2 * Copyright(c) 2015, 2016 Intel Corporation.
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.
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.
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.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
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
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.
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.
48 #include <rdma/ib_mad.h>
49 #include <rdma/ib_user_verbs.h>
51 #include <linux/module.h>
52 #include <linux/utsname.h>
53 #include <linux/rculist.h>
55 #include <linux/vmalloc.h>
62 #include "verbs_txreq.h"
65 static unsigned int hfi1_lkey_table_size = 16;
66 module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
68 MODULE_PARM_DESC(lkey_table_size,
69 "LKEY table size in bits (2^n, 1 <= n <= 23)");
71 static unsigned int hfi1_max_pds = 0xFFFF;
72 module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
73 MODULE_PARM_DESC(max_pds,
74 "Maximum number of protection domains to support");
76 static unsigned int hfi1_max_ahs = 0xFFFF;
77 module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
78 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
80 unsigned int hfi1_max_cqes = 0x2FFFFF;
81 module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
82 MODULE_PARM_DESC(max_cqes,
83 "Maximum number of completion queue entries to support");
85 unsigned int hfi1_max_cqs = 0x1FFFF;
86 module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
87 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
89 unsigned int hfi1_max_qp_wrs = 0x3FFF;
90 module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
91 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
93 unsigned int hfi1_max_qps = 32768;
94 module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
95 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
97 unsigned int hfi1_max_sges = 0x60;
98 module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
99 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
101 unsigned int hfi1_max_mcast_grps = 16384;
102 module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
103 MODULE_PARM_DESC(max_mcast_grps,
104 "Maximum number of multicast groups to support");
106 unsigned int hfi1_max_mcast_qp_attached = 16;
107 module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
109 MODULE_PARM_DESC(max_mcast_qp_attached,
110 "Maximum number of attached QPs to support");
112 unsigned int hfi1_max_srqs = 1024;
113 module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
114 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
116 unsigned int hfi1_max_srq_sges = 128;
117 module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
118 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
120 unsigned int hfi1_max_srq_wrs = 0x1FFFF;
121 module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
122 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
124 unsigned short piothreshold = 256;
125 module_param(piothreshold, ushort, S_IRUGO);
126 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
128 #define COPY_CACHELESS 1
129 #define COPY_ADAPTIVE 2
130 static unsigned int sge_copy_mode;
131 module_param(sge_copy_mode, uint, S_IRUGO);
132 MODULE_PARM_DESC(sge_copy_mode,
133 "Verbs copy mode: 0 use memcpy, 1 use cacheless copy, 2 adapt based on WSS");
135 static void verbs_sdma_complete(
136 struct sdma_txreq *cookie,
139 static int pio_wait(struct rvt_qp *qp,
140 struct send_context *sc,
141 struct hfi1_pkt_state *ps,
144 /* Length of buffer to create verbs txreq cache name */
145 #define TXREQ_NAME_LEN 24
147 static uint wss_threshold;
148 module_param(wss_threshold, uint, S_IRUGO);
149 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
150 static uint wss_clean_period = 256;
151 module_param(wss_clean_period, uint, S_IRUGO);
152 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
154 /* memory working set size */
156 unsigned long *entries;
157 atomic_t total_count;
158 atomic_t clean_counter;
159 atomic_t clean_entry;
166 static struct hfi1_wss wss;
168 int hfi1_wss_init(void)
175 /* check for a valid percent range - default to 80 if none or invalid */
176 if (wss_threshold < 1 || wss_threshold > 100)
178 /* reject a wildly large period */
179 if (wss_clean_period > 1000000)
180 wss_clean_period = 256;
181 /* reject a zero period */
182 if (wss_clean_period == 0)
183 wss_clean_period = 1;
186 * Calculate the table size - the next power of 2 larger than the
187 * LLC size. LLC size is in KiB.
189 llc_size = wss_llc_size() * 1024;
190 table_size = roundup_pow_of_two(llc_size);
192 /* one bit per page in rounded up table */
193 llc_bits = llc_size / PAGE_SIZE;
194 table_bits = table_size / PAGE_SIZE;
195 wss.pages_mask = table_bits - 1;
196 wss.num_entries = table_bits / BITS_PER_LONG;
198 wss.threshold = (llc_bits * wss_threshold) / 100;
199 if (wss.threshold == 0)
202 atomic_set(&wss.clean_counter, wss_clean_period);
204 wss.entries = kcalloc(wss.num_entries, sizeof(*wss.entries),
214 void hfi1_wss_exit(void)
216 /* coded to handle partially initialized and repeat callers */
222 * Advance the clean counter. When the clean period has expired,
225 * This is implemented in atomics to avoid locking. Because multiple
226 * variables are involved, it can be racy which can lead to slightly
227 * inaccurate information. Since this is only a heuristic, this is
228 * OK. Any innaccuracies will clean themselves out as the counter
229 * advances. That said, it is unlikely the entry clean operation will
230 * race - the next possible racer will not start until the next clean
233 * The clean counter is implemented as a decrement to zero. When zero
234 * is reached an entry is cleaned.
236 static void wss_advance_clean_counter(void)
242 /* become the cleaner if we decrement the counter to zero */
243 if (atomic_dec_and_test(&wss.clean_counter)) {
245 * Set, not add, the clean period. This avoids an issue
246 * where the counter could decrement below the clean period.
247 * Doing a set can result in lost decrements, slowing the
248 * clean advance. Since this a heuristic, this possible
251 * An alternative is to loop, advancing the counter by a
252 * clean period until the result is > 0. However, this could
253 * lead to several threads keeping another in the clean loop.
254 * This could be mitigated by limiting the number of times
255 * we stay in the loop.
257 atomic_set(&wss.clean_counter, wss_clean_period);
260 * Uniquely grab the entry to clean and move to next.
261 * The current entry is always the lower bits of
262 * wss.clean_entry. The table size, wss.num_entries,
263 * is always a power-of-2.
265 entry = (atomic_inc_return(&wss.clean_entry) - 1)
266 & (wss.num_entries - 1);
268 /* clear the entry and count the bits */
269 bits = xchg(&wss.entries[entry], 0);
270 weight = hweight64((u64)bits);
271 /* only adjust the contended total count if needed */
273 atomic_sub(weight, &wss.total_count);
278 * Insert the given address into the working set array.
280 static void wss_insert(void *address)
282 u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss.pages_mask;
283 u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
284 u32 nr = page & (BITS_PER_LONG - 1);
286 if (!test_and_set_bit(nr, &wss.entries[entry]))
287 atomic_inc(&wss.total_count);
289 wss_advance_clean_counter();
293 * Is the working set larger than the threshold?
295 static inline bool wss_exceeds_threshold(void)
297 return atomic_read(&wss.total_count) >= wss.threshold;
301 * Translate ib_wr_opcode into ib_wc_opcode.
