xref: /openbmc/linux/drivers/infiniband/hw/hfi1/verbs.c (revision bc05aa6e)
1 /*
2  * Copyright(c) 2015 - 2017 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 #include <rdma/opa_addr.h>
57 
58 #include "hfi.h"
59 #include "common.h"
60 #include "device.h"
61 #include "trace.h"
62 #include "qp.h"
63 #include "verbs_txreq.h"
64 #include "debugfs.h"
65 #include "vnic.h"
66 
67 static unsigned int hfi1_lkey_table_size = 16;
68 module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
69 		   S_IRUGO);
70 MODULE_PARM_DESC(lkey_table_size,
71 		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
72 
73 static unsigned int hfi1_max_pds = 0xFFFF;
74 module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
75 MODULE_PARM_DESC(max_pds,
76 		 "Maximum number of protection domains to support");
77 
78 static unsigned int hfi1_max_ahs = 0xFFFF;
79 module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
80 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
81 
82 unsigned int hfi1_max_cqes = 0x2FFFFF;
83 module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
84 MODULE_PARM_DESC(max_cqes,
85 		 "Maximum number of completion queue entries to support");
86 
87 unsigned int hfi1_max_cqs = 0x1FFFF;
88 module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
89 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
90 
91 unsigned int hfi1_max_qp_wrs = 0x3FFF;
92 module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
93 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
94 
95 unsigned int hfi1_max_qps = 32768;
96 module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
97 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
98 
99 unsigned int hfi1_max_sges = 0x60;
100 module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
101 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
102 
103 unsigned int hfi1_max_mcast_grps = 16384;
104 module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
105 MODULE_PARM_DESC(max_mcast_grps,
106 		 "Maximum number of multicast groups to support");
107 
108 unsigned int hfi1_max_mcast_qp_attached = 16;
109 module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
110 		   uint, S_IRUGO);
111 MODULE_PARM_DESC(max_mcast_qp_attached,
112 		 "Maximum number of attached QPs to support");
113 
114 unsigned int hfi1_max_srqs = 1024;
115 module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
116 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
117 
118 unsigned int hfi1_max_srq_sges = 128;
119 module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
120 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
121 
122 unsigned int hfi1_max_srq_wrs = 0x1FFFF;
123 module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
124 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
125 
126 unsigned short piothreshold = 256;
127 module_param(piothreshold, ushort, S_IRUGO);
128 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
129 
130 #define COPY_CACHELESS 1
131 #define COPY_ADAPTIVE  2
132 static unsigned int sge_copy_mode;
133 module_param(sge_copy_mode, uint, S_IRUGO);
134 MODULE_PARM_DESC(sge_copy_mode,
135 		 "Verbs copy mode: 0 use memcpy, 1 use cacheless copy, 2 adapt based on WSS");
136 
137 static void verbs_sdma_complete(
138 	struct sdma_txreq *cookie,
139 	int status);
140 
141 static int pio_wait(struct rvt_qp *qp,
142 		    struct send_context *sc,
143 		    struct hfi1_pkt_state *ps,
144 		    u32 flag);
145 
146 /* Length of buffer to create verbs txreq cache name */
147 #define TXREQ_NAME_LEN 24
148 
149 /* 16B trailing buffer */
150 static const u8 trail_buf[MAX_16B_PADDING];
151 
152 static uint wss_threshold;
153 module_param(wss_threshold, uint, S_IRUGO);
154 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
155 static uint wss_clean_period = 256;
156 module_param(wss_clean_period, uint, S_IRUGO);
157 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
158 
159 /* memory working set size */
160 struct hfi1_wss {
161 	unsigned long *entries;
162 	atomic_t total_count;
163 	atomic_t clean_counter;
164 	atomic_t clean_entry;
165 
166 	int threshold;
167 	int num_entries;
168 	long pages_mask;
169 };
170 
171 static struct hfi1_wss wss;
172 
173 int hfi1_wss_init(void)
174 {
175 	long llc_size;
176 	long llc_bits;
177 	long table_size;
178 	long table_bits;
179 
180 	/* check for a valid percent range - default to 80 if none or invalid */
181 	if (wss_threshold < 1 || wss_threshold > 100)
182 		wss_threshold = 80;
183 	/* reject a wildly large period */
184 	if (wss_clean_period > 1000000)
185 		wss_clean_period = 256;
186 	/* reject a zero period */
187 	if (wss_clean_period == 0)
188 		wss_clean_period = 1;
189 
190 	/*
191 	 * Calculate the table size - the next power of 2 larger than the
192 	 * LLC size.  LLC size is in KiB.
193 	 */
194 	llc_size = wss_llc_size() * 1024;
195 	table_size = roundup_pow_of_two(llc_size);
196 
197 	/* one bit per page in rounded up table */
198 	llc_bits = llc_size / PAGE_SIZE;
199 	table_bits = table_size / PAGE_SIZE;
200 	wss.pages_mask = table_bits - 1;
201 	wss.num_entries = table_bits / BITS_PER_LONG;
202 
203 	wss.threshold = (llc_bits * wss_threshold) / 100;
204 	if (wss.threshold == 0)
205 		wss.threshold = 1;
206 
207 	atomic_set(&wss.clean_counter, wss_clean_period);
208 
209 	wss.entries = kcalloc(wss.num_entries, sizeof(*wss.entries),
210 			      GFP_KERNEL);
211 	if (!wss.entries) {
212 		hfi1_wss_exit();
213 		return -ENOMEM;
214 	}
215 
216 	return 0;
217 }
218 
219 void hfi1_wss_exit(void)
220 {
221 	/* coded to handle partially initialized and repeat callers */
222 	kfree(wss.entries);
223 	wss.entries = NULL;
224 }
225 
226 /*
227  * Advance the clean counter.  When the clean period has expired,
228  * clean an entry.
229  *
230  * This is implemented in atomics to avoid locking.  Because multiple
231  * variables are involved, it can be racy which can lead to slightly
232  * inaccurate information.  Since this is only a heuristic, this is
233  * OK.  Any innaccuracies will clean themselves out as the counter
234  * advances.  That said, it is unlikely the entry clean operation will
235  * race - the next possible racer will not start until the next clean
236  * period.
237  *
238  * The clean counter is implemented as a decrement to zero.  When zero
239  * is reached an entry is cleaned.
240  */
241 static void wss_advance_clean_counter(void)
242 {
243 	int entry;
244 	int weight;
245 	unsigned long bits;
246 
247 	/* become the cleaner if we decrement the counter to zero */
248 	if (atomic_dec_and_test(&wss.clean_counter)) {
249 		/*
250 		 * Set, not add, the clean period.  This avoids an issue
251 		 * where the counter could decrement below the clean period.
252 		 * Doing a set can result in lost decrements, slowing the
253 		 * clean advance.  Since this a heuristic, this possible
254 		 * slowdown is OK.
255 		 *
256 		 * An alternative is to loop, advancing the counter by a
257 		 * clean period until the result is > 0. However, this could
258 		 * lead to several threads keeping another in the clean loop.
259 		 * This could be mitigated by limiting the number of times
260 		 * we stay in the loop.
261 		 */
262 		atomic_set(&wss.clean_counter, wss_clean_period);
263 
264 		/*
265 		 * Uniquely grab the entry to clean and move to next.
266 		 * The current entry is always the lower bits of
267 		 * wss.clean_entry.  The table size, wss.num_entries,
268 		 * is always a power-of-2.
269 		 */
270 		entry = (atomic_inc_return(&wss.clean_entry) - 1)
271 			& (wss.num_entries - 1);
272 
273 		/* clear the entry and count the bits */
274 		bits = xchg(&wss.entries[entry], 0);
275 		weight = hweight64((u64)bits);
276 		/* only adjust the contended total count if needed */
277 		if (weight)
278 			atomic_sub(weight, &wss.total_count);
279 	}
280 }
281 
282 /*
283  * Insert the given address into the working set array.
284  */
285 static void wss_insert(void *address)
286 {
287 	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss.pages_mask;
288 	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
289 	u32 nr = page & (BITS_PER_LONG - 1);
290 
291 	if (!test_and_set_bit(nr, &wss.entries[entry]))
292 		atomic_inc(&wss.total_count);
293 
294 	wss_advance_clean_counter();
295 }
296 
297 /*
298  * Is the working set larger than the threshold?
299  */
300 static inline bool wss_exceeds_threshold(void)
301 {
302 	return atomic_read(&wss.total_count) >= wss.threshold;
303 }
304 
305 /*
306  * Translate ib_wr_opcode into ib_wc_opcode.
