xref: /openbmc/linux/drivers/infiniband/hw/hfi1/verbs.c (revision 045f77ba)
1 /*
2  * Copyright(c) 2015 - 2018 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 #include "fault.h"
67 #include "affinity.h"
68 
69 static unsigned int hfi1_lkey_table_size = 16;
70 module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
71 		   S_IRUGO);
72 MODULE_PARM_DESC(lkey_table_size,
73 		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
74 
75 static unsigned int hfi1_max_pds = 0xFFFF;
76 module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
77 MODULE_PARM_DESC(max_pds,
78 		 "Maximum number of protection domains to support");
79 
80 static unsigned int hfi1_max_ahs = 0xFFFF;
81 module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
82 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
83 
84 unsigned int hfi1_max_cqes = 0x2FFFFF;
85 module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
86 MODULE_PARM_DESC(max_cqes,
87 		 "Maximum number of completion queue entries to support");
88 
89 unsigned int hfi1_max_cqs = 0x1FFFF;
90 module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
91 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
92 
93 unsigned int hfi1_max_qp_wrs = 0x3FFF;
94 module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
95 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
96 
97 unsigned int hfi1_max_qps = 32768;
98 module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
99 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
100 
101 unsigned int hfi1_max_sges = 0x60;
102 module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
103 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
104 
105 unsigned int hfi1_max_mcast_grps = 16384;
106 module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
107 MODULE_PARM_DESC(max_mcast_grps,
108 		 "Maximum number of multicast groups to support");
109 
110 unsigned int hfi1_max_mcast_qp_attached = 16;
111 module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
112 		   uint, S_IRUGO);
113 MODULE_PARM_DESC(max_mcast_qp_attached,
114 		 "Maximum number of attached QPs to support");
115 
116 unsigned int hfi1_max_srqs = 1024;
117 module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
118 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
119 
120 unsigned int hfi1_max_srq_sges = 128;
121 module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
122 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
123 
124 unsigned int hfi1_max_srq_wrs = 0x1FFFF;
125 module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
126 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
127 
128 unsigned short piothreshold = 256;
129 module_param(piothreshold, ushort, S_IRUGO);
130 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
131 
132 #define COPY_CACHELESS 1
133 #define COPY_ADAPTIVE  2
134 static unsigned int sge_copy_mode;
135 module_param(sge_copy_mode, uint, S_IRUGO);
136 MODULE_PARM_DESC(sge_copy_mode,
137 		 "Verbs copy mode: 0 use memcpy, 1 use cacheless copy, 2 adapt based on WSS");
138 
139 static void verbs_sdma_complete(
140 	struct sdma_txreq *cookie,
141 	int status);
142 
143 static int pio_wait(struct rvt_qp *qp,
144 		    struct send_context *sc,
145 		    struct hfi1_pkt_state *ps,
146 		    u32 flag);
147 
148 /* Length of buffer to create verbs txreq cache name */
149 #define TXREQ_NAME_LEN 24
150 
151 /* 16B trailing buffer */
152 static const u8 trail_buf[MAX_16B_PADDING];
153 
154 static uint wss_threshold;
155 module_param(wss_threshold, uint, S_IRUGO);
156 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
157 static uint wss_clean_period = 256;
158 module_param(wss_clean_period, uint, S_IRUGO);
159 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
160 
161 /* memory working set size */
162 struct hfi1_wss {
163 	unsigned long *entries;
164 	atomic_t total_count;
165 	atomic_t clean_counter;
166 	atomic_t clean_entry;
167 
168 	int threshold;
169 	int num_entries;
170 	long pages_mask;
171 };
172 
173 static struct hfi1_wss wss;
174 
175 int hfi1_wss_init(void)
176 {
177 	long llc_size;
178 	long llc_bits;
179 	long table_size;
180 	long table_bits;
181 
182 	/* check for a valid percent range - default to 80 if none or invalid */
183 	if (wss_threshold < 1 || wss_threshold > 100)
184 		wss_threshold = 80;
185 	/* reject a wildly large period */
186 	if (wss_clean_period > 1000000)
187 		wss_clean_period = 256;
188 	/* reject a zero period */
189 	if (wss_clean_period == 0)
190 		wss_clean_period = 1;
191 
192 	/*
193 	 * Calculate the table size - the next power of 2 larger than the
194 	 * LLC size.  LLC size is in KiB.
195 	 */
196 	llc_size = wss_llc_size() * 1024;
197 	table_size = roundup_pow_of_two(llc_size);
198 
199 	/* one bit per page in rounded up table */
200 	llc_bits = llc_size / PAGE_SIZE;
201 	table_bits = table_size / PAGE_SIZE;
202 	wss.pages_mask = table_bits - 1;
203 	wss.num_entries = table_bits / BITS_PER_LONG;
204 
205 	wss.threshold = (llc_bits * wss_threshold) / 100;
206 	if (wss.threshold == 0)
207 		wss.threshold = 1;
208 
209 	atomic_set(&wss.clean_counter, wss_clean_period);
210 
211 	wss.entries = kcalloc(wss.num_entries, sizeof(*wss.entries),
212 			      GFP_KERNEL);
213 	if (!wss.entries) {
214 		hfi1_wss_exit();
215 		return -ENOMEM;
216 	}
217 
218 	return 0;
219 }
220 
221 void hfi1_wss_exit(void)
222 {
223 	/* coded to handle partially initialized and repeat callers */
224 	kfree(wss.entries);
225 	wss.entries = NULL;
226 }
227 
228 /*
229  * Advance the clean counter.  When the clean period has expired,
230  * clean an entry.
231  *
232  * This is implemented in atomics to avoid locking.  Because multiple
233  * variables are involved, it can be racy which can lead to slightly
234  * inaccurate information.  Since this is only a heuristic, this is
235  * OK.  Any innaccuracies will clean themselves out as the counter
236  * advances.  That said, it is unlikely the entry clean operation will
237  * race - the next possible racer will not start until the next clean
238  * period.
239  *
240  * The clean counter is implemented as a decrement to zero.  When zero
241  * is reached an entry is cleaned.
242  */
243 static void wss_advance_clean_counter(void)
244 {
245 	int entry;
246 	int weight;
247 	unsigned long bits;
248 
249 	/* become the cleaner if we decrement the counter to zero */
250 	if (atomic_dec_and_test(&wss.clean_counter)) {
251 		/*
252 		 * Set, not add, the clean period.  This avoids an issue
253 		 * where the counter could decrement below the clean period.
254 		 * Doing a set can result in lost decrements, slowing the
255 		 * clean advance.  Since this a heuristic, this possible
256 		 * slowdown is OK.
257 		 *
258 		 * An alternative is to loop, advancing the counter by a
259 		 * clean period until the result is > 0. However, this could
260 		 * lead to several threads keeping another in the clean loop.
261 		 * This could be mitigated by limiting the number of times
262 		 * we stay in the loop.
263 		 */
264 		atomic_set(&wss.clean_counter, wss_clean_period);
265 
266 		/*
267 		 * Uniquely grab the entry to clean and move to next.
268 		 * The current entry is always the lower bits of
269 		 * wss.clean_entry.  The table size, wss.num_entries,
270 		 * is always a power-of-2.
271 		 */
272 		entry = (atomic_inc_return(&wss.clean_entry) - 1)
273 			& (wss.num_entries - 1);
274 
275 		/* clear the entry and count the bits */
276 		bits = xchg(&wss.entries[entry], 0);
277 		weight = hweight64((u64)bits);
278 		/* only adjust the contended total count if needed */
279 		if (weight)
280 			atomic_sub(weight, &wss.total_count);
281 	}
282 }
283 
284 /*
285  * Insert the given address into the working set array.
286  */
287 static void wss_insert(void *address)
288 {
289 	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss.pages_mask;
290 	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
291 	u32 nr = page & (BITS_PER_LONG - 1);
292 
293 	if (!test_and_set_bit(nr, &wss.entries[entry]))
294 		atomic_inc(&wss.total_count);
295 
296 	wss_advance_clean_counter();
297 }
298 
299 /*
300  * Is the working set larger than the threshold?
301  */
302 static inline bool wss_exceeds_threshold(void)
303 {
304 	return atomic_read(&wss.total_count) >= wss.threshold;
305 }
306 
307 /*
308  * Translate ib_wr_opcode into ib_wc_opcode.
