xref: /openbmc/linux/drivers/infiniband/hw/hfi1/verbs.c (revision 8fdf9062)
1 /*
2  * Copyright(c) 2015 - 2018 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <rdma/ib_mad.h>
49 #include <rdma/ib_user_verbs.h>
50 #include <linux/io.h>
51 #include <linux/module.h>
52 #include <linux/utsname.h>
53 #include <linux/rculist.h>
54 #include <linux/mm.h>
55 #include <linux/vmalloc.h>
56 #include <rdma/opa_addr.h>
57 
58 #include "hfi.h"
59 #include "common.h"
60 #include "device.h"
61 #include "trace.h"
62 #include "qp.h"
63 #include "verbs_txreq.h"
64 #include "debugfs.h"
65 #include "vnic.h"
66 #include "fault.h"
67 #include "affinity.h"
68 
69 static unsigned int hfi1_lkey_table_size = 16;
70 module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
71 		   S_IRUGO);
72 MODULE_PARM_DESC(lkey_table_size,
73 		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
74 
75 static unsigned int hfi1_max_pds = 0xFFFF;
76 module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
77 MODULE_PARM_DESC(max_pds,
78 		 "Maximum number of protection domains to support");
79 
80 static unsigned int hfi1_max_ahs = 0xFFFF;
81 module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
82 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
83 
84 unsigned int hfi1_max_cqes = 0x2FFFFF;
85 module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
86 MODULE_PARM_DESC(max_cqes,
87 		 "Maximum number of completion queue entries to support");
88 
89 unsigned int hfi1_max_cqs = 0x1FFFF;
90 module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
91 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
92 
93 unsigned int hfi1_max_qp_wrs = 0x3FFF;
94 module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
95 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
96 
97 unsigned int hfi1_max_qps = 32768;
98 module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
99 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
100 
101 unsigned int hfi1_max_sges = 0x60;
102 module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
103 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
104 
105 unsigned int hfi1_max_mcast_grps = 16384;
106 module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
107 MODULE_PARM_DESC(max_mcast_grps,
108 		 "Maximum number of multicast groups to support");
109 
110 unsigned int hfi1_max_mcast_qp_attached = 16;
111 module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
112 		   uint, S_IRUGO);
113 MODULE_PARM_DESC(max_mcast_qp_attached,
114 		 "Maximum number of attached QPs to support");
115 
116 unsigned int hfi1_max_srqs = 1024;
117 module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
118 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
119 
120 unsigned int hfi1_max_srq_sges = 128;
121 module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
122 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
123 
124 unsigned int hfi1_max_srq_wrs = 0x1FFFF;
125 module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
126 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
127 
128 unsigned short piothreshold = 256;
129 module_param(piothreshold, ushort, S_IRUGO);
130 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
131 
132 static unsigned int sge_copy_mode;
133 module_param(sge_copy_mode, uint, S_IRUGO);
134 MODULE_PARM_DESC(sge_copy_mode,
135 		 "Verbs copy mode: 0 use memcpy, 1 use cacheless copy, 2 adapt based on WSS");
136 
137 static void verbs_sdma_complete(
138 	struct sdma_txreq *cookie,
139 	int status);
140 
141 static int pio_wait(struct rvt_qp *qp,
142 		    struct send_context *sc,
143 		    struct hfi1_pkt_state *ps,
144 		    u32 flag);
145 
146 /* Length of buffer to create verbs txreq cache name */
147 #define TXREQ_NAME_LEN 24
148 
149 /* 16B trailing buffer */
150 static const u8 trail_buf[MAX_16B_PADDING];
151 
152 static uint wss_threshold = 80;
153 module_param(wss_threshold, uint, S_IRUGO);
154 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
155 static uint wss_clean_period = 256;
156 module_param(wss_clean_period, uint, S_IRUGO);
157 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
158 
159 /*
160  * Translate ib_wr_opcode into ib_wc_opcode.
161  */
162 const enum ib_wc_opcode ib_hfi1_wc_opcode[] = {
163 	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
164 	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
165 	[IB_WR_SEND] = IB_WC_SEND,
166 	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
167 	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
168 	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
169 	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD,
170 	[IB_WR_SEND_WITH_INV] = IB_WC_SEND,
171 	[IB_WR_LOCAL_INV] = IB_WC_LOCAL_INV,
172 	[IB_WR_REG_MR] = IB_WC_REG_MR
173 };
174 
175 /*
176  * Length of header by opcode, 0 --> not supported
177  */
178 const u8 hdr_len_by_opcode[256] = {
179 	/* RC */
180 	[IB_OPCODE_RC_SEND_FIRST]                     = 12 + 8,
181 	[IB_OPCODE_RC_SEND_MIDDLE]                    = 12 + 8,
182 	[IB_OPCODE_RC_SEND_LAST]                      = 12 + 8,
183 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
184 	[IB_OPCODE_RC_SEND_ONLY]                      = 12 + 8,
185 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
186 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
187 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = 12 + 8,
188 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = 12 + 8,
189 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
190 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
191 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
192 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = 12 + 8 + 16,
193 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = 12 + 8 + 4,
194 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = 12 + 8,
195 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = 12 + 8 + 4,
196 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = 12 + 8 + 4,
197 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = 12 + 8 + 4,
198 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = 12 + 8 + 4 + 8,
199 	[IB_OPCODE_RC_COMPARE_SWAP]                   = 12 + 8 + 28,
200 	[IB_OPCODE_RC_FETCH_ADD]                      = 12 + 8 + 28,
201 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = 12 + 8 + 4,
202 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = 12 + 8 + 4,
203 	/* UC */
204 	[IB_OPCODE_UC_SEND_FIRST]                     = 12 + 8,
205 	[IB_OPCODE_UC_SEND_MIDDLE]                    = 12 + 8,
206 	[IB_OPCODE_UC_SEND_LAST]                      = 12 + 8,
207 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
208 	[IB_OPCODE_UC_SEND_ONLY]                      = 12 + 8,
209 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
210 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
211 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = 12 + 8,
212 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = 12 + 8,
213 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
214 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
215 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
216 	/* UD */
217 	[IB_OPCODE_UD_SEND_ONLY]                      = 12 + 8 + 8,
218 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 12
219 };
220 
221 static const opcode_handler opcode_handler_tbl[256] = {
222 	/* RC */
223 	[IB_OPCODE_RC_SEND_FIRST]                     = &hfi1_rc_rcv,
224 	[IB_OPCODE_RC_SEND_MIDDLE]                    = &hfi1_rc_rcv,
225 	[IB_OPCODE_RC_SEND_LAST]                      = &hfi1_rc_rcv,
226 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
227 	[IB_OPCODE_RC_SEND_ONLY]                      = &hfi1_rc_rcv,
228 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
229 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = &hfi1_rc_rcv,
230 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = &hfi1_rc_rcv,
231 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = &hfi1_rc_rcv,
232 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
233 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = &hfi1_rc_rcv,
234 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
235 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = &hfi1_rc_rcv,
236 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = &hfi1_rc_rcv,
237 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = &hfi1_rc_rcv,
238 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = &hfi1_rc_rcv,
239 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = &hfi1_rc_rcv,
240 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = &hfi1_rc_rcv,
241 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = &hfi1_rc_rcv,
242 	[IB_OPCODE_RC_COMPARE_SWAP]                   = &hfi1_rc_rcv,
243 	[IB_OPCODE_RC_FETCH_ADD]                      = &hfi1_rc_rcv,
244 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = &hfi1_rc_rcv,
245 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = &hfi1_rc_rcv,
246 	/* UC */
247 	[IB_OPCODE_UC_SEND_FIRST]                     = &hfi1_uc_rcv,
248 	[IB_OPCODE_UC_SEND_MIDDLE]                    = &hfi1_uc_rcv,
249 	[IB_OPCODE_UC_SEND_LAST]                      = &hfi1_uc_rcv,
250 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
251 	[IB_OPCODE_UC_SEND_ONLY]                      = &hfi1_uc_rcv,
252 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
253 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = &hfi1_uc_rcv,
254 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = &hfi1_uc_rcv,
255 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = &hfi1_uc_rcv,
256 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
257 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = &hfi1_uc_rcv,
258 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
259 	/* UD */
260 	[IB_OPCODE_UD_SEND_ONLY]                      = &hfi1_ud_rcv,
261 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_ud_rcv,
262 	/* CNP */
263 	[IB_OPCODE_CNP]				      = &hfi1_cnp_rcv
264 };
265 
266 #define OPMASK 0x1f
267 
268 static const u32 pio_opmask[BIT(3)] = {
269 	/* RC */
270 	[IB_OPCODE_RC >> 5] =
271 		BIT(RC_OP(SEND_ONLY) & OPMASK) |
272 		BIT(RC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
273 		BIT(RC_OP(RDMA_WRITE_ONLY) & OPMASK) |
274 		BIT(RC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK) |
275 		BIT(RC_OP(RDMA_READ_REQUEST) & OPMASK) |
276 		BIT(RC_OP(ACKNOWLEDGE) & OPMASK) |
277 		BIT(RC_OP(ATOMIC_ACKNOWLEDGE) & OPMASK) |
278 		BIT(RC_OP(COMPARE_SWAP) & OPMASK) |
279 		BIT(RC_OP(FETCH_ADD) & OPMASK),
280 	/* UC */
281 	[IB_OPCODE_UC >> 5] =
282 		BIT(UC_OP(SEND_ONLY) & OPMASK) |
283 		BIT(UC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
284 		BIT(UC_OP(RDMA_WRITE_ONLY) & OPMASK) |
285 		BIT(UC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK),
286 };
287 
288 /*
289  * System image GUID.
