xref: /openbmc/linux/drivers/infiniband/hw/hfi1/verbs.c (revision 8730046c)
1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <rdma/ib_mad.h>
49 #include <rdma/ib_user_verbs.h>
50 #include <linux/io.h>
51 #include <linux/module.h>
52 #include <linux/utsname.h>
53 #include <linux/rculist.h>
54 #include <linux/mm.h>
55 #include <linux/vmalloc.h>
56 
57 #include "hfi.h"
58 #include "common.h"
59 #include "device.h"
60 #include "trace.h"
61 #include "qp.h"
62 #include "verbs_txreq.h"
63 
64 static unsigned int hfi1_lkey_table_size = 16;
65 module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
66 		   S_IRUGO);
67 MODULE_PARM_DESC(lkey_table_size,
68 		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
69 
70 static unsigned int hfi1_max_pds = 0xFFFF;
71 module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
72 MODULE_PARM_DESC(max_pds,
73 		 "Maximum number of protection domains to support");
74 
75 static unsigned int hfi1_max_ahs = 0xFFFF;
76 module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
77 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
78 
79 unsigned int hfi1_max_cqes = 0x2FFFFF;
80 module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
81 MODULE_PARM_DESC(max_cqes,
82 		 "Maximum number of completion queue entries to support");
83 
84 unsigned int hfi1_max_cqs = 0x1FFFF;
85 module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
86 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
87 
88 unsigned int hfi1_max_qp_wrs = 0x3FFF;
89 module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
90 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
91 
92 unsigned int hfi1_max_qps = 32768;
93 module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
94 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
95 
96 unsigned int hfi1_max_sges = 0x60;
97 module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
98 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
99 
100 unsigned int hfi1_max_mcast_grps = 16384;
101 module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
102 MODULE_PARM_DESC(max_mcast_grps,
103 		 "Maximum number of multicast groups to support");
104 
105 unsigned int hfi1_max_mcast_qp_attached = 16;
106 module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
107 		   uint, S_IRUGO);
108 MODULE_PARM_DESC(max_mcast_qp_attached,
109 		 "Maximum number of attached QPs to support");
110 
111 unsigned int hfi1_max_srqs = 1024;
112 module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
113 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
114 
115 unsigned int hfi1_max_srq_sges = 128;
116 module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
117 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
118 
119 unsigned int hfi1_max_srq_wrs = 0x1FFFF;
120 module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
121 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
122 
123 unsigned short piothreshold = 256;
124 module_param(piothreshold, ushort, S_IRUGO);
125 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
126 
127 #define COPY_CACHELESS 1
128 #define COPY_ADAPTIVE  2
129 static unsigned int sge_copy_mode;
130 module_param(sge_copy_mode, uint, S_IRUGO);
131 MODULE_PARM_DESC(sge_copy_mode,
132 		 "Verbs copy mode: 0 use memcpy, 1 use cacheless copy, 2 adapt based on WSS");
133 
134 static void verbs_sdma_complete(
135 	struct sdma_txreq *cookie,
136 	int status);
137 
138 static int pio_wait(struct rvt_qp *qp,
139 		    struct send_context *sc,
140 		    struct hfi1_pkt_state *ps,
141 		    u32 flag);
142 
143 /* Length of buffer to create verbs txreq cache name */
144 #define TXREQ_NAME_LEN 24
145 
146 static uint wss_threshold;
147 module_param(wss_threshold, uint, S_IRUGO);
148 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
149 static uint wss_clean_period = 256;
150 module_param(wss_clean_period, uint, S_IRUGO);
151 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
152 
153 /* memory working set size */
154 struct hfi1_wss {
155 	unsigned long *entries;
156 	atomic_t total_count;
157 	atomic_t clean_counter;
158 	atomic_t clean_entry;
159 
160 	int threshold;
161 	int num_entries;
162 	long pages_mask;
163 };
164 
165 static struct hfi1_wss wss;
166 
167 int hfi1_wss_init(void)
168 {
169 	long llc_size;
170 	long llc_bits;
171 	long table_size;
172 	long table_bits;
173 
174 	/* check for a valid percent range - default to 80 if none or invalid */
175 	if (wss_threshold < 1 || wss_threshold > 100)
176 		wss_threshold = 80;
177 	/* reject a wildly large period */
178 	if (wss_clean_period > 1000000)
179 		wss_clean_period = 256;
180 	/* reject a zero period */
181 	if (wss_clean_period == 0)
182 		wss_clean_period = 1;
183 
184 	/*
185 	 * Calculate the table size - the next power of 2 larger than the
186 	 * LLC size.  LLC size is in KiB.
187 	 */
188 	llc_size = wss_llc_size() * 1024;
189 	table_size = roundup_pow_of_two(llc_size);
190 
191 	/* one bit per page in rounded up table */
192 	llc_bits = llc_size / PAGE_SIZE;
193 	table_bits = table_size / PAGE_SIZE;
194 	wss.pages_mask = table_bits - 1;
195 	wss.num_entries = table_bits / BITS_PER_LONG;
196 
197 	wss.threshold = (llc_bits * wss_threshold) / 100;
198 	if (wss.threshold == 0)
199 		wss.threshold = 1;
200 
201 	atomic_set(&wss.clean_counter, wss_clean_period);
202 
203 	wss.entries = kcalloc(wss.num_entries, sizeof(*wss.entries),
204 			      GFP_KERNEL);
205 	if (!wss.entries) {
206 		hfi1_wss_exit();
207 		return -ENOMEM;
208 	}
209 
210 	return 0;
211 }
212 
213 void hfi1_wss_exit(void)
214 {
215 	/* coded to handle partially initialized and repeat callers */
216 	kfree(wss.entries);
217 	wss.entries = NULL;
218 }
219 
220 /*
221  * Advance the clean counter.  When the clean period has expired,
222  * clean an entry.
223  *
224  * This is implemented in atomics to avoid locking.  Because multiple
225  * variables are involved, it can be racy which can lead to slightly
226  * inaccurate information.  Since this is only a heuristic, this is
227  * OK.  Any innaccuracies will clean themselves out as the counter
228  * advances.  That said, it is unlikely the entry clean operation will
229  * race - the next possible racer will not start until the next clean
230  * period.
231  *
232  * The clean counter is implemented as a decrement to zero.  When zero
233  * is reached an entry is cleaned.
234  */
235 static void wss_advance_clean_counter(void)
236 {
237 	int entry;
238 	int weight;
239 	unsigned long bits;
240 
241 	/* become the cleaner if we decrement the counter to zero */
242 	if (atomic_dec_and_test(&wss.clean_counter)) {
243 		/*
244 		 * Set, not add, the clean period.  This avoids an issue
245 		 * where the counter could decrement below the clean period.
246 		 * Doing a set can result in lost decrements, slowing the
247 		 * clean advance.  Since this a heuristic, this possible
248 		 * slowdown is OK.
249 		 *
250 		 * An alternative is to loop, advancing the counter by a
251 		 * clean period until the result is > 0. However, this could
252 		 * lead to several threads keeping another in the clean loop.
253 		 * This could be mitigated by limiting the number of times
254 		 * we stay in the loop.
255 		 */
256 		atomic_set(&wss.clean_counter, wss_clean_period);
257 
258 		/*
259 		 * Uniquely grab the entry to clean and move to next.
260 		 * The current entry is always the lower bits of
261 		 * wss.clean_entry.  The table size, wss.num_entries,
262 		 * is always a power-of-2.
263 		 */
264 		entry = (atomic_inc_return(&wss.clean_entry) - 1)
265 			& (wss.num_entries - 1);
266 
267 		/* clear the entry and count the bits */
268 		bits = xchg(&wss.entries[entry], 0);
269 		weight = hweight64((u64)bits);
270 		/* only adjust the contended total count if needed */
271 		if (weight)
272 			atomic_sub(weight, &wss.total_count);
273 	}
274 }
275 
276 /*
277  * Insert the given address into the working set array.
278  */
279 static void wss_insert(void *address)
280 {
281 	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss.pages_mask;
282 	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
283 	u32 nr = page & (BITS_PER_LONG - 1);
284 
285 	if (!test_and_set_bit(nr, &wss.entries[entry]))
286 		atomic_inc(&wss.total_count);
287 
288 	wss_advance_clean_counter();
289 }
290 
291 /*
292  * Is the working set larger than the threshold?