307  */
308 const enum ib_wc_opcode ib_hfi1_wc_opcode[] = {
309 	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
310 	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
311 	[IB_WR_SEND] = IB_WC_SEND,
312 	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
313 	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
314 	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
315 	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD,
316 	[IB_WR_SEND_WITH_INV] = IB_WC_SEND,
317 	[IB_WR_LOCAL_INV] = IB_WC_LOCAL_INV,
318 	[IB_WR_REG_MR] = IB_WC_REG_MR
319 };
320 
321 /*
322  * Length of header by opcode, 0 --> not supported
323  */
324 const u8 hdr_len_by_opcode[256] = {
325 	/* RC */
326 	[IB_OPCODE_RC_SEND_FIRST]                     = 12 + 8,
327 	[IB_OPCODE_RC_SEND_MIDDLE]                    = 12 + 8,
328 	[IB_OPCODE_RC_SEND_LAST]                      = 12 + 8,
329 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
330 	[IB_OPCODE_RC_SEND_ONLY]                      = 12 + 8,
331 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
332 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
333 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = 12 + 8,
334 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = 12 + 8,
335 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
336 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
337 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
338 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = 12 + 8 + 16,
339 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = 12 + 8 + 4,
340 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = 12 + 8,
341 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = 12 + 8 + 4,
342 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = 12 + 8 + 4,
343 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = 12 + 8 + 4,
344 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = 12 + 8 + 4 + 8,
345 	[IB_OPCODE_RC_COMPARE_SWAP]                   = 12 + 8 + 28,
346 	[IB_OPCODE_RC_FETCH_ADD]                      = 12 + 8 + 28,
347 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = 12 + 8 + 4,
348 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = 12 + 8 + 4,
349 	/* UC */
350 	[IB_OPCODE_UC_SEND_FIRST]                     = 12 + 8,
351 	[IB_OPCODE_UC_SEND_MIDDLE]                    = 12 + 8,
352 	[IB_OPCODE_UC_SEND_LAST]                      = 12 + 8,
353 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
354 	[IB_OPCODE_UC_SEND_ONLY]                      = 12 + 8,
355 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
356 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
357 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = 12 + 8,
358 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = 12 + 8,
359 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
360 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
361 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
362 	/* UD */
363 	[IB_OPCODE_UD_SEND_ONLY]                      = 12 + 8 + 8,
364 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 12
365 };
366 
367 static const opcode_handler opcode_handler_tbl[256] = {
368 	/* RC */
369 	[IB_OPCODE_RC_SEND_FIRST]                     = &hfi1_rc_rcv,
370 	[IB_OPCODE_RC_SEND_MIDDLE]                    = &hfi1_rc_rcv,
371 	[IB_OPCODE_RC_SEND_LAST]                      = &hfi1_rc_rcv,
372 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
373 	[IB_OPCODE_RC_SEND_ONLY]                      = &hfi1_rc_rcv,
374 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
375 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = &hfi1_rc_rcv,
376 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = &hfi1_rc_rcv,
377 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = &hfi1_rc_rcv,
378 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
379 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = &hfi1_rc_rcv,
380 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
381 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = &hfi1_rc_rcv,
382 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = &hfi1_rc_rcv,
383 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = &hfi1_rc_rcv,
384 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = &hfi1_rc_rcv,
385 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = &hfi1_rc_rcv,
386 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = &hfi1_rc_rcv,
387 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = &hfi1_rc_rcv,
388 	[IB_OPCODE_RC_COMPARE_SWAP]                   = &hfi1_rc_rcv,
389 	[IB_OPCODE_RC_FETCH_ADD]                      = &hfi1_rc_rcv,
390 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = &hfi1_rc_rcv,
391 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = &hfi1_rc_rcv,
392 	/* UC */
393 	[IB_OPCODE_UC_SEND_FIRST]                     = &hfi1_uc_rcv,
394 	[IB_OPCODE_UC_SEND_MIDDLE]                    = &hfi1_uc_rcv,
395 	[IB_OPCODE_UC_SEND_LAST]                      = &hfi1_uc_rcv,
396 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
397 	[IB_OPCODE_UC_SEND_ONLY]                      = &hfi1_uc_rcv,
398 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
399 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = &hfi1_uc_rcv,
400 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = &hfi1_uc_rcv,
401 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = &hfi1_uc_rcv,
402 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
403 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = &hfi1_uc_rcv,
404 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
405 	/* UD */
406 	[IB_OPCODE_UD_SEND_ONLY]                      = &hfi1_ud_rcv,
407 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_ud_rcv,
408 	/* CNP */
409 	[IB_OPCODE_CNP]				      = &hfi1_cnp_rcv
410 };
411 
412 #define OPMASK 0x1f
413 
414 static const u32 pio_opmask[BIT(3)] = {
415 	/* RC */
416 	[IB_OPCODE_RC >> 5] =
417 		BIT(RC_OP(SEND_ONLY) & OPMASK) |
418 		BIT(RC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
419 		BIT(RC_OP(RDMA_WRITE_ONLY) & OPMASK) |
420 		BIT(RC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK) |
421 		BIT(RC_OP(RDMA_READ_REQUEST) & OPMASK) |
422 		BIT(RC_OP(ACKNOWLEDGE) & OPMASK) |
423 		BIT(RC_OP(ATOMIC_ACKNOWLEDGE) & OPMASK) |
424 		BIT(RC_OP(COMPARE_SWAP) & OPMASK) |
425 		BIT(RC_OP(FETCH_ADD) & OPMASK),
426 	/* UC */
427 	[IB_OPCODE_UC >> 5] =
428 		BIT(UC_OP(SEND_ONLY) & OPMASK) |
429 		BIT(UC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
430 		BIT(UC_OP(RDMA_WRITE_ONLY) & OPMASK) |
431 		BIT(UC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK),
432 };
433 
434 /*
435  * System image GUID.
436  */
437 __be64 ib_hfi1_sys_image_guid;
438 
439 /**
440  * hfi1_copy_sge - copy data to SGE memory
441  * @ss: the SGE state
442  * @data: the data to copy
443  * @length: the length of the data
444  * @release: boolean to release MR
445  * @copy_last: do a separate copy of the last 8 bytes
446  */
447 void hfi1_copy_sge(
448 	struct rvt_sge_state *ss,
449 	void *data, u32 length,
450 	bool release,
451 	bool copy_last)
452 {
453 	struct rvt_sge *sge = &ss->sge;
454 	int i;
455 	bool in_last = false;
456 	bool cacheless_copy = false;
457 
458 	if (sge_copy_mode == COPY_CACHELESS) {
459 		cacheless_copy = length >= PAGE_SIZE;
460 	} else if (sge_copy_mode == COPY_ADAPTIVE) {
461 		if (length >= PAGE_SIZE) {
462 			/*
463 			 * NOTE: this *assumes*:
464 			 * o The first vaddr is the dest.
465 			 * o If multiple pages, then vaddr is sequential.
466 			 */
467 			wss_insert(sge->vaddr);
468 			if (length >= (2 * PAGE_SIZE))
469 				wss_insert(sge->vaddr + PAGE_SIZE);
470 
471 			cacheless_copy = wss_exceeds_threshold();
472 		} else {
473 			wss_advance_clean_counter();
474 		}
475 	}
476 	if (copy_last) {
477 		if (length > 8) {
478 			length -= 8;
479 		} else {
480 			copy_last = false;
481 			in_last = true;
482 		}
483 	}
484 
485 again:
486 	while (length) {
487 		u32 len = rvt_get_sge_length(sge, length);
488 
489 		WARN_ON_ONCE(len == 0);
490 		if (unlikely(in_last)) {
491 			/* enforce byte transfer ordering */
492 			for (i = 0; i < len; i++)
493 				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
494 		} else if (cacheless_copy) {
495 			cacheless_memcpy(sge->vaddr, data, len);
496 		} else {
497 			memcpy(sge->vaddr, data, len);
498 		}
499 		rvt_update_sge(ss, len, release);
500 		data += len;
501 		length -= len;
502 	}
503 
504 	if (copy_last) {
505 		copy_last = false;
506 		in_last = true;
507 		length = 8;
508 		goto again;
509 	}
510 }
511 
512 /*
513  * Make sure the QP is ready and able to accept the given opcode.
514  */
515 static inline opcode_handler qp_ok(struct hfi1_packet *packet)
516 {
517 	if (!(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
518 		return NULL;
519 	if (((packet->opcode & RVT_OPCODE_QP_MASK) ==
520 	     packet->qp->allowed_ops) ||
521 	    (packet->opcode == IB_OPCODE_CNP))
522 		return opcode_handler_tbl[packet->opcode];
523 
524 	return NULL;
525 }
526 
527 static u64 hfi1_fault_tx(struct rvt_qp *qp, u8 opcode, u64 pbc)
528 {
529 #ifdef CONFIG_FAULT_INJECTION
530 	if ((opcode & IB_OPCODE_MSP) == IB_OPCODE_MSP)
531 		/*
532 		 * In order to drop non-IB traffic we
533 		 * set PbcInsertHrc to NONE (0x2).
534 		 * The packet will still be delivered
535 		 * to the receiving node but a
536 		 * KHdrHCRCErr (KDETH packet with a bad
537 		 * HCRC) will be triggered and the
538 		 * packet will not be delivered to the
539 		 * correct context.
540 		 */
541 		pbc |= (u64)PBC_IHCRC_NONE << PBC_INSERT_HCRC_SHIFT;
542 	else
543 		/*
544 		 * In order to drop regular verbs
545 		 * traffic we set the PbcTestEbp
546 		 * flag. The packet will still be
547 		 * delivered to the receiving node but
548 		 * a 'late ebp error' will be
549 		 * triggered and will be dropped.