290  */
291 __be64 ib_hfi1_sys_image_guid;
292 
293 /*
294  * Make sure the QP is ready and able to accept the given opcode.
295  */
296 static inline opcode_handler qp_ok(struct hfi1_packet *packet)
297 {
298 	if (!(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
299 		return NULL;
300 	if (((packet->opcode & RVT_OPCODE_QP_MASK) ==
301 	     packet->qp->allowed_ops) ||
302 	    (packet->opcode == IB_OPCODE_CNP))
303 		return opcode_handler_tbl[packet->opcode];
304 
305 	return NULL;
306 }
307 
308 static u64 hfi1_fault_tx(struct rvt_qp *qp, u8 opcode, u64 pbc)
309 {
310 #ifdef CONFIG_FAULT_INJECTION
311 	if ((opcode & IB_OPCODE_MSP) == IB_OPCODE_MSP)
312 		/*
313 		 * In order to drop non-IB traffic we
314 		 * set PbcInsertHrc to NONE (0x2).
315 		 * The packet will still be delivered
316 		 * to the receiving node but a
317 		 * KHdrHCRCErr (KDETH packet with a bad
318 		 * HCRC) will be triggered and the
319 		 * packet will not be delivered to the
320 		 * correct context.
321 		 */
322 		pbc |= (u64)PBC_IHCRC_NONE << PBC_INSERT_HCRC_SHIFT;
323 	else
324 		/*
325 		 * In order to drop regular verbs
326 		 * traffic we set the PbcTestEbp
327 		 * flag. The packet will still be
328 		 * delivered to the receiving node but
329 		 * a 'late ebp error' will be
330 		 * triggered and will be dropped.
331 		 */
332 		pbc |= PBC_TEST_EBP;
333 #endif
334 	return pbc;
335 }
336 
337 static int hfi1_do_pkey_check(struct hfi1_packet *packet)
338 {
339 	struct hfi1_ctxtdata *rcd = packet->rcd;
340 	struct hfi1_pportdata *ppd = rcd->ppd;
341 	struct hfi1_16b_header *hdr = packet->hdr;
342 	u16 pkey;
343 
344 	/* Pkey check needed only for bypass packets */
345 	if (packet->etype != RHF_RCV_TYPE_BYPASS)
346 		return 0;
347 
348 	/* Perform pkey check */
349 	pkey = hfi1_16B_get_pkey(hdr);
350 	return ingress_pkey_check(ppd, pkey, packet->sc,
351 				  packet->qp->s_pkey_index,
352 				  packet->slid, true);
353 }
354 
355 static inline void hfi1_handle_packet(struct hfi1_packet *packet,
356 				      bool is_mcast)
357 {
358 	u32 qp_num;
359 	struct hfi1_ctxtdata *rcd = packet->rcd;
360 	struct hfi1_pportdata *ppd = rcd->ppd;
361 	struct hfi1_ibport *ibp = rcd_to_iport(rcd);
362 	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
363 	opcode_handler packet_handler;
364 	unsigned long flags;
365 
366 	inc_opstats(packet->tlen, &rcd->opstats->stats[packet->opcode]);
367 
368 	if (unlikely(is_mcast)) {
369 		struct rvt_mcast *mcast;
370 		struct rvt_mcast_qp *p;
371 
372 		if (!packet->grh)
373 			goto drop;
374 		mcast = rvt_mcast_find(&ibp->rvp,
375 				       &packet->grh->dgid,
376 				       opa_get_lid(packet->dlid, 9B));
377 		if (!mcast)
378 			goto drop;
379 		list_for_each_entry_rcu(p, &mcast->qp_list, list) {
380 			packet->qp = p->qp;
381 			if (hfi1_do_pkey_check(packet))
382 				goto drop;
383 			spin_lock_irqsave(&packet->qp->r_lock, flags);
384 			packet_handler = qp_ok(packet);
385 			if (likely(packet_handler))
386 				packet_handler(packet);
387 			else
388 				ibp->rvp.n_pkt_drops++;
389 			spin_unlock_irqrestore(&packet->qp->r_lock, flags);
390 		}
391 		/*
392 		 * Notify rvt_multicast_detach() if it is waiting for us
393 		 * to finish.
394 		 */
395 		if (atomic_dec_return(&mcast->refcount) <= 1)
396 			wake_up(&mcast->wait);
397 	} else {
398 		/* Get the destination QP number. */
399 		if (packet->etype == RHF_RCV_TYPE_BYPASS &&
400 		    hfi1_16B_get_l4(packet->hdr) == OPA_16B_L4_FM)
401 			qp_num = hfi1_16B_get_dest_qpn(packet->mgmt);
402 		else
403 			qp_num = ib_bth_get_qpn(packet->ohdr);
404 
405 		rcu_read_lock();
406 		packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
407 		if (!packet->qp)
408 			goto unlock_drop;
409 
410 		if (hfi1_do_pkey_check(packet))
411 			goto unlock_drop;
412 
413 		spin_lock_irqsave(&packet->qp->r_lock, flags);
414 		packet_handler = qp_ok(packet);
415 		if (likely(packet_handler))
416 			packet_handler(packet);
417 		else
418 			ibp->rvp.n_pkt_drops++;
419 		spin_unlock_irqrestore(&packet->qp->r_lock, flags);
420 		rcu_read_unlock();
421 	}
422 	return;
423 unlock_drop:
424 	rcu_read_unlock();
425 drop:
426 	ibp->rvp.n_pkt_drops++;
427 }
428 
429 /**
430  * hfi1_ib_rcv - process an incoming packet
431  * @packet: data packet information
432  *
433  * This is called to process an incoming packet at interrupt level.
434  */
435 void hfi1_ib_rcv(struct hfi1_packet *packet)
436 {
437 	struct hfi1_ctxtdata *rcd = packet->rcd;
438 
439 	trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));
440 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
441 }
442 
443 void hfi1_16B_rcv(struct hfi1_packet *packet)
444 {
445 	struct hfi1_ctxtdata *rcd = packet->rcd;
446 
447 	trace_input_ibhdr(rcd->dd, packet, false);
448 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
449 }
450 
451 /*
452  * This is called from a timer to check for QPs
453  * which need kernel memory in order to send a packet.
454  */
455 static void mem_timer(struct timer_list *t)
456 {
457 	struct hfi1_ibdev *dev = from_timer(dev, t, mem_timer);
458 	struct list_head *list = &dev->memwait;
459 	struct rvt_qp *qp = NULL;
460 	struct iowait *wait;
461 	unsigned long flags;
462 	struct hfi1_qp_priv *priv;
463 
464 	write_seqlock_irqsave(&dev->iowait_lock, flags);
465 	if (!list_empty(list)) {
466 		wait = list_first_entry(list, struct iowait, list);
467 		qp = iowait_to_qp(wait);
468 		priv = qp->priv;
469 		list_del_init(&priv->s_iowait.list);
470 		priv->s_iowait.lock = NULL;
471 		/* refcount held until actual wake up */
472 		if (!list_empty(list))
473 			mod_timer(&dev->mem_timer, jiffies + 1);
474 	}
475 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
476 
477 	if (qp)
478 		hfi1_qp_wakeup(qp, RVT_S_WAIT_KMEM);
479 }
480 
481 /*
482  * This is called with progress side lock held.