293  */
294 static inline int wss_exceeds_threshold(void)
295 {
296 	return atomic_read(&wss.total_count) >= wss.threshold;
297 }
298 
299 /*
300  * Length of header by opcode, 0 --> not supported
301  */
302 const u8 hdr_len_by_opcode[256] = {
303 	/* RC */
304 	[IB_OPCODE_RC_SEND_FIRST]                     = 12 + 8,
305 	[IB_OPCODE_RC_SEND_MIDDLE]                    = 12 + 8,
306 	[IB_OPCODE_RC_SEND_LAST]                      = 12 + 8,
307 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
308 	[IB_OPCODE_RC_SEND_ONLY]                      = 12 + 8,
309 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
310 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
311 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = 12 + 8,
312 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = 12 + 8,
313 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
314 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
315 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
316 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = 12 + 8 + 16,
317 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = 12 + 8 + 4,
318 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = 12 + 8,
319 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = 12 + 8 + 4,
320 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = 12 + 8 + 4,
321 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = 12 + 8 + 4,
322 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = 12 + 8 + 4 + 8,
323 	[IB_OPCODE_RC_COMPARE_SWAP]                   = 12 + 8 + 28,
324 	[IB_OPCODE_RC_FETCH_ADD]                      = 12 + 8 + 28,
325 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = 12 + 8 + 4,
326 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = 12 + 8 + 4,
327 	/* UC */
328 	[IB_OPCODE_UC_SEND_FIRST]                     = 12 + 8,
329 	[IB_OPCODE_UC_SEND_MIDDLE]                    = 12 + 8,
330 	[IB_OPCODE_UC_SEND_LAST]                      = 12 + 8,
331 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
332 	[IB_OPCODE_UC_SEND_ONLY]                      = 12 + 8,
333 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
334 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
335 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = 12 + 8,
336 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = 12 + 8,
337 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
338 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
339 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
340 	/* UD */
341 	[IB_OPCODE_UD_SEND_ONLY]                      = 12 + 8 + 8,
342 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 12
343 };
344 
345 static const opcode_handler opcode_handler_tbl[256] = {
346 	/* RC */
347 	[IB_OPCODE_RC_SEND_FIRST]                     = &hfi1_rc_rcv,
348 	[IB_OPCODE_RC_SEND_MIDDLE]                    = &hfi1_rc_rcv,
349 	[IB_OPCODE_RC_SEND_LAST]                      = &hfi1_rc_rcv,
350 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
351 	[IB_OPCODE_RC_SEND_ONLY]                      = &hfi1_rc_rcv,
352 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
353 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = &hfi1_rc_rcv,
354 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = &hfi1_rc_rcv,
355 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = &hfi1_rc_rcv,
356 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
357 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = &hfi1_rc_rcv,
358 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
359 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = &hfi1_rc_rcv,
360 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = &hfi1_rc_rcv,
361 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = &hfi1_rc_rcv,
362 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = &hfi1_rc_rcv,
363 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = &hfi1_rc_rcv,
364 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = &hfi1_rc_rcv,
365 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = &hfi1_rc_rcv,
366 	[IB_OPCODE_RC_COMPARE_SWAP]                   = &hfi1_rc_rcv,
367 	[IB_OPCODE_RC_FETCH_ADD]                      = &hfi1_rc_rcv,
368 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = &hfi1_rc_rcv,
369 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = &hfi1_rc_rcv,
370 	/* UC */
371 	[IB_OPCODE_UC_SEND_FIRST]                     = &hfi1_uc_rcv,
372 	[IB_OPCODE_UC_SEND_MIDDLE]                    = &hfi1_uc_rcv,
373 	[IB_OPCODE_UC_SEND_LAST]                      = &hfi1_uc_rcv,
374 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
375 	[IB_OPCODE_UC_SEND_ONLY]                      = &hfi1_uc_rcv,
376 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
377 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = &hfi1_uc_rcv,
378 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = &hfi1_uc_rcv,
379 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = &hfi1_uc_rcv,
380 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
381 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = &hfi1_uc_rcv,
382 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
383 	/* UD */
384 	[IB_OPCODE_UD_SEND_ONLY]                      = &hfi1_ud_rcv,
385 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_ud_rcv,
386 	/* CNP */
387 	[IB_OPCODE_CNP]				      = &hfi1_cnp_rcv
388 };
389 
390 #define OPMASK 0x1f
391 
392 static const u32 pio_opmask[BIT(3)] = {
393 	/* RC */
394 	[IB_OPCODE_RC >> 5] =
395 		BIT(RC_OP(SEND_ONLY) & OPMASK) |
396 		BIT(RC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
397 		BIT(RC_OP(RDMA_WRITE_ONLY) & OPMASK) |
398 		BIT(RC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK) |
399 		BIT(RC_OP(RDMA_READ_REQUEST) & OPMASK) |
400 		BIT(RC_OP(ACKNOWLEDGE) & OPMASK) |
401 		BIT(RC_OP(ATOMIC_ACKNOWLEDGE) & OPMASK) |
402 		BIT(RC_OP(COMPARE_SWAP) & OPMASK) |
403 		BIT(RC_OP(FETCH_ADD) & OPMASK),
404 	/* UC */
405 	[IB_OPCODE_UC >> 5] =
406 		BIT(UC_OP(SEND_ONLY) & OPMASK) |
407 		BIT(UC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
408 		BIT(UC_OP(RDMA_WRITE_ONLY) & OPMASK) |
409 		BIT(UC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK),
410 };
411 
412 /*
413  * System image GUID.
414  */
415 __be64 ib_hfi1_sys_image_guid;
416 
417 /**
418  * hfi1_copy_sge - copy data to SGE memory
419  * @ss: the SGE state
420  * @data: the data to copy
421  * @length: the length of the data
422  * @copy_last: do a separate copy of the last 8 bytes
423  */
424 void hfi1_copy_sge(
425 	struct rvt_sge_state *ss,
426 	void *data, u32 length,
427 	int release,
428 	int copy_last)
429 {
430 	struct rvt_sge *sge = &ss->sge;
431 	int in_last = 0;
432 	int i;
433 	int cacheless_copy = 0;
434 
435 	if (sge_copy_mode == COPY_CACHELESS) {
436 		cacheless_copy = length >= PAGE_SIZE;
437 	} else if (sge_copy_mode == COPY_ADAPTIVE) {
438 		if (length >= PAGE_SIZE) {
439 			/*
440 			 * NOTE: this *assumes*:
441 			 * o The first vaddr is the dest.
442 			 * o If multiple pages, then vaddr is sequential.
443 			 */
444 			wss_insert(sge->vaddr);
445 			if (length >= (2 * PAGE_SIZE))
446 				wss_insert(sge->vaddr + PAGE_SIZE);
447 
448 			cacheless_copy = wss_exceeds_threshold();
449 		} else {
450 			wss_advance_clean_counter();
451 		}
452 	}
453 	if (copy_last) {
454 		if (length > 8) {
455 			length -= 8;
456 		} else {
457 			copy_last = 0;
458 			in_last = 1;
459 		}
460 	}
461 
462 again:
463 	while (length) {
464 		u32 len = sge->length;
465 
466 		if (len > length)
467 			len = length;
468 		if (len > sge->sge_length)
469 			len = sge->sge_length;
470 		WARN_ON_ONCE(len == 0);
471 		if (unlikely(in_last)) {
472 			/* enforce byte transfer ordering */
473 			for (i = 0; i < len; i++)
474 				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
475 		} else if (cacheless_copy) {
476 			cacheless_memcpy(sge->vaddr, data, len);
477 		} else {
478 			memcpy(sge->vaddr, data, len);
479 		}
480 		sge->vaddr += len;
481 		sge->length -= len;
482 		sge->sge_length -= len;
483 		if (sge->sge_length == 0) {
484 			if (release)
485 				rvt_put_mr(sge->mr);
486 			if (--ss->num_sge)
487 				*sge = *ss->sg_list++;
488 		} else if (sge->length == 0 && sge->mr->lkey) {
489 			if (++sge->n >= RVT_SEGSZ) {
490 				if (++sge->m >= sge->mr->mapsz)
491 					break;
492 				sge->n = 0;
493 			}
494 			sge->vaddr =
495 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
496 			sge->length =
497 				sge->mr->map[sge->m]->segs[sge->n].length;
498 		}
499 		data += len;
500 		length -= len;
501 	}
502 
503 	if (copy_last) {
504 		copy_last = 0;
505 		in_last = 1;
506 		length = 8;
507 		goto again;
508 	}
509 }
510 
511 /**
512  * hfi1_skip_sge - skip over SGE memory
513  * @ss: the SGE state
514  * @length: the number of bytes to skip
515  */
516 void hfi1_skip_sge(struct rvt_sge_state *ss, u32 length, int release)
517 {
518 	struct rvt_sge *sge = &ss->sge;
519 
520 	while (length) {
521 		u32 len = sge->length;
522 
523 		if (len > length)
524 			len = length;
525 		if (len > sge->sge_length)
526 			len = sge->sge_length;
527 		WARN_ON_ONCE(len == 0);
528 		sge->vaddr += len;
529 		sge->length -= len;
530 		sge->sge_length -= len;
531 		if (sge->sge_length == 0) {
532 			if (release)
533 				rvt_put_mr(sge->mr);
534 			if (--ss->num_sge)
535 				*sge = *ss->sg_list++;
536 		} else if (sge->length == 0 && sge->mr->lkey) {
537 			if (++sge->n >= RVT_SEGSZ) {
538 				if (++sge->m >= sge->mr->mapsz)
539 					break;
540 				sge->n = 0;
541 			}
542 			sge->vaddr =
543 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
544 			sge->length =
545 				sge->mr->map[sge->m]->segs[sge->n].length;
546 		}
547 		length -= len;
548 	}
549 }
550 
551 /*
552  * Make sure the QP is ready and able to accept the given opcode.
