xref: /openbmc/linux/drivers/infiniband/hw/hfi1/init.c (revision f5b06569)
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/netdevice.h>
50 #include <linux/vmalloc.h>
51 #include <linux/delay.h>
52 #include <linux/idr.h>
53 #include <linux/module.h>
54 #include <linux/printk.h>
55 #include <linux/hrtimer.h>
56 #include <rdma/rdma_vt.h>
57 
58 #include "hfi.h"
59 #include "device.h"
60 #include "common.h"
61 #include "trace.h"
62 #include "mad.h"
63 #include "sdma.h"
64 #include "debugfs.h"
65 #include "verbs.h"
66 #include "aspm.h"
67 #include "affinity.h"
68 
69 #undef pr_fmt
70 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
71 
72 /*
73  * min buffers we want to have per context, after driver
74  */
75 #define HFI1_MIN_USER_CTXT_BUFCNT 7
76 
77 #define HFI1_MIN_HDRQ_EGRBUF_CNT 2
78 #define HFI1_MAX_HDRQ_EGRBUF_CNT 16352
79 #define HFI1_MIN_EAGER_BUFFER_SIZE (4 * 1024) /* 4KB */
80 #define HFI1_MAX_EAGER_BUFFER_SIZE (256 * 1024) /* 256KB */
81 
82 /*
83  * Number of user receive contexts we are configured to use (to allow for more
84  * pio buffers per ctxt, etc.)  Zero means use one user context per CPU.
85  */
86 int num_user_contexts = -1;
87 module_param_named(num_user_contexts, num_user_contexts, uint, S_IRUGO);
88 MODULE_PARM_DESC(
89 	num_user_contexts, "Set max number of user contexts to use");
90 
91 uint krcvqs[RXE_NUM_DATA_VL];
92 int krcvqsset;
93 module_param_array(krcvqs, uint, &krcvqsset, S_IRUGO);
94 MODULE_PARM_DESC(krcvqs, "Array of the number of non-control kernel receive queues by VL");
95 
96 /* computed based on above array */
97 unsigned n_krcvqs;
98 
99 static unsigned hfi1_rcvarr_split = 25;
100 module_param_named(rcvarr_split, hfi1_rcvarr_split, uint, S_IRUGO);
101 MODULE_PARM_DESC(rcvarr_split, "Percent of context's RcvArray entries used for Eager buffers");
102 
103 static uint eager_buffer_size = (2 << 20); /* 2MB */
104 module_param(eager_buffer_size, uint, S_IRUGO);
105 MODULE_PARM_DESC(eager_buffer_size, "Size of the eager buffers, default: 2MB");
106 
107 static uint rcvhdrcnt = 2048; /* 2x the max eager buffer count */
108 module_param_named(rcvhdrcnt, rcvhdrcnt, uint, S_IRUGO);
109 MODULE_PARM_DESC(rcvhdrcnt, "Receive header queue count (default 2048)");
110 
111 static uint hfi1_hdrq_entsize = 32;
112 module_param_named(hdrq_entsize, hfi1_hdrq_entsize, uint, S_IRUGO);
113 MODULE_PARM_DESC(hdrq_entsize, "Size of header queue entries: 2 - 8B, 16 - 64B (default), 32 - 128B");
114 
115 unsigned int user_credit_return_threshold = 33;	/* default is 33% */
116 module_param(user_credit_return_threshold, uint, S_IRUGO);
117 MODULE_PARM_DESC(user_credit_return_threshold, "Credit return threshold for user send contexts, return when unreturned credits passes this many blocks (in percent of allocated blocks, 0 is off)");
118 
119 static inline u64 encode_rcv_header_entry_size(u16);
120 
121 static struct idr hfi1_unit_table;
122 u32 hfi1_cpulist_count;
123 unsigned long *hfi1_cpulist;
124 
125 /*
126  * Common code for creating the receive context array.
127  */
128 int hfi1_create_ctxts(struct hfi1_devdata *dd)
129 {
130 	unsigned i;
131 	int ret;
132 
133 	/* Control context has to be always 0 */
134 	BUILD_BUG_ON(HFI1_CTRL_CTXT != 0);
135 
136 	dd->rcd = kzalloc_node(dd->num_rcv_contexts * sizeof(*dd->rcd),
137 			       GFP_KERNEL, dd->node);
138 	if (!dd->rcd)
139 		goto nomem;
140 
141 	/* create one or more kernel contexts */
142 	for (i = 0; i < dd->first_user_ctxt; ++i) {
143 		struct hfi1_pportdata *ppd;
144 		struct hfi1_ctxtdata *rcd;
145 
146 		ppd = dd->pport + (i % dd->num_pports);
147 		rcd = hfi1_create_ctxtdata(ppd, i, dd->node);
148 		if (!rcd) {
149 			dd_dev_err(dd,
150 				   "Unable to allocate kernel receive context, failing\n");
151 			goto nomem;
152 		}
153 		/*
154 		 * Set up the kernel context flags here and now because they
155 		 * use default values for all receive side memories.  User
156 		 * contexts will be handled as they are created.
157 		 */
158 		rcd->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
159 			HFI1_CAP_KGET(NODROP_RHQ_FULL) |
160 			HFI1_CAP_KGET(NODROP_EGR_FULL) |
161 			HFI1_CAP_KGET(DMA_RTAIL);
162 
163 		/* Control context must use DMA_RTAIL */
164 		if (rcd->ctxt == HFI1_CTRL_CTXT)
165 			rcd->flags |= HFI1_CAP_DMA_RTAIL;
166 		rcd->seq_cnt = 1;
167 
168 		rcd->sc = sc_alloc(dd, SC_ACK, rcd->rcvhdrqentsize, dd->node);
169 		if (!rcd->sc) {
170 			dd_dev_err(dd,
171 				   "Unable to allocate kernel send context, failing\n");
172 			dd->rcd[rcd->ctxt] = NULL;
173 			hfi1_free_ctxtdata(dd, rcd);
174 			goto nomem;
175 		}
176 
177 		ret = hfi1_init_ctxt(rcd->sc);
178 		if (ret < 0) {
179 			dd_dev_err(dd,
180 				   "Failed to setup kernel receive context, failing\n");
181 			sc_free(rcd->sc);
182 			dd->rcd[rcd->ctxt] = NULL;
183 			hfi1_free_ctxtdata(dd, rcd);
184 			ret = -EFAULT;
185 			goto bail;
186 		}
187 	}
188 
189 	/*
190 	 * Initialize aspm, to be done after gen3 transition and setting up
191 	 * contexts and before enabling interrupts
192 	 */
193 	aspm_init(dd);
194 
195 	return 0;
196 nomem:
197 	ret = -ENOMEM;
198 bail:
199 	kfree(dd->rcd);
200 	dd->rcd = NULL;
201 	return ret;
202 }
203 
204 /*
205  * Common code for user and kernel context setup.
206  */
207 struct hfi1_ctxtdata *hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, u32 ctxt,
208 					   int numa)
209 {
210 	struct hfi1_devdata *dd = ppd->dd;
211 	struct hfi1_ctxtdata *rcd;
212 	unsigned kctxt_ngroups = 0;
213 	u32 base;
214 
215 	if (dd->rcv_entries.nctxt_extra >
216 	    dd->num_rcv_contexts - dd->first_user_ctxt)
217 		kctxt_ngroups = (dd->rcv_entries.nctxt_extra -
218 				 (dd->num_rcv_contexts - dd->first_user_ctxt));
219 	rcd = kzalloc(sizeof(*rcd), GFP_KERNEL);
220 	if (rcd) {
221 		u32 rcvtids, max_entries;
222 
223 		hfi1_cdbg(PROC, "setting up context %u\n", ctxt);
224 
225 		INIT_LIST_HEAD(&rcd->qp_wait_list);
226 		rcd->ppd = ppd;
227 		rcd->dd = dd;
228 		rcd->cnt = 1;
229 		rcd->ctxt = ctxt;
230 		dd->rcd[ctxt] = rcd;
231 		rcd->numa_id = numa;
232 		rcd->rcv_array_groups = dd->rcv_entries.ngroups;
233 
234 		mutex_init(&rcd->exp_lock);
235 
236 		/*
237 		 * Calculate the context's RcvArray entry starting point.
238 		 * We do this here because we have to take into account all
239 		 * the RcvArray entries that previous context would have
240 		 * taken and we have to account for any extra groups
241 		 * assigned to the kernel or user contexts.