483  */
484 /* New API */
485 static void verbs_sdma_complete(
486 	struct sdma_txreq *cookie,
487 	int status)
488 {
489 	struct verbs_txreq *tx =
490 		container_of(cookie, struct verbs_txreq, txreq);
491 	struct rvt_qp *qp = tx->qp;
492 
493 	spin_lock(&qp->s_lock);
494 	if (tx->wqe) {
495 		rvt_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
496 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
497 		struct hfi1_opa_header *hdr;
498 
499 		hdr = &tx->phdr.hdr;
500 		hfi1_rc_send_complete(qp, hdr);
501 	}
502 	spin_unlock(&qp->s_lock);
503 
504 	hfi1_put_txreq(tx);
505 }
506 
507 static int wait_kmem(struct hfi1_ibdev *dev,
508 		     struct rvt_qp *qp,
509 		     struct hfi1_pkt_state *ps)
510 {
511 	struct hfi1_qp_priv *priv = qp->priv;
512 	unsigned long flags;
513 	int ret = 0;
514 
515 	spin_lock_irqsave(&qp->s_lock, flags);
516 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
517 		write_seqlock(&dev->iowait_lock);
518 		list_add_tail(&ps->s_txreq->txreq.list,
519 			      &ps->wait->tx_head);
520 		if (list_empty(&priv->s_iowait.list)) {
521 			if (list_empty(&dev->memwait))
522 				mod_timer(&dev->mem_timer, jiffies + 1);
523 			qp->s_flags |= RVT_S_WAIT_KMEM;
524 			list_add_tail(&priv->s_iowait.list, &dev->memwait);
525 			priv->s_iowait.lock = &dev->iowait_lock;
526 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM);
527 			rvt_get_qp(qp);
528 		}
529 		write_sequnlock(&dev->iowait_lock);
530 		hfi1_qp_unbusy(qp, ps->wait);
531 		ret = -EBUSY;
532 	}
533 	spin_unlock_irqrestore(&qp->s_lock, flags);
534 
535 	return ret;
536 }
537 
538 /*
539  * This routine calls txadds for each sg entry.
540  *
541  * Add failures will revert the sge cursor
542  */
543 static noinline int build_verbs_ulp_payload(
544 	struct sdma_engine *sde,
545 	u32 length,
546 	struct verbs_txreq *tx)
547 {
548 	struct rvt_sge_state *ss = tx->ss;
549 	struct rvt_sge *sg_list = ss->sg_list;
550 	struct rvt_sge sge = ss->sge;
551 	u8 num_sge = ss->num_sge;
552 	u32 len;
553 	int ret = 0;
554 
555 	while (length) {
556 		len = ss->sge.length;
557 		if (len > length)
558 			len = length;
559 		if (len > ss->sge.sge_length)
560 			len = ss->sge.sge_length;
561 		WARN_ON_ONCE(len == 0);
562 		ret = sdma_txadd_kvaddr(
563 			sde->dd,
564 			&tx->txreq,
565 			ss->sge.vaddr,
566 			len);
567 		if (ret)
568 			goto bail_txadd;
569 		rvt_update_sge(ss, len, false);
570 		length -= len;
571 	}
572 	return ret;
573 bail_txadd:
574 	/* unwind cursor */
575 	ss->sge = sge;
576 	ss->num_sge = num_sge;
577 	ss->sg_list = sg_list;
578 	return ret;
579 }
580 
581 /**
582  * update_tx_opstats - record stats by opcode
583  * @qp; the qp
584  * @ps: transmit packet state
585  * @plen: the plen in dwords
586  *
587  * This is a routine to record the tx opstats after a
588  * packet has been presented to the egress mechanism.
589  */
590 static void update_tx_opstats(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
591 			      u32 plen)
592 {
593 #ifdef CONFIG_DEBUG_FS
594 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
595 	struct hfi1_opcode_stats_perctx *s = get_cpu_ptr(dd->tx_opstats);
596 
597 	inc_opstats(plen * 4, &s->stats[ps->opcode]);
598 	put_cpu_ptr(s);
599 #endif
600 }
601 
602 /*
603  * Build the number of DMA descriptors needed to send length bytes of data.
604  *
605  * NOTE: DMA mapping is held in the tx until completed in the ring or
606  *       the tx desc is freed without having been submitted to the ring
607  *
608  * This routine ensures all the helper routine calls succeed.
609  */
610 /* New API */
611 static int build_verbs_tx_desc(
612 	struct sdma_engine *sde,
613 	u32 length,
614 	struct verbs_txreq *tx,
615 	struct hfi1_ahg_info *ahg_info,
616 	u64 pbc)
617 {
618 	int ret = 0;
619 	struct hfi1_sdma_header *phdr = &tx->phdr;
620 	u16 hdrbytes = (tx->hdr_dwords + sizeof(pbc) / 4) << 2;
621 	u8 extra_bytes = 0;
622 
623 	if (tx->phdr.hdr.hdr_type) {
624 		/*
625 		 * hdrbytes accounts for PBC. Need to subtract 8 bytes
626 		 * before calculating padding.
627 		 */
628 		extra_bytes = hfi1_get_16b_padding(hdrbytes - 8, length) +
629 			      (SIZE_OF_CRC << 2) + SIZE_OF_LT;
630 	}
631 	if (!ahg_info->ahgcount) {
632 		ret = sdma_txinit_ahg(
633 			&tx->txreq,
634 			ahg_info->tx_flags,
635 			hdrbytes + length +
636 			extra_bytes,
637 			ahg_info->ahgidx,
638 			0,
639 			NULL,
640 			0,
641 			verbs_sdma_complete);
642 		if (ret)
643 			goto bail_txadd;
644 		phdr->pbc = cpu_to_le64(pbc);
645 		ret = sdma_txadd_kvaddr(
646 			sde->dd,
647 			&tx->txreq,
648 			phdr,
649 			hdrbytes);
650 		if (ret)
651 			goto bail_txadd;
652 	} else {
653 		ret = sdma_txinit_ahg(
654 			&tx->txreq,
655 			ahg_info->tx_flags,
656 			length,
657 			ahg_info->ahgidx,
658 			ahg_info->ahgcount,
659 			ahg_info->ahgdesc,
660 			hdrbytes,
661 			verbs_sdma_complete);
662 		if (ret)
663 			goto bail_txadd;
664 	}
665 	/* add the ulp payload - if any. tx->ss can be NULL for acks */
666 	if (tx->ss) {
667 		ret = build_verbs_ulp_payload(sde, length, tx);
668 		if (ret)
669 			goto bail_txadd;
670 	}
671 
672 	/* add icrc, lt byte, and padding to flit */
673 	if (extra_bytes)
674 		ret = sdma_txadd_kvaddr(sde->dd, &tx->txreq,
675 					(void *)trail_buf, extra_bytes);
676 
677 bail_txadd:
678 	return ret;
679 }
680 
681 int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
682 			u64 pbc)
683 {
684 	struct hfi1_qp_priv *priv = qp->priv;
685 	struct hfi1_ahg_info *ahg_info = priv->s_ahg;
686 	u32 hdrwords = ps->s_txreq->hdr_dwords;
687 	u32 len = ps->s_txreq->s_cur_size;
688 	u32 plen;
689 	struct hfi1_ibdev *dev = ps->dev;
690 	struct hfi1_pportdata *ppd = ps->ppd;
691 	struct verbs_txreq *tx;
692 	u8 sc5 = priv->s_sc;
693 	int ret;
694 	u32 dwords;
695 
696 	if (ps->s_txreq->phdr.hdr.hdr_type) {
697 		u8 extra_bytes = hfi1_get_16b_padding((hdrwords << 2), len);
698 
699 		dwords = (len + extra_bytes + (SIZE_OF_CRC << 2) +
700 			  SIZE_OF_LT) >> 2;
701 	} else {
702 		dwords = (len + 3) >> 2;
703 	}
704 	plen = hdrwords + dwords + sizeof(pbc) / 4;
705 
706 	tx = ps->s_txreq;
707 	if (!sdma_txreq_built(&tx->txreq)) {
708 		if (likely(pbc == 0)) {
709 			u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
710 
711 			/* No vl15 here */
712 			/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
713 			if (ps->s_txreq->phdr.hdr.hdr_type)
714 				pbc |= PBC_PACKET_BYPASS |
715 				       PBC_INSERT_BYPASS_ICRC;
716 			else
717 				pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
718 
719 			if (unlikely(hfi1_dbg_should_fault_tx(qp, ps->opcode)))
720 				pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
721 			pbc = create_pbc(ppd,
722 					 pbc,
723 					 qp->srate_mbps,
724 					 vl,
725 					 plen);
726 		}
727 		tx->wqe = qp->s_wqe;
728 		ret = build_verbs_tx_desc(tx->sde, len, tx, ahg_info, pbc);
729 		if (unlikely(ret))
730 			goto bail_build;
731 	}
732 	ret =  sdma_send_txreq(tx->sde, ps->wait, &tx->txreq, ps->pkts_sent);
733 	if (unlikely(ret < 0)) {
734 		if (ret == -ECOMM)
735 			goto bail_ecomm;
736 		return ret;
737 	}
738 
739 	update_tx_opstats(qp, ps, plen);
740 	trace_sdma_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
741 				&ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
742 	return ret;
743 
744 bail_ecomm:
745 	/* The current one got "sent" */
746 	return 0;
747 bail_build:
748 	ret = wait_kmem(dev, qp, ps);
749 	if (!ret) {
750 		/* free txreq - bad state */
751 		hfi1_put_txreq(ps->s_txreq);
752 		ps->s_txreq = NULL;
753 	}
754 	return ret;
755 }
756 
757 /*
758  * If we are now in the error state, return zero to flush the
759  * send work request.
