• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2020 Intel Corporation.
4  */
5 
6 #include <linux/pci.h>
7 #include <linux/netdevice.h>
8 #include <linux/vmalloc.h>
9 #include <linux/delay.h>
10 #include <linux/xarray.h>
11 #include <linux/module.h>
12 #include <linux/printk.h>
13 #include <linux/hrtimer.h>
14 #include <linux/bitmap.h>
15 #include <linux/numa.h>
16 #include <rdma/rdma_vt.h>
17 
18 #include "hfi.h"
19 #include "device.h"
20 #include "common.h"
21 #include "trace.h"
22 #include "mad.h"
23 #include "sdma.h"
24 #include "debugfs.h"
25 #include "verbs.h"
26 #include "aspm.h"
27 #include "affinity.h"
28 #include "vnic.h"
29 #include "exp_rcv.h"
30 #include "netdev.h"
31 
32 #undef pr_fmt
33 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
34 
35 /*
36  * min buffers we want to have per context, after driver
37  */
38 #define HFI1_MIN_USER_CTXT_BUFCNT 7
39 
40 #define HFI1_MIN_EAGER_BUFFER_SIZE (4 * 1024) /* 4KB */
41 #define HFI1_MAX_EAGER_BUFFER_SIZE (256 * 1024) /* 256KB */
42 
43 #define NUM_IB_PORTS 1
44 
45 /*
46  * Number of user receive contexts we are configured to use (to allow for more
47  * pio buffers per ctxt, etc.)  Zero means use one user context per CPU.
48  */
49 int num_user_contexts = -1;
50 module_param_named(num_user_contexts, num_user_contexts, int, 0444);
51 MODULE_PARM_DESC(
52 	num_user_contexts, "Set max number of user contexts to use (default: -1 will use the real (non-HT) CPU count)");
53 
54 uint krcvqs[RXE_NUM_DATA_VL];
55 int krcvqsset;
56 module_param_array(krcvqs, uint, &krcvqsset, S_IRUGO);
57 MODULE_PARM_DESC(krcvqs, "Array of the number of non-control kernel receive queues by VL");
58 
59 /* computed based on above array */
60 unsigned long n_krcvqs;
61 
62 static unsigned hfi1_rcvarr_split = 25;
63 module_param_named(rcvarr_split, hfi1_rcvarr_split, uint, S_IRUGO);
64 MODULE_PARM_DESC(rcvarr_split, "Percent of context's RcvArray entries used for Eager buffers");
65 
66 static uint eager_buffer_size = (8 << 20); /* 8MB */
67 module_param(eager_buffer_size, uint, S_IRUGO);
68 MODULE_PARM_DESC(eager_buffer_size, "Size of the eager buffers, default: 8MB");
69 
70 static uint rcvhdrcnt = 2048; /* 2x the max eager buffer count */
71 module_param_named(rcvhdrcnt, rcvhdrcnt, uint, S_IRUGO);
72 MODULE_PARM_DESC(rcvhdrcnt, "Receive header queue count (default 2048)");
73 
74 static uint hfi1_hdrq_entsize = 32;
75 module_param_named(hdrq_entsize, hfi1_hdrq_entsize, uint, 0444);
76 MODULE_PARM_DESC(hdrq_entsize, "Size of header queue entries: 2 - 8B, 16 - 64B, 32 - 128B (default)");
77 
78 unsigned int user_credit_return_threshold = 33;	/* default is 33% */
79 module_param(user_credit_return_threshold, uint, S_IRUGO);
80 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)");
81 
82 DEFINE_XARRAY_FLAGS(hfi1_dev_table, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
83 
hfi1_create_kctxt(struct hfi1_devdata * dd,struct hfi1_pportdata * ppd)84 static int hfi1_create_kctxt(struct hfi1_devdata *dd,
85 			     struct hfi1_pportdata *ppd)
86 {
87 	struct hfi1_ctxtdata *rcd;
88 	int ret;
89 
90 	/* Control context has to be always 0 */
91 	BUILD_BUG_ON(HFI1_CTRL_CTXT != 0);
92 
93 	ret = hfi1_create_ctxtdata(ppd, dd->node, &rcd);
94 	if (ret < 0) {
95 		dd_dev_err(dd, "Kernel receive context allocation failed\n");
96 		return ret;
97 	}
98 
99 	/*
100 	 * Set up the kernel context flags here and now because they use
101 	 * default values for all receive side memories.  User contexts will
102 	 * be handled as they are created.
103 	 */
104 	rcd->flags = HFI1_CAP_KGET(MULTI_PKT_EGR) |
105 		HFI1_CAP_KGET(NODROP_RHQ_FULL) |
106 		HFI1_CAP_KGET(NODROP_EGR_FULL) |
107 		HFI1_CAP_KGET(DMA_RTAIL);
108 
109 	/* Control context must use DMA_RTAIL */
110 	if (rcd->ctxt == HFI1_CTRL_CTXT)
111 		rcd->flags |= HFI1_CAP_DMA_RTAIL;
112 	rcd->fast_handler = get_dma_rtail_setting(rcd) ?
113 				handle_receive_interrupt_dma_rtail :
114 				handle_receive_interrupt_nodma_rtail;
115 
116 	hfi1_set_seq_cnt(rcd, 1);
117 
118 	rcd->sc = sc_alloc(dd, SC_ACK, rcd->rcvhdrqentsize, dd->node);
119 	if (!rcd->sc) {
120 		dd_dev_err(dd, "Kernel send context allocation failed\n");
121 		return -ENOMEM;
122 	}
123 	hfi1_init_ctxt(rcd->sc);
124 
125 	return 0;
126 }
127 
128 /*
129  * Create the receive context array and one or more kernel contexts
130  */
hfi1_create_kctxts(struct hfi1_devdata * dd)131 int hfi1_create_kctxts(struct hfi1_devdata *dd)
132 {
133 	u16 i;
134 	int ret;
135 
136 	dd->rcd = kcalloc_node(dd->num_rcv_contexts, sizeof(*dd->rcd),
137 			       GFP_KERNEL, dd->node);
138 	if (!dd->rcd)
139 		return -ENOMEM;
140 
141 	for (i = 0; i < dd->first_dyn_alloc_ctxt; ++i) {
142 		ret = hfi1_create_kctxt(dd, dd->pport);
143 		if (ret)
144 			goto bail;
145 	}
146 
147 	return 0;
148 bail:
149 	for (i = 0; dd->rcd && i < dd->first_dyn_alloc_ctxt; ++i)
150 		hfi1_free_ctxt(dd->rcd[i]);
151 
152 	/* All the contexts should be freed, free the array */
153 	kfree(dd->rcd);
154 	dd->rcd = NULL;
155 	return ret;
156 }
157 
158 /*
159  * Helper routines for the receive context reference count (rcd and uctxt).
160  */
hfi1_rcd_init(struct hfi1_ctxtdata * rcd)161 static void hfi1_rcd_init(struct hfi1_ctxtdata *rcd)
162 {
163 	kref_init(&rcd->kref);
164 }
165 
166 /**
167  * hfi1_rcd_free - When reference is zero clean up.
168  * @kref: pointer to an initialized rcd data structure
169  *
170  */
hfi1_rcd_free(struct kref * kref)171 static void hfi1_rcd_free(struct kref *kref)
172 {
173 	unsigned long flags;
174 	struct hfi1_ctxtdata *rcd =
175 		container_of(kref, struct hfi1_ctxtdata, kref);
176 
177 	spin_lock_irqsave(&rcd->dd->uctxt_lock, flags);
178 	rcd->dd->rcd[rcd->ctxt] = NULL;
179 	spin_unlock_irqrestore(&rcd->dd->uctxt_lock, flags);
180 
181 	hfi1_free_ctxtdata(rcd->dd, rcd);
182 
183 	kfree(rcd);
184 }
185 
186 /**
187  * hfi1_rcd_put - decrement reference for rcd
188  * @rcd: pointer to an initialized rcd data structure
189  *
190  * Use this to put a reference after the init.
191  */
hfi1_rcd_put(struct hfi1_ctxtdata * rcd)192 int hfi1_rcd_put(struct hfi1_ctxtdata *rcd)
193 {
194 	if (rcd)
195 		return kref_put(&rcd->kref, hfi1_rcd_free);
196 
197 	return 0;
198 }
199 
200 /**
201  * hfi1_rcd_get - increment reference for rcd
202  * @rcd: pointer to an initialized rcd data structure
203  *
204  * Use this to get a reference after the init.
205  *
206  * Return : reflect kref_get_unless_zero(), which returns non-zero on
207  * increment, otherwise 0.
208  */
hfi1_rcd_get(struct hfi1_ctxtdata * rcd)209 int hfi1_rcd_get(struct hfi1_ctxtdata *rcd)
210 {
211 	return kref_get_unless_zero(&rcd->kref);
212 }
213 
214 /**
215  * allocate_rcd_index - allocate an rcd index from the rcd array
216  * @dd: pointer to a valid devdata structure
217  * @rcd: rcd data structure to assign
218  * @index: pointer to index that is allocated
219  *
220  * Find an empty index in the rcd array, and assign the given rcd to it.
221  * If the array is full, we are EBUSY.
222  *
223  */
allocate_rcd_index(struct hfi1_devdata * dd,struct hfi1_ctxtdata * rcd,u16 * index)224 static int allocate_rcd_index(struct hfi1_devdata *dd,
225 			      struct hfi1_ctxtdata *rcd, u16 *index)
226 {
227 	unsigned long flags;
228 	u16 ctxt;
229 
230 	spin_lock_irqsave(&dd->uctxt_lock, flags);
231 	for (ctxt = 0; ctxt < dd->num_rcv_contexts; ctxt++)
232 		if (!dd->rcd[ctxt])
233 			break;
234 
235 	if (ctxt < dd->num_rcv_contexts) {
236 		rcd->ctxt = ctxt;
237 		dd->rcd[ctxt] = rcd;
238 		hfi1_rcd_init(rcd);
239 	}
240 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
241 
242 	if (ctxt >= dd->num_rcv_contexts)
243 		return -EBUSY;
244 
245 	*index = ctxt;
246 
247 	return 0;
248 }
249 
250 /**
251  * hfi1_rcd_get_by_index_safe - validate the ctxt index before accessing the
252  * array
253  * @dd: pointer to a valid devdata structure
254  * @ctxt: the index of an possilbe rcd
255  *
256  * This is a wrapper for hfi1_rcd_get_by_index() to validate that the given
257  * ctxt index is valid.
258  *
259  * The caller is responsible for making the _put().
