xref: /openbmc/linux/drivers/infiniband/hw/hfi1/init.c (revision e15a5365)
1 /*
2  * Copyright(c) 2015 - 2020 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/xarray.h>
53 #include <linux/module.h>
54 #include <linux/printk.h>
55 #include <linux/hrtimer.h>
56 #include <linux/bitmap.h>
57 #include <linux/numa.h>
58 #include <rdma/rdma_vt.h>
59 
60 #include "hfi.h"
61 #include "device.h"
62 #include "common.h"
63 #include "trace.h"
64 #include "mad.h"
65 #include "sdma.h"
66 #include "debugfs.h"
67 #include "verbs.h"
68 #include "aspm.h"
69 #include "affinity.h"
70 #include "vnic.h"
71 #include "exp_rcv.h"
72 #include "netdev.h"
73 
74 #undef pr_fmt
75 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
76 
77 /*
78  * min buffers we want to have per context, after driver
79  */
80 #define HFI1_MIN_USER_CTXT_BUFCNT 7
81 
82 #define HFI1_MIN_EAGER_BUFFER_SIZE (4 * 1024) /* 4KB */
83 #define HFI1_MAX_EAGER_BUFFER_SIZE (256 * 1024) /* 256KB */
84 
85 #define NUM_IB_PORTS 1
86 
87 /*
88  * Number of user receive contexts we are configured to use (to allow for more
89  * pio buffers per ctxt, etc.)  Zero means use one user context per CPU.
90  */
91 int num_user_contexts = -1;
92 module_param_named(num_user_contexts, num_user_contexts, int, 0444);
93 MODULE_PARM_DESC(
94 	num_user_contexts, "Set max number of user contexts to use (default: -1 will use the real (non-HT) CPU count)");
95 
96 uint krcvqs[RXE_NUM_DATA_VL];
97 int krcvqsset;
98 module_param_array(krcvqs, uint, &krcvqsset, S_IRUGO);
99 MODULE_PARM_DESC(krcvqs, "Array of the number of non-control kernel receive queues by VL");
100 
101 /* computed based on above array */
102 unsigned long n_krcvqs;
103 
104 static unsigned hfi1_rcvarr_split = 25;
105 module_param_named(rcvarr_split, hfi1_rcvarr_split, uint, S_IRUGO);
106 MODULE_PARM_DESC(rcvarr_split, "Percent of context's RcvArray entries used for Eager buffers");
107 
108 static uint eager_buffer_size = (8 << 20); /* 8MB */
109 module_param(eager_buffer_size, uint, S_IRUGO);
110 MODULE_PARM_DESC(eager_buffer_size, "Size of the eager buffers, default: 8MB");
111 
112 static uint rcvhdrcnt = 2048; /* 2x the max eager buffer count */
113 module_param_named(rcvhdrcnt, rcvhdrcnt, uint, S_IRUGO);
114 MODULE_PARM_DESC(rcvhdrcnt, "Receive header queue count (default 2048)");
115 
116 static uint hfi1_hdrq_entsize = 32;
117 module_param_named(hdrq_entsize, hfi1_hdrq_entsize, uint, 0444);
118 MODULE_PARM_DESC(hdrq_entsize, "Size of header queue entries: 2 - 8B, 16 - 64B, 32 - 128B (default)");
119 
120 unsigned int user_credit_return_threshold = 33;	/* default is 33% */
121 module_param(user_credit_return_threshold, uint, S_IRUGO);
122 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)");
123 
124 DEFINE_XARRAY_FLAGS(hfi1_dev_table, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
125 
126 static int hfi1_create_kctxt(struct hfi1_devdata *dd,
127 			     struct hfi1_pportdata *ppd)
128 {
129 	struct hfi1_ctxtdata *rcd;
130 	int ret;
131 
132 	/* Control context has to be always 0 */
133 	BUILD_BUG_ON(HFI1_CTRL_CTXT != 0);
134 
135 	ret = hfi1_create_ctxtdata(ppd, dd->node, &rcd);
136 	if (ret < 0) {
137 		dd_dev_err(dd, "Kernel receive context allocation failed\n");
138 		return ret;
139 	}
140 
141 	/*
142 	 * Set up the kernel context flags here and now because they use
143 	 * default values for all receive side memories.  User contexts will
144 	 * be handled as they are created.
145 	 */
146 	rcd->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
147 		HFI1_CAP_KGET(NODROP_RHQ_FULL) |
148 		HFI1_CAP_KGET(NODROP_EGR_FULL) |
149 		HFI1_CAP_KGET(DMA_RTAIL);
150 
151 	/* Control context must use DMA_RTAIL */
152 	if (rcd->ctxt == HFI1_CTRL_CTXT)
153 		rcd->flags |= HFI1_CAP_DMA_RTAIL;
154 	rcd->fast_handler = get_dma_rtail_setting(rcd) ?
155 				handle_receive_interrupt_dma_rtail :
156 				handle_receive_interrupt_nodma_rtail;
157 	rcd->slow_handler = handle_receive_interrupt;
158 
159 	hfi1_set_seq_cnt(rcd, 1);
160 
161 	rcd->sc = sc_alloc(dd, SC_ACK, rcd->rcvhdrqentsize, dd->node);
162 	if (!rcd->sc) {
163 		dd_dev_err(dd, "Kernel send context allocation failed\n");
164 		return -ENOMEM;
165 	}
166 	hfi1_init_ctxt(rcd->sc);
167 
168 	return 0;
169 }
170 
171 /*
172  * Create the receive context array and one or more kernel contexts
173  */
174 int hfi1_create_kctxts(struct hfi1_devdata *dd)
175 {
176 	u16 i;
177 	int ret;
178 
179 	dd->rcd = kcalloc_node(dd->num_rcv_contexts, sizeof(*dd->rcd),
180 			       GFP_KERNEL, dd->node);
181 	if (!dd->rcd)
182 		return -ENOMEM;
183 
184 	for (i = 0; i < dd->first_dyn_alloc_ctxt; ++i) {
185 		ret = hfi1_create_kctxt(dd, dd->pport);
186 		if (ret)
187 			goto bail;
188 	}
189 
190 	return 0;
191 bail:
192 	for (i = 0; dd->rcd && i < dd->first_dyn_alloc_ctxt; ++i)
193 		hfi1_free_ctxt(dd->rcd[i]);
194 
195 	/* All the contexts should be freed, free the array */
196 	kfree(dd->rcd);
197 	dd->rcd = NULL;
198 	return ret;
199 }
200 
201 /*
202  * Helper routines for the receive context reference count (rcd and uctxt).
203  */
204 static void hfi1_rcd_init(struct hfi1_ctxtdata *rcd)
205 {
206 	kref_init(&rcd->kref);
207 }
208 
209 /**
210  * hfi1_rcd_free - When reference is zero clean up.
211  * @kref: pointer to an initialized rcd data structure
212  *
213  */
214 static void hfi1_rcd_free(struct kref *kref)
215 {
216 	unsigned long flags;
217 	struct hfi1_ctxtdata *rcd =
218 		container_of(kref, struct hfi1_ctxtdata, kref);
219 
220 	spin_lock_irqsave(&rcd->dd->uctxt_lock, flags);
221 	rcd->dd->rcd[rcd->ctxt] = NULL;
222 	spin_unlock_irqrestore(&rcd->dd->uctxt_lock, flags);
223 
224 	hfi1_free_ctxtdata(rcd->dd, rcd);
225 
226 	kfree(rcd);
227 }
228 
229 /**
230  * hfi1_rcd_put - decrement reference for rcd
231  * @rcd: pointer to an initialized rcd data structure
232  *
233  * Use this to put a reference after the init.
234  */
235 int hfi1_rcd_put(struct hfi1_ctxtdata *rcd)
236 {
237 	if (rcd)
238 		return kref_put(&rcd->kref, hfi1_rcd_free);
239 
240 	return 0;
241 }
242 
243 /**
244  * hfi1_rcd_get - increment reference for rcd
245  * @rcd: pointer to an initialized rcd data structure
246  *
247  * Use this to get a reference after the init.
248  *
249  * Return : reflect kref_get_unless_zero(), which returns non-zero on
250  * increment, otherwise 0.
251  */
252 int hfi1_rcd_get(struct hfi1_ctxtdata *rcd)
253 {
254 	return kref_get_unless_zero(&rcd->kref);
255 }
256 
257 /**
258  * allocate_rcd_index - allocate an rcd index from the rcd array
259  * @dd: pointer to a valid devdata structure
260  * @rcd: rcd data structure to assign
261  * @index: pointer to index that is allocated
262  *
263  * Find an empty index in the rcd array, and assign the given rcd to it.
264  * If the array is full, we are EBUSY.
265  *
266  */
267 static int allocate_rcd_index(struct hfi1_devdata *dd,
268 			      struct hfi1_ctxtdata *rcd, u16 *index)
269 {
270 	unsigned long flags;
271 	u16 ctxt;
272 
273 	spin_lock_irqsave(&dd->uctxt_lock, flags);
274 	for (ctxt = 0; ctxt < dd->num_rcv_contexts; ctxt++)
275 		if (!dd->rcd[ctxt])
276 			break;
277 
278 	if (ctxt < dd->num_rcv_contexts) {
279 		rcd->ctxt = ctxt;
280 		dd->rcd[ctxt] = rcd;
281 		hfi1_rcd_init(rcd);
282 	}
283 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
284 
285 	if (ctxt >= dd->num_rcv_contexts)
286 		return -EBUSY;
287 
288 	*index = ctxt;
289 
290 	return 0;
291 }
292 
293 /**
294  * hfi1_rcd_get_by_index_safe - validate the ctxt index before accessing the
295  * array
296  * @dd: pointer to a valid devdata structure
297  * @ctxt: the index of an possilbe rcd
298  *
299  * This is a wrapper for hfi1_rcd_get_by_index() to validate that the given
300  * ctxt index is valid.
301  *
302  * The caller is responsible for making the _put().