760  */
761 static int pio_wait(struct rvt_qp *qp,
762 		    struct send_context *sc,
763 		    struct hfi1_pkt_state *ps,
764 		    u32 flag)
765 {
766 	struct hfi1_qp_priv *priv = qp->priv;
767 	struct hfi1_devdata *dd = sc->dd;
768 	unsigned long flags;
769 	int ret = 0;
770 
771 	/*
772 	 * Note that as soon as want_buffer() is called and
773 	 * possibly before it returns, sc_piobufavail()
774 	 * could be called. Therefore, put QP on the I/O wait list before
775 	 * enabling the PIO avail interrupt.
776 	 */
777 	spin_lock_irqsave(&qp->s_lock, flags);
778 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
779 		write_seqlock(&sc->waitlock);
780 		list_add_tail(&ps->s_txreq->txreq.list,
781 			      &ps->wait->tx_head);
782 		if (list_empty(&priv->s_iowait.list)) {
783 			struct hfi1_ibdev *dev = &dd->verbs_dev;
784 			int was_empty;
785 
786 			dev->n_piowait += !!(flag & RVT_S_WAIT_PIO);
787 			dev->n_piodrain += !!(flag & HFI1_S_WAIT_PIO_DRAIN);
788 			qp->s_flags |= flag;
789 			was_empty = list_empty(&sc->piowait);
790 			iowait_queue(ps->pkts_sent, &priv->s_iowait,
791 				     &sc->piowait);
792 			priv->s_iowait.lock = &sc->waitlock;
793 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_PIO);
794 			rvt_get_qp(qp);
795 			/* counting: only call wantpiobuf_intr if first user */
796 			if (was_empty)
797 				hfi1_sc_wantpiobuf_intr(sc, 1);
798 		}
799 		write_sequnlock(&sc->waitlock);
800 		hfi1_qp_unbusy(qp, ps->wait);
801 		ret = -EBUSY;
802 	}
803 	spin_unlock_irqrestore(&qp->s_lock, flags);
804 	return ret;
805 }
806 
807 static void verbs_pio_complete(void *arg, int code)
808 {
809 	struct rvt_qp *qp = (struct rvt_qp *)arg;
810 	struct hfi1_qp_priv *priv = qp->priv;
811 
812 	if (iowait_pio_dec(&priv->s_iowait))
813 		iowait_drain_wakeup(&priv->s_iowait);
814 }
815 
816 int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
817 			u64 pbc)
818 {
819 	struct hfi1_qp_priv *priv = qp->priv;
820 	u32 hdrwords = ps->s_txreq->hdr_dwords;
821 	struct rvt_sge_state *ss = ps->s_txreq->ss;
822 	u32 len = ps->s_txreq->s_cur_size;
823 	u32 dwords;
824 	u32 plen;
825 	struct hfi1_pportdata *ppd = ps->ppd;
826 	u32 *hdr;
827 	u8 sc5;
828 	unsigned long flags = 0;
829 	struct send_context *sc;
830 	struct pio_buf *pbuf;
831 	int wc_status = IB_WC_SUCCESS;
832 	int ret = 0;
833 	pio_release_cb cb = NULL;
834 	u8 extra_bytes = 0;
835 
836 	if (ps->s_txreq->phdr.hdr.hdr_type) {
837 		u8 pad_size = hfi1_get_16b_padding((hdrwords << 2), len);
838 
839 		extra_bytes = pad_size + (SIZE_OF_CRC << 2) + SIZE_OF_LT;
840 		dwords = (len + extra_bytes) >> 2;
841 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.opah;
842 	} else {
843 		dwords = (len + 3) >> 2;
844 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.ibh;
845 	}
846 	plen = hdrwords + dwords + sizeof(pbc) / 4;
847 
848 	/* only RC/UC use complete */
849 	switch (qp->ibqp.qp_type) {
850 	case IB_QPT_RC:
851 	case IB_QPT_UC:
852 		cb = verbs_pio_complete;
853 		break;
854 	default:
855 		break;
856 	}
857 
858 	/* vl15 special case taken care of in ud.c */
859 	sc5 = priv->s_sc;
860 	sc = ps->s_txreq->psc;
861 
862 	if (likely(pbc == 0)) {
863 		u8 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
864 
865 		/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
866 		if (ps->s_txreq->phdr.hdr.hdr_type)
867 			pbc |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
868 		else
869 			pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
870 
871 		if (unlikely(hfi1_dbg_should_fault_tx(qp, ps->opcode)))
872 			pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
873 		pbc = create_pbc(ppd, pbc, qp->srate_mbps, vl, plen);
874 	}
875 	if (cb)
876 		iowait_pio_inc(&priv->s_iowait);
877 	pbuf = sc_buffer_alloc(sc, plen, cb, qp);
878 	if (unlikely(!pbuf)) {
879 		if (cb)
880 			verbs_pio_complete(qp, 0);
881 		if (ppd->host_link_state != HLS_UP_ACTIVE) {
882 			/*
883 			 * If we have filled the PIO buffers to capacity and are
884 			 * not in an active state this request is not going to
885 			 * go out to so just complete it with an error or else a
886 			 * ULP or the core may be stuck waiting.
887 			 */
888 			hfi1_cdbg(
889 				PIO,
890 				"alloc failed. state not active, completing");
891 			wc_status = IB_WC_GENERAL_ERR;
892 			goto pio_bail;
893 		} else {
894 			/*
895 			 * This is a normal occurrence. The PIO buffs are full
896 			 * up but we are still happily sending, well we could be
897 			 * so lets continue to queue the request.
898 			 */
899 			hfi1_cdbg(PIO, "alloc failed. state active, queuing");
900 			ret = pio_wait(qp, sc, ps, RVT_S_WAIT_PIO);
901 			if (!ret)
902 				/* txreq not queued - free */
903 				goto bail;
904 			/* tx consumed in wait */
905 			return ret;
906 		}
907 	}
908 
909 	if (dwords == 0) {
910 		pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
911 	} else {
912 		seg_pio_copy_start(pbuf, pbc,
913 				   hdr, hdrwords * 4);
914 		if (ss) {
915 			while (len) {
916 				void *addr = ss->sge.vaddr;
917 				u32 slen = ss->sge.length;
918 
919 				if (slen > len)
920 					slen = len;
921 				if (slen > ss->sge.sge_length)
922 					slen = ss->sge.sge_length;
923 				rvt_update_sge(ss, slen, false);
924 				seg_pio_copy_mid(pbuf, addr, slen);
925 				len -= slen;
926 			}
927 		}
928 		/* add icrc, lt byte, and padding to flit */
929 		if (extra_bytes)
930 			seg_pio_copy_mid(pbuf, trail_buf, extra_bytes);
931 
932 		seg_pio_copy_end(pbuf);
933 	}
934 
935 	update_tx_opstats(qp, ps, plen);
936 	trace_pio_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
937 			       &ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
938 
939 pio_bail:
940 	if (qp->s_wqe) {
941 		spin_lock_irqsave(&qp->s_lock, flags);
942 		rvt_send_complete(qp, qp->s_wqe, wc_status);
943 		spin_unlock_irqrestore(&qp->s_lock, flags);
944 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
945 		spin_lock_irqsave(&qp->s_lock, flags);
946 		hfi1_rc_send_complete(qp, &ps->s_txreq->phdr.hdr);
947 		spin_unlock_irqrestore(&qp->s_lock, flags);
948 	}
949 
950 	ret = 0;
951 
952 bail:
953 	hfi1_put_txreq(ps->s_txreq);
954 	return ret;
955 }
956 
957 /*
958  * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
959  * being an entry from the partition key table), return 0
960  * otherwise. Use the matching criteria for egress partition keys
961  * specified in the OPAv1 spec., section 9.1l.7.
962  */
963 static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
964 {
965 	u16 mkey = pkey & PKEY_LOW_15_MASK;
966 	u16 mentry = ent & PKEY_LOW_15_MASK;
967 
968 	if (mkey == mentry) {
969 		/*
970 		 * If pkey[15] is set (full partition member),
971 		 * is bit 15 in the corresponding table element
972 		 * clear (limited member)?
973 		 */
974 		if (pkey & PKEY_MEMBER_MASK)
975 			return !!(ent & PKEY_MEMBER_MASK);
976 		return 1;
977 	}
978 	return 0;
979 }
980 
981 /**
982  * egress_pkey_check - check P_KEY of a packet
983  * @ppd:  Physical IB port data
984  * @slid: SLID for packet
985  * @bkey: PKEY for header
986  * @sc5:  SC for packet
987  * @s_pkey_index: It will be used for look up optimization for kernel contexts
988  * only. If it is negative value, then it means user contexts is calling this
989  * function.
990  *
991  * It checks if hdr's pkey is valid.