242 		 */
243 		if (ctxt < dd->first_user_ctxt) {
244 			if (ctxt < kctxt_ngroups) {
245 				base = ctxt * (dd->rcv_entries.ngroups + 1);
246 				rcd->rcv_array_groups++;
247 			} else
248 				base = kctxt_ngroups +
249 					(ctxt * dd->rcv_entries.ngroups);
250 		} else {
251 			u16 ct = ctxt - dd->first_user_ctxt;
252 
253 			base = ((dd->n_krcv_queues * dd->rcv_entries.ngroups) +
254 				kctxt_ngroups);
255 			if (ct < dd->rcv_entries.nctxt_extra) {
256 				base += ct * (dd->rcv_entries.ngroups + 1);
257 				rcd->rcv_array_groups++;
258 			} else
259 				base += dd->rcv_entries.nctxt_extra +
260 					(ct * dd->rcv_entries.ngroups);
261 		}
262 		rcd->eager_base = base * dd->rcv_entries.group_size;
263 
264 		/* Validate and initialize Rcv Hdr Q variables */
265 		if (rcvhdrcnt % HDRQ_INCREMENT) {
266 			dd_dev_err(dd,
267 				   "ctxt%u: header queue count %d must be divisible by %lu\n",
268 				   rcd->ctxt, rcvhdrcnt, HDRQ_INCREMENT);
269 			goto bail;
270 		}
271 		rcd->rcvhdrq_cnt = rcvhdrcnt;
272 		rcd->rcvhdrqentsize = hfi1_hdrq_entsize;
273 		/*
274 		 * Simple Eager buffer allocation: we have already pre-allocated
275 		 * the number of RcvArray entry groups. Each ctxtdata structure
276 		 * holds the number of groups for that context.
277 		 *
278 		 * To follow CSR requirements and maintain cacheline alignment,
279 		 * make sure all sizes and bases are multiples of group_size.
280 		 *
281 		 * The expected entry count is what is left after assigning
282 		 * eager.
283 		 */
284 		max_entries = rcd->rcv_array_groups *
285 			dd->rcv_entries.group_size;
286 		rcvtids = ((max_entries * hfi1_rcvarr_split) / 100);
287 		rcd->egrbufs.count = round_down(rcvtids,
288 						dd->rcv_entries.group_size);
289 		if (rcd->egrbufs.count > MAX_EAGER_ENTRIES) {
290 			dd_dev_err(dd, "ctxt%u: requested too many RcvArray entries.\n",
291 				   rcd->ctxt);
292 			rcd->egrbufs.count = MAX_EAGER_ENTRIES;
293 		}
294 		hfi1_cdbg(PROC,
295 			  "ctxt%u: max Eager buffer RcvArray entries: %u\n",
296 			  rcd->ctxt, rcd->egrbufs.count);
297 
298 		/*
299 		 * Allocate array that will hold the eager buffer accounting
300 		 * data.
301 		 * This will allocate the maximum possible buffer count based
302 		 * on the value of the RcvArray split parameter.
303 		 * The resulting value will be rounded down to the closest
304 		 * multiple of dd->rcv_entries.group_size.
305 		 */
306 		rcd->egrbufs.buffers = kcalloc(rcd->egrbufs.count,
307 					       sizeof(*rcd->egrbufs.buffers),
308 					       GFP_KERNEL);
309 		if (!rcd->egrbufs.buffers)
310 			goto bail;
311 		rcd->egrbufs.rcvtids = kcalloc(rcd->egrbufs.count,
312 					       sizeof(*rcd->egrbufs.rcvtids),
313 					       GFP_KERNEL);
314 		if (!rcd->egrbufs.rcvtids)
315 			goto bail;
316 		rcd->egrbufs.size = eager_buffer_size;
317 		/*
318 		 * The size of the buffers programmed into the RcvArray
319 		 * entries needs to be big enough to handle the highest
320 		 * MTU supported.
321 		 */
322 		if (rcd->egrbufs.size < hfi1_max_mtu) {
323 			rcd->egrbufs.size = __roundup_pow_of_two(hfi1_max_mtu);
324 			hfi1_cdbg(PROC,
325 				  "ctxt%u: eager bufs size too small. Adjusting to %zu\n",
326 				    rcd->ctxt, rcd->egrbufs.size);
327 		}
328 		rcd->egrbufs.rcvtid_size = HFI1_MAX_EAGER_BUFFER_SIZE;
329 
330 		if (ctxt < dd->first_user_ctxt) { /* N/A for PSM contexts */
331 			rcd->opstats = kzalloc(sizeof(*rcd->opstats),
332 				GFP_KERNEL);
333 			if (!rcd->opstats)
334 				goto bail;
335 		}
336 	}
337 	return rcd;
338 bail:
339 	kfree(rcd->egrbufs.rcvtids);
340 	kfree(rcd->egrbufs.buffers);
341 	kfree(rcd);
342 	return NULL;
343 }
344 
345 /*
346  * Convert a receive header entry size that to the encoding used in the CSR.
347  *
348  * Return a zero if the given size is invalid.
349  */
350 static inline u64 encode_rcv_header_entry_size(u16 size)
351 {
352 	/* there are only 3 valid receive header entry sizes */
353 	if (size == 2)
354 		return 1;
355 	if (size == 16)
356 		return 2;
357 	else if (size == 32)
358 		return 4;
359 	return 0; /* invalid */
360 }
361 
362 /*
363  * Select the largest ccti value over all SLs to determine the intra-
364  * packet gap for the link.
365  *
366  * called with cca_timer_lock held (to protect access to cca_timer
367  * array), and rcu_read_lock() (to protect access to cc_state).
368  */
369 void set_link_ipg(struct hfi1_pportdata *ppd)
370 {
371 	struct hfi1_devdata *dd = ppd->dd;
372 	struct cc_state *cc_state;
373 	int i;
374 	u16 cce, ccti_limit, max_ccti = 0;
375 	u16 shift, mult;
376 	u64 src;
377 	u32 current_egress_rate; /* Mbits /sec */
378 	u32 max_pkt_time;
379 	/*
380 	 * max_pkt_time is the maximum packet egress time in units
381 	 * of the fabric clock period 1/(805 MHz).
382 	 */
383 
384 	cc_state = get_cc_state(ppd);
385 
386 	if (!cc_state)
387 		/*
388 		 * This should _never_ happen - rcu_read_lock() is held,
389 		 * and set_link_ipg() should not be called if cc_state
390 		 * is NULL.
391 		 */
392 		return;
393 
394 	for (i = 0; i < OPA_MAX_SLS; i++) {
395 		u16 ccti = ppd->cca_timer[i].ccti;
396 
397 		if (ccti > max_ccti)
398 			max_ccti = ccti;
399 	}
400 
401 	ccti_limit = cc_state->cct.ccti_limit;
402 	if (max_ccti > ccti_limit)
403 		max_ccti = ccti_limit;
404 
405 	cce = cc_state->cct.entries[max_ccti].entry;
406 	shift = (cce & 0xc000) >> 14;
407 	mult = (cce & 0x3fff);
408 
409 	current_egress_rate = active_egress_rate(ppd);
410 
411 	max_pkt_time = egress_cycles(ppd->ibmaxlen, current_egress_rate);
412 
413 	src = (max_pkt_time >> shift) * mult;
414 
415 	src &= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SMASK;
416 	src <<= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SHIFT;
417 
418 	write_csr(dd, SEND_STATIC_RATE_CONTROL, src);
419 }
420 
421 static enum hrtimer_restart cca_timer_fn(struct hrtimer *t)
422 {
423 	struct cca_timer *cca_timer;
424 	struct hfi1_pportdata *ppd;
425 	int sl;
426 	u16 ccti_timer, ccti_min;
427 	struct cc_state *cc_state;
428 	unsigned long flags;
429 	enum hrtimer_restart ret = HRTIMER_NORESTART;
430 
431 	cca_timer = container_of(t, struct cca_timer, hrtimer);
432 	ppd = cca_timer->ppd;
433 	sl = cca_timer->sl;
434 
435 	rcu_read_lock();
436 
437 	cc_state = get_cc_state(ppd);
438 
439 	if (!cc_state) {
440 		rcu_read_unlock();
441 		return HRTIMER_NORESTART;
442 	}
443 
444 	/*
445 	 * 1) decrement ccti for SL
446 	 * 2) calculate IPG for link (set_link_ipg())
447 	 * 3) restart timer, unless ccti is at min value
448 	 */
449 
450 	ccti_min = cc_state->cong_setting.entries[sl].ccti_min;
451 	ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
452 
453 	spin_lock_irqsave(&ppd->cca_timer_lock, flags);
454 
455 	if (cca_timer->ccti > ccti_min) {
456 		cca_timer->ccti--;
457 		set_link_ipg(ppd);
458 	}
459 
460 	if (cca_timer->ccti > ccti_min) {
461 		unsigned long nsec = 1024 * ccti_timer;
462 		/* ccti_timer is in units of 1.024 usec */
463 		hrtimer_forward_now(t, ns_to_ktime(nsec));
464 		ret = HRTIMER_RESTART;
465 	}
466 
467 	spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
468 	rcu_read_unlock();
469 	return ret;
470 }
471 
472 /*
473  * Common code for initializing the physical port structure.