553  */
554 static inline opcode_handler qp_ok(int opcode, struct hfi1_packet *packet)
555 {
556 	if (!(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
557 		return NULL;
558 	if (((opcode & RVT_OPCODE_QP_MASK) == packet->qp->allowed_ops) ||
559 	    (opcode == IB_OPCODE_CNP))
560 		return opcode_handler_tbl[opcode];
561 
562 	return NULL;
563 }
564 
565 /**
566  * hfi1_ib_rcv - process an incoming packet
567  * @packet: data packet information
568  *
569  * This is called to process an incoming packet at interrupt level.
570  *
571  * Tlen is the length of the header + data + CRC in bytes.
572  */
573 void hfi1_ib_rcv(struct hfi1_packet *packet)
574 {
575 	struct hfi1_ctxtdata *rcd = packet->rcd;
576 	struct ib_header *hdr = packet->hdr;
577 	u32 tlen = packet->tlen;
578 	struct hfi1_pportdata *ppd = rcd->ppd;
579 	struct hfi1_ibport *ibp = &ppd->ibport_data;
580 	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
581 	opcode_handler packet_handler;
582 	unsigned long flags;
583 	u32 qp_num;
584 	int lnh;
585 	u8 opcode;
586 	u16 lid;
587 
588 	/* Check for GRH */
589 	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
590 	if (lnh == HFI1_LRH_BTH) {
591 		packet->ohdr = &hdr->u.oth;
592 	} else if (lnh == HFI1_LRH_GRH) {
593 		u32 vtf;
594 
595 		packet->ohdr = &hdr->u.l.oth;
596 		if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
597 			goto drop;
598 		vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
599 		if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
600 			goto drop;
601 		packet->rcv_flags |= HFI1_HAS_GRH;
602 	} else {
603 		goto drop;
604 	}
605 
606 	trace_input_ibhdr(rcd->dd, hdr);
607 
608 	opcode = (be32_to_cpu(packet->ohdr->bth[0]) >> 24);
609 	inc_opstats(tlen, &rcd->opstats->stats[opcode]);
610 
611 	/* Get the destination QP number. */
612 	qp_num = be32_to_cpu(packet->ohdr->bth[1]) & RVT_QPN_MASK;
613 	lid = be16_to_cpu(hdr->lrh[1]);
614 	if (unlikely((lid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
615 		     (lid != be16_to_cpu(IB_LID_PERMISSIVE)))) {
616 		struct rvt_mcast *mcast;
617 		struct rvt_mcast_qp *p;
618 
619 		if (lnh != HFI1_LRH_GRH)
620 			goto drop;
621 		mcast = rvt_mcast_find(&ibp->rvp, &hdr->u.l.grh.dgid);
622 		if (!mcast)
623 			goto drop;
624 		list_for_each_entry_rcu(p, &mcast->qp_list, list) {
625 			packet->qp = p->qp;
626 			spin_lock_irqsave(&packet->qp->r_lock, flags);
627 			packet_handler = qp_ok(opcode, packet);
628 			if (likely(packet_handler))
629 				packet_handler(packet);
630 			else
631 				ibp->rvp.n_pkt_drops++;
632 			spin_unlock_irqrestore(&packet->qp->r_lock, flags);
633 		}
634 		/*
635 		 * Notify rvt_multicast_detach() if it is waiting for us
636 		 * to finish.
637 		 */
638 		if (atomic_dec_return(&mcast->refcount) <= 1)
639 			wake_up(&mcast->wait);
640 	} else {
641 		rcu_read_lock();
642 		packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
643 		if (!packet->qp) {
644 			rcu_read_unlock();
645 			goto drop;
646 		}
647 		spin_lock_irqsave(&packet->qp->r_lock, flags);
648 		packet_handler = qp_ok(opcode, packet);
649 		if (likely(packet_handler))
650 			packet_handler(packet);
651 		else
652 			ibp->rvp.n_pkt_drops++;
653 		spin_unlock_irqrestore(&packet->qp->r_lock, flags);
654 		rcu_read_unlock();
655 	}
656 	return;
657 
658 drop:
659 	ibp->rvp.n_pkt_drops++;
660 }
661 
662 /*
663  * This is called from a timer to check for QPs
664  * which need kernel memory in order to send a packet.
665  */
666 static void mem_timer(unsigned long data)
667 {
668 	struct hfi1_ibdev *dev = (struct hfi1_ibdev *)data;
669 	struct list_head *list = &dev->memwait;
670 	struct rvt_qp *qp = NULL;
671 	struct iowait *wait;
672 	unsigned long flags;
673 	struct hfi1_qp_priv *priv;
674 
675 	write_seqlock_irqsave(&dev->iowait_lock, flags);
676 	if (!list_empty(list)) {
677 		wait = list_first_entry(list, struct iowait, list);
678 		qp = iowait_to_qp(wait);
679 		priv = qp->priv;
680 		list_del_init(&priv->s_iowait.list);
681 		priv->s_iowait.lock = NULL;
682 		/* refcount held until actual wake up */
683 		if (!list_empty(list))
684 			mod_timer(&dev->mem_timer, jiffies + 1);
685 	}
686 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
687 
688 	if (qp)
689 		hfi1_qp_wakeup(qp, RVT_S_WAIT_KMEM);
690 }
691 
692 void update_sge(struct rvt_sge_state *ss, u32 length)
693 {
694 	struct rvt_sge *sge = &ss->sge;
695 
696 	sge->vaddr += length;
697 	sge->length -= length;
698 	sge->sge_length -= length;
699 	if (sge->sge_length == 0) {
700 		if (--ss->num_sge)
701 			*sge = *ss->sg_list++;
702 	} else if (sge->length == 0 && sge->mr->lkey) {
703 		if (++sge->n >= RVT_SEGSZ) {
704 			if (++sge->m >= sge->mr->mapsz)
705 				return;
706 			sge->n = 0;
707 		}
708 		sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
709 		sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
710 	}
711 }
712 
713 /*
714  * This is called with progress side lock held.
715  */
716 /* New API */
717 static void verbs_sdma_complete(
718 	struct sdma_txreq *cookie,
719 	int status)
720 {
721 	struct verbs_txreq *tx =
722 		container_of(cookie, struct verbs_txreq, txreq);
723 	struct rvt_qp *qp = tx->qp;
724 
725 	spin_lock(&qp->s_lock);
726 	if (tx->wqe) {
727 		hfi1_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
728 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
729 		struct ib_header *hdr;
730 
731 		hdr = &tx->phdr.hdr;
732 		hfi1_rc_send_complete(qp, hdr);
733 	}
734 	spin_unlock(&qp->s_lock);
735 
736 	hfi1_put_txreq(tx);
737 }
738 
739 static int wait_kmem(struct hfi1_ibdev *dev,
740 		     struct rvt_qp *qp,
741 		     struct hfi1_pkt_state *ps)
742 {
743 	struct hfi1_qp_priv *priv = qp->priv;
744 	unsigned long flags;
745 	int ret = 0;
746 
747 	spin_lock_irqsave(&qp->s_lock, flags);
748 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
749 		write_seqlock(&dev->iowait_lock);
750 		list_add_tail(&ps->s_txreq->txreq.list,
751 			      &priv->s_iowait.tx_head);
752 		if (list_empty(&priv->s_iowait.list)) {
753 			if (list_empty(&dev->memwait))
754 				mod_timer(&dev->mem_timer, jiffies + 1);
755 			qp->s_flags |= RVT_S_WAIT_KMEM;
756 			list_add_tail(&priv->s_iowait.list, &dev->memwait);
757 			priv->s_iowait.lock = &dev->iowait_lock;
758 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM);
759 			rvt_get_qp(qp);
760 		}
761 		write_sequnlock(&dev->iowait_lock);
762 		qp->s_flags &= ~RVT_S_BUSY;
763 		ret = -EBUSY;
764 	}
765 	spin_unlock_irqrestore(&qp->s_lock, flags);
766 
767 	return ret;
768 }
769 
770 /*
771  * This routine calls txadds for each sg entry.
772  *
773  * Add failures will revert the sge cursor
774  */
775 static noinline int build_verbs_ulp_payload(
776 	struct sdma_engine *sde,
777 	u32 length,
778 	struct verbs_txreq *tx)
779 {
780 	struct rvt_sge_state *ss = tx->ss;
781 	struct rvt_sge *sg_list = ss->sg_list;
782 	struct rvt_sge sge = ss->sge;
783 	u8 num_sge = ss->num_sge;
784 	u32 len;
785 	int ret = 0;
786 
787 	while (length) {
788 		len = ss->sge.length;
789 		if (len > length)
790 			len = length;
791 		if (len > ss->sge.sge_length)
792 			len = ss->sge.sge_length;
793 		WARN_ON_ONCE(len == 0);
794 		ret = sdma_txadd_kvaddr(
795 			sde->dd,
796 			&tx->txreq,
797 			ss->sge.vaddr,
798 			len);
799 		if (ret)
800 			goto bail_txadd;
801 		update_sge(ss, len);
802 		length -= len;
803 	}
804 	return ret;
805 bail_txadd:
806 	/* unwind cursor */
807 	ss->sge = sge;
808 	ss->num_sge = num_sge;
809 	ss->sg_list = sg_list;
810 	return ret;
811 }
812 
813 /*
814  * Build the number of DMA descriptors needed to send length bytes of data.