303 const enum ib_wc_opcode ib_hfi1_wc_opcode[] = {
304 [IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
305 [IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
306 [IB_WR_SEND] = IB_WC_SEND,
307 [IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
308 [IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
309 [IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
310 [IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD,
311 [IB_WR_SEND_WITH_INV] = IB_WC_SEND,
312 [IB_WR_LOCAL_INV] = IB_WC_LOCAL_INV,
313 [IB_WR_REG_MR] = IB_WC_REG_MR
317 * Length of header by opcode, 0 --> not supported
319 const u8 hdr_len_by_opcode[256] = {
321 [IB_OPCODE_RC_SEND_FIRST] = 12 + 8,
322 [IB_OPCODE_RC_SEND_MIDDLE] = 12 + 8,
323 [IB_OPCODE_RC_SEND_LAST] = 12 + 8,
324 [IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
325 [IB_OPCODE_RC_SEND_ONLY] = 12 + 8,
326 [IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE] = 12 + 8 + 4,
327 [IB_OPCODE_RC_RDMA_WRITE_FIRST] = 12 + 8 + 16,
328 [IB_OPCODE_RC_RDMA_WRITE_MIDDLE] = 12 + 8,
329 [IB_OPCODE_RC_RDMA_WRITE_LAST] = 12 + 8,
330 [IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
331 [IB_OPCODE_RC_RDMA_WRITE_ONLY] = 12 + 8 + 16,
332 [IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
333 [IB_OPCODE_RC_RDMA_READ_REQUEST] = 12 + 8 + 16,
334 [IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST] = 12 + 8 + 4,
335 [IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE] = 12 + 8,
336 [IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST] = 12 + 8 + 4,
337 [IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY] = 12 + 8 + 4,
338 [IB_OPCODE_RC_ACKNOWLEDGE] = 12 + 8 + 4,
339 [IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE] = 12 + 8 + 4 + 8,
340 [IB_OPCODE_RC_COMPARE_SWAP] = 12 + 8 + 28,
341 [IB_OPCODE_RC_FETCH_ADD] = 12 + 8 + 28,
342 [IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE] = 12 + 8 + 4,
343 [IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE] = 12 + 8 + 4,
345 [IB_OPCODE_UC_SEND_FIRST] = 12 + 8,
346 [IB_OPCODE_UC_SEND_MIDDLE] = 12 + 8,
347 [IB_OPCODE_UC_SEND_LAST] = 12 + 8,
348 [IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
349 [IB_OPCODE_UC_SEND_ONLY] = 12 + 8,
350 [IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE] = 12 + 8 + 4,
351 [IB_OPCODE_UC_RDMA_WRITE_FIRST] = 12 + 8 + 16,
352 [IB_OPCODE_UC_RDMA_WRITE_MIDDLE] = 12 + 8,
353 [IB_OPCODE_UC_RDMA_WRITE_LAST] = 12 + 8,
354 [IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
355 [IB_OPCODE_UC_RDMA_WRITE_ONLY] = 12 + 8 + 16,
356 [IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
358 [IB_OPCODE_UD_SEND_ONLY] = 12 + 8 + 8,
359 [IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE] = 12 + 8 + 12
362 static const opcode_handler opcode_handler_tbl[256] = {
364 [IB_OPCODE_RC_SEND_FIRST] = &hfi1_rc_rcv,
365 [IB_OPCODE_RC_SEND_MIDDLE] = &hfi1_rc_rcv,
366 [IB_OPCODE_RC_SEND_LAST] = &hfi1_rc_rcv,
367 [IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
368 [IB_OPCODE_RC_SEND_ONLY] = &hfi1_rc_rcv,
369 [IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
370 [IB_OPCODE_RC_RDMA_WRITE_FIRST] = &hfi1_rc_rcv,
371 [IB_OPCODE_RC_RDMA_WRITE_MIDDLE] = &hfi1_rc_rcv,
372 [IB_OPCODE_RC_RDMA_WRITE_LAST] = &hfi1_rc_rcv,
373 [IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
374 [IB_OPCODE_RC_RDMA_WRITE_ONLY] = &hfi1_rc_rcv,
375 [IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
376 [IB_OPCODE_RC_RDMA_READ_REQUEST] = &hfi1_rc_rcv,
377 [IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST] = &hfi1_rc_rcv,
378 [IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE] = &hfi1_rc_rcv,
379 [IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST] = &hfi1_rc_rcv,
380 [IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY] = &hfi1_rc_rcv,
381 [IB_OPCODE_RC_ACKNOWLEDGE] = &hfi1_rc_rcv,
382 [IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE] = &hfi1_rc_rcv,
383 [IB_OPCODE_RC_COMPARE_SWAP] = &hfi1_rc_rcv,
384 [IB_OPCODE_RC_FETCH_ADD] = &hfi1_rc_rcv,
385 [IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE] = &hfi1_rc_rcv,
386 [IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE] = &hfi1_rc_rcv,
388 [IB_OPCODE_UC_SEND_FIRST] = &hfi1_uc_rcv,
389 [IB_OPCODE_UC_SEND_MIDDLE] = &hfi1_uc_rcv,
390 [IB_OPCODE_UC_SEND_LAST] = &hfi1_uc_rcv,
391 [IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
392 [IB_OPCODE_UC_SEND_ONLY] = &hfi1_uc_rcv,
393 [IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
394 [IB_OPCODE_UC_RDMA_WRITE_FIRST] = &hfi1_uc_rcv,
395 [IB_OPCODE_UC_RDMA_WRITE_MIDDLE] = &hfi1_uc_rcv,
396 [IB_OPCODE_UC_RDMA_WRITE_LAST] = &hfi1_uc_rcv,
397 [IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
398 [IB_OPCODE_UC_RDMA_WRITE_ONLY] = &hfi1_uc_rcv,
399 [IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
401 [IB_OPCODE_UD_SEND_ONLY] = &hfi1_ud_rcv,
402 [IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE] = &hfi1_ud_rcv,
404 [IB_OPCODE_CNP] = &hfi1_cnp_rcv
409 static const u32 pio_opmask[BIT(3)] = {
411 [IB_OPCODE_RC >> 5] =
412 BIT(RC_OP(SEND_ONLY) & OPMASK) |
413 BIT(RC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
414 BIT(RC_OP(RDMA_WRITE_ONLY) & OPMASK) |
415 BIT(RC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK) |
416 BIT(RC_OP(RDMA_READ_REQUEST) & OPMASK) |
417 BIT(RC_OP(ACKNOWLEDGE) & OPMASK) |
418 BIT(RC_OP(ATOMIC_ACKNOWLEDGE) & OPMASK) |
419 BIT(RC_OP(COMPARE_SWAP) & OPMASK) |
420 BIT(RC_OP(FETCH_ADD) & OPMASK),
422 [IB_OPCODE_UC >> 5] =
423 BIT(UC_OP(SEND_ONLY) & OPMASK) |
424 BIT(UC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
425 BIT(UC_OP(RDMA_WRITE_ONLY) & OPMASK) |
426 BIT(UC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK),
432 __be64 ib_hfi1_sys_image_guid;
435 * hfi1_copy_sge - copy data to SGE memory
437 * @data: the data to copy
438 * @length: the length of the data
439 * @release: boolean to release MR
440 * @copy_last: do a separate copy of the last 8 bytes
443 struct rvt_sge_state *ss,
444 void *data, u32 length,
448 struct rvt_sge *sge = &ss->sge;
450 bool in_last = false;
451 bool cacheless_copy = false;
453 if (sge_copy_mode == COPY_CACHELESS) {
454 cacheless_copy = length >= PAGE_SIZE;
455 } else if (sge_copy_mode == COPY_ADAPTIVE) {
456 if (length >= PAGE_SIZE) {
458 * NOTE: this *assumes*:
459 * o The first vaddr is the dest.