474  */
475 void hfi1_init_pportdata(struct pci_dev *pdev, struct hfi1_pportdata *ppd,
476 			 struct hfi1_devdata *dd, u8 hw_pidx, u8 port)
477 {
478 	int i;
479 	uint default_pkey_idx;
480 	struct cc_state *cc_state;
481 
482 	ppd->dd = dd;
483 	ppd->hw_pidx = hw_pidx;
484 	ppd->port = port; /* IB port number, not index */
485 
486 	default_pkey_idx = 1;
487 
488 	ppd->pkeys[default_pkey_idx] = DEFAULT_P_KEY;
489 	if (loopback) {
490 		hfi1_early_err(&pdev->dev,
491 			       "Faking data partition 0x8001 in idx %u\n",
492 			       !default_pkey_idx);
493 		ppd->pkeys[!default_pkey_idx] = 0x8001;
494 	}
495 
496 	INIT_WORK(&ppd->link_vc_work, handle_verify_cap);
497 	INIT_WORK(&ppd->link_up_work, handle_link_up);
498 	INIT_WORK(&ppd->link_down_work, handle_link_down);
499 	INIT_WORK(&ppd->freeze_work, handle_freeze);
500 	INIT_WORK(&ppd->link_downgrade_work, handle_link_downgrade);
501 	INIT_WORK(&ppd->sma_message_work, handle_sma_message);
502 	INIT_WORK(&ppd->link_bounce_work, handle_link_bounce);
503 	INIT_WORK(&ppd->linkstate_active_work, receive_interrupt_work);
504 	INIT_WORK(&ppd->qsfp_info.qsfp_work, qsfp_event);
505 
506 	mutex_init(&ppd->hls_lock);
507 	spin_lock_init(&ppd->sdma_alllock);
508 	spin_lock_init(&ppd->qsfp_info.qsfp_lock);
509 
510 	ppd->qsfp_info.ppd = ppd;
511 	ppd->sm_trap_qp = 0x0;
512 	ppd->sa_qp = 0x1;
513 
514 	ppd->hfi1_wq = NULL;
515 
516 	spin_lock_init(&ppd->cca_timer_lock);
517 
518 	for (i = 0; i < OPA_MAX_SLS; i++) {
519 		hrtimer_init(&ppd->cca_timer[i].hrtimer, CLOCK_MONOTONIC,
520 			     HRTIMER_MODE_REL);
521 		ppd->cca_timer[i].ppd = ppd;
522 		ppd->cca_timer[i].sl = i;
523 		ppd->cca_timer[i].ccti = 0;
524 		ppd->cca_timer[i].hrtimer.function = cca_timer_fn;
525 	}
526 
527 	ppd->cc_max_table_entries = IB_CC_TABLE_CAP_DEFAULT;
528 
529 	spin_lock_init(&ppd->cc_state_lock);
530 	spin_lock_init(&ppd->cc_log_lock);
531 	cc_state = kzalloc(sizeof(*cc_state), GFP_KERNEL);
532 	RCU_INIT_POINTER(ppd->cc_state, cc_state);
533 	if (!cc_state)
534 		goto bail;
535 	return;
536 
537 bail:
538 
539 	hfi1_early_err(&pdev->dev,
540 		       "Congestion Control Agent disabled for port %d\n", port);
541 }
542 
543 /*
544  * Do initialization for device that is only needed on
545  * first detect, not on resets.
546  */
547 static int loadtime_init(struct hfi1_devdata *dd)
548 {
549 	return 0;
550 }
551 
552 /**
553  * init_after_reset - re-initialize after a reset
554  * @dd: the hfi1_ib device
555  *
556  * sanity check at least some of the values after reset, and
557  * ensure no receive or transmit (explicitly, in case reset
558  * failed
559  */
560 static int init_after_reset(struct hfi1_devdata *dd)
561 {
562 	int i;
563 
564 	/*
565 	 * Ensure chip does no sends or receives, tail updates, or
566 	 * pioavail updates while we re-initialize.  This is mostly
567 	 * for the driver data structures, not chip registers.
568 	 */
569 	for (i = 0; i < dd->num_rcv_contexts; i++)
570 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
571 				  HFI1_RCVCTRL_INTRAVAIL_DIS |
572 				  HFI1_RCVCTRL_TAILUPD_DIS, i);
573 	pio_send_control(dd, PSC_GLOBAL_DISABLE);
574 	for (i = 0; i < dd->num_send_contexts; i++)
575 		sc_disable(dd->send_contexts[i].sc);
576 
577 	return 0;
578 }
579 
580 static void enable_chip(struct hfi1_devdata *dd)
581 {
582 	u32 rcvmask;
583 	u32 i;
584 
585 	/* enable PIO send */
586 	pio_send_control(dd, PSC_GLOBAL_ENABLE);
587 
588 	/*
589 	 * Enable kernel ctxts' receive and receive interrupt.
590 	 * Other ctxts done as user opens and initializes them.
591 	 */
592 	for (i = 0; i < dd->first_user_ctxt; ++i) {
593 		rcvmask = HFI1_RCVCTRL_CTXT_ENB | HFI1_RCVCTRL_INTRAVAIL_ENB;
594 		rcvmask |= HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, DMA_RTAIL) ?
595 			HFI1_RCVCTRL_TAILUPD_ENB : HFI1_RCVCTRL_TAILUPD_DIS;
596 		if (!HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, MULTI_PKT_EGR))
597 			rcvmask |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
598 		if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_RHQ_FULL))
599 			rcvmask |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
600 		if (HFI1_CAP_KGET_MASK(dd->rcd[i]->flags, NODROP_EGR_FULL))
601 			rcvmask |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
602 		hfi1_rcvctrl(dd, rcvmask, i);
603 		sc_enable(dd->rcd[i]->sc);
604 	}
605 }
606 
607 /**
608  * create_workqueues - create per port workqueues
609  * @dd: the hfi1_ib device
610  */
611 static int create_workqueues(struct hfi1_devdata *dd)
612 {
613 	int pidx;
614 	struct hfi1_pportdata *ppd;
615 
616 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
617 		ppd = dd->pport + pidx;
618 		if (!ppd->hfi1_wq) {
619 			ppd->hfi1_wq =
620 				alloc_workqueue(
621 				    "hfi%d_%d",
622 				    WQ_SYSFS | WQ_HIGHPRI | WQ_CPU_INTENSIVE,
623 				    dd->num_sdma,
624 				    dd->unit, pidx);
625 			if (!ppd->hfi1_wq)
626 				goto wq_error;
627 		}
628 	}
629 	return 0;
630 wq_error:
631 	pr_err("alloc_workqueue failed for port %d\n", pidx + 1);
632 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
633 		ppd = dd->pport + pidx;
634 		if (ppd->hfi1_wq) {
635 			destroy_workqueue(ppd->hfi1_wq);
636 			ppd->hfi1_wq = NULL;
637 		}
638 	}
639 	return -ENOMEM;
640 }
641 
642 /**
643  * hfi1_init - do the actual initialization sequence on the chip
644  * @dd: the hfi1_ib device
645  * @reinit: re-initializing, so don't allocate new memory
646  *
647  * Do the actual initialization sequence on the chip.  This is done
648  * both from the init routine called from the PCI infrastructure, and
649  * when we reset the chip, or detect that it was reset internally,
650  * or it's administratively re-enabled.
651  *
652  * Memory allocation here and in called routines is only done in
653  * the first case (reinit == 0).  We have to be careful, because even
654  * without memory allocation, we need to re-write all the chip registers
655  * TIDs, etc. after the reset or enable has completed.
656  */
657 int hfi1_init(struct hfi1_devdata *dd, int reinit)
658 {
659 	int ret = 0, pidx, lastfail = 0;
660 	unsigned i, len;
661 	struct hfi1_ctxtdata *rcd;
662 	struct hfi1_pportdata *ppd;
663 
664 	/* Set up recv low level handlers */
665 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_EXPECTED] =
666 						kdeth_process_expected;
667 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_EAGER] =
668 						kdeth_process_eager;
669 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_IB] = process_receive_ib;
670 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_ERROR] =
671 						process_receive_error;
672 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_BYPASS] =
673 						process_receive_bypass;
674 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID5] =
675 						process_receive_invalid;
676 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID6] =
677 						process_receive_invalid;
678 	dd->normal_rhf_rcv_functions[RHF_RCV_TYPE_INVALID7] =
679 						process_receive_invalid;
680 	dd->rhf_rcv_function_map = dd->normal_rhf_rcv_functions;
681 
682 	/* Set up send low level handlers */
683 	dd->process_pio_send = hfi1_verbs_send_pio;
684 	dd->process_dma_send = hfi1_verbs_send_dma;
685 	dd->pio_inline_send = pio_copy;
686 
687 	if (is_ax(dd)) {
688 		atomic_set(&dd->drop_packet, DROP_PACKET_ON);
689 		dd->do_drop = 1;
690 	} else {
691 		atomic_set(&dd->drop_packet, DROP_PACKET_OFF);
692 		dd->do_drop = 0;
693 	}
694 
695 	/* make sure the link is not "up" */
696 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
697 		ppd = dd->pport + pidx;
698 		ppd->linkup = 0;
699 	}
700 
701 	if (reinit)
702 		ret = init_after_reset(dd);
703 	else
704 		ret = loadtime_init(dd);
705 	if (ret)
706 		goto done;
707 
708 	/* allocate dummy tail memory for all receive contexts */
709 	dd->rcvhdrtail_dummy_kvaddr = dma_zalloc_coherent(
710 		&dd->pcidev->dev, sizeof(u64),
711 		&dd->rcvhdrtail_dummy_physaddr,
712 		GFP_KERNEL);
713 
714 	if (!dd->rcvhdrtail_dummy_kvaddr) {
715 		dd_dev_err(dd, "cannot allocate dummy tail memory\n");
716 		ret = -ENOMEM;
717 		goto done;
718 	}
719 
720 	/* dd->rcd can be NULL if early initialization failed */
721 	for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
722 		/*
723 		 * Set up the (kernel) rcvhdr queue and egr TIDs.  If doing
724 		 * re-init, the simplest way to handle this is to free
725 		 * existing, and re-allocate.
