verbs.c 54.8 KB
Newer Older
M
Mike Marciniszyn 已提交
1
/*
J
Jubin John 已提交
2
 * Copyright(c) 2015, 2016 Intel Corporation.
M
Mike Marciniszyn 已提交
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61
 *
 * This file is provided under a dual BSD/GPLv2 license.  When using or
 * redistributing this file, you may do so under either license.
 *
 * GPL LICENSE SUMMARY
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of version 2 of the GNU General Public License as
 * published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * BSD LICENSE
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 *  - Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 *  - Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in
 *    the documentation and/or other materials provided with the
 *    distribution.
 *  - Neither the name of Intel Corporation nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 */

#include <rdma/ib_mad.h>
#include <rdma/ib_user_verbs.h>
#include <linux/io.h>
#include <linux/module.h>
#include <linux/utsname.h>
#include <linux/rculist.h>
#include <linux/mm.h>
#include <linux/vmalloc.h>

#include "hfi.h"
#include "common.h"
#include "device.h"
#include "trace.h"
#include "qp.h"
62
#include "verbs_txreq.h"
M
Mike Marciniszyn 已提交
63

64
static unsigned int hfi1_lkey_table_size = 16;
M
Mike Marciniszyn 已提交
65 66 67 68 69 70 71 72 73 74 75 76 77 78
module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
		   S_IRUGO);
MODULE_PARM_DESC(lkey_table_size,
		 "LKEY table size in bits (2^n, 1 <= n <= 23)");

static unsigned int hfi1_max_pds = 0xFFFF;
module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
MODULE_PARM_DESC(max_pds,
		 "Maximum number of protection domains to support");

static unsigned int hfi1_max_ahs = 0xFFFF;
module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");

79
unsigned int hfi1_max_cqes = 0x2FFFFF;
M
Mike Marciniszyn 已提交
80 81 82 83 84 85 86 87 88 89 90 91
module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
MODULE_PARM_DESC(max_cqes,
		 "Maximum number of completion queue entries to support");

unsigned int hfi1_max_cqs = 0x1FFFF;
module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");

unsigned int hfi1_max_qp_wrs = 0x3FFF;
module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");

92
unsigned int hfi1_max_qps = 32768;
M
Mike Marciniszyn 已提交
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");

unsigned int hfi1_max_sges = 0x60;
module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");

unsigned int hfi1_max_mcast_grps = 16384;
module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
MODULE_PARM_DESC(max_mcast_grps,
		 "Maximum number of multicast groups to support");

unsigned int hfi1_max_mcast_qp_attached = 16;
module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
		   uint, S_IRUGO);
MODULE_PARM_DESC(max_mcast_qp_attached,
		 "Maximum number of attached QPs to support");

unsigned int hfi1_max_srqs = 1024;
module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");

unsigned int hfi1_max_srq_sges = 128;
module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");

unsigned int hfi1_max_srq_wrs = 0x1FFFF;
module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");

123
unsigned short piothreshold = 256;
124 125 126
module_param(piothreshold, ushort, S_IRUGO);
MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");

127 128 129 130 131 132 133
#define COPY_CACHELESS 1
#define COPY_ADAPTIVE  2
static unsigned int sge_copy_mode;
module_param(sge_copy_mode, uint, S_IRUGO);
MODULE_PARM_DESC(sge_copy_mode,
		 "Verbs copy mode: 0 use memcpy, 1 use cacheless copy, 2 adapt based on WSS");

M
Mike Marciniszyn 已提交
134 135
static void verbs_sdma_complete(
	struct sdma_txreq *cookie,
136
	int status);
M
Mike Marciniszyn 已提交
137

138 139 140 141 142
static int pio_wait(struct rvt_qp *qp,
		    struct send_context *sc,
		    struct hfi1_pkt_state *ps,
		    u32 flag);

143 144 145
/* Length of buffer to create verbs txreq cache name */
#define TXREQ_NAME_LEN 24

146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298
static uint wss_threshold;
module_param(wss_threshold, uint, S_IRUGO);
MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
static uint wss_clean_period = 256;
module_param(wss_clean_period, uint, S_IRUGO);
MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");

/* memory working set size */
struct hfi1_wss {
	unsigned long *entries;
	atomic_t total_count;
	atomic_t clean_counter;
	atomic_t clean_entry;

	int threshold;
	int num_entries;
	long pages_mask;
};

static struct hfi1_wss wss;

int hfi1_wss_init(void)
{
	long llc_size;
	long llc_bits;
	long table_size;
	long table_bits;

	/* check for a valid percent range - default to 80 if none or invalid */
	if (wss_threshold < 1 || wss_threshold > 100)
		wss_threshold = 80;
	/* reject a wildly large period */
	if (wss_clean_period > 1000000)
		wss_clean_period = 256;
	/* reject a zero period */
	if (wss_clean_period == 0)
		wss_clean_period = 1;

	/*
	 * Calculate the table size - the next power of 2 larger than the
	 * LLC size.  LLC size is in KiB.
	 */
	llc_size = wss_llc_size() * 1024;
	table_size = roundup_pow_of_two(llc_size);

	/* one bit per page in rounded up table */
	llc_bits = llc_size / PAGE_SIZE;
	table_bits = table_size / PAGE_SIZE;
	wss.pages_mask = table_bits - 1;
	wss.num_entries = table_bits / BITS_PER_LONG;

	wss.threshold = (llc_bits * wss_threshold) / 100;
	if (wss.threshold == 0)
		wss.threshold = 1;

	atomic_set(&wss.clean_counter, wss_clean_period);

	wss.entries = kcalloc(wss.num_entries, sizeof(*wss.entries),
			      GFP_KERNEL);
	if (!wss.entries) {
		hfi1_wss_exit();
		return -ENOMEM;
	}

	return 0;
}

void hfi1_wss_exit(void)
{
	/* coded to handle partially initialized and repeat callers */
	kfree(wss.entries);
	wss.entries = NULL;
}

/*
 * Advance the clean counter.  When the clean period has expired,
 * clean an entry.
 *
 * This is implemented in atomics to avoid locking.  Because multiple
 * variables are involved, it can be racy which can lead to slightly
 * inaccurate information.  Since this is only a heuristic, this is
 * OK.  Any innaccuracies will clean themselves out as the counter
 * advances.  That said, it is unlikely the entry clean operation will
 * race - the next possible racer will not start until the next clean
 * period.
 *
 * The clean counter is implemented as a decrement to zero.  When zero
 * is reached an entry is cleaned.
 */
static void wss_advance_clean_counter(void)
{
	int entry;
	int weight;
	unsigned long bits;

	/* become the cleaner if we decrement the counter to zero */
	if (atomic_dec_and_test(&wss.clean_counter)) {
		/*
		 * Set, not add, the clean period.  This avoids an issue
		 * where the counter could decrement below the clean period.
		 * Doing a set can result in lost decrements, slowing the
		 * clean advance.  Since this a heuristic, this possible
		 * slowdown is OK.
		 *
		 * An alternative is to loop, advancing the counter by a
		 * clean period until the result is > 0. However, this could
		 * lead to several threads keeping another in the clean loop.
		 * This could be mitigated by limiting the number of times
		 * we stay in the loop.
		 */
		atomic_set(&wss.clean_counter, wss_clean_period);

		/*
		 * Uniquely grab the entry to clean and move to next.
		 * The current entry is always the lower bits of
		 * wss.clean_entry.  The table size, wss.num_entries,
		 * is always a power-of-2.
		 */
		entry = (atomic_inc_return(&wss.clean_entry) - 1)
			& (wss.num_entries - 1);

		/* clear the entry and count the bits */
		bits = xchg(&wss.entries[entry], 0);
		weight = hweight64((u64)bits);
		/* only adjust the contended total count if needed */
		if (weight)
			atomic_sub(weight, &wss.total_count);
	}
}

/*
 * Insert the given address into the working set array.
 */
static void wss_insert(void *address)
{
	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss.pages_mask;
	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
	u32 nr = page & (BITS_PER_LONG - 1);

	if (!test_and_set_bit(nr, &wss.entries[entry]))
		atomic_inc(&wss.total_count);

	wss_advance_clean_counter();
}

/*
 * Is the working set larger than the threshold?
 */
static inline int wss_exceeds_threshold(void)
{
	return atomic_read(&wss.total_count) >= wss.threshold;
}

