasket block status: 0x%llx\n", __func__, reg);
1325 	if (reg == ~0ull) {	/* PCIe read failed/timeout */
1326 		dd_dev_err(dd, "SBR failed - unable to read from device\n");
1327 		return_error = 1;
1328 		ret = -ENOSYS;
1329 		goto done;
1330 	}
1331 
1332 	/* clear the DC reset */
1333 	write_csr(dd, CCE_DC_CTRL, 0);
1334 
1335 	/* Set the LED off */
1336 	setextled(dd, 0);
1337 
1338 	/* check for any per-lane errors */
1339 	pci_read_config_dword(dd->pcidev, PCIE_CFG_SPCIE2, &reg32);
1340 	dd_dev_info(dd, "%s: per-lane errors: 0x%x\n", __func__, reg32);
1341 
1342 	/* extract status, look for our HFI */
1343 	status = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_SHIFT)
1344 			& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_STS_MASK;
1345 	if ((status & (1 << dd->hfi1_id)) == 0) {
1346 		dd_dev_err(dd,
1347 			   "%s: gasket status 0x%x, expecting 0x%x\n",
1348 			   __func__, status, 1 << dd->hfi1_id);
1349 		ret = -EIO;
1350 		goto done;
1351 	}
1352 
1353 	/* extract error */
1354 	err = (reg >> ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_SHIFT)
1355 		& ASIC_PCIE_SD_HOST_STATUS_FW_DNLD_ERR_MASK;
1356 	if (err) {
1357 		dd_dev_err(dd, "%s: gasket error %d\n", __func__, err);
1358 		ret = -EIO;
1359 		goto done;
1360 	}
1361 
1362 	/* update our link information cache */
1363 	update_lbus_info(dd);
1364 	dd_dev_info(dd, "%s: new speed and width: %s\n", __func__,
1365 		    dd->lbus_info);
1366 
1367 	if (dd->lbus_speed != target_speed) { /* not target */
1368 		/* maybe retry */
1369 		do_retry = retry_count < pcie_retry;
1370 		dd_dev_err(dd, "PCIe link speed did not switch to Gen%d%s\n",
1371 			   pcie_target, do_retry ? ", retrying" : "");
1372 		retry_count++;
1373 		if (do_retry) {
1374 			msleep(100); /* allow time to settle */
1375 			goto retry;
1376 		}
1377 		ret = -EIO;
1378 	}
1379 
1380 done:
1381 	if (therm) {
1382 		write_csr(dd, ASIC_CFG_THERM_POLL_EN, 0x1);
1383 		msleep(100);
1384 		dd_dev_info(dd, "%s: Re-enable therm polling\n",
1385 			    __func__);
1386 	}
1387 	release_chip_resource(dd, CR_SBUS);
1388 done_no_mutex:
1389 	/* return no error if it is OK to be at current speed */
1390 	if (ret && !return_error) {
1391 		dd_dev_err(dd, "Proceeding at current speed PCIe speed\n");
1392 		ret = 0;
1393 	}
1394 
1395 	dd_dev_info(dd, "%s: done\n", __func__);
1396 	return ret;
1397 }
1398