303  *
304  */
305 struct hfi1_ctxtdata *hfi1_rcd_get_by_index_safe(struct hfi1_devdata *dd,
306 						 u16 ctxt)
307 {
308 	if (ctxt < dd->num_rcv_contexts)
309 		return hfi1_rcd_get_by_index(dd, ctxt);
310 
311 	return NULL;
312 }
313 
314 /**
315  * hfi1_rcd_get_by_index
316  * @dd: pointer to a valid devdata structure
317  * @ctxt: the index of an possilbe rcd
318  *
319  * We need to protect access to the rcd array.  If access is needed to
320  * one or more index, get the protecting spinlock and then increment the
321  * kref.
322  *
323  * The caller is responsible for making the _put().
324  *
325  */
326 struct hfi1_ctxtdata *hfi1_rcd_get_by_index(struct hfi1_devdata *dd, u16 ctxt)
327 {
328 	unsigned long flags;
329 	struct hfi1_ctxtdata *rcd = NULL;
330 
331 	spin_lock_irqsave(&dd->uctxt_lock, flags);
332 	if (dd->rcd[ctxt]) {
333 		rcd = dd->rcd[ctxt];
334 		if (!hfi1_rcd_get(rcd))
335 			rcd = NULL;
336 	}
337 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
338 
339 	return rcd;
340 }
341 
342 /*
343  * Common code for user and kernel context create and setup.
344  * NOTE: the initial kref is done here (hf1_rcd_init()).
345  */
346 int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa,
347 			 struct hfi1_ctxtdata **context)
348 {
349 	struct hfi1_devdata *dd = ppd->dd;
350 	struct hfi1_ctxtdata *rcd;
351 	unsigned kctxt_ngroups = 0;
352 	u32 base;
353 
354 	if (dd->rcv_entries.nctxt_extra >
355 	    dd->num_rcv_contexts - dd->first_dyn_alloc_ctxt)
356 		kctxt_ngroups = (dd->rcv_entries.nctxt_extra -
357 			 (dd->num_rcv_contexts - dd->first_dyn_alloc_ctxt));
358 	rcd = kzalloc_node(sizeof(*rcd), GFP_KERNEL, numa);
359 	if (rcd) {
360 		u32 rcvtids, max_entries;
361 		u16 ctxt;
362 		int ret;
363 
364 		ret = allocate_rcd_index(dd, rcd, &ctxt);
365 		if (ret) {
366 			*context = NULL;
367 			kfree(rcd);
368 			return ret;
369 		}
370 
371 		INIT_LIST_HEAD(&rcd->qp_wait_list);
372 		hfi1_exp_tid_group_init(rcd);
373 		rcd->ppd = ppd;
374 		rcd->dd = dd;
375 		rcd->numa_id = numa;
376 		rcd->rcv_array_groups = dd->rcv_entries.ngroups;
377 		rcd->rhf_rcv_function_map = normal_rhf_rcv_functions;
378 		rcd->msix_intr = CCE_NUM_MSIX_VECTORS;
379 
380 		mutex_init(&rcd->exp_mutex);
381 		spin_lock_init(&rcd->exp_lock);
382 		INIT_LIST_HEAD(&rcd->flow_queue.queue_head);
383 		INIT_LIST_HEAD(&rcd->rarr_queue.queue_head);
384 
385 		hfi1_cdbg(PROC, "setting up context %u\n", rcd->ctxt);
386 
387 		/*
388 		 * Calculate the context's RcvArray entry starting point.
389 		 * We do this here because we have to take into account all
390 		 * the RcvArray entries that previous context would have
391 		 * taken and we have to account for any extra groups assigned
392 		 * to the static (kernel) or dynamic (vnic/user) contexts.
393 		 */
394 		if (ctxt < dd->first_dyn_alloc_ctxt) {
395 			if (ctxt < kctxt_ngroups) {
396 				base = ctxt * (dd->rcv_entries.ngroups + 1);
397 				rcd->rcv_array_groups++;
398 			} else {
399 				base = kctxt_ngroups +
400 					(ctxt * dd->rcv_entries.ngroups);
401 			}
402 		} else {
403 			u16 ct = ctxt - dd->first_dyn_alloc_ctxt;
404 
405 			base = ((dd->n_krcv_queues * dd->rcv_entries.ngroups) +
406 				kctxt_ngroups);
407 			if (ct < dd->rcv_entries.nctxt_extra) {
408 				base += ct * (dd->rcv_entries.ngroups + 1);
409 				rcd->rcv_array_groups++;
410 			} else {
411 				base += dd->rcv_entries.nctxt_extra +
412 					(ct * dd->rcv_entries.ngroups);
413 			}
414 		}
415 		rcd->eager_base = base * dd->rcv_entries.group_size;
416 
417 		rcd->rcvhdrq_cnt = rcvhdrcnt;
418 		rcd->rcvhdrqentsize = hfi1_hdrq_entsize;
419 		rcd->rhf_offset =
420 			rcd->rcvhdrqentsize - sizeof(u64) / sizeof(u32);
421 		/*
422 		 * Simple Eager buffer allocation: we have already pre-allocated
423 		 * the number of RcvArray entry groups. Each ctxtdata structure
424 		 * holds the number of groups for that context.
425 		 *
426 		 * To follow CSR requirements and maintain cacheline alignment,
427 		 * make sure all sizes and bases are multiples of group_size.
428 		 *
429 		 * The expected entry count is what is left after assigning
430 		 * eager.
431 		 */
432 		max_entries = rcd->rcv_array_groups *
433 			dd->rcv_entries.group_size;
434 		rcvtids = ((max_entries * hfi1_rcvarr_split) / 100);
435 		rcd->egrbufs.count = round_down(rcvtids,
436 						dd->rcv_entries.group_size);
437 		if (rcd->egrbufs.count > MAX_EAGER_ENTRIES) {
438 			dd_dev_err(dd, "ctxt%u: requested too many RcvArray entries.\n",
439 				   rcd->ctxt);
440 			rcd->egrbufs.count = MAX_EAGER_ENTRIES;
441 		}
442 		hfi1_cdbg(PROC,
443 			  "ctxt%u: max Eager buffer RcvArray entries: %u\n",
444 			  rcd->ctxt, rcd->egrbufs.count);
445 
446 		/*
447 		 * Allocate array that will hold the eager buffer accounting
448 		 * data.
449 		 * This will allocate the maximum possible buffer count based
450 		 * on the value of the RcvArray split parameter.
451 		 * The resulting value will be rounded down to the closest
452 		 * multiple of dd->rcv_entries.group_size.
453 		 */
454 		rcd->egrbufs.buffers =
455 			kcalloc_node(rcd->egrbufs.count,
456 				     sizeof(*rcd->egrbufs.buffers),
457 				     GFP_KERNEL, numa);
458 		if (!rcd->egrbufs.buffers)
459 			goto bail;
460 		rcd->egrbufs.rcvtids =
461 			kcalloc_node(rcd->egrbufs.count,
462 				     sizeof(*rcd->egrbufs.rcvtids),
463 				     GFP_KERNEL, numa);
464 		if (!rcd->egrbufs.rcvtids)
465 			goto bail;
466 		rcd->egrbufs.size = eager_buffer_size;
467 		/*
468 		 * The size of the buffers programmed into the RcvArray
469 		 * entries needs to be big enough to handle the highest
470 		 * MTU supported.
471 		 */
472 		if (rcd->egrbufs.size < hfi1_max_mtu) {
473 			rcd->egrbufs.size = __roundup_pow_of_two(hfi1_max_mtu);
474 			hfi1_cdbg(PROC,
475 				  "ctxt%u: eager bufs size too small. Adjusting to %u\n",
476 				    rcd->ctxt, rcd->egrbufs.size);
477 		}
478 		rcd->egrbufs.rcvtid_size = HFI1_MAX_EAGER_BUFFER_SIZE;
479 
480 		/* Applicable only for statically created kernel contexts */
481 		if (ctxt < dd->first_dyn_alloc_ctxt) {
482 			rcd->opstats = kzalloc_node(sizeof(*rcd->opstats),
483 						    GFP_KERNEL, numa);
484 			if (!rcd->opstats)
485 				goto bail;
486 
487 			/* Initialize TID flow generations for the context */
488 			hfi1_kern_init_ctxt_generations(rcd);
489 		}
490 
491 		*context = rcd;
492 		return 0;
493 	}
494 
495 bail:
496 	*context = NULL;
497 	hfi1_free_ctxt(rcd);
498 	return -ENOMEM;
499 }
500 
501 /**
502  * hfi1_free_ctxt
503  * @rcd: pointer to an initialized rcd data structure
504  *
505  * This wrapper is the free function that matches hfi1_create_ctxtdata().
506  * When a context is done being used (kernel or user), this function is called
507  * for the "final" put to match the kref init from hf1i_create_ctxtdata().
508  * Other users of the context do a get/put sequence to make sure that the
509  * structure isn't removed while in use.
510  */
511 void hfi1_free_ctxt(struct hfi1_ctxtdata *rcd)
512 {
513 	hfi1_rcd_put(rcd);
514 }
515 
516 /*
517  * Select the largest ccti value over all SLs to determine the intra-
518  * packet gap for the link.
519  *
520  * called with cca_timer_lock held (to protect access to cca_timer
521  * array), and rcu_read_lock() (to protect access to cc_state).
522  */
523 void set_link_ipg(struct hfi1_pportdata *ppd)
524 {
525 	struct hfi1_devdata *dd = ppd->dd;
526 	struct cc_state *cc_state;
527 	int i;
528 	u16 cce, ccti_limit, max_ccti = 0;
529 	u16 shift, mult;
530 	u64 src;
531 	u32 current_egress_rate; /* Mbits /sec */
532 	u32 max_pkt_time;
533 	/*
534 	 * max_pkt_time is the maximum packet egress time in units
535 	 * of the fabric clock period 1/(805 MHz).
536 	 */
537 
538 	cc_state = get_cc_state(ppd);
539 
540 	if (!cc_state)
541 		/*
542 		 * This should _never_ happen - rcu_read_lock() is held,
543 		 * and set_link_ipg() should not be called if cc_state
544 		 * is NULL.