M
Mike Marciniszyn 已提交
299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
/*
 * Length of header by opcode, 0 --> not supported
 */
const u8 hdr_len_by_opcode[256] = {
	/* RC */
	[IB_OPCODE_RC_SEND_FIRST]                     = 12 + 8,
	[IB_OPCODE_RC_SEND_MIDDLE]                    = 12 + 8,
	[IB_OPCODE_RC_SEND_LAST]                      = 12 + 8,
	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
	[IB_OPCODE_RC_SEND_ONLY]                      = 12 + 8,
	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = 12 + 8,
	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = 12 + 8,
	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = 12 + 8 + 16,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = 12 + 8 + 4,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = 12 + 8,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = 12 + 8 + 4,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = 12 + 8 + 4,
	[IB_OPCODE_RC_ACKNOWLEDGE]                    = 12 + 8 + 4,
322
	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = 12 + 8 + 4 + 8,
M
Mike Marciniszyn 已提交
323 324
	[IB_OPCODE_RC_COMPARE_SWAP]                   = 12 + 8 + 28,
	[IB_OPCODE_RC_FETCH_ADD]                      = 12 + 8 + 28,
325 326
	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = 12 + 8 + 4,
	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = 12 + 8 + 4,
M
Mike Marciniszyn 已提交
327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367
	/* UC */
	[IB_OPCODE_UC_SEND_FIRST]                     = 12 + 8,
	[IB_OPCODE_UC_SEND_MIDDLE]                    = 12 + 8,
	[IB_OPCODE_UC_SEND_LAST]                      = 12 + 8,
	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
	[IB_OPCODE_UC_SEND_ONLY]                      = 12 + 8,
	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = 12 + 8,
	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = 12 + 8,
	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
	/* UD */
	[IB_OPCODE_UD_SEND_ONLY]                      = 12 + 8 + 8,
	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 12
};

static const opcode_handler opcode_handler_tbl[256] = {
	/* RC */
	[IB_OPCODE_RC_SEND_FIRST]                     = &hfi1_rc_rcv,
	[IB_OPCODE_RC_SEND_MIDDLE]                    = &hfi1_rc_rcv,
	[IB_OPCODE_RC_SEND_LAST]                      = &hfi1_rc_rcv,
	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
	[IB_OPCODE_RC_SEND_ONLY]                      = &hfi1_rc_rcv,
	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = &hfi1_rc_rcv,
	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = &hfi1_rc_rcv,
	[IB_OPCODE_RC_ACKNOWLEDGE]                    = &hfi1_rc_rcv,
	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = &hfi1_rc_rcv,
	[IB_OPCODE_RC_COMPARE_SWAP]                   = &hfi1_rc_rcv,
	[IB_OPCODE_RC_FETCH_ADD]                      = &hfi1_rc_rcv,
368 369
	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = &hfi1_rc_rcv,
	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = &hfi1_rc_rcv,
M
Mike Marciniszyn 已提交
370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
	/* UC */
	[IB_OPCODE_UC_SEND_FIRST]                     = &hfi1_uc_rcv,
	[IB_OPCODE_UC_SEND_MIDDLE]                    = &hfi1_uc_rcv,
	[IB_OPCODE_UC_SEND_LAST]                      = &hfi1_uc_rcv,
	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
	[IB_OPCODE_UC_SEND_ONLY]                      = &hfi1_uc_rcv,
	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = &hfi1_uc_rcv,
	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = &hfi1_uc_rcv,
	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = &hfi1_uc_rcv,
	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = &hfi1_uc_rcv,
	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
	/* UD */
	[IB_OPCODE_UD_SEND_ONLY]                      = &hfi1_ud_rcv,
	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_ud_rcv,
	/* CNP */
	[IB_OPCODE_CNP]				      = &hfi1_cnp_rcv
};

390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411
#define OPMASK 0x1f

static const u32 pio_opmask[BIT(3)] = {
	/* RC */
	[IB_OPCODE_RC >> 5] =
		BIT(RC_OP(SEND_ONLY) & OPMASK) |
		BIT(RC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
		BIT(RC_OP(RDMA_WRITE_ONLY) & OPMASK) |
		BIT(RC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK) |
		BIT(RC_OP(RDMA_READ_REQUEST) & OPMASK) |
		BIT(RC_OP(ACKNOWLEDGE) & OPMASK) |
		BIT(RC_OP(ATOMIC_ACKNOWLEDGE) & OPMASK) |
		BIT(RC_OP(COMPARE_SWAP) & OPMASK) |
		BIT(RC_OP(FETCH_ADD) & OPMASK),
	/* UC */
	[IB_OPCODE_UC >> 5] =
		BIT(UC_OP(SEND_ONLY) & OPMASK) |
		BIT(UC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
		BIT(UC_OP(RDMA_WRITE_ONLY) & OPMASK) |
		BIT(UC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK),
};

M
Mike Marciniszyn 已提交
412 413 414 415 416 417 418 419 420 421
/*
 * System image GUID.
 */
__be64 ib_hfi1_sys_image_guid;

/**
 * hfi1_copy_sge - copy data to SGE memory
 * @ss: the SGE state
 * @data: the data to copy
 * @length: the length of the data
422
 * @copy_last: do a separate copy of the last 8 bytes
M
Mike Marciniszyn 已提交
423 424
 */
void hfi1_copy_sge(
425
	struct rvt_sge_state *ss,
M
Mike Marciniszyn 已提交
426
	void *data, u32 length,
427 428
	int release,
	int copy_last)
M
Mike Marciniszyn 已提交
429
{
430
	struct rvt_sge *sge = &ss->sge;
431 432
	int in_last = 0;
	int i;
433 434 435 436 437 438 439 440 441 442 443 444 445 446
	int cacheless_copy = 0;

	if (sge_copy_mode == COPY_CACHELESS) {
		cacheless_copy = length >= PAGE_SIZE;
	} else if (sge_copy_mode == COPY_ADAPTIVE) {
		if (length >= PAGE_SIZE) {
			/*
			 * NOTE: this *assumes*:
			 * o The first vaddr is the dest.
			 * o If multiple pages, then vaddr is sequential.
			 */
			wss_insert(sge->vaddr);
			if (length >= (2 * PAGE_SIZE))
				wss_insert(sge->vaddr + PAGE_SIZE);
M
Mike Marciniszyn 已提交
447

448 449 450 451 452
			cacheless_copy = wss_exceeds_threshold();
		} else {
			wss_advance_clean_counter();
		}
	}
453 454 455 456 457 458 459 460 461 462
	if (copy_last) {
		if (length > 8) {
			length -= 8;
		} else {
			copy_last = 0;
			in_last = 1;
		}
	}

again:
M
Mike Marciniszyn 已提交
463 464 465 466 467 468 469 470
	while (length) {
		u32 len = sge->length;

		if (len > length)
			len = length;
		if (len > sge->sge_length)
			len = sge->sge_length;
		WARN_ON_ONCE(len == 0);
471 472
		if (unlikely(in_last)) {
			/* enforce byte transfer ordering */
473 474
			for (i = 0; i < len; i++)
				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
475 476
		} else if (cacheless_copy) {
			cacheless_memcpy(sge->vaddr, data, len);
477 478 479
		} else {
			memcpy(sge->vaddr, data, len);
		}
M
Mike Marciniszyn 已提交
480 481 482 483 484
		sge->vaddr += len;
		sge->length -= len;
		sge->sge_length -= len;
		if (sge->sge_length == 0) {
			if (release)
485
				rvt_put_mr(sge->mr);
M
Mike Marciniszyn 已提交
486 487 488
			if (--ss->num_sge)
				*sge = *ss->sg_list++;
		} else if (sge->length == 0 && sge->mr->lkey) {
489
			if (++sge->n >= RVT_SEGSZ) {
M
Mike Marciniszyn 已提交
490 491 492 493 494 495 496 497 498 499 500 501
				if (++sge->m >= sge->mr->mapsz)
					break;
				sge->n = 0;
			}
			sge->vaddr =
				sge->mr->map[sge->m]->segs[sge->n].vaddr;
			sge->length =
				sge->mr->map[sge->m]->segs[sge->n].length;
		}
		data += len;
		length -= len;
	}
502 503 504 505 506 507 508

	if (copy_last) {
		copy_last = 0;
		in_last = 1;
		length = 8;
		goto again;
	}
M
Mike Marciniszyn 已提交
509 510 511 512 513 514 515
}

/**
 * hfi1_skip_sge - skip over SGE memory
 * @ss: the SGE state
 * @length: the number of bytes to skip
 */
516
void hfi1_skip_sge(struct rvt_sge_state *ss, u32 length, int release)
M
Mike Marciniszyn 已提交
517
{
518
	struct rvt_sge *sge = &ss->sge;
M
Mike Marciniszyn 已提交
519 520 521 522 523 524 525 526 527 528 529 530 531 532

	while (length) {
		u32 len = sge->length;

		if (len > length)
			len = length;
		if (len > sge->sge_length)
			len = sge->sge_length;
		WARN_ON_ONCE(len == 0);
		sge->vaddr += len;
		sge->length -= len;
		sge->sge_length -= len;
		if (sge->sge_length == 0) {
			if (release)
533
				rvt_put_mr(sge->mr);
M
Mike Marciniszyn 已提交
534 535 536
			if (--ss->num_sge)
				*sge = *ss->sg_list++;
		} else if (sge->length == 0 && sge->mr->lkey) {
537
			if (++sge->n >= RVT_SEGSZ) {
M
Mike Marciniszyn 已提交
538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553
				if (++sge->m >= sge->mr->mapsz)
					break;
				sge->n = 0;
			}
			sge->vaddr =
				sge->mr->map[sge->m]->segs[sge->n].vaddr;
			sge->length =
				sge->mr->map[sge->m]->segs[sge->n].length;
		}
		length -= len;
	}
}

/*
 * Make sure the QP is ready and able to accept the given opcode.
 */
554
static inline opcode_handler qp_ok(int opcode, struct hfi1_packet *packet)
M
Mike Marciniszyn 已提交
555
{
556
	if (!(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
557
		return NULL;
558
	if (((opcode & RVT_OPCODE_QP_MASK) == packet->qp->allowed_ops) ||
M
Mike Marciniszyn 已提交
559
	    (opcode == IB_OPCODE_CNP))
560 561 562
		return opcode_handler_tbl[opcode];

	return NULL;
M
Mike Marciniszyn 已提交
563 564 565 566 567 568 569 570 571 572 573 574 575
}