992  *
993  * Return: 0 on success, otherwise, 1
994  */
995 int egress_pkey_check(struct hfi1_pportdata *ppd, u32 slid, u16 pkey,
996 		      u8 sc5, int8_t s_pkey_index)
997 {
998 	struct hfi1_devdata *dd;
999 	int i;
1000 	int is_user_ctxt_mechanism = (s_pkey_index < 0);
1001 
1002 	if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
1003 		return 0;
1004 
1005 	/* If SC15, pkey[0:14] must be 0x7fff */
1006 	if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
1007 		goto bad;
1008 
1009 	/* Is the pkey = 0x0, or 0x8000? */
1010 	if ((pkey & PKEY_LOW_15_MASK) == 0)
1011 		goto bad;
1012 
1013 	/*
1014 	 * For the kernel contexts only, if a qp is passed into the function,
1015 	 * the most likely matching pkey has index qp->s_pkey_index
1016 	 */
1017 	if (!is_user_ctxt_mechanism &&
1018 	    egress_pkey_matches_entry(pkey, ppd->pkeys[s_pkey_index])) {
1019 		return 0;
1020 	}
1021 
1022 	for (i = 0; i < MAX_PKEY_VALUES; i++) {
1023 		if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
1024 			return 0;
1025 	}
1026 bad:
1027 	/*
1028 	 * For the user-context mechanism, the P_KEY check would only happen
1029 	 * once per SDMA request, not once per packet.  Therefore, there's no
1030 	 * need to increment the counter for the user-context mechanism.
1031 	 */
1032 	if (!is_user_ctxt_mechanism) {
1033 		incr_cntr64(&ppd->port_xmit_constraint_errors);
1034 		dd = ppd->dd;
1035 		if (!(dd->err_info_xmit_constraint.status &
1036 		      OPA_EI_STATUS_SMASK)) {
1037 			dd->err_info_xmit_constraint.status |=
1038 				OPA_EI_STATUS_SMASK;
1039 			dd->err_info_xmit_constraint.slid = slid;
1040 			dd->err_info_xmit_constraint.pkey = pkey;
1041 		}
1042 	}
1043 	return 1;
1044 }
1045 
1046 /**
1047  * get_send_routine - choose an egress routine
1048  *
1049  * Choose an egress routine based on QP type
1050  * and size
1051  */
1052 static inline send_routine get_send_routine(struct rvt_qp *qp,
1053 					    struct hfi1_pkt_state *ps)
1054 {
1055 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1056 	struct hfi1_qp_priv *priv = qp->priv;
1057 	struct verbs_txreq *tx = ps->s_txreq;
1058 
1059 	if (unlikely(!(dd->flags & HFI1_HAS_SEND_DMA)))
1060 		return dd->process_pio_send;
1061 	switch (qp->ibqp.qp_type) {
1062 	case IB_QPT_SMI:
1063 		return dd->process_pio_send;
1064 	case IB_QPT_GSI:
1065 	case IB_QPT_UD:
1066 		break;
1067 	case IB_QPT_UC:
1068 	case IB_QPT_RC: {
1069 		if (piothreshold &&
1070 		    tx->s_cur_size <= min(piothreshold, qp->pmtu) &&
1071 		    (BIT(ps->opcode & OPMASK) & pio_opmask[ps->opcode >> 5]) &&
1072 		    iowait_sdma_pending(&priv->s_iowait) == 0 &&
1073 		    !sdma_txreq_built(&tx->txreq))
1074 			return dd->process_pio_send;
1075 		break;
1076 	}
1077 	default:
1078 		break;
1079 	}
1080 	return dd->process_dma_send;
1081 }
1082 
1083 /**
1084  * hfi1_verbs_send - send a packet
1085  * @qp: the QP to send on
1086  * @ps: the state of the packet to send
1087  *
1088  * Return zero if packet is sent or queued OK.
1089  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1090  */
1091 int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
1092 {
1093 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1094 	struct hfi1_qp_priv *priv = qp->priv;
1095 	struct ib_other_headers *ohdr = NULL;
1096 	send_routine sr;
1097 	int ret;
1098 	u16 pkey;
1099 	u32 slid;
1100 	u8 l4 = 0;
1101 
1102 	/* locate the pkey within the headers */
1103 	if (ps->s_txreq->phdr.hdr.hdr_type) {
1104 		struct hfi1_16b_header *hdr = &ps->s_txreq->phdr.hdr.opah;
1105 
1106 		l4 = hfi1_16B_get_l4(hdr);
1107 		if (l4 == OPA_16B_L4_IB_LOCAL)
1108 			ohdr = &hdr->u.oth;
1109 		else if (l4 == OPA_16B_L4_IB_GLOBAL)
1110 			ohdr = &hdr->u.l.oth;
1111 
1112 		slid = hfi1_16B_get_slid(hdr);
1113 		pkey = hfi1_16B_get_pkey(hdr);
1114 	} else {
1115 		struct ib_header *hdr = &ps->s_txreq->phdr.hdr.ibh;
1116 		u8 lnh = ib_get_lnh(hdr);
1117 
1118 		if (lnh == HFI1_LRH_GRH)
1119 			ohdr = &hdr->u.l.oth;
1120 		else
1121 			ohdr = &hdr->u.oth;
1122 		slid = ib_get_slid(hdr);
1123 		pkey = ib_bth_get_pkey(ohdr);
1124 	}
1125 
1126 	if (likely(l4 != OPA_16B_L4_FM))
1127 		ps->opcode = ib_bth_get_opcode(ohdr);
1128 	else
1129 		ps->opcode = IB_OPCODE_UD_SEND_ONLY;
1130 
1131 	sr = get_send_routine(qp, ps);
1132 	ret = egress_pkey_check(dd->pport, slid, pkey,
1133 				priv->s_sc, qp->s_pkey_index);
1134 	if (unlikely(ret)) {
1135 		/*
1136 		 * The value we are returning here does not get propagated to
1137 		 * the verbs caller. Thus we need to complete the request with
1138 		 * error otherwise the caller could be sitting waiting on the
1139 		 * completion event. Only do this for PIO. SDMA has its own
1140 		 * mechanism for handling the errors. So for SDMA we can just
1141 		 * return.
1142 		 */
1143 		if (sr == dd->process_pio_send) {
1144 			unsigned long flags;
1145 
1146 			hfi1_cdbg(PIO, "%s() Failed. Completing with err",
1147 				  __func__);
1148 			spin_lock_irqsave(&qp->s_lock, flags);
1149 			rvt_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
1150 			spin_unlock_irqrestore(&qp->s_lock, flags);
1151 		}
1152 		return -EINVAL;
1153 	}
1154 	if (sr == dd->process_dma_send && iowait_pio_pending(&priv->s_iowait))
1155 		return pio_wait(qp,
1156 				ps->s_txreq->psc,
1157 				ps,
1158 				HFI1_S_WAIT_PIO_DRAIN);
1159 	return sr(qp, ps, 0);
1160 }
1161 
1162 /**
1163  * hfi1_fill_device_attr - Fill in rvt dev info device attributes.
1164  * @dd: the device data structure
1165  */
1166 static void hfi1_fill_device_attr(struct hfi1_devdata *dd)
1167 {
1168 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1169 	u32 ver = dd->dc8051_ver;
1170 
1171 	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1172 
1173 	rdi->dparms.props.fw_ver = ((u64)(dc8051_ver_maj(ver)) << 32) |
1174 		((u64)(dc8051_ver_min(ver)) << 16) |
1175 		(u64)dc8051_ver_patch(ver);
1176 
1177 	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1178 			IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1179 			IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1180 			IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE |
1181 			IB_DEVICE_MEM_MGT_EXTENSIONS |
1182 			IB_DEVICE_RDMA_NETDEV_OPA_VNIC;
1183 	rdi->dparms.props.page_size_cap = PAGE_SIZE;
1184 	rdi->dparms.props.vendor_id = dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
1185 	rdi->dparms.props.vendor_part_id = dd->pcidev->device;
1186 	rdi->dparms.props.hw_ver = dd->minrev;
1187 	rdi->dparms.props.sys_image_guid = ib_hfi1_sys_image_guid;
1188 	rdi->dparms.props.max_mr_size = U64_MAX;
1189 	rdi->dparms.props.max_fast_reg_page_list_len = UINT_MAX;
1190 	rdi->dparms.props.max_qp = hfi1_max_qps;
1191 	rdi->dparms.props.max_qp_wr = hfi1_max_qp_wrs;
1192 	rdi->dparms.props.max_send_sge = hfi1_max_sges;
1193 	rdi->dparms.props.max_recv_sge = hfi1_max_sges;
1194 	rdi->dparms.props.max_sge_rd = hfi1_max_sges;
1195 	rdi->dparms.props.max_cq = hfi1_max_cqs;
1196 	rdi->dparms.props.max_ah = hfi1_max_ahs;
1197 	rdi->dparms.props.max_cqe = hfi1_max_cqes;
1198 	rdi->dparms.props.max_mr = rdi->lkey_table.max;
1199 	rdi->dparms.props.max_fmr = rdi->lkey_table.max;
1200 	rdi->dparms.props.max_map_per_fmr = 32767;
1201 	rdi->dparms.props.max_pd = hfi1_max_pds;
1202 	rdi->dparms.props.max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
1203 	rdi->dparms.props.max_qp_init_rd_atom = 255;
1204 	rdi->dparms.props.max_srq = hfi1_max_srqs;
1205 	rdi->dparms.props.max_srq_wr = hfi1_max_srq_wrs;
1206 	rdi->dparms.props.max_srq_sge = hfi1_max_srq_sges;
1207 	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1208 	rdi->dparms.props.max_pkeys = hfi1_get_npkeys(dd);
1209 	rdi->dparms.props.max_mcast_grp = hfi1_max_mcast_grps;
1210 	rdi->dparms.props.max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
1211 	rdi->dparms.props.max_total_mcast_qp_attach =
1212 					rdi->dparms.props.max_mcast_qp_attach *
1213 					rdi->dparms.props.max_mcast_grp;
1214 }
1215 
1216 static inline u16 opa_speed_to_ib(u16 in)
1217 {
1218 	u16 out = 0;
1219 
1220 	if (in & OPA_LINK_SPEED_25G)
1221 		out |= IB_SPEED_EDR;
1222 	if (in & OPA_LINK_SPEED_12_5G)
1223 		out |= IB_SPEED_FDR;
1224 
1225 	return out;
1226 }
1227 
1228 /*
1229  * Convert a single OPA link width (no multiple flags) to an IB value.