460 * o If multiple pages, then vaddr is sequential.
462 wss_insert(sge->vaddr);
463 if (length >= (2 * PAGE_SIZE))
464 wss_insert(sge->vaddr + PAGE_SIZE);
466 cacheless_copy = wss_exceeds_threshold();
468 wss_advance_clean_counter();
482 u32 len = rvt_get_sge_length(sge, length);
484 WARN_ON_ONCE(len == 0);
485 if (unlikely(in_last)) {
486 /* enforce byte transfer ordering */
487 for (i = 0; i < len; i++)
488 ((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
489 } else if (cacheless_copy) {
490 cacheless_memcpy(sge->vaddr, data, len);
492 memcpy(sge->vaddr, data, len);
494 rvt_update_sge(ss, len, release);
508 * Make sure the QP is ready and able to accept the given opcode.
510 static inline opcode_handler qp_ok(int opcode, struct hfi1_packet *packet)
512 if (!(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
514 if (((opcode & RVT_OPCODE_QP_MASK) == packet->qp->allowed_ops) ||
515 (opcode == IB_OPCODE_CNP))
516 return opcode_handler_tbl[opcode];
522 * hfi1_ib_rcv - process an incoming packet
523 * @packet: data packet information
525 * This is called to process an incoming packet at interrupt level.
527 * Tlen is the length of the header + data + CRC in bytes.
529 void hfi1_ib_rcv(struct hfi1_packet *packet)
531 struct hfi1_ctxtdata *rcd = packet->rcd;
532 struct ib_header *hdr = packet->hdr;
533 u32 tlen = packet->tlen;
534 struct hfi1_pportdata *ppd = rcd->ppd;
535 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
536 struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
537 opcode_handler packet_handler;
545 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
546 if (lnh == HFI1_LRH_BTH) {
547 packet->ohdr = &hdr->u.oth;
548 } else if (lnh == HFI1_LRH_GRH) {
551 packet->ohdr = &hdr->u.l.oth;
552 if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
554 vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
555 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
557 packet->rcv_flags |= HFI1_HAS_GRH;
562 trace_input_ibhdr(rcd->dd, hdr);
564 opcode = (be32_to_cpu(packet->ohdr->bth[0]) >> 24);
565 inc_opstats(tlen, &rcd->opstats->stats[opcode]);
567 /* Get the destination QP number. */
568 qp_num = be32_to_cpu(packet->ohdr->bth[1]) & RVT_QPN_MASK;
569 lid = be16_to_cpu(hdr->lrh[1]);
570 if (unlikely((lid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
571 (lid != be16_to_cpu(IB_LID_PERMISSIVE)))) {
572 struct rvt_mcast *mcast;
573 struct rvt_mcast_qp *p;
575 if (lnh != HFI1_LRH_GRH)
577 mcast = rvt_mcast_find(&ibp->rvp, &hdr->u.l.grh.dgid);
580 list_for_each_entry_rcu(p, &mcast->qp_list, list) {
582 spin_lock_irqsave(&packet->qp->r_lock, flags);
583 packet_handler = qp_ok(opcode, packet);
584 if (likely(packet_handler))
585 packet_handler(packet);
587 ibp->rvp.n_pkt_drops++;
588 spin_unlock_irqrestore(&packet->qp->r_lock, flags);
591 * Notify rvt_multicast_detach() if it is waiting for us
594 if (atomic_dec_return(&mcast->refcount) <= 1)
595 wake_up(&mcast->wait);
598 packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
603 if (unlikely(hfi1_dbg_fault_opcode(packet->qp, opcode,
608 spin_lock_irqsave(&packet->qp->r_lock, flags);
609 packet_handler = qp_ok(opcode, packet);
610 if (likely(packet_handler))
611 packet_handler(packet);
613 ibp->rvp.n_pkt_drops++;
614 spin_unlock_irqrestore(&packet->qp->r_lock, flags);
620 ibp->rvp.n_pkt_drops++;
624 * This is called from a timer to check for QPs
625 * which need kernel memory in order to send a packet.
627 static void mem_timer(unsigned long data)
629 struct hfi1_ibdev *dev = (struct hfi1_ibdev *)data;
630 struct list_head *list = &dev->memwait;
631 struct rvt_qp *qp = NULL;
634 struct hfi1_qp_priv *priv;
636 write_seqlock_irqsave(&dev->iowait_lock, flags);
637 if (!list_empty(list)) {
638 wait = list_first_entry(list, struct iowait, list);
639 qp = iowait_to_qp(wait);
641 list_del_init(&priv->s_iowait.list);
642 priv->s_iowait.lock = NULL;
643 /* refcount held until actual wake up */
644 if (!list_empty(list))
645 mod_timer(&dev->mem_timer, jiffies + 1);
647 write_sequnlock_irqrestore(&dev->iowait_lock, flags);
650 hfi1_qp_wakeup(qp, RVT_S_WAIT_KMEM);
654 * This is called with progress side lock held.
657 static void verbs_sdma_complete(
658 struct sdma_txreq *cookie,
661 struct verbs_txreq *tx =
662 container_of(cookie, struct verbs_txreq, txreq);
663 struct rvt_qp *qp = tx->qp;
665 spin_lock(&qp->s_lock);
667 hfi1_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
668 } else if (qp->ibqp.qp_type == IB_QPT_RC) {
669 struct ib_header *hdr;
672 hfi1_rc_send_complete(qp, hdr);
674 spin_unlock(&qp->s_lock);
679 static int wait_kmem(struct hfi1_ibdev *dev,
681 struct hfi1_pkt_state *ps)
683 struct hfi1_qp_priv *priv = qp->priv;
687 spin_lock_irqsave(&qp->s_lock, flags);
688 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
689 write_seqlock(&dev->iowait_lock);
690 list_add_tail(&ps->s_txreq->txreq.list,
691 &priv->s_iowait.tx_head);
692 if (list_empty(&priv->s_iowait.list)) {
693 if (list_empty(&dev->memwait))
694 mod_timer(&dev->mem_timer, jiffies + 1);
695 qp->s_flags |= RVT_S_WAIT_KMEM;
696 list_add_tail(&priv->s_iowait.list, &dev->memwait);
697 priv->s_iowait.lock = &dev->iowait_lock;
698 trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM);
701 write_sequnlock(&dev->iowait_lock);
702 qp->s_flags &= ~RVT_S_BUSY;
705 spin_unlock_irqrestore(&qp->s_lock, flags);
711 * This routine calls txadds for each sg entry.