550 		 */
551 		pbc |= PBC_TEST_EBP;
552 #endif
553 	return pbc;
554 }
555 
556 static int hfi1_do_pkey_check(struct hfi1_packet *packet)
557 {
558 	struct hfi1_ctxtdata *rcd = packet->rcd;
559 	struct hfi1_pportdata *ppd = rcd->ppd;
560 	struct hfi1_16b_header *hdr = packet->hdr;
561 	u16 pkey;
562 
563 	/* Pkey check needed only for bypass packets */
564 	if (packet->etype != RHF_RCV_TYPE_BYPASS)
565 		return 0;
566 
567 	/* Perform pkey check */
568 	pkey = hfi1_16B_get_pkey(hdr);
569 	return ingress_pkey_check(ppd, pkey, packet->sc,
570 				  packet->qp->s_pkey_index,
571 				  packet->slid, true);
572 }
573 
574 static inline void hfi1_handle_packet(struct hfi1_packet *packet,
575 				      bool is_mcast)
576 {
577 	u32 qp_num;
578 	struct hfi1_ctxtdata *rcd = packet->rcd;
579 	struct hfi1_pportdata *ppd = rcd->ppd;
580 	struct hfi1_ibport *ibp = rcd_to_iport(rcd);
581 	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
582 	opcode_handler packet_handler;
583 	unsigned long flags;
584 
585 	inc_opstats(packet->tlen, &rcd->opstats->stats[packet->opcode]);
586 
587 	if (unlikely(is_mcast)) {
588 		struct rvt_mcast *mcast;
589 		struct rvt_mcast_qp *p;
590 
591 		if (!packet->grh)
592 			goto drop;
593 		mcast = rvt_mcast_find(&ibp->rvp,
594 				       &packet->grh->dgid,
595 				       opa_get_lid(packet->dlid, 9B));
596 		if (!mcast)
597 			goto drop;
598 		list_for_each_entry_rcu(p, &mcast->qp_list, list) {
599 			packet->qp = p->qp;
600 			if (hfi1_do_pkey_check(packet))
601 				goto drop;
602 			spin_lock_irqsave(&packet->qp->r_lock, flags);
603 			packet_handler = qp_ok(packet);
604 			if (likely(packet_handler))
605 				packet_handler(packet);
606 			else
607 				ibp->rvp.n_pkt_drops++;
608 			spin_unlock_irqrestore(&packet->qp->r_lock, flags);
609 		}
610 		/*
611 		 * Notify rvt_multicast_detach() if it is waiting for us
612 		 * to finish.
613 		 */
614 		if (atomic_dec_return(&mcast->refcount) <= 1)
615 			wake_up(&mcast->wait);
616 	} else {
617 		/* Get the destination QP number. */
618 		qp_num = ib_bth_get_qpn(packet->ohdr);
619 		rcu_read_lock();
620 		packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
621 		if (!packet->qp)
622 			goto unlock_drop;
623 
624 		if (hfi1_do_pkey_check(packet))
625 			goto unlock_drop;
626 
627 		if (unlikely(hfi1_dbg_fault_opcode(packet->qp, packet->opcode,
628 						   true)))
629 			goto unlock_drop;
630 
631 		spin_lock_irqsave(&packet->qp->r_lock, flags);
632 		packet_handler = qp_ok(packet);
633 		if (likely(packet_handler))
634 			packet_handler(packet);
635 		else
636 			ibp->rvp.n_pkt_drops++;
637 		spin_unlock_irqrestore(&packet->qp->r_lock, flags);
638 		rcu_read_unlock();
639 	}
640 	return;
641 unlock_drop:
642 	rcu_read_unlock();
643 drop:
644 	ibp->rvp.n_pkt_drops++;
645 }
646 
647 /**
648  * hfi1_ib_rcv - process an incoming packet
649  * @packet: data packet information
650  *
651  * This is called to process an incoming packet at interrupt level.
652  */
653 void hfi1_ib_rcv(struct hfi1_packet *packet)
654 {
655 	struct hfi1_ctxtdata *rcd = packet->rcd;
656 
657 	trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));
658 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
659 }
660 
661 void hfi1_16B_rcv(struct hfi1_packet *packet)
662 {
663 	struct hfi1_ctxtdata *rcd = packet->rcd;
664 
665 	trace_input_ibhdr(rcd->dd, packet, false);
666 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
667 }
668 
669 /*
670  * This is called from a timer to check for QPs
671  * which need kernel memory in order to send a packet.
672  */
673 static void mem_timer(struct timer_list *t)
674 {
675 	struct hfi1_ibdev *dev = from_timer(dev, t, mem_timer);
676 	struct list_head *list = &dev->memwait;
677 	struct rvt_qp *qp = NULL;
678 	struct iowait *wait;
679 	unsigned long flags;
680 	struct hfi1_qp_priv *priv;
681 
682 	write_seqlock_irqsave(&dev->iowait_lock, flags);
683 	if (!list_empty(list)) {
684 		wait = list_first_entry(list, struct iowait, list);
685 		qp = iowait_to_qp(wait);
686 		priv = qp->priv;
687 		list_del_init(&priv->s_iowait.list);
688 		priv->s_iowait.lock = NULL;
689 		/* refcount held until actual wake up */
690 		if (!list_empty(list))
691 			mod_timer(&dev->mem_timer, jiffies + 1);
692 	}
693 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
694 
695 	if (qp)
696 		hfi1_qp_wakeup(qp, RVT_S_WAIT_KMEM);
697 }
698 
699 /*
700  * This is called with progress side lock held.
701  */
702 /* New API */
703 static void verbs_sdma_complete(
704 	struct sdma_txreq *cookie,
705 	int status)
706 {
707 	struct verbs_txreq *tx =
708 		container_of(cookie, struct verbs_txreq, txreq);
709 	struct rvt_qp *qp = tx->qp;
710 
711 	spin_lock(&qp->s_lock);
712 	if (tx->wqe) {
713 		hfi1_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
714 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
715 		struct hfi1_opa_header *hdr;
716 
717 		hdr = &tx->phdr.hdr;
718 		hfi1_rc_send_complete(qp, hdr);
719 	}
720 	spin_unlock(&qp->s_lock);
721 
722 	hfi1_put_txreq(tx);
723 }
724 
725 static int wait_kmem(struct hfi1_ibdev *dev,
726 		     struct rvt_qp *qp,
727 		     struct hfi1_pkt_state *ps)
728 {
729 	struct hfi1_qp_priv *priv = qp->priv;
730 	unsigned long flags;
731 	int ret = 0;
732 
733 	spin_lock_irqsave(&qp->s_lock, flags);
734 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
735 		write_seqlock(&dev->iowait_lock);
736 		list_add_tail(&ps->s_txreq->txreq.list,
737 			      &priv->s_iowait.tx_head);
738 		if (list_empty(&priv->s_iowait.list)) {
739 			if (list_empty(&dev->memwait))
740 				mod_timer(&dev->mem_timer, jiffies + 1);
741 			qp->s_flags |= RVT_S_WAIT_KMEM;
742 			list_add_tail(&priv->s_iowait.list, &dev->memwait);
743 			priv->s_iowait.lock = &dev->iowait_lock;
744 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM);
745 			rvt_get_qp(qp);
746 		}
747 		write_sequnlock(&dev->iowait_lock);
748 		qp->s_flags &= ~RVT_S_BUSY;
749 		ret = -EBUSY;
750 	}
751 	spin_unlock_irqrestore(&qp->s_lock, flags);
752 
753 	return ret;
754 }
755 
756 /*
757  * This routine calls txadds for each sg entry.
758  *
759  * Add failures will revert the sge cursor
760  */
761 static noinline int build_verbs_ulp_payload(
762 	struct sdma_engine *sde,
763 	u32 length,
764 	struct verbs_txreq *tx)
765 {
766 	struct rvt_sge_state *ss = tx->ss;
767 	struct rvt_sge *sg_list = ss->sg_list;
768 	struct rvt_sge sge = ss->sge;
769 	u8 num_sge = ss->num_sge;
770 	u32 len;
771 	int ret = 0;
772 
773 	while (length) {
774 		len = ss->sge.length;
775 		if (len > length)
776 			len = length;
777 		if (len > ss->sge.sge_length)
778 			len = ss->sge.sge_length;
779 		WARN_ON_ONCE(len == 0);
780 		ret = sdma_txadd_kvaddr(
781 			sde->dd,
782 			&tx->txreq,
783 			ss->sge.vaddr,
784 			len);
785 		if (ret)
786 			goto bail_txadd;
787 		rvt_update_sge(ss, len, false);
788 		length -= len;
789 	}
790 	return ret;
791 bail_txadd:
792 	/* unwind cursor */
793 	ss->sge = sge;
794 	ss->num_sge = num_sge;
795 	ss->sg_list = sg_list;
796 	return ret;
797 }
798 
799 /**
800  * update_tx_opstats - record stats by opcode
801  * @qp; the qp
802  * @ps: transmit packet state
803  * @plen: the plen in dwords
804  *
805  * This is a routine to record the tx opstats after a
806  * packet has been presented to the egress mechanism.
807  */
808 static void update_tx_opstats(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
809 			      u32 plen)
810 {
811 #ifdef CONFIG_DEBUG_FS
812 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
813 	struct hfi1_opcode_stats_perctx *s = get_cpu_ptr(dd->tx_opstats);
814 
815 	inc_opstats(plen * 4, &s->stats[ps->opcode]);
816 	put_cpu_ptr(s);
817 #endif
818 }
819 
820 /*
821  * Build the number of DMA descriptors needed to send length bytes of data.
822  *
823  * NOTE: DMA mapping is held in the tx until completed in the ring or
824  *       the tx desc is freed without having been submitted to the ring
825  *
826  * This routine ensures all the helper routine calls succeed.
827  */
828 /* New API */
829 static int build_verbs_tx_desc(
830 	struct sdma_engine *sde,
831 	u32 length,
832 	struct verbs_txreq *tx,
833 	struct hfi1_ahg_info *ahg_info,
834 	u64 pbc)
835 {
836 	int ret = 0;
837 	struct hfi1_sdma_header *phdr = &tx->phdr;
838 	u16 hdrbytes = (tx->hdr_dwords + sizeof(pbc) / 4) << 2;
839 	u8 extra_bytes = 0;
840 
841 	if (tx->phdr.hdr.hdr_type) {
842 		/*
843 		 * hdrbytes accounts for PBC. Need to subtract 8 bytes
844 		 * before calculating padding.