309  */
310 const enum ib_wc_opcode ib_hfi1_wc_opcode[] = {
311 	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
312 	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
313 	[IB_WR_SEND] = IB_WC_SEND,
314 	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
315 	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
316 	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
317 	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD,
318 	[IB_WR_SEND_WITH_INV] = IB_WC_SEND,
319 	[IB_WR_LOCAL_INV] = IB_WC_LOCAL_INV,
320 	[IB_WR_REG_MR] = IB_WC_REG_MR
321 };
322 
323 /*
324  * Length of header by opcode, 0 --> not supported
325  */
326 const u8 hdr_len_by_opcode[256] = {
327 	/* RC */
328 	[IB_OPCODE_RC_SEND_FIRST]                     = 12 + 8,
329 	[IB_OPCODE_RC_SEND_MIDDLE]                    = 12 + 8,
330 	[IB_OPCODE_RC_SEND_LAST]                      = 12 + 8,
331 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
332 	[IB_OPCODE_RC_SEND_ONLY]                      = 12 + 8,
333 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
334 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
335 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = 12 + 8,
336 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = 12 + 8,
337 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
338 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
339 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
340 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = 12 + 8 + 16,
341 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = 12 + 8 + 4,
342 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = 12 + 8,
343 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = 12 + 8 + 4,
344 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = 12 + 8 + 4,
345 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = 12 + 8 + 4,
346 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = 12 + 8 + 4 + 8,
347 	[IB_OPCODE_RC_COMPARE_SWAP]                   = 12 + 8 + 28,
348 	[IB_OPCODE_RC_FETCH_ADD]                      = 12 + 8 + 28,
349 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = 12 + 8 + 4,
350 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = 12 + 8 + 4,
351 	/* UC */
352 	[IB_OPCODE_UC_SEND_FIRST]                     = 12 + 8,
353 	[IB_OPCODE_UC_SEND_MIDDLE]                    = 12 + 8,
354 	[IB_OPCODE_UC_SEND_LAST]                      = 12 + 8,
355 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
356 	[IB_OPCODE_UC_SEND_ONLY]                      = 12 + 8,
357 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
358 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
359 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = 12 + 8,
360 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = 12 + 8,
361 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
362 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
363 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
364 	/* UD */
365 	[IB_OPCODE_UD_SEND_ONLY]                      = 12 + 8 + 8,
366 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 12
367 };
368 
369 static const opcode_handler opcode_handler_tbl[256] = {
370 	/* RC */
371 	[IB_OPCODE_RC_SEND_FIRST]                     = &hfi1_rc_rcv,
372 	[IB_OPCODE_RC_SEND_MIDDLE]                    = &hfi1_rc_rcv,
373 	[IB_OPCODE_RC_SEND_LAST]                      = &hfi1_rc_rcv,
374 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
375 	[IB_OPCODE_RC_SEND_ONLY]                      = &hfi1_rc_rcv,
376 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
377 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = &hfi1_rc_rcv,
378 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = &hfi1_rc_rcv,
379 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = &hfi1_rc_rcv,
380 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
381 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = &hfi1_rc_rcv,
382 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
383 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = &hfi1_rc_rcv,
384 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = &hfi1_rc_rcv,
385 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = &hfi1_rc_rcv,
386 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = &hfi1_rc_rcv,
387 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = &hfi1_rc_rcv,
388 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = &hfi1_rc_rcv,
389 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = &hfi1_rc_rcv,
390 	[IB_OPCODE_RC_COMPARE_SWAP]                   = &hfi1_rc_rcv,
391 	[IB_OPCODE_RC_FETCH_ADD]                      = &hfi1_rc_rcv,
392 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = &hfi1_rc_rcv,
393 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = &hfi1_rc_rcv,
394 	/* UC */
395 	[IB_OPCODE_UC_SEND_FIRST]                     = &hfi1_uc_rcv,
396 	[IB_OPCODE_UC_SEND_MIDDLE]                    = &hfi1_uc_rcv,
397 	[IB_OPCODE_UC_SEND_LAST]                      = &hfi1_uc_rcv,
398 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
399 	[IB_OPCODE_UC_SEND_ONLY]                      = &hfi1_uc_rcv,
400 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
401 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = &hfi1_uc_rcv,
402 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = &hfi1_uc_rcv,
403 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = &hfi1_uc_rcv,
404 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
405 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = &hfi1_uc_rcv,
406 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
407 	/* UD */
408 	[IB_OPCODE_UD_SEND_ONLY]                      = &hfi1_ud_rcv,
409 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_ud_rcv,
410 	/* CNP */
411 	[IB_OPCODE_CNP]				      = &hfi1_cnp_rcv
412 };
413 
414 #define OPMASK 0x1f
415 
416 static const u32 pio_opmask[BIT(3)] = {
417 	/* RC */
418 	[IB_OPCODE_RC >> 5] =
419 		BIT(RC_OP(SEND_ONLY) & OPMASK) |
420 		BIT(RC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
421 		BIT(RC_OP(RDMA_WRITE_ONLY) & OPMASK) |
422 		BIT(RC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK) |
423 		BIT(RC_OP(RDMA_READ_REQUEST) & OPMASK) |
424 		BIT(RC_OP(ACKNOWLEDGE) & OPMASK) |
425 		BIT(RC_OP(ATOMIC_ACKNOWLEDGE) & OPMASK) |
426 		BIT(RC_OP(COMPARE_SWAP) & OPMASK) |
427 		BIT(RC_OP(FETCH_ADD) & OPMASK),
428 	/* UC */
429 	[IB_OPCODE_UC >> 5] =
430 		BIT(UC_OP(SEND_ONLY) & OPMASK) |
431 		BIT(UC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
432 		BIT(UC_OP(RDMA_WRITE_ONLY) & OPMASK) |
433 		BIT(UC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK),
434 };
435 
436 /*
437  * System image GUID.
438  */
439 __be64 ib_hfi1_sys_image_guid;
440 
441 /**
442  * hfi1_copy_sge - copy data to SGE memory
443  * @ss: the SGE state
444  * @data: the data to copy
445  * @length: the length of the data
446  * @release: boolean to release MR
447  * @copy_last: do a separate copy of the last 8 bytes
448  */
449 void hfi1_copy_sge(
450 	struct rvt_sge_state *ss,
451 	void *data, u32 length,
452 	bool release,
453 	bool copy_last)
454 {
455 	struct rvt_sge *sge = &ss->sge;
456 	int i;
457 	bool in_last = false;
458 	bool cacheless_copy = false;
459 
460 	if (sge_copy_mode == COPY_CACHELESS) {
461 		cacheless_copy = length >= PAGE_SIZE;
462 	} else if (sge_copy_mode == COPY_ADAPTIVE) {
463 		if (length >= PAGE_SIZE) {
464 			/*
465 			 * NOTE: this *assumes*:
466 			 * o The first vaddr is the dest.
467 			 * o If multiple pages, then vaddr is sequential.
468 			 */
469 			wss_insert(sge->vaddr);
470 			if (length >= (2 * PAGE_SIZE))
471 				wss_insert(sge->vaddr + PAGE_SIZE);
472 
473 			cacheless_copy = wss_exceeds_threshold();
474 		} else {
475 			wss_advance_clean_counter();
476 		}
477 	}
478 	if (copy_last) {
479 		if (length > 8) {
480 			length -= 8;
481 		} else {
482 			copy_last = false;
483 			in_last = true;
484 		}
485 	}
486 
487 again:
488 	while (length) {
489 		u32 len = rvt_get_sge_length(sge, length);
490 
491 		WARN_ON_ONCE(len == 0);
492 		if (unlikely(in_last)) {
493 			/* enforce byte transfer ordering */
494 			for (i = 0; i < len; i++)
495 				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
496 		} else if (cacheless_copy) {
497 			cacheless_memcpy(sge->vaddr, data, len);
498 		} else {
499 			memcpy(sge->vaddr, data, len);
500 		}
501 		rvt_update_sge(ss, len, release);
502 		data += len;
503 		length -= len;
504 	}
505 
506 	if (copy_last) {
507 		copy_last = false;
508 		in_last = true;
509 		length = 8;
510 		goto again;
511 	}
512 }
513 
514 /*
515  * Make sure the QP is ready and able to accept the given opcode.
516  */
517 static inline opcode_handler qp_ok(struct hfi1_packet *packet)
518 {
519 	if (!(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
520 		return NULL;
521 	if (((packet->opcode & RVT_OPCODE_QP_MASK) ==
522 	     packet->qp->allowed_ops) ||
523 	    (packet->opcode == IB_OPCODE_CNP))
524 		return opcode_handler_tbl[packet->opcode];
525 
526 	return NULL;
527 }
528 
529 static u64 hfi1_fault_tx(struct rvt_qp *qp, u8 opcode, u64 pbc)
530 {
531 #ifdef CONFIG_FAULT_INJECTION
532 	if ((opcode & IB_OPCODE_MSP) == IB_OPCODE_MSP)
533 		/*
534 		 * In order to drop non-IB traffic we
535 		 * set PbcInsertHrc to NONE (0x2).
536 		 * The packet will still be delivered
537 		 * to the receiving node but a
538 		 * KHdrHCRCErr (KDETH packet with a bad
539 		 * HCRC) will be triggered and the
540 		 * packet will not be delivered to the
541 		 * correct context.
542 		 */
543 		pbc |= (u64)PBC_IHCRC_NONE << PBC_INSERT_HCRC_SHIFT;
544 	else
545 		/*
546 		 * In order to drop regular verbs
547 		 * traffic we set the PbcTestEbp
548 		 * flag. The packet will still be
549 		 * delivered to the receiving node but
550 		 * a 'late ebp error' will be
551 		 * triggered and will be dropped.