260  *
261  */
hfi1_rcd_get_by_index_safe(struct hfi1_devdata * dd,u16 ctxt)262 struct hfi1_ctxtdata *hfi1_rcd_get_by_index_safe(struct hfi1_devdata *dd,
263 						 u16 ctxt)
264 {
265 	if (ctxt < dd->num_rcv_contexts)
266 		return hfi1_rcd_get_by_index(dd, ctxt);
267 
268 	return NULL;
269 }
270 
271 /**
272  * hfi1_rcd_get_by_index - get by index
273  * @dd: pointer to a valid devdata structure
274  * @ctxt: the index of an possilbe rcd
275  *
276  * We need to protect access to the rcd array.  If access is needed to
277  * one or more index, get the protecting spinlock and then increment the
278  * kref.
279  *
280  * The caller is responsible for making the _put().
281  *
282  */
hfi1_rcd_get_by_index(struct hfi1_devdata * dd,u16 ctxt)283 struct hfi1_ctxtdata *hfi1_rcd_get_by_index(struct hfi1_devdata *dd, u16 ctxt)
284 {
285 	unsigned long flags;
286 	struct hfi1_ctxtdata *rcd = NULL;
287 
288 	spin_lock_irqsave(&dd->uctxt_lock, flags);
289 	if (dd->rcd[ctxt]) {
290 		rcd = dd->rcd[ctxt];
291 		if (!hfi1_rcd_get(rcd))
292 			rcd = NULL;
293 	}
294 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
295 
296 	return rcd;
297 }
298 
299 /*
300  * Common code for user and kernel context create and setup.
301  * NOTE: the initial kref is done here (hf1_rcd_init()).
302  */
hfi1_create_ctxtdata(struct hfi1_pportdata * ppd,int numa,struct hfi1_ctxtdata ** context)303 int hfi1_create_ctxtdata(struct hfi1_pportdata *ppd, int numa,
304 			 struct hfi1_ctxtdata **context)
305 {
306 	struct hfi1_devdata *dd = ppd->dd;
307 	struct hfi1_ctxtdata *rcd;
308 	unsigned kctxt_ngroups = 0;
309 	u32 base;
310 
311 	if (dd->rcv_entries.nctxt_extra >
312 	    dd->num_rcv_contexts - dd->first_dyn_alloc_ctxt)
313 		kctxt_ngroups = (dd->rcv_entries.nctxt_extra -
314 			 (dd->num_rcv_contexts - dd->first_dyn_alloc_ctxt));
315 	rcd = kzalloc_node(sizeof(*rcd), GFP_KERNEL, numa);
316 	if (rcd) {
317 		u32 rcvtids, max_entries;
318 		u16 ctxt;
319 		int ret;
320 
321 		ret = allocate_rcd_index(dd, rcd, &ctxt);
322 		if (ret) {
323 			*context = NULL;
324 			kfree(rcd);
325 			return ret;
326 		}
327 
328 		INIT_LIST_HEAD(&rcd->qp_wait_list);
329 		hfi1_exp_tid_group_init(rcd);
330 		rcd->ppd = ppd;
331 		rcd->dd = dd;
332 		rcd->numa_id = numa;
333 		rcd->rcv_array_groups = dd->rcv_entries.ngroups;
334 		rcd->rhf_rcv_function_map = normal_rhf_rcv_functions;
335 		rcd->slow_handler = handle_receive_interrupt;
336 		rcd->do_interrupt = rcd->slow_handler;
337 		rcd->msix_intr = CCE_NUM_MSIX_VECTORS;
338 
339 		mutex_init(&rcd->exp_mutex);
340 		spin_lock_init(&rcd->exp_lock);
341 		INIT_LIST_HEAD(&rcd->flow_queue.queue_head);
342 		INIT_LIST_HEAD(&rcd->rarr_queue.queue_head);
343 
344 		hfi1_cdbg(PROC, "setting up context %u\n", rcd->ctxt);
345 
346 		/*
347 		 * Calculate the context's RcvArray entry starting point.
348 		 * We do this here because we have to take into account all
349 		 * the RcvArray entries that previous context would have
350 		 * taken and we have to account for any extra groups assigned
351 		 * to the static (kernel) or dynamic (vnic/user) contexts.
352 		 */
353 		if (ctxt < dd->first_dyn_alloc_ctxt) {
354 			if (ctxt < kctxt_ngroups) {
355 				base = ctxt * (dd->rcv_entries.ngroups + 1);
356 				rcd->rcv_array_groups++;
357 			} else {
358 				base = kctxt_ngroups +
359 					(ctxt * dd->rcv_entries.ngroups);
360 			}
361 		} else {
362 			u16 ct = ctxt - dd->first_dyn_alloc_ctxt;
363 
364 			base = ((dd->n_krcv_queues * dd->rcv_entries.ngroups) +
365 				kctxt_ngroups);
366 			if (ct < dd->rcv_entries.nctxt_extra) {
367 				base += ct * (dd->rcv_entries.ngroups + 1);
368 				rcd->rcv_array_groups++;
369 			} else {
370 				base += dd->rcv_entries.nctxt_extra +
371 					(ct * dd->rcv_entries.ngroups);
372 			}
373 		}
374 		rcd->eager_base = base * dd->rcv_entries.group_size;
375 
376 		rcd->rcvhdrq_cnt = rcvhdrcnt;
377 		rcd->rcvhdrqentsize = hfi1_hdrq_entsize;
378 		rcd->rhf_offset =
379 			rcd->rcvhdrqentsize - sizeof(u64) / sizeof(u32);
380 		/*
381 		 * Simple Eager buffer allocation: we have already pre-allocated
382 		 * the number of RcvArray entry groups. Each ctxtdata structure
383 		 * holds the number of groups for that context.
384 		 *
385 		 * To follow CSR requirements and maintain cacheline alignment,
386 		 * make sure all sizes and bases are multiples of group_size.
387 		 *
388 		 * The expected entry count is what is left after assigning
389 		 * eager.
390 		 */
391 		max_entries = rcd->rcv_array_groups *
392 			dd->rcv_entries.group_size;
393 		rcvtids = ((max_entries * hfi1_rcvarr_split) / 100);
394 		rcd->egrbufs.count = round_down(rcvtids,
395 						dd->rcv_entries.group_size);
396 		if (rcd->egrbufs.count > MAX_EAGER_ENTRIES) {
397 			dd_dev_err(dd, "ctxt%u: requested too many RcvArray entries.\n",
398 				   rcd->ctxt);
399 			rcd->egrbufs.count = MAX_EAGER_ENTRIES;
400 		}
401 		hfi1_cdbg(PROC,
402 			  "ctxt%u: max Eager buffer RcvArray entries: %u\n",
403 			  rcd->ctxt, rcd->egrbufs.count);
404 
405 		/*
406 		 * Allocate array that will hold the eager buffer accounting
407 		 * data.
408 		 * This will allocate the maximum possible buffer count based
409 		 * on the value of the RcvArray split parameter.
410 		 * The resulting value will be rounded down to the closest
411 		 * multiple of dd->rcv_entries.group_size.
412 		 */
413 		rcd->egrbufs.buffers =
414 			kcalloc_node(rcd->egrbufs.count,
415 				     sizeof(*rcd->egrbufs.buffers),
416 				     GFP_KERNEL, numa);
417 		if (!rcd->egrbufs.buffers)
418 			goto bail;
419 		rcd->egrbufs.rcvtids =
420 			kcalloc_node(rcd->egrbufs.count,
421 				     sizeof(*rcd->egrbufs.rcvtids),
422 				     GFP_KERNEL, numa);
423 		if (!rcd->egrbufs.rcvtids)
424 			goto bail;
425 		rcd->egrbufs.size = eager_buffer_size;
426 		/*
427 		 * The size of the buffers programmed into the RcvArray
428 		 * entries needs to be big enough to handle the highest
429 		 * MTU supported.
430 		 */
431 		if (rcd->egrbufs.size < hfi1_max_mtu) {
432 			rcd->egrbufs.size = __roundup_pow_of_two(hfi1_max_mtu);
433 			hfi1_cdbg(PROC,
434 				  "ctxt%u: eager bufs size too small. Adjusting to %u\n",
435 				    rcd->ctxt, rcd->egrbufs.size);
436 		}
437 		rcd->egrbufs.rcvtid_size = HFI1_MAX_EAGER_BUFFER_SIZE;
438 
439 		/* Applicable only for statically created kernel contexts */
440 		if (ctxt < dd->first_dyn_alloc_ctxt) {
441 			rcd->opstats = kzalloc_node(sizeof(*rcd->opstats),
442 						    GFP_KERNEL, numa);
443 			if (!rcd->opstats)
444 				goto bail;
445 
446 			/* Initialize TID flow generations for the context */
447 			hfi1_kern_init_ctxt_generations(rcd);
448 		}
449 
450 		*context = rcd;
451 		return 0;
452 	}
453 
454 bail:
455 	*context = NULL;
456 	hfi1_free_ctxt(rcd);
457 	return -ENOMEM;
458 }
459 
460 /**
461  * hfi1_free_ctxt - free context
462  * @rcd: pointer to an initialized rcd data structure
463  *
464  * This wrapper is the free function that matches hfi1_create_ctxtdata().
465  * When a context is done being used (kernel or user), this function is called
466  * for the "final" put to match the kref init from hf1i_create_ctxtdata().
467  * Other users of the context do a get/put sequence to make sure that the
468  * structure isn't removed while in use.
469  */
hfi1_free_ctxt(struct hfi1_ctxtdata * rcd)470 void hfi1_free_ctxt(struct hfi1_ctxtdata *rcd)
471 {
472 	hfi1_rcd_put(rcd);
473 }
474 
475 /*
476  * Select the largest ccti value over all SLs to determine the intra-
477  * packet gap for the link.
478  *
479  * called with cca_timer_lock held (to protect access to cca_timer
480  * array), and rcu_read_lock() (to protect access to cc_state).
481  */
set_link_ipg(struct hfi1_pportdata * ppd)482 void set_link_ipg(struct hfi1_pportdata *ppd)
483 {
484 	struct hfi1_devdata *dd = ppd->dd;
485 	struct cc_state *cc_state;
486 	int i;
487 	u16 cce, ccti_limit, max_ccti = 0;
488 	u16 shift, mult;
489 	u64 src;
490 	u32 current_egress_rate; /* Mbits /sec */
491 	u64 max_pkt_time;
492 	/*
493 	 * max_pkt_time is the maximum packet egress time in units
494 	 * of the fabric clock period 1/(805 MHz).