/**
 * hfi1_ib_rcv - process an incoming packet
 * @packet: data packet information
 *
 * This is called to process an incoming packet at interrupt level.
 *
 * Tlen is the length of the header + data + CRC in bytes.
 */
void hfi1_ib_rcv(struct hfi1_packet *packet)
{
	struct hfi1_ctxtdata *rcd = packet->rcd;
576
	struct ib_header *hdr = packet->hdr;
M
Mike Marciniszyn 已提交
577 578
	u32 tlen = packet->tlen;
	struct hfi1_pportdata *ppd = rcd->ppd;
579
	struct hfi1_ibport *ibp = rcd_to_iport(rcd);
580
	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
581
	opcode_handler packet_handler;
582
	unsigned long flags;
M
Mike Marciniszyn 已提交
583 584 585 586 587 588 589
	u32 qp_num;
	int lnh;
	u8 opcode;
	u16 lid;

	/* Check for GRH */
	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
590
	if (lnh == HFI1_LRH_BTH) {
M
Mike Marciniszyn 已提交
591
		packet->ohdr = &hdr->u.oth;
592
	} else if (lnh == HFI1_LRH_GRH) {
M
Mike Marciniszyn 已提交
593 594 595 596 597 598 599 600 601
		u32 vtf;

		packet->ohdr = &hdr->u.l.oth;
		if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
			goto drop;
		vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
		if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
			goto drop;
		packet->rcv_flags |= HFI1_HAS_GRH;
602
	} else {
M
Mike Marciniszyn 已提交
603
		goto drop;
604
	}
M
Mike Marciniszyn 已提交
605 606 607 608 609 610 611

	trace_input_ibhdr(rcd->dd, hdr);

	opcode = (be32_to_cpu(packet->ohdr->bth[0]) >> 24);
	inc_opstats(tlen, &rcd->opstats->stats[opcode]);

	/* Get the destination QP number. */
612
	qp_num = be32_to_cpu(packet->ohdr->bth[1]) & RVT_QPN_MASK;
M
Mike Marciniszyn 已提交
613
	lid = be16_to_cpu(hdr->lrh[1]);
614 615
	if (unlikely((lid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
		     (lid != be16_to_cpu(IB_LID_PERMISSIVE)))) {
616 617
		struct rvt_mcast *mcast;
		struct rvt_mcast_qp *p;
M
Mike Marciniszyn 已提交
618 619 620

		if (lnh != HFI1_LRH_GRH)
			goto drop;
621
		mcast = rvt_mcast_find(&ibp->rvp, &hdr->u.l.grh.dgid);
622
		if (!mcast)
M
Mike Marciniszyn 已提交
623 624 625
			goto drop;
		list_for_each_entry_rcu(p, &mcast->qp_list, list) {
			packet->qp = p->qp;
626
			spin_lock_irqsave(&packet->qp->r_lock, flags);
627 628 629 630 631
			packet_handler = qp_ok(opcode, packet);
			if (likely(packet_handler))
				packet_handler(packet);
			else
				ibp->rvp.n_pkt_drops++;
632
			spin_unlock_irqrestore(&packet->qp->r_lock, flags);
M
Mike Marciniszyn 已提交
633 634
		}
		/*
635
		 * Notify rvt_multicast_detach() if it is waiting for us
M
Mike Marciniszyn 已提交
636 637 638 639 640 641
		 * to finish.
		 */
		if (atomic_dec_return(&mcast->refcount) <= 1)
			wake_up(&mcast->wait);
	} else {
		rcu_read_lock();
642
		packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
M
Mike Marciniszyn 已提交
643 644 645 646
		if (!packet->qp) {
			rcu_read_unlock();
			goto drop;
		}
647
		spin_lock_irqsave(&packet->qp->r_lock, flags);
648 649 650 651 652
		packet_handler = qp_ok(opcode, packet);
		if (likely(packet_handler))
			packet_handler(packet);
		else
			ibp->rvp.n_pkt_drops++;
653
		spin_unlock_irqrestore(&packet->qp->r_lock, flags);
M
Mike Marciniszyn 已提交
654 655 656 657 658
		rcu_read_unlock();
	}
	return;

drop:
659
	ibp->rvp.n_pkt_drops++;
M
Mike Marciniszyn 已提交
660 661 662 663 664 665 666 667 668 669
}

/*
 * This is called from a timer to check for QPs
 * which need kernel memory in order to send a packet.
 */
static void mem_timer(unsigned long data)
{
	struct hfi1_ibdev *dev = (struct hfi1_ibdev *)data;
	struct list_head *list = &dev->memwait;
670
	struct rvt_qp *qp = NULL;
M
Mike Marciniszyn 已提交
671 672
	struct iowait *wait;
	unsigned long flags;
673
	struct hfi1_qp_priv *priv;
M
Mike Marciniszyn 已提交
674 675 676 677

	write_seqlock_irqsave(&dev->iowait_lock, flags);
	if (!list_empty(list)) {
		wait = list_first_entry(list, struct iowait, list);
678 679 680
		qp = iowait_to_qp(wait);
		priv = qp->priv;
		list_del_init(&priv->s_iowait.list);
681
		priv->s_iowait.lock = NULL;
M
Mike Marciniszyn 已提交
682 683 684 685 686 687 688
		/* refcount held until actual wake up */
		if (!list_empty(list))
			mod_timer(&dev->mem_timer, jiffies + 1);
	}
	write_sequnlock_irqrestore(&dev->iowait_lock, flags);

	if (qp)
689
		hfi1_qp_wakeup(qp, RVT_S_WAIT_KMEM);
M
Mike Marciniszyn 已提交
690 691
}

692
void update_sge(struct rvt_sge_state *ss, u32 length)
M
Mike Marciniszyn 已提交
693
{
694
	struct rvt_sge *sge = &ss->sge;
M
Mike Marciniszyn 已提交
695 696 697 698 699 700 701 702

	sge->vaddr += length;
	sge->length -= length;
	sge->sge_length -= length;
	if (sge->sge_length == 0) {
		if (--ss->num_sge)
			*sge = *ss->sg_list++;
	} else if (sge->length == 0 && sge->mr->lkey) {
703
		if (++sge->n >= RVT_SEGSZ) {
M
Mike Marciniszyn 已提交
704 705 706 707 708 709 710 711 712 713 714 715 716 717 718
			if (++sge->m >= sge->mr->mapsz)
				return;
			sge->n = 0;
		}
		sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr;
		sge->length = sge->mr->map[sge->m]->segs[sge->n].length;
	}
}

/*
 * This is called with progress side lock held.
 */
/* New API */
static void verbs_sdma_complete(
	struct sdma_txreq *cookie,
719
	int status)
M
Mike Marciniszyn 已提交
720 721 722
{
	struct verbs_txreq *tx =
		container_of(cookie, struct verbs_txreq, txreq);
723
	struct rvt_qp *qp = tx->qp;
M
Mike Marciniszyn 已提交
724 725

	spin_lock(&qp->s_lock);
726
	if (tx->wqe) {
M
Mike Marciniszyn 已提交
727
		hfi1_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
728
	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
729
		struct ib_header *hdr;
M
Mike Marciniszyn 已提交
730 731 732 733 734 735 736 737 738

		hdr = &tx->phdr.hdr;
		hfi1_rc_send_complete(qp, hdr);
	}
	spin_unlock(&qp->s_lock);

	hfi1_put_txreq(tx);
}

739 740 741
static int wait_kmem(struct hfi1_ibdev *dev,
		     struct rvt_qp *qp,
		     struct hfi1_pkt_state *ps)
M
Mike Marciniszyn 已提交
742
{
743
	struct hfi1_qp_priv *priv = qp->priv;
M
Mike Marciniszyn 已提交
744 745 746 747
	unsigned long flags;
	int ret = 0;

	spin_lock_irqsave(&qp->s_lock, flags);
748
	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
M
Mike Marciniszyn 已提交
749
		write_seqlock(&dev->iowait_lock);
750 751
		list_add_tail(&ps->s_txreq->txreq.list,
			      &priv->s_iowait.tx_head);
752
		if (list_empty(&priv->s_iowait.list)) {
M
Mike Marciniszyn 已提交
753 754
			if (list_empty(&dev->memwait))
				mod_timer(&dev->mem_timer, jiffies + 1);
755
			qp->s_flags |= RVT_S_WAIT_KMEM;
756
			list_add_tail(&priv->s_iowait.list, &dev->memwait);
757
			priv->s_iowait.lock = &dev->iowait_lock;
758
			trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM);
759
			rvt_get_qp(qp);
M
Mike Marciniszyn 已提交
760 761
		}
		write_sequnlock(&dev->iowait_lock);
762
		qp->s_flags &= ~RVT_S_BUSY;
M
Mike Marciniszyn 已提交
763 764 765 766 767 768 769 770 771 772 773 774
		ret = -EBUSY;
	}
	spin_unlock_irqrestore(&qp->s_lock, flags);

	return ret;
}

/*
 * This routine calls txadds for each sg entry.
 *
 * Add failures will revert the sge cursor
 */
775
static noinline int build_verbs_ulp_payload(
M
Mike Marciniszyn 已提交
776 777 778 779
	struct sdma_engine *sde,
	u32 length,
	struct verbs_txreq *tx)
{
780
	struct rvt_sge_state *ss = tx->ss;
781 782
	struct rvt_sge *sg_list = ss->sg_list;
	struct rvt_sge sge = ss->sge;
M
Mike Marciniszyn 已提交
783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818
	u8 num_sge = ss->num_sge;
	u32 len;
	int ret = 0;

	while (length) {
		len = ss->sge.length;
		if (len > length)
			len = length;
		if (len > ss->sge.sge_length)
			len = ss->sge.sge_length;
		WARN_ON_ONCE(len == 0);
		ret = sdma_txadd_kvaddr(
			sde->dd,
			&tx->txreq,
			ss->sge.vaddr,
			len);
		if (ret)
			goto bail_txadd;
		update_sge(ss, len);
		length -= len;
	}
	return ret;
bail_txadd:
	/* unwind cursor */
	ss->sge = sge;
	ss->num_sge = num_sge;
	ss->sg_list = sg_list;
	return ret;
}