713 * Add failures will revert the sge cursor
715 static noinline int build_verbs_ulp_payload(
716 struct sdma_engine *sde,
718 struct verbs_txreq *tx)
720 struct rvt_sge_state *ss = tx->ss;
721 struct rvt_sge *sg_list = ss->sg_list;
722 struct rvt_sge sge = ss->sge;
723 u8 num_sge = ss->num_sge;
728 len = ss->sge.length;
731 if (len > ss->sge.sge_length)
732 len = ss->sge.sge_length;
733 WARN_ON_ONCE(len == 0);
734 ret = sdma_txadd_kvaddr(
741 rvt_update_sge(ss, len, false);
748 ss->num_sge = num_sge;
749 ss->sg_list = sg_list;
754 * Build the number of DMA descriptors needed to send length bytes of data.
756 * NOTE: DMA mapping is held in the tx until completed in the ring or
757 * the tx desc is freed without having been submitted to the ring
759 * This routine ensures all the helper routine calls succeed.
762 static int build_verbs_tx_desc(
763 struct sdma_engine *sde,
765 struct verbs_txreq *tx,
766 struct hfi1_ahg_info *ahg_info,
770 struct hfi1_sdma_header *phdr = &tx->phdr;
771 u16 hdrbytes = tx->hdr_dwords << 2;
773 if (!ahg_info->ahgcount) {
774 ret = sdma_txinit_ahg(
782 verbs_sdma_complete);
785 phdr->pbc = cpu_to_le64(pbc);
786 ret = sdma_txadd_kvaddr(
794 ret = sdma_txinit_ahg(
802 verbs_sdma_complete);
807 /* add the ulp payload - if any. tx->ss can be NULL for acks */
809 ret = build_verbs_ulp_payload(sde, length, tx);
814 int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
817 struct hfi1_qp_priv *priv = qp->priv;
818 struct hfi1_ahg_info *ahg_info = priv->s_ahg;
819 u32 hdrwords = qp->s_hdrwords;
820 u32 len = ps->s_txreq->s_cur_size;
821 u32 plen = hdrwords + ((len + 3) >> 2) + 2; /* includes pbc */
822 struct hfi1_ibdev *dev = ps->dev;
823 struct hfi1_pportdata *ppd = ps->ppd;
824 struct verbs_txreq *tx;
831 if (!sdma_txreq_built(&tx->txreq)) {
832 if (likely(pbc == 0)) {
833 u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
835 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
836 pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
838 pbc = create_pbc(ppd,
845 ret = build_verbs_tx_desc(tx->sde, len, tx, ahg_info, pbc);
849 ret = sdma_send_txreq(tx->sde, &priv->s_iowait, &tx->txreq);
850 if (unlikely(ret < 0)) {
855 trace_sdma_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
856 &ps->s_txreq->phdr.hdr);
860 /* The current one got "sent" */
863 ret = wait_kmem(dev, qp, ps);
865 /* free txreq - bad state */
866 hfi1_put_txreq(ps->s_txreq);
873 * If we are now in the error state, return zero to flush the
876 static int pio_wait(struct rvt_qp *qp,
877 struct send_context *sc,
878 struct hfi1_pkt_state *ps,
881 struct hfi1_qp_priv *priv = qp->priv;
882 struct hfi1_devdata *dd = sc->dd;
883 struct hfi1_ibdev *dev = &dd->verbs_dev;
888 * Note that as soon as want_buffer() is called and
889 * possibly before it returns, sc_piobufavail()
890 * could be called. Therefore, put QP on the I/O wait list before
891 * enabling the PIO avail interrupt.
893 spin_lock_irqsave(&qp->s_lock, flags);
894 if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
895 write_seqlock(&dev->iowait_lock);
896 list_add_tail(&ps->s_txreq->txreq.list,
897 &priv->s_iowait.tx_head);
898 if (list_empty(&priv->s_iowait.list)) {
899 struct hfi1_ibdev *dev = &dd->verbs_dev;
902 dev->n_piowait += !!(flag & RVT_S_WAIT_PIO);
903 dev->n_piodrain += !!(flag & RVT_S_WAIT_PIO_DRAIN);
905 was_empty = list_empty(&sc->piowait);
906 list_add_tail(&priv->s_iowait.list, &sc->piowait);
907 priv->s_iowait.lock = &dev->iowait_lock;
908 trace_hfi1_qpsleep(qp, RVT_S_WAIT_PIO);
910 /* counting: only call wantpiobuf_intr if first user */
912 hfi1_sc_wantpiobuf_intr(sc, 1);
914 write_sequnlock(&dev->iowait_lock);
915 qp->s_flags &= ~RVT_S_BUSY;
918 spin_unlock_irqrestore(&qp->s_lock, flags);
922 static void verbs_pio_complete(void *arg, int code)
924 struct rvt_qp *qp = (struct rvt_qp *)arg;
925 struct hfi1_qp_priv *priv = qp->priv;
927 if (iowait_pio_dec(&priv->s_iowait))
928 iowait_drain_wakeup(&priv->s_iowait);
931 int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
934 struct hfi1_qp_priv *priv = qp->priv;
935 u32 hdrwords = qp->s_hdrwords;
936 struct rvt_sge_state *ss = ps->s_txreq->ss;
937 u32 len = ps->s_txreq->s_cur_size;
938 u32 dwords = (len + 3) >> 2;
939 u32 plen = hdrwords + dwords + 2; /* includes pbc */
940 struct hfi1_pportdata *ppd = ps->ppd;
941 u32 *hdr = (u32 *)&ps->s_txreq->phdr.hdr;
944 unsigned long flags = 0;
945 struct send_context *sc;
946 struct pio_buf *pbuf;
947 int wc_status = IB_WC_SUCCESS;
949 pio_release_cb cb = NULL;
951 /* only RC/UC use complete */
952 switch (qp->ibqp.qp_type) {
955 cb = verbs_pio_complete;
961 /* vl15 special case taken care of in ud.c */
963 sc = ps->s_txreq->psc;
965 if (likely(pbc == 0)) {
966 u8 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
967 /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
968 pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
969 pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
972 iowait_pio_inc(&priv->s_iowait);
973 pbuf = sc_buffer_alloc(sc, plen, cb, qp);
974 if (unlikely(!pbuf)) {
976 verbs_pio_complete(qp, 0);
977 if (ppd->host_link_state != HLS_UP_ACTIVE) {
979 * If we have filled the PIO buffers to capacity and are
980 * not in an active state this request is not going to
981 * go out to so just complete it with an error or else a
982 * ULP or the core may be stuck waiting.
986 "alloc failed. state not active, completing");
987 wc_status = IB_WC_GENERAL_ERR;
991 * This is a normal occurrence. The PIO buffs are full
992 * up but we are still happily sending, well we could be
993 * so lets continue to queue the request.