545 		 */
546 		return;
547 
548 	for (i = 0; i < OPA_MAX_SLS; i++) {
549 		u16 ccti = ppd->cca_timer[i].ccti;
550 
551 		if (ccti > max_ccti)
552 			max_ccti = ccti;
553 	}
554 
555 	ccti_limit = cc_state->cct.ccti_limit;
556 	if (max_ccti > ccti_limit)
557 		max_ccti = ccti_limit;
558 
559 	cce = cc_state->cct.entries[max_ccti].entry;
560 	shift = (cce & 0xc000) >> 14;
561 	mult = (cce & 0x3fff);
562 
563 	current_egress_rate = active_egress_rate(ppd);
564 
565 	max_pkt_time = egress_cycles(ppd->ibmaxlen, current_egress_rate);
566 
567 	src = (max_pkt_time >> shift) * mult;
568 
569 	src &= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SMASK;
570 	src <<= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SHIFT;
571 
572 	write_csr(dd, SEND_STATIC_RATE_CONTROL, src);
573 }
574 
575 static enum hrtimer_restart cca_timer_fn(struct hrtimer *t)
576 {
577 	struct cca_timer *cca_timer;
578 	struct hfi1_pportdata *ppd;
579 	int sl;
580 	u16 ccti_timer, ccti_min;
581 	struct cc_state *cc_state;
582 	unsigned long flags;
583 	enum hrtimer_restart ret = HRTIMER_NORESTART;
584 
585 	cca_timer = container_of(t, struct cca_timer, hrtimer);
586 	ppd = cca_timer->ppd;
587 	sl = cca_timer->sl;
588 
589 	rcu_read_lock();
590 
591 	cc_state = get_cc_state(ppd);
592 
593 	if (!cc_state) {
594 		rcu_read_unlock();
595 		return HRTIMER_NORESTART;
596 	}
597 
598 	/*
599 	 * 1) decrement ccti for SL
600 	 * 2) calculate IPG for link (set_link_ipg())
601 	 * 3) restart timer, unless ccti is at min value
602 	 */
603 
604 	ccti_min = cc_state->cong_setting.entries[sl].ccti_min;
605 	ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
606 
607 	spin_lock_irqsave(&ppd->cca_timer_lock, flags);
608 
609 	if (cca_timer->ccti > ccti_min) {
610 		cca_timer->ccti--;
611 		set_link_ipg(ppd);
612 	}
613 
614 	if (cca_timer->ccti > ccti_min) {
615 		unsigned long nsec = 1024 * ccti_timer;
616 		/* ccti_timer is in units of 1.024 usec */
617 		hrtimer_forward_now(t, ns_to_ktime(nsec));
618 		ret = HRTIMER_RESTART;
619 	}
620 
621 	spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
622 	rcu_read_unlock();
623 	return ret;
624 }
625 
626 /*
627  * Common code for initializing the physical port structure.
628  */
629 void hfi1_init_pportdata(struct pci_dev *pdev, struct hfi1_pportdata *ppd,
630 			 struct hfi1_devdata *dd, u8 hw_pidx, u8 port)
631 {
632 	int i;
633 	uint default_pkey_idx;
634 	struct cc_state *cc_state;
635 
636 	ppd->dd = dd;
637 	ppd->hw_pidx = hw_pidx;
638 	ppd->port = port; /* IB port number, not index */
639 	ppd->prev_link_width = LINK_WIDTH_DEFAULT;
640 	/*
641 	 * There are C_VL_COUNT number of PortVLXmitWait counters.
642 	 * Adding 1 to C_VL_COUNT to include the PortXmitWait counter.
643 	 */
644 	for (i = 0; i < C_VL_COUNT + 1; i++) {
645 		ppd->port_vl_xmit_wait_last[i] = 0;
646 		ppd->vl_xmit_flit_cnt[i] = 0;
647 	}
648 
649 	default_pkey_idx = 1;
650 
651 	ppd->pkeys[default_pkey_idx] = DEFAULT_P_KEY;
652 	ppd->part_enforce |= HFI1_PART_ENFORCE_IN;
653 
654 	if (loopback) {
655 		dd_dev_err(dd, "Faking data partition 0x8001 in idx %u\n",
656 			   !default_pkey_idx);
657 		ppd->pkeys[!default_pkey_idx] = 0x8001;
658 	}
659 
660 	INIT_WORK(&ppd->link_vc_work, handle_verify_cap);
661 	INIT_WORK(&ppd->link_up_work, handle_link_up);
662 	INIT_WORK(&ppd->link_down_work, handle_link_down);
663 	INIT_WORK(&ppd->freeze_work, handle_freeze);
664 	INIT_WORK(&ppd->link_downgrade_work, handle_link_downgrade);
665 	INIT_WORK(&ppd->sma_message_work, handle_sma_message);
666 	INIT_WORK(&ppd->link_bounce_work, handle_link_bounce);
667 	INIT_DELAYED_WORK(&ppd->start_link_work, handle_start_link);
668 	INIT_WORK(&ppd->linkstate_active_work, receive_interrupt_work);
669 	INIT_WORK(&ppd->qsfp_info.qsfp_work, qsfp_event);
670 
671 	mutex_init(&ppd->hls_lock);
672 	spin_lock_init(&ppd->qsfp_info.qsfp_lock);
673 
674 	ppd->qsfp_info.ppd = ppd;
675 	ppd->sm_trap_qp = 0x0;
676 	ppd->sa_qp = 0x1;
677 
678 	ppd->hfi1_wq = NULL;
679 
680 	spin_lock_init(&ppd->cca_timer_lock);
681 
682 	for (i = 0; i < OPA_MAX_SLS; i++) {
683 		hrtimer_init(&ppd->cca_timer[i].hrtimer, CLOCK_MONOTONIC,
684 			     HRTIMER_MODE_REL);
685 		ppd->cca_timer[i].ppd = ppd;
686 		ppd->cca_timer[i].sl = i;
687 		ppd->cca_timer[i].ccti = 0;
688 		ppd->cca_timer[i].hrtimer.function = cca_timer_fn;
689 	}
690 
691 	ppd->cc_max_table_entries = IB_CC_TABLE_CAP_DEFAULT;
692 
693 	spin_lock_init(&ppd->cc_state_lock);
694 	spin_lock_init(&ppd->cc_log_lock);
695 	cc_state = kzalloc(sizeof(*cc_state), GFP_KERNEL);
696 	RCU_INIT_POINTER(ppd->cc_state, cc_state);
697 	if (!cc_state)
698 		goto bail;
699 	return;
700 
701 bail:
702 	dd_dev_err(dd, "Congestion Control Agent disabled for port %d\n", port);
703 }
704 
705 /*
706  * Do initialization for device that is only needed on
707  * first detect, not on resets.
708  */
709 static int loadtime_init(struct hfi1_devdata *dd)
710 {
711 	return 0;
712 }
713 
714 /**
715  * init_after_reset - re-initialize after a reset
716  * @dd: the hfi1_ib device
717  *
718  * sanity check at least some of the values after reset, and
719  * ensure no receive or transmit (explicitly, in case reset
720  * failed
721  */
722 static int init_after_reset(struct hfi1_devdata *dd)
723 {
724 	int i;
725 	struct hfi1_ctxtdata *rcd;
726 	/*
727 	 * Ensure chip does no sends or receives, tail updates, or
728 	 * pioavail updates while we re-initialize.  This is mostly
729 	 * for the driver data structures, not chip registers.
730 	 */
731 	for (i = 0; i < dd->num_rcv_contexts; i++) {
732 		rcd = hfi1_rcd_get_by_index(dd, i);
733 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
734 			     HFI1_RCVCTRL_INTRAVAIL_DIS |
735 			     HFI1_RCVCTRL_TAILUPD_DIS, rcd);
736 		hfi1_rcd_put(rcd);
737 	}
738 	pio_send_control(dd, PSC_GLOBAL_DISABLE);
739 	for (i = 0; i < dd->num_send_contexts; i++)
740 		sc_disable(dd->send_contexts[i].sc);
741 
742 	return 0;
743 }
744 
745 static void enable_chip(struct hfi1_devdata *dd)
746 {
747 	struct hfi1_ctxtdata *rcd;
748 	u32 rcvmask;
749 	u16 i;
750 
751 	/* enable PIO send */
752 	pio_send_control(dd, PSC_GLOBAL_ENABLE);
753 
754 	/*
755 	 * Enable kernel ctxts' receive and receive interrupt.
756 	 * Other ctxts done as user opens and initializes them.
757 	 */
758 	for (i = 0; i < dd->first_dyn_alloc_ctxt; ++i) {
759 		rcd = hfi1_rcd_get_by_index(dd, i);
760 		if (!rcd)
761 			continue;
762 		rcvmask = HFI1_RCVCTRL_CTXT_ENB | HFI1_RCVCTRL_INTRAVAIL_ENB;
763 		rcvmask |= HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ?
764 			HFI1_RCVCTRL_TAILUPD_ENB : HFI1_RCVCTRL_TAILUPD_DIS;
765 		if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
766 			rcvmask |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
767 		if (HFI1_CAP_KGET_MASK(rcd->flags, NODROP_RHQ_FULL))
768 			rcvmask |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
769 		if (HFI1_CAP_KGET_MASK(rcd->flags, NODROP_EGR_FULL))
770 			rcvmask |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
771 		if (HFI1_CAP_IS_KSET(TID_RDMA))
772 			rcvmask |= HFI1_RCVCTRL_TIDFLOW_ENB;
773 		hfi1_rcvctrl(dd, rcvmask, rcd);
774 		sc_enable(rcd->sc);
775 		hfi1_rcd_put(rcd);
776 	}
777 }
778 
779 /**
780  * create_workqueues - create per port workqueues
781  * @dd: the hfi1_ib device
782  */
783 static int create_workqueues(struct hfi1_devdata *dd)
784 {
785 	int pidx;
786 	struct hfi1_pportdata *ppd;
787 
788 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
789 		ppd = dd->pport + pidx;
790 		if (!ppd->hfi1_wq) {
791 			ppd->hfi1_wq =
792 				alloc_workqueue(
793 				    "hfi%d_%d",
794 				    WQ_SYSFS | WQ_HIGHPRI | WQ_CPU_INTENSIVE |
795 				    WQ_MEM_RECLAIM,
796 				    HFI1_MAX_ACTIVE_WORKQUEUE_ENTRIES,
797 				    dd->unit, pidx);
798 			if (!ppd->hfi1_wq)
799 				goto wq_error;
800 		}
801 		if (!ppd->link_wq) {
802 			/*
803 			 * Make the link workqueue single-threaded to enforce
804 			 * serialization.