/*
 * Build the number of DMA descriptors needed to send length bytes of data.
 *
 * NOTE: DMA mapping is held in the tx until completed in the ring or
 *       the tx desc is freed without having been submitted to the ring
 *
819
 * This routine ensures all the helper routine calls succeed.
M
Mike Marciniszyn 已提交
820 821 822 823 824 825
 */
/* New API */
static int build_verbs_tx_desc(
	struct sdma_engine *sde,
	u32 length,
	struct verbs_txreq *tx,
826
	struct hfi1_ahg_info *ahg_info,
M
Mike Marciniszyn 已提交
827 828 829
	u64 pbc)
{
	int ret = 0;
830
	struct hfi1_sdma_header *phdr = &tx->phdr;
M
Mike Marciniszyn 已提交
831 832
	u16 hdrbytes = tx->hdr_dwords << 2;

833
	if (!ahg_info->ahgcount) {
M
Mike Marciniszyn 已提交
834 835
		ret = sdma_txinit_ahg(
			&tx->txreq,
836
			ahg_info->tx_flags,
M
Mike Marciniszyn 已提交
837
			hdrbytes + length,
838
			ahg_info->ahgidx,
M
Mike Marciniszyn 已提交
839 840 841 842 843 844 845 846 847 848
			0,
			NULL,
			0,
			verbs_sdma_complete);
		if (ret)
			goto bail_txadd;
		phdr->pbc = cpu_to_le64(pbc);
		ret = sdma_txadd_kvaddr(
			sde->dd,
			&tx->txreq,
849 850
			phdr,
			hdrbytes);
M
Mike Marciniszyn 已提交
851 852 853 854 855
		if (ret)
			goto bail_txadd;
	} else {
		ret = sdma_txinit_ahg(
			&tx->txreq,
856
			ahg_info->tx_flags,
M
Mike Marciniszyn 已提交
857
			length,
858 859 860
			ahg_info->ahgidx,
			ahg_info->ahgcount,
			ahg_info->ahgdesc,
M
Mike Marciniszyn 已提交
861 862 863 864 865 866
			hdrbytes,
			verbs_sdma_complete);
		if (ret)
			goto bail_txadd;
	}

867 868 869
	/* add the ulp payload - if any. tx->ss can be NULL for acks */
	if (tx->ss)
		ret = build_verbs_ulp_payload(sde, length, tx);
M
Mike Marciniszyn 已提交
870 871 872 873
bail_txadd:
	return ret;
}

874
int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
875
			u64 pbc)
M
Mike Marciniszyn 已提交
876
{
877
	struct hfi1_qp_priv *priv = qp->priv;
878
	struct hfi1_ahg_info *ahg_info = priv->s_ahg;
879
	u32 hdrwords = qp->s_hdrwords;
880
	u32 len = ps->s_txreq->s_cur_size;
881 882 883
	u32 plen = hdrwords + ((len + 3) >> 2) + 2; /* includes pbc */
	struct hfi1_ibdev *dev = ps->dev;
	struct hfi1_pportdata *ppd = ps->ppd;
M
Mike Marciniszyn 已提交
884 885
	struct verbs_txreq *tx;
	u64 pbc_flags = 0;
886 887
	u8 sc5 = priv->s_sc;

M
Mike Marciniszyn 已提交
888 889
	int ret;

890
	tx = ps->s_txreq;
891 892 893 894 895 896 897 898 899 900 901 902 903 904
	if (!sdma_txreq_built(&tx->txreq)) {
		if (likely(pbc == 0)) {
			u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
			/* No vl15 here */
			/* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
			pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;

			pbc = create_pbc(ppd,
					 pbc_flags,
					 qp->srate_mbps,
					 vl,
					 plen);
		}
		tx->wqe = qp->s_wqe;
905
		ret = build_verbs_tx_desc(tx->sde, len, tx, ahg_info, pbc);
906 907
		if (unlikely(ret))
			goto bail_build;
M
Mike Marciniszyn 已提交
908
	}
909 910 911 912 913 914
	ret =  sdma_send_txreq(tx->sde, &priv->s_iowait, &tx->txreq);
	if (unlikely(ret < 0)) {
		if (ret == -ECOMM)
			goto bail_ecomm;
		return ret;
	}
915 916
	trace_sdma_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
				&ps->s_txreq->phdr.hdr);
M
Mike Marciniszyn 已提交
917 918 919 920 921 922
	return ret;

bail_ecomm:
	/* The current one got "sent" */
	return 0;
bail_build:
923 924 925 926 927 928 929
	ret = wait_kmem(dev, qp, ps);
	if (!ret) {
		/* free txreq - bad state */
		hfi1_put_txreq(ps->s_txreq);
		ps->s_txreq = NULL;
	}
	return ret;
M
Mike Marciniszyn 已提交
930 931 932 933 934 935
}

/*
 * If we are now in the error state, return zero to flush the
 * send work request.
 */
936 937 938 939
static int pio_wait(struct rvt_qp *qp,
		    struct send_context *sc,
		    struct hfi1_pkt_state *ps,
		    u32 flag)
M
Mike Marciniszyn 已提交
940
{
941
	struct hfi1_qp_priv *priv = qp->priv;
M
Mike Marciniszyn 已提交
942 943 944 945 946 947 948 949 950 951 952 953
	struct hfi1_devdata *dd = sc->dd;
	struct hfi1_ibdev *dev = &dd->verbs_dev;
	unsigned long flags;
	int ret = 0;

	/*
	 * Note that as soon as want_buffer() is called and
	 * possibly before it returns, sc_piobufavail()
	 * could be called. Therefore, put QP on the I/O wait list before
	 * enabling the PIO avail interrupt.
	 */
	spin_lock_irqsave(&qp->s_lock, flags);
954
	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
M
Mike Marciniszyn 已提交
955
		write_seqlock(&dev->iowait_lock);
956 957
		list_add_tail(&ps->s_txreq->txreq.list,
			      &priv->s_iowait.tx_head);
958
		if (list_empty(&priv->s_iowait.list)) {
M
Mike Marciniszyn 已提交
959 960 961
			struct hfi1_ibdev *dev = &dd->verbs_dev;
			int was_empty;

962 963 964
			dev->n_piowait += !!(flag & RVT_S_WAIT_PIO);
			dev->n_piodrain += !!(flag & RVT_S_WAIT_PIO_DRAIN);
			qp->s_flags |= flag;
M
Mike Marciniszyn 已提交
965
			was_empty = list_empty(&sc->piowait);
966
			list_add_tail(&priv->s_iowait.list, &sc->piowait);
967
			priv->s_iowait.lock = &dev->iowait_lock;
968
			trace_hfi1_qpsleep(qp, RVT_S_WAIT_PIO);
969
			rvt_get_qp(qp);
M
Mike Marciniszyn 已提交
970 971 972 973 974
			/* counting: only call wantpiobuf_intr if first user */
			if (was_empty)
				hfi1_sc_wantpiobuf_intr(sc, 1);
		}
		write_sequnlock(&dev->iowait_lock);
975
		qp->s_flags &= ~RVT_S_BUSY;
M
Mike Marciniszyn 已提交
976 977 978 979 980 981
		ret = -EBUSY;
	}
	spin_unlock_irqrestore(&qp->s_lock, flags);
	return ret;
}

982 983 984 985 986 987 988 989 990
static void verbs_pio_complete(void *arg, int code)
{
	struct rvt_qp *qp = (struct rvt_qp *)arg;
	struct hfi1_qp_priv *priv = qp->priv;

	if (iowait_pio_dec(&priv->s_iowait))
		iowait_drain_wakeup(&priv->s_iowait);
}

991
int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
992
			u64 pbc)
M
Mike Marciniszyn 已提交
993
{
994
	struct hfi1_qp_priv *priv = qp->priv;
995
	u32 hdrwords = qp->s_hdrwords;
996
	struct rvt_sge_state *ss = ps->s_txreq->ss;
997
	u32 len = ps->s_txreq->s_cur_size;
998 999 1000
	u32 dwords = (len + 3) >> 2;
	u32 plen = hdrwords + dwords + 2; /* includes pbc */
	struct hfi1_pportdata *ppd = ps->ppd;
1001
	u32 *hdr = (u32 *)&ps->s_txreq->phdr.hdr;
M
Mike Marciniszyn 已提交
1002
	u64 pbc_flags = 0;
1003
	u8 sc5;
M
Mike Marciniszyn 已提交
1004 1005 1006 1007
	unsigned long flags = 0;
	struct send_context *sc;
	struct pio_buf *pbuf;
	int wc_status = IB_WC_SUCCESS;
1008
	int ret = 0;
1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019
	pio_release_cb cb = NULL;