845 		 */
846 		extra_bytes = hfi1_get_16b_padding(hdrbytes - 8, length) +
847 			      (SIZE_OF_CRC << 2) + SIZE_OF_LT;
848 	}
849 	if (!ahg_info->ahgcount) {
850 		ret = sdma_txinit_ahg(
851 			&tx->txreq,
852 			ahg_info->tx_flags,
853 			hdrbytes + length +
854 			extra_bytes,
855 			ahg_info->ahgidx,
856 			0,
857 			NULL,
858 			0,
859 			verbs_sdma_complete);
860 		if (ret)
861 			goto bail_txadd;
862 		phdr->pbc = cpu_to_le64(pbc);
863 		ret = sdma_txadd_kvaddr(
864 			sde->dd,
865 			&tx->txreq,
866 			phdr,
867 			hdrbytes);
868 		if (ret)
869 			goto bail_txadd;
870 	} else {
871 		ret = sdma_txinit_ahg(
872 			&tx->txreq,
873 			ahg_info->tx_flags,
874 			length,
875 			ahg_info->ahgidx,
876 			ahg_info->ahgcount,
877 			ahg_info->ahgdesc,
878 			hdrbytes,
879 			verbs_sdma_complete);
880 		if (ret)
881 			goto bail_txadd;
882 	}
883 	/* add the ulp payload - if any. tx->ss can be NULL for acks */
884 	if (tx->ss) {
885 		ret = build_verbs_ulp_payload(sde, length, tx);
886 		if (ret)
887 			goto bail_txadd;
888 	}
889 
890 	/* add icrc, lt byte, and padding to flit */
891 	if (extra_bytes)
892 		ret = sdma_txadd_kvaddr(sde->dd, &tx->txreq,
893 					(void *)trail_buf, extra_bytes);
894 
895 bail_txadd:
896 	return ret;
897 }
898 
899 int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
900 			u64 pbc)
901 {
902 	struct hfi1_qp_priv *priv = qp->priv;
903 	struct hfi1_ahg_info *ahg_info = priv->s_ahg;
904 	u32 hdrwords = ps->s_txreq->hdr_dwords;
905 	u32 len = ps->s_txreq->s_cur_size;
906 	u32 plen;
907 	struct hfi1_ibdev *dev = ps->dev;
908 	struct hfi1_pportdata *ppd = ps->ppd;
909 	struct verbs_txreq *tx;
910 	u8 sc5 = priv->s_sc;
911 	int ret;
912 	u32 dwords;
913 
914 	if (ps->s_txreq->phdr.hdr.hdr_type) {
915 		u8 extra_bytes = hfi1_get_16b_padding((hdrwords << 2), len);
916 
917 		dwords = (len + extra_bytes + (SIZE_OF_CRC << 2) +
918 			  SIZE_OF_LT) >> 2;
919 	} else {
920 		dwords = (len + 3) >> 2;
921 	}
922 	plen = hdrwords + dwords + sizeof(pbc) / 4;
923 
924 	tx = ps->s_txreq;
925 	if (!sdma_txreq_built(&tx->txreq)) {
926 		if (likely(pbc == 0)) {
927 			u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
928 
929 			/* No vl15 here */
930 			/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
931 			if (ps->s_txreq->phdr.hdr.hdr_type)
932 				pbc |= PBC_PACKET_BYPASS |
933 				       PBC_INSERT_BYPASS_ICRC;
934 			else
935 				pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
936 
937 			if (unlikely(hfi1_dbg_fault_opcode(qp, ps->opcode,
938 							   false)))
939 				pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
940 			pbc = create_pbc(ppd,
941 					 pbc,
942 					 qp->srate_mbps,
943 					 vl,
944 					 plen);
945 		}
946 		tx->wqe = qp->s_wqe;
947 		ret = build_verbs_tx_desc(tx->sde, len, tx, ahg_info, pbc);
948 		if (unlikely(ret))
949 			goto bail_build;
950 	}
951 	ret =  sdma_send_txreq(tx->sde, &priv->s_iowait, &tx->txreq,
952 			       ps->pkts_sent);
953 	if (unlikely(ret < 0)) {
954 		if (ret == -ECOMM)
955 			goto bail_ecomm;
956 		return ret;
957 	}
958 
959 	update_tx_opstats(qp, ps, plen);
960 	trace_sdma_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
961 				&ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
962 	return ret;
963 
964 bail_ecomm:
965 	/* The current one got "sent" */
966 	return 0;
967 bail_build:
968 	ret = wait_kmem(dev, qp, ps);
969 	if (!ret) {
970 		/* free txreq - bad state */
971 		hfi1_put_txreq(ps->s_txreq);
972 		ps->s_txreq = NULL;
973 	}
974 	return ret;
975 }
976 
977 /*
978  * If we are now in the error state, return zero to flush the
979  * send work request.
980  */
981 static int pio_wait(struct rvt_qp *qp,
982 		    struct send_context *sc,
983 		    struct hfi1_pkt_state *ps,
984 		    u32 flag)
985 {
986 	struct hfi1_qp_priv *priv = qp->priv;
987 	struct hfi1_devdata *dd = sc->dd;
988 	struct hfi1_ibdev *dev = &dd->verbs_dev;
989 	unsigned long flags;
990 	int ret = 0;
991 
992 	/*
993 	 * Note that as soon as want_buffer() is called and
994 	 * possibly before it returns, sc_piobufavail()
995 	 * could be called. Therefore, put QP on the I/O wait list before
996 	 * enabling the PIO avail interrupt.
997 	 */
998 	spin_lock_irqsave(&qp->s_lock, flags);
999 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
1000 		write_seqlock(&dev->iowait_lock);
1001 		list_add_tail(&ps->s_txreq->txreq.list,
1002 			      &priv->s_iowait.tx_head);
1003 		if (list_empty(&priv->s_iowait.list)) {
1004 			struct hfi1_ibdev *dev = &dd->verbs_dev;
1005 			int was_empty;
1006 
1007 			dev->n_piowait += !!(flag & RVT_S_WAIT_PIO);
1008 			dev->n_piodrain += !!(flag & RVT_S_WAIT_PIO_DRAIN);
1009 			qp->s_flags |= flag;
1010 			was_empty = list_empty(&sc->piowait);
1011 			iowait_queue(ps->pkts_sent, &priv->s_iowait,
1012 				     &sc->piowait);
1013 			priv->s_iowait.lock = &dev->iowait_lock;
1014 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_PIO);
1015 			rvt_get_qp(qp);
1016 			/* counting: only call wantpiobuf_intr if first user */
1017 			if (was_empty)
1018 				hfi1_sc_wantpiobuf_intr(sc, 1);
1019 		}
1020 		write_sequnlock(&dev->iowait_lock);
1021 		qp->s_flags &= ~RVT_S_BUSY;
1022 		ret = -EBUSY;
1023 	}
1024 	spin_unlock_irqrestore(&qp->s_lock, flags);
1025 	return ret;
1026 }
1027 
1028 static void verbs_pio_complete(void *arg, int code)
1029 {
1030 	struct rvt_qp *qp = (struct rvt_qp *)arg;
1031 	struct hfi1_qp_priv *priv = qp->priv;
1032 
1033 	if (iowait_pio_dec(&priv->s_iowait))
1034 		iowait_drain_wakeup(&priv->s_iowait);
1035 }
1036 
1037 int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
1038 			u64 pbc)
1039 {
1040 	struct hfi1_qp_priv *priv = qp->priv;
1041 	u32 hdrwords = ps->s_txreq->hdr_dwords;
1042 	struct rvt_sge_state *ss = ps->s_txreq->ss;
1043 	u32 len = ps->s_txreq->s_cur_size;
1044 	u32 dwords;
1045 	u32 plen;
1046 	struct hfi1_pportdata *ppd = ps->ppd;
1047 	u32 *hdr;
1048 	u8 sc5;
1049 	unsigned long flags = 0;
1050 	struct send_context *sc;
1051 	struct pio_buf *pbuf;
1052 	int wc_status = IB_WC_SUCCESS;
1053 	int ret = 0;
1054 	pio_release_cb cb = NULL;
1055 	u8 extra_bytes = 0;
1056 
1057 	if (ps->s_txreq->phdr.hdr.hdr_type) {
1058 		u8 pad_size = hfi1_get_16b_padding((hdrwords << 2), len);
1059 
1060 		extra_bytes = pad_size + (SIZE_OF_CRC << 2) + SIZE_OF_LT;
1061 		dwords = (len + extra_bytes) >> 2;
1062 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.opah;
1063 	} else {
1064 		dwords = (len + 3) >> 2;
1065 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.ibh;
1066 	}
1067 	plen = hdrwords + dwords + sizeof(pbc) / 4;
1068 
1069 	/* only RC/UC use complete */
1070 	switch (qp->ibqp.qp_type) {
1071 	case IB_QPT_RC:
1072 	case IB_QPT_UC:
1073 		cb = verbs_pio_complete;
1074 		break;
1075 	default:
1076 		break;
1077 	}
1078 
1079 	/* vl15 special case taken care of in ud.c */
1080 	sc5 = priv->s_sc;
1081 	sc = ps->s_txreq->psc;
1082 
1083 	if (likely(pbc == 0)) {
1084 		u8 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
1085 
1086 		/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
1087 		if (ps->s_txreq->phdr.hdr.hdr_type)
1088 			pbc |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
1089 		else
1090 			pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
1091 		if (unlikely(hfi1_dbg_fault_opcode(qp, ps->opcode, false)))
1092 			pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
1093 		pbc = create_pbc(ppd, pbc, qp->srate_mbps, vl, plen);
1094 	}
1095 	if (cb)
1096 		iowait_pio_inc(&priv->s_iowait);
1097 	pbuf = sc_buffer_alloc(sc, plen, cb, qp);
1098 	if (unlikely(!pbuf)) {
1099 		if (cb)
1100 			verbs_pio_complete(qp, 0);
1101 		if (ppd->host_link_state != HLS_UP_ACTIVE) {
1102 			/*
1103 			 * If we have filled the PIO buffers to capacity and are
1104 			 * not in an active state this request is not going to
1105 			 * go out to so just complete it with an error or else a
1106 			 * ULP or the core may be stuck waiting.
