xref: /openbmc/linux/drivers/infiniband/hw/hfi1/verbs.c (revision 6726d552)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2020 Intel Corporation.
4  */
5 
6 #include <rdma/ib_mad.h>
7 #include <rdma/ib_user_verbs.h>
8 #include <linux/io.h>
9 #include <linux/module.h>
10 #include <linux/utsname.h>
11 #include <linux/rculist.h>
12 #include <linux/mm.h>
13 #include <linux/vmalloc.h>
14 #include <rdma/opa_addr.h>
15 #include <linux/nospec.h>
16 
17 #include "hfi.h"
18 #include "common.h"
19 #include "device.h"
20 #include "trace.h"
21 #include "qp.h"
22 #include "verbs_txreq.h"
23 #include "debugfs.h"
24 #include "vnic.h"
25 #include "fault.h"
26 #include "affinity.h"
27 #include "ipoib.h"
28 
29 static unsigned int hfi1_lkey_table_size = 16;
30 module_param_named(lkey_table_size, hfi1_lkey_table_size, uint,
31 		   S_IRUGO);
32 MODULE_PARM_DESC(lkey_table_size,
33 		 "LKEY table size in bits (2^n, 1 <= n <= 23)");
34 
35 static unsigned int hfi1_max_pds = 0xFFFF;
36 module_param_named(max_pds, hfi1_max_pds, uint, S_IRUGO);
37 MODULE_PARM_DESC(max_pds,
38 		 "Maximum number of protection domains to support");
39 
40 static unsigned int hfi1_max_ahs = 0xFFFF;
41 module_param_named(max_ahs, hfi1_max_ahs, uint, S_IRUGO);
42 MODULE_PARM_DESC(max_ahs, "Maximum number of address handles to support");
43 
44 unsigned int hfi1_max_cqes = 0x2FFFFF;
45 module_param_named(max_cqes, hfi1_max_cqes, uint, S_IRUGO);
46 MODULE_PARM_DESC(max_cqes,
47 		 "Maximum number of completion queue entries to support");
48 
49 unsigned int hfi1_max_cqs = 0x1FFFF;
50 module_param_named(max_cqs, hfi1_max_cqs, uint, S_IRUGO);
51 MODULE_PARM_DESC(max_cqs, "Maximum number of completion queues to support");
52 
53 unsigned int hfi1_max_qp_wrs = 0x3FFF;
54 module_param_named(max_qp_wrs, hfi1_max_qp_wrs, uint, S_IRUGO);
55 MODULE_PARM_DESC(max_qp_wrs, "Maximum number of QP WRs to support");
56 
57 unsigned int hfi1_max_qps = 32768;
58 module_param_named(max_qps, hfi1_max_qps, uint, S_IRUGO);
59 MODULE_PARM_DESC(max_qps, "Maximum number of QPs to support");
60 
61 unsigned int hfi1_max_sges = 0x60;
62 module_param_named(max_sges, hfi1_max_sges, uint, S_IRUGO);
63 MODULE_PARM_DESC(max_sges, "Maximum number of SGEs to support");
64 
65 unsigned int hfi1_max_mcast_grps = 16384;
66 module_param_named(max_mcast_grps, hfi1_max_mcast_grps, uint, S_IRUGO);
67 MODULE_PARM_DESC(max_mcast_grps,
68 		 "Maximum number of multicast groups to support");
69 
70 unsigned int hfi1_max_mcast_qp_attached = 16;
71 module_param_named(max_mcast_qp_attached, hfi1_max_mcast_qp_attached,
72 		   uint, S_IRUGO);
73 MODULE_PARM_DESC(max_mcast_qp_attached,
74 		 "Maximum number of attached QPs to support");
75 
76 unsigned int hfi1_max_srqs = 1024;
77 module_param_named(max_srqs, hfi1_max_srqs, uint, S_IRUGO);
78 MODULE_PARM_DESC(max_srqs, "Maximum number of SRQs to support");
79 
80 unsigned int hfi1_max_srq_sges = 128;
81 module_param_named(max_srq_sges, hfi1_max_srq_sges, uint, S_IRUGO);
82 MODULE_PARM_DESC(max_srq_sges, "Maximum number of SRQ SGEs to support");
83 
84 unsigned int hfi1_max_srq_wrs = 0x1FFFF;
85 module_param_named(max_srq_wrs, hfi1_max_srq_wrs, uint, S_IRUGO);
86 MODULE_PARM_DESC(max_srq_wrs, "Maximum number of SRQ WRs support");
87 
88 unsigned short piothreshold = 256;