495 	 */
496 
497 	cc_state = get_cc_state(ppd);
498 
499 	if (!cc_state)
500 		/*
501 		 * This should _never_ happen - rcu_read_lock() is held,
502 		 * and set_link_ipg() should not be called if cc_state
503 		 * is NULL.
504 		 */
505 		return;
506 
507 	for (i = 0; i < OPA_MAX_SLS; i++) {
508 		u16 ccti = ppd->cca_timer[i].ccti;
509 
510 		if (ccti > max_ccti)
511 			max_ccti = ccti;
512 	}
513 
514 	ccti_limit = cc_state->cct.ccti_limit;
515 	if (max_ccti > ccti_limit)
516 		max_ccti = ccti_limit;
517 
518 	cce = cc_state->cct.entries[max_ccti].entry;
519 	shift = (cce & 0xc000) >> 14;
520 	mult = (cce & 0x3fff);
521 
522 	current_egress_rate = active_egress_rate(ppd);
523 
524 	max_pkt_time = egress_cycles(ppd->ibmaxlen, current_egress_rate);
525 
526 	src = (max_pkt_time >> shift) * mult;
527 
528 	src &= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SMASK;
529 	src <<= SEND_STATIC_RATE_CONTROL_CSR_SRC_RELOAD_SHIFT;
530 
531 	write_csr(dd, SEND_STATIC_RATE_CONTROL, src);
532 }
533 
cca_timer_fn(struct hrtimer * t)534 static enum hrtimer_restart cca_timer_fn(struct hrtimer *t)
535 {
536 	struct cca_timer *cca_timer;
537 	struct hfi1_pportdata *ppd;
538 	int sl;
539 	u16 ccti_timer, ccti_min;
540 	struct cc_state *cc_state;
541 	unsigned long flags;
542 	enum hrtimer_restart ret = HRTIMER_NORESTART;
543 
544 	cca_timer = container_of(t, struct cca_timer, hrtimer);
545 	ppd = cca_timer->ppd;
546 	sl = cca_timer->sl;
547 
548 	rcu_read_lock();
549 
550 	cc_state = get_cc_state(ppd);
551 
552 	if (!cc_state) {
553 		rcu_read_unlock();
554 		return HRTIMER_NORESTART;
555 	}
556 
557 	/*
558 	 * 1) decrement ccti for SL
559 	 * 2) calculate IPG for link (set_link_ipg())
560 	 * 3) restart timer, unless ccti is at min value
561 	 */
562 
563 	ccti_min = cc_state->cong_setting.entries[sl].ccti_min;
564 	ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
565 
566 	spin_lock_irqsave(&ppd->cca_timer_lock, flags);
567 
568 	if (cca_timer->ccti > ccti_min) {
569 		cca_timer->ccti--;
570 		set_link_ipg(ppd);
571 	}
572 
573 	if (cca_timer->ccti > ccti_min) {
574 		unsigned long nsec = 1024 * ccti_timer;
575 		/* ccti_timer is in units of 1.024 usec */
576 		hrtimer_forward_now(t, ns_to_ktime(nsec));
577 		ret = HRTIMER_RESTART;
578 	}
579 
580 	spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
581 	rcu_read_unlock();
582 	return ret;
583 }
584 
585 /*
586  * Common code for initializing the physical port structure.
587  */
hfi1_init_pportdata(struct pci_dev * pdev,struct hfi1_pportdata * ppd,struct hfi1_devdata * dd,u8 hw_pidx,u32 port)588 void hfi1_init_pportdata(struct pci_dev *pdev, struct hfi1_pportdata *ppd,
589 			 struct hfi1_devdata *dd, u8 hw_pidx, u32 port)
590 {
591 	int i;
592 	uint default_pkey_idx;
593 	struct cc_state *cc_state;
594 
595 	ppd->dd = dd;
596 	ppd->hw_pidx = hw_pidx;
597 	ppd->port = port; /* IB port number, not index */
598 	ppd->prev_link_width = LINK_WIDTH_DEFAULT;
599 	/*
600 	 * There are C_VL_COUNT number of PortVLXmitWait counters.
601 	 * Adding 1 to C_VL_COUNT to include the PortXmitWait counter.
602 	 */
603 	for (i = 0; i < C_VL_COUNT + 1; i++) {
604 		ppd->port_vl_xmit_wait_last[i] = 0;
605 		ppd->vl_xmit_flit_cnt[i] = 0;
606 	}
607 
608 	default_pkey_idx = 1;
609 
610 	ppd->pkeys[default_pkey_idx] = DEFAULT_P_KEY;
611 	ppd->part_enforce |= HFI1_PART_ENFORCE_IN;
612 	ppd->pkeys[0] = 0x8001;
613 
614 	INIT_WORK(&ppd->link_vc_work, handle_verify_cap);
615 	INIT_WORK(&ppd->link_up_work, handle_link_up);
616 	INIT_WORK(&ppd->link_down_work, handle_link_down);
617 	INIT_WORK(&ppd->freeze_work, handle_freeze);
618 	INIT_WORK(&ppd->link_downgrade_work, handle_link_downgrade);
619 	INIT_WORK(&ppd->sma_message_work, handle_sma_message);
620 	INIT_WORK(&ppd->link_bounce_work, handle_link_bounce);
621 	INIT_DELAYED_WORK(&ppd->start_link_work, handle_start_link);
622 	INIT_WORK(&ppd->linkstate_active_work, receive_interrupt_work);
623 	INIT_WORK(&ppd->qsfp_info.qsfp_work, qsfp_event);
624 
625 	mutex_init(&ppd->hls_lock);
626 	spin_lock_init(&ppd->qsfp_info.qsfp_lock);
627 
628 	ppd->qsfp_info.ppd = ppd;
629 	ppd->sm_trap_qp = 0x0;
630 	ppd->sa_qp = 0x1;
631 
632 	ppd->hfi1_wq = NULL;
633 
634 	spin_lock_init(&ppd->cca_timer_lock);
635 
636 	for (i = 0; i < OPA_MAX_SLS; i++) {
637 		hrtimer_init(&ppd->cca_timer[i].hrtimer, CLOCK_MONOTONIC,
638 			     HRTIMER_MODE_REL);
639 		ppd->cca_timer[i].ppd = ppd;
640 		ppd->cca_timer[i].sl = i;
641 		ppd->cca_timer[i].ccti = 0;
642 		ppd->cca_timer[i].hrtimer.function = cca_timer_fn;
643 	}
644 
645 	ppd->cc_max_table_entries = IB_CC_TABLE_CAP_DEFAULT;
646 
647 	spin_lock_init(&ppd->cc_state_lock);
648 	spin_lock_init(&ppd->cc_log_lock);
649 	cc_state = kzalloc(sizeof(*cc_state), GFP_KERNEL);
650 	RCU_INIT_POINTER(ppd->cc_state, cc_state);
651 	if (!cc_state)
652 		goto bail;
653 	return;
654 
655 bail:
656 	dd_dev_err(dd, "Congestion Control Agent disabled for port %d\n", port);
657 }
658 
659 /*
660  * Do initialization for device that is only needed on
661  * first detect, not on resets.
662  */
loadtime_init(struct hfi1_devdata * dd)663 static int loadtime_init(struct hfi1_devdata *dd)
664 {
665 	return 0;
666 }
667 
668 /**
669  * init_after_reset - re-initialize after a reset
670  * @dd: the hfi1_ib device
671  *
672  * sanity check at least some of the values after reset, and
673  * ensure no receive or transmit (explicitly, in case reset
674  * failed
675  */
init_after_reset(struct hfi1_devdata * dd)676 static int init_after_reset(struct hfi1_devdata *dd)
677 {
678 	int i;
679 	struct hfi1_ctxtdata *rcd;
680 	/*
681 	 * Ensure chip does no sends or receives, tail updates, or
682 	 * pioavail updates while we re-initialize.  This is mostly
683 	 * for the driver data structures, not chip registers.
684 	 */
685 	for (i = 0; i < dd->num_rcv_contexts; i++) {
686 		rcd = hfi1_rcd_get_by_index(dd, i);
687 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
688 			     HFI1_RCVCTRL_INTRAVAIL_DIS |
689 			     HFI1_RCVCTRL_TAILUPD_DIS, rcd);
690 		hfi1_rcd_put(rcd);
691 	}
692 	pio_send_control(dd, PSC_GLOBAL_DISABLE);
693 	for (i = 0; i < dd->num_send_contexts; i++)
694 		sc_disable(dd->send_contexts[i].sc);
695 
696 	return 0;
697 }
698 
enable_chip(struct hfi1_devdata * dd)699 static void enable_chip(struct hfi1_devdata *dd)
700 {
701 	struct hfi1_ctxtdata *rcd;
702 	u32 rcvmask;
703 	u16 i;
704 
705 	/* enable PIO send */
706 	pio_send_control(dd, PSC_GLOBAL_ENABLE);
707 
708 	/*
709 	 * Enable kernel ctxts' receive and receive interrupt.
710 	 * Other ctxts done as user opens and initializes them.
711 	 */
712 	for (i = 0; i < dd->first_dyn_alloc_ctxt; ++i) {
713 		rcd = hfi1_rcd_get_by_index(dd, i);
714 		if (!rcd)
715 			continue;
716 		rcvmask = HFI1_RCVCTRL_CTXT_ENB | HFI1_RCVCTRL_INTRAVAIL_ENB;
717 		rcvmask |= HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ?
718 			HFI1_RCVCTRL_TAILUPD_ENB : HFI1_RCVCTRL_TAILUPD_DIS;
719 		if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
720 			rcvmask |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
721 		if (HFI1_CAP_KGET_MASK(rcd->flags, NODROP_RHQ_FULL))
722 			rcvmask |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
723 		if (HFI1_CAP_KGET_MASK(rcd->flags, NODROP_EGR_FULL))
724 			rcvmask |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
725 		if (HFI1_CAP_IS_KSET(TID_RDMA))
726 			rcvmask |= HFI1_RCVCTRL_TIDFLOW_ENB;
727 		hfi1_rcvctrl(dd, rcvmask, rcd);
728 		sc_enable(rcd->sc);
729 		hfi1_rcd_put(rcd);
730 	}
731 }
732 
733 /**
734  * create_workqueues - create per port workqueues
735  * @dd: the hfi1_ib device
736  */
create_workqueues(struct hfi1_devdata * dd)737 static int create_workqueues(struct hfi1_devdata *dd)
738 {
739 	int pidx;
740 	struct hfi1_pportdata *ppd;
741 
742 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
743 		ppd = dd->pport + pidx;
744 		if (!ppd->hfi1_wq) {
745 			ppd->hfi1_wq =
746 				alloc_workqueue(
747 				    "hfi%d_%d",
748 				    WQ_SYSFS | WQ_HIGHPRI | WQ_CPU_INTENSIVE |
749 				    WQ_MEM_RECLAIM,
750 				    HFI1_MAX_ACTIVE_WORKQUEUE_ENTRIES,
751 				    dd->unit, pidx);
752 			if (!ppd->hfi1_wq)
753 				goto wq_error;
754 		}
755 		if (!ppd->link_wq) {
756 			/*
757 			 * Make the link workqueue single-threaded to enforce
758 			 * serialization.