	/* only RC/UC use complete */
	switch (qp->ibqp.qp_type) {
	case IB_QPT_RC:
	case IB_QPT_UC:
		cb = verbs_pio_complete;
		break;
	default:
		break;
	}
M
Mike Marciniszyn 已提交
1020 1021

	/* vl15 special case taken care of in ud.c */
1022
	sc5 = priv->s_sc;
1023
	sc = ps->s_txreq->psc;
M
Mike Marciniszyn 已提交
1024 1025

	if (likely(pbc == 0)) {
1026
		u8 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
M
Mike Marciniszyn 已提交
1027 1028 1029 1030
		/* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
		pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
		pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
	}
1031 1032 1033
	if (cb)
		iowait_pio_inc(&priv->s_iowait);
	pbuf = sc_buffer_alloc(sc, plen, cb, qp);
1034
	if (unlikely(!pbuf)) {
1035 1036
		if (cb)
			verbs_pio_complete(qp, 0);
M
Mike Marciniszyn 已提交
1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055
		if (ppd->host_link_state != HLS_UP_ACTIVE) {
			/*
			 * If we have filled the PIO buffers to capacity and are
			 * not in an active state this request is not going to
			 * go out to so just complete it with an error or else a
			 * ULP or the core may be stuck waiting.
			 */
			hfi1_cdbg(
				PIO,
				"alloc failed. state not active, completing");
			wc_status = IB_WC_GENERAL_ERR;
			goto pio_bail;
		} else {
			/*
			 * This is a normal occurrence. The PIO buffs are full
			 * up but we are still happily sending, well we could be
			 * so lets continue to queue the request.
			 */
			hfi1_cdbg(PIO, "alloc failed. state active, queuing");
1056
			ret = pio_wait(qp, sc, ps, RVT_S_WAIT_PIO);
1057
			if (!ret)
1058
				/* txreq not queued - free */
1059 1060 1061
				goto bail;
			/* tx consumed in wait */
			return ret;
M
Mike Marciniszyn 已提交
1062 1063 1064 1065 1066 1067 1068
		}
	}

	if (len == 0) {
		pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
	} else {
		if (ss) {
1069
			seg_pio_copy_start(pbuf, pbc, hdr, hdrwords * 4);
M
Mike Marciniszyn 已提交
1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
			while (len) {
				void *addr = ss->sge.vaddr;
				u32 slen = ss->sge.length;

				if (slen > len)
					slen = len;
				update_sge(ss, slen);
				seg_pio_copy_mid(pbuf, addr, slen);
				len -= slen;
			}
			seg_pio_copy_end(pbuf);
		}
	}

1084 1085
	trace_pio_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
			       &ps->s_txreq->phdr.hdr);
M
Mike Marciniszyn 已提交
1086 1087 1088 1089 1090 1091 1092 1093

pio_bail:
	if (qp->s_wqe) {
		spin_lock_irqsave(&qp->s_lock, flags);
		hfi1_send_complete(qp, qp->s_wqe, wc_status);
		spin_unlock_irqrestore(&qp->s_lock, flags);
	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
		spin_lock_irqsave(&qp->s_lock, flags);
1094
		hfi1_rc_send_complete(qp, &ps->s_txreq->phdr.hdr);
M
Mike Marciniszyn 已提交
1095 1096
		spin_unlock_irqrestore(&qp->s_lock, flags);
	}
1097 1098 1099 1100 1101 1102

	ret = 0;

bail:
	hfi1_put_txreq(ps->s_txreq);
	return ret;
M
Mike Marciniszyn 已提交
1103
}
1104

M
Mike Marciniszyn 已提交
1105 1106
/*
 * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1107
 * being an entry from the partition key table), return 0
M
Mike Marciniszyn 已提交
1108 1109 1110 1111 1112 1113
 * otherwise. Use the matching criteria for egress partition keys
 * specified in the OPAv1 spec., section 9.1l.7.
 */
static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
{
	u16 mkey = pkey & PKEY_LOW_15_MASK;
1114
	u16 mentry = ent & PKEY_LOW_15_MASK;
M
Mike Marciniszyn 已提交
1115

1116
	if (mkey == mentry) {
M
Mike Marciniszyn 已提交
1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128
		/*
		 * If pkey[15] is set (full partition member),
		 * is bit 15 in the corresponding table element
		 * clear (limited member)?
		 */
		if (pkey & PKEY_MEMBER_MASK)
			return !!(ent & PKEY_MEMBER_MASK);
		return 1;
	}
	return 0;
}

1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141
/**
 * egress_pkey_check - check P_KEY of a packet
 * @ppd:    Physical IB port data
 * @lrh: Local route header
 * @bth: Base transport header
 * @sc5:    SC for packet
 * @s_pkey_index: It will be used for look up optimization for kernel contexts
 * only. If it is negative value, then it means user contexts is calling this
 * function.
 *
 * It checks if hdr's pkey is valid.
 *
 * Return: 0 on success, otherwise, 1
M
Mike Marciniszyn 已提交
1142
 */
1143 1144
int egress_pkey_check(struct hfi1_pportdata *ppd, __be16 *lrh, __be32 *bth,
		      u8 sc5, int8_t s_pkey_index)
M
Mike Marciniszyn 已提交
1145 1146
{
	struct hfi1_devdata *dd;
1147
	int i;
M
Mike Marciniszyn 已提交
1148
	u16 pkey;
1149
	int is_user_ctxt_mechanism = (s_pkey_index < 0);
M
Mike Marciniszyn 已提交
1150 1151 1152 1153

	if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
		return 0;

1154
	pkey = (u16)be32_to_cpu(bth[0]);
M
Mike Marciniszyn 已提交
1155 1156 1157 1158 1159 1160 1161 1162 1163

	/* If SC15, pkey[0:14] must be 0x7fff */
	if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
		goto bad;

	/* Is the pkey = 0x0, or 0x8000? */
	if ((pkey & PKEY_LOW_15_MASK) == 0)
		goto bad;

1164 1165 1166 1167 1168 1169 1170
	/*
	 * For the kernel contexts only, if a qp is passed into the function,
	 * the most likely matching pkey has index qp->s_pkey_index
	 */
	if (!is_user_ctxt_mechanism &&
	    egress_pkey_matches_entry(pkey, ppd->pkeys[s_pkey_index])) {
		return 0;
M
Mike Marciniszyn 已提交
1171 1172
	}

1173 1174 1175 1176
	for (i = 0; i < MAX_PKEY_VALUES; i++) {
		if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
			return 0;
	}
M
Mike Marciniszyn 已提交
1177
bad:
1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194
	/*
	 * For the user-context mechanism, the P_KEY check would only happen
	 * once per SDMA request, not once per packet.  Therefore, there's no
	 * need to increment the counter for the user-context mechanism.
	 */
	if (!is_user_ctxt_mechanism) {
		incr_cntr64(&ppd->port_xmit_constraint_errors);
		dd = ppd->dd;
		if (!(dd->err_info_xmit_constraint.status &
		      OPA_EI_STATUS_SMASK)) {
			u16 slid = be16_to_cpu(lrh[3]);

			dd->err_info_xmit_constraint.status |=
				OPA_EI_STATUS_SMASK;
			dd->err_info_xmit_constraint.slid = slid;
			dd->err_info_xmit_constraint.pkey = pkey;
		}
M
Mike Marciniszyn 已提交
1195 1196 1197 1198
	}
	return 1;
}

1199 1200 1201 1202 1203 1204 1205
/**
 * get_send_routine - choose an egress routine
 *
 * Choose an egress routine based on QP type
 * and size
 */
static inline send_routine get_send_routine(struct rvt_qp *qp,
1206
					    struct verbs_txreq *tx)
1207 1208 1209
{
	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
	struct hfi1_qp_priv *priv = qp->priv;
1210
	struct ib_header *h = &tx->phdr.hdr;
1211 1212 1213 1214 1215 1216 1217 1218 1219 1220

	if (unlikely(!(dd->flags & HFI1_HAS_SEND_DMA)))
		return dd->process_pio_send;
	switch (qp->ibqp.qp_type) {
	case IB_QPT_SMI:
		return dd->process_pio_send;
	case IB_QPT_GSI:
	case IB_QPT_UD:
		break;
	case IB_QPT_UC:
1221 1222 1223
	case IB_QPT_RC: {
		u8 op = get_opcode(h);

1224
		if (piothreshold &&
1225
		    tx->s_cur_size <= min(piothreshold, qp->pmtu) &&
1226
		    (BIT(op & OPMASK) & pio_opmask[op >> 5]) &&
1227 1228
		    iowait_sdma_pending(&priv->s_iowait) == 0 &&
		    !sdma_txreq_built(&tx->txreq))
1229 1230
			return dd->process_pio_send;
		break;
1231
	}
1232 1233 1234 1235 1236 1237
	default:
		break;
	}
	return dd->process_dma_send;
}

M
Mike Marciniszyn 已提交
1238 1239 1240
/**
 * hfi1_verbs_send - send a packet
 * @qp: the QP to send on
1241
 * @ps: the state of the packet to send
M
Mike Marciniszyn 已提交
1242 1243
 *
 * Return zero if packet is sent or queued OK.
1244
 * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
M
Mike Marciniszyn 已提交
1245
 */
1246
int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
M
Mike Marciniszyn 已提交
1247 1248
{
	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1249
	struct hfi1_qp_priv *priv = qp->priv;
1250 1251
	struct ib_other_headers *ohdr;
	struct ib_header *hdr;
1252
	send_routine sr;
M
Mike Marciniszyn 已提交
1253
	int ret;
1254 1255 1256 1257 1258 1259 1260 1261 1262
	u8 lnh;

	hdr = &ps->s_txreq->phdr.hdr;
	/* locate the pkey within the headers */
	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
	if (lnh == HFI1_LRH_GRH)
		ohdr = &hdr->u.l.oth;
	else
		ohdr = &hdr->u.oth;
M
Mike Marciniszyn 已提交
1263