89 module_param(piothreshold, ushort, S_IRUGO);
90 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
91 
92 static unsigned int sge_copy_mode;
93 module_param(sge_copy_mode, uint, S_IRUGO);
94 MODULE_PARM_DESC(sge_copy_mode,
95 		 "Verbs copy mode: 0 use memcpy, 1 use cacheless copy, 2 adapt based on WSS");
96 
97 static void verbs_sdma_complete(
98 	struct sdma_txreq *cookie,
99 	int status);
100 
101 static int pio_wait(struct rvt_qp *qp,
102 		    struct send_context *sc,
103 		    struct hfi1_pkt_state *ps,
104 		    u32 flag);
105 
106 /* Length of buffer to create verbs txreq cache name */
107 #define TXREQ_NAME_LEN 24
108 
109 static uint wss_threshold = 80;
110 module_param(wss_threshold, uint, S_IRUGO);
111 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
112 static uint wss_clean_period = 256;
113 module_param(wss_clean_period, uint, S_IRUGO);
114 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
115 
116 /*
117  * Translate ib_wr_opcode into ib_wc_opcode.
118  */
119 const enum ib_wc_opcode ib_hfi1_wc_opcode[] = {
120 	[IB_WR_RDMA_WRITE] = IB_WC_RDMA_WRITE,
121 	[IB_WR_TID_RDMA_WRITE] = IB_WC_RDMA_WRITE,
122 	[IB_WR_RDMA_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
123 	[IB_WR_SEND] = IB_WC_SEND,
124 	[IB_WR_SEND_WITH_IMM] = IB_WC_SEND,
125 	[IB_WR_RDMA_READ] = IB_WC_RDMA_READ,
126 	[IB_WR_TID_RDMA_READ] = IB_WC_RDMA_READ,
127 	[IB_WR_ATOMIC_CMP_AND_SWP] = IB_WC_COMP_SWAP,
128 	[IB_WR_ATOMIC_FETCH_AND_ADD] = IB_WC_FETCH_ADD,
129 	[IB_WR_SEND_WITH_INV] = IB_WC_SEND,
130 	[IB_WR_LOCAL_INV] = IB_WC_LOCAL_INV,
131 	[IB_WR_REG_MR] = IB_WC_REG_MR
132 };
133 
134 /*
135  * Length of header by opcode, 0 --> not supported
136  */
137 const u8 hdr_len_by_opcode[256] = {
138 	/* RC */
139 	[IB_OPCODE_RC_SEND_FIRST]                     = 12 + 8,
140 	[IB_OPCODE_RC_SEND_MIDDLE]                    = 12 + 8,
141 	[IB_OPCODE_RC_SEND_LAST]                      = 12 + 8,
142 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
143 	[IB_OPCODE_RC_SEND_ONLY]                      = 12 + 8,
144 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
145 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
146 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = 12 + 8,
147 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = 12 + 8,
148 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
149 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
150 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
151 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = 12 + 8 + 16,
152 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = 12 + 8 + 4,
153 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = 12 + 8,
154 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = 12 + 8 + 4,
155 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = 12 + 8 + 4,
156 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = 12 + 8 + 4,
157 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = 12 + 8 + 4 + 8,
158 	[IB_OPCODE_RC_COMPARE_SWAP]                   = 12 + 8 + 28,
159 	[IB_OPCODE_RC_FETCH_ADD]                      = 12 + 8 + 28,
160 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = 12 + 8 + 4,
161 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = 12 + 8 + 4,
162 	[IB_OPCODE_TID_RDMA_READ_REQ]                 = 12 + 8 + 36,