759 			 */
760 			ppd->link_wq =
761 				alloc_workqueue(
762 				    "hfi_link_%d_%d",
763 				    WQ_SYSFS | WQ_MEM_RECLAIM | WQ_UNBOUND,
764 				    1, /* max_active */
765 				    dd->unit, pidx);
766 			if (!ppd->link_wq)
767 				goto wq_error;
768 		}
769 	}
770 	return 0;
771 wq_error:
772 	pr_err("alloc_workqueue failed for port %d\n", pidx + 1);
773 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
774 		ppd = dd->pport + pidx;
775 		if (ppd->hfi1_wq) {
776 			destroy_workqueue(ppd->hfi1_wq);
777 			ppd->hfi1_wq = NULL;
778 		}
779 		if (ppd->link_wq) {
780 			destroy_workqueue(ppd->link_wq);
781 			ppd->link_wq = NULL;
782 		}
783 	}
784 	return -ENOMEM;
785 }
786 
787 /**
788  * destroy_workqueues - destroy per port workqueues
789  * @dd: the hfi1_ib device
790  */
destroy_workqueues(struct hfi1_devdata * dd)791 static void destroy_workqueues(struct hfi1_devdata *dd)
792 {
793 	int pidx;
794 	struct hfi1_pportdata *ppd;
795 
796 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
797 		ppd = dd->pport + pidx;
798 
799 		if (ppd->hfi1_wq) {
800 			destroy_workqueue(ppd->hfi1_wq);
801 			ppd->hfi1_wq = NULL;
802 		}
803 		if (ppd->link_wq) {
804 			destroy_workqueue(ppd->link_wq);
805 			ppd->link_wq = NULL;
806 		}
807 	}
808 }
809 
810 /**
811  * enable_general_intr() - Enable the IRQs that will be handled by the
812  * general interrupt handler.
813  * @dd: valid devdata
814  *
815  */
enable_general_intr(struct hfi1_devdata * dd)816 static void enable_general_intr(struct hfi1_devdata *dd)
817 {
818 	set_intr_bits(dd, CCE_ERR_INT, MISC_ERR_INT, true);
819 	set_intr_bits(dd, PIO_ERR_INT, TXE_ERR_INT, true);
820 	set_intr_bits(dd, IS_SENDCTXT_ERR_START, IS_SENDCTXT_ERR_END, true);
821 	set_intr_bits(dd, PBC_INT, GPIO_ASSERT_INT, true);
822 	set_intr_bits(dd, TCRIT_INT, TCRIT_INT, true);
823 	set_intr_bits(dd, IS_DC_START, IS_DC_END, true);
824 	set_intr_bits(dd, IS_SENDCREDIT_START, IS_SENDCREDIT_END, true);
825 }
826 
827 /**
828  * hfi1_init - do the actual initialization sequence on the chip
829  * @dd: the hfi1_ib device
830  * @reinit: re-initializing, so don't allocate new memory
831  *
832  * Do the actual initialization sequence on the chip.  This is done
833  * both from the init routine called from the PCI infrastructure, and
834  * when we reset the chip, or detect that it was reset internally,
835  * or it's administratively re-enabled.
836  *
837  * Memory allocation here and in called routines is only done in
838  * the first case (reinit == 0).  We have to be careful, because even
839  * without memory allocation, we need to re-write all the chip registers
840  * TIDs, etc. after the reset or enable has completed.
841  */
hfi1_init(struct hfi1_devdata * dd,int reinit)842 int hfi1_init(struct hfi1_devdata *dd, int reinit)
843 {
844 	int ret = 0, pidx, lastfail = 0;
845 	unsigned long len;
846 	u16 i;
847 	struct hfi1_ctxtdata *rcd;
848 	struct hfi1_pportdata *ppd;
849 
850 	/* Set up send low level handlers */
851 	dd->process_pio_send = hfi1_verbs_send_pio;
852 	dd->process_dma_send = hfi1_verbs_send_dma;
853 	dd->pio_inline_send = pio_copy;
854 	dd->process_vnic_dma_send = hfi1_vnic_send_dma;
855 
856 	if (is_ax(dd)) {
857 		atomic_set(&dd->drop_packet, DROP_PACKET_ON);
858 		dd->do_drop = true;
859 	} else {
860 		atomic_set(&dd->drop_packet, DROP_PACKET_OFF);
861 		dd->do_drop = false;
862 	}
863 
864 	/* make sure the link is not "up" */
865 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
866 		ppd = dd->pport + pidx;
867 		ppd->linkup = 0;
868 	}
869 
870 	if (reinit)
871 		ret = init_after_reset(dd);
872 	else
873 		ret = loadtime_init(dd);
874 	if (ret)
875 		goto done;
876 
877 	/* dd->rcd can be NULL if early initialization failed */
878 	for (i = 0; dd->rcd && i < dd->first_dyn_alloc_ctxt; ++i) {
879 		/*
880 		 * Set up the (kernel) rcvhdr queue and egr TIDs.  If doing
881 		 * re-init, the simplest way to handle this is to free
882 		 * existing, and re-allocate.
883 		 * Need to re-create rest of ctxt 0 ctxtdata as well.
884 		 */
885 		rcd = hfi1_rcd_get_by_index(dd, i);
886 		if (!rcd)
887 			continue;
888 
889 		lastfail = hfi1_create_rcvhdrq(dd, rcd);
890 		if (!lastfail)
891 			lastfail = hfi1_setup_eagerbufs(rcd);
892 		if (!lastfail)
893 			lastfail = hfi1_kern_exp_rcv_init(rcd, reinit);
894 		if (lastfail) {
895 			dd_dev_err(dd,
896 				   "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
897 			ret = lastfail;
898 		}
899 		/* enable IRQ */
900 		hfi1_rcd_put(rcd);
901 	}
902 
903 	/* Allocate enough memory for user event notification. */
904 	len = PAGE_ALIGN(chip_rcv_contexts(dd) * HFI1_MAX_SHARED_CTXTS *
905 			 sizeof(*dd->events));
906 	dd->events = vmalloc_user(len);
907 	if (!dd->events)
908 		dd_dev_err(dd, "Failed to allocate user events page\n");
909 	/*
910 	 * Allocate a page for device and port status.
911 	 * Page will be shared amongst all user processes.
912 	 */
913 	dd->status = vmalloc_user(PAGE_SIZE);
914 	if (!dd->status)
915 		dd_dev_err(dd, "Failed to allocate dev status page\n");
916 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
917 		ppd = dd->pport + pidx;
918 		if (dd->status)
919 			/* Currently, we only have one port */
920 			ppd->statusp = &dd->status->port;
921 
922 		set_mtu(ppd);
923 	}
924 
925 	/* enable chip even if we have an error, so we can debug cause */
926 	enable_chip(dd);
927 
928 done:
929 	/*
930 	 * Set status even if port serdes is not initialized
931 	 * so that diags will work.
932 	 */
933 	if (dd->status)
934 		dd->status->dev |= HFI1_STATUS_CHIP_PRESENT |
935 			HFI1_STATUS_INITTED;
936 	if (!ret) {
937 		/* enable all interrupts from the chip */
938 		enable_general_intr(dd);
939 		init_qsfp_int(dd);
940 
941 		/* chip is OK for user apps; mark it as initialized */
942 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
943 			ppd = dd->pport + pidx;
944 
945 			/*
946 			 * start the serdes - must be after interrupts are
947 			 * enabled so we are notified when the link goes up
948 			 */
949 			lastfail = bringup_serdes(ppd);
950 			if (lastfail)
951 				dd_dev_info(dd,
952 					    "Failed to bring up port %u\n",
953 					    ppd->port);
954 
955 			/*
956 			 * Set status even if port serdes is not initialized
957 			 * so that diags will work.
958 			 */
959 			if (ppd->statusp)
960 				*ppd->statusp |= HFI1_STATUS_CHIP_PRESENT |
961 							HFI1_STATUS_INITTED;
962 			if (!ppd->link_speed_enabled)
963 				continue;
964 		}
965 	}
966 
967 	/* if ret is non-zero, we probably should do some cleanup here... */
968 	return ret;
969 }
970 
hfi1_lookup(int unit)971 struct hfi1_devdata *hfi1_lookup(int unit)
972 {
973 	return xa_load(&hfi1_dev_table, unit);
974 }
975 
976 /*
977  * Stop the timers during unit shutdown, or after an error late
978  * in initialization.
979  */
stop_timers(struct hfi1_devdata * dd)980 static void stop_timers(struct hfi1_devdata *dd)
981 {
982 	struct hfi1_pportdata *ppd;
983 	int pidx;
984 
985 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
986 		ppd = dd->pport + pidx;
987 		if (ppd->led_override_timer.function) {
988 			del_timer_sync(&ppd->led_override_timer);
989 			atomic_set(&ppd->led_override_timer_active, 0);
990 		}
991 	}
992 }
993 
994 /**
995  * shutdown_device - shut down a device
996  * @dd: the hfi1_ib device
997  *
998  * This is called to make the device quiet when we are about to
999  * unload the driver, and also when the device is administratively
1000  * disabled.   It does not free any data structures.