805 			 */
806 			ppd->link_wq =
807 				alloc_workqueue(
808 				    "hfi_link_%d_%d",
809 				    WQ_SYSFS | WQ_MEM_RECLAIM | WQ_UNBOUND,
810 				    1, /* max_active */
811 				    dd->unit, pidx);
812 			if (!ppd->link_wq)
813 				goto wq_error;
814 		}
815 	}
816 	return 0;
817 wq_error:
818 	pr_err("alloc_workqueue failed for port %d\n", pidx + 1);
819 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
820 		ppd = dd->pport + pidx;
821 		if (ppd->hfi1_wq) {
822 			destroy_workqueue(ppd->hfi1_wq);
823 			ppd->hfi1_wq = NULL;
824 		}
825 		if (ppd->link_wq) {
826 			destroy_workqueue(ppd->link_wq);
827 			ppd->link_wq = NULL;
828 		}
829 	}
830 	return -ENOMEM;
831 }
832 
833 /**
834  * destroy_workqueues - destroy per port workqueues
835  * @dd: the hfi1_ib device
836  */
837 static void destroy_workqueues(struct hfi1_devdata *dd)
838 {
839 	int pidx;
840 	struct hfi1_pportdata *ppd;
841 
842 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
843 		ppd = dd->pport + pidx;
844 
845 		if (ppd->hfi1_wq) {
846 			destroy_workqueue(ppd->hfi1_wq);
847 			ppd->hfi1_wq = NULL;
848 		}
849 		if (ppd->link_wq) {
850 			destroy_workqueue(ppd->link_wq);
851 			ppd->link_wq = NULL;
852 		}
853 	}
854 }
855 
856 /**
857  * enable_general_intr() - Enable the IRQs that will be handled by the
858  * general interrupt handler.
859  * @dd: valid devdata
860  *
861  */
862 static void enable_general_intr(struct hfi1_devdata *dd)
863 {
864 	set_intr_bits(dd, CCE_ERR_INT, MISC_ERR_INT, true);
865 	set_intr_bits(dd, PIO_ERR_INT, TXE_ERR_INT, true);
866 	set_intr_bits(dd, IS_SENDCTXT_ERR_START, IS_SENDCTXT_ERR_END, true);
867 	set_intr_bits(dd, PBC_INT, GPIO_ASSERT_INT, true);
868 	set_intr_bits(dd, TCRIT_INT, TCRIT_INT, true);
869 	set_intr_bits(dd, IS_DC_START, IS_DC_END, true);
870 	set_intr_bits(dd, IS_SENDCREDIT_START, IS_SENDCREDIT_END, true);
871 }
872 
873 /**
874  * hfi1_init - do the actual initialization sequence on the chip
875  * @dd: the hfi1_ib device
876  * @reinit: re-initializing, so don't allocate new memory
877  *
878  * Do the actual initialization sequence on the chip.  This is done
879  * both from the init routine called from the PCI infrastructure, and
880  * when we reset the chip, or detect that it was reset internally,
881  * or it's administratively re-enabled.
882  *
883  * Memory allocation here and in called routines is only done in
884  * the first case (reinit == 0).  We have to be careful, because even
885  * without memory allocation, we need to re-write all the chip registers
886  * TIDs, etc. after the reset or enable has completed.
887  */
888 int hfi1_init(struct hfi1_devdata *dd, int reinit)
889 {
890 	int ret = 0, pidx, lastfail = 0;
891 	unsigned long len;
892 	u16 i;
893 	struct hfi1_ctxtdata *rcd;
894 	struct hfi1_pportdata *ppd;
895 
896 	/* Set up send low level handlers */
897 	dd->process_pio_send = hfi1_verbs_send_pio;
898 	dd->process_dma_send = hfi1_verbs_send_dma;
899 	dd->pio_inline_send = pio_copy;
900 	dd->process_vnic_dma_send = hfi1_vnic_send_dma;
901 
902 	if (is_ax(dd)) {
903 		atomic_set(&dd->drop_packet, DROP_PACKET_ON);
904 		dd->do_drop = true;
905 	} else {
906 		atomic_set(&dd->drop_packet, DROP_PACKET_OFF);
907 		dd->do_drop = false;
908 	}
909 
910 	/* make sure the link is not "up" */
911 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
912 		ppd = dd->pport + pidx;
913 		ppd->linkup = 0;
914 	}
915 
916 	if (reinit)
917 		ret = init_after_reset(dd);
918 	else
919 		ret = loadtime_init(dd);
920 	if (ret)
921 		goto done;
922 
923 	/* allocate dummy tail memory for all receive contexts */
924 	dd->rcvhdrtail_dummy_kvaddr = dma_alloc_coherent(&dd->pcidev->dev,
925 							 sizeof(u64),
926 							 &dd->rcvhdrtail_dummy_dma,
927 							 GFP_KERNEL);
928 
929 	if (!dd->rcvhdrtail_dummy_kvaddr) {
930 		dd_dev_err(dd, "cannot allocate dummy tail memory\n");
931 		ret = -ENOMEM;
932 		goto done;
933 	}
934 
935 	/* dd->rcd can be NULL if early initialization failed */
936 	for (i = 0; dd->rcd && i < dd->first_dyn_alloc_ctxt; ++i) {
937 		/*
938 		 * Set up the (kernel) rcvhdr queue and egr TIDs.  If doing
939 		 * re-init, the simplest way to handle this is to free
940 		 * existing, and re-allocate.
941 		 * Need to re-create rest of ctxt 0 ctxtdata as well.
942 		 */
943 		rcd = hfi1_rcd_get_by_index(dd, i);
944 		if (!rcd)
945 			continue;
946 
947 		rcd->do_interrupt = &handle_receive_interrupt;
948 
949 		lastfail = hfi1_create_rcvhdrq(dd, rcd);
950 		if (!lastfail)
951 			lastfail = hfi1_setup_eagerbufs(rcd);
952 		if (!lastfail)
953 			lastfail = hfi1_kern_exp_rcv_init(rcd, reinit);
954 		if (lastfail) {
955 			dd_dev_err(dd,
956 				   "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
957 			ret = lastfail;
958 		}
959 		/* enable IRQ */
960 		hfi1_rcd_put(rcd);
961 	}
962 
963 	/* Allocate enough memory for user event notification. */
964 	len = PAGE_ALIGN(chip_rcv_contexts(dd) * HFI1_MAX_SHARED_CTXTS *
965 			 sizeof(*dd->events));
966 	dd->events = vmalloc_user(len);
967 	if (!dd->events)
968 		dd_dev_err(dd, "Failed to allocate user events page\n");
969 	/*
970 	 * Allocate a page for device and port status.
971 	 * Page will be shared amongst all user processes.
972 	 */
973 	dd->status = vmalloc_user(PAGE_SIZE);
974 	if (!dd->status)
975 		dd_dev_err(dd, "Failed to allocate dev status page\n");
976 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
977 		ppd = dd->pport + pidx;
978 		if (dd->status)
979 			/* Currently, we only have one port */
980 			ppd->statusp = &dd->status->port;
981 
982 		set_mtu(ppd);
983 	}
984 
985 	/* enable chip even if we have an error, so we can debug cause */
986 	enable_chip(dd);
987 
988 done:
989 	/*
990 	 * Set status even if port serdes is not initialized
991 	 * so that diags will work.
992 	 */
993 	if (dd->status)
994 		dd->status->dev |= HFI1_STATUS_CHIP_PRESENT |
995 			HFI1_STATUS_INITTED;
996 	if (!ret) {
997 		/* enable all interrupts from the chip */
998 		enable_general_intr(dd);
999 		init_qsfp_int(dd);
1000 
1001 		/* chip is OK for user apps; mark it as initialized */
1002 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1003 			ppd = dd->pport + pidx;
1004 
1005 			/*
1006 			 * start the serdes - must be after interrupts are
1007 			 * enabled so we are notified when the link goes up
1008 			 */
1009 			lastfail = bringup_serdes(ppd);
1010 			if (lastfail)
1011 				dd_dev_info(dd,
1012 					    "Failed to bring up port %u\n",
1013 					    ppd->port);
1014 
1015 			/*
1016 			 * Set status even if port serdes is not initialized
1017 			 * so that diags will work.
1018 			 */
1019 			if (ppd->statusp)
1020 				*ppd->statusp |= HFI1_STATUS_CHIP_PRESENT |
1021 							HFI1_STATUS_INITTED;
1022 			if (!ppd->link_speed_enabled)
1023 				continue;
1024 		}
1025 	}
1026 
1027 	/* if ret is non-zero, we probably should do some cleanup here... */
1028 	return ret;
1029 }
1030 
1031 struct hfi1_devdata *hfi1_lookup(int unit)
1032 {
1033 	return xa_load(&hfi1_dev_table, unit);
1034 }
1035 
1036 /*
1037  * Stop the timers during unit shutdown, or after an error late
1038  * in initialization.
1039  */
1040 static void stop_timers(struct hfi1_devdata *dd)
1041 {
1042 	struct hfi1_pportdata *ppd;
1043 	int pidx;
1044 
1045 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1046 		ppd = dd->pport + pidx;
1047 		if (ppd->led_override_timer.function) {
1048 			del_timer_sync(&ppd->led_override_timer);
1049 			atomic_set(&ppd->led_override_timer_active, 0);
1050 		}
1051 	}
1052 }
1053 
1054 /**
1055  * shutdown_device - shut down a device
1056  * @dd: the hfi1_ib device
1057  *
1058  * This is called to make the device quiet when we are about to
1059  * unload the driver, and also when the device is administratively
1060  * disabled.   It does not free any data structures.