815  *
816  * NOTE: DMA mapping is held in the tx until completed in the ring or
817  *       the tx desc is freed without having been submitted to the ring
818  *
819  * This routine ensures all the helper routine calls succeed.
820  */
821 /* New API */
822 static int build_verbs_tx_desc(
823 	struct sdma_engine *sde,
824 	u32 length,
825 	struct verbs_txreq *tx,
826 	struct hfi1_ahg_info *ahg_info,
827 	u64 pbc)
828 {
829 	int ret = 0;
830 	struct hfi1_sdma_header *phdr = &tx->phdr;
831 	u16 hdrbytes = tx->hdr_dwords << 2;
832 
833 	if (!ahg_info->ahgcount) {
834 		ret = sdma_txinit_ahg(
835 			&tx->txreq,
836 			ahg_info->tx_flags,
837 			hdrbytes + length,
838 			ahg_info->ahgidx,
839 			0,
840 			NULL,
841 			0,
842 			verbs_sdma_complete);
843 		if (ret)
844 			goto bail_txadd;
845 		phdr->pbc = cpu_to_le64(pbc);
846 		ret = sdma_txadd_kvaddr(
847 			sde->dd,
848 			&tx->txreq,
849 			phdr,
850 			hdrbytes);
851 		if (ret)
852 			goto bail_txadd;
853 	} else {
854 		ret = sdma_txinit_ahg(
855 			&tx->txreq,
856 			ahg_info->tx_flags,
857 			length,
858 			ahg_info->ahgidx,
859 			ahg_info->ahgcount,
860 			ahg_info->ahgdesc,
861 			hdrbytes,
862 			verbs_sdma_complete);
863 		if (ret)
864 			goto bail_txadd;
865 	}
866 
867 	/* add the ulp payload - if any. tx->ss can be NULL for acks */
868 	if (tx->ss)
869 		ret = build_verbs_ulp_payload(sde, length, tx);
870 bail_txadd:
871 	return ret;
872 }
873 
874 int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
875 			u64 pbc)
876 {
877 	struct hfi1_qp_priv *priv = qp->priv;
878 	struct hfi1_ahg_info *ahg_info = priv->s_ahg;
879 	u32 hdrwords = qp->s_hdrwords;
880 	u32 len = ps->s_txreq->s_cur_size;
881 	u32 plen = hdrwords + ((len + 3) >> 2) + 2; /* includes pbc */
882 	struct hfi1_ibdev *dev = ps->dev;
883 	struct hfi1_pportdata *ppd = ps->ppd;
884 	struct verbs_txreq *tx;
885 	u64 pbc_flags = 0;
886 	u8 sc5 = priv->s_sc;
887 
888 	int ret;
889 
890 	tx = ps->s_txreq;
891 	if (!sdma_txreq_built(&tx->txreq)) {
892 		if (likely(pbc == 0)) {
893 			u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
894 			/* No vl15 here */
895 			/* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
896 			pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
897 
898 			pbc = create_pbc(ppd,
899 					 pbc_flags,
900 					 qp->srate_mbps,
901 					 vl,
902 					 plen);
903 		}
904 		tx->wqe = qp->s_wqe;
905 		ret = build_verbs_tx_desc(tx->sde, len, tx, ahg_info, pbc);
906 		if (unlikely(ret))
907 			goto bail_build;
908 	}
909 	ret =  sdma_send_txreq(tx->sde, &priv->s_iowait, &tx->txreq);
910 	if (unlikely(ret < 0)) {
911 		if (ret == -ECOMM)
912 			goto bail_ecomm;
913 		return ret;
914 	}
915 	trace_sdma_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
916 				&ps->s_txreq->phdr.hdr);
917 	return ret;
918 
919 bail_ecomm:
920 	/* The current one got "sent" */
921 	return 0;
922 bail_build:
923 	ret = wait_kmem(dev, qp, ps);
924 	if (!ret) {
925 		/* free txreq - bad state */
926 		hfi1_put_txreq(ps->s_txreq);
927 		ps->s_txreq = NULL;
928 	}
929 	return ret;
930 }
931 
932 /*
933  * If we are now in the error state, return zero to flush the
934  * send work request.
935  */
936 static int pio_wait(struct rvt_qp *qp,
937 		    struct send_context *sc,
938 		    struct hfi1_pkt_state *ps,
939 		    u32 flag)
940 {
941 	struct hfi1_qp_priv *priv = qp->priv;
942 	struct hfi1_devdata *dd = sc->dd;
943 	struct hfi1_ibdev *dev = &dd->verbs_dev;
944 	unsigned long flags;
945 	int ret = 0;
946 
947 	/*
948 	 * Note that as soon as want_buffer() is called and
949 	 * possibly before it returns, sc_piobufavail()
950 	 * could be called. Therefore, put QP on the I/O wait list before
951 	 * enabling the PIO avail interrupt.
952 	 */
953 	spin_lock_irqsave(&qp->s_lock, flags);
954 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
955 		write_seqlock(&dev->iowait_lock);
956 		list_add_tail(&ps->s_txreq->txreq.list,
957 			      &priv->s_iowait.tx_head);
958 		if (list_empty(&priv->s_iowait.list)) {
959 			struct hfi1_ibdev *dev = &dd->verbs_dev;
960 			int was_empty;
961 
962 			dev->n_piowait += !!(flag & RVT_S_WAIT_PIO);
963 			dev->n_piodrain += !!(flag & RVT_S_WAIT_PIO_DRAIN);
964 			qp->s_flags |= flag;
965 			was_empty = list_empty(&sc->piowait);
966 			list_add_tail(&priv->s_iowait.list, &sc->piowait);
967 			priv->s_iowait.lock = &dev->iowait_lock;
968 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_PIO);
969 			rvt_get_qp(qp);
970 			/* counting: only call wantpiobuf_intr if first user */
971 			if (was_empty)
972 				hfi1_sc_wantpiobuf_intr(sc, 1);
973 		}
974 		write_sequnlock(&dev->iowait_lock);
975 		qp->s_flags &= ~RVT_S_BUSY;
976 		ret = -EBUSY;
977 	}
978 	spin_unlock_irqrestore(&qp->s_lock, flags);
979 	return ret;
980 }
981 
982 static void verbs_pio_complete(void *arg, int code)
983 {
984 	struct rvt_qp *qp = (struct rvt_qp *)arg;
985 	struct hfi1_qp_priv *priv = qp->priv;
986 
987 	if (iowait_pio_dec(&priv->s_iowait))
988 		iowait_drain_wakeup(&priv->s_iowait);
989 }
990 
991 int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
992 			u64 pbc)
993 {
994 	struct hfi1_qp_priv *priv = qp->priv;
995 	u32 hdrwords = qp->s_hdrwords;
996 	struct rvt_sge_state *ss = ps->s_txreq->ss;
997 	u32 len = ps->s_txreq->s_cur_size;
998 	u32 dwords = (len + 3) >> 2;
999 	u32 plen = hdrwords + dwords + 2; /* includes pbc */
1000 	struct hfi1_pportdata *ppd = ps->ppd;
1001 	u32 *hdr = (u32 *)&ps->s_txreq->phdr.hdr;
1002 	u64 pbc_flags = 0;
1003 	u8 sc5;
1004 	unsigned long flags = 0;
1005 	struct send_context *sc;
1006 	struct pio_buf *pbuf;
1007 	int wc_status = IB_WC_SUCCESS;
1008 	int ret = 0;
1009 	pio_release_cb cb = NULL;
1010 
1011 	/* only RC/UC use complete */
1012 	switch (qp->ibqp.qp_type) {
1013 	case IB_QPT_RC:
1014 	case IB_QPT_UC:
1015 		cb = verbs_pio_complete;
1016 		break;
1017 	default:
1018 		break;
1019 	}
1020 
1021 	/* vl15 special case taken care of in ud.c */
1022 	sc5 = priv->s_sc;
1023 	sc = ps->s_txreq->psc;
1024 
1025 	if (likely(pbc == 0)) {
1026 		u8 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
1027 		/* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
1028 		pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
1029 		pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
1030 	}
1031 	if (cb)
1032 		iowait_pio_inc(&priv->s_iowait);
1033 	pbuf = sc_buffer_alloc(sc, plen, cb, qp);
1034 	if (unlikely(!pbuf)) {
1035 		if (cb)
1036 			verbs_pio_complete(qp, 0);
1037 		if (ppd->host_link_state != HLS_UP_ACTIVE) {
1038 			/*
1039 			 * If we have filled the PIO buffers to capacity and are
1040 			 * not in an active state this request is not going to
1041 			 * go out to so just complete it with an error or else a
1042 			 * ULP or the core may be stuck waiting.