1264
	sr = get_send_routine(qp, ps->s_txreq);
1265 1266 1267 1268 1269
	ret = egress_pkey_check(dd->pport,
				hdr->lrh,
				ohdr->bth,
				priv->s_sc,
				qp->s_pkey_index);
M
Mike Marciniszyn 已提交
1270 1271 1272 1273 1274 1275 1276 1277 1278
	if (unlikely(ret)) {
		/*
		 * The value we are returning here does not get propagated to
		 * the verbs caller. Thus we need to complete the request with
		 * error otherwise the caller could be sitting waiting on the
		 * completion event. Only do this for PIO. SDMA has its own
		 * mechanism for handling the errors. So for SDMA we can just
		 * return.
		 */
1279 1280 1281
		if (sr == dd->process_pio_send) {
			unsigned long flags;

M
Mike Marciniszyn 已提交
1282 1283 1284 1285 1286 1287 1288 1289
			hfi1_cdbg(PIO, "%s() Failed. Completing with err",
				  __func__);
			spin_lock_irqsave(&qp->s_lock, flags);
			hfi1_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
			spin_unlock_irqrestore(&qp->s_lock, flags);
		}
		return -EINVAL;
	}
1290 1291 1292 1293 1294
	if (sr == dd->process_dma_send && iowait_pio_pending(&priv->s_iowait))
		return pio_wait(qp,
				ps->s_txreq->psc,
				ps,
				RVT_S_WAIT_PIO_DRAIN);
1295
	return sr(qp, ps, 0);
M
Mike Marciniszyn 已提交
1296 1297
}

1298 1299 1300 1301 1302
/**
 * hfi1_fill_device_attr - Fill in rvt dev info device attributes.
 * @dd: the device data structure
 */
static void hfi1_fill_device_attr(struct hfi1_devdata *dd)
M
Mike Marciniszyn 已提交
1303
{
1304
	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1305
	u16 ver = dd->dc8051_ver;
1306 1307 1308

	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));

1309 1310
	rdi->dparms.props.fw_ver = ((u64)(dc8051_ver_maj(ver)) << 16) |
				    (u64)dc8051_ver_min(ver);
1311 1312 1313
	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
			IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
			IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1314 1315
			IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE |
			IB_DEVICE_MEM_MGT_EXTENSIONS;
1316 1317 1318 1319 1320
	rdi->dparms.props.page_size_cap = PAGE_SIZE;
	rdi->dparms.props.vendor_id = dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
	rdi->dparms.props.vendor_part_id = dd->pcidev->device;
	rdi->dparms.props.hw_ver = dd->minrev;
	rdi->dparms.props.sys_image_guid = ib_hfi1_sys_image_guid;
1321 1322
	rdi->dparms.props.max_mr_size = U64_MAX;
	rdi->dparms.props.max_fast_reg_page_list_len = UINT_MAX;
1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345
	rdi->dparms.props.max_qp = hfi1_max_qps;
	rdi->dparms.props.max_qp_wr = hfi1_max_qp_wrs;
	rdi->dparms.props.max_sge = hfi1_max_sges;
	rdi->dparms.props.max_sge_rd = hfi1_max_sges;
	rdi->dparms.props.max_cq = hfi1_max_cqs;
	rdi->dparms.props.max_ah = hfi1_max_ahs;
	rdi->dparms.props.max_cqe = hfi1_max_cqes;
	rdi->dparms.props.max_mr = rdi->lkey_table.max;
	rdi->dparms.props.max_fmr = rdi->lkey_table.max;
	rdi->dparms.props.max_map_per_fmr = 32767;
	rdi->dparms.props.max_pd = hfi1_max_pds;
	rdi->dparms.props.max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
	rdi->dparms.props.max_qp_init_rd_atom = 255;
	rdi->dparms.props.max_srq = hfi1_max_srqs;
	rdi->dparms.props.max_srq_wr = hfi1_max_srq_wrs;
	rdi->dparms.props.max_srq_sge = hfi1_max_srq_sges;
	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
	rdi->dparms.props.max_pkeys = hfi1_get_npkeys(dd);
	rdi->dparms.props.max_mcast_grp = hfi1_max_mcast_grps;
	rdi->dparms.props.max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
	rdi->dparms.props.max_total_mcast_qp_attach =
					rdi->dparms.props.max_mcast_qp_attach *
					rdi->dparms.props.max_mcast_grp;
M
Mike Marciniszyn 已提交
1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378
}

static inline u16 opa_speed_to_ib(u16 in)
{
	u16 out = 0;

	if (in & OPA_LINK_SPEED_25G)
		out |= IB_SPEED_EDR;
	if (in & OPA_LINK_SPEED_12_5G)
		out |= IB_SPEED_FDR;

	return out;
}

/*
 * Convert a single OPA link width (no multiple flags) to an IB value.
 * A zero OPA link width means link down, which means the IB width value
 * is a don't care.
 */
static inline u16 opa_width_to_ib(u16 in)
{
	switch (in) {
	case OPA_LINK_WIDTH_1X:
	/* map 2x and 3x to 1x as they don't exist in IB */
	case OPA_LINK_WIDTH_2X:
	case OPA_LINK_WIDTH_3X:
		return IB_WIDTH_1X;
	default: /* link down or unknown, return our largest width */
	case OPA_LINK_WIDTH_4X:
		return IB_WIDTH_4X;
	}
}

1379
static int query_port(struct rvt_dev_info *rdi, u8 port_num,
M
Mike Marciniszyn 已提交
1380 1381
		      struct ib_port_attr *props)
{
1382 1383 1384
	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
M
Mike Marciniszyn 已提交
1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428
	u16 lid = ppd->lid;

	props->lid = lid ? lid : 0;
	props->lmc = ppd->lmc;
	/* OPA logical states match IB logical states */
	props->state = driver_lstate(ppd);
	props->phys_state = hfi1_ibphys_portstate(ppd);
	props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
	props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
	/* see rate_show() in ib core/sysfs.c */
	props->active_speed = (u8)opa_speed_to_ib(ppd->link_speed_active);
	props->max_vl_num = ppd->vls_supported;

	/* Once we are a "first class" citizen and have added the OPA MTUs to
	 * the core we can advertise the larger MTU enum to the ULPs, for now
	 * advertise only 4K.
	 *
	 * Those applications which are either OPA aware or pass the MTU enum
	 * from the Path Records to us will get the new 8k MTU.  Those that
	 * attempt to process the MTU enum may fail in various ways.
	 */
	props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
				      4096 : hfi1_max_mtu), IB_MTU_4096);
	props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
		mtu_to_enum(ppd->ibmtu, IB_MTU_2048);

	return 0;
}

static int modify_device(struct ib_device *device,
			 int device_modify_mask,
			 struct ib_device_modify *device_modify)
{
	struct hfi1_devdata *dd = dd_from_ibdev(device);
	unsigned i;
	int ret;

	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
				   IB_DEVICE_MODIFY_NODE_DESC)) {
		ret = -EOPNOTSUPP;
		goto bail;
	}

	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1429 1430
		memcpy(device->node_desc, device_modify->node_desc,
		       IB_DEVICE_NODE_DESC_MAX);
M
Mike Marciniszyn 已提交
1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453
		for (i = 0; i < dd->num_pports; i++) {
			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;

			hfi1_node_desc_chg(ibp);
		}
	}

	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
		ib_hfi1_sys_image_guid =
			cpu_to_be64(device_modify->sys_image_guid);
		for (i = 0; i < dd->num_pports; i++) {
			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;

			hfi1_sys_guid_chg(ibp);
		}
	}

	ret = 0;

bail:
	return ret;
}

1454
static int shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
M
Mike Marciniszyn 已提交
1455
{
1456 1457 1458 1459
	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
	int ret;
M
Mike Marciniszyn 已提交
1460

1461 1462 1463
	set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
			     OPA_LINKDOWN_REASON_UNKNOWN);
	ret = set_link_state(ppd, HLS_DN_DOWNDEF);
M
Mike Marciniszyn 已提交
1464 1465 1466
	return ret;
}

1467 1468
static int hfi1_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
			    int guid_index, __be64 *guid)
M
Mike Marciniszyn 已提交
1469
{
1470
	struct hfi1_ibport *ibp = container_of(rvp, struct hfi1_ibport, rvp);
M
Mike Marciniszyn 已提交
1471

1472
	if (guid_index >= HFI1_GUIDS_PER_PORT)
1473
		return -EINVAL;
M
Mike Marciniszyn 已提交
1474

1475
	*guid = get_sguid(ibp, guid_index);
1476
	return 0;
M
Mike Marciniszyn 已提交
1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488
}

/*
 * convert ah port,sl to sc
 */
u8 ah_to_sc(struct ib_device *ibdev, struct ib_ah_attr *ah)
{
	struct hfi1_ibport *ibp = to_iport(ibdev, ah->port_num);

	return ibp->sl_to_sc[ah->sl];
}

1489
static int hfi1_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr)
M
Mike Marciniszyn 已提交
1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501
{
	struct hfi1_ibport *ibp;
	struct hfi1_pportdata *ppd;
	struct hfi1_devdata *dd;
	u8 sc5;