1061  * Everything it does has to be setup again by hfi1_init(dd, 1)
1062  */
1063 static void shutdown_device(struct hfi1_devdata *dd)
1064 {
1065 	struct hfi1_pportdata *ppd;
1066 	struct hfi1_ctxtdata *rcd;
1067 	unsigned pidx;
1068 	int i;
1069 
1070 	if (dd->flags & HFI1_SHUTDOWN)
1071 		return;
1072 	dd->flags |= HFI1_SHUTDOWN;
1073 
1074 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1075 		ppd = dd->pport + pidx;
1076 
1077 		ppd->linkup = 0;
1078 		if (ppd->statusp)
1079 			*ppd->statusp &= ~(HFI1_STATUS_IB_CONF |
1080 					   HFI1_STATUS_IB_READY);
1081 	}
1082 	dd->flags &= ~HFI1_INITTED;
1083 
1084 	/* mask and clean up interrupts */
1085 	set_intr_bits(dd, IS_FIRST_SOURCE, IS_LAST_SOURCE, false);
1086 	msix_clean_up_interrupts(dd);
1087 
1088 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1089 		ppd = dd->pport + pidx;
1090 		for (i = 0; i < dd->num_rcv_contexts; i++) {
1091 			rcd = hfi1_rcd_get_by_index(dd, i);
1092 			hfi1_rcvctrl(dd, HFI1_RCVCTRL_TAILUPD_DIS |
1093 				     HFI1_RCVCTRL_CTXT_DIS |
1094 				     HFI1_RCVCTRL_INTRAVAIL_DIS |
1095 				     HFI1_RCVCTRL_PKEY_DIS |
1096 				     HFI1_RCVCTRL_ONE_PKT_EGR_DIS, rcd);
1097 			hfi1_rcd_put(rcd);
1098 		}
1099 		/*
1100 		 * Gracefully stop all sends allowing any in progress to
1101 		 * trickle out first.
1102 		 */
1103 		for (i = 0; i < dd->num_send_contexts; i++)
1104 			sc_flush(dd->send_contexts[i].sc);
1105 	}
1106 
1107 	/*
1108 	 * Enough for anything that's going to trickle out to have actually
1109 	 * done so.
1110 	 */
1111 	udelay(20);
1112 
1113 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1114 		ppd = dd->pport + pidx;
1115 
1116 		/* disable all contexts */
1117 		for (i = 0; i < dd->num_send_contexts; i++)
1118 			sc_disable(dd->send_contexts[i].sc);
1119 		/* disable the send device */
1120 		pio_send_control(dd, PSC_GLOBAL_DISABLE);
1121 
1122 		shutdown_led_override(ppd);
1123 
1124 		/*
1125 		 * Clear SerdesEnable.
1126 		 * We can't count on interrupts since we are stopping.
1127 		 */
1128 		hfi1_quiet_serdes(ppd);
1129 		if (ppd->hfi1_wq)
1130 			flush_workqueue(ppd->hfi1_wq);
1131 		if (ppd->link_wq)
1132 			flush_workqueue(ppd->link_wq);
1133 	}
1134 	sdma_exit(dd);
1135 }
1136 
1137 /**
1138  * hfi1_free_ctxtdata - free a context's allocated data
1139  * @dd: the hfi1_ib device
1140  * @rcd: the ctxtdata structure
1141  *
1142  * free up any allocated data for a context
1143  * It should never change any chip state, or global driver state.
1144  */
1145 void hfi1_free_ctxtdata(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
1146 {
1147 	u32 e;
1148 
1149 	if (!rcd)
1150 		return;
1151 
1152 	if (rcd->rcvhdrq) {
1153 		dma_free_coherent(&dd->pcidev->dev, rcvhdrq_size(rcd),
1154 				  rcd->rcvhdrq, rcd->rcvhdrq_dma);
1155 		rcd->rcvhdrq = NULL;
1156 		if (hfi1_rcvhdrtail_kvaddr(rcd)) {
1157 			dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1158 					  (void *)hfi1_rcvhdrtail_kvaddr(rcd),
1159 					  rcd->rcvhdrqtailaddr_dma);
1160 			rcd->rcvhdrtail_kvaddr = NULL;
1161 		}
1162 	}
1163 
1164 	/* all the RcvArray entries should have been cleared by now */
1165 	kfree(rcd->egrbufs.rcvtids);
1166 	rcd->egrbufs.rcvtids = NULL;
1167 
1168 	for (e = 0; e < rcd->egrbufs.alloced; e++) {
1169 		if (rcd->egrbufs.buffers[e].dma)
1170 			dma_free_coherent(&dd->pcidev->dev,
1171 					  rcd->egrbufs.buffers[e].len,
1172 					  rcd->egrbufs.buffers[e].addr,
1173 					  rcd->egrbufs.buffers[e].dma);
1174 	}
1175 	kfree(rcd->egrbufs.buffers);
1176 	rcd->egrbufs.alloced = 0;
1177 	rcd->egrbufs.buffers = NULL;
1178 
1179 	sc_free(rcd->sc);
1180 	rcd->sc = NULL;
1181 
1182 	vfree(rcd->subctxt_uregbase);
1183 	vfree(rcd->subctxt_rcvegrbuf);
1184 	vfree(rcd->subctxt_rcvhdr_base);
1185 	kfree(rcd->opstats);
1186 
1187 	rcd->subctxt_uregbase = NULL;
1188 	rcd->subctxt_rcvegrbuf = NULL;
1189 	rcd->subctxt_rcvhdr_base = NULL;
1190 	rcd->opstats = NULL;
1191 }
1192 
1193 /*
1194  * Release our hold on the shared asic data.  If we are the last one,
1195  * return the structure to be finalized outside the lock.  Must be
1196  * holding hfi1_dev_table lock.
1197  */
1198 static struct hfi1_asic_data *release_asic_data(struct hfi1_devdata *dd)
1199 {
1200 	struct hfi1_asic_data *ad;
1201 	int other;
1202 
1203 	if (!dd->asic_data)
1204 		return NULL;
1205 	dd->asic_data->dds[dd->hfi1_id] = NULL;
1206 	other = dd->hfi1_id ? 0 : 1;
1207 	ad = dd->asic_data;
1208 	dd->asic_data = NULL;
1209 	/* return NULL if the other dd still has a link */
1210 	return ad->dds[other] ? NULL : ad;
1211 }
1212 
1213 static void finalize_asic_data(struct hfi1_devdata *dd,
1214 			       struct hfi1_asic_data *ad)
1215 {
1216 	clean_up_i2c(dd, ad);
1217 	kfree(ad);
1218 }
1219 
1220 /**
1221  * hfi1_free_devdata - cleans up and frees per-unit data structure
1222  * @dd: pointer to a valid devdata structure
1223  *
1224  * It cleans up and frees all data structures set up by
1225  * by hfi1_alloc_devdata().
1226  */
1227 void hfi1_free_devdata(struct hfi1_devdata *dd)
1228 {
1229 	struct hfi1_asic_data *ad;
1230 	unsigned long flags;
1231 
1232 	xa_lock_irqsave(&hfi1_dev_table, flags);
1233 	__xa_erase(&hfi1_dev_table, dd->unit);
1234 	ad = release_asic_data(dd);
1235 	xa_unlock_irqrestore(&hfi1_dev_table, flags);
1236 
1237 	finalize_asic_data(dd, ad);
1238 	free_platform_config(dd);
1239 	rcu_barrier(); /* wait for rcu callbacks to complete */
1240 	free_percpu(dd->int_counter);
1241 	free_percpu(dd->rcv_limit);
1242 	free_percpu(dd->send_schedule);
1243 	free_percpu(dd->tx_opstats);
1244 	dd->int_counter   = NULL;
1245 	dd->rcv_limit     = NULL;
1246 	dd->send_schedule = NULL;
1247 	dd->tx_opstats    = NULL;
1248 	kfree(dd->comp_vect);
1249 	dd->comp_vect = NULL;
1250 	sdma_clean(dd, dd->num_sdma);
1251 	rvt_dealloc_device(&dd->verbs_dev.rdi);
1252 }
1253 
1254 /**
1255  * hfi1_alloc_devdata - Allocate our primary per-unit data structure.
1256  * @pdev: Valid PCI device
1257  * @extra: How many bytes to alloc past the default
1258  *
1259  * Must be done via verbs allocator, because the verbs cleanup process
1260  * both does cleanup and free of the data structure.
1261  * "extra" is for chip-specific data.
1262  */
1263 static struct hfi1_devdata *hfi1_alloc_devdata(struct pci_dev *pdev,
1264 					       size_t extra)
1265 {
1266 	struct hfi1_devdata *dd;
1267 	int ret, nports;
1268 
1269 	/* extra is * number of ports */
1270 	nports = extra / sizeof(struct hfi1_pportdata);
1271 
1272 	dd = (struct hfi1_devdata *)rvt_alloc_device(sizeof(*dd) + extra,
1273 						     nports);
1274 	if (!dd)
1275 		return ERR_PTR(-ENOMEM);
1276 	dd->num_pports = nports;
1277 	dd->pport = (struct hfi1_pportdata *)(dd + 1);
1278 	dd->pcidev = pdev;
1279 	pci_set_drvdata(pdev, dd);
1280 	dd->node = NUMA_NO_NODE;
1281 
1282 	ret = xa_alloc_irq(&hfi1_dev_table, &dd->unit, dd, xa_limit_32b,
1283 			GFP_KERNEL);
1284 	if (ret < 0) {
1285 		dev_err(&pdev->dev,
1286 			"Could not allocate unit ID: error %d\n", -ret);
1287 		goto bail;
1288 	}
1289 	rvt_set_ibdev_name(&dd->verbs_dev.rdi, "%s_%d", class_name(), dd->unit);
1290 
1291 	/*
1292 	 * Initialize all locks for the device. This needs to be as early as
1293 	 * possible so locks are usable.