552 		 */
553 		pbc |= PBC_TEST_EBP;
554 #endif
555 	return pbc;
556 }
557 
558 static int hfi1_do_pkey_check(struct hfi1_packet *packet)
559 {
560 	struct hfi1_ctxtdata *rcd = packet->rcd;
561 	struct hfi1_pportdata *ppd = rcd->ppd;
562 	struct hfi1_16b_header *hdr = packet->hdr;
563 	u16 pkey;
564 
565 	/* Pkey check needed only for bypass packets */
566 	if (packet->etype != RHF_RCV_TYPE_BYPASS)
567 		return 0;
568 
569 	/* Perform pkey check */
570 	pkey = hfi1_16B_get_pkey(hdr);
571 	return ingress_pkey_check(ppd, pkey, packet->sc,
572 				  packet->qp->s_pkey_index,
573 				  packet->slid, true);
574 }
575 
576 static inline void hfi1_handle_packet(struct hfi1_packet *packet,
577 				      bool is_mcast)
578 {
579 	u32 qp_num;
580 	struct hfi1_ctxtdata *rcd = packet->rcd;
581 	struct hfi1_pportdata *ppd = rcd->ppd;
582 	struct hfi1_ibport *ibp = rcd_to_iport(rcd);
583 	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
584 	opcode_handler packet_handler;
585 	unsigned long flags;
586 
587 	inc_opstats(packet->tlen, &rcd->opstats->stats[packet->opcode]);
588 
589 	if (unlikely(is_mcast)) {
590 		struct rvt_mcast *mcast;
591 		struct rvt_mcast_qp *p;
592 
593 		if (!packet->grh)
594 			goto drop;
595 		mcast = rvt_mcast_find(&ibp->rvp,
596 				       &packet->grh->dgid,
597 				       opa_get_lid(packet->dlid, 9B));
598 		if (!mcast)
599 			goto drop;
600 		list_for_each_entry_rcu(p, &mcast->qp_list, list) {
601 			packet->qp = p->qp;
602 			if (hfi1_do_pkey_check(packet))
603 				goto drop;
604 			spin_lock_irqsave(&packet->qp->r_lock, flags);
605 			packet_handler = qp_ok(packet);
606 			if (likely(packet_handler))
607 				packet_handler(packet);
608 			else
609 				ibp->rvp.n_pkt_drops++;
610 			spin_unlock_irqrestore(&packet->qp->r_lock, flags);
611 		}
612 		/*
613 		 * Notify rvt_multicast_detach() if it is waiting for us
614 		 * to finish.
615 		 */
616 		if (atomic_dec_return(&mcast->refcount) <= 1)
617 			wake_up(&mcast->wait);
618 	} else {
619 		/* Get the destination QP number. */
620 		if (packet->etype == RHF_RCV_TYPE_BYPASS &&
621 		    hfi1_16B_get_l4(packet->hdr) == OPA_16B_L4_FM)
622 			qp_num = hfi1_16B_get_dest_qpn(packet->mgmt);
623 		else
624 			qp_num = ib_bth_get_qpn(packet->ohdr);
625 
626 		rcu_read_lock();
627 		packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
628 		if (!packet->qp)
629 			goto unlock_drop;
630 
631 		if (hfi1_do_pkey_check(packet))
632 			goto unlock_drop;
633 
634 		spin_lock_irqsave(&packet->qp->r_lock, flags);
635 		packet_handler = qp_ok(packet);
636 		if (likely(packet_handler))
637 			packet_handler(packet);
638 		else
639 			ibp->rvp.n_pkt_drops++;
640 		spin_unlock_irqrestore(&packet->qp->r_lock, flags);
641 		rcu_read_unlock();
642 	}
643 	return;
644 unlock_drop:
645 	rcu_read_unlock();
646 drop:
647 	ibp->rvp.n_pkt_drops++;
648 }
649 
650 /**
651  * hfi1_ib_rcv - process an incoming packet
652  * @packet: data packet information
653  *
654  * This is called to process an incoming packet at interrupt level.
655  */
656 void hfi1_ib_rcv(struct hfi1_packet *packet)
657 {
658 	struct hfi1_ctxtdata *rcd = packet->rcd;
659 
660 	trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));
661 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
662 }
663 
664 void hfi1_16B_rcv(struct hfi1_packet *packet)
665 {
666 	struct hfi1_ctxtdata *rcd = packet->rcd;
667 
668 	trace_input_ibhdr(rcd->dd, packet, false);
669 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
670 }
671 
672 /*
673  * This is called from a timer to check for QPs
674  * which need kernel memory in order to send a packet.
675  */
676 static void mem_timer(struct timer_list *t)
677 {
678 	struct hfi1_ibdev *dev = from_timer(dev, t, mem_timer);
679 	struct list_head *list = &dev->memwait;
680 	struct rvt_qp *qp = NULL;
681 	struct iowait *wait;
682 	unsigned long flags;
683 	struct hfi1_qp_priv *priv;
684 
685 	write_seqlock_irqsave(&dev->iowait_lock, flags);
686 	if (!list_empty(list)) {
687 		wait = list_first_entry(list, struct iowait, list);
688 		qp = iowait_to_qp(wait);
689 		priv = qp->priv;
690 		list_del_init(&priv->s_iowait.list);
691 		priv->s_iowait.lock = NULL;
692 		/* refcount held until actual wake up */
693 		if (!list_empty(list))
694 			mod_timer(&dev->mem_timer, jiffies + 1);
695 	}
696 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
697 
698 	if (qp)
699 		hfi1_qp_wakeup(qp, RVT_S_WAIT_KMEM);
700 }
701 
702 /*
703  * This is called with progress side lock held.
704  */
705 /* New API */
706 static void verbs_sdma_complete(
707 	struct sdma_txreq *cookie,
708 	int status)
709 {
710 	struct verbs_txreq *tx =
711 		container_of(cookie, struct verbs_txreq, txreq);
712 	struct rvt_qp *qp = tx->qp;
713 
714 	spin_lock(&qp->s_lock);
715 	if (tx->wqe) {
716 		hfi1_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
717 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
718 		struct hfi1_opa_header *hdr;
719 
720 		hdr = &tx->phdr.hdr;
721 		hfi1_rc_send_complete(qp, hdr);
722 	}
723 	spin_unlock(&qp->s_lock);
724 
725 	hfi1_put_txreq(tx);
726 }
727 
728 static int wait_kmem(struct hfi1_ibdev *dev,
729 		     struct rvt_qp *qp,
730 		     struct hfi1_pkt_state *ps)
731 {
732 	struct hfi1_qp_priv *priv = qp->priv;
733 	unsigned long flags;
734 	int ret = 0;
735 
736 	spin_lock_irqsave(&qp->s_lock, flags);
737 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
738 		write_seqlock(&dev->iowait_lock);
739 		list_add_tail(&ps->s_txreq->txreq.list,
740 			      &priv->s_iowait.tx_head);
741 		if (list_empty(&priv->s_iowait.list)) {
742 			if (list_empty(&dev->memwait))
743 				mod_timer(&dev->mem_timer, jiffies + 1);
744 			qp->s_flags |= RVT_S_WAIT_KMEM;
745 			list_add_tail(&priv->s_iowait.list, &dev->memwait);
746 			priv->s_iowait.lock = &dev->iowait_lock;
747 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM);
748 			rvt_get_qp(qp);
749 		}
750 		write_sequnlock(&dev->iowait_lock);
751 		qp->s_flags &= ~RVT_S_BUSY;
752 		ret = -EBUSY;
753 	}
754 	spin_unlock_irqrestore(&qp->s_lock, flags);
755 
756 	return ret;
757 }
758 
759 /*
760  * This routine calls txadds for each sg entry.
761  *
762  * Add failures will revert the sge cursor
763  */
764 static noinline int build_verbs_ulp_payload(
765 	struct sdma_engine *sde,
766 	u32 length,
767 	struct verbs_txreq *tx)
768 {
769 	struct rvt_sge_state *ss = tx->ss;
770 	struct rvt_sge *sg_list = ss->sg_list;
771 	struct rvt_sge sge = ss->sge;
772 	u8 num_sge = ss->num_sge;
773 	u32 len;
774 	int ret = 0;
775 
776 	while (length) {
777 		len = ss->sge.length;
778 		if (len > length)
779 			len = length;
780 		if (len > ss->sge.sge_length)
781 			len = ss->sge.sge_length;
782 		WARN_ON_ONCE(len == 0);
783 		ret = sdma_txadd_kvaddr(
784 			sde->dd,
785 			&tx->txreq,
786 			ss->sge.vaddr,
787 			len);
788 		if (ret)
789 			goto bail_txadd;
790 		rvt_update_sge(ss, len, false);
791 		length -= len;
792 	}
793 	return ret;
794 bail_txadd:
795 	/* unwind cursor */
796 	ss->sge = sge;
797 	ss->num_sge = num_sge;
798 	ss->sg_list = sg_list;
799 	return ret;
800 }
801 
802 /**
803  * update_tx_opstats - record stats by opcode
804  * @qp; the qp
805  * @ps: transmit packet state
806  * @plen: the plen in dwords
807  *
808  * This is a routine to record the tx opstats after a
809  * packet has been presented to the egress mechanism.
810  */
811 static void update_tx_opstats(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
812 			      u32 plen)
813 {
814 #ifdef CONFIG_DEBUG_FS
815 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
816 	struct hfi1_opcode_stats_perctx *s = get_cpu_ptr(dd->tx_opstats);
817 
818 	inc_opstats(plen * 4, &s->stats[ps->opcode]);
819 	put_cpu_ptr(s);
820 #endif
821 }
822 
823 /*
824  * Build the number of DMA descriptors needed to send length bytes of data.
825  *
826  * NOTE: DMA mapping is held in the tx until completed in the ring or
827  *       the tx desc is freed without having been submitted to the ring
828  *
829  * This routine ensures all the helper routine calls succeed.
830  */
831 /* New API */
832 static int build_verbs_tx_desc(
833 	struct sdma_engine *sde,
834 	u32 length,
835 	struct verbs_txreq *tx,
836 	struct hfi1_ahg_info *ahg_info,
837 	u64 pbc)
838 {
839 	int ret = 0;
840 	struct hfi1_sdma_header *phdr = &tx->phdr;
841 	u16 hdrbytes = (tx->hdr_dwords + sizeof(pbc) / 4) << 2;
842 	u8 extra_bytes = 0;
843 
844 	if (tx->phdr.hdr.hdr_type) {
845 		/*
846 		 * hdrbytes accounts for PBC. Need to subtract 8 bytes
847 		 * before calculating padding.