1001  * Everything it does has to be setup again by hfi1_init(dd, 1)
1002  */
shutdown_device(struct hfi1_devdata * dd)1003 static void shutdown_device(struct hfi1_devdata *dd)
1004 {
1005 	struct hfi1_pportdata *ppd;
1006 	struct hfi1_ctxtdata *rcd;
1007 	unsigned pidx;
1008 	int i;
1009 
1010 	if (dd->flags & HFI1_SHUTDOWN)
1011 		return;
1012 	dd->flags |= HFI1_SHUTDOWN;
1013 
1014 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1015 		ppd = dd->pport + pidx;
1016 
1017 		ppd->linkup = 0;
1018 		if (ppd->statusp)
1019 			*ppd->statusp &= ~(HFI1_STATUS_IB_CONF |
1020 					   HFI1_STATUS_IB_READY);
1021 	}
1022 	dd->flags &= ~HFI1_INITTED;
1023 
1024 	/* mask and clean up interrupts */
1025 	set_intr_bits(dd, IS_FIRST_SOURCE, IS_LAST_SOURCE, false);
1026 	msix_clean_up_interrupts(dd);
1027 
1028 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1029 		ppd = dd->pport + pidx;
1030 		for (i = 0; i < dd->num_rcv_contexts; i++) {
1031 			rcd = hfi1_rcd_get_by_index(dd, i);
1032 			hfi1_rcvctrl(dd, HFI1_RCVCTRL_TAILUPD_DIS |
1033 				     HFI1_RCVCTRL_CTXT_DIS |
1034 				     HFI1_RCVCTRL_INTRAVAIL_DIS |
1035 				     HFI1_RCVCTRL_PKEY_DIS |
1036 				     HFI1_RCVCTRL_ONE_PKT_EGR_DIS, rcd);
1037 			hfi1_rcd_put(rcd);
1038 		}
1039 		/*
1040 		 * Gracefully stop all sends allowing any in progress to
1041 		 * trickle out first.
1042 		 */
1043 		for (i = 0; i < dd->num_send_contexts; i++)
1044 			sc_flush(dd->send_contexts[i].sc);
1045 	}
1046 
1047 	/*
1048 	 * Enough for anything that's going to trickle out to have actually
1049 	 * done so.
1050 	 */
1051 	udelay(20);
1052 
1053 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1054 		ppd = dd->pport + pidx;
1055 
1056 		/* disable all contexts */
1057 		for (i = 0; i < dd->num_send_contexts; i++)
1058 			sc_disable(dd->send_contexts[i].sc);
1059 		/* disable the send device */
1060 		pio_send_control(dd, PSC_GLOBAL_DISABLE);
1061 
1062 		shutdown_led_override(ppd);
1063 
1064 		/*
1065 		 * Clear SerdesEnable.
1066 		 * We can't count on interrupts since we are stopping.
1067 		 */
1068 		hfi1_quiet_serdes(ppd);
1069 		if (ppd->hfi1_wq)
1070 			flush_workqueue(ppd->hfi1_wq);
1071 		if (ppd->link_wq)
1072 			flush_workqueue(ppd->link_wq);
1073 	}
1074 	sdma_exit(dd);
1075 }
1076 
1077 /**
1078  * hfi1_free_ctxtdata - free a context's allocated data
1079  * @dd: the hfi1_ib device
1080  * @rcd: the ctxtdata structure
1081  *
1082  * free up any allocated data for a context
1083  * It should never change any chip state, or global driver state.
1084  */
hfi1_free_ctxtdata(struct hfi1_devdata * dd,struct hfi1_ctxtdata * rcd)1085 void hfi1_free_ctxtdata(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
1086 {
1087 	u32 e;
1088 
1089 	if (!rcd)
1090 		return;
1091 
1092 	if (rcd->rcvhdrq) {
1093 		dma_free_coherent(&dd->pcidev->dev, rcvhdrq_size(rcd),
1094 				  rcd->rcvhdrq, rcd->rcvhdrq_dma);
1095 		rcd->rcvhdrq = NULL;
1096 		if (hfi1_rcvhdrtail_kvaddr(rcd)) {
1097 			dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1098 					  (void *)hfi1_rcvhdrtail_kvaddr(rcd),
1099 					  rcd->rcvhdrqtailaddr_dma);
1100 			rcd->rcvhdrtail_kvaddr = NULL;
1101 		}
1102 	}
1103 
1104 	/* all the RcvArray entries should have been cleared by now */
1105 	kfree(rcd->egrbufs.rcvtids);
1106 	rcd->egrbufs.rcvtids = NULL;
1107 
1108 	for (e = 0; e < rcd->egrbufs.alloced; e++) {
1109 		if (rcd->egrbufs.buffers[e].addr)
1110 			dma_free_coherent(&dd->pcidev->dev,
1111 					  rcd->egrbufs.buffers[e].len,
1112 					  rcd->egrbufs.buffers[e].addr,
1113 					  rcd->egrbufs.buffers[e].dma);
1114 	}
1115 	kfree(rcd->egrbufs.buffers);
1116 	rcd->egrbufs.alloced = 0;
1117 	rcd->egrbufs.buffers = NULL;
1118 
1119 	sc_free(rcd->sc);
1120 	rcd->sc = NULL;
1121 
1122 	vfree(rcd->subctxt_uregbase);
1123 	vfree(rcd->subctxt_rcvegrbuf);
1124 	vfree(rcd->subctxt_rcvhdr_base);
1125 	kfree(rcd->opstats);
1126 
1127 	rcd->subctxt_uregbase = NULL;
1128 	rcd->subctxt_rcvegrbuf = NULL;
1129 	rcd->subctxt_rcvhdr_base = NULL;
1130 	rcd->opstats = NULL;
1131 }
1132 
1133 /*
1134  * Release our hold on the shared asic data.  If we are the last one,
1135  * return the structure to be finalized outside the lock.  Must be
1136  * holding hfi1_dev_table lock.
1137  */
release_asic_data(struct hfi1_devdata * dd)1138 static struct hfi1_asic_data *release_asic_data(struct hfi1_devdata *dd)
1139 {
1140 	struct hfi1_asic_data *ad;
1141 	int other;
1142 
1143 	if (!dd->asic_data)
1144 		return NULL;
1145 	dd->asic_data->dds[dd->hfi1_id] = NULL;
1146 	other = dd->hfi1_id ? 0 : 1;
1147 	ad = dd->asic_data;
1148 	dd->asic_data = NULL;
1149 	/* return NULL if the other dd still has a link */
1150 	return ad->dds[other] ? NULL : ad;
1151 }
1152 
finalize_asic_data(struct hfi1_devdata * dd,struct hfi1_asic_data * ad)1153 static void finalize_asic_data(struct hfi1_devdata *dd,
1154 			       struct hfi1_asic_data *ad)
1155 {
1156 	clean_up_i2c(dd, ad);
1157 	kfree(ad);
1158 }
1159 
1160 /**
1161  * hfi1_free_devdata - cleans up and frees per-unit data structure
1162  * @dd: pointer to a valid devdata structure
1163  *
1164  * It cleans up and frees all data structures set up by
1165  * by hfi1_alloc_devdata().
1166  */
hfi1_free_devdata(struct hfi1_devdata * dd)1167 void hfi1_free_devdata(struct hfi1_devdata *dd)
1168 {
1169 	struct hfi1_asic_data *ad;
1170 	unsigned long flags;
1171 
1172 	xa_lock_irqsave(&hfi1_dev_table, flags);
1173 	__xa_erase(&hfi1_dev_table, dd->unit);
1174 	ad = release_asic_data(dd);
1175 	xa_unlock_irqrestore(&hfi1_dev_table, flags);
1176 
1177 	finalize_asic_data(dd, ad);
1178 	free_platform_config(dd);
1179 	rcu_barrier(); /* wait for rcu callbacks to complete */
1180 	free_percpu(dd->int_counter);
1181 	free_percpu(dd->rcv_limit);
1182 	free_percpu(dd->send_schedule);
1183 	free_percpu(dd->tx_opstats);
1184 	dd->int_counter   = NULL;
1185 	dd->rcv_limit     = NULL;
1186 	dd->send_schedule = NULL;
1187 	dd->tx_opstats    = NULL;
1188 	kfree(dd->comp_vect);
1189 	dd->comp_vect = NULL;
1190 	if (dd->rcvhdrtail_dummy_kvaddr)
1191 		dma_free_coherent(&dd->pcidev->dev, sizeof(u64),
1192 				  (void *)dd->rcvhdrtail_dummy_kvaddr,
1193 				  dd->rcvhdrtail_dummy_dma);
1194 	dd->rcvhdrtail_dummy_kvaddr = NULL;
1195 	sdma_clean(dd, dd->num_sdma);
1196 	rvt_dealloc_device(&dd->verbs_dev.rdi);
1197 }
1198 
1199 /**
1200  * hfi1_alloc_devdata - Allocate our primary per-unit data structure.
1201  * @pdev: Valid PCI device
1202  * @extra: How many bytes to alloc past the default
1203  *
1204  * Must be done via verbs allocator, because the verbs cleanup process
1205  * both does cleanup and free of the data structure.
1206  * "extra" is for chip-specific data.
1207  */
hfi1_alloc_devdata(struct pci_dev * pdev,size_t extra)1208 static struct hfi1_devdata *hfi1_alloc_devdata(struct pci_dev *pdev,
1209 					       size_t extra)
1210 {
1211 	struct hfi1_devdata *dd;
1212 	int ret, nports;
1213 
1214 	/* extra is * number of ports */
1215 	nports = extra / sizeof(struct hfi1_pportdata);
1216 
1217 	dd = (struct hfi1_devdata *)rvt_alloc_device(sizeof(*dd) + extra,
1218 						     nports);
1219 	if (!dd)
1220 		return ERR_PTR(-ENOMEM);
1221 	dd->num_pports = nports;
1222 	dd->pport = (struct hfi1_pportdata *)(dd + 1);
1223 	dd->pcidev = pdev;
1224 	pci_set_drvdata(pdev, dd);
1225 
1226 	ret = xa_alloc_irq(&hfi1_dev_table, &dd->unit, dd, xa_limit_32b,
1227 			GFP_KERNEL);
1228 	if (ret < 0) {
1229 		dev_err(&pdev->dev,
1230 			"Could not allocate unit ID: error %d\n", -ret);
1231 		goto bail;
1232 	}
1233 	rvt_set_ibdev_name(&dd->verbs_dev.rdi, "%s_%d", class_name(), dd->unit);
1234 	/*
1235 	 * If the BIOS does not have the NUMA node information set, select
1236 	 * NUMA 0 so we get consistent performance.
1237 	 */
1238 	dd->node = pcibus_to_node(pdev->bus);
1239 	if (dd->node == NUMA_NO_NODE) {
1240 		dd_dev_err(dd, "Invalid PCI NUMA node. Performance may be affected\n");
1241 		dd->node = 0;
1242 	}
1243 
1244 	/*
1245 	 * Initialize all locks for the device. This needs to be as early as
1246 	 * possible so locks are usable.