1294 	 */
1295 	spin_lock_init(&dd->sc_lock);
1296 	spin_lock_init(&dd->sendctrl_lock);
1297 	spin_lock_init(&dd->rcvctrl_lock);
1298 	spin_lock_init(&dd->uctxt_lock);
1299 	spin_lock_init(&dd->hfi1_diag_trans_lock);
1300 	spin_lock_init(&dd->sc_init_lock);
1301 	spin_lock_init(&dd->dc8051_memlock);
1302 	seqlock_init(&dd->sc2vl_lock);
1303 	spin_lock_init(&dd->sde_map_lock);
1304 	spin_lock_init(&dd->pio_map_lock);
1305 	mutex_init(&dd->dc8051_lock);
1306 	init_waitqueue_head(&dd->event_queue);
1307 	spin_lock_init(&dd->irq_src_lock);
1308 
1309 	dd->int_counter = alloc_percpu(u64);
1310 	if (!dd->int_counter) {
1311 		ret = -ENOMEM;
1312 		goto bail;
1313 	}
1314 
1315 	dd->rcv_limit = alloc_percpu(u64);
1316 	if (!dd->rcv_limit) {
1317 		ret = -ENOMEM;
1318 		goto bail;
1319 	}
1320 
1321 	dd->send_schedule = alloc_percpu(u64);
1322 	if (!dd->send_schedule) {
1323 		ret = -ENOMEM;
1324 		goto bail;
1325 	}
1326 
1327 	dd->tx_opstats = alloc_percpu(struct hfi1_opcode_stats_perctx);
1328 	if (!dd->tx_opstats) {
1329 		ret = -ENOMEM;
1330 		goto bail;
1331 	}
1332 
1333 	dd->comp_vect = kzalloc(sizeof(*dd->comp_vect), GFP_KERNEL);
1334 	if (!dd->comp_vect) {
1335 		ret = -ENOMEM;
1336 		goto bail;
1337 	}
1338 
1339 	atomic_set(&dd->ipoib_rsm_usr_num, 0);
1340 	return dd;
1341 
1342 bail:
1343 	hfi1_free_devdata(dd);
1344 	return ERR_PTR(ret);
1345 }
1346 
1347 /*
1348  * Called from freeze mode handlers, and from PCI error
1349  * reporting code.  Should be paranoid about state of
1350  * system and data structures.
1351  */
1352 void hfi1_disable_after_error(struct hfi1_devdata *dd)
1353 {
1354 	if (dd->flags & HFI1_INITTED) {
1355 		u32 pidx;
1356 
1357 		dd->flags &= ~HFI1_INITTED;
1358 		if (dd->pport)
1359 			for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1360 				struct hfi1_pportdata *ppd;
1361 
1362 				ppd = dd->pport + pidx;
1363 				if (dd->flags & HFI1_PRESENT)
1364 					set_link_state(ppd, HLS_DN_DISABLE);
1365 
1366 				if (ppd->statusp)
1367 					*ppd->statusp &= ~HFI1_STATUS_IB_READY;
1368 			}
1369 	}
1370 
1371 	/*
1372 	 * Mark as having had an error for driver, and also
1373 	 * for /sys and status word mapped to user programs.
1374 	 * This marks unit as not usable, until reset.
1375 	 */
1376 	if (dd->status)
1377 		dd->status->dev |= HFI1_STATUS_HWERROR;
1378 }
1379 
1380 static void remove_one(struct pci_dev *);
1381 static int init_one(struct pci_dev *, const struct pci_device_id *);
1382 static void shutdown_one(struct pci_dev *);
1383 
1384 #define DRIVER_LOAD_MSG "Intel " DRIVER_NAME " loaded: "
1385 #define PFX DRIVER_NAME ": "
1386 
1387 const struct pci_device_id hfi1_pci_tbl[] = {
1388 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL0) },
1389 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL1) },
1390 	{ 0, }
1391 };
1392 
1393 MODULE_DEVICE_TABLE(pci, hfi1_pci_tbl);
1394 
1395 static struct pci_driver hfi1_pci_driver = {
1396 	.name = DRIVER_NAME,
1397 	.probe = init_one,
1398 	.remove = remove_one,
1399 	.shutdown = shutdown_one,
1400 	.id_table = hfi1_pci_tbl,
1401 	.err_handler = &hfi1_pci_err_handler,
1402 };
1403 
1404 static void __init compute_krcvqs(void)
1405 {
1406 	int i;
1407 
1408 	for (i = 0; i < krcvqsset; i++)
1409 		n_krcvqs += krcvqs[i];
1410 }
1411 
1412 /*
1413  * Do all the generic driver unit- and chip-independent memory
1414  * allocation and initialization.
1415  */
1416 static int __init hfi1_mod_init(void)
1417 {
1418 	int ret;
1419 
1420 	ret = dev_init();
1421 	if (ret)
1422 		goto bail;
1423 
1424 	ret = node_affinity_init();
1425 	if (ret)
1426 		goto bail;
1427 
1428 	/* validate max MTU before any devices start */
1429 	if (!valid_opa_max_mtu(hfi1_max_mtu)) {
1430 		pr_err("Invalid max_mtu 0x%x, using 0x%x instead\n",
1431 		       hfi1_max_mtu, HFI1_DEFAULT_MAX_MTU);
1432 		hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
1433 	}
1434 	/* valid CUs run from 1-128 in powers of 2 */
1435 	if (hfi1_cu > 128 || !is_power_of_2(hfi1_cu))
1436 		hfi1_cu = 1;
1437 	/* valid credit return threshold is 0-100, variable is unsigned */
1438 	if (user_credit_return_threshold > 100)
1439 		user_credit_return_threshold = 100;
1440 
1441 	compute_krcvqs();
1442 	/*
1443 	 * sanitize receive interrupt count, time must wait until after
1444 	 * the hardware type is known
1445 	 */
1446 	if (rcv_intr_count > RCV_HDR_HEAD_COUNTER_MASK)
1447 		rcv_intr_count = RCV_HDR_HEAD_COUNTER_MASK;
1448 	/* reject invalid combinations */
1449 	if (rcv_intr_count == 0 && rcv_intr_timeout == 0) {
1450 		pr_err("Invalid mode: both receive interrupt count and available timeout are zero - setting interrupt count to 1\n");
1451 		rcv_intr_count = 1;
1452 	}
1453 	if (rcv_intr_count > 1 && rcv_intr_timeout == 0) {
1454 		/*
1455 		 * Avoid indefinite packet delivery by requiring a timeout
1456 		 * if count is > 1.
1457 		 */
1458 		pr_err("Invalid mode: receive interrupt count greater than 1 and available timeout is zero - setting available timeout to 1\n");
1459 		rcv_intr_timeout = 1;
1460 	}
1461 	if (rcv_intr_dynamic && !(rcv_intr_count > 1 && rcv_intr_timeout > 0)) {
1462 		/*
1463 		 * The dynamic algorithm expects a non-zero timeout
1464 		 * and a count > 1.
1465 		 */
1466 		pr_err("Invalid mode: dynamic receive interrupt mitigation with invalid count and timeout - turning dynamic off\n");
1467 		rcv_intr_dynamic = 0;
1468 	}
1469 
1470 	/* sanitize link CRC options */
1471 	link_crc_mask &= SUPPORTED_CRCS;
1472 
1473 	ret = opfn_init();
1474 	if (ret < 0) {
1475 		pr_err("Failed to allocate opfn_wq");
1476 		goto bail_dev;
1477 	}
1478 
1479 	/*
1480 	 * These must be called before the driver is registered with
1481 	 * the PCI subsystem.
1482 	 */
1483 	hfi1_dbg_init();
1484 	ret = pci_register_driver(&hfi1_pci_driver);
1485 	if (ret < 0) {
1486 		pr_err("Unable to register driver: error %d\n", -ret);
1487 		goto bail_dev;
1488 	}
1489 	goto bail; /* all OK */
1490 
1491 bail_dev:
1492 	hfi1_dbg_exit();
1493 	dev_cleanup();
1494 bail:
1495 	return ret;
1496 }
1497 
1498 module_init(hfi1_mod_init);
1499 
1500 /*
1501  * Do the non-unit driver cleanup, memory free, etc. at unload.
1502  */
1503 static void __exit hfi1_mod_cleanup(void)
1504 {
1505 	pci_unregister_driver(&hfi1_pci_driver);
1506 	opfn_exit();
1507 	node_affinity_destroy_all();
1508 	hfi1_dbg_exit();
1509 
1510 	WARN_ON(!xa_empty(&hfi1_dev_table));
1511 	dispose_firmware();	/* asymmetric with obtain_firmware() */
1512 	dev_cleanup();
1513 }
1514 
1515 module_exit(hfi1_mod_cleanup);
1516 
1517 /* this can only be called after a successful initialization */
1518 static void cleanup_device_data(struct hfi1_devdata *dd)
1519 {
1520 	int ctxt;
1521 	int pidx;
1522 
1523 	/* users can't do anything more with chip */
1524 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1525 		struct hfi1_pportdata *ppd = &dd->pport[pidx];
1526 		struct cc_state *cc_state;
1527 		int i;
1528 
1529 		if (ppd->statusp)
1530 			*ppd->statusp &= ~HFI1_STATUS_CHIP_PRESENT;
1531 
1532 		for (i = 0; i < OPA_MAX_SLS; i++)
1533 			hrtimer_cancel(&ppd->cca_timer[i].hrtimer);
1534 
1535 		spin_lock(&ppd->cc_state_lock);
1536 		cc_state = get_cc_state_protected(ppd);
1537 		RCU_INIT_POINTER(ppd->cc_state, NULL);
1538 		spin_unlock(&ppd->cc_state_lock);
1539 
1540 		if (cc_state)
1541 			kfree_rcu(cc_state, rcu);
1542 	}
1543 
1544 	free_credit_return(dd);
1545 
1546 	if (dd->rcvhdrtail_dummy_kvaddr) {
1547 		dma_free_coherent(&dd->pcidev->dev, sizeof(u64),
1548 				  (void *)dd->rcvhdrtail_dummy_kvaddr,
1549 				  dd->rcvhdrtail_dummy_dma);
1550 		dd->rcvhdrtail_dummy_kvaddr = NULL;
1551 	}
1552 
1553 	/*
1554 	 * Free any resources still in use (usually just kernel contexts)
1555 	 * at unload; we do for ctxtcnt, because that's what we allocate.