1107 			 */
1108 			hfi1_cdbg(
1109 				PIO,
1110 				"alloc failed. state not active, completing");
1111 			wc_status = IB_WC_GENERAL_ERR;
1112 			goto pio_bail;
1113 		} else {
1114 			/*
1115 			 * This is a normal occurrence. The PIO buffs are full
1116 			 * up but we are still happily sending, well we could be
1117 			 * so lets continue to queue the request.
1118 			 */
1119 			hfi1_cdbg(PIO, "alloc failed. state active, queuing");
1120 			ret = pio_wait(qp, sc, ps, RVT_S_WAIT_PIO);
1121 			if (!ret)
1122 				/* txreq not queued - free */
1123 				goto bail;
1124 			/* tx consumed in wait */
1125 			return ret;
1126 		}
1127 	}
1128 
1129 	if (dwords == 0) {
1130 		pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
1131 	} else {
1132 		seg_pio_copy_start(pbuf, pbc,
1133 				   hdr, hdrwords * 4);
1134 		if (ss) {
1135 			while (len) {
1136 				void *addr = ss->sge.vaddr;
1137 				u32 slen = ss->sge.length;
1138 
1139 				if (slen > len)
1140 					slen = len;
1141 				rvt_update_sge(ss, slen, false);
1142 				seg_pio_copy_mid(pbuf, addr, slen);
1143 				len -= slen;
1144 			}
1145 		}
1146 		/* add icrc, lt byte, and padding to flit */
1147 		if (extra_bytes)
1148 			seg_pio_copy_mid(pbuf, trail_buf, extra_bytes);
1149 
1150 		seg_pio_copy_end(pbuf);
1151 	}
1152 
1153 	update_tx_opstats(qp, ps, plen);
1154 	trace_pio_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
1155 			       &ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
1156 
1157 pio_bail:
1158 	if (qp->s_wqe) {
1159 		spin_lock_irqsave(&qp->s_lock, flags);
1160 		hfi1_send_complete(qp, qp->s_wqe, wc_status);
1161 		spin_unlock_irqrestore(&qp->s_lock, flags);
1162 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1163 		spin_lock_irqsave(&qp->s_lock, flags);
1164 		hfi1_rc_send_complete(qp, &ps->s_txreq->phdr.hdr);
1165 		spin_unlock_irqrestore(&qp->s_lock, flags);
1166 	}
1167 
1168 	ret = 0;
1169 
1170 bail:
1171 	hfi1_put_txreq(ps->s_txreq);
1172 	return ret;
1173 }
1174 
1175 /*
1176  * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1177  * being an entry from the partition key table), return 0
1178  * otherwise. Use the matching criteria for egress partition keys
1179  * specified in the OPAv1 spec., section 9.1l.7.
1180  */
1181 static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
1182 {
1183 	u16 mkey = pkey & PKEY_LOW_15_MASK;
1184 	u16 mentry = ent & PKEY_LOW_15_MASK;
1185 
1186 	if (mkey == mentry) {
1187 		/*
1188 		 * If pkey[15] is set (full partition member),
1189 		 * is bit 15 in the corresponding table element
1190 		 * clear (limited member)?
1191 		 */
1192 		if (pkey & PKEY_MEMBER_MASK)
1193 			return !!(ent & PKEY_MEMBER_MASK);
1194 		return 1;
1195 	}
1196 	return 0;
1197 }
1198 
1199 /**
1200  * egress_pkey_check - check P_KEY of a packet
1201  * @ppd:  Physical IB port data
1202  * @slid: SLID for packet
1203  * @bkey: PKEY for header
1204  * @sc5:  SC for packet
1205  * @s_pkey_index: It will be used for look up optimization for kernel contexts
1206  * only. If it is negative value, then it means user contexts is calling this
1207  * function.
1208  *
1209  * It checks if hdr's pkey is valid.
1210  *
1211  * Return: 0 on success, otherwise, 1
1212  */
1213 int egress_pkey_check(struct hfi1_pportdata *ppd, u32 slid, u16 pkey,
1214 		      u8 sc5, int8_t s_pkey_index)
1215 {
1216 	struct hfi1_devdata *dd;
1217 	int i;
1218 	int is_user_ctxt_mechanism = (s_pkey_index < 0);
1219 
1220 	if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
1221 		return 0;
1222 
1223 	/* If SC15, pkey[0:14] must be 0x7fff */
1224 	if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
1225 		goto bad;
1226 
1227 	/* Is the pkey = 0x0, or 0x8000? */
1228 	if ((pkey & PKEY_LOW_15_MASK) == 0)
1229 		goto bad;
1230 
1231 	/*
1232 	 * For the kernel contexts only, if a qp is passed into the function,
1233 	 * the most likely matching pkey has index qp->s_pkey_index
1234 	 */
1235 	if (!is_user_ctxt_mechanism &&
1236 	    egress_pkey_matches_entry(pkey, ppd->pkeys[s_pkey_index])) {
1237 		return 0;
1238 	}
1239 
1240 	for (i = 0; i < MAX_PKEY_VALUES; i++) {
1241 		if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
1242 			return 0;
1243 	}
1244 bad:
1245 	/*
1246 	 * For the user-context mechanism, the P_KEY check would only happen
1247 	 * once per SDMA request, not once per packet.  Therefore, there's no
1248 	 * need to increment the counter for the user-context mechanism.
1249 	 */
1250 	if (!is_user_ctxt_mechanism) {
1251 		incr_cntr64(&ppd->port_xmit_constraint_errors);
1252 		dd = ppd->dd;
1253 		if (!(dd->err_info_xmit_constraint.status &
1254 		      OPA_EI_STATUS_SMASK)) {
1255 			dd->err_info_xmit_constraint.status |=
1256 				OPA_EI_STATUS_SMASK;
1257 			dd->err_info_xmit_constraint.slid = slid;
1258 			dd->err_info_xmit_constraint.pkey = pkey;
1259 		}
1260 	}
1261 	return 1;
1262 }
1263 
1264 /**
1265  * get_send_routine - choose an egress routine
1266  *
1267  * Choose an egress routine based on QP type
1268  * and size
1269  */
1270 static inline send_routine get_send_routine(struct rvt_qp *qp,
1271 					    struct hfi1_pkt_state *ps)
1272 {
1273 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1274 	struct hfi1_qp_priv *priv = qp->priv;
1275 	struct verbs_txreq *tx = ps->s_txreq;
1276 
1277 	if (unlikely(!(dd->flags & HFI1_HAS_SEND_DMA)))
1278 		return dd->process_pio_send;
1279 	switch (qp->ibqp.qp_type) {
1280 	case IB_QPT_SMI:
1281 		return dd->process_pio_send;
1282 	case IB_QPT_GSI:
1283 	case IB_QPT_UD:
1284 		break;
1285 	case IB_QPT_UC:
1286 	case IB_QPT_RC: {
1287 		if (piothreshold &&
1288 		    tx->s_cur_size <= min(piothreshold, qp->pmtu) &&
1289 		    (BIT(ps->opcode & OPMASK) & pio_opmask[ps->opcode >> 5]) &&
1290 		    iowait_sdma_pending(&priv->s_iowait) == 0 &&
1291 		    !sdma_txreq_built(&tx->txreq))
1292 			return dd->process_pio_send;
1293 		break;
1294 	}
1295 	default:
1296 		break;
1297 	}
1298 	return dd->process_dma_send;
1299 }
1300 
1301 /**
1302  * hfi1_verbs_send - send a packet
1303  * @qp: the QP to send on
1304  * @ps: the state of the packet to send
1305  *
1306  * Return zero if packet is sent or queued OK.
1307  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1308  */
1309 int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
1310 {
1311 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1312 	struct hfi1_qp_priv *priv = qp->priv;
1313 	struct ib_other_headers *ohdr;
1314 	send_routine sr;
1315 	int ret;
1316 	u16 pkey;
1317 	u32 slid;
1318 
1319 	/* locate the pkey within the headers */
1320 	if (ps->s_txreq->phdr.hdr.hdr_type) {
1321 		struct hfi1_16b_header *hdr = &ps->s_txreq->phdr.hdr.opah;
1322 		u8 l4 = hfi1_16B_get_l4(hdr);
1323 
1324 		if (l4 == OPA_16B_L4_IB_GLOBAL)
1325 			ohdr = &hdr->u.l.oth;
1326 		else
1327 			ohdr = &hdr->u.oth;
1328 		slid = hfi1_16B_get_slid(hdr);
1329 		pkey = hfi1_16B_get_pkey(hdr);
1330 	} else {
1331 		struct ib_header *hdr = &ps->s_txreq->phdr.hdr.ibh;
1332 		u8 lnh = ib_get_lnh(hdr);
1333 
1334 		if (lnh == HFI1_LRH_GRH)
1335 			ohdr = &hdr->u.l.oth;
1336 		else
1337 			ohdr = &hdr->u.oth;
1338 		slid = ib_get_slid(hdr);
1339 		pkey = ib_bth_get_pkey(ohdr);
1340 	}
1341 
1342 	ps->opcode = ib_bth_get_opcode(ohdr);
1343 	sr = get_send_routine(qp, ps);
1344 	ret = egress_pkey_check(dd->pport, slid, pkey,
1345 				priv->s_sc, qp->s_pkey_index);
1346 	if (unlikely(ret)) {
1347 		/*
1348 		 * The value we are returning here does not get propagated to
1349 		 * the verbs caller. Thus we need to complete the request with
1350 		 * error otherwise the caller could be sitting waiting on the
1351 		 * completion event. Only do this for PIO. SDMA has its own
1352 		 * mechanism for handling the errors. So for SDMA we can just
1353 		 * return.