163 	[IB_OPCODE_TID_RDMA_READ_RESP]                = 12 + 8 + 36,
164 	[IB_OPCODE_TID_RDMA_WRITE_REQ]                = 12 + 8 + 36,
165 	[IB_OPCODE_TID_RDMA_WRITE_RESP]               = 12 + 8 + 36,
166 	[IB_OPCODE_TID_RDMA_WRITE_DATA]               = 12 + 8 + 36,
167 	[IB_OPCODE_TID_RDMA_WRITE_DATA_LAST]          = 12 + 8 + 36,
168 	[IB_OPCODE_TID_RDMA_ACK]                      = 12 + 8 + 36,
169 	[IB_OPCODE_TID_RDMA_RESYNC]                   = 12 + 8 + 36,
170 	/* UC */
171 	[IB_OPCODE_UC_SEND_FIRST]                     = 12 + 8,
172 	[IB_OPCODE_UC_SEND_MIDDLE]                    = 12 + 8,
173 	[IB_OPCODE_UC_SEND_LAST]                      = 12 + 8,
174 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = 12 + 8 + 4,
175 	[IB_OPCODE_UC_SEND_ONLY]                      = 12 + 8,
176 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 4,
177 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = 12 + 8 + 16,
178 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = 12 + 8,
179 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = 12 + 8,
180 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = 12 + 8 + 4,
181 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = 12 + 8 + 16,
182 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = 12 + 8 + 20,
183 	/* UD */
184 	[IB_OPCODE_UD_SEND_ONLY]                      = 12 + 8 + 8,
185 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = 12 + 8 + 12
186 };
187 
188 static const opcode_handler opcode_handler_tbl[256] = {
189 	/* RC */
190 	[IB_OPCODE_RC_SEND_FIRST]                     = &hfi1_rc_rcv,
191 	[IB_OPCODE_RC_SEND_MIDDLE]                    = &hfi1_rc_rcv,
192 	[IB_OPCODE_RC_SEND_LAST]                      = &hfi1_rc_rcv,
193 	[IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
194 	[IB_OPCODE_RC_SEND_ONLY]                      = &hfi1_rc_rcv,
195 	[IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_rc_rcv,
196 	[IB_OPCODE_RC_RDMA_WRITE_FIRST]               = &hfi1_rc_rcv,
197 	[IB_OPCODE_RC_RDMA_WRITE_MIDDLE]              = &hfi1_rc_rcv,
198 	[IB_OPCODE_RC_RDMA_WRITE_LAST]                = &hfi1_rc_rcv,
199 	[IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_rc_rcv,
200 	[IB_OPCODE_RC_RDMA_WRITE_ONLY]                = &hfi1_rc_rcv,
201 	[IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_rc_rcv,
202 	[IB_OPCODE_RC_RDMA_READ_REQUEST]              = &hfi1_rc_rcv,
203 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST]       = &hfi1_rc_rcv,
204 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE]      = &hfi1_rc_rcv,
205 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST]        = &hfi1_rc_rcv,
206 	[IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY]        = &hfi1_rc_rcv,
207 	[IB_OPCODE_RC_ACKNOWLEDGE]                    = &hfi1_rc_rcv,
208 	[IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE]             = &hfi1_rc_rcv,
209 	[IB_OPCODE_RC_COMPARE_SWAP]                   = &hfi1_rc_rcv,
210 	[IB_OPCODE_RC_FETCH_ADD]                      = &hfi1_rc_rcv,
211 	[IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE]      = &hfi1_rc_rcv,
212 	[IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE]      = &hfi1_rc_rcv,
213 
214 	/* TID RDMA has separate handlers for different opcodes.*/