1247 	 */
1248 	spin_lock_init(&dd->sc_lock);
1249 	spin_lock_init(&dd->sendctrl_lock);
1250 	spin_lock_init(&dd->rcvctrl_lock);
1251 	spin_lock_init(&dd->uctxt_lock);
1252 	spin_lock_init(&dd->hfi1_diag_trans_lock);
1253 	spin_lock_init(&dd->sc_init_lock);
1254 	spin_lock_init(&dd->dc8051_memlock);
1255 	seqlock_init(&dd->sc2vl_lock);
1256 	spin_lock_init(&dd->sde_map_lock);
1257 	spin_lock_init(&dd->pio_map_lock);
1258 	mutex_init(&dd->dc8051_lock);
1259 	init_waitqueue_head(&dd->event_queue);
1260 	spin_lock_init(&dd->irq_src_lock);
1261 
1262 	dd->int_counter = alloc_percpu(u64);
1263 	if (!dd->int_counter) {
1264 		ret = -ENOMEM;
1265 		goto bail;
1266 	}
1267 
1268 	dd->rcv_limit = alloc_percpu(u64);
1269 	if (!dd->rcv_limit) {
1270 		ret = -ENOMEM;
1271 		goto bail;
1272 	}
1273 
1274 	dd->send_schedule = alloc_percpu(u64);
1275 	if (!dd->send_schedule) {
1276 		ret = -ENOMEM;
1277 		goto bail;
1278 	}
1279 
1280 	dd->tx_opstats = alloc_percpu(struct hfi1_opcode_stats_perctx);
1281 	if (!dd->tx_opstats) {
1282 		ret = -ENOMEM;
1283 		goto bail;
1284 	}
1285 
1286 	dd->comp_vect = kzalloc(sizeof(*dd->comp_vect), GFP_KERNEL);
1287 	if (!dd->comp_vect) {
1288 		ret = -ENOMEM;
1289 		goto bail;
1290 	}
1291 
1292 	/* allocate dummy tail memory for all receive contexts */
1293 	dd->rcvhdrtail_dummy_kvaddr =
1294 		dma_alloc_coherent(&dd->pcidev->dev, sizeof(u64),
1295 				   &dd->rcvhdrtail_dummy_dma, GFP_KERNEL);
1296 	if (!dd->rcvhdrtail_dummy_kvaddr) {
1297 		ret = -ENOMEM;
1298 		goto bail;
1299 	}
1300 
1301 	atomic_set(&dd->ipoib_rsm_usr_num, 0);
1302 	return dd;
1303 
1304 bail:
1305 	hfi1_free_devdata(dd);
1306 	return ERR_PTR(ret);
1307 }
1308 
1309 /*
1310  * Called from freeze mode handlers, and from PCI error
1311  * reporting code.  Should be paranoid about state of
1312  * system and data structures.
1313  */
hfi1_disable_after_error(struct hfi1_devdata * dd)1314 void hfi1_disable_after_error(struct hfi1_devdata *dd)
1315 {
1316 	if (dd->flags & HFI1_INITTED) {
1317 		u32 pidx;
1318 
1319 		dd->flags &= ~HFI1_INITTED;
1320 		if (dd->pport)
1321 			for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1322 				struct hfi1_pportdata *ppd;
1323 
1324 				ppd = dd->pport + pidx;
1325 				if (dd->flags & HFI1_PRESENT)
1326 					set_link_state(ppd, HLS_DN_DISABLE);
1327 
1328 				if (ppd->statusp)
1329 					*ppd->statusp &= ~HFI1_STATUS_IB_READY;
1330 			}
1331 	}
1332 
1333 	/*
1334 	 * Mark as having had an error for driver, and also
1335 	 * for /sys and status word mapped to user programs.
1336 	 * This marks unit as not usable, until reset.
1337 	 */
1338 	if (dd->status)
1339 		dd->status->dev |= HFI1_STATUS_HWERROR;
1340 }
1341 
1342 static void remove_one(struct pci_dev *);
1343 static int init_one(struct pci_dev *, const struct pci_device_id *);
1344 static void shutdown_one(struct pci_dev *);
1345 
1346 #define DRIVER_LOAD_MSG "Intel " DRIVER_NAME " loaded: "
1347 #define PFX DRIVER_NAME ": "
1348 
1349 const struct pci_device_id hfi1_pci_tbl[] = {
1350 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL0) },
1351 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL1) },
1352 	{ 0, }
1353 };
1354 
1355 MODULE_DEVICE_TABLE(pci, hfi1_pci_tbl);
1356 
1357 static struct pci_driver hfi1_pci_driver = {
1358 	.name = DRIVER_NAME,
1359 	.probe = init_one,
1360 	.remove = remove_one,
1361 	.shutdown = shutdown_one,
1362 	.id_table = hfi1_pci_tbl,
1363 	.err_handler = &hfi1_pci_err_handler,
1364 };
1365 
compute_krcvqs(void)1366 static void __init compute_krcvqs(void)
1367 {
1368 	int i;
1369 
1370 	for (i = 0; i < krcvqsset; i++)
1371 		n_krcvqs += krcvqs[i];
1372 }
1373 
1374 /*
1375  * Do all the generic driver unit- and chip-independent memory
1376  * allocation and initialization.
1377  */
hfi1_mod_init(void)1378 static int __init hfi1_mod_init(void)
1379 {
1380 	int ret;
1381 
1382 	ret = dev_init();
1383 	if (ret)
1384 		goto bail;
1385 
1386 	ret = node_affinity_init();
1387 	if (ret)
1388 		goto bail;
1389 
1390 	/* validate max MTU before any devices start */
1391 	if (!valid_opa_max_mtu(hfi1_max_mtu)) {
1392 		pr_err("Invalid max_mtu 0x%x, using 0x%x instead\n",
1393 		       hfi1_max_mtu, HFI1_DEFAULT_MAX_MTU);
1394 		hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
1395 	}
1396 	/* valid CUs run from 1-128 in powers of 2 */
1397 	if (hfi1_cu > 128 || !is_power_of_2(hfi1_cu))
1398 		hfi1_cu = 1;
1399 	/* valid credit return threshold is 0-100, variable is unsigned */
1400 	if (user_credit_return_threshold > 100)
1401 		user_credit_return_threshold = 100;
1402 
1403 	compute_krcvqs();
1404 	/*
1405 	 * sanitize receive interrupt count, time must wait until after
1406 	 * the hardware type is known
1407 	 */
1408 	if (rcv_intr_count > RCV_HDR_HEAD_COUNTER_MASK)
1409 		rcv_intr_count = RCV_HDR_HEAD_COUNTER_MASK;
1410 	/* reject invalid combinations */
1411 	if (rcv_intr_count == 0 && rcv_intr_timeout == 0) {
1412 		pr_err("Invalid mode: both receive interrupt count and available timeout are zero - setting interrupt count to 1\n");
1413 		rcv_intr_count = 1;
1414 	}
1415 	if (rcv_intr_count > 1 && rcv_intr_timeout == 0) {
1416 		/*
1417 		 * Avoid indefinite packet delivery by requiring a timeout
1418 		 * if count is > 1.
1419 		 */
1420 		pr_err("Invalid mode: receive interrupt count greater than 1 and available timeout is zero - setting available timeout to 1\n");
1421 		rcv_intr_timeout = 1;
1422 	}
1423 	if (rcv_intr_dynamic && !(rcv_intr_count > 1 && rcv_intr_timeout > 0)) {
1424 		/*
1425 		 * The dynamic algorithm expects a non-zero timeout
1426 		 * and a count > 1.
1427 		 */
1428 		pr_err("Invalid mode: dynamic receive interrupt mitigation with invalid count and timeout - turning dynamic off\n");
1429 		rcv_intr_dynamic = 0;
1430 	}
1431 
1432 	/* sanitize link CRC options */
1433 	link_crc_mask &= SUPPORTED_CRCS;
1434 
1435 	ret = opfn_init();
1436 	if (ret < 0) {
1437 		pr_err("Failed to allocate opfn_wq");
1438 		goto bail_dev;
1439 	}
1440 
1441 	/*
1442 	 * These must be called before the driver is registered with
1443 	 * the PCI subsystem.
1444 	 */
1445 	hfi1_dbg_init();
1446 	ret = pci_register_driver(&hfi1_pci_driver);
1447 	if (ret < 0) {
1448 		pr_err("Unable to register driver: error %d\n", -ret);
1449 		goto bail_dev;
1450 	}
1451 	goto bail; /* all OK */
1452 
1453 bail_dev:
1454 	hfi1_dbg_exit();
1455 	dev_cleanup();
1456 bail:
1457 	return ret;
1458 }
1459 
1460 module_init(hfi1_mod_init);
1461 
1462 /*
1463  * Do the non-unit driver cleanup, memory free, etc. at unload.
1464  */
hfi1_mod_cleanup(void)1465 static void __exit hfi1_mod_cleanup(void)
1466 {
1467 	pci_unregister_driver(&hfi1_pci_driver);
1468 	opfn_exit();
1469 	node_affinity_destroy_all();
1470 	hfi1_dbg_exit();
1471 
1472 	WARN_ON(!xa_empty(&hfi1_dev_table));
1473 	dispose_firmware();	/* asymmetric with obtain_firmware() */
1474 	dev_cleanup();
1475 }
1476 
1477 module_exit(hfi1_mod_cleanup);
1478 
1479 /* this can only be called after a successful initialization */
cleanup_device_data(struct hfi1_devdata * dd)1480 static void cleanup_device_data(struct hfi1_devdata *dd)
1481 {
1482 	int ctxt;
1483 	int pidx;
1484 
1485 	/* users can't do anything more with chip */
1486 	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1487 		struct hfi1_pportdata *ppd = &dd->pport[pidx];
1488 		struct cc_state *cc_state;
1489 		int i;
1490 
1491 		if (ppd->statusp)
1492 			*ppd->statusp &= ~HFI1_STATUS_CHIP_PRESENT;
1493 
1494 		for (i = 0; i < OPA_MAX_SLS; i++)
1495 			hrtimer_cancel(&ppd->cca_timer[i].hrtimer);
1496 
1497 		spin_lock(&ppd->cc_state_lock);
1498 		cc_state = get_cc_state_protected(ppd);
1499 		RCU_INIT_POINTER(ppd->cc_state, NULL);
1500 		spin_unlock(&ppd->cc_state_lock);
1501 
1502 		if (cc_state)
1503 			kfree_rcu(cc_state, rcu);
1504 	}
1505 
1506 	free_credit_return(dd);
1507 
1508 	/*
1509 	 * Free any resources still in use (usually just kernel contexts)
1510 	 * at unload; we do for ctxtcnt, because that's what we allocate.