726 		 * Need to re-create rest of ctxt 0 ctxtdata as well.
727 		 */
728 		rcd = dd->rcd[i];
729 		if (!rcd)
730 			continue;
731 
732 		rcd->do_interrupt = &handle_receive_interrupt;
733 
734 		lastfail = hfi1_create_rcvhdrq(dd, rcd);
735 		if (!lastfail)
736 			lastfail = hfi1_setup_eagerbufs(rcd);
737 		if (lastfail) {
738 			dd_dev_err(dd,
739 				   "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
740 			ret = lastfail;
741 		}
742 	}
743 
744 	/* Allocate enough memory for user event notification. */
745 	len = PAGE_ALIGN(dd->chip_rcv_contexts * HFI1_MAX_SHARED_CTXTS *
746 			 sizeof(*dd->events));
747 	dd->events = vmalloc_user(len);
748 	if (!dd->events)
749 		dd_dev_err(dd, "Failed to allocate user events page\n");
750 	/*
751 	 * Allocate a page for device and port status.
752 	 * Page will be shared amongst all user processes.
753 	 */
754 	dd->status = vmalloc_user(PAGE_SIZE);
755 	if (!dd->status)
756 		dd_dev_err(dd, "Failed to allocate dev status page\n");
757 	else
758 		dd->freezelen = PAGE_SIZE - (sizeof(*dd->status) -
759 					     sizeof(dd->status->freezemsg));
760 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
761 		ppd = dd->pport + pidx;
762 		if (dd->status)
763 			/* Currently, we only have one port */
764 			ppd->statusp = &dd->status->port;
765 
766 		set_mtu(ppd);
767 	}
768 
769 	/* enable chip even if we have an error, so we can debug cause */
770 	enable_chip(dd);
771 
772 done:
773 	/*
774 	 * Set status even if port serdes is not initialized
775 	 * so that diags will work.
776 	 */
777 	if (dd->status)
778 		dd->status->dev |= HFI1_STATUS_CHIP_PRESENT |
779 			HFI1_STATUS_INITTED;
780 	if (!ret) {
781 		/* enable all interrupts from the chip */
782 		set_intr_state(dd, 1);
783 
784 		/* chip is OK for user apps; mark it as initialized */
785 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
786 			ppd = dd->pport + pidx;
787 
788 			/*
789 			 * start the serdes - must be after interrupts are
790 			 * enabled so we are notified when the link goes up
791 			 */
792 			lastfail = bringup_serdes(ppd);
793 			if (lastfail)
794 				dd_dev_info(dd,
795 					    "Failed to bring up port %u\n",
796 					    ppd->port);
797 
798 			/*
799 			 * Set status even if port serdes is not initialized
800 			 * so that diags will work.
801 			 */
802 			if (ppd->statusp)
803 				*ppd->statusp |= HFI1_STATUS_CHIP_PRESENT |
804 							HFI1_STATUS_INITTED;
805 			if (!ppd->link_speed_enabled)
806 				continue;
807 		}
808 	}
809 
810 	/* if ret is non-zero, we probably should do some cleanup here... */
811 	return ret;
812 }
813 
814 static inline struct hfi1_devdata *__hfi1_lookup(int unit)
815 {
816 	return idr_find(&hfi1_unit_table, unit);
817 }
818 
819 struct hfi1_devdata *hfi1_lookup(int unit)
820 {
821 	struct hfi1_devdata *dd;
822 	unsigned long flags;
823 
824 	spin_lock_irqsave(&hfi1_devs_lock, flags);
825 	dd = __hfi1_lookup(unit);
826 	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
827 
828 	return dd;
829 }
830 
831 /*
832  * Stop the timers during unit shutdown, or after an error late
833  * in initialization.
834  */
835 static void stop_timers(struct hfi1_devdata *dd)
836 {
837 	struct hfi1_pportdata *ppd;
838 	int pidx;
839 
840 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
841 		ppd = dd->pport + pidx;
842 		if (ppd->led_override_timer.data) {
843 			del_timer_sync(&ppd->led_override_timer);
844 			atomic_set(&ppd->led_override_timer_active, 0);
845 		}
846 	}
847 }
848 
849 /**
850  * shutdown_device - shut down a device
851  * @dd: the hfi1_ib device
852  *
853  * This is called to make the device quiet when we are about to
854  * unload the driver, and also when the device is administratively
855  * disabled.   It does not free any data structures.
856  * Everything it does has to be setup again by hfi1_init(dd, 1)
857  */
858 static void shutdown_device(struct hfi1_devdata *dd)
859 {
860 	struct hfi1_pportdata *ppd;
861 	unsigned pidx;
862 	int i;
863 
864 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
865 		ppd = dd->pport + pidx;
866 
867 		ppd->linkup = 0;
868 		if (ppd->statusp)
869 			*ppd->statusp &= ~(HFI1_STATUS_IB_CONF |
870 					   HFI1_STATUS_IB_READY);
871 	}
872 	dd->flags &= ~HFI1_INITTED;
873 
874 	/* mask interrupts, but not errors */
875 	set_intr_state(dd, 0);
876 
877 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
878 		ppd = dd->pport + pidx;
879 		for (i = 0; i < dd->num_rcv_contexts; i++)
880 			hfi1_rcvctrl(dd, HFI1_RCVCTRL_TAILUPD_DIS |
881 					  HFI1_RCVCTRL_CTXT_DIS |
882 					  HFI1_RCVCTRL_INTRAVAIL_DIS |
883 					  HFI1_RCVCTRL_PKEY_DIS |
884 					  HFI1_RCVCTRL_ONE_PKT_EGR_DIS, i);
885 		/*
886 		 * Gracefully stop all sends allowing any in progress to
887 		 * trickle out first.
888 		 */
889 		for (i = 0; i < dd->num_send_contexts; i++)
890 			sc_flush(dd->send_contexts[i].sc);
891 	}
892 
893 	/*
894 	 * Enough for anything that's going to trickle out to have actually
895 	 * done so.
896 	 */
897 	udelay(20);
898 
899 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
900 		ppd = dd->pport + pidx;
901 
902 		/* disable all contexts */
903 		for (i = 0; i < dd->num_send_contexts; i++)
904 			sc_disable(dd->send_contexts[i].sc);
905 		/* disable the send device */
906 		pio_send_control(dd, PSC_GLOBAL_DISABLE);
907 
908 		shutdown_led_override(ppd);
909 
910 		/*
911 		 * Clear SerdesEnable.
912 		 * We can't count on interrupts since we are stopping.
913 		 */
914 		hfi1_quiet_serdes(ppd);
915 
916 		if (ppd->hfi1_wq) {
917 			destroy_workqueue(ppd->hfi1_wq);
918 			ppd->hfi1_wq = NULL;
919 		}
920 	}
921 	sdma_exit(dd);
922 }
923 
924 /**
925  * hfi1_free_ctxtdata - free a context's allocated data
926  * @dd: the hfi1_ib device
927  * @rcd: the ctxtdata structure
928  *
929  * free up any allocated data for a context
930  * This should not touch anything that would affect a simultaneous
931  * re-allocation of context data, because it is called after hfi1_mutex
932  * is released (and can be called from reinit as well).
933  * It should never change any chip state, or global driver state.
934  */
935 void hfi1_free_ctxtdata(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
936 {
937 	unsigned e;
938 
939 	if (!rcd)
940 		return;
941 
942 	if (rcd->rcvhdrq) {
943 		dma_free_coherent(&dd->pcidev->dev, rcd->rcvhdrq_size,
944 				  rcd->rcvhdrq, rcd->rcvhdrq_phys);
945 		rcd->rcvhdrq = NULL;
946 		if (rcd->rcvhdrtail_kvaddr) {
947 			dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
948 					  (void *)rcd->rcvhdrtail_kvaddr,
949 					  rcd->rcvhdrqtailaddr_phys);
950 			rcd->rcvhdrtail_kvaddr = NULL;
951 		}
952 	}
953 
954 	/* all the RcvArray entries should have been cleared by now */
955 	kfree(rcd->egrbufs.rcvtids);
956 
957 	for (e = 0; e < rcd->egrbufs.alloced; e++) {
958 		if (rcd->egrbufs.buffers[e].phys)
959 			dma_free_coherent(&dd->pcidev->dev,
960 					  rcd->egrbufs.buffers[e].len,
961 					  rcd->egrbufs.buffers[e].addr,
962 					  rcd->egrbufs.buffers[e].phys);
963 	}
964 	kfree(rcd->egrbufs.buffers);
965 
966 	sc_free(rcd->sc);
967 	vfree(rcd->user_event_mask);
968 	vfree(rcd->subctxt_uregbase);
969 	vfree(rcd->subctxt_rcvegrbuf);
970 	vfree(rcd->subctxt_rcvhdr_base);
971 	kfree(rcd->opstats);
972 	kfree(rcd);
973 }
974 
975 /*
976  * Release our hold on the shared asic data.  If we are the last one,
977  * return the structure to be finalized outside the lock.  Must be
978  * holding hfi1_devs_lock.