848 		 */
849 		extra_bytes = hfi1_get_16b_padding(hdrbytes - 8, length) +
850 			      (SIZE_OF_CRC << 2) + SIZE_OF_LT;
851 	}
852 	if (!ahg_info->ahgcount) {
853 		ret = sdma_txinit_ahg(
854 			&tx->txreq,
855 			ahg_info->tx_flags,
856 			hdrbytes + length +
857 			extra_bytes,
858 			ahg_info->ahgidx,
859 			0,
860 			NULL,
861 			0,
862 			verbs_sdma_complete);
863 		if (ret)
864 			goto bail_txadd;
865 		phdr->pbc = cpu_to_le64(pbc);
866 		ret = sdma_txadd_kvaddr(
867 			sde->dd,
868 			&tx->txreq,
869 			phdr,
870 			hdrbytes);
871 		if (ret)
872 			goto bail_txadd;
873 	} else {
874 		ret = sdma_txinit_ahg(
875 			&tx->txreq,
876 			ahg_info->tx_flags,
877 			length,
878 			ahg_info->ahgidx,
879 			ahg_info->ahgcount,
880 			ahg_info->ahgdesc,
881 			hdrbytes,
882 			verbs_sdma_complete);
883 		if (ret)
884 			goto bail_txadd;
885 	}
886 	/* add the ulp payload - if any. tx->ss can be NULL for acks */
887 	if (tx->ss) {
888 		ret = build_verbs_ulp_payload(sde, length, tx);
889 		if (ret)
890 			goto bail_txadd;
891 	}
892 
893 	/* add icrc, lt byte, and padding to flit */
894 	if (extra_bytes)
895 		ret = sdma_txadd_kvaddr(sde->dd, &tx->txreq,
896 					(void *)trail_buf, extra_bytes);
897 
898 bail_txadd:
899 	return ret;
900 }
901 
902 int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
903 			u64 pbc)
904 {
905 	struct hfi1_qp_priv *priv = qp->priv;
906 	struct hfi1_ahg_info *ahg_info = priv->s_ahg;
907 	u32 hdrwords = ps->s_txreq->hdr_dwords;
908 	u32 len = ps->s_txreq->s_cur_size;
909 	u32 plen;
910 	struct hfi1_ibdev *dev = ps->dev;
911 	struct hfi1_pportdata *ppd = ps->ppd;
912 	struct verbs_txreq *tx;
913 	u8 sc5 = priv->s_sc;
914 	int ret;
915 	u32 dwords;
916 
917 	if (ps->s_txreq->phdr.hdr.hdr_type) {
918 		u8 extra_bytes = hfi1_get_16b_padding((hdrwords << 2), len);
919 
920 		dwords = (len + extra_bytes + (SIZE_OF_CRC << 2) +
921 			  SIZE_OF_LT) >> 2;
922 	} else {
923 		dwords = (len + 3) >> 2;
924 	}
925 	plen = hdrwords + dwords + sizeof(pbc) / 4;
926 
927 	tx = ps->s_txreq;
928 	if (!sdma_txreq_built(&tx->txreq)) {
929 		if (likely(pbc == 0)) {
930 			u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
931 
932 			/* No vl15 here */
933 			/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
934 			if (ps->s_txreq->phdr.hdr.hdr_type)
935 				pbc |= PBC_PACKET_BYPASS |
936 				       PBC_INSERT_BYPASS_ICRC;
937 			else
938 				pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
939 
940 			if (unlikely(hfi1_dbg_should_fault_tx(qp, ps->opcode)))
941 				pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
942 			pbc = create_pbc(ppd,
943 					 pbc,
944 					 qp->srate_mbps,
945 					 vl,
946 					 plen);
947 		}
948 		tx->wqe = qp->s_wqe;
949 		ret = build_verbs_tx_desc(tx->sde, len, tx, ahg_info, pbc);
950 		if (unlikely(ret))
951 			goto bail_build;
952 	}
953 	ret =  sdma_send_txreq(tx->sde, &priv->s_iowait, &tx->txreq,
954 			       ps->pkts_sent);
955 	if (unlikely(ret < 0)) {
956 		if (ret == -ECOMM)
957 			goto bail_ecomm;
958 		return ret;
959 	}
960 
961 	update_tx_opstats(qp, ps, plen);
962 	trace_sdma_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
963 				&ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
964 	return ret;
965 
966 bail_ecomm:
967 	/* The current one got "sent" */
968 	return 0;
969 bail_build:
970 	ret = wait_kmem(dev, qp, ps);
971 	if (!ret) {
972 		/* free txreq - bad state */
973 		hfi1_put_txreq(ps->s_txreq);
974 		ps->s_txreq = NULL;
975 	}
976 	return ret;
977 }
978 
979 /*
980  * If we are now in the error state, return zero to flush the
981  * send work request.
982  */
983 static int pio_wait(struct rvt_qp *qp,
984 		    struct send_context *sc,
985 		    struct hfi1_pkt_state *ps,
986 		    u32 flag)
987 {
988 	struct hfi1_qp_priv *priv = qp->priv;
989 	struct hfi1_devdata *dd = sc->dd;
990 	struct hfi1_ibdev *dev = &dd->verbs_dev;
991 	unsigned long flags;
992 	int ret = 0;
993 
994 	/*
995 	 * Note that as soon as want_buffer() is called and
996 	 * possibly before it returns, sc_piobufavail()
997 	 * could be called. Therefore, put QP on the I/O wait list before
998 	 * enabling the PIO avail interrupt.
999 	 */
1000 	spin_lock_irqsave(&qp->s_lock, flags);
1001 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
1002 		write_seqlock(&dev->iowait_lock);
1003 		list_add_tail(&ps->s_txreq->txreq.list,
1004 			      &priv->s_iowait.tx_head);
1005 		if (list_empty(&priv->s_iowait.list)) {
1006 			struct hfi1_ibdev *dev = &dd->verbs_dev;
1007 			int was_empty;
1008 
1009 			dev->n_piowait += !!(flag & RVT_S_WAIT_PIO);
1010 			dev->n_piodrain += !!(flag & HFI1_S_WAIT_PIO_DRAIN);
1011 			qp->s_flags |= flag;
1012 			was_empty = list_empty(&sc->piowait);
1013 			iowait_queue(ps->pkts_sent, &priv->s_iowait,
1014 				     &sc->piowait);
1015 			priv->s_iowait.lock = &dev->iowait_lock;
1016 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_PIO);
1017 			rvt_get_qp(qp);
1018 			/* counting: only call wantpiobuf_intr if first user */
1019 			if (was_empty)
1020 				hfi1_sc_wantpiobuf_intr(sc, 1);
1021 		}
1022 		write_sequnlock(&dev->iowait_lock);
1023 		qp->s_flags &= ~RVT_S_BUSY;
1024 		ret = -EBUSY;
1025 	}
1026 	spin_unlock_irqrestore(&qp->s_lock, flags);
1027 	return ret;
1028 }
1029 
1030 static void verbs_pio_complete(void *arg, int code)
1031 {
1032 	struct rvt_qp *qp = (struct rvt_qp *)arg;
1033 	struct hfi1_qp_priv *priv = qp->priv;
1034 
1035 	if (iowait_pio_dec(&priv->s_iowait))
1036 		iowait_drain_wakeup(&priv->s_iowait);
1037 }
1038 
1039 int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
1040 			u64 pbc)
1041 {
1042 	struct hfi1_qp_priv *priv = qp->priv;
1043 	u32 hdrwords = ps->s_txreq->hdr_dwords;
1044 	struct rvt_sge_state *ss = ps->s_txreq->ss;
1045 	u32 len = ps->s_txreq->s_cur_size;
1046 	u32 dwords;
1047 	u32 plen;
1048 	struct hfi1_pportdata *ppd = ps->ppd;
1049 	u32 *hdr;
1050 	u8 sc5;
1051 	unsigned long flags = 0;
1052 	struct send_context *sc;
1053 	struct pio_buf *pbuf;
1054 	int wc_status = IB_WC_SUCCESS;
1055 	int ret = 0;
1056 	pio_release_cb cb = NULL;
1057 	u8 extra_bytes = 0;
1058 
1059 	if (ps->s_txreq->phdr.hdr.hdr_type) {
1060 		u8 pad_size = hfi1_get_16b_padding((hdrwords << 2), len);
1061 
1062 		extra_bytes = pad_size + (SIZE_OF_CRC << 2) + SIZE_OF_LT;
1063 		dwords = (len + extra_bytes) >> 2;
1064 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.opah;
1065 	} else {
1066 		dwords = (len + 3) >> 2;
1067 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.ibh;
1068 	}
1069 	plen = hdrwords + dwords + sizeof(pbc) / 4;
1070 
1071 	/* only RC/UC use complete */
1072 	switch (qp->ibqp.qp_type) {
1073 	case IB_QPT_RC:
1074 	case IB_QPT_UC:
1075 		cb = verbs_pio_complete;
1076 		break;
1077 	default:
1078 		break;
1079 	}
1080 
1081 	/* vl15 special case taken care of in ud.c */
1082 	sc5 = priv->s_sc;
1083 	sc = ps->s_txreq->psc;
1084 
1085 	if (likely(pbc == 0)) {
1086 		u8 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
1087 
1088 		/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
1089 		if (ps->s_txreq->phdr.hdr.hdr_type)
1090 			pbc |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
1091 		else
1092 			pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
1093 
1094 		if (unlikely(hfi1_dbg_should_fault_tx(qp, ps->opcode)))
1095 			pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
1096 		pbc = create_pbc(ppd, pbc, qp->srate_mbps, vl, plen);
1097 	}
1098 	if (cb)
1099 		iowait_pio_inc(&priv->s_iowait);
1100 	pbuf = sc_buffer_alloc(sc, plen, cb, qp);
1101 	if (unlikely(!pbuf)) {
1102 		if (cb)
1103 			verbs_pio_complete(qp, 0);
1104 		if (ppd->host_link_state != HLS_UP_ACTIVE) {
1105 			/*
1106 			 * If we have filled the PIO buffers to capacity and are
1107 			 * not in an active state this request is not going to
1108 			 * go out to so just complete it with an error or else a
1109 			 * ULP or the core may be stuck waiting.