1354 		 */
1355 		if (sr == dd->process_pio_send) {
1356 			unsigned long flags;
1357 
1358 			hfi1_cdbg(PIO, "%s() Failed. Completing with err",
1359 				  __func__);
1360 			spin_lock_irqsave(&qp->s_lock, flags);
1361 			hfi1_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
1362 			spin_unlock_irqrestore(&qp->s_lock, flags);
1363 		}
1364 		return -EINVAL;
1365 	}
1366 	if (sr == dd->process_dma_send && iowait_pio_pending(&priv->s_iowait))
1367 		return pio_wait(qp,
1368 				ps->s_txreq->psc,
1369 				ps,
1370 				RVT_S_WAIT_PIO_DRAIN);
1371 	return sr(qp, ps, 0);
1372 }
1373 
1374 /**
1375  * hfi1_fill_device_attr - Fill in rvt dev info device attributes.
1376  * @dd: the device data structure
1377  */
1378 static void hfi1_fill_device_attr(struct hfi1_devdata *dd)
1379 {
1380 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1381 	u32 ver = dd->dc8051_ver;
1382 
1383 	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1384 
1385 	rdi->dparms.props.fw_ver = ((u64)(dc8051_ver_maj(ver)) << 32) |
1386 		((u64)(dc8051_ver_min(ver)) << 16) |
1387 		(u64)dc8051_ver_patch(ver);
1388 
1389 	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1390 			IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1391 			IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1392 			IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE |
1393 			IB_DEVICE_MEM_MGT_EXTENSIONS |
1394 			IB_DEVICE_RDMA_NETDEV_OPA_VNIC;
1395 	rdi->dparms.props.page_size_cap = PAGE_SIZE;
1396 	rdi->dparms.props.vendor_id = dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
1397 	rdi->dparms.props.vendor_part_id = dd->pcidev->device;
1398 	rdi->dparms.props.hw_ver = dd->minrev;
1399 	rdi->dparms.props.sys_image_guid = ib_hfi1_sys_image_guid;
1400 	rdi->dparms.props.max_mr_size = U64_MAX;
1401 	rdi->dparms.props.max_fast_reg_page_list_len = UINT_MAX;
1402 	rdi->dparms.props.max_qp = hfi1_max_qps;
1403 	rdi->dparms.props.max_qp_wr = hfi1_max_qp_wrs;
1404 	rdi->dparms.props.max_sge = hfi1_max_sges;
1405 	rdi->dparms.props.max_sge_rd = hfi1_max_sges;
1406 	rdi->dparms.props.max_cq = hfi1_max_cqs;
1407 	rdi->dparms.props.max_ah = hfi1_max_ahs;
1408 	rdi->dparms.props.max_cqe = hfi1_max_cqes;
1409 	rdi->dparms.props.max_mr = rdi->lkey_table.max;
1410 	rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1411 	rdi->dparms.props.max_map_per_fmr = 32767;
1412 	rdi->dparms.props.max_pd = hfi1_max_pds;
1413 	rdi->dparms.props.max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
1414 	rdi->dparms.props.max_qp_init_rd_atom = 255;
1415 	rdi->dparms.props.max_srq = hfi1_max_srqs;
1416 	rdi->dparms.props.max_srq_wr = hfi1_max_srq_wrs;
1417 	rdi->dparms.props.max_srq_sge = hfi1_max_srq_sges;
1418 	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1419 	rdi->dparms.props.max_pkeys = hfi1_get_npkeys(dd);
1420 	rdi->dparms.props.max_mcast_grp = hfi1_max_mcast_grps;
1421 	rdi->dparms.props.max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
1422 	rdi->dparms.props.max_total_mcast_qp_attach =
1423 					rdi->dparms.props.max_mcast_qp_attach *
1424 					rdi->dparms.props.max_mcast_grp;
1425 }
1426 
1427 static inline u16 opa_speed_to_ib(u16 in)
1428 {
1429 	u16 out = 0;
1430 
1431 	if (in & OPA_LINK_SPEED_25G)
1432 		out |= IB_SPEED_EDR;
1433 	if (in & OPA_LINK_SPEED_12_5G)
1434 		out |= IB_SPEED_FDR;
1435 
1436 	return out;
1437 }
1438 
1439 /*
1440  * Convert a single OPA link width (no multiple flags) to an IB value.
1441  * A zero OPA link width means link down, which means the IB width value
1442  * is a don't care.
1443  */
1444 static inline u16 opa_width_to_ib(u16 in)
1445 {
1446 	switch (in) {
1447 	case OPA_LINK_WIDTH_1X:
1448 	/* map 2x and 3x to 1x as they don't exist in IB */
1449 	case OPA_LINK_WIDTH_2X:
1450 	case OPA_LINK_WIDTH_3X:
1451 		return IB_WIDTH_1X;
1452 	default: /* link down or unknown, return our largest width */
1453 	case OPA_LINK_WIDTH_4X:
1454 		return IB_WIDTH_4X;
1455 	}
1456 }
1457 
1458 static int query_port(struct rvt_dev_info *rdi, u8 port_num,
1459 		      struct ib_port_attr *props)
1460 {
1461 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1462 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1463 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1464 	u32 lid = ppd->lid;
1465 
1466 	/* props being zeroed by the caller, avoid zeroing it here */
1467 	props->lid = lid ? lid : 0;
1468 	props->lmc = ppd->lmc;
1469 	/* OPA logical states match IB logical states */
1470 	props->state = driver_lstate(ppd);
1471 	props->phys_state = driver_pstate(ppd);
1472 	props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
1473 	props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
1474 	/* see rate_show() in ib core/sysfs.c */
1475 	props->active_speed = (u8)opa_speed_to_ib(ppd->link_speed_active);
1476 	props->max_vl_num = ppd->vls_supported;
1477 
1478 	/* Once we are a "first class" citizen and have added the OPA MTUs to
1479 	 * the core we can advertise the larger MTU enum to the ULPs, for now
1480 	 * advertise only 4K.
1481 	 *
1482 	 * Those applications which are either OPA aware or pass the MTU enum
1483 	 * from the Path Records to us will get the new 8k MTU.  Those that
1484 	 * attempt to process the MTU enum may fail in various ways.
1485 	 */
1486 	props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
1487 				      4096 : hfi1_max_mtu), IB_MTU_4096);
1488 	props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
1489 		mtu_to_enum(ppd->ibmtu, IB_MTU_4096);
1490 
1491 	/*
1492 	 * sm_lid of 0xFFFF needs special handling so that it can
1493 	 * be differentiated from a permissve LID of 0xFFFF.
1494 	 * We set the grh_required flag here so the SA can program
1495 	 * the DGID in the address handle appropriately
1496 	 */
1497 	if (props->sm_lid == be16_to_cpu(IB_LID_PERMISSIVE))
1498 		props->grh_required = true;
1499 
1500 	return 0;
1501 }
1502 
1503 static int modify_device(struct ib_device *device,
1504 			 int device_modify_mask,
1505 			 struct ib_device_modify *device_modify)
1506 {
1507 	struct hfi1_devdata *dd = dd_from_ibdev(device);
1508 	unsigned i;
1509 	int ret;
1510 
1511 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1512 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1513 		ret = -EOPNOTSUPP;
1514 		goto bail;
1515 	}
1516 
1517 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1518 		memcpy(device->node_desc, device_modify->node_desc,
1519 		       IB_DEVICE_NODE_DESC_MAX);
1520 		for (i = 0; i < dd->num_pports; i++) {
1521 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1522 
1523 			hfi1_node_desc_chg(ibp);
1524 		}
1525 	}
1526 
1527 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1528 		ib_hfi1_sys_image_guid =
1529 			cpu_to_be64(device_modify->sys_image_guid);
1530 		for (i = 0; i < dd->num_pports; i++) {
1531 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1532 
1533 			hfi1_sys_guid_chg(ibp);
1534 		}
1535 	}
1536 
1537 	ret = 0;
1538 
1539 bail:
1540 	return ret;
1541 }
1542 
1543 static int shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1544 {
1545 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1546 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1547 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1548 	int ret;
1549 
1550 	set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
1551 			     OPA_LINKDOWN_REASON_UNKNOWN);
1552 	ret = set_link_state(ppd, HLS_DN_DOWNDEF);
1553 	return ret;
1554 }
1555 
1556 static int hfi1_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1557 			    int guid_index, __be64 *guid)
1558 {
1559 	struct hfi1_ibport *ibp = container_of(rvp, struct hfi1_ibport, rvp);
1560 
1561 	if (guid_index >= HFI1_GUIDS_PER_PORT)
1562 		return -EINVAL;
1563 
1564 	*guid = get_sguid(ibp, guid_index);
1565 	return 0;
1566 }
1567 
1568 /*
1569  * convert ah port,sl to sc
1570  */
1571 u8 ah_to_sc(struct ib_device *ibdev, struct rdma_ah_attr *ah)
1572 {
1573 	struct hfi1_ibport *ibp = to_iport(ibdev, rdma_ah_get_port_num(ah));
1574 
1575 	return ibp->sl_to_sc[rdma_ah_get_sl(ah)];
1576 }
1577 
1578 static int hfi1_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1579 {
1580 	struct hfi1_ibport *ibp;
1581 	struct hfi1_pportdata *ppd;
1582 	struct hfi1_devdata *dd;
1583 	u8 sc5;
1584 
1585 	if (hfi1_check_mcast(rdma_ah_get_dlid(ah_attr)) &&
1586 	    !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1587 		return -EINVAL;
1588 
1589 	/* test the mapping for validity */
1590 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1591 	ppd = ppd_from_ibp(ibp);
1592 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
1593 	dd = dd_from_ppd(ppd);
1594 	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1595 		return -EINVAL;
1596 	return 0;
1597 }
1598 
1599 static void hfi1_notify_new_ah(struct ib_device *ibdev,
1600 			       struct rdma_ah_attr *ah_attr,
1601 			       struct rvt_ah *ah)
1602 {
1603 	struct hfi1_ibport *ibp;
1604 	struct hfi1_pportdata *ppd;
1605 	struct hfi1_devdata *dd;
1606 	u8 sc5;
1607 	struct rdma_ah_attr *attr = &ah->attr;
1608 
1609 	/*
1610 	 * Do not trust reading anything from rvt_ah at this point as it is not
1611 	 * done being setup. We can however modify things which we need to set.