	/* test the mapping for validity */
	ibp = to_iport(ibdev, ah_attr->port_num);
	ppd = ppd_from_ibp(ibp);
	sc5 = ibp->sl_to_sc[ah_attr->sl];
	dd = dd_from_ppd(ppd);
	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1502
		return -EINVAL;
M
Mike Marciniszyn 已提交
1503 1504 1505
	return 0;
}

1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528
static void hfi1_notify_new_ah(struct ib_device *ibdev,
			       struct ib_ah_attr *ah_attr,
			       struct rvt_ah *ah)
{
	struct hfi1_ibport *ibp;
	struct hfi1_pportdata *ppd;
	struct hfi1_devdata *dd;
	u8 sc5;

	/*
	 * Do not trust reading anything from rvt_ah at this point as it is not
	 * done being setup. We can however modify things which we need to set.
	 */

	ibp = to_iport(ibdev, ah_attr->port_num);
	ppd = ppd_from_ibp(ibp);
	sc5 = ibp->sl_to_sc[ah->attr.sl];
	dd = dd_from_ppd(ppd);
	ah->vl = sc_to_vlt(dd, sc5);
	if (ah->vl < num_vls || ah->vl == 15)
		ah->log_pmtu = ilog2(dd->vld[ah->vl].mtu);
}

M
Mike Marciniszyn 已提交
1529 1530 1531 1532
struct ib_ah *hfi1_create_qp0_ah(struct hfi1_ibport *ibp, u16 dlid)
{
	struct ib_ah_attr attr;
	struct ib_ah *ah = ERR_PTR(-EINVAL);
1533
	struct rvt_qp *qp0;
M
Mike Marciniszyn 已提交
1534 1535 1536 1537 1538

	memset(&attr, 0, sizeof(attr));
	attr.dlid = dlid;
	attr.port_num = ppd_from_ibp(ibp)->port;
	rcu_read_lock();
1539
	qp0 = rcu_dereference(ibp->rvp.qp[0]);
M
Mike Marciniszyn 已提交
1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565
	if (qp0)
		ah = ib_create_ah(qp0->ibqp.pd, &attr);
	rcu_read_unlock();
	return ah;
}

/**
 * hfi1_get_npkeys - return the size of the PKEY table for context 0
 * @dd: the hfi1_ib device
 */
unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
{
	return ARRAY_SIZE(dd->pport[0].pkeys);
}

static void init_ibport(struct hfi1_pportdata *ppd)
{
	struct hfi1_ibport *ibp = &ppd->ibport_data;
	size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
	int i;

	for (i = 0; i < sz; i++) {
		ibp->sl_to_sc[i] = i;
		ibp->sc_to_sl[i] = i;
	}

1566
	spin_lock_init(&ibp->rvp.lock);
M
Mike Marciniszyn 已提交
1567
	/* Set the prefix to the default value (see ch. 4.1.1) */
1568 1569
	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
	ibp->rvp.sm_lid = 0;
M
Mike Marciniszyn 已提交
1570
	/* Below should only set bits defined in OPA PortInfo.CapabilityMask */
1571
	ibp->rvp.port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
M
Mike Marciniszyn 已提交
1572
		IB_PORT_CAP_MASK_NOTICE_SUP;
1573 1574 1575 1576 1577 1578 1579 1580
	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;

	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
M
Mike Marciniszyn 已提交
1581 1582
}

1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593
static void hfi1_get_dev_fw_str(struct ib_device *ibdev, char *str,
				size_t str_len)
{
	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
	struct hfi1_ibdev *dev = dev_from_rdi(rdi);
	u16 ver = dd_from_dev(dev)->dc8051_ver;

	snprintf(str, str_len, "%u.%u", dc8051_ver_maj(ver),
		 dc8051_ver_min(ver));
}

1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741
static const char * const driver_cntr_names[] = {
	/* must be element 0*/
	"DRIVER_KernIntr",
	"DRIVER_ErrorIntr",
	"DRIVER_Tx_Errs",
	"DRIVER_Rcv_Errs",
	"DRIVER_HW_Errs",
	"DRIVER_NoPIOBufs",
	"DRIVER_CtxtsOpen",
	"DRIVER_RcvLen_Errs",
	"DRIVER_EgrBufFull",
	"DRIVER_EgrHdrFull"
};

static const char **dev_cntr_names;
static const char **port_cntr_names;
static int num_driver_cntrs = ARRAY_SIZE(driver_cntr_names);
static int num_dev_cntrs;
static int num_port_cntrs;
static int cntr_names_initialized;

/*
 * Convert a list of names separated by '\n' into an array of NULL terminated
 * strings. Optionally some entries can be reserved in the array to hold extra
 * external strings.
 */
static int init_cntr_names(const char *names_in,
			   const int names_len,
			   int num_extra_names,
			   int *num_cntrs,
			   const char ***cntr_names)
{
	char *names_out, *p, **q;
	int i, n;

	n = 0;
	for (i = 0; i < names_len; i++)
		if (names_in[i] == '\n')
			n++;

	names_out = kmalloc((n + num_extra_names) * sizeof(char *) + names_len,
			    GFP_KERNEL);
	if (!names_out) {
		*num_cntrs = 0;
		*cntr_names = NULL;
		return -ENOMEM;
	}

	p = names_out + (n + num_extra_names) * sizeof(char *);
	memcpy(p, names_in, names_len);

	q = (char **)names_out;
	for (i = 0; i < n; i++) {
		q[i] = p;
		p = strchr(p, '\n');
		*p++ = '\0';
	}

	*num_cntrs = n;
	*cntr_names = (const char **)names_out;
	return 0;
}

static struct rdma_hw_stats *alloc_hw_stats(struct ib_device *ibdev,
					    u8 port_num)
{
	int i, err;

	if (!cntr_names_initialized) {
		struct hfi1_devdata *dd = dd_from_ibdev(ibdev);

		err = init_cntr_names(dd->cntrnames,
				      dd->cntrnameslen,
				      num_driver_cntrs,
				      &num_dev_cntrs,
				      &dev_cntr_names);
		if (err)
			return NULL;

		for (i = 0; i < num_driver_cntrs; i++)
			dev_cntr_names[num_dev_cntrs + i] =
				driver_cntr_names[i];

		err = init_cntr_names(dd->portcntrnames,
				      dd->portcntrnameslen,
				      0,
				      &num_port_cntrs,
				      &port_cntr_names);
		if (err) {
			kfree(dev_cntr_names);
			dev_cntr_names = NULL;
			return NULL;
		}
		cntr_names_initialized = 1;
	}

	if (!port_num)
		return rdma_alloc_hw_stats_struct(
				dev_cntr_names,
				num_dev_cntrs + num_driver_cntrs,
				RDMA_HW_STATS_DEFAULT_LIFESPAN);
	else
		return rdma_alloc_hw_stats_struct(
				port_cntr_names,
				num_port_cntrs,
				RDMA_HW_STATS_DEFAULT_LIFESPAN);
}

static u64 hfi1_sps_ints(void)
{
	unsigned long flags;
	struct hfi1_devdata *dd;
	u64 sps_ints = 0;

	spin_lock_irqsave(&hfi1_devs_lock, flags);
	list_for_each_entry(dd, &hfi1_dev_list, list) {
		sps_ints += get_all_cpu_total(dd->int_counter);
	}
	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
	return sps_ints;
}

static int get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
			u8 port, int index)
{
	u64 *values;
	int count;

	if (!port) {
		u64 *stats = (u64 *)&hfi1_stats;
		int i;

		hfi1_read_cntrs(dd_from_ibdev(ibdev), NULL, &values);
		values[num_dev_cntrs] = hfi1_sps_ints();
		for (i = 1; i < num_driver_cntrs; i++)
			values[num_dev_cntrs + i] = stats[i];
		count = num_dev_cntrs + num_driver_cntrs;
	} else {
		struct hfi1_ibport *ibp = to_iport(ibdev, port);

		hfi1_read_portcntrs(ppd_from_ibp(ibp), NULL, &values);
		count = num_port_cntrs;
	}

	memcpy(stats->value, values, count * sizeof(u64));
	return count;
}

M
Mike Marciniszyn 已提交
1742 1743 1744 1745 1746 1747 1748 1749
/**
 * hfi1_register_ib_device - register our device with the infiniband core
 * @dd: the device data structure
 * Return 0 if successful, errno if unsuccessful.
 */
int hfi1_register_ib_device(struct hfi1_devdata *dd)
{
	struct hfi1_ibdev *dev = &dd->verbs_dev;
1750
	struct ib_device *ibdev = &dev->rdi.ibdev;
M
Mike Marciniszyn 已提交
1751
	struct hfi1_pportdata *ppd = dd->pport;
1752
	struct hfi1_ibport *ibp = &ppd->ibport_data;
1753
	unsigned i;
M
Mike Marciniszyn 已提交
1754 1755 1756 1757 1758 1759 1760
	int ret;
	size_t lcpysz = IB_DEVICE_NAME_MAX;

	for (i = 0; i < dd->num_pports; i++)
		init_ibport(ppd + i);