1556 	 */
1557 	for (ctxt = 0; dd->rcd && ctxt < dd->num_rcv_contexts; ctxt++) {
1558 		struct hfi1_ctxtdata *rcd = dd->rcd[ctxt];
1559 
1560 		if (rcd) {
1561 			hfi1_free_ctxt_rcv_groups(rcd);
1562 			hfi1_free_ctxt(rcd);
1563 		}
1564 	}
1565 
1566 	kfree(dd->rcd);
1567 	dd->rcd = NULL;
1568 
1569 	free_pio_map(dd);
1570 	/* must follow rcv context free - need to remove rcv's hooks */
1571 	for (ctxt = 0; ctxt < dd->num_send_contexts; ctxt++)
1572 		sc_free(dd->send_contexts[ctxt].sc);
1573 	dd->num_send_contexts = 0;
1574 	kfree(dd->send_contexts);
1575 	dd->send_contexts = NULL;
1576 	kfree(dd->hw_to_sw);
1577 	dd->hw_to_sw = NULL;
1578 	kfree(dd->boardname);
1579 	vfree(dd->events);
1580 	vfree(dd->status);
1581 }
1582 
1583 /*
1584  * Clean up on unit shutdown, or error during unit load after
1585  * successful initialization.
1586  */
1587 static void postinit_cleanup(struct hfi1_devdata *dd)
1588 {
1589 	hfi1_start_cleanup(dd);
1590 	hfi1_comp_vectors_clean_up(dd);
1591 	hfi1_dev_affinity_clean_up(dd);
1592 
1593 	hfi1_pcie_ddcleanup(dd);
1594 	hfi1_pcie_cleanup(dd->pcidev);
1595 
1596 	cleanup_device_data(dd);
1597 
1598 	hfi1_free_devdata(dd);
1599 }
1600 
1601 static int init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
1602 {
1603 	int ret = 0, j, pidx, initfail;
1604 	struct hfi1_devdata *dd;
1605 	struct hfi1_pportdata *ppd;
1606 
1607 	/* First, lock the non-writable module parameters */
1608 	HFI1_CAP_LOCK();
1609 
1610 	/* Validate dev ids */
1611 	if (!(ent->device == PCI_DEVICE_ID_INTEL0 ||
1612 	      ent->device == PCI_DEVICE_ID_INTEL1)) {
1613 		dev_err(&pdev->dev, "Failing on unknown Intel deviceid 0x%x\n",
1614 			ent->device);
1615 		ret = -ENODEV;
1616 		goto bail;
1617 	}
1618 
1619 	/* Allocate the dd so we can get to work */
1620 	dd = hfi1_alloc_devdata(pdev, NUM_IB_PORTS *
1621 				sizeof(struct hfi1_pportdata));
1622 	if (IS_ERR(dd)) {
1623 		ret = PTR_ERR(dd);
1624 		goto bail;
1625 	}
1626 
1627 	/* Validate some global module parameters */
1628 	ret = hfi1_validate_rcvhdrcnt(dd, rcvhdrcnt);
1629 	if (ret)
1630 		goto bail;
1631 
1632 	/* use the encoding function as a sanitization check */
1633 	if (!encode_rcv_header_entry_size(hfi1_hdrq_entsize)) {
1634 		dd_dev_err(dd, "Invalid HdrQ Entry size %u\n",
1635 			   hfi1_hdrq_entsize);
1636 		ret = -EINVAL;
1637 		goto bail;
1638 	}
1639 
1640 	/* The receive eager buffer size must be set before the receive
1641 	 * contexts are created.
1642 	 *
1643 	 * Set the eager buffer size.  Validate that it falls in a range
1644 	 * allowed by the hardware - all powers of 2 between the min and
1645 	 * max.  The maximum valid MTU is within the eager buffer range
1646 	 * so we do not need to cap the max_mtu by an eager buffer size
1647 	 * setting.
1648 	 */
1649 	if (eager_buffer_size) {
1650 		if (!is_power_of_2(eager_buffer_size))
1651 			eager_buffer_size =
1652 				roundup_pow_of_two(eager_buffer_size);
1653 		eager_buffer_size =
1654 			clamp_val(eager_buffer_size,
1655 				  MIN_EAGER_BUFFER * 8,
1656 				  MAX_EAGER_BUFFER_TOTAL);
1657 		dd_dev_info(dd, "Eager buffer size %u\n",
1658 			    eager_buffer_size);
1659 	} else {
1660 		dd_dev_err(dd, "Invalid Eager buffer size of 0\n");
1661 		ret = -EINVAL;
1662 		goto bail;
1663 	}
1664 
1665 	/* restrict value of hfi1_rcvarr_split */
1666 	hfi1_rcvarr_split = clamp_val(hfi1_rcvarr_split, 0, 100);
1667 
1668 	ret = hfi1_pcie_init(dd);
1669 	if (ret)
1670 		goto bail;
1671 
1672 	/*
1673 	 * Do device-specific initialization, function table setup, dd
1674 	 * allocation, etc.
1675 	 */
1676 	ret = hfi1_init_dd(dd);
1677 	if (ret)
1678 		goto clean_bail; /* error already printed */
1679 
1680 	ret = create_workqueues(dd);
1681 	if (ret)
1682 		goto clean_bail;
1683 
1684 	/* do the generic initialization */
1685 	initfail = hfi1_init(dd, 0);
1686 
1687 	ret = hfi1_register_ib_device(dd);
1688 
1689 	/*
1690 	 * Now ready for use.  this should be cleared whenever we
1691 	 * detect a reset, or initiate one.  If earlier failure,
1692 	 * we still create devices, so diags, etc. can be used
1693 	 * to determine cause of problem.
1694 	 */
1695 	if (!initfail && !ret) {
1696 		dd->flags |= HFI1_INITTED;
1697 		/* create debufs files after init and ib register */
1698 		hfi1_dbg_ibdev_init(&dd->verbs_dev);
1699 	}
1700 
1701 	j = hfi1_device_create(dd);
1702 	if (j)
1703 		dd_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1704 
1705 	if (initfail || ret) {
1706 		msix_clean_up_interrupts(dd);
1707 		stop_timers(dd);
1708 		flush_workqueue(ib_wq);
1709 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1710 			hfi1_quiet_serdes(dd->pport + pidx);
1711 			ppd = dd->pport + pidx;
1712 			if (ppd->hfi1_wq) {
1713 				destroy_workqueue(ppd->hfi1_wq);
1714 				ppd->hfi1_wq = NULL;
1715 			}
1716 			if (ppd->link_wq) {
1717 				destroy_workqueue(ppd->link_wq);
1718 				ppd->link_wq = NULL;
1719 			}
1720 		}
1721 		if (!j)
1722 			hfi1_device_remove(dd);
1723 		if (!ret)
1724 			hfi1_unregister_ib_device(dd);
1725 		postinit_cleanup(dd);
1726 		if (initfail)
1727 			ret = initfail;
1728 		goto bail;	/* everything already cleaned */
1729 	}
1730 
1731 	sdma_start(dd);
1732 
1733 	return 0;
1734 
1735 clean_bail:
1736 	hfi1_pcie_cleanup(pdev);
1737 bail:
1738 	return ret;
1739 }
1740 
1741 static void wait_for_clients(struct hfi1_devdata *dd)
1742 {
1743 	/*
1744 	 * Remove the device init value and complete the device if there is
1745 	 * no clients or wait for active clients to finish.
1746 	 */
1747 	if (atomic_dec_and_test(&dd->user_refcount))
1748 		complete(&dd->user_comp);
1749 
1750 	wait_for_completion(&dd->user_comp);
1751 }
1752 
1753 static void remove_one(struct pci_dev *pdev)
1754 {
1755 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
1756 
1757 	/* close debugfs files before ib unregister */
1758 	hfi1_dbg_ibdev_exit(&dd->verbs_dev);
1759 
1760 	/* remove the /dev hfi1 interface */
1761 	hfi1_device_remove(dd);
1762 
1763 	/* wait for existing user space clients to finish */
1764 	wait_for_clients(dd);
1765 
1766 	/* unregister from IB core */
1767 	hfi1_unregister_ib_device(dd);
1768 
1769 	/* free netdev data */
1770 	hfi1_netdev_free(dd);
1771 
1772 	/*
1773 	 * Disable the IB link, disable interrupts on the device,
1774 	 * clear dma engines, etc.
1775 	 */
1776 	shutdown_device(dd);
1777 	destroy_workqueues(dd);
1778 
1779 	stop_timers(dd);
1780 
1781 	/* wait until all of our (qsfp) queue_work() calls complete */
1782 	flush_workqueue(ib_wq);
1783 
1784 	postinit_cleanup(dd);
1785 }
1786 
1787 static void shutdown_one(struct pci_dev *pdev)
1788 {
1789 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
1790 
1791 	shutdown_device(dd);
1792 }
1793 
1794 /**
1795  * hfi1_create_rcvhdrq - create a receive header queue
1796  * @dd: the hfi1_ib device
1797  * @rcd: the context data
1798  *
1799  * This must be contiguous memory (from an i/o perspective), and must be
1800  * DMA'able (which means for some systems, it will go through an IOMMU,
1801  * or be forced into a low address range).