1612 	 */
1613 
1614 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1615 	ppd = ppd_from_ibp(ibp);
1616 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)];
1617 	hfi1_update_ah_attr(ibdev, attr);
1618 	hfi1_make_opa_lid(attr);
1619 	dd = dd_from_ppd(ppd);
1620 	ah->vl = sc_to_vlt(dd, sc5);
1621 	if (ah->vl < num_vls || ah->vl == 15)
1622 		ah->log_pmtu = ilog2(dd->vld[ah->vl].mtu);
1623 }
1624 
1625 /**
1626  * hfi1_get_npkeys - return the size of the PKEY table for context 0
1627  * @dd: the hfi1_ib device
1628  */
1629 unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
1630 {
1631 	return ARRAY_SIZE(dd->pport[0].pkeys);
1632 }
1633 
1634 static void init_ibport(struct hfi1_pportdata *ppd)
1635 {
1636 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1637 	size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
1638 	int i;
1639 
1640 	for (i = 0; i < sz; i++) {
1641 		ibp->sl_to_sc[i] = i;
1642 		ibp->sc_to_sl[i] = i;
1643 	}
1644 
1645 	for (i = 0; i < RVT_MAX_TRAP_LISTS ; i++)
1646 		INIT_LIST_HEAD(&ibp->rvp.trap_lists[i].list);
1647 	timer_setup(&ibp->rvp.trap_timer, hfi1_handle_trap_timer, 0);
1648 
1649 	spin_lock_init(&ibp->rvp.lock);
1650 	/* Set the prefix to the default value (see ch. 4.1.1) */
1651 	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1652 	ibp->rvp.sm_lid = 0;
1653 	/*
1654 	 * Below should only set bits defined in OPA PortInfo.CapabilityMask
1655 	 * and PortInfo.CapabilityMask3
1656 	 */
1657 	ibp->rvp.port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
1658 		IB_PORT_CAP_MASK_NOTICE_SUP;
1659 	ibp->rvp.port_cap3_flags = OPA_CAP_MASK3_IsSharedSpaceSupported;
1660 	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1661 	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1662 	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1663 	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1664 	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1665 
1666 	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1667 	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1668 }
1669 
1670 static void hfi1_get_dev_fw_str(struct ib_device *ibdev, char *str)
1671 {
1672 	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
1673 	struct hfi1_ibdev *dev = dev_from_rdi(rdi);
1674 	u32 ver = dd_from_dev(dev)->dc8051_ver;
1675 
1676 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u.%u", dc8051_ver_maj(ver),
1677 		 dc8051_ver_min(ver), dc8051_ver_patch(ver));
1678 }
1679 
1680 static const char * const driver_cntr_names[] = {
1681 	/* must be element 0*/
1682 	"DRIVER_KernIntr",
1683 	"DRIVER_ErrorIntr",
1684 	"DRIVER_Tx_Errs",
1685 	"DRIVER_Rcv_Errs",
1686 	"DRIVER_HW_Errs",
1687 	"DRIVER_NoPIOBufs",
1688 	"DRIVER_CtxtsOpen",
1689 	"DRIVER_RcvLen_Errs",
1690 	"DRIVER_EgrBufFull",
1691 	"DRIVER_EgrHdrFull"
1692 };
1693 
1694 static DEFINE_MUTEX(cntr_names_lock); /* protects the *_cntr_names bufers */
1695 static const char **dev_cntr_names;
1696 static const char **port_cntr_names;
1697 static int num_driver_cntrs = ARRAY_SIZE(driver_cntr_names);
1698 static int num_dev_cntrs;
1699 static int num_port_cntrs;
1700 static int cntr_names_initialized;
1701 
1702 /*
1703  * Convert a list of names separated by '\n' into an array of NULL terminated
1704  * strings. Optionally some entries can be reserved in the array to hold extra
1705  * external strings.
1706  */
1707 static int init_cntr_names(const char *names_in,
1708 			   const size_t names_len,
1709 			   int num_extra_names,
1710 			   int *num_cntrs,
1711 			   const char ***cntr_names)
1712 {
1713 	char *names_out, *p, **q;
1714 	int i, n;
1715 
1716 	n = 0;
1717 	for (i = 0; i < names_len; i++)
1718 		if (names_in[i] == '\n')
1719 			n++;
1720 
1721 	names_out = kmalloc((n + num_extra_names) * sizeof(char *) + names_len,
1722 			    GFP_KERNEL);
1723 	if (!names_out) {
1724 		*num_cntrs = 0;
1725 		*cntr_names = NULL;
1726 		return -ENOMEM;
1727 	}
1728 
1729 	p = names_out + (n + num_extra_names) * sizeof(char *);
1730 	memcpy(p, names_in, names_len);
1731 
1732 	q = (char **)names_out;
1733 	for (i = 0; i < n; i++) {
1734 		q[i] = p;
1735 		p = strchr(p, '\n');
1736 		*p++ = '\0';
1737 	}
1738 
1739 	*num_cntrs = n;
1740 	*cntr_names = (const char **)names_out;
1741 	return 0;
1742 }
1743 
1744 static struct rdma_hw_stats *alloc_hw_stats(struct ib_device *ibdev,
1745 					    u8 port_num)
1746 {
1747 	int i, err;
1748 
1749 	mutex_lock(&cntr_names_lock);
1750 	if (!cntr_names_initialized) {
1751 		struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1752 
1753 		err = init_cntr_names(dd->cntrnames,
1754 				      dd->cntrnameslen,
1755 				      num_driver_cntrs,
1756 				      &num_dev_cntrs,
1757 				      &dev_cntr_names);
1758 		if (err) {
1759 			mutex_unlock(&cntr_names_lock);
1760 			return NULL;
1761 		}
1762 
1763 		for (i = 0; i < num_driver_cntrs; i++)
1764 			dev_cntr_names[num_dev_cntrs + i] =
1765 				driver_cntr_names[i];
1766 
1767 		err = init_cntr_names(dd->portcntrnames,
1768 				      dd->portcntrnameslen,
1769 				      0,
1770 				      &num_port_cntrs,
1771 				      &port_cntr_names);
1772 		if (err) {
1773 			kfree(dev_cntr_names);
1774 			dev_cntr_names = NULL;
1775 			mutex_unlock(&cntr_names_lock);
1776 			return NULL;
1777 		}
1778 		cntr_names_initialized = 1;
1779 	}
1780 	mutex_unlock(&cntr_names_lock);
1781 
1782 	if (!port_num)
1783 		return rdma_alloc_hw_stats_struct(
1784 				dev_cntr_names,
1785 				num_dev_cntrs + num_driver_cntrs,
1786 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1787 	else
1788 		return rdma_alloc_hw_stats_struct(
1789 				port_cntr_names,
1790 				num_port_cntrs,
1791 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1792 }
1793 
1794 static u64 hfi1_sps_ints(void)
1795 {
1796 	unsigned long flags;
1797 	struct hfi1_devdata *dd;
1798 	u64 sps_ints = 0;
1799 
1800 	spin_lock_irqsave(&hfi1_devs_lock, flags);
1801 	list_for_each_entry(dd, &hfi1_dev_list, list) {
1802 		sps_ints += get_all_cpu_total(dd->int_counter);
1803 	}
1804 	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1805 	return sps_ints;
1806 }
1807 
1808 static int get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1809 			u8 port, int index)
1810 {
1811 	u64 *values;
1812 	int count;
1813 
1814 	if (!port) {
1815 		u64 *stats = (u64 *)&hfi1_stats;
1816 		int i;
1817 
1818 		hfi1_read_cntrs(dd_from_ibdev(ibdev), NULL, &values);
1819 		values[num_dev_cntrs] = hfi1_sps_ints();
1820 		for (i = 1; i < num_driver_cntrs; i++)
1821 			values[num_dev_cntrs + i] = stats[i];
1822 		count = num_dev_cntrs + num_driver_cntrs;
1823 	} else {
1824 		struct hfi1_ibport *ibp = to_iport(ibdev, port);
1825 
1826 		hfi1_read_portcntrs(ppd_from_ibp(ibp), NULL, &values);
1827 		count = num_port_cntrs;
1828 	}
1829 
1830 	memcpy(stats->value, values, count * sizeof(u64));
1831 	return count;
1832 }
1833 
1834 /**
1835  * hfi1_register_ib_device - register our device with the infiniband core
1836  * @dd: the device data structure
1837  * Return 0 if successful, errno if unsuccessful.