	/* Only need to initialize non-zero fields. */
1761

1762
	setup_timer(&dev->mem_timer, mem_timer, (unsigned long)dev);
M
Mike Marciniszyn 已提交
1763 1764

	seqlock_init(&dev->iowait_lock);
1765
	seqlock_init(&dev->txwait_lock);
M
Mike Marciniszyn 已提交
1766 1767 1768
	INIT_LIST_HEAD(&dev->txwait);
	INIT_LIST_HEAD(&dev->memwait);

1769 1770
	ret = verbs_txreq_init(dev);
	if (ret)
M
Mike Marciniszyn 已提交
1771 1772
		goto err_verbs_txreq;

1773 1774 1775
	/* Use first-port GUID as node guid */
	ibdev->node_guid = get_sguid(ibp, HFI1_PORT_GUID_INDEX);

M
Mike Marciniszyn 已提交
1776 1777 1778 1779 1780 1781
	/*
	 * The system image GUID is supposed to be the same for all
	 * HFIs in a single system but since there can be other
	 * device types in the system, we can't be sure this is unique.
	 */
	if (!ib_hfi1_sys_image_guid)
1782
		ib_hfi1_sys_image_guid = ibdev->node_guid;
M
Mike Marciniszyn 已提交
1783 1784 1785 1786 1787 1788
	lcpysz = strlcpy(ibdev->name, class_name(), lcpysz);
	strlcpy(ibdev->name + lcpysz, "_%d", IB_DEVICE_NAME_MAX - lcpysz);
	ibdev->owner = THIS_MODULE;
	ibdev->phys_port_cnt = dd->num_pports;
	ibdev->dma_device = &dd->pcidev->dev;
	ibdev->modify_device = modify_device;
1789 1790
	ibdev->alloc_hw_stats = alloc_hw_stats;
	ibdev->get_hw_stats = get_hw_stats;
1791 1792

	/* keep process mad in the driver */
M
Mike Marciniszyn 已提交
1793
	ibdev->process_mad = hfi1_process_mad;
1794
	ibdev->get_dev_fw_str = hfi1_get_dev_fw_str;
M
Mike Marciniszyn 已提交
1795 1796 1797 1798

	strncpy(ibdev->node_desc, init_utsname()->nodename,
		sizeof(ibdev->node_desc));

1799 1800 1801 1802
	/*
	 * Fill in rvt info object.
	 */
	dd->verbs_dev.rdi.driver_f.port_callback = hfi1_create_port_files;
1803 1804
	dd->verbs_dev.rdi.driver_f.get_card_name = get_card_name;
	dd->verbs_dev.rdi.driver_f.get_pci_dev = get_pci_dev;
1805
	dd->verbs_dev.rdi.driver_f.check_ah = hfi1_check_ah;
1806
	dd->verbs_dev.rdi.driver_f.notify_new_ah = hfi1_notify_new_ah;
1807
	dd->verbs_dev.rdi.driver_f.get_guid_be = hfi1_get_guid_be;
1808 1809 1810
	dd->verbs_dev.rdi.driver_f.query_port_state = query_port;
	dd->verbs_dev.rdi.driver_f.shut_down_port = shut_down_port;
	dd->verbs_dev.rdi.driver_f.cap_mask_chg = hfi1_cap_mask_chg;
1811 1812 1813 1814
	/*
	 * Fill in rvt info device attributes.
	 */
	hfi1_fill_device_attr(dd);
1815 1816 1817 1818 1819 1820 1821 1822

	/* queue pair */
	dd->verbs_dev.rdi.dparms.qp_table_size = hfi1_qp_table_size;
	dd->verbs_dev.rdi.dparms.qpn_start = 0;
	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
	dd->verbs_dev.rdi.dparms.qos_shift = dd->qos_shift;
	dd->verbs_dev.rdi.dparms.qpn_res_start = kdeth_qp << 16;
	dd->verbs_dev.rdi.dparms.qpn_res_end =
1823
	dd->verbs_dev.rdi.dparms.qpn_res_start + 65535;
1824 1825 1826 1827
	dd->verbs_dev.rdi.dparms.max_rdma_atomic = HFI1_MAX_RDMA_ATOMIC;
	dd->verbs_dev.rdi.dparms.psn_mask = PSN_MASK;
	dd->verbs_dev.rdi.dparms.psn_shift = PSN_SHIFT;
	dd->verbs_dev.rdi.dparms.psn_modify_mask = PSN_MODIFY_MASK;
1828 1829 1830
	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_INTEL_OPA;
	dd->verbs_dev.rdi.dparms.max_mad_size = OPA_MGMT_MAD_SIZE;

1831 1832 1833 1834
	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qp_priv_alloc;
	dd->verbs_dev.rdi.driver_f.qp_priv_free = qp_priv_free;
	dd->verbs_dev.rdi.driver_f.free_all_qps = free_all_qps;
	dd->verbs_dev.rdi.driver_f.notify_qp_reset = notify_qp_reset;
1835 1836
	dd->verbs_dev.rdi.driver_f.do_send = hfi1_do_send;
	dd->verbs_dev.rdi.driver_f.schedule_send = hfi1_schedule_send;
1837
	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _hfi1_schedule_send;
1838 1839 1840 1841 1842 1843 1844 1845 1846 1847
	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = get_pmtu_from_attr;
	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = flush_qp_waiters;
	dd->verbs_dev.rdi.driver_f.stop_send_queue = stop_send_queue;
	dd->verbs_dev.rdi.driver_f.quiesce_qp = quiesce_qp;
	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
	dd->verbs_dev.rdi.driver_f.mtu_from_qp = mtu_from_qp;
	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = mtu_to_path_mtu;
	dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
	dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
1848
	dd->verbs_dev.rdi.driver_f.notify_restart_rc = hfi1_restart_rc;
1849
	dd->verbs_dev.rdi.driver_f.check_send_wqe = hfi1_check_send_wqe;
1850

1851 1852 1853 1854
	/* completeion queue */
	snprintf(dd->verbs_dev.rdi.dparms.cq_name,
		 sizeof(dd->verbs_dev.rdi.dparms.cq_name),
		 "hfi1_cq%d", dd->unit);
1855
	dd->verbs_dev.rdi.dparms.node = dd->node;
1856

1857
	/* misc settings */
1858
	dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
1859
	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
1860 1861 1862
	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);

1863 1864 1865
	/* post send table */
	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;

1866 1867 1868 1869 1870 1871
	ppd = dd->pport;
	for (i = 0; i < dd->num_pports; i++, ppd++)
		rvt_init_port(&dd->verbs_dev.rdi,
			      &ppd->ibport_data.rvp,
			      i,
			      ppd->pkeys);
1872 1873

	ret = rvt_register_device(&dd->verbs_dev.rdi);
M
Mike Marciniszyn 已提交
1874
	if (ret)
1875
		goto err_verbs_txreq;
M
Mike Marciniszyn 已提交
1876 1877 1878 1879 1880

	ret = hfi1_verbs_register_sysfs(dd);
	if (ret)
		goto err_class;

1881
	return ret;
M
Mike Marciniszyn 已提交
1882 1883

err_class:
1884
	rvt_unregister_device(&dd->verbs_dev.rdi);
M
Mike Marciniszyn 已提交
1885
err_verbs_txreq:
1886
	verbs_txreq_exit(dev);
M
Mike Marciniszyn 已提交
1887 1888 1889 1890 1891 1892 1893 1894 1895 1896
	dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
	return ret;
}

void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
{
	struct hfi1_ibdev *dev = &dd->verbs_dev;

	hfi1_verbs_unregister_sysfs(dd);

1897
	rvt_unregister_device(&dd->verbs_dev.rdi);
M
Mike Marciniszyn 已提交
1898 1899 1900 1901 1902 1903 1904

	if (!list_empty(&dev->txwait))
		dd_dev_err(dd, "txwait list not empty!\n");
	if (!list_empty(&dev->memwait))
		dd_dev_err(dd, "memwait list not empty!\n");

	del_timer_sync(&dev->mem_timer);
1905
	verbs_txreq_exit(dev);
1906 1907 1908 1909

	kfree(dev_cntr_names);
	kfree(port_cntr_names);
	cntr_names_initialized = 0;
M
Mike Marciniszyn 已提交
1910 1911 1912 1913
}

void hfi1_cnp_rcv(struct hfi1_packet *packet)
{
1914
	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1915
	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1916
	struct ib_header *hdr = packet->hdr;
1917
	struct rvt_qp *qp = packet->qp;
1918 1919
	u32 lqpn, rqpn = 0;
	u16 rlid = 0;
1920
	u8 sl, sc5, svc_type;
1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938

	switch (packet->qp->ibqp.qp_type) {
	case IB_QPT_UC:
		rlid = qp->remote_ah_attr.dlid;
		rqpn = qp->remote_qpn;
		svc_type = IB_CC_SVCTYPE_UC;
		break;
	case IB_QPT_RC:
		rlid = qp->remote_ah_attr.dlid;
		rqpn = qp->remote_qpn;
		svc_type = IB_CC_SVCTYPE_RC;
		break;
	case IB_QPT_SMI:
	case IB_QPT_GSI:
	case IB_QPT_UD:
		svc_type = IB_CC_SVCTYPE_UD;
		break;
	default:
1939
		ibp->rvp.n_pkt_drops++;
1940 1941 1942
		return;
	}

1943
	sc5 = hdr2sc(hdr, packet->rhf);
1944 1945 1946 1947
	sl = ibp->sc_to_sl[sc5];
	lqpn = qp->ibqp.qp_num;

	process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
M
Mike Marciniszyn 已提交
1948
}