1511 	 */
1512 	for (ctxt = 0; dd->rcd && ctxt < dd->num_rcv_contexts; ctxt++) {
1513 		struct hfi1_ctxtdata *rcd = dd->rcd[ctxt];
1514 
1515 		if (rcd) {
1516 			hfi1_free_ctxt_rcv_groups(rcd);
1517 			hfi1_free_ctxt(rcd);
1518 		}
1519 	}
1520 
1521 	kfree(dd->rcd);
1522 	dd->rcd = NULL;
1523 
1524 	free_pio_map(dd);
1525 	/* must follow rcv context free - need to remove rcv's hooks */
1526 	for (ctxt = 0; ctxt < dd->num_send_contexts; ctxt++)
1527 		sc_free(dd->send_contexts[ctxt].sc);
1528 	dd->num_send_contexts = 0;
1529 	kfree(dd->send_contexts);
1530 	dd->send_contexts = NULL;
1531 	kfree(dd->hw_to_sw);
1532 	dd->hw_to_sw = NULL;
1533 	kfree(dd->boardname);
1534 	vfree(dd->events);
1535 	vfree(dd->status);
1536 }
1537 
1538 /*
1539  * Clean up on unit shutdown, or error during unit load after
1540  * successful initialization.
1541  */
postinit_cleanup(struct hfi1_devdata * dd)1542 static void postinit_cleanup(struct hfi1_devdata *dd)
1543 {
1544 	hfi1_start_cleanup(dd);
1545 	hfi1_comp_vectors_clean_up(dd);
1546 	hfi1_dev_affinity_clean_up(dd);
1547 
1548 	hfi1_pcie_ddcleanup(dd);
1549 	hfi1_pcie_cleanup(dd->pcidev);
1550 
1551 	cleanup_device_data(dd);
1552 
1553 	hfi1_free_devdata(dd);
1554 }
1555 
init_one(struct pci_dev * pdev,const struct pci_device_id * ent)1556 static int init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
1557 {
1558 	int ret = 0, j, pidx, initfail;
1559 	struct hfi1_devdata *dd;
1560 	struct hfi1_pportdata *ppd;
1561 
1562 	/* First, lock the non-writable module parameters */
1563 	HFI1_CAP_LOCK();
1564 
1565 	/* Validate dev ids */
1566 	if (!(ent->device == PCI_DEVICE_ID_INTEL0 ||
1567 	      ent->device == PCI_DEVICE_ID_INTEL1)) {
1568 		dev_err(&pdev->dev, "Failing on unknown Intel deviceid 0x%x\n",
1569 			ent->device);
1570 		ret = -ENODEV;
1571 		goto bail;
1572 	}
1573 
1574 	/* Allocate the dd so we can get to work */
1575 	dd = hfi1_alloc_devdata(pdev, NUM_IB_PORTS *
1576 				sizeof(struct hfi1_pportdata));
1577 	if (IS_ERR(dd)) {
1578 		ret = PTR_ERR(dd);
1579 		goto bail;
1580 	}
1581 
1582 	/* Validate some global module parameters */
1583 	ret = hfi1_validate_rcvhdrcnt(dd, rcvhdrcnt);
1584 	if (ret)
1585 		goto bail;
1586 
1587 	/* use the encoding function as a sanitization check */
1588 	if (!encode_rcv_header_entry_size(hfi1_hdrq_entsize)) {
1589 		dd_dev_err(dd, "Invalid HdrQ Entry size %u\n",
1590 			   hfi1_hdrq_entsize);
1591 		ret = -EINVAL;
1592 		goto bail;
1593 	}
1594 
1595 	/* The receive eager buffer size must be set before the receive
1596 	 * contexts are created.
1597 	 *
1598 	 * Set the eager buffer size.  Validate that it falls in a range
1599 	 * allowed by the hardware - all powers of 2 between the min and
1600 	 * max.  The maximum valid MTU is within the eager buffer range
1601 	 * so we do not need to cap the max_mtu by an eager buffer size
1602 	 * setting.
1603 	 */
1604 	if (eager_buffer_size) {
1605 		if (!is_power_of_2(eager_buffer_size))
1606 			eager_buffer_size =
1607 				roundup_pow_of_two(eager_buffer_size);
1608 		eager_buffer_size =
1609 			clamp_val(eager_buffer_size,
1610 				  MIN_EAGER_BUFFER * 8,
1611 				  MAX_EAGER_BUFFER_TOTAL);
1612 		dd_dev_info(dd, "Eager buffer size %u\n",
1613 			    eager_buffer_size);
1614 	} else {
1615 		dd_dev_err(dd, "Invalid Eager buffer size of 0\n");
1616 		ret = -EINVAL;
1617 		goto bail;
1618 	}
1619 
1620 	/* restrict value of hfi1_rcvarr_split */
1621 	hfi1_rcvarr_split = clamp_val(hfi1_rcvarr_split, 0, 100);
1622 
1623 	ret = hfi1_pcie_init(dd);
1624 	if (ret)
1625 		goto bail;
1626 
1627 	/*
1628 	 * Do device-specific initialization, function table setup, dd
1629 	 * allocation, etc.
1630 	 */
1631 	ret = hfi1_init_dd(dd);
1632 	if (ret)
1633 		goto clean_bail; /* error already printed */
1634 
1635 	ret = create_workqueues(dd);
1636 	if (ret)
1637 		goto clean_bail;
1638 
1639 	/* do the generic initialization */
1640 	initfail = hfi1_init(dd, 0);
1641 
1642 	ret = hfi1_register_ib_device(dd);
1643 
1644 	/*
1645 	 * Now ready for use.  this should be cleared whenever we
1646 	 * detect a reset, or initiate one.  If earlier failure,
1647 	 * we still create devices, so diags, etc. can be used
1648 	 * to determine cause of problem.
1649 	 */
1650 	if (!initfail && !ret) {
1651 		dd->flags |= HFI1_INITTED;
1652 		/* create debufs files after init and ib register */
1653 		hfi1_dbg_ibdev_init(&dd->verbs_dev);
1654 	}
1655 
1656 	j = hfi1_device_create(dd);
1657 	if (j)
1658 		dd_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1659 
1660 	if (initfail || ret) {
1661 		msix_clean_up_interrupts(dd);
1662 		stop_timers(dd);
1663 		flush_workqueue(ib_wq);
1664 		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1665 			hfi1_quiet_serdes(dd->pport + pidx);
1666 			ppd = dd->pport + pidx;
1667 			if (ppd->hfi1_wq) {
1668 				destroy_workqueue(ppd->hfi1_wq);
1669 				ppd->hfi1_wq = NULL;
1670 			}
1671 			if (ppd->link_wq) {
1672 				destroy_workqueue(ppd->link_wq);
1673 				ppd->link_wq = NULL;
1674 			}
1675 		}
1676 		if (!j)
1677 			hfi1_device_remove(dd);
1678 		if (!ret)
1679 			hfi1_unregister_ib_device(dd);
1680 		postinit_cleanup(dd);
1681 		if (initfail)
1682 			ret = initfail;
1683 		goto bail;	/* everything already cleaned */
1684 	}
1685 
1686 	sdma_start(dd);
1687 
1688 	return 0;
1689 
1690 clean_bail:
1691 	hfi1_pcie_cleanup(pdev);
1692 bail:
1693 	return ret;
1694 }
1695 
wait_for_clients(struct hfi1_devdata * dd)1696 static void wait_for_clients(struct hfi1_devdata *dd)
1697 {
1698 	/*
1699 	 * Remove the device init value and complete the device if there is
1700 	 * no clients or wait for active clients to finish.
1701 	 */
1702 	if (refcount_dec_and_test(&dd->user_refcount))
1703 		complete(&dd->user_comp);
1704 
1705 	wait_for_completion(&dd->user_comp);
1706 }
1707 
remove_one(struct pci_dev * pdev)1708 static void remove_one(struct pci_dev *pdev)
1709 {
1710 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
1711 
1712 	/* close debugfs files before ib unregister */
1713 	hfi1_dbg_ibdev_exit(&dd->verbs_dev);
1714 
1715 	/* remove the /dev hfi1 interface */
1716 	hfi1_device_remove(dd);
1717 
1718 	/* wait for existing user space clients to finish */
1719 	wait_for_clients(dd);
1720 
1721 	/* unregister from IB core */
1722 	hfi1_unregister_ib_device(dd);
1723 
1724 	/* free netdev data */
1725 	hfi1_free_rx(dd);
1726 
1727 	/*
1728 	 * Disable the IB link, disable interrupts on the device,
1729 	 * clear dma engines, etc.
1730 	 */
1731 	shutdown_device(dd);
1732 	destroy_workqueues(dd);
1733 
1734 	stop_timers(dd);
1735 
1736 	/* wait until all of our (qsfp) queue_work() calls complete */
1737 	flush_workqueue(ib_wq);
1738 
1739 	postinit_cleanup(dd);
1740 }
1741 
shutdown_one(struct pci_dev * pdev)1742 static void shutdown_one(struct pci_dev *pdev)
1743 {
1744 	struct hfi1_devdata *dd = pci_get_drvdata(pdev);
1745 
1746 	shutdown_device(dd);
1747 }
1748 
1749 /**
1750  * hfi1_create_rcvhdrq - create a receive header queue
1751  * @dd: the hfi1_ib device
1752  * @rcd: the context data
1753  *
1754  * This must be contiguous memory (from an i/o perspective), and must be
1755  * DMA'able (which means for some systems, it will go through an IOMMU,
1756  * or be forced into a low address range).