215 	[IB_OPCODE_TID_RDMA_WRITE_REQ]       = &hfi1_rc_rcv_tid_rdma_write_req,
216 	[IB_OPCODE_TID_RDMA_WRITE_RESP]      = &hfi1_rc_rcv_tid_rdma_write_resp,
217 	[IB_OPCODE_TID_RDMA_WRITE_DATA]      = &hfi1_rc_rcv_tid_rdma_write_data,
218 	[IB_OPCODE_TID_RDMA_WRITE_DATA_LAST] = &hfi1_rc_rcv_tid_rdma_write_data,
219 	[IB_OPCODE_TID_RDMA_READ_REQ]        = &hfi1_rc_rcv_tid_rdma_read_req,
220 	[IB_OPCODE_TID_RDMA_READ_RESP]       = &hfi1_rc_rcv_tid_rdma_read_resp,
221 	[IB_OPCODE_TID_RDMA_RESYNC]          = &hfi1_rc_rcv_tid_rdma_resync,
222 	[IB_OPCODE_TID_RDMA_ACK]             = &hfi1_rc_rcv_tid_rdma_ack,
223 
224 	/* UC */
225 	[IB_OPCODE_UC_SEND_FIRST]                     = &hfi1_uc_rcv,
226 	[IB_OPCODE_UC_SEND_MIDDLE]                    = &hfi1_uc_rcv,
227 	[IB_OPCODE_UC_SEND_LAST]                      = &hfi1_uc_rcv,
228 	[IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
229 	[IB_OPCODE_UC_SEND_ONLY]                      = &hfi1_uc_rcv,
230 	[IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_uc_rcv,
231 	[IB_OPCODE_UC_RDMA_WRITE_FIRST]               = &hfi1_uc_rcv,
232 	[IB_OPCODE_UC_RDMA_WRITE_MIDDLE]              = &hfi1_uc_rcv,
233 	[IB_OPCODE_UC_RDMA_WRITE_LAST]                = &hfi1_uc_rcv,
234 	[IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE] = &hfi1_uc_rcv,
235 	[IB_OPCODE_UC_RDMA_WRITE_ONLY]                = &hfi1_uc_rcv,
236 	[IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE] = &hfi1_uc_rcv,
237 	/* UD */
238 	[IB_OPCODE_UD_SEND_ONLY]                      = &hfi1_ud_rcv,
239 	[IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE]       = &hfi1_ud_rcv,
240 	/* CNP */
241 	[IB_OPCODE_CNP]				      = &hfi1_cnp_rcv
242 };
243 
244 #define OPMASK 0x1f
245 
246 static const u32 pio_opmask[BIT(3)] = {
247 	/* RC */
248 	[IB_OPCODE_RC >> 5] =
249 		BIT(RC_OP(SEND_ONLY) & OPMASK) |
250 		BIT(RC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
251 		BIT(RC_OP(RDMA_WRITE_ONLY) & OPMASK) |
252 		BIT(RC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK) |
253 		BIT(RC_OP(RDMA_READ_REQUEST) & OPMASK) |
254 		BIT(RC_OP(ACKNOWLEDGE) & OPMASK) |
255 		BIT(RC_OP(ATOMIC_ACKNOWLEDGE) & OPMASK) |
256 		BIT(RC_OP(COMPARE_SWAP) & OPMASK) |
257 		BIT(RC_OP(FETCH_ADD) & OPMASK),
258 	/* UC */
259 	[IB_OPCODE_UC >> 5] =
260 		BIT(UC_OP(SEND_ONLY) & OPMASK) |
261 		BIT(UC_OP(SEND_ONLY_WITH_IMMEDIATE) & OPMASK) |
262 		BIT(UC_OP(RDMA_WRITE_ONLY) & OPMASK) |
263 		BIT(UC_OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE) & OPMASK),
264 };
265 
266 /*
267  * System image GUID.
268  */
269 __be64 ib_hfi1_sys_image_guid;
270 
271 /*
272  * Make sure the QP is ready and able to accept the given opcode.
273  */
274 static inline opcode_handler qp_ok(struct hfi1_packet *packet)
275 {
276 	if (!(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
277 		return NULL;
278 	if (((packet->opcode & RVT_OPCODE_QP_MASK) ==
279 	     packet->qp->allowed_ops) ||
280 	    (packet->opcode == IB_OPCODE_CNP))
281 		return opcode_handler_tbl[packet->opcode];
282 
283 	return NULL;
284 }
285 
286 static u64 hfi1_fault_tx(struct rvt_qp *qp, u8 opcode, u64 pbc)
287 {
288 #ifdef CONFIG_FAULT_INJECTION
289 	if ((opcode & IB_OPCODE_MSP) == IB_OPCODE_MSP) {
290 		/*
291 		 * In order to drop non-IB traffic we
292 		 * set PbcInsertHrc to NONE (0x2).