995 hfi1_cdbg(PIO, "alloc failed. state active, queuing");
996 ret = pio_wait(qp, sc, ps, RVT_S_WAIT_PIO);
998 /* txreq not queued - free */
1000 /* tx consumed in wait */
1006 pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
1009 seg_pio_copy_start(pbuf, pbc, hdr, hdrwords * 4);
1011 void *addr = ss->sge.vaddr;
1012 u32 slen = ss->sge.length;
1016 rvt_update_sge(ss, slen, false);
1017 seg_pio_copy_mid(pbuf, addr, slen);
1020 seg_pio_copy_end(pbuf);
1024 trace_pio_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
1025 &ps->s_txreq->phdr.hdr);
1029 spin_lock_irqsave(&qp->s_lock, flags);
1030 hfi1_send_complete(qp, qp->s_wqe, wc_status);
1031 spin_unlock_irqrestore(&qp->s_lock, flags);
1032 } else if (qp->ibqp.qp_type == IB_QPT_RC) {
1033 spin_lock_irqsave(&qp->s_lock, flags);
1034 hfi1_rc_send_complete(qp, &ps->s_txreq->phdr.hdr);
1035 spin_unlock_irqrestore(&qp->s_lock, flags);
1041 hfi1_put_txreq(ps->s_txreq);
1046 * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1047 * being an entry from the partition key table), return 0
1048 * otherwise. Use the matching criteria for egress partition keys
1049 * specified in the OPAv1 spec., section 9.1l.7.
1051 static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
1053 u16 mkey = pkey & PKEY_LOW_15_MASK;
1054 u16 mentry = ent & PKEY_LOW_15_MASK;
1056 if (mkey == mentry) {
1058 * If pkey[15] is set (full partition member),
1059 * is bit 15 in the corresponding table element
1060 * clear (limited member)?
1062 if (pkey & PKEY_MEMBER_MASK)
1063 return !!(ent & PKEY_MEMBER_MASK);
1070 * egress_pkey_check - check P_KEY of a packet
1071 * @ppd: Physical IB port data
1072 * @lrh: Local route header
1073 * @bth: Base transport header
1074 * @sc5: SC for packet
1075 * @s_pkey_index: It will be used for look up optimization for kernel contexts
1076 * only. If it is negative value, then it means user contexts is calling this
1079 * It checks if hdr's pkey is valid.
1081 * Return: 0 on success, otherwise, 1
1083 int egress_pkey_check(struct hfi1_pportdata *ppd, __be16 *lrh, __be32 *bth,
1084 u8 sc5, int8_t s_pkey_index)
1086 struct hfi1_devdata *dd;
1089 int is_user_ctxt_mechanism = (s_pkey_index < 0);
1091 if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
1094 pkey = (u16)be32_to_cpu(bth[0]);
1096 /* If SC15, pkey[0:14] must be 0x7fff */
1097 if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
1100 /* Is the pkey = 0x0, or 0x8000? */
1101 if ((pkey & PKEY_LOW_15_MASK) == 0)
1105 * For the kernel contexts only, if a qp is passed into the function,
1106 * the most likely matching pkey has index qp->s_pkey_index
1108 if (!is_user_ctxt_mechanism &&
1109 egress_pkey_matches_entry(pkey, ppd->pkeys[s_pkey_index])) {
1113 for (i = 0; i < MAX_PKEY_VALUES; i++) {
1114 if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
1119 * For the user-context mechanism, the P_KEY check would only happen
1120 * once per SDMA request, not once per packet. Therefore, there's no
1121 * need to increment the counter for the user-context mechanism.
1123 if (!is_user_ctxt_mechanism) {
1124 incr_cntr64(&ppd->port_xmit_constraint_errors);
1126 if (!(dd->err_info_xmit_constraint.status &
1127 OPA_EI_STATUS_SMASK)) {
1128 u16 slid = be16_to_cpu(lrh[3]);
1130 dd->err_info_xmit_constraint.status |=
1131 OPA_EI_STATUS_SMASK;
1132 dd->err_info_xmit_constraint.slid = slid;
1133 dd->err_info_xmit_constraint.pkey = pkey;
1140 * get_send_routine - choose an egress routine
1142 * Choose an egress routine based on QP type
1145 static inline send_routine get_send_routine(struct rvt_qp *qp,
1146 struct verbs_txreq *tx)
1148 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1149 struct hfi1_qp_priv *priv = qp->priv;
1150 struct ib_header *h = &tx->phdr.hdr;
1152 if (unlikely(!(dd->flags & HFI1_HAS_SEND_DMA)))
1153 return dd->process_pio_send;
1154 switch (qp->ibqp.qp_type) {
1156 return dd->process_pio_send;
1162 u8 op = get_opcode(h);
1165 tx->s_cur_size <= min(piothreshold, qp->pmtu) &&
1166 (BIT(op & OPMASK) & pio_opmask[op >> 5]) &&
1167 iowait_sdma_pending(&priv->s_iowait) == 0 &&
1168 !sdma_txreq_built(&tx->txreq))
1169 return dd->process_pio_send;
1175 return dd->process_dma_send;
1179 * hfi1_verbs_send - send a packet
1180 * @qp: the QP to send on
1181 * @ps: the state of the packet to send
1183 * Return zero if packet is sent or queued OK.
1184 * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1186 int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
1188 struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1189 struct hfi1_qp_priv *priv = qp->priv;
1190 struct ib_other_headers *ohdr;
1191 struct ib_header *hdr;
1196 hdr = &ps->s_txreq->phdr.hdr;
1197 /* locate the pkey within the headers */
1198 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
1199 if (lnh == HFI1_LRH_GRH)
1200 ohdr = &hdr->u.l.oth;
1204 sr = get_send_routine(qp, ps->s_txreq);
1205 ret = egress_pkey_check(dd->pport,
1210 if (unlikely(ret)) {
1212 * The value we are returning here does not get propagated to
1213 * the verbs caller. Thus we need to complete the request with
1214 * error otherwise the caller could be sitting waiting on the
1215 * completion event. Only do this for PIO. SDMA has its own
1216 * mechanism for handling the errors. So for SDMA we can just
1219 if (sr == dd->process_pio_send) {
1220 unsigned long flags;
1222 hfi1_cdbg(PIO, "%s() Failed. Completing with err",
1224 spin_lock_irqsave(&qp->s_lock, flags);
1225 hfi1_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
1226 spin_unlock_irqrestore(&qp->s_lock, flags);
1230 if (sr == dd->process_dma_send && iowait_pio_pending(&priv->s_iowait))
1234 RVT_S_WAIT_PIO_DRAIN);
1235 return sr(qp, ps, 0);
1239 * hfi1_fill_device_attr - Fill in rvt dev info device attributes.