979  */
980 static struct hfi1_asic_data *release_asic_data(struct hfi1_devdata *dd)
981 {
982 	struct hfi1_asic_data *ad;
983 	int other;
984 
985 	if (!dd->asic_data)
986 		return NULL;
987 	dd->asic_data->dds[dd->hfi1_id] = NULL;
988 	other = dd->hfi1_id ? 0 : 1;
989 	ad = dd->asic_data;
990 	dd->asic_data = NULL;
991 	/* return NULL if the other dd still has a link */
992 	return ad->dds[other] ? NULL : ad;
993 }
994 
995 static void finalize_asic_data(struct hfi1_devdata *dd,
996 			       struct hfi1_asic_data *ad)
997 {
998 	clean_up_i2c(dd, ad);
999 	kfree(ad);
1000 }
1001 
1002 static void __hfi1_free_devdata(struct kobject *kobj)
1003 {
1004 	struct hfi1_devdata *dd =
1005 		container_of(kobj, struct hfi1_devdata, kobj);
1006 	struct hfi1_asic_data *ad;
1007 	unsigned long flags;
1008 
1009 	spin_lock_irqsave(&hfi1_devs_lock, flags);
1010 	idr_remove(&hfi1_unit_table, dd->unit);
1011 	list_del(&dd->list);
1012 	ad = release_asic_data(dd);
1013 	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1014 	if (ad)
1015 		finalize_asic_data(dd, ad);
1016 	free_platform_config(dd);
1017 	rcu_barrier(); /* wait for rcu callbacks to complete */
1018 	free_percpu(dd->int_counter);
1019 	free_percpu(dd->rcv_limit);
1020 	free_percpu(dd->send_schedule);
1021 	rvt_dealloc_device(&dd->verbs_dev.rdi);
1022 }
1023 
1024 static struct kobj_type hfi1_devdata_type = {
1025 	.release = __hfi1_free_devdata,
1026 };
1027 
1028 void hfi1_free_devdata(struct hfi1_devdata *dd)
1029 {
1030 	kobject_put(&dd->kobj);
1031 }
1032 
1033 /*
1034  * Allocate our primary per-unit data structure.  Must be done via verbs
1035  * allocator, because the verbs cleanup process both does cleanup and
1036  * free of the data structure.
1037  * "extra" is for chip-specific data.
1038  *
1039  * Use the idr mechanism to get a unit number for this unit.
1040  */
1041 struct hfi1_devdata *hfi1_alloc_devdata(struct pci_dev *pdev, size_t extra)
1042 {
1043 	unsigned long flags;
1044 	struct hfi1_devdata *dd;
1045 	int ret, nports;
1046 
1047 	/* extra is * number of ports */
1048 	nports = extra / sizeof(struct hfi1_pportdata);
1049 
1050 	dd = (struct hfi1_devdata *)rvt_alloc_device(sizeof(*dd) + extra,
1051 						     nports);
1052 	if (!dd)
1053 		return ERR_PTR(-ENOMEM);
1054 	dd->num_pports = nports;
1055 	dd->pport = (struct hfi1_pportdata *)(dd + 1);
1056 
1057 	INIT_LIST_HEAD(&dd->list);
1058 	idr_preload(GFP_KERNEL);
1059 	spin_lock_irqsave(&hfi1_devs_lock, flags);
1060 
1061 	ret = idr_alloc(&hfi1_unit_table, dd, 0, 0, GFP_NOWAIT);
1062 	if (ret >= 0) {
1063 		dd->unit = ret;
1064 		list_add(&dd->list, &hfi1_dev_list);
1065 	}
1066 
1067 	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1068 	idr_preload_end();
1069 
1070 	if (ret < 0) {
1071 		hfi1_early_err(&pdev->dev,
1072 			       "Could not allocate unit ID: error %d\n", -ret);
1073 		goto bail;
1074 	}
1075 	/*
1076 	 * Initialize all locks for the device. This needs to be as early as
1077 	 * possible so locks are usable.
1078 	 */
1079 	spin_lock_init(&dd->sc_lock);
1080 	spin_lock_init(&dd->sendctrl_lock);
1081 	spin_lock_init(&dd->rcvctrl_lock);
1082 	spin_lock_init(&dd->uctxt_lock);
1083 	spin_lock_init(&dd->hfi1_diag_trans_lock);
1084 	spin_lock_init(&dd->sc_init_lock);
1085 	spin_lock_init(&dd->dc8051_lock);
1086 	spin_lock_init(&dd->dc8051_memlock);
1087 	seqlock_init(&dd->sc2vl_lock);
1088 	spin_lock_init(&dd->sde_map_lock);
1089 	spin_lock_init(&dd->pio_map_lock);
1090 	init_waitqueue_head(&dd->event_queue);
1091 
1092 	dd->int_counter = alloc_percpu(u64);
1093 	if (!dd->int_counter) {
1094 		ret = -ENOMEM;
1095 		hfi1_early_err(&pdev->dev,
1096 			       "Could not allocate per-cpu int_counter\n");
1097 		goto bail;
1098 	}
1099 
1100 	dd->rcv_limit = alloc_percpu(u64);
1101 	if (!dd->rcv_limit) {
1102 		ret = -ENOMEM;
1103 		hfi1_early_err(&pdev->dev,
1104 			       "Could not allocate per-cpu rcv_limit\n");
1105 		goto bail;
1106 	}
1107 
1108 	dd->send_schedule = alloc_percpu(u64);
1109 	if (!dd->send_schedule) {
1110 		ret = -ENOMEM;
1111 		hfi1_early_err(&pdev->dev,
1112 			       "Could not allocate per-cpu int_counter\n");
1113 		goto bail;
1114 	}
1115 
1116 	if (!hfi1_cpulist_count) {
1117 		u32 count = num_online_cpus();
1118 
1119 		hfi1_cpulist = kcalloc(BITS_TO_LONGS(count), sizeof(long),
1120 				       GFP_KERNEL);
1121 		if (hfi1_cpulist)
1122 			hfi1_cpulist_count = count;
1123 		else
1124 			hfi1_early_err(
1125 			&pdev->dev,
1126 			"Could not alloc cpulist info, cpu affinity might be wrong\n");
1127 	}
1128 	kobject_init(&dd->kobj, &hfi1_devdata_type);
1129 	return dd;
1130 
1131 bail:
1132 	if (!list_empty(&dd->list))
1133 		list_del_init(&dd->list);
1134 	rvt_dealloc_device(&dd->verbs_dev.rdi);
1135 	return ERR_PTR(ret);
1136 }
1137 
1138 /*
1139  * Called from freeze mode handlers, and from PCI error
1140  * reporting code.  Should be paranoid about state of
1141  * system and data structures.
1142  */
1143 void hfi1_disable_after_error(struct hfi1_devdata *dd)
1144 {
1145 	if (dd->flags & HFI1_INITTED) {
1146 		u32 pidx;
1147 
1148 		dd->flags &= ~HFI1_INITTED;
1149 		if (dd->pport)
1150 			for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1151 				struct hfi1_pportdata *ppd;
1152 
1153 				ppd = dd->pport + pidx;
1154 				if (dd->flags & HFI1_PRESENT)
1155 					set_link_state(ppd, HLS_DN_DISABLE);
1156 
1157 				if (ppd->statusp)
1158 					*ppd->statusp &= ~HFI1_STATUS_IB_READY;
1159 			}
1160 	}
1161 
1162 	/*
1163 	 * Mark as having had an error for driver, and also
1164 	 * for /sys and status word mapped to user programs.
1165 	 * This marks unit as not usable, until reset.
1166 	 */
1167 	if (dd->status)
1168 		dd->status->dev |= HFI1_STATUS_HWERROR;
1169 }
1170 
1171 static void remove_one(struct pci_dev *);
1172 static int init_one(struct pci_dev *, const struct pci_device_id *);
1173 
1174 #define DRIVER_LOAD_MSG "Intel " DRIVER_NAME " loaded: "
1175 #define PFX DRIVER_NAME ": "
1176 
1177 const struct pci_device_id hfi1_pci_tbl[] = {
1178 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL0) },
1179 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL1) },
1180 	{ 0, }
1181 };
1182 
1183 MODULE_DEVICE_TABLE(pci, hfi1_pci_tbl);
1184 
1185 static struct pci_driver hfi1_pci_driver = {
1186 	.name = DRIVER_NAME,
1187 	.probe = init_one,
1188 	.remove = remove_one,
1189 	.id_table = hfi1_pci_tbl,
1190 	.err_handler = &hfi1_pci_err_handler,
1191 };
1192 
1193 static void __init compute_krcvqs(void)
1194 {
1195 	int i;
1196 
1197 	for (i = 0; i < krcvqsset; i++)
1198 		n_krcvqs += krcvqs[i];
1199 }
1200 
1201 /*
1202  * Do all the generic driver unit- and chip-independent memory
1203  * allocation and initialization.