1230  * A zero OPA link width means link down, which means the IB width value
1231  * is a don't care.
1232  */
1233 static inline u16 opa_width_to_ib(u16 in)
1234 {
1235 	switch (in) {
1236 	case OPA_LINK_WIDTH_1X:
1237 	/* map 2x and 3x to 1x as they don't exist in IB */
1238 	case OPA_LINK_WIDTH_2X:
1239 	case OPA_LINK_WIDTH_3X:
1240 		return IB_WIDTH_1X;
1241 	default: /* link down or unknown, return our largest width */
1242 	case OPA_LINK_WIDTH_4X:
1243 		return IB_WIDTH_4X;
1244 	}
1245 }
1246 
1247 static int query_port(struct rvt_dev_info *rdi, u8 port_num,
1248 		      struct ib_port_attr *props)
1249 {
1250 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1251 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1252 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1253 	u32 lid = ppd->lid;
1254 
1255 	/* props being zeroed by the caller, avoid zeroing it here */
1256 	props->lid = lid ? lid : 0;
1257 	props->lmc = ppd->lmc;
1258 	/* OPA logical states match IB logical states */
1259 	props->state = driver_lstate(ppd);
1260 	props->phys_state = driver_pstate(ppd);
1261 	props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
1262 	props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
1263 	/* see rate_show() in ib core/sysfs.c */
1264 	props->active_speed = (u8)opa_speed_to_ib(ppd->link_speed_active);
1265 	props->max_vl_num = ppd->vls_supported;
1266 
1267 	/* Once we are a "first class" citizen and have added the OPA MTUs to
1268 	 * the core we can advertise the larger MTU enum to the ULPs, for now
1269 	 * advertise only 4K.
1270 	 *
1271 	 * Those applications which are either OPA aware or pass the MTU enum
1272 	 * from the Path Records to us will get the new 8k MTU.  Those that
1273 	 * attempt to process the MTU enum may fail in various ways.
1274 	 */
1275 	props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
1276 				      4096 : hfi1_max_mtu), IB_MTU_4096);
1277 	props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
1278 		mtu_to_enum(ppd->ibmtu, IB_MTU_4096);
1279 
1280 	return 0;
1281 }
1282 
1283 static int modify_device(struct ib_device *device,
1284 			 int device_modify_mask,
1285 			 struct ib_device_modify *device_modify)
1286 {
1287 	struct hfi1_devdata *dd = dd_from_ibdev(device);
1288 	unsigned i;
1289 	int ret;
1290 
1291 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1292 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1293 		ret = -EOPNOTSUPP;
1294 		goto bail;
1295 	}
1296 
1297 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1298 		memcpy(device->node_desc, device_modify->node_desc,
1299 		       IB_DEVICE_NODE_DESC_MAX);
1300 		for (i = 0; i < dd->num_pports; i++) {
1301 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1302 
1303 			hfi1_node_desc_chg(ibp);
1304 		}
1305 	}
1306 
1307 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1308 		ib_hfi1_sys_image_guid =
1309 			cpu_to_be64(device_modify->sys_image_guid);
1310 		for (i = 0; i < dd->num_pports; i++) {
1311 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1312 
1313 			hfi1_sys_guid_chg(ibp);
1314 		}
1315 	}
1316 
1317 	ret = 0;
1318 
1319 bail:
1320 	return ret;
1321 }
1322 
1323 static int shut_down_port(struct rvt_dev_info *rdi, u8 port_num)
1324 {
1325 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1326 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1327 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1328 	int ret;
1329 
1330 	set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
1331 			     OPA_LINKDOWN_REASON_UNKNOWN);
1332 	ret = set_link_state(ppd, HLS_DN_DOWNDEF);
1333 	return ret;
1334 }
1335 
1336 static int hfi1_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1337 			    int guid_index, __be64 *guid)
1338 {
1339 	struct hfi1_ibport *ibp = container_of(rvp, struct hfi1_ibport, rvp);
1340 
1341 	if (guid_index >= HFI1_GUIDS_PER_PORT)
1342 		return -EINVAL;
1343 
1344 	*guid = get_sguid(ibp, guid_index);
1345 	return 0;
1346 }
1347 
1348 /*
1349  * convert ah port,sl to sc
1350  */
1351 u8 ah_to_sc(struct ib_device *ibdev, struct rdma_ah_attr *ah)
1352 {
1353 	struct hfi1_ibport *ibp = to_iport(ibdev, rdma_ah_get_port_num(ah));
1354 
1355 	return ibp->sl_to_sc[rdma_ah_get_sl(ah)];
1356 }
1357 
1358 static int hfi1_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1359 {
1360 	struct hfi1_ibport *ibp;
1361 	struct hfi1_pportdata *ppd;
1362 	struct hfi1_devdata *dd;
1363 	u8 sc5;
1364 	u8 sl;
1365 
1366 	if (hfi1_check_mcast(rdma_ah_get_dlid(ah_attr)) &&
1367 	    !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1368 		return -EINVAL;
1369 
1370 	/* test the mapping for validity */
1371 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1372 	ppd = ppd_from_ibp(ibp);
1373 	dd = dd_from_ppd(ppd);
1374 
1375 	sl = rdma_ah_get_sl(ah_attr);
1376 	if (sl >= ARRAY_SIZE(ibp->sl_to_sc))
1377 		return -EINVAL;
1378 
1379 	sc5 = ibp->sl_to_sc[sl];
1380 	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1381 		return -EINVAL;
1382 	return 0;
1383 }
1384 
1385 static void hfi1_notify_new_ah(struct ib_device *ibdev,
1386 			       struct rdma_ah_attr *ah_attr,
1387 			       struct rvt_ah *ah)
1388 {
1389 	struct hfi1_ibport *ibp;
1390 	struct hfi1_pportdata *ppd;
1391 	struct hfi1_devdata *dd;
1392 	u8 sc5;
1393 	struct rdma_ah_attr *attr = &ah->attr;
1394 
1395 	/*
1396 	 * Do not trust reading anything from rvt_ah at this point as it is not
1397 	 * done being setup. We can however modify things which we need to set.