1110 			 */
1111 			hfi1_cdbg(
1112 				PIO,
1113 				"alloc failed. state not active, completing");
1114 			wc_status = IB_WC_GENERAL_ERR;
1115 			goto pio_bail;
1116 		} else {
1117 			/*
1118 			 * This is a normal occurrence. The PIO buffs are full
1119 			 * up but we are still happily sending, well we could be
1120 			 * so lets continue to queue the request.
1121 			 */
1122 			hfi1_cdbg(PIO, "alloc failed. state active, queuing");
1123 			ret = pio_wait(qp, sc, ps, RVT_S_WAIT_PIO);
1124 			if (!ret)
1125 				/* txreq not queued - free */
1126 				goto bail;
1127 			/* tx consumed in wait */
1128 			return ret;
1129 		}
1130 	}
1131 
1132 	if (dwords == 0) {
1133 		pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
1134 	} else {
1135 		seg_pio_copy_start(pbuf, pbc,
1136 				   hdr, hdrwords * 4);
1137 		if (ss) {
1138 			while (len) {
1139 				void *addr = ss->sge.vaddr;
1140 				u32 slen = ss->sge.length;
1141 
1142 				if (slen > len)
1143 					slen = len;
1144 				rvt_update_sge(ss, slen, false);
1145 				seg_pio_copy_mid(pbuf, addr, slen);
1146 				len -= slen;
1147 			}
1148 		}
1149 		/* add icrc, lt byte, and padding to flit */
1150 		if (extra_bytes)
1151 			seg_pio_copy_mid(pbuf, trail_buf, extra_bytes);
1152 
1153 		seg_pio_copy_end(pbuf);
1154 	}
1155 
1156 	update_tx_opstats(qp, ps, plen);
1157 	trace_pio_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
1158 			       &ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
1159 
1160 pio_bail:
1161 	if (qp->s_wqe) {
1162 		spin_lock_irqsave(&qp->s_lock, flags);
1163 		hfi1_send_complete(qp, qp->s_wqe, wc_status);
1164 		spin_unlock_irqrestore(&qp->s_lock, flags);
1165 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1166 		spin_lock_irqsave(&qp->s_lock, flags);
1167 		hfi1_rc_send_complete(qp, &ps->s_txreq->phdr.hdr);
1168 		spin_unlock_irqrestore(&qp->s_lock, flags);
1169 	}
1170 
1171 	ret = 0;
1172 
1173 bail:
1174 	hfi1_put_txreq(ps->s_txreq);
1175 	return ret;
1176 }
1177 
1178 /*
1179  * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1180  * being an entry from the partition key table), return 0
1181  * otherwise. Use the matching criteria for egress partition keys
1182  * specified in the OPAv1 spec., section 9.1l.7.
1183  */
1184 static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
1185 {
1186 	u16 mkey = pkey & PKEY_LOW_15_MASK;
1187 	u16 mentry = ent & PKEY_LOW_15_MASK;
1188 
1189 	if (mkey == mentry) {
1190 		/*
1191 		 * If pkey[15] is set (full partition member),
1192 		 * is bit 15 in the corresponding table element
1193 		 * clear (limited member)?
1194 		 */
1195 		if (pkey & PKEY_MEMBER_MASK)
1196 			return !!(ent & PKEY_MEMBER_MASK);
1197 		return 1;
1198 	}
1199 	return 0;
1200 }
1201 
1202 /**
1203  * egress_pkey_check - check P_KEY of a packet
1204  * @ppd:  Physical IB port data
1205  * @slid: SLID for packet
1206  * @bkey: PKEY for header
1207  * @sc5:  SC for packet
1208  * @s_pkey_index: It will be used for look up optimization for kernel contexts
1209  * only. If it is negative value, then it means user contexts is calling this
1210  * function.
1211  *
1212  * It checks if hdr's pkey is valid.
1213  *
1214  * Return: 0 on success, otherwise, 1
1215  */
1216 int egress_pkey_check(struct hfi1_pportdata *ppd, u32 slid, u16 pkey,
1217 		      u8 sc5, int8_t s_pkey_index)
1218 {
1219 	struct hfi1_devdata *dd;
1220 	int i;
1221 	int is_user_ctxt_mechanism = (s_pkey_index < 0);
1222 
1223 	if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
1224 		return 0;
1225 
1226 	/* If SC15, pkey[0:14] must be 0x7fff */
1227 	if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
1228 		goto bad;
1229 
1230 	/* Is the pkey = 0x0, or 0x8000? */
1231 	if ((pkey & PKEY_LOW_15_MASK) == 0)
1232 		goto bad;
1233 
1234 	/*
1235 	 * For the kernel contexts only, if a qp is passed into the function,
1236 	 * the most likely matching pkey has index qp->s_pkey_index
1237 	 */
1238 	if (!is_user_ctxt_mechanism &&
1239 	    egress_pkey_matches_entry(pkey, ppd->pkeys[s_pkey_index])) {
1240 		return 0;
1241 	}
1242 
1243 	for (i = 0; i < MAX_PKEY_VALUES; i++) {
1244 		if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
1245 			return 0;
1246 	}
1247 bad:
1248 	/*
1249 	 * For the user-context mechanism, the P_KEY check would only happen
1250 	 * once per SDMA request, not once per packet.  Therefore, there's no
1251 	 * need to increment the counter for the user-context mechanism.
1252 	 */
1253 	if (!is_user_ctxt_mechanism) {
1254 		incr_cntr64(&ppd->port_xmit_constraint_errors);
1255 		dd = ppd->dd;
1256 		if (!(dd->err_info_xmit_constraint.status &
1257 		      OPA_EI_STATUS_SMASK)) {
1258 			dd->err_info_xmit_constraint.status |=
1259 				OPA_EI_STATUS_SMASK;
1260 			dd->err_info_xmit_constraint.slid = slid;
1261 			dd->err_info_xmit_constraint.pkey = pkey;
1262 		}
1263 	}
1264 	return 1;
1265 }
1266 
1267 /**
1268  * get_send_routine - choose an egress routine
1269  *
1270  * Choose an egress routine based on QP type
1271  * and size
1272  */
1273 static inline send_routine get_send_routine(struct rvt_qp *qp,
1274 					    struct hfi1_pkt_state *ps)
1275 {
1276 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1277 	struct hfi1_qp_priv *priv = qp->priv;
1278 	struct verbs_txreq *tx = ps->s_txreq;
1279 
1280 	if (unlikely(!(dd->flags & HFI1_HAS_SEND_DMA)))
1281 		return dd->process_pio_send;
1282 	switch (qp->ibqp.qp_type) {
1283 	case IB_QPT_SMI:
1284 		return dd->process_pio_send;
1285 	case IB_QPT_GSI:
1286 	case IB_QPT_UD:
1287 		break;
1288 	case IB_QPT_UC:
1289 	case IB_QPT_RC: {
1290 		if (piothreshold &&
1291 		    tx->s_cur_size <= min(piothreshold, qp->pmtu) &&
1292 		    (BIT(ps->opcode & OPMASK) & pio_opmask[ps->opcode >> 5]) &&
1293 		    iowait_sdma_pending(&priv->s_iowait) == 0 &&
1294 		    !sdma_txreq_built(&tx->txreq))
1295 			return dd->process_pio_send;
1296 		break;
1297 	}
1298 	default:
1299 		break;
1300 	}
1301 	return dd->process_dma_send;
1302 }
1303 
1304 /**
1305  * hfi1_verbs_send - send a packet
1306  * @qp: the QP to send on
1307  * @ps: the state of the packet to send
1308  *
1309  * Return zero if packet is sent or queued OK.
1310  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1311  */
1312 int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
1313 {
1314 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1315 	struct hfi1_qp_priv *priv = qp->priv;
1316 	struct ib_other_headers *ohdr = NULL;
1317 	send_routine sr;
1318 	int ret;
1319 	u16 pkey;
1320 	u32 slid;
1321 	u8 l4 = 0;
1322 
1323 	/* locate the pkey within the headers */
1324 	if (ps->s_txreq->phdr.hdr.hdr_type) {
1325 		struct hfi1_16b_header *hdr = &ps->s_txreq->phdr.hdr.opah;
1326 
1327 		l4 = hfi1_16B_get_l4(hdr);
1328 		if (l4 == OPA_16B_L4_IB_LOCAL)
1329 			ohdr = &hdr->u.oth;
1330 		else if (l4 == OPA_16B_L4_IB_GLOBAL)
1331 			ohdr = &hdr->u.l.oth;
1332 
1333 		slid = hfi1_16B_get_slid(hdr);
1334 		pkey = hfi1_16B_get_pkey(hdr);
1335 	} else {
1336 		struct ib_header *hdr = &ps->s_txreq->phdr.hdr.ibh;
1337 		u8 lnh = ib_get_lnh(hdr);
1338 
1339 		if (lnh == HFI1_LRH_GRH)
1340 			ohdr = &hdr->u.l.oth;
1341 		else
1342 			ohdr = &hdr->u.oth;
1343 		slid = ib_get_slid(hdr);
1344 		pkey = ib_bth_get_pkey(ohdr);
1345 	}
1346 
1347 	if (likely(l4 != OPA_16B_L4_FM))
1348 		ps->opcode = ib_bth_get_opcode(ohdr);
1349 	else
1350 		ps->opcode = IB_OPCODE_UD_SEND_ONLY;
1351 
1352 	sr = get_send_routine(qp, ps);
1353 	ret = egress_pkey_check(dd->pport, slid, pkey,
1354 				priv->s_sc, qp->s_pkey_index);
1355 	if (unlikely(ret)) {
1356 		/*
1357 		 * The value we are returning here does not get propagated to
1358 		 * the verbs caller. Thus we need to complete the request with
1359 		 * error otherwise the caller could be sitting waiting on the
1360 		 * completion event. Only do this for PIO. SDMA has its own
1361 		 * mechanism for handling the errors. So for SDMA we can just
1362 		 * return.