1043 			 */
1044 			hfi1_cdbg(
1045 				PIO,
1046 				"alloc failed. state not active, completing");
1047 			wc_status = IB_WC_GENERAL_ERR;
1048 			goto pio_bail;
1049 		} else {
1050 			/*
1051 			 * This is a normal occurrence. The PIO buffs are full
1052 			 * up but we are still happily sending, well we could be
1053 			 * so lets continue to queue the request.
1054 			 */
1055 			hfi1_cdbg(PIO, "alloc failed. state active, queuing");
1056 			ret = pio_wait(qp, sc, ps, RVT_S_WAIT_PIO);
1057 			if (!ret)
1058 				/* txreq not queued - free */
1059 				goto bail;
1060 			/* tx consumed in wait */
1061 			return ret;
1062 		}
1063 	}
1064 
1065 	if (len == 0) {
1066 		pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
1067 	} else {
1068 		if (ss) {
1069 			seg_pio_copy_start(pbuf, pbc, hdr, hdrwords * 4);
1070 			while (len) {
1071 				void *addr = ss->sge.vaddr;
1072 				u32 slen = ss->sge.length;
1073 
1074 				if (slen > len)
1075 					slen = len;
1076 				update_sge(ss, slen);
1077 				seg_pio_copy_mid(pbuf, addr, slen);
1078 				len -= slen;
1079 			}
1080 			seg_pio_copy_end(pbuf);
1081 		}
1082 	}
1083 
1084 	trace_pio_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
1085 			       &ps->s_txreq->phdr.hdr);
1086 
1087 pio_bail:
1088 	if (qp->s_wqe) {
1089 		spin_lock_irqsave(&qp->s_lock, flags);
1090 		hfi1_send_complete(qp, qp->s_wqe, wc_status);
1091 		spin_unlock_irqrestore(&qp->s_lock, flags);
1092 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1093 		spin_lock_irqsave(&qp->s_lock, flags);
1094 		hfi1_rc_send_complete(qp, &ps->s_txreq->phdr.hdr);
1095 		spin_unlock_irqrestore(&qp->s_lock, flags);
1096 	}
1097 
1098 	ret = 0;
1099 
1100 bail:
1101 	hfi1_put_txreq(ps->s_txreq);
1102 	return ret;
1103 }
1104 
1105 /*
1106  * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1107  * being an entry from the partition key table), return 0
1108  * otherwise. Use the matching criteria for egress partition keys
1109  * specified in the OPAv1 spec., section 9.1l.7.
1110  */
1111 static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
1112 {
1113 	u16 mkey = pkey & PKEY_LOW_15_MASK;
1114 	u16 mentry = ent & PKEY_LOW_15_MASK;
1115 
1116 	if (mkey == mentry) {
1117 		/*
1118 		 * If pkey[15] is set (full partition member),
1119 		 * is bit 15 in the corresponding table element
1120 		 * clear (limited member)?
1121 		 */
1122 		if (pkey & PKEY_MEMBER_MASK)
1123 			return !!(ent & PKEY_MEMBER_MASK);
1124 		return 1;
1125 	}
1126 	return 0;
1127 }
1128 
1129 /**
1130  * egress_pkey_check - check P_KEY of a packet
1131  * @ppd:    Physical IB port data
1132  * @lrh: Local route header
1133  * @bth: Base transport header
1134  * @sc5:    SC for packet
1135  * @s_pkey_index: It will be used for look up optimization for kernel contexts
1136  * only. If it is negative value, then it means user contexts is calling this
1137  * function.
1138  *
1139  * It checks if hdr's pkey is valid.
1140  *
1141  * Return: 0 on success, otherwise, 1
1142  */
1143 int egress_pkey_check(struct hfi1_pportdata *ppd, __be16 *lrh, __be32 *bth,
1144 		      u8 sc5, int8_t s_pkey_index)
1145 {
1146 	struct hfi1_devdata *dd;
1147 	int i;
1148 	u16 pkey;
1149 	int is_user_ctxt_mechanism = (s_pkey_index < 0);
1150 
1151 	if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
1152 		return 0;
1153 
1154 	pkey = (u16)be32_to_cpu(bth[0]);
1155 
1156 	/* If SC15, pkey[0:14] must be 0x7fff */
1157 	if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
1158 		goto bad;
1159 
1160 	/* Is the pkey = 0x0, or 0x8000? */
1161 	if ((pkey & PKEY_LOW_15_MASK) == 0)
1162 		goto bad;
1163 
1164 	/*
1165 	 * For the kernel contexts only, if a qp is passed into the function,
1166 	 * the most likely matching pkey has index qp->s_pkey_index
1167 	 */
1168 	if (!is_user_ctxt_mechanism &&
1169 	    egress_pkey_matches_entry(pkey, ppd->pkeys[s_pkey_index])) {
1170 		return 0;
1171 	}
1172 
1173 	for (i = 0; i < MAX_PKEY_VALUES; i++) {
1174 		if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
1175 			return 0;
1176 	}
1177 bad:
1178 	/*
1179 	 * For the user-context mechanism, the P_KEY check would only happen
1180 	 * once per SDMA request, not once per packet.  Therefore, there's no
1181 	 * need to increment the counter for the user-context mechanism.
1182 	 */
1183 	if (!is_user_ctxt_mechanism) {
1184 		incr_cntr64(&ppd->port_xmit_constraint_errors);
1185 		dd = ppd->dd;
1186 		if (!(dd->err_info_xmit_constraint.status &
1187 		      OPA_EI_STATUS_SMASK)) {
1188 			u16 slid = be16_to_cpu(lrh[3]);
1189 
1190 			dd->err_info_xmit_constraint.status |=
1191 				OPA_EI_STATUS_SMASK;
1192 			dd->err_info_xmit_constraint.slid = slid;
1193 			dd->err_info_xmit_constraint.pkey = pkey;
1194 		}
1195 	}
1196 	return 1;
1197 }
1198 
1199 /**
1200  * get_send_routine - choose an egress routine
1201  *
1202  * Choose an egress routine based on QP type
1203  * and size
1204  */
1205 static inline send_routine get_send_routine(struct rvt_qp *qp,
1206 					    struct verbs_txreq *tx)
1207 {
1208 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1209 	struct hfi1_qp_priv *priv = qp->priv;
1210 	struct ib_header *h = &tx->phdr.hdr;
1211 
1212 	if (unlikely(!(dd->flags & HFI1_HAS_SEND_DMA)))
1213 		return dd->process_pio_send;
1214 	switch (qp->ibqp.qp_type) {
1215 	case IB_QPT_SMI:
1216 		return dd->process_pio_send;
1217 	case IB_QPT_GSI:
1218 	case IB_QPT_UD:
1219 		break;
1220 	case IB_QPT_UC:
1221 	case IB_QPT_RC: {
1222 		u8 op = get_opcode(h);
1223 
1224 		if (piothreshold &&
1225 		    tx->s_cur_size <= min(piothreshold, qp->pmtu) &&
1226 		    (BIT(op & OPMASK) & pio_opmask[op >> 5]) &&
1227 		    iowait_sdma_pending(&priv->s_iowait) == 0 &&
1228 		    !sdma_txreq_built(&tx->txreq))
1229 			return dd->process_pio_send;
1230 		break;
1231 	}
1232 	default:
1233 		break;
1234 	}
1235 	return dd->process_dma_send;
1236 }
1237 
1238 /**
1239  * hfi1_verbs_send - send a packet
1240  * @qp: the QP to send on
1241  * @ps: the state of the packet to send
1242  *
1243  * Return zero if packet is sent or queued OK.
1244  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1245  */
1246 int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
1247 {
1248 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1249 	struct hfi1_qp_priv *priv = qp->priv;
1250 	struct ib_other_headers *ohdr;
1251 	struct ib_header *hdr;
1252 	send_routine sr;
1253 	int ret;
1254 	u8 lnh;
1255 
1256 	hdr = &ps->s_txreq->phdr.hdr;
1257 	/* locate the pkey within the headers */
1258 	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
1259 	if (lnh == HFI1_LRH_GRH)
1260 		ohdr = &hdr->u.l.oth;
1261 	else
1262 		ohdr = &hdr->u.oth;
1263 
1264 	sr = get_send_routine(qp, ps->s_txreq);
1265 	ret = egress_pkey_check(dd->pport,
1266 				hdr->lrh,
1267 				ohdr->bth,
1268 				priv->s_sc,
1269 				qp->s_pkey_index);
1270 	if (unlikely(ret)) {
1271 		/*
1272 		 * The value we are returning here does not get propagated to
1273 		 * the verbs caller. Thus we need to complete the request with
1274 		 * error otherwise the caller could be sitting waiting on the
1275 		 * completion event. Only do this for PIO. SDMA has its own
1276 		 * mechanism for handling the errors. So for SDMA we can just
1277 		 * return.