1398 	 */
1399 
1400 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1401 	ppd = ppd_from_ibp(ibp);
1402 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)];
1403 	hfi1_update_ah_attr(ibdev, attr);
1404 	hfi1_make_opa_lid(attr);
1405 	dd = dd_from_ppd(ppd);
1406 	ah->vl = sc_to_vlt(dd, sc5);
1407 	if (ah->vl < num_vls || ah->vl == 15)
1408 		ah->log_pmtu = ilog2(dd->vld[ah->vl].mtu);
1409 }
1410 
1411 /**
1412  * hfi1_get_npkeys - return the size of the PKEY table for context 0
1413  * @dd: the hfi1_ib device
1414  */
1415 unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
1416 {
1417 	return ARRAY_SIZE(dd->pport[0].pkeys);
1418 }
1419 
1420 static void init_ibport(struct hfi1_pportdata *ppd)
1421 {
1422 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1423 	size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
1424 	int i;
1425 
1426 	for (i = 0; i < sz; i++) {
1427 		ibp->sl_to_sc[i] = i;
1428 		ibp->sc_to_sl[i] = i;
1429 	}
1430 
1431 	for (i = 0; i < RVT_MAX_TRAP_LISTS ; i++)
1432 		INIT_LIST_HEAD(&ibp->rvp.trap_lists[i].list);
1433 	timer_setup(&ibp->rvp.trap_timer, hfi1_handle_trap_timer, 0);
1434 
1435 	spin_lock_init(&ibp->rvp.lock);
1436 	/* Set the prefix to the default value (see ch. 4.1.1) */
1437 	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1438 	ibp->rvp.sm_lid = 0;
1439 	/*
1440 	 * Below should only set bits defined in OPA PortInfo.CapabilityMask
1441 	 * and PortInfo.CapabilityMask3
1442 	 */
1443 	ibp->rvp.port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
1444 		IB_PORT_CAP_MASK_NOTICE_SUP;
1445 	ibp->rvp.port_cap3_flags = OPA_CAP_MASK3_IsSharedSpaceSupported;
1446 	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1447 	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1448 	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1449 	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1450 	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1451 
1452 	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1453 	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1454 }
1455 
1456 static void hfi1_get_dev_fw_str(struct ib_device *ibdev, char *str)
1457 {
1458 	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
1459 	struct hfi1_ibdev *dev = dev_from_rdi(rdi);
1460 	u32 ver = dd_from_dev(dev)->dc8051_ver;
1461 
1462 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u.%u", dc8051_ver_maj(ver),
1463 		 dc8051_ver_min(ver), dc8051_ver_patch(ver));
1464 }
1465 
1466 static const char * const driver_cntr_names[] = {
1467 	/* must be element 0*/
1468 	"DRIVER_KernIntr",
1469 	"DRIVER_ErrorIntr",
1470 	"DRIVER_Tx_Errs",
1471 	"DRIVER_Rcv_Errs",
1472 	"DRIVER_HW_Errs",
1473 	"DRIVER_NoPIOBufs",
1474 	"DRIVER_CtxtsOpen",
1475 	"DRIVER_RcvLen_Errs",
1476 	"DRIVER_EgrBufFull",
1477 	"DRIVER_EgrHdrFull"
1478 };
1479 
1480 static DEFINE_MUTEX(cntr_names_lock); /* protects the *_cntr_names bufers */
1481 static const char **dev_cntr_names;
1482 static const char **port_cntr_names;
1483 int num_driver_cntrs = ARRAY_SIZE(driver_cntr_names);
1484 static int num_dev_cntrs;
1485 static int num_port_cntrs;
1486 static int cntr_names_initialized;
1487 
1488 /*
1489  * Convert a list of names separated by '\n' into an array of NULL terminated
1490  * strings. Optionally some entries can be reserved in the array to hold extra
1491  * external strings.
1492  */
1493 static int init_cntr_names(const char *names_in,
1494 			   const size_t names_len,
1495 			   int num_extra_names,
1496 			   int *num_cntrs,
1497 			   const char ***cntr_names)
1498 {
1499 	char *names_out, *p, **q;
1500 	int i, n;
1501 
1502 	n = 0;
1503 	for (i = 0; i < names_len; i++)
1504 		if (names_in[i] == '\n')
1505 			n++;
1506 
1507 	names_out = kmalloc((n + num_extra_names) * sizeof(char *) + names_len,
1508 			    GFP_KERNEL);
1509 	if (!names_out) {
1510 		*num_cntrs = 0;
1511 		*cntr_names = NULL;
1512 		return -ENOMEM;
1513 	}
1514 
1515 	p = names_out + (n + num_extra_names) * sizeof(char *);
1516 	memcpy(p, names_in, names_len);
1517 
1518 	q = (char **)names_out;
1519 	for (i = 0; i < n; i++) {
1520 		q[i] = p;
1521 		p = strchr(p, '\n');
1522 		*p++ = '\0';
1523 	}
1524 
1525 	*num_cntrs = n;
1526 	*cntr_names = (const char **)names_out;
1527 	return 0;
1528 }
1529 
1530 static struct rdma_hw_stats *alloc_hw_stats(struct ib_device *ibdev,
1531 					    u8 port_num)
1532 {
1533 	int i, err;
1534 
1535 	mutex_lock(&cntr_names_lock);
1536 	if (!cntr_names_initialized) {
1537 		struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1538 
1539 		err = init_cntr_names(dd->cntrnames,
1540 				      dd->cntrnameslen,
1541 				      num_driver_cntrs,
1542 				      &num_dev_cntrs,
1543 				      &dev_cntr_names);
1544 		if (err) {
1545 			mutex_unlock(&cntr_names_lock);
1546 			return NULL;
1547 		}
1548 
1549 		for (i = 0; i < num_driver_cntrs; i++)
1550 			dev_cntr_names[num_dev_cntrs + i] =
1551 				driver_cntr_names[i];
1552 
1553 		err = init_cntr_names(dd->portcntrnames,
1554 				      dd->portcntrnameslen,
1555 				      0,
1556 				      &num_port_cntrs,
1557 				      &port_cntr_names);
1558 		if (err) {
1559 			kfree(dev_cntr_names);
1560 			dev_cntr_names = NULL;
1561 			mutex_unlock(&cntr_names_lock);
1562 			return NULL;
1563 		}
1564 		cntr_names_initialized = 1;
1565 	}
1566 	mutex_unlock(&cntr_names_lock);
1567 
1568 	if (!port_num)
1569 		return rdma_alloc_hw_stats_struct(
1570 				dev_cntr_names,
1571 				num_dev_cntrs + num_driver_cntrs,
1572 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1573 	else
1574 		return rdma_alloc_hw_stats_struct(
1575 				port_cntr_names,
1576 				num_port_cntrs,
1577 				RDMA_HW_STATS_DEFAULT_LIFESPAN);
1578 }
1579 
1580 static u64 hfi1_sps_ints(void)
1581 {
1582 	unsigned long flags;
1583 	struct hfi1_devdata *dd;
1584 	u64 sps_ints = 0;
1585 
1586 	spin_lock_irqsave(&hfi1_devs_lock, flags);
1587 	list_for_each_entry(dd, &hfi1_dev_list, list) {
1588 		sps_ints += get_all_cpu_total(dd->int_counter);
1589 	}
1590 	spin_unlock_irqrestore(&hfi1_devs_lock, flags);
1591 	return sps_ints;
1592 }
1593 
1594 static int get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1595 			u8 port, int index)
1596 {
1597 	u64 *values;
1598 	int count;
1599 
1600 	if (!port) {
1601 		u64 *stats = (u64 *)&hfi1_stats;
1602 		int i;
1603 
1604 		hfi1_read_cntrs(dd_from_ibdev(ibdev), NULL, &values);
1605 		values[num_dev_cntrs] = hfi1_sps_ints();
1606 		for (i = 1; i < num_driver_cntrs; i++)
1607 			values[num_dev_cntrs + i] = stats[i];
1608 		count = num_dev_cntrs + num_driver_cntrs;
1609 	} else {
1610 		struct hfi1_ibport *ibp = to_iport(ibdev, port);
1611 
1612 		hfi1_read_portcntrs(ppd_from_ibp(ibp), NULL, &values);
1613 		count = num_port_cntrs;
1614 	}
1615 
1616 	memcpy(stats->value, values, count * sizeof(u64));
1617 	return count;
1618 }
1619 
1620 static const struct ib_device_ops hfi1_dev_ops = {
1621 	.alloc_hw_stats = alloc_hw_stats,
1622 	.alloc_rdma_netdev = hfi1_vnic_alloc_rn,
1623 	.get_dev_fw_str = hfi1_get_dev_fw_str,
1624 	.get_hw_stats = get_hw_stats,
1625 	.modify_device = modify_device,
1626 	/* keep process mad in the driver */
1627 	.process_mad = hfi1_process_mad,
1628 };
1629 
1630 /**
1631  * hfi1_register_ib_device - register our device with the infiniband core
1632  * @dd: the device data structure
1633  * Return 0 if successful, errno if unsuccessful.
1634  */
1635 int hfi1_register_ib_device(struct hfi1_devdata *dd)
1636 {
1637 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1638 	struct ib_device *ibdev = &dev->rdi.ibdev;
1639 	struct hfi1_pportdata *ppd = dd->pport;
1640 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1641 	unsigned i;
1642 	int ret;
1643 
1644 	for (i = 0; i < dd->num_pports; i++)
1645 		init_ibport(ppd + i);
1646 
1647 	/* Only need to initialize non-zero fields. */
1648 
1649 	timer_setup(&dev->mem_timer, mem_timer, 0);
1650 
1651 	seqlock_init(&dev->iowait_lock);
1652 	seqlock_init(&dev->txwait_lock);
1653 	INIT_LIST_HEAD(&dev->txwait);
1654 	INIT_LIST_HEAD(&dev->memwait);
1655 
1656 	ret = verbs_txreq_init(dev);
1657 	if (ret)
1658 		goto err_verbs_txreq;
1659 
1660 	/* Use first-port GUID as node guid */
1661 	ibdev->node_guid = get_sguid(ibp, HFI1_PORT_GUID_INDEX);
1662 
1663 	/*
1664 	 * The system image GUID is supposed to be the same for all
1665 	 * HFIs in a single system but since there can be other
1666 	 * device types in the system, we can't be sure this is unique.