293 		 * The packet will still be delivered
294 		 * to the receiving node but a
295 		 * KHdrHCRCErr (KDETH packet with a bad
296 		 * HCRC) will be triggered and the
297 		 * packet will not be delivered to the
298 		 * correct context.
299 		 */
300 		pbc &= ~PBC_INSERT_HCRC_SMASK;
301 		pbc |= (u64)PBC_IHCRC_NONE << PBC_INSERT_HCRC_SHIFT;
302 	} else {
303 		/*
304 		 * In order to drop regular verbs
305 		 * traffic we set the PbcTestEbp
306 		 * flag. The packet will still be
307 		 * delivered to the receiving node but
308 		 * a 'late ebp error' will be
309 		 * triggered and will be dropped.
310 		 */
311 		pbc |= PBC_TEST_EBP;
312 	}
313 #endif
314 	return pbc;
315 }
316 
317 static opcode_handler tid_qp_ok(int opcode, struct hfi1_packet *packet)
318 {
319 	if (packet->qp->ibqp.qp_type != IB_QPT_RC ||
320 	    !(ib_rvt_state_ops[packet->qp->state] & RVT_PROCESS_RECV_OK))
321 		return NULL;
322 	if ((opcode & RVT_OPCODE_QP_MASK) == IB_OPCODE_TID_RDMA)
323 		return opcode_handler_tbl[opcode];
324 	return NULL;
325 }
326 
327 void hfi1_kdeth_eager_rcv(struct hfi1_packet *packet)
328 {
329 	struct hfi1_ctxtdata *rcd = packet->rcd;
330 	struct ib_header *hdr = packet->hdr;
331 	u32 tlen = packet->tlen;
332 	struct hfi1_pportdata *ppd = rcd->ppd;
333 	struct hfi1_ibport *ibp = &ppd->ibport_data;
334 	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
335 	opcode_handler opcode_handler;
336 	unsigned long flags;
337 	u32 qp_num;
338 	int lnh;
339 	u8 opcode;
340 
341 	/* DW == LRH (2) + BTH (3) + KDETH (9) + CRC (1) */
342 	if (unlikely(tlen < 15 * sizeof(u32)))
343 		goto drop;
344 
345 	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
346 	if (lnh != HFI1_LRH_BTH)
347 		goto drop;
348 
349 	packet->ohdr = &hdr->u.oth;
350 	trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));
351 
352 	opcode = (be32_to_cpu(packet->ohdr->bth[0]) >> 24);
353 	inc_opstats(tlen, &rcd->opstats->stats[opcode]);
354 
355 	/* verbs_qp can be picked up from any tid_rdma header struct */
356 	qp_num = be32_to_cpu(packet->ohdr->u.tid_rdma.r_req.verbs_qp) &
357 		RVT_QPN_MASK;
358 
359 	rcu_read_lock();
360 	packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
361 	if (!packet->qp)
362 		goto drop_rcu;
363 	spin_lock_irqsave(&packet->qp->r_lock, flags);
364 	opcode_handler = tid_qp_ok(opcode, packet);
365 	if (likely(opcode_handler))
366 		opcode_handler(packet);
367 	else
368 		goto drop_unlock;
369 	spin_unlock_irqrestore(&packet->qp->r_lock, flags);
370 	rcu_read_unlock();
371 
372 	return;
373 drop_unlock:
374 	spin_unlock_irqrestore(&packet->qp->r_lock, flags);
375 drop_rcu:
376 	rcu_read_unlock();
377 drop:
378 	ibp->rvp.n_pkt_drops++;
379 }
380 
381 void hfi1_kdeth_expected_rcv(struct hfi1_packet *packet)
382 {
383 	struct hfi1_ctxtdata *rcd = packet->rcd;
384 	struct ib_header *hdr = packet->hdr;
385 	u32 tlen = packet->tlen;
386 	struct hfi1_pportdata *ppd = rcd->ppd;
387 	struct hfi1_ibport *ibp = &ppd->ibport_data;
388 	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
389 	opcode_handler opcode_handler;
390 	unsigned long flags;