1363 		 */
1364 		if (sr == dd->process_pio_send) {
1365 			unsigned long flags;
1366 
1367 			hfi1_cdbg(PIO, "%s() Failed. Completing with err",
1368 				  __func__);
1369 			spin_lock_irqsave(&qp->s_lock, flags);
1370 			hfi1_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
1371 			spin_unlock_irqrestore(&qp->s_lock, flags);
1372 		}
1373 		return -EINVAL;
1374 	}
1375 	if (sr == dd->process_dma_send && iowait_pio_pending(&priv->s_iowait))
1376 		return pio_wait(qp,
1377 				ps->s_txreq->psc,
1378 				ps,
1379 				HFI1_S_WAIT_PIO_DRAIN);
1380 	return sr(qp, ps, 0);
1381 }
1382 
1383 /**
1384  * hfi1_fill_device_attr - Fill in rvt dev info device attributes.
1385  * @dd: the device data structure
1386  */
1387 static void hfi1_fill_device_attr(struct hfi1_devdata *dd)
1388 {
1389 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1390 	u32 ver = dd->dc8051_ver;
1391 
1392 	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1393 
1394 	rdi->dparms.props.fw_ver = ((u64)(dc8051_ver_maj(ver)) << 32) |
1395 		((u64)(dc8051_ver_min(ver)) << 16) |
1396 		(u64)dc8051_ver_patch(ver);
1397 
1398 	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1399 			IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1400 			IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1401 			IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE |
1402 			IB_DEVICE_MEM_MGT_EXTENSIONS |
1403 			IB_DEVICE_RDMA_NETDEV_OPA_VNIC;
1404 	rdi->dparms.props.page_size_cap = PAGE_SIZE;
1405 	rdi->dparms.props.vendor_id = dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
1406 	rdi->dparms.props.vendor_part_id = dd->pcidev->device;
1407 	rdi->dparms.props.hw_ver = dd->minrev;
1408 	rdi->dparms.props.sys_image_guid = ib_hfi1_sys_image_guid;
1409 	rdi->dparms.props.max_mr_size = U64_MAX;
1410 	rdi->dparms.props.max_fast_reg_page_list_len = UINT_MAX;
1411 	rdi->dparms.props.max_qp = hfi1_max_qps;
1412 	rdi->dparms.props.max_qp_wr = hfi1_max_qp_wrs;
1413 	rdi->dparms.props.max_send_sge = hfi1_max_sges;
1414 	rdi->dparms.props.max_recv_sge = hfi1_max_sges;
1415 	rdi->dparms.props.max_sge_rd = hfi1_max_sges;
1416 	rdi->dparms.props.max_cq = hfi1_max_cqs;
1417 	rdi->dparms.props.max_ah = hfi1_max_ahs;
1418 	rdi->dparms.props.max_cqe = hfi1_max_cqes;
1419 	rdi->dparms.props.max_mr = rdi->lkey_table.max;
1420 	rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1421 	rdi->dparms.props.max_map_per_fmr = 32767;
1422 	rdi->dparms.props.max_pd = hfi1_max_pds;
1423 	rdi->dparms.props.max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
1424 	rdi->dparms.props.max_qp_init_rd_atom = 255;
1425 	rdi->dparms.props.max_srq = hfi1_max_srqs;
1426 	rdi->dparms.props.max_srq_wr = hfi1_max_srq_wrs;
1427 	rdi->dparms.props.max_srq_sge = hfi1_max_srq_sges;
1428 	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1429 	rdi->dparms.props.max_pkeys = hfi1_get_npkeys(dd);
1430 	rdi->dparms.props.max_mcast_grp = hfi1_max_mcast_grps;
1431 	rdi->dparms.props.max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
1432 	rdi->dparms.props.max_total_mcast_qp_attach =
1433 					rdi->dparms.props.max_mcast_qp_attach *
1434 					rdi->dparms.props.max_mcast_grp;
1435 }
1436 
1437 static inline u16 opa_speed_to_ib(u16 in)
1438 {
1439 	u16 out = 0;
1440 
1441 	if (in & OPA_LINK_SPEED_25G)
1442 		out |= IB_SPEED_EDR;
1443 	if (in & OPA_LINK_SPEED_12_5G)
1444 		out |= IB_SPEED_FDR;
1445 
1446 	return out;
1447 }
1448 
1449 /*
1450  * Convert a single OPA link width (no multiple flags) to an IB value.
1451  * A zero OPA link width means link down, which means the IB width value
1452  * is a don't care.
1453  */
1454 static inline u16 opa_width_to_ib(u16 in)
1455 {
1456 	switch (in) {
1457 	case OPA_LINK_WIDTH_1X:
1458 	/* map 2x and 3x to 1x as they don't exist in IB */
1459 	case OPA_LINK_WIDTH_2X:
1460 	case OPA_LINK_WIDTH_3X:
1461 		return IB_WIDTH_1X;
1462 	default: /* link down or unknown, return our largest width */
1463 	case OPA_LINK_WIDTH_4X:
1464 		return IB_WIDTH_4X;
1465 	}
1466 }
1467 
1468 static int query_port(struct rvt_dev_info *rdi, u8 port_num,
1469 		      struct ib_port_attr *props)
1470 {
1471 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1472 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1473 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1474 	u32 lid = ppd->lid;
1475 
1476 	/* props being zeroed by the caller, avoid zeroing it here */
1477 	props->lid = lid ? lid : 0;
1478 	props->lmc = ppd->lmc;
1479 	/* OPA logical states match IB logical states */
1480 	props->state = driver_lstate(ppd);
1481 	props->phys_state = driver_pstate(ppd);
1482 	props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
1483 	props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
1484 	/* see rate_show() in ib core/sysfs.c */
1485 	props->active_speed = (u8)opa_speed_to_ib(ppd->link_speed_active);
1486 	props->max_vl_num = ppd->vls_supported;
1487 
1488 	/* Once we are a "first class" citizen and have added the OPA MTUs to
1489 	 * the core we can advertise the larger MTU enum to the ULPs, for now
1490 	 * advertise only 4K.
1491 	 *
1492 	 * Those applications which are either OPA aware or pass the MTU enum
1493 	 * from the Path Records to us will get the new 8k MTU.  Those that
1494 	 * attempt to process the MTU enum may fail in various ways.
1495 	 */
1496 	props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
1497 				      4096 : hfi1_max_mtu), IB_MTU_4096);
1498 	props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
1499 		mtu_to_enum(ppd->ibmtu, IB_MTU_4096);
1500 
1501 	return 0;
1502 }
1503 
1504 static int modify_device(struct ib_device *device,
1505 			 int device_modify_mask,
1506 			 struct ib_device_modify *device_modify)
1507 {
1508 	struct hfi1_devdata *dd = dd_from_ibdev(device);
1509 	unsigned i;
1510 	int ret;
1511 
1512 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1513 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1514 		ret = -EOPNOTSUPP;
1515 		goto bail;
1516 	}
1517 
1518 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1519 		memcpy(device->node_desc, device_modify->node_desc,
1520 		       IB_DEVICE_NODE_DESC_MAX);
1521 		for (i = 0; i < dd->num_pports; i++) {
1522 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1523 
1524 			hfi1_node_desc_chg(ibp);
1525 		}
1526 	}
1527 
1528 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1529 		ib_hfi1_sys_image_guid =
1530 			cpu_to_be64(device_modify->sys_image_guid);
1531 		for (i = 0; i < dd->num_pports; i++) {
1532 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1533 
1534 			hfi1_sys_guid_chg(ibp);
1535 		}
1536 	}
1537 
1538 	ret = 0;
1539 
1540 bail:
1541 	return ret;
1542 }
1543 
1544 static int shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1545 {
1546 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1547 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1548 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1549 	int ret;
1550 
1551 	set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
1552 			     OPA_LINKDOWN_REASON_UNKNOWN);
1553 	ret = set_link_state(ppd, HLS_DN_DOWNDEF);
1554 	return ret;
1555 }
1556 
1557 static int hfi1_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1558 			    int guid_index, __be64 *guid)
1559 {
1560 	struct hfi1_ibport *ibp = container_of(rvp, struct hfi1_ibport, rvp);
1561 
1562 	if (guid_index >= HFI1_GUIDS_PER_PORT)
1563 		return -EINVAL;
1564 
1565 	*guid = get_sguid(ibp, guid_index);
1566 	return 0;
1567 }
1568 
1569 /*
1570  * convert ah port,sl to sc
1571  */
1572 u8 ah_to_sc(struct ib_device *ibdev, struct rdma_ah_attr *ah)
1573 {
1574 	struct hfi1_ibport *ibp = to_iport(ibdev, rdma_ah_get_port_num(ah));
1575 
1576 	return ibp->sl_to_sc[rdma_ah_get_sl(ah)];
1577 }
1578 
1579 static int hfi1_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1580 {
1581 	struct hfi1_ibport *ibp;
1582 	struct hfi1_pportdata *ppd;
1583 	struct hfi1_devdata *dd;
1584 	u8 sc5;
1585 
1586 	if (hfi1_check_mcast(rdma_ah_get_dlid(ah_attr)) &&
1587 	    !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1588 		return -EINVAL;
1589 
1590 	/* test the mapping for validity */
1591 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1592 	ppd = ppd_from_ibp(ibp);
1593 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(ah_attr)];
1594 	dd = dd_from_ppd(ppd);
1595 	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1596 		return -EINVAL;
1597 	return 0;
1598 }
1599 
1600 static void hfi1_notify_new_ah(struct ib_device *ibdev,
1601 			       struct rdma_ah_attr *ah_attr,
1602 			       struct rvt_ah *ah)
1603 {
1604 	struct hfi1_ibport *ibp;
1605 	struct hfi1_pportdata *ppd;
1606 	struct hfi1_devdata *dd;
1607 	u8 sc5;
1608 	struct rdma_ah_attr *attr = &ah->attr;
1609 
1610 	/*
1611 	 * Do not trust reading anything from rvt_ah at this point as it is not
1612 	 * done being setup. We can however modify things which we need to set.