1278 		 */
1279 		if (sr == dd->process_pio_send) {
1280 			unsigned long flags;
1281 
1282 			hfi1_cdbg(PIO, "%s() Failed. Completing with err",
1283 				  __func__);
1284 			spin_lock_irqsave(&qp->s_lock, flags);
1285 			hfi1_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
1286 			spin_unlock_irqrestore(&qp->s_lock, flags);
1287 		}
1288 		return -EINVAL;
1289 	}
1290 	if (sr == dd->process_dma_send && iowait_pio_pending(&priv->s_iowait))
1291 		return pio_wait(qp,
1292 				ps->s_txreq->psc,
1293 				ps,
1294 				RVT_S_WAIT_PIO_DRAIN);
1295 	return sr(qp, ps, 0);
1296 }
1297 
1298 /**
1299  * hfi1_fill_device_attr - Fill in rvt dev info device attributes.
1300  * @dd: the device data structure
1301  */
1302 static void hfi1_fill_device_attr(struct hfi1_devdata *dd)
1303 {
1304 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1305 	u16 ver = dd->dc8051_ver;
1306 
1307 	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1308 
1309 	rdi->dparms.props.fw_ver = ((u64)(dc8051_ver_maj(ver)) << 16) |
1310 				    (u64)dc8051_ver_min(ver);
1311 	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1312 			IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1313 			IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1314 			IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE |
1315 			IB_DEVICE_MEM_MGT_EXTENSIONS;
1316 	rdi->dparms.props.page_size_cap = PAGE_SIZE;
1317 	rdi->dparms.props.vendor_id = dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
1318 	rdi->dparms.props.vendor_part_id = dd->pcidev->device;
1319 	rdi->dparms.props.hw_ver = dd->minrev;
1320 	rdi->dparms.props.sys_image_guid = ib_hfi1_sys_image_guid;
1321 	rdi->dparms.props.max_mr_size = U64_MAX;
1322 	rdi->dparms.props.max_fast_reg_page_list_len = UINT_MAX;
1323 	rdi->dparms.props.max_qp = hfi1_max_qps;
1324 	rdi->dparms.props.max_qp_wr = hfi1_max_qp_wrs;
1325 	rdi->dparms.props.max_sge = hfi1_max_sges;
1326 	rdi->dparms.props.max_sge_rd = hfi1_max_sges;
1327 	rdi->dparms.props.max_cq = hfi1_max_cqs;
1328 	rdi->dparms.props.max_ah = hfi1_max_ahs;
1329 	rdi->dparms.props.max_cqe = hfi1_max_cqes;
1330 	rdi->dparms.props.max_mr = rdi->lkey_table.max;
1331 	rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1332 	rdi->dparms.props.max_map_per_fmr = 32767;
1333 	rdi->dparms.props.max_pd = hfi1_max_pds;
1334 	rdi->dparms.props.max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
1335 	rdi->dparms.props.max_qp_init_rd_atom = 255;
1336 	rdi->dparms.props.max_srq = hfi1_max_srqs;
1337 	rdi->dparms.props.max_srq_wr = hfi1_max_srq_wrs;
1338 	rdi->dparms.props.max_srq_sge = hfi1_max_srq_sges;
1339 	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1340 	rdi->dparms.props.max_pkeys = hfi1_get_npkeys(dd);
1341 	rdi->dparms.props.max_mcast_grp = hfi1_max_mcast_grps;
1342 	rdi->dparms.props.max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
1343 	rdi->dparms.props.max_total_mcast_qp_attach =
1344 					rdi->dparms.props.max_mcast_qp_attach *
1345 					rdi->dparms.props.max_mcast_grp;
1346 }
1347 
1348 static inline u16 opa_speed_to_ib(u16 in)
1349 {
1350 	u16 out = 0;
1351 
1352 	if (in & OPA_LINK_SPEED_25G)
1353 		out |= IB_SPEED_EDR;
1354 	if (in & OPA_LINK_SPEED_12_5G)
1355 		out |= IB_SPEED_FDR;
1356 
1357 	return out;
1358 }
1359 
1360 /*
1361  * Convert a single OPA link width (no multiple flags) to an IB value.
1362  * A zero OPA link width means link down, which means the IB width value
1363  * is a don't care.
1364  */
1365 static inline u16 opa_width_to_ib(u16 in)
1366 {
1367 	switch (in) {
1368 	case OPA_LINK_WIDTH_1X:
1369 	/* map 2x and 3x to 1x as they don't exist in IB */
1370 	case OPA_LINK_WIDTH_2X:
1371 	case OPA_LINK_WIDTH_3X:
1372 		return IB_WIDTH_1X;
1373 	default: /* link down or unknown, return our largest width */
1374 	case OPA_LINK_WIDTH_4X:
1375 		return IB_WIDTH_4X;
1376 	}
1377 }
1378 
1379 static int query_port(struct rvt_dev_info *rdi, u8 port_num,
1380 		      struct ib_port_attr *props)
1381 {
1382 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1383 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1384 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1385 	u16 lid = ppd->lid;
1386 
1387 	props->lid = lid ? lid : 0;
1388 	props->lmc = ppd->lmc;
1389 	/* OPA logical states match IB logical states */
1390 	props->state = driver_lstate(ppd);
1391 	props->phys_state = hfi1_ibphys_portstate(ppd);
1392 	props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
1393 	props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
1394 	/* see rate_show() in ib core/sysfs.c */
1395 	props->active_speed = (u8)opa_speed_to_ib(ppd->link_speed_active);
1396 	props->max_vl_num = ppd->vls_supported;
1397 
1398 	/* Once we are a "first class" citizen and have added the OPA MTUs to
1399 	 * the core we can advertise the larger MTU enum to the ULPs, for now
1400 	 * advertise only 4K.
1401 	 *
1402 	 * Those applications which are either OPA aware or pass the MTU enum
1403 	 * from the Path Records to us will get the new 8k MTU.  Those that
1404 	 * attempt to process the MTU enum may fail in various ways.
1405 	 */
1406 	props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
1407 				      4096 : hfi1_max_mtu), IB_MTU_4096);
1408 	props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
1409 		mtu_to_enum(ppd->ibmtu, IB_MTU_2048);
1410 
1411 	return 0;
1412 }
1413 
1414 static int modify_device(struct ib_device *device,
1415 			 int device_modify_mask,
1416 			 struct ib_device_modify *device_modify)
1417 {
1418 	struct hfi1_devdata *dd = dd_from_ibdev(device);
1419 	unsigned i;
1420 	int ret;
1421 
1422 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1423 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1424 		ret = -EOPNOTSUPP;
1425 		goto bail;
1426 	}
1427 
1428 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1429 		memcpy(device->node_desc, device_modify->node_desc,
1430 		       IB_DEVICE_NODE_DESC_MAX);
1431 		for (i = 0; i < dd->num_pports; i++) {
1432 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1433 
1434 			hfi1_node_desc_chg(ibp);
1435 		}
1436 	}
1437 
1438 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1439 		ib_hfi1_sys_image_guid =
1440 			cpu_to_be64(device_modify->sys_image_guid);
1441 		for (i = 0; i < dd->num_pports; i++) {
1442 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1443 
1444 			hfi1_sys_guid_chg(ibp);
1445 		}
1446 	}
1447 
1448 	ret = 0;
1449 
1450 bail:
1451 	return ret;
1452 }
1453 
1454 static int shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1455 {
1456 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1457 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1458 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1459 	int ret;
1460 
1461 	set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
1462 			     OPA_LINKDOWN_REASON_UNKNOWN);
1463 	ret = set_link_state(ppd, HLS_DN_DOWNDEF);
1464 	return ret;
1465 }
1466 
1467 static int hfi1_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1468 			    int guid_index, __be64 *guid)
1469 {
1470 	struct hfi1_ibport *ibp = container_of(rvp, struct hfi1_ibport, rvp);
1471 
1472 	if (guid_index >= HFI1_GUIDS_PER_PORT)
1473 		return -EINVAL;
1474 
1475 	*guid = get_sguid(ibp, guid_index);
1476 	return 0;
1477 }
1478 
1479 /*
1480  * convert ah port,sl to sc
1481  */
1482 u8 ah_to_sc(struct ib_device *ibdev, struct ib_ah_attr *ah)
1483 {
1484 	struct hfi1_ibport *ibp = to_iport(ibdev, ah->port_num);
1485 
1486 	return ibp->sl_to_sc[ah->sl];
1487 }
1488 
1489 static int hfi1_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr)
1490 {
1491 	struct hfi1_ibport *ibp;
1492 	struct hfi1_pportdata *ppd;
1493 	struct hfi1_devdata *dd;
1494 	u8 sc5;
1495 
1496 	/* test the mapping for validity */
1497 	ibp = to_iport(ibdev, ah_attr->port_num);
1498 	ppd = ppd_from_ibp(ibp);
1499 	sc5 = ibp->sl_to_sc[ah_attr->sl];
1500 	dd = dd_from_ppd(ppd);
1501 	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1502 		return -EINVAL;
1503 	return 0;
1504 }
1505 
1506 static void hfi1_notify_new_ah(struct ib_device *ibdev,
1507 			       struct ib_ah_attr *ah_attr,
1508 			       struct rvt_ah *ah)
1509 {
1510 	struct hfi1_ibport *ibp;
1511 	struct hfi1_pportdata *ppd;
1512 	struct hfi1_devdata *dd;
1513 	u8 sc5;
1514 
1515 	/*
1516 	 * Do not trust reading anything from rvt_ah at this point as it is not
1517 	 * done being setup. We can however modify things which we need to set.