1240 * @dd: the device data structure
1242 static void hfi1_fill_device_attr(struct hfi1_devdata *dd)
1244 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1245 u32 ver = dd->dc8051_ver;
1247 memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1249 rdi->dparms.props.fw_ver = ((u64)(dc8051_ver_maj(ver)) << 32) |
1250 ((u64)(dc8051_ver_min(ver)) << 16) |
1251 (u64)dc8051_ver_patch(ver);
1253 rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1254 IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1255 IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1256 IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE |
1257 IB_DEVICE_MEM_MGT_EXTENSIONS;
1258 rdi->dparms.props.page_size_cap = PAGE_SIZE;
1259 rdi->dparms.props.vendor_id = dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
1260 rdi->dparms.props.vendor_part_id = dd->pcidev->device;
1261 rdi->dparms.props.hw_ver = dd->minrev;
1262 rdi->dparms.props.sys_image_guid = ib_hfi1_sys_image_guid;
1263 rdi->dparms.props.max_mr_size = U64_MAX;
1264 rdi->dparms.props.max_fast_reg_page_list_len = UINT_MAX;
1265 rdi->dparms.props.max_qp = hfi1_max_qps;
1266 rdi->dparms.props.max_qp_wr = hfi1_max_qp_wrs;
1267 rdi->dparms.props.max_sge = hfi1_max_sges;
1268 rdi->dparms.props.max_sge_rd = hfi1_max_sges;
1269 rdi->dparms.props.max_cq = hfi1_max_cqs;
1270 rdi->dparms.props.max_ah = hfi1_max_ahs;
1271 rdi->dparms.props.max_cqe = hfi1_max_cqes;
1272 rdi->dparms.props.max_mr = rdi->lkey_table.max;
1273 rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1274 rdi->dparms.props.max_map_per_fmr = 32767;
1275 rdi->dparms.props.max_pd = hfi1_max_pds;
1276 rdi->dparms.props.max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
1277 rdi->dparms.props.max_qp_init_rd_atom = 255;
1278 rdi->dparms.props.max_srq = hfi1_max_srqs;
1279 rdi->dparms.props.max_srq_wr = hfi1_max_srq_wrs;
1280 rdi->dparms.props.max_srq_sge = hfi1_max_srq_sges;
1281 rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1282 rdi->dparms.props.max_pkeys = hfi1_get_npkeys(dd);
1283 rdi->dparms.props.max_mcast_grp = hfi1_max_mcast_grps;
1284 rdi->dparms.props.max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
1285 rdi->dparms.props.max_total_mcast_qp_attach =
1286 rdi->dparms.props.max_mcast_qp_attach *
1287 rdi->dparms.props.max_mcast_grp;
1290 static inline u16 opa_speed_to_ib(u16 in)
1294 if (in & OPA_LINK_SPEED_25G)
1295 out |= IB_SPEED_EDR;
1296 if (in & OPA_LINK_SPEED_12_5G)
1297 out |= IB_SPEED_FDR;
1303 * Convert a single OPA link width (no multiple flags) to an IB value.
1304 * A zero OPA link width means link down, which means the IB width value
1307 static inline u16 opa_width_to_ib(u16 in)
1310 case OPA_LINK_WIDTH_1X:
1311 /* map 2x and 3x to 1x as they don't exist in IB */
1312 case OPA_LINK_WIDTH_2X:
1313 case OPA_LINK_WIDTH_3X:
1315 default: /* link down or unknown, return our largest width */
1316 case OPA_LINK_WIDTH_4X:
1321 static int query_port(struct rvt_dev_info *rdi, u8 port_num,
1322 struct ib_port_attr *props)
1324 struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1325 struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1326 struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1329 /* props being zeroed by the caller, avoid zeroing it here */
1330 props->lid = lid ? lid : 0;
1331 props->lmc = ppd->lmc;
1332 /* OPA logical states match IB logical states */
1333 props->state = driver_lstate(ppd);
1334 props->phys_state = hfi1_ibphys_portstate(ppd);
1335 props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
1336 props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
1337 /* see rate_show() in ib core/sysfs.c */
1338 props->active_speed = (u8)opa_speed_to_ib(ppd->link_speed_active);
1339 props->max_vl_num = ppd->vls_supported;
1341 /* Once we are a "first class" citizen and have added the OPA MTUs to
1342 * the core we can advertise the larger MTU enum to the ULPs, for now
1343 * advertise only 4K.
1345 * Those applications which are either OPA aware or pass the MTU enum
1346 * from the Path Records to us will get the new 8k MTU. Those that
1347 * attempt to process the MTU enum may fail in various ways.
1349 props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
1350 4096 : hfi1_max_mtu), IB_MTU_4096);
1351 props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
1352 mtu_to_enum(ppd->ibmtu, IB_MTU_2048);
1357 static int modify_device(struct ib_device *device,
1358 int device_modify_mask,
1359 struct ib_device_modify *device_modify)
1361 struct hfi1_devdata *dd = dd_from_ibdev(device);
1365 if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1366 IB_DEVICE_MODIFY_NODE_DESC)) {
1371 if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1372 memcpy(device->node_desc, device_modify->node_desc,
1373 IB_DEVICE_NODE_DESC_MAX);
1374 for (i = 0; i < dd->num_pports; i++) {
1375 struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1377 hfi1_node_desc_chg(ibp);
1381 if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1382 ib_hfi1_sys_image_guid =
1383 cpu_to_be64(device_modify->sys_image_guid);
1384 for (i = 0; i < dd->num_pports; i++) {
1385 struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1387 hfi1_sys_guid_chg(ibp);
1397 static int shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1399 struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1400 struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1401 struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1404 set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
1405 OPA_LINKDOWN_REASON_UNKNOWN);
1406 ret = set_link_state(ppd, HLS_DN_DOWNDEF);
1410 static int hfi1_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1411 int guid_index, __be64 *guid)
1413 struct hfi1_ibport *ibp = container_of(rvp, struct hfi1_ibport, rvp);
1415 if (guid_index >= HFI1_GUIDS_PER_PORT)
1418 *guid = get_sguid(ibp, guid_index);
1423 * convert ah port,sl to sc
1425 u8 ah_to_sc(struct ib_device *ibdev, struct ib_ah_attr *ah)
1427 struct hfi1_ibport *ibp = to_iport(ibdev, ah->port_num);
1429 return ibp->sl_to_sc[ah->sl];
1432 static int hfi1_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr)
1434 struct hfi1_ibport *ibp;
1435 struct hfi1_pportdata *ppd;
1436 struct hfi1_devdata *dd;
1439 /* test the mapping for validity */
1440 ibp = to_iport(ibdev, ah_attr->port_num);
1441 ppd = ppd_from_ibp(ibp);
1442 sc5 = ibp->sl_to_sc[ah_attr->sl];
1443 dd = dd_from_ppd(ppd);
1444 if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1449 static void hfi1_notify_new_ah(struct ib_device *ibdev,
1450 struct ib_ah_attr *ah_attr,
1453 struct hfi1_ibport *ibp;
1454 struct hfi1_pportdata *ppd;
1455 struct hfi1_devdata *dd;
1459 * Do not trust reading anything from rvt_ah at this point as it is not
1460 * done being setup. We can however modify things which we need to set.