1204  */
1205 static int __init hfi1_mod_init(void)
1206 {
1207 	int ret;
1208 
1209 	ret = dev_init();
1210 	if (ret)
1211 		goto bail;
1212 
1213 	ret = node_affinity_init();
1214 	if (ret)
1215 		goto bail;
1216 
1217 	/* validate max MTU before any devices start */
1218 	if (!valid_opa_max_mtu(hfi1_max_mtu)) {
1219 		pr_err("Invalid max_mtu 0x%x, using 0x%x instead\n",
1220 		       hfi1_max_mtu, HFI1_DEFAULT_MAX_MTU);
1221 		hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
1222 	}
1223 	/* valid CUs run from 1-128 in powers of 2 */
1224 	if (hfi1_cu > 128 || !is_power_of_2(hfi1_cu))
1225 		hfi1_cu = 1;
1226 	/* valid credit return threshold is 0-100, variable is unsigned */
1227 	if (user_credit_return_threshold > 100)
1228 		user_credit_return_threshold = 100;
1229 
1230 	compute_krcvqs();
1231 	/*
1232 	 * sanitize receive interrupt count, time must wait until after
1233 	 * the hardware type is known
1234 	 */
1235 	if (rcv_intr_count > RCV_HDR_HEAD_COUNTER_MASK)
1236 		rcv_intr_count = RCV_HDR_HEAD_COUNTER_MASK;
1237 	/* reject invalid combinations */
1238 	if (rcv_intr_count == 0 && rcv_intr_timeout == 0) {
1239 		pr_err("Invalid mode: both receive interrupt count and available timeout are zero - setting interrupt count to 1\n");
1240 		rcv_intr_count = 1;
1241 	}
1242 	if (rcv_intr_count > 1 && rcv_intr_timeout == 0) {
1243 		/*
1244 		 * Avoid indefinite packet delivery by requiring a timeout
1245 		 * if count is > 1.
1246 		 */
1247 		pr_err("Invalid mode: receive interrupt count greater than 1 and available timeout is zero - setting available timeout to 1\n");
1248 		rcv_intr_timeout = 1;
1249 	}
1250 	if (rcv_intr_dynamic && !(rcv_intr_count > 1 && rcv_intr_timeout > 0)) {
1251 		/*
1252 		 * The dynamic algorithm expects a non-zero timeout
1253 		 * and a count > 1.
1254 		 */
1255 		pr_err("Invalid mode: dynamic receive interrupt mitigation with invalid count and timeout - turning dynamic off\n");
1256 		rcv_intr_dynamic = 0;
1257 	}
1258 
1259 	/* sanitize link CRC options */
1260 	link_crc_mask &= SUPPORTED_CRCS;
1261 
1262 	/*
1263 	 * These must be called before the driver is registered with
1264 	 * the PCI subsystem.
1265 	 */
1266 	idr_init(&hfi1_unit_table);
1267 
1268 	hfi1_dbg_init();
1269 	ret = hfi1_wss_init();
1270 	if (ret < 0)
1271 		goto bail_wss;
1272 	ret = pci_register_driver(&hfi1_pci_driver);
1273 	if (ret < 0) {
1274 		pr_err("Unable to register driver: error %d\n", -ret);
1275 		goto bail_dev;
1276 	}
1277 	goto bail; /* all OK */
1278 
1279 bail_dev:
1280 	hfi1_wss_exit();
1281 bail_wss:
1282 	hfi1_dbg_exit();
1283 	idr_destroy(&hfi1_unit_table);
1284 	dev_cleanup();
1285 bail:
1286 	return ret;
1287 }
1288 
1289 module_init(hfi1_mod_init);
1290 
1291 /*
1292  * Do the non-unit driver cleanup, memory free, etc. at unload.
1293  */
1294 static void __exit hfi1_mod_cleanup(void)
1295 {
1296 	pci_unregister_driver(&hfi1_pci_driver);
1297 	node_affinity_destroy();
1298 	hfi1_wss_exit();
1299 	hfi1_dbg_exit();
1300 	hfi1_cpulist_count = 0;
1301 	kfree(hfi1_cpulist);
1302 
1303 	idr_destroy(&hfi1_unit_table);
1304 	dispose_firmware();	/* asymmetric with obtain_firmware() */
1305 	dev_cleanup();
1306 }
1307 
1308 module_exit(hfi1_mod_cleanup);
1309 
1310 /* this can only be called after a successful initialization */
1311 static void cleanup_device_data(struct hfi1_devdata *dd)
1312 {
1313 	int ctxt;
1314 	int pidx;
1315 	struct hfi1_ctxtdata **tmp;
1316 	unsigned long flags;
1317 
1318 	/* users can't do anything more with chip */
1319 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1320 		struct hfi1_pportdata *ppd = &dd->pport[pidx];
1321 		struct cc_state *cc_state;
1322 		int i;
1323 
1324 		if (ppd->statusp)
1325 			*ppd->statusp &= ~HFI1_STATUS_CHIP_PRESENT;
1326 
1327 		for (i = 0; i < OPA_MAX_SLS; i++)
1328 			hrtimer_cancel(&ppd->cca_timer[i].hrtimer);
1329 
1330 		spin_lock(&ppd->cc_state_lock);
1331 		cc_state = get_cc_state_protected(ppd);
1332 		RCU_INIT_POINTER(ppd->cc_state, NULL);
1333 		spin_unlock(&ppd->cc_state_lock);
1334 
1335 		if (cc_state)
1336 			kfree_rcu(cc_state, rcu);
1337 	}
1338 
1339 	free_credit_return(dd);
1340 
1341 	/*
1342 	 * Free any resources still in use (usually just kernel contexts)
1343 	 * at unload; we do for ctxtcnt, because that's what we allocate.
1344 	 * We acquire lock to be really paranoid that rcd isn't being
1345 	 * accessed from some interrupt-related code (that should not happen,
1346 	 * but best to be sure).
1347 	 */
1348 	spin_lock_irqsave(&dd->uctxt_lock, flags);
1349 	tmp = dd->rcd;
1350 	dd->rcd = NULL;
1351 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1352 
1353 	if (dd->rcvhdrtail_dummy_kvaddr) {
1354 		dma_free_coherent(&dd->pcidev->dev, sizeof(u64),
1355 				  (void *)dd->rcvhdrtail_dummy_kvaddr,
1356 				  dd->rcvhdrtail_dummy_physaddr);
1357 		dd->rcvhdrtail_dummy_kvaddr = NULL;
1358 	}
1359 
1360 	for (ctxt = 0; tmp && ctxt < dd->num_rcv_contexts; ctxt++) {
1361 		struct hfi1_ctxtdata *rcd = tmp[ctxt];
1362 
1363 		tmp[ctxt] = NULL; /* debugging paranoia */
1364 		if (rcd) {
1365 			hfi1_clear_tids(rcd);
1366 			hfi1_free_ctxtdata(dd, rcd);
1367 		}
1368 	}
1369 	kfree(tmp);
1370 	free_pio_map(dd);
1371 	/* must follow rcv context free - need to remove rcv's hooks */
1372 	for (ctxt = 0; ctxt < dd->num_send_contexts; ctxt++)
1373 		sc_free(dd->send_contexts[ctxt].sc);
1374 	dd->num_send_contexts = 0;
1375 	kfree(dd->send_contexts);
1376 	dd->send_contexts = NULL;
1377 	kfree(dd->hw_to_sw);
1378 	dd->hw_to_sw = NULL;
1379 	kfree(dd->boardname);
1380 	vfree(dd->events);
1381 	vfree(dd->status);
1382 }
1383 
1384 /*
1385  * Clean up on unit shutdown, or error during unit load after
1386  * successful initialization.
1387  */
1388 static void postinit_cleanup(struct hfi1_devdata *dd)
1389 {
1390 	hfi1_start_cleanup(dd);
1391 
1392 	hfi1_pcie_ddcleanup(dd);
1393 	hfi1_pcie_cleanup(dd->pcidev);
1394 
1395 	cleanup_device_data(dd);
1396 
1397 	hfi1_free_devdata(dd);
1398 }
1399 
1400 static int init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
1401 {
1402 	int ret = 0, j, pidx, initfail;
1403 	struct hfi1_devdata *dd = ERR_PTR(-EINVAL);
1404 	struct hfi1_pportdata *ppd;
1405 
1406 	/* First, lock the non-writable module parameters */
1407 	HFI1_CAP_LOCK();
1408 
1409 	/* Validate some global module parameters */
1410 	if (rcvhdrcnt <= HFI1_MIN_HDRQ_EGRBUF_CNT) {
1411 		hfi1_early_err(&pdev->dev, "Header queue  count too small\n");
1412 		ret = -EINVAL;
1413 		goto bail;
1414 	}
1415 	if (rcvhdrcnt > HFI1_MAX_HDRQ_EGRBUF_CNT) {
1416 		hfi1_early_err(&pdev->dev,
1417 			       "Receive header queue count cannot be greater than %u\n",
1418 			       HFI1_MAX_HDRQ_EGRBUF_CNT);
1419 		ret = -EINVAL;
1420 		goto bail;
1421 	}
1422 	/* use the encoding function as a sanitization check */
1423 	if (!encode_rcv_header_entry_size(hfi1_hdrq_entsize)) {
1424 		hfi1_early_err(&pdev->dev, "Invalid HdrQ Entry size %u\n",
1425 			       hfi1_hdrq_entsize);
1426 		ret = -EINVAL;
1427 		goto bail;
1428 	}
1429 
1430 	/* The receive eager buffer size must be set before the receive
1431 	 * contexts are created.