1838  */
1839 int hfi1_register_ib_device(struct hfi1_devdata *dd)
1840 {
1841 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1842 	struct ib_device *ibdev = &dev->rdi.ibdev;
1843 	struct hfi1_pportdata *ppd = dd->pport;
1844 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1845 	unsigned i;
1846 	int ret;
1847 
1848 	for (i = 0; i < dd->num_pports; i++)
1849 		init_ibport(ppd + i);
1850 
1851 	/* Only need to initialize non-zero fields. */
1852 
1853 	timer_setup(&dev->mem_timer, mem_timer, 0);
1854 
1855 	seqlock_init(&dev->iowait_lock);
1856 	seqlock_init(&dev->txwait_lock);
1857 	INIT_LIST_HEAD(&dev->txwait);
1858 	INIT_LIST_HEAD(&dev->memwait);
1859 
1860 	ret = verbs_txreq_init(dev);
1861 	if (ret)
1862 		goto err_verbs_txreq;
1863 
1864 	/* Use first-port GUID as node guid */
1865 	ibdev->node_guid = get_sguid(ibp, HFI1_PORT_GUID_INDEX);
1866 
1867 	/*
1868 	 * The system image GUID is supposed to be the same for all
1869 	 * HFIs in a single system but since there can be other
1870 	 * device types in the system, we can't be sure this is unique.
1871 	 */
1872 	if (!ib_hfi1_sys_image_guid)
1873 		ib_hfi1_sys_image_guid = ibdev->node_guid;
1874 	ibdev->owner = THIS_MODULE;
1875 	ibdev->phys_port_cnt = dd->num_pports;
1876 	ibdev->dev.parent = &dd->pcidev->dev;
1877 	ibdev->modify_device = modify_device;
1878 	ibdev->alloc_hw_stats = alloc_hw_stats;
1879 	ibdev->get_hw_stats = get_hw_stats;
1880 	ibdev->alloc_rdma_netdev = hfi1_vnic_alloc_rn;
1881 
1882 	/* keep process mad in the driver */
1883 	ibdev->process_mad = hfi1_process_mad;
1884 	ibdev->get_dev_fw_str = hfi1_get_dev_fw_str;
1885 
1886 	strncpy(ibdev->node_desc, init_utsname()->nodename,
1887 		sizeof(ibdev->node_desc));
1888 
1889 	/*
1890 	 * Fill in rvt info object.
1891 	 */
1892 	dd->verbs_dev.rdi.driver_f.port_callback = hfi1_create_port_files;
1893 	dd->verbs_dev.rdi.driver_f.get_pci_dev = get_pci_dev;
1894 	dd->verbs_dev.rdi.driver_f.check_ah = hfi1_check_ah;
1895 	dd->verbs_dev.rdi.driver_f.notify_new_ah = hfi1_notify_new_ah;
1896 	dd->verbs_dev.rdi.driver_f.get_guid_be = hfi1_get_guid_be;
1897 	dd->verbs_dev.rdi.driver_f.query_port_state = query_port;
1898 	dd->verbs_dev.rdi.driver_f.shut_down_port = shut_down_port;
1899 	dd->verbs_dev.rdi.driver_f.cap_mask_chg = hfi1_cap_mask_chg;
1900 	/*
1901 	 * Fill in rvt info device attributes.
1902 	 */
1903 	hfi1_fill_device_attr(dd);
1904 
1905 	/* queue pair */
1906 	dd->verbs_dev.rdi.dparms.qp_table_size = hfi1_qp_table_size;
1907 	dd->verbs_dev.rdi.dparms.qpn_start = 0;
1908 	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1909 	dd->verbs_dev.rdi.dparms.qos_shift = dd->qos_shift;
1910 	dd->verbs_dev.rdi.dparms.qpn_res_start = kdeth_qp << 16;
1911 	dd->verbs_dev.rdi.dparms.qpn_res_end =
1912 	dd->verbs_dev.rdi.dparms.qpn_res_start + 65535;
1913 	dd->verbs_dev.rdi.dparms.max_rdma_atomic = HFI1_MAX_RDMA_ATOMIC;
1914 	dd->verbs_dev.rdi.dparms.psn_mask = PSN_MASK;
1915 	dd->verbs_dev.rdi.dparms.psn_shift = PSN_SHIFT;
1916 	dd->verbs_dev.rdi.dparms.psn_modify_mask = PSN_MODIFY_MASK;
1917 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_INTEL_OPA |
1918 						RDMA_CORE_CAP_OPA_AH;
1919 	dd->verbs_dev.rdi.dparms.max_mad_size = OPA_MGMT_MAD_SIZE;
1920 
1921 	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qp_priv_alloc;
1922 	dd->verbs_dev.rdi.driver_f.qp_priv_free = qp_priv_free;
1923 	dd->verbs_dev.rdi.driver_f.free_all_qps = free_all_qps;
1924 	dd->verbs_dev.rdi.driver_f.notify_qp_reset = notify_qp_reset;
1925 	dd->verbs_dev.rdi.driver_f.do_send = hfi1_do_send_from_rvt;
1926 	dd->verbs_dev.rdi.driver_f.schedule_send = hfi1_schedule_send;
1927 	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _hfi1_schedule_send;
1928 	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = get_pmtu_from_attr;
1929 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1930 	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = flush_qp_waiters;
1931 	dd->verbs_dev.rdi.driver_f.stop_send_queue = stop_send_queue;
1932 	dd->verbs_dev.rdi.driver_f.quiesce_qp = quiesce_qp;
1933 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1934 	dd->verbs_dev.rdi.driver_f.mtu_from_qp = mtu_from_qp;
1935 	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = mtu_to_path_mtu;
1936 	dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
1937 	dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
1938 	dd->verbs_dev.rdi.driver_f.notify_restart_rc = hfi1_restart_rc;
1939 	dd->verbs_dev.rdi.driver_f.check_send_wqe = hfi1_check_send_wqe;
1940 
1941 	/* completeion queue */
1942 	snprintf(dd->verbs_dev.rdi.dparms.cq_name,
1943 		 sizeof(dd->verbs_dev.rdi.dparms.cq_name),
1944 		 "hfi1_cq%d", dd->unit);
1945 	dd->verbs_dev.rdi.dparms.node = dd->node;
1946 
1947 	/* misc settings */
1948 	dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
1949 	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
1950 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1951 	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
1952 
1953 	/* post send table */
1954 	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;
1955 
1956 	ppd = dd->pport;
1957 	for (i = 0; i < dd->num_pports; i++, ppd++)
1958 		rvt_init_port(&dd->verbs_dev.rdi,
1959 			      &ppd->ibport_data.rvp,
1960 			      i,
1961 			      ppd->pkeys);
1962 
1963 	ret = rvt_register_device(&dd->verbs_dev.rdi);
1964 	if (ret)
1965 		goto err_verbs_txreq;
1966 
1967 	ret = hfi1_verbs_register_sysfs(dd);
1968 	if (ret)
1969 		goto err_class;
1970 
1971 	return ret;
1972 
1973 err_class:
1974 	rvt_unregister_device(&dd->verbs_dev.rdi);
1975 err_verbs_txreq:
1976 	verbs_txreq_exit(dev);
1977 	dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1978 	return ret;
1979 }
1980 
1981 void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
1982 {
1983 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1984 
1985 	hfi1_verbs_unregister_sysfs(dd);
1986 
1987 	rvt_unregister_device(&dd->verbs_dev.rdi);
1988 
1989 	if (!list_empty(&dev->txwait))
1990 		dd_dev_err(dd, "txwait list not empty!\n");
1991 	if (!list_empty(&dev->memwait))
1992 		dd_dev_err(dd, "memwait list not empty!\n");
1993 
1994 	del_timer_sync(&dev->mem_timer);
1995 	verbs_txreq_exit(dev);
1996 
1997 	mutex_lock(&cntr_names_lock);
1998 	kfree(dev_cntr_names);
1999 	kfree(port_cntr_names);
2000 	dev_cntr_names = NULL;
2001 	port_cntr_names = NULL;
2002 	cntr_names_initialized = 0;
2003 	mutex_unlock(&cntr_names_lock);
2004 }
2005 
2006 void hfi1_cnp_rcv(struct hfi1_packet *packet)
2007 {
2008 	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
2009 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
2010 	struct ib_header *hdr = packet->hdr;
2011 	struct rvt_qp *qp = packet->qp;
2012 	u32 lqpn, rqpn = 0;
2013 	u16 rlid = 0;
2014 	u8 sl, sc5, svc_type;
2015 
2016 	switch (packet->qp->ibqp.qp_type) {
2017 	case IB_QPT_UC:
2018 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
2019 		rqpn = qp->remote_qpn;
2020 		svc_type = IB_CC_SVCTYPE_UC;
2021 		break;
2022 	case IB_QPT_RC:
2023 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
2024 		rqpn = qp->remote_qpn;
2025 		svc_type = IB_CC_SVCTYPE_RC;
2026 		break;
2027 	case IB_QPT_SMI:
2028 	case IB_QPT_GSI:
2029 	case IB_QPT_UD:
2030 		svc_type = IB_CC_SVCTYPE_UD;
2031 		break;
2032 	default:
2033 		ibp->rvp.n_pkt_drops++;
2034 		return;
2035 	}
2036 
2037 	sc5 = hfi1_9B_get_sc5(hdr, packet->rhf);
2038 	sl = ibp->sc_to_sl[sc5];
2039 	lqpn = qp->ibqp.qp_num;
2040 
2041 	process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
2042 }
2043