1802  */
1803 int hfi1_create_rcvhdrq(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
1804 {
1805 	unsigned amt;
1806 
1807 	if (!rcd->rcvhdrq) {
1808 		gfp_t gfp_flags;
1809 
1810 		amt = rcvhdrq_size(rcd);
1811 
1812 		if (rcd->ctxt < dd->first_dyn_alloc_ctxt || rcd->is_vnic)
1813 			gfp_flags = GFP_KERNEL;
1814 		else
1815 			gfp_flags = GFP_USER;
1816 		rcd->rcvhdrq = dma_alloc_coherent(&dd->pcidev->dev, amt,
1817 						  &rcd->rcvhdrq_dma,
1818 						  gfp_flags | __GFP_COMP);
1819 
1820 		if (!rcd->rcvhdrq) {
1821 			dd_dev_err(dd,
1822 				   "attempt to allocate %d bytes for ctxt %u rcvhdrq failed\n",
1823 				   amt, rcd->ctxt);
1824 			goto bail;
1825 		}
1826 
1827 		if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ||
1828 		    HFI1_CAP_UGET_MASK(rcd->flags, DMA_RTAIL)) {
1829 			rcd->rcvhdrtail_kvaddr = dma_alloc_coherent(&dd->pcidev->dev,
1830 								    PAGE_SIZE,
1831 								    &rcd->rcvhdrqtailaddr_dma,
1832 								    gfp_flags);
1833 			if (!rcd->rcvhdrtail_kvaddr)
1834 				goto bail_free;
1835 		}
1836 	}
1837 
1838 	set_hdrq_regs(rcd->dd, rcd->ctxt, rcd->rcvhdrqentsize,
1839 		      rcd->rcvhdrq_cnt);
1840 
1841 	return 0;
1842 
1843 bail_free:
1844 	dd_dev_err(dd,
1845 		   "attempt to allocate 1 page for ctxt %u rcvhdrqtailaddr failed\n",
1846 		   rcd->ctxt);
1847 	dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1848 			  rcd->rcvhdrq_dma);
1849 	rcd->rcvhdrq = NULL;
1850 bail:
1851 	return -ENOMEM;
1852 }
1853 
1854 /**
1855  * allocate eager buffers, both kernel and user contexts.
1856  * @rcd: the context we are setting up.
1857  *
1858  * Allocate the eager TID buffers and program them into hip.
1859  * They are no longer completely contiguous, we do multiple allocation
1860  * calls.  Otherwise we get the OOM code involved, by asking for too
1861  * much per call, with disastrous results on some kernels.
1862  */
1863 int hfi1_setup_eagerbufs(struct hfi1_ctxtdata *rcd)
1864 {
1865 	struct hfi1_devdata *dd = rcd->dd;
1866 	u32 max_entries, egrtop, alloced_bytes = 0;
1867 	gfp_t gfp_flags;
1868 	u16 order, idx = 0;
1869 	int ret = 0;
1870 	u16 round_mtu = roundup_pow_of_two(hfi1_max_mtu);
1871 
1872 	/*
1873 	 * GFP_USER, but without GFP_FS, so buffer cache can be
1874 	 * coalesced (we hope); otherwise, even at order 4,
1875 	 * heavy filesystem activity makes these fail, and we can
1876 	 * use compound pages.
1877 	 */
1878 	gfp_flags = __GFP_RECLAIM | __GFP_IO | __GFP_COMP;
1879 
1880 	/*
1881 	 * The minimum size of the eager buffers is a groups of MTU-sized
1882 	 * buffers.
1883 	 * The global eager_buffer_size parameter is checked against the
1884 	 * theoretical lower limit of the value. Here, we check against the
1885 	 * MTU.
1886 	 */
1887 	if (rcd->egrbufs.size < (round_mtu * dd->rcv_entries.group_size))
1888 		rcd->egrbufs.size = round_mtu * dd->rcv_entries.group_size;
1889 	/*
1890 	 * If using one-pkt-per-egr-buffer, lower the eager buffer
1891 	 * size to the max MTU (page-aligned).
1892 	 */
1893 	if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
1894 		rcd->egrbufs.rcvtid_size = round_mtu;
1895 
1896 	/*
1897 	 * Eager buffers sizes of 1MB or less require smaller TID sizes
1898 	 * to satisfy the "multiple of 8 RcvArray entries" requirement.
1899 	 */
1900 	if (rcd->egrbufs.size <= (1 << 20))
1901 		rcd->egrbufs.rcvtid_size = max((unsigned long)round_mtu,
1902 			rounddown_pow_of_two(rcd->egrbufs.size / 8));
1903 
1904 	while (alloced_bytes < rcd->egrbufs.size &&
1905 	       rcd->egrbufs.alloced < rcd->egrbufs.count) {
1906 		rcd->egrbufs.buffers[idx].addr =
1907 			dma_alloc_coherent(&dd->pcidev->dev,
1908 					   rcd->egrbufs.rcvtid_size,
1909 					   &rcd->egrbufs.buffers[idx].dma,
1910 					   gfp_flags);
1911 		if (rcd->egrbufs.buffers[idx].addr) {
1912 			rcd->egrbufs.buffers[idx].len =
1913 				rcd->egrbufs.rcvtid_size;
1914 			rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].addr =
1915 				rcd->egrbufs.buffers[idx].addr;
1916 			rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].dma =
1917 				rcd->egrbufs.buffers[idx].dma;
1918 			rcd->egrbufs.alloced++;
1919 			alloced_bytes += rcd->egrbufs.rcvtid_size;
1920 			idx++;
1921 		} else {
1922 			u32 new_size, i, j;
1923 			u64 offset = 0;
1924 
1925 			/*
1926 			 * Fail the eager buffer allocation if:
1927 			 *   - we are already using the lowest acceptable size
1928 			 *   - we are using one-pkt-per-egr-buffer (this implies
1929 			 *     that we are accepting only one size)
1930 			 */
1931 			if (rcd->egrbufs.rcvtid_size == round_mtu ||
1932 			    !HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR)) {
1933 				dd_dev_err(dd, "ctxt%u: Failed to allocate eager buffers\n",
1934 					   rcd->ctxt);
1935 				ret = -ENOMEM;
1936 				goto bail_rcvegrbuf_phys;
1937 			}
1938 
1939 			new_size = rcd->egrbufs.rcvtid_size / 2;
1940 
1941 			/*
1942 			 * If the first attempt to allocate memory failed, don't
1943 			 * fail everything but continue with the next lower
1944 			 * size.
1945 			 */
1946 			if (idx == 0) {
1947 				rcd->egrbufs.rcvtid_size = new_size;
1948 				continue;
1949 			}
1950 
1951 			/*
1952 			 * Re-partition already allocated buffers to a smaller
1953 			 * size.
1954 			 */
1955 			rcd->egrbufs.alloced = 0;
1956 			for (i = 0, j = 0, offset = 0; j < idx; i++) {
1957 				if (i >= rcd->egrbufs.count)
1958 					break;
1959 				rcd->egrbufs.rcvtids[i].dma =
1960 					rcd->egrbufs.buffers[j].dma + offset;
1961 				rcd->egrbufs.rcvtids[i].addr =
1962 					rcd->egrbufs.buffers[j].addr + offset;
1963 				rcd->egrbufs.alloced++;
1964 				if ((rcd->egrbufs.buffers[j].dma + offset +
1965 				     new_size) ==
1966 				    (rcd->egrbufs.buffers[j].dma +
1967 				     rcd->egrbufs.buffers[j].len)) {
1968 					j++;
1969 					offset = 0;
1970 				} else {
1971 					offset += new_size;
1972 				}
1973 			}
1974 			rcd->egrbufs.rcvtid_size = new_size;
1975 		}
1976 	}
1977 	rcd->egrbufs.numbufs = idx;
1978 	rcd->egrbufs.size = alloced_bytes;
1979 
1980 	hfi1_cdbg(PROC,
1981 		  "ctxt%u: Alloced %u rcv tid entries @ %uKB, total %uKB\n",
1982 		  rcd->ctxt, rcd->egrbufs.alloced,
1983 		  rcd->egrbufs.rcvtid_size / 1024, rcd->egrbufs.size / 1024);
1984 
1985 	/*
1986 	 * Set the contexts rcv array head update threshold to the closest
1987 	 * power of 2 (so we can use a mask instead of modulo) below half
1988 	 * the allocated entries.
1989 	 */
1990 	rcd->egrbufs.threshold =
1991 		rounddown_pow_of_two(rcd->egrbufs.alloced / 2);
1992 	/*
1993 	 * Compute the expected RcvArray entry base. This is done after
1994 	 * allocating the eager buffers in order to maximize the
1995 	 * expected RcvArray entries for the context.
1996 	 */
1997 	max_entries = rcd->rcv_array_groups * dd->rcv_entries.group_size;
1998 	egrtop = roundup(rcd->egrbufs.alloced, dd->rcv_entries.group_size);
1999 	rcd->expected_count = max_entries - egrtop;
2000 	if (rcd->expected_count > MAX_TID_PAIR_ENTRIES * 2)
2001 		rcd->expected_count = MAX_TID_PAIR_ENTRIES * 2;
2002 
2003 	rcd->expected_base = rcd->eager_base + egrtop;
2004 	hfi1_cdbg(PROC, "ctxt%u: eager:%u, exp:%u, egrbase:%u, expbase:%u\n",
2005 		  rcd->ctxt, rcd->egrbufs.alloced, rcd->expected_count,
2006 		  rcd->eager_base, rcd->expected_base);
2007 
2008 	if (!hfi1_rcvbuf_validate(rcd->egrbufs.rcvtid_size, PT_EAGER, &order)) {
2009 		hfi1_cdbg(PROC,
2010 			  "ctxt%u: current Eager buffer size is invalid %u\n",
2011 			  rcd->ctxt, rcd->egrbufs.rcvtid_size);
2012 		ret = -EINVAL;
2013 		goto bail_rcvegrbuf_phys;
2014 	}
2015 
2016 	for (idx = 0; idx < rcd->egrbufs.alloced; idx++) {
2017 		hfi1_put_tid(dd, rcd->eager_base + idx, PT_EAGER,
2018 			     rcd->egrbufs.rcvtids[idx].dma, order);
2019 		cond_resched();
2020 	}
2021 
2022 	return 0;
2023 
2024 bail_rcvegrbuf_phys:
2025 	for (idx = 0; idx < rcd->egrbufs.alloced &&
2026 	     rcd->egrbufs.buffers[idx].addr;
2027 	     idx++) {
2028 		dma_free_coherent(&dd->pcidev->dev,
2029 				  rcd->egrbufs.buffers[idx].len,
2030 				  rcd->egrbufs.buffers[idx].addr,
2031 				  rcd->egrbufs.buffers[idx].dma);
2032 		rcd->egrbufs.buffers[idx].addr = NULL;
2033 		rcd->egrbufs.buffers[idx].dma = 0;
2034 		rcd->egrbufs.buffers[idx].len = 0;
2035 	}
2036 
2037 	return ret;
2038 }
2039