1432 	 *
1433 	 * Set the eager buffer size.  Validate that it falls in a range
1434 	 * allowed by the hardware - all powers of 2 between the min and
1435 	 * max.  The maximum valid MTU is within the eager buffer range
1436 	 * so we do not need to cap the max_mtu by an eager buffer size
1437 	 * setting.
1438 	 */
1439 	if (eager_buffer_size) {
1440 		if (!is_power_of_2(eager_buffer_size))
1441 			eager_buffer_size =
1442 				roundup_pow_of_two(eager_buffer_size);
1443 		eager_buffer_size =
1444 			clamp_val(eager_buffer_size,
1445 				  MIN_EAGER_BUFFER * 8,
1446 				  MAX_EAGER_BUFFER_TOTAL);
1447 		hfi1_early_info(&pdev->dev, "Eager buffer size %u\n",
1448 				eager_buffer_size);
1449 	} else {
1450 		hfi1_early_err(&pdev->dev, "Invalid Eager buffer size of 0\n");
1451 		ret = -EINVAL;
1452 		goto bail;
1453 	}
1454 
1455 	/* restrict value of hfi1_rcvarr_split */
1456 	hfi1_rcvarr_split = clamp_val(hfi1_rcvarr_split, 0, 100);
1457 
1458 	ret = hfi1_pcie_init(pdev, ent);
1459 	if (ret)
1460 		goto bail;
1461 
1462 	/*
1463 	 * Do device-specific initialization, function table setup, dd
1464 	 * allocation, etc.
1465 	 */
1466 	switch (ent->device) {
1467 	case PCI_DEVICE_ID_INTEL0:
1468 	case PCI_DEVICE_ID_INTEL1:
1469 		dd = hfi1_init_dd(pdev, ent);
1470 		break;
1471 	default:
1472 		hfi1_early_err(&pdev->dev,
1473 			       "Failing on unknown Intel deviceid 0x%x\n",
1474 			       ent->device);
1475 		ret = -ENODEV;
1476 	}
1477 
1478 	if (IS_ERR(dd))
1479 		ret = PTR_ERR(dd);
1480 	if (ret)
1481 		goto clean_bail; /* error already printed */
1482 
1483 	ret = create_workqueues(dd);
1484 	if (ret)
1485 		goto clean_bail;
1486 
1487 	/* do the generic initialization */
1488 	initfail = hfi1_init(dd, 0);
1489 
1490 	ret = hfi1_register_ib_device(dd);
1491 
1492 	/*
1493 	 * Now ready for use.  this should be cleared whenever we
1494 	 * detect a reset, or initiate one.  If earlier failure,
1495 	 * we still create devices, so diags, etc. can be used
1496 	 * to determine cause of problem.
1497 	 */
1498 	if (!initfail && !ret) {
1499 		dd->flags |= HFI1_INITTED;
1500 		/* create debufs files after init and ib register */
1501 		hfi1_dbg_ibdev_init(&dd->verbs_dev);
1502 	}
1503 
1504 	j = hfi1_device_create(dd);
1505 	if (j)
1506 		dd_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1507 
1508 	if (initfail || ret) {
1509 		stop_timers(dd);
1510 		flush_workqueue(ib_wq);
1511 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1512 			hfi1_quiet_serdes(dd->pport + pidx);
1513 			ppd = dd->pport + pidx;
1514 			if (ppd->hfi1_wq) {
1515 				destroy_workqueue(ppd->hfi1_wq);
1516 				ppd->hfi1_wq = NULL;
1517 			}
1518 		}
1519 		if (!j)
1520 			hfi1_device_remove(dd);
1521 		if (!ret)
1522 			hfi1_unregister_ib_device(dd);
1523 		postinit_cleanup(dd);
1524 		if (initfail)
1525 			ret = initfail;
1526 		goto bail;	/* everything already cleaned */
1527 	}
1528 
1529 	sdma_start(dd);
1530 
1531 	return 0;
1532 
1533 clean_bail:
1534 	hfi1_pcie_cleanup(pdev);
1535 bail:
1536 	return ret;
1537 }
1538 
1539 static void remove_one(struct pci_dev *pdev)
1540 {
1541 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
1542 
1543 	/* close debugfs files before ib unregister */
1544 	hfi1_dbg_ibdev_exit(&dd->verbs_dev);
1545 	/* unregister from IB core */
1546 	hfi1_unregister_ib_device(dd);
1547 
1548 	/*
1549 	 * Disable the IB link, disable interrupts on the device,
1550 	 * clear dma engines, etc.
1551 	 */
1552 	shutdown_device(dd);
1553 
1554 	stop_timers(dd);
1555 
1556 	/* wait until all of our (qsfp) queue_work() calls complete */
1557 	flush_workqueue(ib_wq);
1558 
1559 	hfi1_device_remove(dd);
1560 
1561 	postinit_cleanup(dd);
1562 }
1563 
1564 /**
1565  * hfi1_create_rcvhdrq - create a receive header queue
1566  * @dd: the hfi1_ib device
1567  * @rcd: the context data
1568  *
1569  * This must be contiguous memory (from an i/o perspective), and must be
1570  * DMA'able (which means for some systems, it will go through an IOMMU,
1571  * or be forced into a low address range).
1572  */
1573 int hfi1_create_rcvhdrq(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
1574 {
1575 	unsigned amt;
1576 	u64 reg;
1577 
1578 	if (!rcd->rcvhdrq) {
1579 		dma_addr_t phys_hdrqtail;
1580 		gfp_t gfp_flags;
1581 
1582 		/*
1583 		 * rcvhdrqentsize is in DWs, so we have to convert to bytes
1584 		 * (* sizeof(u32)).
1585 		 */
1586 		amt = PAGE_ALIGN(rcd->rcvhdrq_cnt * rcd->rcvhdrqentsize *
1587 				 sizeof(u32));
1588 
1589 		gfp_flags = (rcd->ctxt >= dd->first_user_ctxt) ?
1590 			GFP_USER : GFP_KERNEL;
1591 		rcd->rcvhdrq = dma_zalloc_coherent(
1592 			&dd->pcidev->dev, amt, &rcd->rcvhdrq_phys,
1593 			gfp_flags | __GFP_COMP);
1594 
1595 		if (!rcd->rcvhdrq) {
1596 			dd_dev_err(dd,
1597 				   "attempt to allocate %d bytes for ctxt %u rcvhdrq failed\n",
1598 				   amt, rcd->ctxt);
1599 			goto bail;
1600 		}
1601 
1602 		if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
1603 			rcd->rcvhdrtail_kvaddr = dma_zalloc_coherent(
1604 				&dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
1605 				gfp_flags);
1606 			if (!rcd->rcvhdrtail_kvaddr)
1607 				goto bail_free;
1608 			rcd->rcvhdrqtailaddr_phys = phys_hdrqtail;
1609 		}
1610 
1611 		rcd->rcvhdrq_size = amt;
1612 	}
1613 	/*
1614 	 * These values are per-context:
1615 	 *	RcvHdrCnt
1616 	 *	RcvHdrEntSize
1617 	 *	RcvHdrSize
1618 	 */
1619 	reg = ((u64)(rcd->rcvhdrq_cnt >> HDRQ_SIZE_SHIFT)
1620 			& RCV_HDR_CNT_CNT_MASK)
1621 		<< RCV_HDR_CNT_CNT_SHIFT;
1622 	write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_CNT, reg);
1623 	reg = (encode_rcv_header_entry_size(rcd->rcvhdrqentsize)
1624 			& RCV_HDR_ENT_SIZE_ENT_SIZE_MASK)
1625 		<< RCV_HDR_ENT_SIZE_ENT_SIZE_SHIFT;
1626 	write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_ENT_SIZE, reg);
1627 	reg = (dd->rcvhdrsize & RCV_HDR_SIZE_HDR_SIZE_MASK)
1628 		<< RCV_HDR_SIZE_HDR_SIZE_SHIFT;
1629 	write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_SIZE, reg);
1630 
1631 	/*
1632 	 * Program dummy tail address for every receive context
1633 	 * before enabling any receive context
1634 	 */
1635 	write_kctxt_csr(dd, rcd->ctxt, RCV_HDR_TAIL_ADDR,
1636 			dd->rcvhdrtail_dummy_physaddr);
1637 
1638 	return 0;
1639 
1640 bail_free:
1641 	dd_dev_err(dd,
1642 		   "attempt to allocate 1 page for ctxt %u rcvhdrqtailaddr failed\n",
1643 		   rcd->ctxt);
1644 	vfree(rcd->user_event_mask);
1645 	rcd->user_event_mask = NULL;
1646 	dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1647 			  rcd->rcvhdrq_phys);
1648 	rcd->rcvhdrq = NULL;
1649 bail:
1650 	return -ENOMEM;
1651 }
1652 
1653 /**
1654  * allocate eager buffers, both kernel and user contexts.