391 	u32 qp_num;
392 	int lnh;
393 	u8 opcode;
394 
395 	/* DW == LRH (2) + BTH (3) + KDETH (9) + CRC (1) */
396 	if (unlikely(tlen < 15 * sizeof(u32)))
397 		goto drop;
398 
399 	lnh = be16_to_cpu(hdr->lrh[0]) & 3;
400 	if (lnh != HFI1_LRH_BTH)
401 		goto drop;
402 
403 	packet->ohdr = &hdr->u.oth;
404 	trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));
405 
406 	opcode = (be32_to_cpu(packet->ohdr->bth[0]) >> 24);
407 	inc_opstats(tlen, &rcd->opstats->stats[opcode]);
408 
409 	/* verbs_qp can be picked up from any tid_rdma header struct */
410 	qp_num = be32_to_cpu(packet->ohdr->u.tid_rdma.r_rsp.verbs_qp) &
411 		RVT_QPN_MASK;
412 
413 	rcu_read_lock();
414 	packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
415 	if (!packet->qp)
416 		goto drop_rcu;
417 	spin_lock_irqsave(&packet->qp->r_lock, flags);
418 	opcode_handler = tid_qp_ok(opcode, packet);
419 	if (likely(opcode_handler))
420 		opcode_handler(packet);
421 	else
422 		goto drop_unlock;
423 	spin_unlock_irqrestore(&packet->qp->r_lock, flags);
424 	rcu_read_unlock();
425 
426 	return;
427 drop_unlock:
428 	spin_unlock_irqrestore(&packet->qp->r_lock, flags);
429 drop_rcu:
430 	rcu_read_unlock();
431 drop:
432 	ibp->rvp.n_pkt_drops++;
433 }
434 
435 static int hfi1_do_pkey_check(struct hfi1_packet *packet)
436 {
437 	struct hfi1_ctxtdata *rcd = packet->rcd;
438 	struct hfi1_pportdata *ppd = rcd->ppd;
439 	struct hfi1_16b_header *hdr = packet->hdr;
440 	u16 pkey;
441 
442 	/* Pkey check needed only for bypass packets */
443 	if (packet->etype != RHF_RCV_TYPE_BYPASS)
444 		return 0;
445 
446 	/* Perform pkey check */
447 	pkey = hfi1_16B_get_pkey(hdr);
448 	return ingress_pkey_check(ppd, pkey, packet->sc,
449 				  packet->qp->s_pkey_index,
450 				  packet->slid, true);
451 }
452 
453 static inline void hfi1_handle_packet(struct hfi1_packet *packet,
454 				      bool is_mcast)
455 {
456 	u32 qp_num;
457 	struct hfi1_ctxtdata *rcd = packet->rcd;
458 	struct hfi1_pportdata *ppd = rcd->ppd;
459 	struct hfi1_ibport *ibp = rcd_to_iport(rcd);
460 	struct rvt_dev_info *rdi = &ppd->dd->verbs_dev.rdi;
461 	opcode_handler packet_handler;
462 	unsigned long flags;
463 
464 	inc_opstats(packet->tlen, &rcd->opstats->stats[packet->opcode]);
465 
466 	if (unlikely(is_mcast)) {
467 		struct rvt_mcast *mcast;
468 		struct rvt_mcast_qp *p;
469 
470 		if (!packet->grh)
471 			goto drop;
472 		mcast = rvt_mcast_find(&ibp->rvp,
473 				       &packet->grh->dgid,
474 				       opa_get_lid(packet->dlid, 9B));
475 		if (!mcast)
476 			goto drop;
477 		rcu_read_lock();
478 		list_for_each_entry_rcu(p, &mcast->qp_list, list) {
479 			packet->qp = p->qp;
480 			if (hfi1_do_pkey_check(packet))
481 				goto unlock_drop;
482 			spin_lock_irqsave(&packet->qp->r_lock, flags);
483 			packet_handler = qp_ok(packet);
484 			if (likely(packet_handler))
485 				packet_handler(packet);
486 			else
487 				ibp->rvp.n_pkt_drops++;
488 			spin_unlock_irqrestore(&packet->qp->r_lock, flags);
489 		}
490 		rcu_read_unlock();
491 		/*
492 		 * Notify rvt_multicast_detach() if it is waiting for us
493 		 * to finish.