1613 	 */
1614 
1615 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1616 	ppd = ppd_from_ibp(ibp);
1617 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)];
1618 	hfi1_update_ah_attr(ibdev, attr);
1619 	hfi1_make_opa_lid(attr);
1620 	dd = dd_from_ppd(ppd);
1621 	ah->vl = sc_to_vlt(dd, sc5);
1622 	if (ah->vl < num_vls || ah->vl == 15)
1623 		ah->log_pmtu = ilog2(dd->vld[ah->vl].mtu);
1624 }
1625 
1626 /**
1627  * hfi1_get_npkeys - return the size of the PKEY table for context 0
1628  * @dd: the hfi1_ib device
1629  */
1630 unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
1631 {
1632 	return ARRAY_SIZE(dd->pport[0].pkeys);
1633 }
1634 
1635 static void init_ibport(struct hfi1_pportdata *ppd)
1636 {
1637 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1638 	size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
1639 	int i;
1640 
1641 	for (i = 0; i < sz; i++) {
1642 		ibp->sl_to_sc[i] = i;
1643 		ibp->sc_to_sl[i] = i;
1644 	}
1645 
1646 	for (i = 0; i < RVT_MAX_TRAP_LISTS ; i++)
1647 		INIT_LIST_HEAD(&ibp->rvp.trap_lists[i].list);
1648 	timer_setup(&ibp->rvp.trap_timer, hfi1_handle_trap_timer, 0);
1649 
1650 	spin_lock_init(&ibp->rvp.lock);
1651 	/* Set the prefix to the default value (see ch. 4.1.1) */
1652 	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1653 	ibp->rvp.sm_lid = 0;
1654 	/*
1655 	 * Below should only set bits defined in OPA PortInfo.CapabilityMask
1656 	 * and PortInfo.CapabilityMask3
1657 	 */
1658 	ibp->rvp.port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
1659 		IB_PORT_CAP_MASK_NOTICE_SUP;
1660 	ibp->rvp.port_cap3_flags = OPA_CAP_MASK3_IsSharedSpaceSupported;
1661 	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1662 	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1663 	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1664 	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1665 	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1666 
1667 	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1668 	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1669 }
1670 
1671 static void hfi1_get_dev_fw_str(struct ib_device *ibdev, char *str)
1672 {
1673 	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
1674 	struct hfi1_ibdev *dev = dev_from_rdi(rdi);
1675 	u32 ver = dd_from_dev(dev)->dc8051_ver;
1676 
1677 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u.%u", dc8051_ver_maj(ver),
1678 		 dc8051_ver_min(ver), dc8051_ver_patch(ver));
1679 }
1680 
1681 static const char * const driver_cntr_names[] = {
1682 	/* must be element 0*/
1683 	"DRIVER_KernIntr",
1684 	"DRIVER_ErrorIntr",
1685 	"DRIVER_Tx_Errs",
1686 	"DRIVER_Rcv_Errs",
1687 	"DRIVER_HW_Errs",
1688 	"DRIVER_NoPIOBufs",
1689 	"DRIVER_CtxtsOpen",
1690 	"DRIVER_RcvLen_Errs",
1691 	"DRIVER_EgrBufFull",
1692 	"DRIVER_EgrHdrFull"
1693 };
1694 
1695 static DEFINE_MUTEX(cntr_names_lock); /* protects the *_cntr_names bufers */
1696 static const char **dev_cntr_names;
1697 static const char **port_cntr_names;
1698 static int num_driver_cntrs = ARRAY_SIZE(driver_cntr_names);
1699 static int num_dev_cntrs;
1700 static int num_port_cntrs;
1701 static int cntr_names_initialized;
1702 
1703 /*
1704  * Convert a list of names separated by '\n' into an array of NULL terminated
1705  * strings. Optionally some entries can be reserved in the array to hold extra
1706  * external strings.
1707  */
1708 static int init_cntr_names(const char *names_in,
1709 			   const size_t names_len,
1710 			   int num_extra_names,
1711 			   int *num_cntrs,
1712 			   const char ***cntr_names)
1713 {
1714 	char *names_out, *p, **q;
1715 	int i, n;
1716 
1717 	n = 0;
1718 	for (i = 0; i < names_len; i++)
1719 		if (names_in[i] == '\n')
1720 			n++;
1721 
1722 	names_out = kmalloc((n + num_extra_names) * sizeof(char *) + names_len,
1723 			    GFP_KERNEL);
1724 	if (!names_out) {
1725 		*num_cntrs = 0;
1726 		*cntr_names = NULL;
1727 		return -ENOMEM;
1728 	}
1729 
1730 	p = names_out + (n + num_extra_names) * sizeof(char *);
1731 	memcpy(p, names_in, names_len);
1732 
1733 	q = (char **)names_out;
1734 	for (i = 0; i < n; i++) {
1735 		q[i] = p;
1736 		p = strchr(p, '\n');
1737 		*p++ = '\0';
1738 	}
1739 
1740 	*num_cntrs = n;
1741 	*cntr_names = (const char **)names_out;
1742 	return 0;
1743 }
1744 
1745 static struct rdma_hw_stats *alloc_hw_stats(struct ib_device *ibdev,
1746 					    u8 port_num)
1747 {
1748 	int i, err;
1749 
1750 	mutex_lock(&cntr_names_lock);
1751 	if (!cntr_names_initialized) {
1752 		struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1753 
1754 		err = init_cntr_names(dd->cntrnames,
1755 				      dd->cntrnameslen,
1756 				      num_driver_cntrs,
1757 				      &num_dev_cntrs,
1758 				      &dev_cntr_names);
1759 		if (err) {
1760 			mutex_unlock(&cntr_names_lock);
1761 			return NULL;
1762 		}
1763 
1764 		for (i = 0; i < num_driver_cntrs; i++)
1765 			dev_cntr_names[num_dev_cntrs + i] =
1766 				driver_cntr_names[i];
1767 
1768 		err = init_cntr_names(dd->portcntrnames,
1769 				      dd->portcntrnameslen,
1770 				      0,
1771 				      &num_port_cntrs,
1772 				      &port_cntr_names);
1773 		if (err) {
1774 			kfree(dev_cntr_names);
1775 			dev_cntr_names = NULL;
1776 			mutex_unlock(&cntr_names_lock);
1777 			return NULL;
1778 		}
1779 		cntr_names_initialized = 1;
1780 	}
1781 	mutex_unlock(&cntr_names_lock);
1782 
1783 	if (!port_num)
1784 		return rdma_alloc_hw_stats_struct(
1785 				dev_cntr_names,
1786 				num_dev_cntrs + num_driver_cntrs,
1787 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1788 	else
1789 		return rdma_alloc_hw_stats_struct(
1790 				port_cntr_names,
1791 				num_port_cntrs,
1792 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1793 }
1794 
1795 static u64 hfi1_sps_ints(void)
1796 {
1797 	unsigned long flags;
1798 	struct hfi1_devdata *dd;
1799 	u64 sps_ints = 0;
1800 
1801 	spin_lock_irqsave(&hfi1_devs_lock, flags);
1802 	list_for_each_entry(dd, &hfi1_dev_list, list) {
1803 		sps_ints += get_all_cpu_total(dd->int_counter);
1804 	}
1805 	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1806 	return sps_ints;
1807 }
1808 
1809 static int get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1810 			u8 port, int index)
1811 {
1812 	u64 *values;
1813 	int count;
1814 
1815 	if (!port) {
1816 		u64 *stats = (u64 *)&hfi1_stats;
1817 		int i;
1818 
1819 		hfi1_read_cntrs(dd_from_ibdev(ibdev), NULL, &values);
1820 		values[num_dev_cntrs] = hfi1_sps_ints();
1821 		for (i = 1; i < num_driver_cntrs; i++)
1822 			values[num_dev_cntrs + i] = stats[i];
1823 		count = num_dev_cntrs + num_driver_cntrs;
1824 	} else {
1825 		struct hfi1_ibport *ibp = to_iport(ibdev, port);
1826 
1827 		hfi1_read_portcntrs(ppd_from_ibp(ibp), NULL, &values);
1828 		count = num_port_cntrs;
1829 	}
1830 
1831 	memcpy(stats->value, values, count * sizeof(u64));
1832 	return count;
1833 }
1834 
1835 /**
1836  * hfi1_register_ib_device - register our device with the infiniband core
1837  * @dd: the device data structure
1838  * Return 0 if successful, errno if unsuccessful.