1655  * @rcd: the context we are setting up.
1656  *
1657  * Allocate the eager TID buffers and program them into hip.
1658  * They are no longer completely contiguous, we do multiple allocation
1659  * calls.  Otherwise we get the OOM code involved, by asking for too
1660  * much per call, with disastrous results on some kernels.
1661  */
1662 int hfi1_setup_eagerbufs(struct hfi1_ctxtdata *rcd)
1663 {
1664 	struct hfi1_devdata *dd = rcd->dd;
1665 	u32 max_entries, egrtop, alloced_bytes = 0, idx = 0;
1666 	gfp_t gfp_flags;
1667 	u16 order;
1668 	int ret = 0;
1669 	u16 round_mtu = roundup_pow_of_two(hfi1_max_mtu);
1670 
1671 	/*
1672 	 * GFP_USER, but without GFP_FS, so buffer cache can be
1673 	 * coalesced (we hope); otherwise, even at order 4,
1674 	 * heavy filesystem activity makes these fail, and we can
1675 	 * use compound pages.
1676 	 */
1677 	gfp_flags = __GFP_RECLAIM | __GFP_IO | __GFP_COMP;
1678 
1679 	/*
1680 	 * The minimum size of the eager buffers is a groups of MTU-sized
1681 	 * buffers.
1682 	 * The global eager_buffer_size parameter is checked against the
1683 	 * theoretical lower limit of the value. Here, we check against the
1684 	 * MTU.
1685 	 */
1686 	if (rcd->egrbufs.size < (round_mtu * dd->rcv_entries.group_size))
1687 		rcd->egrbufs.size = round_mtu * dd->rcv_entries.group_size;
1688 	/*
1689 	 * If using one-pkt-per-egr-buffer, lower the eager buffer
1690 	 * size to the max MTU (page-aligned).
1691 	 */
1692 	if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
1693 		rcd->egrbufs.rcvtid_size = round_mtu;
1694 
1695 	/*
1696 	 * Eager buffers sizes of 1MB or less require smaller TID sizes
1697 	 * to satisfy the "multiple of 8 RcvArray entries" requirement.
1698 	 */
1699 	if (rcd->egrbufs.size <= (1 << 20))
1700 		rcd->egrbufs.rcvtid_size = max((unsigned long)round_mtu,
1701 			rounddown_pow_of_two(rcd->egrbufs.size / 8));
1702 
1703 	while (alloced_bytes < rcd->egrbufs.size &&
1704 	       rcd->egrbufs.alloced < rcd->egrbufs.count) {
1705 		rcd->egrbufs.buffers[idx].addr =
1706 			dma_zalloc_coherent(&dd->pcidev->dev,
1707 					    rcd->egrbufs.rcvtid_size,
1708 					    &rcd->egrbufs.buffers[idx].phys,
1709 					    gfp_flags);
1710 		if (rcd->egrbufs.buffers[idx].addr) {
1711 			rcd->egrbufs.buffers[idx].len =
1712 				rcd->egrbufs.rcvtid_size;
1713 			rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].addr =
1714 				rcd->egrbufs.buffers[idx].addr;
1715 			rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].phys =
1716 				rcd->egrbufs.buffers[idx].phys;
1717 			rcd->egrbufs.alloced++;
1718 			alloced_bytes += rcd->egrbufs.rcvtid_size;
1719 			idx++;
1720 		} else {
1721 			u32 new_size, i, j;
1722 			u64 offset = 0;
1723 
1724 			/*
1725 			 * Fail the eager buffer allocation if:
1726 			 *   - we are already using the lowest acceptable size
1727 			 *   - we are using one-pkt-per-egr-buffer (this implies
1728 			 *     that we are accepting only one size)
1729 			 */
1730 			if (rcd->egrbufs.rcvtid_size == round_mtu ||
1731 			    !HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR)) {
1732 				dd_dev_err(dd, "ctxt%u: Failed to allocate eager buffers\n",
1733 					   rcd->ctxt);
1734 				goto bail_rcvegrbuf_phys;
1735 			}
1736 
1737 			new_size = rcd->egrbufs.rcvtid_size / 2;
1738 
1739 			/*
1740 			 * If the first attempt to allocate memory failed, don't
1741 			 * fail everything but continue with the next lower
1742 			 * size.
1743 			 */
1744 			if (idx == 0) {
1745 				rcd->egrbufs.rcvtid_size = new_size;
1746 				continue;
1747 			}
1748 
1749 			/*
1750 			 * Re-partition already allocated buffers to a smaller
1751 			 * size.
1752 			 */
1753 			rcd->egrbufs.alloced = 0;
1754 			for (i = 0, j = 0, offset = 0; j < idx; i++) {
1755 				if (i >= rcd->egrbufs.count)
1756 					break;
1757 				rcd->egrbufs.rcvtids[i].phys =
1758 					rcd->egrbufs.buffers[j].phys + offset;
1759 				rcd->egrbufs.rcvtids[i].addr =
1760 					rcd->egrbufs.buffers[j].addr + offset;
1761 				rcd->egrbufs.alloced++;
1762 				if ((rcd->egrbufs.buffers[j].phys + offset +
1763 				     new_size) ==
1764 				    (rcd->egrbufs.buffers[j].phys +
1765 				     rcd->egrbufs.buffers[j].len)) {
1766 					j++;
1767 					offset = 0;
1768 				} else {
1769 					offset += new_size;
1770 				}
1771 			}
1772 			rcd->egrbufs.rcvtid_size = new_size;
1773 		}
1774 	}
1775 	rcd->egrbufs.numbufs = idx;
1776 	rcd->egrbufs.size = alloced_bytes;
1777 
1778 	hfi1_cdbg(PROC,
1779 		  "ctxt%u: Alloced %u rcv tid entries @ %uKB, total %zuKB\n",
1780 		  rcd->ctxt, rcd->egrbufs.alloced,
1781 		  rcd->egrbufs.rcvtid_size / 1024, rcd->egrbufs.size / 1024);
1782 
1783 	/*
1784 	 * Set the contexts rcv array head update threshold to the closest
1785 	 * power of 2 (so we can use a mask instead of modulo) below half
1786 	 * the allocated entries.
1787 	 */
1788 	rcd->egrbufs.threshold =
1789 		rounddown_pow_of_two(rcd->egrbufs.alloced / 2);
1790 	/*
1791 	 * Compute the expected RcvArray entry base. This is done after
1792 	 * allocating the eager buffers in order to maximize the
1793 	 * expected RcvArray entries for the context.
1794 	 */
1795 	max_entries = rcd->rcv_array_groups * dd->rcv_entries.group_size;
1796 	egrtop = roundup(rcd->egrbufs.alloced, dd->rcv_entries.group_size);
1797 	rcd->expected_count = max_entries - egrtop;
1798 	if (rcd->expected_count > MAX_TID_PAIR_ENTRIES * 2)
1799 		rcd->expected_count = MAX_TID_PAIR_ENTRIES * 2;
1800 
1801 	rcd->expected_base = rcd->eager_base + egrtop;
1802 	hfi1_cdbg(PROC, "ctxt%u: eager:%u, exp:%u, egrbase:%u, expbase:%u\n",
1803 		  rcd->ctxt, rcd->egrbufs.alloced, rcd->expected_count,
1804 		  rcd->eager_base, rcd->expected_base);
1805 
1806 	if (!hfi1_rcvbuf_validate(rcd->egrbufs.rcvtid_size, PT_EAGER, &order)) {
1807 		hfi1_cdbg(PROC,
1808 			  "ctxt%u: current Eager buffer size is invalid %u\n",
1809 			  rcd->ctxt, rcd->egrbufs.rcvtid_size);
1810 		ret = -EINVAL;
1811 		goto bail;
1812 	}
1813 
1814 	for (idx = 0; idx < rcd->egrbufs.alloced; idx++) {
1815 		hfi1_put_tid(dd, rcd->eager_base + idx, PT_EAGER,
1816 			     rcd->egrbufs.rcvtids[idx].phys, order);
1817 		cond_resched();
1818 	}
1819 	goto bail;
1820 
1821 bail_rcvegrbuf_phys:
1822 	for (idx = 0; idx < rcd->egrbufs.alloced &&
1823 	     rcd->egrbufs.buffers[idx].addr;
1824 	     idx++) {
1825 		dma_free_coherent(&dd->pcidev->dev,
1826 				  rcd->egrbufs.buffers[idx].len,
1827 				  rcd->egrbufs.buffers[idx].addr,
1828 				  rcd->egrbufs.buffers[idx].phys);
1829 		rcd->egrbufs.buffers[idx].addr = NULL;
1830 		rcd->egrbufs.buffers[idx].phys = 0;
1831 		rcd->egrbufs.buffers[idx].len = 0;
1832 	}
1833 bail:
1834 	return ret;
1835 }
1836