1463 ibp = to_iport(ibdev, ah_attr->port_num);
1464 ppd = ppd_from_ibp(ibp);
1465 sc5 = ibp->sl_to_sc[ah->attr.sl];
1466 dd = dd_from_ppd(ppd);
1467 ah->vl = sc_to_vlt(dd, sc5);
1468 if (ah->vl < num_vls || ah->vl == 15)
1469 ah->log_pmtu = ilog2(dd->vld[ah->vl].mtu);
1472 struct ib_ah *hfi1_create_qp0_ah(struct hfi1_ibport *ibp, u16 dlid)
1474 struct ib_ah_attr attr;
1475 struct ib_ah *ah = ERR_PTR(-EINVAL);
1478 memset(&attr, 0, sizeof(attr));
1480 attr.port_num = ppd_from_ibp(ibp)->port;
1482 qp0 = rcu_dereference(ibp->rvp.qp[0]);
1484 ah = ib_create_ah(qp0->ibqp.pd, &attr);
1490 * hfi1_get_npkeys - return the size of the PKEY table for context 0
1491 * @dd: the hfi1_ib device
1493 unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
1495 return ARRAY_SIZE(dd->pport[0].pkeys);
1498 static void init_ibport(struct hfi1_pportdata *ppd)
1500 struct hfi1_ibport *ibp = &ppd->ibport_data;
1501 size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
1504 for (i = 0; i < sz; i++) {
1505 ibp->sl_to_sc[i] = i;
1506 ibp->sc_to_sl[i] = i;
1509 spin_lock_init(&ibp->rvp.lock);
1510 /* Set the prefix to the default value (see ch. 4.1.1) */
1511 ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1512 ibp->rvp.sm_lid = 0;
1513 /* Below should only set bits defined in OPA PortInfo.CapabilityMask */
1514 ibp->rvp.port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
1515 IB_PORT_CAP_MASK_NOTICE_SUP;
1516 ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1517 ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1518 ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1519 ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1520 ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1522 RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1523 RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1526 static void hfi1_get_dev_fw_str(struct ib_device *ibdev, char *str,
1529 struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
1530 struct hfi1_ibdev *dev = dev_from_rdi(rdi);
1531 u32 ver = dd_from_dev(dev)->dc8051_ver;
1533 snprintf(str, str_len, "%u.%u.%u", dc8051_ver_maj(ver),
1534 dc8051_ver_min(ver), dc8051_ver_patch(ver));
1537 static const char * const driver_cntr_names[] = {
1538 /* must be element 0*/
1546 "DRIVER_RcvLen_Errs",
1547 "DRIVER_EgrBufFull",
1551 static DEFINE_MUTEX(cntr_names_lock); /* protects the *_cntr_names bufers */
1552 static const char **dev_cntr_names;
1553 static const char **port_cntr_names;
1554 static int num_driver_cntrs = ARRAY_SIZE(driver_cntr_names);
1555 static int num_dev_cntrs;
1556 static int num_port_cntrs;
1557 static int cntr_names_initialized;
1560 * Convert a list of names separated by '\n' into an array of NULL terminated
1561 * strings. Optionally some entries can be reserved in the array to hold extra
1564 static int init_cntr_names(const char *names_in,
1565 const size_t names_len,
1566 int num_extra_names,
1568 const char ***cntr_names)
1570 char *names_out, *p, **q;
1574 for (i = 0; i < names_len; i++)
1575 if (names_in[i] == '\n')
1578 names_out = kmalloc((n + num_extra_names) * sizeof(char *) + names_len,
1586 p = names_out + (n + num_extra_names) * sizeof(char *);
1587 memcpy(p, names_in, names_len);
1589 q = (char **)names_out;
1590 for (i = 0; i < n; i++) {
1592 p = strchr(p, '\n');
1597 *cntr_names = (const char **)names_out;
1601 static struct rdma_hw_stats *alloc_hw_stats(struct ib_device *ibdev,
1606 mutex_lock(&cntr_names_lock);
1607 if (!cntr_names_initialized) {
1608 struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1610 err = init_cntr_names(dd->cntrnames,
1616 mutex_unlock(&cntr_names_lock);
1620 for (i = 0; i < num_driver_cntrs; i++)
1621 dev_cntr_names[num_dev_cntrs + i] =
1622 driver_cntr_names[i];
1624 err = init_cntr_names(dd->portcntrnames,
1625 dd->portcntrnameslen,
1630 kfree(dev_cntr_names);
1631 dev_cntr_names = NULL;
1632 mutex_unlock(&cntr_names_lock);
1635 cntr_names_initialized = 1;
1637 mutex_unlock(&cntr_names_lock);
1640 return rdma_alloc_hw_stats_struct(
1642 num_dev_cntrs + num_driver_cntrs,
1643 RDMA_HW_STATS_DEFAULT_LIFESPAN);
1645 return rdma_alloc_hw_stats_struct(
1648 RDMA_HW_STATS_DEFAULT_LIFESPAN);
1651 static u64 hfi1_sps_ints(void)
1653 unsigned long flags;
1654 struct hfi1_devdata *dd;
1657 spin_lock_irqsave(&hfi1_devs_lock, flags);
1658 list_for_each_entry(dd, &hfi1_dev_list, list) {
1659 sps_ints += get_all_cpu_total(dd->int_counter);
1661 spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1665 static int get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1672 u64 *stats = (u64 *)&hfi1_stats;
1675 hfi1_read_cntrs(dd_from_ibdev(ibdev), NULL, &values);
1676 values[num_dev_cntrs] = hfi1_sps_ints();
1677 for (i = 1; i < num_driver_cntrs; i++)
1678 values[num_dev_cntrs + i] = stats[i];
1679 count = num_dev_cntrs + num_driver_cntrs;
1681 struct hfi1_ibport *ibp = to_iport(ibdev, port);
1683 hfi1_read_portcntrs(ppd_from_ibp(ibp), NULL, &values);
1684 count = num_port_cntrs;
1687 memcpy(stats->value, values, count * sizeof(u64));
1692 * hfi1_register_ib_device - register our device with the infiniband core
1693 * @dd: the device data structure
1694 * Return 0 if successful, errno if unsuccessful.
1696 int hfi1_register_ib_device(struct hfi1_devdata *dd)
1698 struct hfi1_ibdev *dev = &dd->verbs_dev;
1699 struct ib_device *ibdev = &dev->rdi.ibdev;
1700 struct hfi1_pportdata *ppd = dd->pport;
1701 struct hfi1_ibport *ibp = &ppd->ibport_data;
1704 size_t lcpysz = IB_DEVICE_NAME_MAX;
1706 for (i = 0; i < dd->num_pports; i++)
1707 init_ibport(ppd + i);
1709 /* Only need to initialize non-zero fields. */
1711 setup_timer(&dev->mem_timer, mem_timer, (unsigned long)dev);
1713 seqlock_init(&dev->iowait_lock);
1714 seqlock_init(&dev->txwait_lock);
1715 INIT_LIST_HEAD(&dev->txwait);
1716 INIT_LIST_HEAD(&dev->memwait);
1718 ret = verbs_txreq_init(dev);
1720 goto err_verbs_txreq;
1722 /* Use first-port GUID as node guid */
1723 ibdev->node_guid = get_sguid(ibp, HFI1_PORT_GUID_INDEX);
1726 * The system image GUID is supposed to be the same for all
1727 * HFIs in a single system but since there can be other
1728 * device types in the system, we can't be sure this is unique.