1757  */
hfi1_create_rcvhdrq(struct hfi1_devdata * dd,struct hfi1_ctxtdata * rcd)1758 int hfi1_create_rcvhdrq(struct hfi1_devdata *dd, struct hfi1_ctxtdata *rcd)
1759 {
1760 	unsigned amt;
1761 
1762 	if (!rcd->rcvhdrq) {
1763 		gfp_t gfp_flags;
1764 
1765 		amt = rcvhdrq_size(rcd);
1766 
1767 		if (rcd->ctxt < dd->first_dyn_alloc_ctxt || rcd->is_vnic)
1768 			gfp_flags = GFP_KERNEL;
1769 		else
1770 			gfp_flags = GFP_USER;
1771 		rcd->rcvhdrq = dma_alloc_coherent(&dd->pcidev->dev, amt,
1772 						  &rcd->rcvhdrq_dma,
1773 						  gfp_flags | __GFP_COMP);
1774 
1775 		if (!rcd->rcvhdrq) {
1776 			dd_dev_err(dd,
1777 				   "attempt to allocate %d bytes for ctxt %u rcvhdrq failed\n",
1778 				   amt, rcd->ctxt);
1779 			goto bail;
1780 		}
1781 
1782 		if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ||
1783 		    HFI1_CAP_UGET_MASK(rcd->flags, DMA_RTAIL)) {
1784 			rcd->rcvhdrtail_kvaddr = dma_alloc_coherent(&dd->pcidev->dev,
1785 								    PAGE_SIZE,
1786 								    &rcd->rcvhdrqtailaddr_dma,
1787 								    gfp_flags);
1788 			if (!rcd->rcvhdrtail_kvaddr)
1789 				goto bail_free;
1790 		}
1791 	}
1792 
1793 	set_hdrq_regs(rcd->dd, rcd->ctxt, rcd->rcvhdrqentsize,
1794 		      rcd->rcvhdrq_cnt);
1795 
1796 	return 0;
1797 
1798 bail_free:
1799 	dd_dev_err(dd,
1800 		   "attempt to allocate 1 page for ctxt %u rcvhdrqtailaddr failed\n",
1801 		   rcd->ctxt);
1802 	dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1803 			  rcd->rcvhdrq_dma);
1804 	rcd->rcvhdrq = NULL;
1805 bail:
1806 	return -ENOMEM;
1807 }
1808 
1809 /**
1810  * hfi1_setup_eagerbufs - llocate eager buffers, both kernel and user
1811  * contexts.
1812  * @rcd: the context we are setting up.
1813  *
1814  * Allocate the eager TID buffers and program them into hip.
1815  * They are no longer completely contiguous, we do multiple allocation
1816  * calls.  Otherwise we get the OOM code involved, by asking for too
1817  * much per call, with disastrous results on some kernels.
1818  */
hfi1_setup_eagerbufs(struct hfi1_ctxtdata * rcd)1819 int hfi1_setup_eagerbufs(struct hfi1_ctxtdata *rcd)
1820 {
1821 	struct hfi1_devdata *dd = rcd->dd;
1822 	u32 max_entries, egrtop, alloced_bytes = 0;
1823 	gfp_t gfp_flags;
1824 	u16 order, idx = 0;
1825 	int ret = 0;
1826 	u16 round_mtu = roundup_pow_of_two(hfi1_max_mtu);
1827 
1828 	/*
1829 	 * GFP_USER, but without GFP_FS, so buffer cache can be
1830 	 * coalesced (we hope); otherwise, even at order 4,
1831 	 * heavy filesystem activity makes these fail, and we can
1832 	 * use compound pages.
1833 	 */
1834 	gfp_flags = __GFP_RECLAIM | __GFP_IO | __GFP_COMP;
1835 
1836 	/*
1837 	 * The minimum size of the eager buffers is a groups of MTU-sized
1838 	 * buffers.
1839 	 * The global eager_buffer_size parameter is checked against the
1840 	 * theoretical lower limit of the value. Here, we check against the
1841 	 * MTU.
1842 	 */
1843 	if (rcd->egrbufs.size < (round_mtu * dd->rcv_entries.group_size))
1844 		rcd->egrbufs.size = round_mtu * dd->rcv_entries.group_size;
1845 	/*
1846 	 * If using one-pkt-per-egr-buffer, lower the eager buffer
1847 	 * size to the max MTU (page-aligned).
1848 	 */
1849 	if (!HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR))
1850 		rcd->egrbufs.rcvtid_size = round_mtu;
1851 
1852 	/*
1853 	 * Eager buffers sizes of 1MB or less require smaller TID sizes
1854 	 * to satisfy the "multiple of 8 RcvArray entries" requirement.
1855 	 */
1856 	if (rcd->egrbufs.size <= (1 << 20))
1857 		rcd->egrbufs.rcvtid_size = max((unsigned long)round_mtu,
1858 			rounddown_pow_of_two(rcd->egrbufs.size / 8));
1859 
1860 	while (alloced_bytes < rcd->egrbufs.size &&
1861 	       rcd->egrbufs.alloced < rcd->egrbufs.count) {
1862 		rcd->egrbufs.buffers[idx].addr =
1863 			dma_alloc_coherent(&dd->pcidev->dev,
1864 					   rcd->egrbufs.rcvtid_size,
1865 					   &rcd->egrbufs.buffers[idx].dma,
1866 					   gfp_flags);
1867 		if (rcd->egrbufs.buffers[idx].addr) {
1868 			rcd->egrbufs.buffers[idx].len =
1869 				rcd->egrbufs.rcvtid_size;
1870 			rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].addr =
1871 				rcd->egrbufs.buffers[idx].addr;
1872 			rcd->egrbufs.rcvtids[rcd->egrbufs.alloced].dma =
1873 				rcd->egrbufs.buffers[idx].dma;
1874 			rcd->egrbufs.alloced++;
1875 			alloced_bytes += rcd->egrbufs.rcvtid_size;
1876 			idx++;
1877 		} else {
1878 			u32 new_size, i, j;
1879 			u64 offset = 0;
1880 
1881 			/*
1882 			 * Fail the eager buffer allocation if:
1883 			 *   - we are already using the lowest acceptable size
1884 			 *   - we are using one-pkt-per-egr-buffer (this implies
1885 			 *     that we are accepting only one size)
1886 			 */
1887 			if (rcd->egrbufs.rcvtid_size == round_mtu ||
1888 			    !HFI1_CAP_KGET_MASK(rcd->flags, MULTI_PKT_EGR)) {
1889 				dd_dev_err(dd, "ctxt%u: Failed to allocate eager buffers\n",
1890 					   rcd->ctxt);
1891 				ret = -ENOMEM;
1892 				goto bail_rcvegrbuf_phys;
1893 			}
1894 
1895 			new_size = rcd->egrbufs.rcvtid_size / 2;
1896 
1897 			/*
1898 			 * If the first attempt to allocate memory failed, don't
1899 			 * fail everything but continue with the next lower
1900 			 * size.
1901 			 */
1902 			if (idx == 0) {
1903 				rcd->egrbufs.rcvtid_size = new_size;
1904 				continue;
1905 			}
1906 
1907 			/*
1908 			 * Re-partition already allocated buffers to a smaller
1909 			 * size.
1910 			 */
1911 			rcd->egrbufs.alloced = 0;
1912 			for (i = 0, j = 0, offset = 0; j < idx; i++) {
1913 				if (i >= rcd->egrbufs.count)
1914 					break;
1915 				rcd->egrbufs.rcvtids[i].dma =
1916 					rcd->egrbufs.buffers[j].dma + offset;
1917 				rcd->egrbufs.rcvtids[i].addr =
1918 					rcd->egrbufs.buffers[j].addr + offset;
1919 				rcd->egrbufs.alloced++;
1920 				if ((rcd->egrbufs.buffers[j].dma + offset +
1921 				     new_size) ==
1922 				    (rcd->egrbufs.buffers[j].dma +
1923 				     rcd->egrbufs.buffers[j].len)) {
1924 					j++;
1925 					offset = 0;
1926 				} else {
1927 					offset += new_size;
1928 				}
1929 			}
1930 			rcd->egrbufs.rcvtid_size = new_size;
1931 		}
1932 	}
1933 	rcd->egrbufs.numbufs = idx;
1934 	rcd->egrbufs.size = alloced_bytes;
1935 
1936 	hfi1_cdbg(PROC,
1937 		  "ctxt%u: Alloced %u rcv tid entries @ %uKB, total %uKB\n",
1938 		  rcd->ctxt, rcd->egrbufs.alloced,
1939 		  rcd->egrbufs.rcvtid_size / 1024, rcd->egrbufs.size / 1024);
1940 
1941 	/*
1942 	 * Set the contexts rcv array head update threshold to the closest
1943 	 * power of 2 (so we can use a mask instead of modulo) below half
1944 	 * the allocated entries.
1945 	 */
1946 	rcd->egrbufs.threshold =
1947 		rounddown_pow_of_two(rcd->egrbufs.alloced / 2);
1948 	/*
1949 	 * Compute the expected RcvArray entry base. This is done after
1950 	 * allocating the eager buffers in order to maximize the
1951 	 * expected RcvArray entries for the context.
1952 	 */
1953 	max_entries = rcd->rcv_array_groups * dd->rcv_entries.group_size;
1954 	egrtop = roundup(rcd->egrbufs.alloced, dd->rcv_entries.group_size);
1955 	rcd->expected_count = max_entries - egrtop;
1956 	if (rcd->expected_count > MAX_TID_PAIR_ENTRIES * 2)
1957 		rcd->expected_count = MAX_TID_PAIR_ENTRIES * 2;
1958 
1959 	rcd->expected_base = rcd->eager_base + egrtop;
1960 	hfi1_cdbg(PROC, "ctxt%u: eager:%u, exp:%u, egrbase:%u, expbase:%u\n",
1961 		  rcd->ctxt, rcd->egrbufs.alloced, rcd->expected_count,
1962 		  rcd->eager_base, rcd->expected_base);
1963 
1964 	if (!hfi1_rcvbuf_validate(rcd->egrbufs.rcvtid_size, PT_EAGER, &order)) {
1965 		hfi1_cdbg(PROC,
1966 			  "ctxt%u: current Eager buffer size is invalid %u\n",
1967 			  rcd->ctxt, rcd->egrbufs.rcvtid_size);
1968 		ret = -EINVAL;
1969 		goto bail_rcvegrbuf_phys;
1970 	}
1971 
1972 	for (idx = 0; idx < rcd->egrbufs.alloced; idx++) {
1973 		hfi1_put_tid(dd, rcd->eager_base + idx, PT_EAGER,
1974 			     rcd->egrbufs.rcvtids[idx].dma, order);
1975 		cond_resched();
1976 	}
1977 
1978 	return 0;
1979 
1980 bail_rcvegrbuf_phys:
1981 	for (idx = 0; idx < rcd->egrbufs.alloced &&
1982 	     rcd->egrbufs.buffers[idx].addr;
1983 	     idx++) {
1984 		dma_free_coherent(&dd->pcidev->dev,
1985 				  rcd->egrbufs.buffers[idx].len,
1986 				  rcd->egrbufs.buffers[idx].addr,
1987 				  rcd->egrbufs.buffers[idx].dma);
1988 		rcd->egrbufs.buffers[idx].addr = NULL;
1989 		rcd->egrbufs.buffers[idx].dma = 0;
1990 		rcd->egrbufs.buffers[idx].len = 0;
1991 	}
1992 
1993 	return ret;
1994 }
1995