1518 	 */
1519 
1520 	ibp = to_iport(ibdev, ah_attr->port_num);
1521 	ppd = ppd_from_ibp(ibp);
1522 	sc5 = ibp->sl_to_sc[ah->attr.sl];
1523 	dd = dd_from_ppd(ppd);
1524 	ah->vl = sc_to_vlt(dd, sc5);
1525 	if (ah->vl < num_vls || ah->vl == 15)
1526 		ah->log_pmtu = ilog2(dd->vld[ah->vl].mtu);
1527 }
1528 
1529 struct ib_ah *hfi1_create_qp0_ah(struct hfi1_ibport *ibp, u16 dlid)
1530 {
1531 	struct ib_ah_attr attr;
1532 	struct ib_ah *ah = ERR_PTR(-EINVAL);
1533 	struct rvt_qp *qp0;
1534 
1535 	memset(&attr, 0, sizeof(attr));
1536 	attr.dlid = dlid;
1537 	attr.port_num = ppd_from_ibp(ibp)->port;
1538 	rcu_read_lock();
1539 	qp0 = rcu_dereference(ibp->rvp.qp[0]);
1540 	if (qp0)
1541 		ah = ib_create_ah(qp0->ibqp.pd, &attr);
1542 	rcu_read_unlock();
1543 	return ah;
1544 }
1545 
1546 /**
1547  * hfi1_get_npkeys - return the size of the PKEY table for context 0
1548  * @dd: the hfi1_ib device
1549  */
1550 unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
1551 {
1552 	return ARRAY_SIZE(dd->pport[0].pkeys);
1553 }
1554 
1555 static void init_ibport(struct hfi1_pportdata *ppd)
1556 {
1557 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1558 	size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
1559 	int i;
1560 
1561 	for (i = 0; i < sz; i++) {
1562 		ibp->sl_to_sc[i] = i;
1563 		ibp->sc_to_sl[i] = i;
1564 	}
1565 
1566 	spin_lock_init(&ibp->rvp.lock);
1567 	/* Set the prefix to the default value (see ch. 4.1.1) */
1568 	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1569 	ibp->rvp.sm_lid = 0;
1570 	/* Below should only set bits defined in OPA PortInfo.CapabilityMask */
1571 	ibp->rvp.port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
1572 		IB_PORT_CAP_MASK_NOTICE_SUP;
1573 	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1574 	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1575 	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1576 	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1577 	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1578 
1579 	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1580 	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1581 }
1582 
1583 static void hfi1_get_dev_fw_str(struct ib_device *ibdev, char *str,
1584 				size_t str_len)
1585 {
1586 	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
1587 	struct hfi1_ibdev *dev = dev_from_rdi(rdi);
1588 	u16 ver = dd_from_dev(dev)->dc8051_ver;
1589 
1590 	snprintf(str, str_len, "%u.%u", dc8051_ver_maj(ver),
1591 		 dc8051_ver_min(ver));
1592 }
1593 
1594 static const char * const driver_cntr_names[] = {
1595 	/* must be element 0*/
1596 	"DRIVER_KernIntr",
1597 	"DRIVER_ErrorIntr",
1598 	"DRIVER_Tx_Errs",
1599 	"DRIVER_Rcv_Errs",
1600 	"DRIVER_HW_Errs",
1601 	"DRIVER_NoPIOBufs",
1602 	"DRIVER_CtxtsOpen",
1603 	"DRIVER_RcvLen_Errs",
1604 	"DRIVER_EgrBufFull",
1605 	"DRIVER_EgrHdrFull"
1606 };
1607 
1608 static const char **dev_cntr_names;
1609 static const char **port_cntr_names;
1610 static int num_driver_cntrs = ARRAY_SIZE(driver_cntr_names);
1611 static int num_dev_cntrs;
1612 static int num_port_cntrs;
1613 static int cntr_names_initialized;
1614 
1615 /*
1616  * Convert a list of names separated by '\n' into an array of NULL terminated
1617  * strings. Optionally some entries can be reserved in the array to hold extra
1618  * external strings.
1619  */
1620 static int init_cntr_names(const char *names_in,
1621 			   const int names_len,
1622 			   int num_extra_names,
1623 			   int *num_cntrs,
1624 			   const char ***cntr_names)
1625 {
1626 	char *names_out, *p, **q;
1627 	int i, n;
1628 
1629 	n = 0;
1630 	for (i = 0; i < names_len; i++)
1631 		if (names_in[i] == '\n')
1632 			n++;
1633 
1634 	names_out = kmalloc((n + num_extra_names) * sizeof(char *) + names_len,
1635 			    GFP_KERNEL);
1636 	if (!names_out) {
1637 		*num_cntrs = 0;
1638 		*cntr_names = NULL;
1639 		return -ENOMEM;
1640 	}
1641 
1642 	p = names_out + (n + num_extra_names) * sizeof(char *);
1643 	memcpy(p, names_in, names_len);
1644 
1645 	q = (char **)names_out;
1646 	for (i = 0; i < n; i++) {
1647 		q[i] = p;
1648 		p = strchr(p, '\n');
1649 		*p++ = '\0';
1650 	}
1651 
1652 	*num_cntrs = n;
1653 	*cntr_names = (const char **)names_out;
1654 	return 0;
1655 }
1656 
1657 static struct rdma_hw_stats *alloc_hw_stats(struct ib_device *ibdev,
1658 					    u8 port_num)
1659 {
1660 	int i, err;
1661 
1662 	if (!cntr_names_initialized) {
1663 		struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1664 
1665 		err = init_cntr_names(dd->cntrnames,
1666 				      dd->cntrnameslen,
1667 				      num_driver_cntrs,
1668 				      &num_dev_cntrs,
1669 				      &dev_cntr_names);
1670 		if (err)
1671 			return NULL;
1672 
1673 		for (i = 0; i < num_driver_cntrs; i++)
1674 			dev_cntr_names[num_dev_cntrs + i] =
1675 				driver_cntr_names[i];
1676 
1677 		err = init_cntr_names(dd->portcntrnames,
1678 				      dd->portcntrnameslen,
1679 				      0,
1680 				      &num_port_cntrs,
1681 				      &port_cntr_names);
1682 		if (err) {
1683 			kfree(dev_cntr_names);
1684 			dev_cntr_names = NULL;
1685 			return NULL;
1686 		}
1687 		cntr_names_initialized = 1;
1688 	}
1689 
1690 	if (!port_num)
1691 		return rdma_alloc_hw_stats_struct(
1692 				dev_cntr_names,
1693 				num_dev_cntrs + num_driver_cntrs,
1694 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1695 	else
1696 		return rdma_alloc_hw_stats_struct(
1697 				port_cntr_names,
1698 				num_port_cntrs,
1699 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1700 }
1701 
1702 static u64 hfi1_sps_ints(void)
1703 {
1704 	unsigned long flags;
1705 	struct hfi1_devdata *dd;
1706 	u64 sps_ints = 0;
1707 
1708 	spin_lock_irqsave(&hfi1_devs_lock, flags);
1709 	list_for_each_entry(dd, &hfi1_dev_list, list) {
1710 		sps_ints += get_all_cpu_total(dd->int_counter);
1711 	}
1712 	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1713 	return sps_ints;
1714 }
1715 
1716 static int get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1717 			u8 port, int index)
1718 {
1719 	u64 *values;
1720 	int count;
1721 
1722 	if (!port) {
1723 		u64 *stats = (u64 *)&hfi1_stats;
1724 		int i;
1725 
1726 		hfi1_read_cntrs(dd_from_ibdev(ibdev), NULL, &values);
1727 		values[num_dev_cntrs] = hfi1_sps_ints();
1728 		for (i = 1; i < num_driver_cntrs; i++)
1729 			values[num_dev_cntrs + i] = stats[i];
1730 		count = num_dev_cntrs + num_driver_cntrs;
1731 	} else {
1732 		struct hfi1_ibport *ibp = to_iport(ibdev, port);
1733 
1734 		hfi1_read_portcntrs(ppd_from_ibp(ibp), NULL, &values);
1735 		count = num_port_cntrs;
1736 	}
1737 
1738 	memcpy(stats->value, values, count * sizeof(u64));
1739 	return count;
1740 }
1741 
1742 /**
1743  * hfi1_register_ib_device - register our device with the infiniband core
1744  * @dd: the device data structure
1745  * Return 0 if successful, errno if unsuccessful.