1730 if (!ib_hfi1_sys_image_guid)
1731 ib_hfi1_sys_image_guid = ibdev->node_guid;
1732 lcpysz = strlcpy(ibdev->name, class_name(), lcpysz);
1733 strlcpy(ibdev->name + lcpysz, "_%d", IB_DEVICE_NAME_MAX - lcpysz);
1734 ibdev->owner = THIS_MODULE;
1735 ibdev->phys_port_cnt = dd->num_pports;
1736 ibdev->dev.parent = &dd->pcidev->dev;
1737 ibdev->modify_device = modify_device;
1738 ibdev->alloc_hw_stats = alloc_hw_stats;
1739 ibdev->get_hw_stats = get_hw_stats;
1741 /* keep process mad in the driver */
1742 ibdev->process_mad = hfi1_process_mad;
1743 ibdev->get_dev_fw_str = hfi1_get_dev_fw_str;
1745 strncpy(ibdev->node_desc, init_utsname()->nodename,
1746 sizeof(ibdev->node_desc));
1749 * Fill in rvt info object.
1751 dd->verbs_dev.rdi.driver_f.port_callback = hfi1_create_port_files;
1752 dd->verbs_dev.rdi.driver_f.get_card_name = get_card_name;
1753 dd->verbs_dev.rdi.driver_f.get_pci_dev = get_pci_dev;
1754 dd->verbs_dev.rdi.driver_f.check_ah = hfi1_check_ah;
1755 dd->verbs_dev.rdi.driver_f.notify_new_ah = hfi1_notify_new_ah;
1756 dd->verbs_dev.rdi.driver_f.get_guid_be = hfi1_get_guid_be;
1757 dd->verbs_dev.rdi.driver_f.query_port_state = query_port;
1758 dd->verbs_dev.rdi.driver_f.shut_down_port = shut_down_port;
1759 dd->verbs_dev.rdi.driver_f.cap_mask_chg = hfi1_cap_mask_chg;
1761 * Fill in rvt info device attributes.
1763 hfi1_fill_device_attr(dd);
1766 dd->verbs_dev.rdi.dparms.qp_table_size = hfi1_qp_table_size;
1767 dd->verbs_dev.rdi.dparms.qpn_start = 0;
1768 dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1769 dd->verbs_dev.rdi.dparms.qos_shift = dd->qos_shift;
1770 dd->verbs_dev.rdi.dparms.qpn_res_start = kdeth_qp << 16;
1771 dd->verbs_dev.rdi.dparms.qpn_res_end =
1772 dd->verbs_dev.rdi.dparms.qpn_res_start + 65535;
1773 dd->verbs_dev.rdi.dparms.max_rdma_atomic = HFI1_MAX_RDMA_ATOMIC;
1774 dd->verbs_dev.rdi.dparms.psn_mask = PSN_MASK;
1775 dd->verbs_dev.rdi.dparms.psn_shift = PSN_SHIFT;
1776 dd->verbs_dev.rdi.dparms.psn_modify_mask = PSN_MODIFY_MASK;
1777 dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_INTEL_OPA;
1778 dd->verbs_dev.rdi.dparms.max_mad_size = OPA_MGMT_MAD_SIZE;
1780 dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qp_priv_alloc;
1781 dd->verbs_dev.rdi.driver_f.qp_priv_free = qp_priv_free;
1782 dd->verbs_dev.rdi.driver_f.free_all_qps = free_all_qps;
1783 dd->verbs_dev.rdi.driver_f.notify_qp_reset = notify_qp_reset;
1784 dd->verbs_dev.rdi.driver_f.do_send = hfi1_do_send;
1785 dd->verbs_dev.rdi.driver_f.schedule_send = hfi1_schedule_send;
1786 dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _hfi1_schedule_send;
1787 dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = get_pmtu_from_attr;
1788 dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1789 dd->verbs_dev.rdi.driver_f.flush_qp_waiters = flush_qp_waiters;
1790 dd->verbs_dev.rdi.driver_f.stop_send_queue = stop_send_queue;
1791 dd->verbs_dev.rdi.driver_f.quiesce_qp = quiesce_qp;
1792 dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1793 dd->verbs_dev.rdi.driver_f.mtu_from_qp = mtu_from_qp;
1794 dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = mtu_to_path_mtu;
1795 dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
1796 dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
1797 dd->verbs_dev.rdi.driver_f.notify_restart_rc = hfi1_restart_rc;
1798 dd->verbs_dev.rdi.driver_f.check_send_wqe = hfi1_check_send_wqe;
1800 /* completeion queue */
1801 snprintf(dd->verbs_dev.rdi.dparms.cq_name,
1802 sizeof(dd->verbs_dev.rdi.dparms.cq_name),
1803 "hfi1_cq%d", dd->unit);
1804 dd->verbs_dev.rdi.dparms.node = dd->node;
1807 dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
1808 dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
1809 dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1810 dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
1812 /* post send table */
1813 dd->verbs_dev.rdi.post_parms = hfi1_post_parms;
1816 for (i = 0; i < dd->num_pports; i++, ppd++)
1817 rvt_init_port(&dd->verbs_dev.rdi,
1818 &ppd->ibport_data.rvp,
1822 ret = rvt_register_device(&dd->verbs_dev.rdi);
1824 goto err_verbs_txreq;
1826 ret = hfi1_verbs_register_sysfs(dd);
1833 rvt_unregister_device(&dd->verbs_dev.rdi);
1835 verbs_txreq_exit(dev);
1836 dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1840 void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
1842 struct hfi1_ibdev *dev = &dd->verbs_dev;
1844 hfi1_verbs_unregister_sysfs(dd);
1846 rvt_unregister_device(&dd->verbs_dev.rdi);
1848 if (!list_empty(&dev->txwait))
1849 dd_dev_err(dd, "txwait list not empty!\n");
1850 if (!list_empty(&dev->memwait))
1851 dd_dev_err(dd, "memwait list not empty!\n");
1853 del_timer_sync(&dev->mem_timer);
1854 verbs_txreq_exit(dev);
1856 mutex_lock(&cntr_names_lock);
1857 kfree(dev_cntr_names);
1858 kfree(port_cntr_names);
1859 dev_cntr_names = NULL;
1860 port_cntr_names = NULL;
1861 cntr_names_initialized = 0;
1862 mutex_unlock(&cntr_names_lock);
1865 void hfi1_cnp_rcv(struct hfi1_packet *packet)
1867 struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1868 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1869 struct ib_header *hdr = packet->hdr;
1870 struct rvt_qp *qp = packet->qp;
1873 u8 sl, sc5, svc_type;
1875 switch (packet->qp->ibqp.qp_type) {
1877 rlid = qp->remote_ah_attr.dlid;
1878 rqpn = qp->remote_qpn;
1879 svc_type = IB_CC_SVCTYPE_UC;
1882 rlid = qp->remote_ah_attr.dlid;
1883 rqpn = qp->remote_qpn;
1884 svc_type = IB_CC_SVCTYPE_RC;
1889 svc_type = IB_CC_SVCTYPE_UD;
1892 ibp->rvp.n_pkt_drops++;
1896 sc5 = hdr2sc(hdr, packet->rhf);
1897 sl = ibp->sc_to_sl[sc5];
1898 lqpn = qp->ibqp.qp_num;
1900 process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);