494 		 */
495 		if (atomic_dec_return(&mcast->refcount) <= 1)
496 			wake_up(&mcast->wait);
497 	} else {
498 		/* Get the destination QP number. */
499 		if (packet->etype == RHF_RCV_TYPE_BYPASS &&
500 		    hfi1_16B_get_l4(packet->hdr) == OPA_16B_L4_FM)
501 			qp_num = hfi1_16B_get_dest_qpn(packet->mgmt);
502 		else
503 			qp_num = ib_bth_get_qpn(packet->ohdr);
504 
505 		rcu_read_lock();
506 		packet->qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
507 		if (!packet->qp)
508 			goto unlock_drop;
509 
510 		if (hfi1_do_pkey_check(packet))
511 			goto unlock_drop;
512 
513 		spin_lock_irqsave(&packet->qp->r_lock, flags);
514 		packet_handler = qp_ok(packet);
515 		if (likely(packet_handler))
516 			packet_handler(packet);
517 		else
518 			ibp->rvp.n_pkt_drops++;
519 		spin_unlock_irqrestore(&packet->qp->r_lock, flags);
520 		rcu_read_unlock();
521 	}
522 	return;
523 unlock_drop:
524 	rcu_read_unlock();
525 drop:
526 	ibp->rvp.n_pkt_drops++;
527 }
528 
529 /**
530  * hfi1_ib_rcv - process an incoming packet
531  * @packet: data packet information
532  *
533  * This is called to process an incoming packet at interrupt level.
534  */
535 void hfi1_ib_rcv(struct hfi1_packet *packet)
536 {
537 	struct hfi1_ctxtdata *rcd = packet->rcd;
538 
539 	trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf)));
540 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
541 }
542 
543 void hfi1_16B_rcv(struct hfi1_packet *packet)
544 {
545 	struct hfi1_ctxtdata *rcd = packet->rcd;
546 
547 	trace_input_ibhdr(rcd->dd, packet, false);
548 	hfi1_handle_packet(packet, hfi1_check_mcast(packet->dlid));
549 }
550 
551 /*
552  * This is called from a timer to check for QPs
553  * which need kernel memory in order to send a packet.
554  */
555 static void mem_timer(struct timer_list *t)
556 {
557 	struct hfi1_ibdev *dev = from_timer(dev, t, mem_timer);
558 	struct list_head *list = &dev->memwait;
559 	struct rvt_qp *qp = NULL;
560 	struct iowait *wait;
561 	unsigned long flags;
562 	struct hfi1_qp_priv *priv;
563 
564 	write_seqlock_irqsave(&dev->iowait_lock, flags);
565 	if (!list_empty(list)) {
566 		wait = list_first_entry(list, struct iowait, list);
567 		qp = iowait_to_qp(wait);
568 		priv = qp->priv;
569 		list_del_init(&priv->s_iowait.list);
570 		priv->s_iowait.lock = NULL;
571 		/* refcount held until actual wake up */
572 		if (!list_empty(list))
573 			mod_timer(&dev->mem_timer, jiffies + 1);
574 	}
575 	write_sequnlock_irqrestore(&dev->iowait_lock, flags);
576 
577 	if (qp)
578 		hfi1_qp_wakeup(qp, RVT_S_WAIT_KMEM);
579 }
580 
581 /*
582  * This is called with progress side lock held.