1667 	 */
1668 	if (!ib_hfi1_sys_image_guid)
1669 		ib_hfi1_sys_image_guid = ibdev->node_guid;
1670 	ibdev->owner = THIS_MODULE;
1671 	ibdev->phys_port_cnt = dd->num_pports;
1672 	ibdev->dev.parent = &dd->pcidev->dev;
1673 
1674 	ib_set_device_ops(ibdev, &hfi1_dev_ops);
1675 
1676 	strlcpy(ibdev->node_desc, init_utsname()->nodename,
1677 		sizeof(ibdev->node_desc));
1678 
1679 	/*
1680 	 * Fill in rvt info object.
1681 	 */
1682 	dd->verbs_dev.rdi.driver_f.port_callback = hfi1_create_port_files;
1683 	dd->verbs_dev.rdi.driver_f.get_pci_dev = get_pci_dev;
1684 	dd->verbs_dev.rdi.driver_f.check_ah = hfi1_check_ah;
1685 	dd->verbs_dev.rdi.driver_f.notify_new_ah = hfi1_notify_new_ah;
1686 	dd->verbs_dev.rdi.driver_f.get_guid_be = hfi1_get_guid_be;
1687 	dd->verbs_dev.rdi.driver_f.query_port_state = query_port;
1688 	dd->verbs_dev.rdi.driver_f.shut_down_port = shut_down_port;
1689 	dd->verbs_dev.rdi.driver_f.cap_mask_chg = hfi1_cap_mask_chg;
1690 	/*
1691 	 * Fill in rvt info device attributes.
1692 	 */
1693 	hfi1_fill_device_attr(dd);
1694 
1695 	/* queue pair */
1696 	dd->verbs_dev.rdi.dparms.qp_table_size = hfi1_qp_table_size;
1697 	dd->verbs_dev.rdi.dparms.qpn_start = 0;
1698 	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1699 	dd->verbs_dev.rdi.dparms.qos_shift = dd->qos_shift;
1700 	dd->verbs_dev.rdi.dparms.qpn_res_start = kdeth_qp << 16;
1701 	dd->verbs_dev.rdi.dparms.qpn_res_end =
1702 	dd->verbs_dev.rdi.dparms.qpn_res_start + 65535;
1703 	dd->verbs_dev.rdi.dparms.max_rdma_atomic = HFI1_MAX_RDMA_ATOMIC;
1704 	dd->verbs_dev.rdi.dparms.psn_mask = PSN_MASK;
1705 	dd->verbs_dev.rdi.dparms.psn_shift = PSN_SHIFT;
1706 	dd->verbs_dev.rdi.dparms.psn_modify_mask = PSN_MODIFY_MASK;
1707 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_INTEL_OPA |
1708 						RDMA_CORE_CAP_OPA_AH;
1709 	dd->verbs_dev.rdi.dparms.max_mad_size = OPA_MGMT_MAD_SIZE;
1710 
1711 	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qp_priv_alloc;
1712 	dd->verbs_dev.rdi.driver_f.qp_priv_init = hfi1_qp_priv_init;
1713 	dd->verbs_dev.rdi.driver_f.qp_priv_free = qp_priv_free;
1714 	dd->verbs_dev.rdi.driver_f.free_all_qps = free_all_qps;
1715 	dd->verbs_dev.rdi.driver_f.notify_qp_reset = notify_qp_reset;
1716 	dd->verbs_dev.rdi.driver_f.do_send = hfi1_do_send_from_rvt;
1717 	dd->verbs_dev.rdi.driver_f.schedule_send = hfi1_schedule_send;
1718 	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _hfi1_schedule_send;
1719 	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = get_pmtu_from_attr;
1720 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1721 	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = flush_qp_waiters;
1722 	dd->verbs_dev.rdi.driver_f.stop_send_queue = stop_send_queue;
1723 	dd->verbs_dev.rdi.driver_f.quiesce_qp = quiesce_qp;
1724 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1725 	dd->verbs_dev.rdi.driver_f.mtu_from_qp = mtu_from_qp;
1726 	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = mtu_to_path_mtu;
1727 	dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
1728 	dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
1729 	dd->verbs_dev.rdi.driver_f.notify_restart_rc = hfi1_restart_rc;
1730 	dd->verbs_dev.rdi.driver_f.setup_wqe = hfi1_setup_wqe;
1731 	dd->verbs_dev.rdi.driver_f.comp_vect_cpu_lookup =
1732 						hfi1_comp_vect_mappings_lookup;
1733 
1734 	/* completeion queue */
1735 	dd->verbs_dev.rdi.ibdev.num_comp_vectors = dd->comp_vect_possible_cpus;
1736 	dd->verbs_dev.rdi.dparms.node = dd->node;
1737 
1738 	/* misc settings */
1739 	dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
1740 	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
1741 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1742 	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
1743 	dd->verbs_dev.rdi.dparms.sge_copy_mode = sge_copy_mode;
1744 	dd->verbs_dev.rdi.dparms.wss_threshold = wss_threshold;
1745 	dd->verbs_dev.rdi.dparms.wss_clean_period = wss_clean_period;
1746 
1747 	/* post send table */
1748 	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;
1749 
1750 	/* opcode translation table */
1751 	dd->verbs_dev.rdi.wc_opcode = ib_hfi1_wc_opcode;
1752 
1753 	ppd = dd->pport;
1754 	for (i = 0; i < dd->num_pports; i++, ppd++)
1755 		rvt_init_port(&dd->verbs_dev.rdi,
1756 			      &ppd->ibport_data.rvp,
1757 			      i,
1758 			      ppd->pkeys);
1759 
1760 	rdma_set_device_sysfs_group(&dd->verbs_dev.rdi.ibdev,
1761 				    &ib_hfi1_attr_group);
1762 
1763 	ret = rvt_register_device(&dd->verbs_dev.rdi, RDMA_DRIVER_HFI1);
1764 	if (ret)
1765 		goto err_verbs_txreq;
1766 
1767 	ret = hfi1_verbs_register_sysfs(dd);
1768 	if (ret)
1769 		goto err_class;
1770 
1771 	return ret;
1772 
1773 err_class:
1774 	rvt_unregister_device(&dd->verbs_dev.rdi);
1775 err_verbs_txreq:
1776 	verbs_txreq_exit(dev);
1777 	dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1778 	return ret;
1779 }
1780 
1781 void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
1782 {
1783 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1784 
1785 	hfi1_verbs_unregister_sysfs(dd);
1786 
1787 	rvt_unregister_device(&dd->verbs_dev.rdi);
1788 
1789 	if (!list_empty(&dev->txwait))
1790 		dd_dev_err(dd, "txwait list not empty!\n");
1791 	if (!list_empty(&dev->memwait))
1792 		dd_dev_err(dd, "memwait list not empty!\n");
1793 
1794 	del_timer_sync(&dev->mem_timer);
1795 	verbs_txreq_exit(dev);
1796 
1797 	mutex_lock(&cntr_names_lock);
1798 	kfree(dev_cntr_names);
1799 	kfree(port_cntr_names);
1800 	dev_cntr_names = NULL;
1801 	port_cntr_names = NULL;
1802 	cntr_names_initialized = 0;
1803 	mutex_unlock(&cntr_names_lock);
1804 }
1805 
1806 void hfi1_cnp_rcv(struct hfi1_packet *packet)
1807 {
1808 	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1809 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1810 	struct ib_header *hdr = packet->hdr;
1811 	struct rvt_qp *qp = packet->qp;
1812 	u32 lqpn, rqpn = 0;
1813 	u16 rlid = 0;
1814 	u8 sl, sc5, svc_type;
1815 
1816 	switch (packet->qp->ibqp.qp_type) {
1817 	case IB_QPT_UC:
1818 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
1819 		rqpn = qp->remote_qpn;
1820 		svc_type = IB_CC_SVCTYPE_UC;
1821 		break;
1822 	case IB_QPT_RC:
1823 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
1824 		rqpn = qp->remote_qpn;
1825 		svc_type = IB_CC_SVCTYPE_RC;
1826 		break;
1827 	case IB_QPT_SMI:
1828 	case IB_QPT_GSI:
1829 	case IB_QPT_UD:
1830 		svc_type = IB_CC_SVCTYPE_UD;
1831 		break;
1832 	default:
1833 		ibp->rvp.n_pkt_drops++;
1834 		return;
1835 	}
1836 
1837 	sc5 = hfi1_9B_get_sc5(hdr, packet->rhf);
1838 	sl = ibp->sc_to_sl[sc5];
1839 	lqpn = qp->ibqp.qp_num;
1840 
1841 	process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
1842 }
1843