1839  */
1840 int hfi1_register_ib_device(struct hfi1_devdata *dd)
1841 {
1842 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1843 	struct ib_device *ibdev = &dev->rdi.ibdev;
1844 	struct hfi1_pportdata *ppd = dd->pport;
1845 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1846 	unsigned i;
1847 	int ret;
1848 
1849 	for (i = 0; i < dd->num_pports; i++)
1850 		init_ibport(ppd + i);
1851 
1852 	/* Only need to initialize non-zero fields. */
1853 
1854 	timer_setup(&dev->mem_timer, mem_timer, 0);
1855 
1856 	seqlock_init(&dev->iowait_lock);
1857 	seqlock_init(&dev->txwait_lock);
1858 	INIT_LIST_HEAD(&dev->txwait);
1859 	INIT_LIST_HEAD(&dev->memwait);
1860 
1861 	ret = verbs_txreq_init(dev);
1862 	if (ret)
1863 		goto err_verbs_txreq;
1864 
1865 	/* Use first-port GUID as node guid */
1866 	ibdev->node_guid = get_sguid(ibp, HFI1_PORT_GUID_INDEX);
1867 
1868 	/*
1869 	 * The system image GUID is supposed to be the same for all
1870 	 * HFIs in a single system but since there can be other
1871 	 * device types in the system, we can't be sure this is unique.
1872 	 */
1873 	if (!ib_hfi1_sys_image_guid)
1874 		ib_hfi1_sys_image_guid = ibdev->node_guid;
1875 	ibdev->owner = THIS_MODULE;
1876 	ibdev->phys_port_cnt = dd->num_pports;
1877 	ibdev->dev.parent = &dd->pcidev->dev;
1878 	ibdev->modify_device = modify_device;
1879 	ibdev->alloc_hw_stats = alloc_hw_stats;
1880 	ibdev->get_hw_stats = get_hw_stats;
1881 	ibdev->alloc_rdma_netdev = hfi1_vnic_alloc_rn;
1882 
1883 	/* keep process mad in the driver */
1884 	ibdev->process_mad = hfi1_process_mad;
1885 	ibdev->get_dev_fw_str = hfi1_get_dev_fw_str;
1886 
1887 	strlcpy(ibdev->node_desc, init_utsname()->nodename,
1888 		sizeof(ibdev->node_desc));
1889 
1890 	/*
1891 	 * Fill in rvt info object.
1892 	 */
1893 	dd->verbs_dev.rdi.driver_f.port_callback = hfi1_create_port_files;
1894 	dd->verbs_dev.rdi.driver_f.get_pci_dev = get_pci_dev;
1895 	dd->verbs_dev.rdi.driver_f.check_ah = hfi1_check_ah;
1896 	dd->verbs_dev.rdi.driver_f.notify_new_ah = hfi1_notify_new_ah;
1897 	dd->verbs_dev.rdi.driver_f.get_guid_be = hfi1_get_guid_be;
1898 	dd->verbs_dev.rdi.driver_f.query_port_state = query_port;
1899 	dd->verbs_dev.rdi.driver_f.shut_down_port = shut_down_port;
1900 	dd->verbs_dev.rdi.driver_f.cap_mask_chg = hfi1_cap_mask_chg;
1901 	/*
1902 	 * Fill in rvt info device attributes.
1903 	 */
1904 	hfi1_fill_device_attr(dd);
1905 
1906 	/* queue pair */
1907 	dd->verbs_dev.rdi.dparms.qp_table_size = hfi1_qp_table_size;
1908 	dd->verbs_dev.rdi.dparms.qpn_start = 0;
1909 	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1910 	dd->verbs_dev.rdi.dparms.qos_shift = dd->qos_shift;
1911 	dd->verbs_dev.rdi.dparms.qpn_res_start = kdeth_qp << 16;
1912 	dd->verbs_dev.rdi.dparms.qpn_res_end =
1913 	dd->verbs_dev.rdi.dparms.qpn_res_start + 65535;
1914 	dd->verbs_dev.rdi.dparms.max_rdma_atomic = HFI1_MAX_RDMA_ATOMIC;
1915 	dd->verbs_dev.rdi.dparms.psn_mask = PSN_MASK;
1916 	dd->verbs_dev.rdi.dparms.psn_shift = PSN_SHIFT;
1917 	dd->verbs_dev.rdi.dparms.psn_modify_mask = PSN_MODIFY_MASK;
1918 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_INTEL_OPA |
1919 						RDMA_CORE_CAP_OPA_AH;
1920 	dd->verbs_dev.rdi.dparms.max_mad_size = OPA_MGMT_MAD_SIZE;
1921 
1922 	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qp_priv_alloc;
1923 	dd->verbs_dev.rdi.driver_f.qp_priv_free = qp_priv_free;
1924 	dd->verbs_dev.rdi.driver_f.free_all_qps = free_all_qps;
1925 	dd->verbs_dev.rdi.driver_f.notify_qp_reset = notify_qp_reset;
1926 	dd->verbs_dev.rdi.driver_f.do_send = hfi1_do_send_from_rvt;
1927 	dd->verbs_dev.rdi.driver_f.schedule_send = hfi1_schedule_send;
1928 	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _hfi1_schedule_send;
1929 	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = get_pmtu_from_attr;
1930 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1931 	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = flush_qp_waiters;
1932 	dd->verbs_dev.rdi.driver_f.stop_send_queue = stop_send_queue;
1933 	dd->verbs_dev.rdi.driver_f.quiesce_qp = quiesce_qp;
1934 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1935 	dd->verbs_dev.rdi.driver_f.mtu_from_qp = mtu_from_qp;
1936 	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = mtu_to_path_mtu;
1937 	dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
1938 	dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
1939 	dd->verbs_dev.rdi.driver_f.notify_restart_rc = hfi1_restart_rc;
1940 	dd->verbs_dev.rdi.driver_f.check_send_wqe = hfi1_check_send_wqe;
1941 	dd->verbs_dev.rdi.driver_f.comp_vect_cpu_lookup =
1942 						hfi1_comp_vect_mappings_lookup;
1943 
1944 	/* completeion queue */
1945 	dd->verbs_dev.rdi.ibdev.num_comp_vectors = dd->comp_vect_possible_cpus;
1946 	dd->verbs_dev.rdi.dparms.node = dd->node;
1947 
1948 	/* misc settings */
1949 	dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
1950 	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
1951 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1952 	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
1953 
1954 	/* post send table */
1955 	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;
1956 
1957 	ppd = dd->pport;
1958 	for (i = 0; i < dd->num_pports; i++, ppd++)
1959 		rvt_init_port(&dd->verbs_dev.rdi,
1960 			      &ppd->ibport_data.rvp,
1961 			      i,
1962 			      ppd->pkeys);
1963 
1964 	ret = rvt_register_device(&dd->verbs_dev.rdi, RDMA_DRIVER_HFI1);
1965 	if (ret)
1966 		goto err_verbs_txreq;
1967 
1968 	ret = hfi1_verbs_register_sysfs(dd);
1969 	if (ret)
1970 		goto err_class;
1971 
1972 	return ret;
1973 
1974 err_class:
1975 	rvt_unregister_device(&dd->verbs_dev.rdi);
1976 err_verbs_txreq:
1977 	verbs_txreq_exit(dev);
1978 	dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1979 	return ret;
1980 }
1981 
1982 void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
1983 {
1984 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1985 
1986 	hfi1_verbs_unregister_sysfs(dd);
1987 
1988 	rvt_unregister_device(&dd->verbs_dev.rdi);
1989 
1990 	if (!list_empty(&dev->txwait))
1991 		dd_dev_err(dd, "txwait list not empty!\n");
1992 	if (!list_empty(&dev->memwait))
1993 		dd_dev_err(dd, "memwait list not empty!\n");
1994 
1995 	del_timer_sync(&dev->mem_timer);
1996 	verbs_txreq_exit(dev);
1997 
1998 	mutex_lock(&cntr_names_lock);
1999 	kfree(dev_cntr_names);
2000 	kfree(port_cntr_names);
2001 	dev_cntr_names = NULL;
2002 	port_cntr_names = NULL;
2003 	cntr_names_initialized = 0;
2004 	mutex_unlock(&cntr_names_lock);
2005 }
2006 
2007 void hfi1_cnp_rcv(struct hfi1_packet *packet)
2008 {
2009 	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
2010 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
2011 	struct ib_header *hdr = packet->hdr;
2012 	struct rvt_qp *qp = packet->qp;
2013 	u32 lqpn, rqpn = 0;
2014 	u16 rlid = 0;
2015 	u8 sl, sc5, svc_type;
2016 
2017 	switch (packet->qp->ibqp.qp_type) {
2018 	case IB_QPT_UC:
2019 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
2020 		rqpn = qp->remote_qpn;
2021 		svc_type = IB_CC_SVCTYPE_UC;
2022 		break;
2023 	case IB_QPT_RC:
2024 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
2025 		rqpn = qp->remote_qpn;
2026 		svc_type = IB_CC_SVCTYPE_RC;
2027 		break;
2028 	case IB_QPT_SMI:
2029 	case IB_QPT_GSI:
2030 	case IB_QPT_UD:
2031 		svc_type = IB_CC_SVCTYPE_UD;
2032 		break;
2033 	default:
2034 		ibp->rvp.n_pkt_drops++;
2035 		return;
2036 	}
2037 
2038 	sc5 = hfi1_9B_get_sc5(hdr, packet->rhf);
2039 	sl = ibp->sc_to_sl[sc5];
2040 	lqpn = qp->ibqp.qp_num;
2041 
2042 	process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
2043 }
2044