1746  */
1747 int hfi1_register_ib_device(struct hfi1_devdata *dd)
1748 {
1749 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1750 	struct ib_device *ibdev = &dev->rdi.ibdev;
1751 	struct hfi1_pportdata *ppd = dd->pport;
1752 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1753 	unsigned i;
1754 	int ret;
1755 	size_t lcpysz = IB_DEVICE_NAME_MAX;
1756 
1757 	for (i = 0; i < dd->num_pports; i++)
1758 		init_ibport(ppd + i);
1759 
1760 	/* Only need to initialize non-zero fields. */
1761 
1762 	setup_timer(&dev->mem_timer, mem_timer, (unsigned long)dev);
1763 
1764 	seqlock_init(&dev->iowait_lock);
1765 	seqlock_init(&dev->txwait_lock);
1766 	INIT_LIST_HEAD(&dev->txwait);
1767 	INIT_LIST_HEAD(&dev->memwait);
1768 
1769 	ret = verbs_txreq_init(dev);
1770 	if (ret)
1771 		goto err_verbs_txreq;
1772 
1773 	/* Use first-port GUID as node guid */
1774 	ibdev->node_guid = get_sguid(ibp, HFI1_PORT_GUID_INDEX);
1775 
1776 	/*
1777 	 * The system image GUID is supposed to be the same for all
1778 	 * HFIs in a single system but since there can be other
1779 	 * device types in the system, we can't be sure this is unique.
1780 	 */
1781 	if (!ib_hfi1_sys_image_guid)
1782 		ib_hfi1_sys_image_guid = ibdev->node_guid;
1783 	lcpysz = strlcpy(ibdev->name, class_name(), lcpysz);
1784 	strlcpy(ibdev->name + lcpysz, "_%d", IB_DEVICE_NAME_MAX - lcpysz);
1785 	ibdev->owner = THIS_MODULE;
1786 	ibdev->phys_port_cnt = dd->num_pports;
1787 	ibdev->dma_device = &dd->pcidev->dev;
1788 	ibdev->modify_device = modify_device;
1789 	ibdev->alloc_hw_stats = alloc_hw_stats;
1790 	ibdev->get_hw_stats = get_hw_stats;
1791 
1792 	/* keep process mad in the driver */
1793 	ibdev->process_mad = hfi1_process_mad;
1794 	ibdev->get_dev_fw_str = hfi1_get_dev_fw_str;
1795 
1796 	strncpy(ibdev->node_desc, init_utsname()->nodename,
1797 		sizeof(ibdev->node_desc));
1798 
1799 	/*
1800 	 * Fill in rvt info object.
1801 	 */
1802 	dd->verbs_dev.rdi.driver_f.port_callback = hfi1_create_port_files;
1803 	dd->verbs_dev.rdi.driver_f.get_card_name = get_card_name;
1804 	dd->verbs_dev.rdi.driver_f.get_pci_dev = get_pci_dev;
1805 	dd->verbs_dev.rdi.driver_f.check_ah = hfi1_check_ah;
1806 	dd->verbs_dev.rdi.driver_f.notify_new_ah = hfi1_notify_new_ah;
1807 	dd->verbs_dev.rdi.driver_f.get_guid_be = hfi1_get_guid_be;
1808 	dd->verbs_dev.rdi.driver_f.query_port_state = query_port;
1809 	dd->verbs_dev.rdi.driver_f.shut_down_port = shut_down_port;
1810 	dd->verbs_dev.rdi.driver_f.cap_mask_chg = hfi1_cap_mask_chg;
1811 	/*
1812 	 * Fill in rvt info device attributes.
1813 	 */
1814 	hfi1_fill_device_attr(dd);
1815 
1816 	/* queue pair */
1817 	dd->verbs_dev.rdi.dparms.qp_table_size = hfi1_qp_table_size;
1818 	dd->verbs_dev.rdi.dparms.qpn_start = 0;
1819 	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1820 	dd->verbs_dev.rdi.dparms.qos_shift = dd->qos_shift;
1821 	dd->verbs_dev.rdi.dparms.qpn_res_start = kdeth_qp << 16;
1822 	dd->verbs_dev.rdi.dparms.qpn_res_end =
1823 	dd->verbs_dev.rdi.dparms.qpn_res_start + 65535;
1824 	dd->verbs_dev.rdi.dparms.max_rdma_atomic = HFI1_MAX_RDMA_ATOMIC;
1825 	dd->verbs_dev.rdi.dparms.psn_mask = PSN_MASK;
1826 	dd->verbs_dev.rdi.dparms.psn_shift = PSN_SHIFT;
1827 	dd->verbs_dev.rdi.dparms.psn_modify_mask = PSN_MODIFY_MASK;
1828 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_INTEL_OPA;
1829 	dd->verbs_dev.rdi.dparms.max_mad_size = OPA_MGMT_MAD_SIZE;
1830 
1831 	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qp_priv_alloc;
1832 	dd->verbs_dev.rdi.driver_f.qp_priv_free = qp_priv_free;
1833 	dd->verbs_dev.rdi.driver_f.free_all_qps = free_all_qps;
1834 	dd->verbs_dev.rdi.driver_f.notify_qp_reset = notify_qp_reset;
1835 	dd->verbs_dev.rdi.driver_f.do_send = hfi1_do_send;
1836 	dd->verbs_dev.rdi.driver_f.schedule_send = hfi1_schedule_send;
1837 	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _hfi1_schedule_send;
1838 	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = get_pmtu_from_attr;
1839 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1840 	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = flush_qp_waiters;
1841 	dd->verbs_dev.rdi.driver_f.stop_send_queue = stop_send_queue;
1842 	dd->verbs_dev.rdi.driver_f.quiesce_qp = quiesce_qp;
1843 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1844 	dd->verbs_dev.rdi.driver_f.mtu_from_qp = mtu_from_qp;
1845 	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = mtu_to_path_mtu;
1846 	dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
1847 	dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
1848 	dd->verbs_dev.rdi.driver_f.check_send_wqe = hfi1_check_send_wqe;
1849 
1850 	/* completeion queue */
1851 	snprintf(dd->verbs_dev.rdi.dparms.cq_name,
1852 		 sizeof(dd->verbs_dev.rdi.dparms.cq_name),
1853 		 "hfi1_cq%d", dd->unit);
1854 	dd->verbs_dev.rdi.dparms.node = dd->node;
1855 
1856 	/* misc settings */
1857 	dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
1858 	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
1859 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1860 	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
1861 
1862 	/* post send table */
1863 	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;
1864 
1865 	ppd = dd->pport;
1866 	for (i = 0; i < dd->num_pports; i++, ppd++)
1867 		rvt_init_port(&dd->verbs_dev.rdi,
1868 			      &ppd->ibport_data.rvp,
1869 			      i,
1870 			      ppd->pkeys);
1871 
1872 	ret = rvt_register_device(&dd->verbs_dev.rdi);
1873 	if (ret)
1874 		goto err_verbs_txreq;
1875 
1876 	ret = hfi1_verbs_register_sysfs(dd);
1877 	if (ret)
1878 		goto err_class;
1879 
1880 	return ret;
1881 
1882 err_class:
1883 	rvt_unregister_device(&dd->verbs_dev.rdi);
1884 err_verbs_txreq:
1885 	verbs_txreq_exit(dev);
1886 	dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1887 	return ret;
1888 }
1889 
1890 void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
1891 {
1892 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1893 
1894 	hfi1_verbs_unregister_sysfs(dd);
1895 
1896 	rvt_unregister_device(&dd->verbs_dev.rdi);
1897 
1898 	if (!list_empty(&dev->txwait))
1899 		dd_dev_err(dd, "txwait list not empty!\n");
1900 	if (!list_empty(&dev->memwait))
1901 		dd_dev_err(dd, "memwait list not empty!\n");
1902 
1903 	del_timer_sync(&dev->mem_timer);
1904 	verbs_txreq_exit(dev);
1905 
1906 	kfree(dev_cntr_names);
1907 	kfree(port_cntr_names);
1908 	cntr_names_initialized = 0;
1909 }
1910 
1911 void hfi1_cnp_rcv(struct hfi1_packet *packet)
1912 {
1913 	struct hfi1_ibport *ibp = &packet->rcd->ppd->ibport_data;
1914 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1915 	struct ib_header *hdr = packet->hdr;
1916 	struct rvt_qp *qp = packet->qp;
1917 	u32 lqpn, rqpn = 0;
1918 	u16 rlid = 0;
1919 	u8 sl, sc5, svc_type;
1920 
1921 	switch (packet->qp->ibqp.qp_type) {
1922 	case IB_QPT_UC:
1923 		rlid = qp->remote_ah_attr.dlid;
1924 		rqpn = qp->remote_qpn;
1925 		svc_type = IB_CC_SVCTYPE_UC;
1926 		break;
1927 	case IB_QPT_RC:
1928 		rlid = qp->remote_ah_attr.dlid;
1929 		rqpn = qp->remote_qpn;
1930 		svc_type = IB_CC_SVCTYPE_RC;
1931 		break;
1932 	case IB_QPT_SMI:
1933 	case IB_QPT_GSI:
1934 	case IB_QPT_UD:
1935 		svc_type = IB_CC_SVCTYPE_UD;
1936 		break;
1937 	default:
1938 		ibp->rvp.n_pkt_drops++;
1939 		return;
1940 	}
1941 
1942 	sc5 = hdr2sc(hdr, packet->rhf);
1943 	sl = ibp->sc_to_sl[sc5];
1944 	lqpn = qp->ibqp.qp_num;
1945 
1946 	process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
1947 }
1948