583  */
584 /* New API */
585 static void verbs_sdma_complete(
586 	struct sdma_txreq *cookie,
587 	int status)
588 {
589 	struct verbs_txreq *tx =
590 		container_of(cookie, struct verbs_txreq, txreq);
591 	struct rvt_qp *qp = tx->qp;
592 
593 	spin_lock(&qp->s_lock);
594 	if (tx->wqe) {
595 		rvt_send_complete(qp, tx->wqe, IB_WC_SUCCESS);
596 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
597 		struct hfi1_opa_header *hdr;
598 
599 		hdr = &tx->phdr.hdr;
600 		if (unlikely(status == SDMA_TXREQ_S_ABORTED))
601 			hfi1_rc_verbs_aborted(qp, hdr);
602 		hfi1_rc_send_complete(qp, hdr);
603 	}
604 	spin_unlock(&qp->s_lock);
605 
606 	hfi1_put_txreq(tx);
607 }
608 
609 void hfi1_wait_kmem(struct rvt_qp *qp)
610 {
611 	struct hfi1_qp_priv *priv = qp->priv;
612 	struct ib_qp *ibqp = &qp->ibqp;
613 	struct ib_device *ibdev = ibqp->device;
614 	struct hfi1_ibdev *dev = to_idev(ibdev);
615 
616 	if (list_empty(&priv->s_iowait.list)) {
617 		if (list_empty(&dev->memwait))
618 			mod_timer(&dev->mem_timer, jiffies + 1);
619 		qp->s_flags |= RVT_S_WAIT_KMEM;
620 		list_add_tail(&priv->s_iowait.list, &dev->memwait);
621 		priv->s_iowait.lock = &dev->iowait_lock;
622 		trace_hfi1_qpsleep(qp, RVT_S_WAIT_KMEM);
623 		rvt_get_qp(qp);
624 	}
625 }
626 
627 static int wait_kmem(struct hfi1_ibdev *dev,
628 		     struct rvt_qp *qp,
629 		     struct hfi1_pkt_state *ps)
630 {
631 	unsigned long flags;
632 	int ret = 0;
633 
634 	spin_lock_irqsave(&qp->s_lock, flags);
635 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
636 		write_seqlock(&dev->iowait_lock);
637 		list_add_tail(&ps->s_txreq->txreq.list,
638 			      &ps->wait->tx_head);
639 		hfi1_wait_kmem(qp);
640 		write_sequnlock(&dev->iowait_lock);
641 		hfi1_qp_unbusy(qp, ps->wait);
642 		ret = -EBUSY;
643 	}
644 	spin_unlock_irqrestore(&qp->s_lock, flags);
645 
646 	return ret;
647 }
648 
649 /*
650  * This routine calls txadds for each sg entry.
651  *
652  * Add failures will revert the sge cursor
653  */
654 static noinline int build_verbs_ulp_payload(
655 	struct sdma_engine *sde,
656 	u32 length,
657 	struct verbs_txreq *tx)
658 {
659 	struct rvt_sge_state *ss = tx->ss;
660 	struct rvt_sge *sg_list = ss->sg_list;
661 	struct rvt_sge sge = ss->sge;
662 	u8 num_sge = ss->num_sge;
663 	u32 len;
664 	int ret = 0;
665 
666 	while (length) {
667 		len = rvt_get_sge_length(&ss->sge, length);
668 		WARN_ON_ONCE(len == 0);
669 		ret = sdma_txadd_kvaddr(
670 			sde->dd,
671 			&tx->txreq,
672 			ss->sge.vaddr,
673 			len);
674 		if (ret)
675 			goto bail_txadd;
676 		rvt_update_sge(ss, len, false);
677 		length -= len;
678 	}
679 	return ret;
680 bail_txadd:
681 	/* unwind cursor */
682 	ss->sge = sge;
683 	ss->num_sge = num_sge;
684 	ss->sg_list = sg_list;
685 	return ret;
686 }
687 
688 /**
689  * update_tx_opstats - record stats by opcode
690  * @qp: the qp
691  * @ps: transmit packet state
692  * @plen: the plen in dwords
693  *
694  * This is a routine to record the tx opstats after a
695  * packet has been presented to the egress mechanism.
696  */
697 static void update_tx_opstats(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
698 			      u32 plen)
699 {
700 #ifdef CONFIG_DEBUG_FS
701 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
702 	struct hfi1_opcode_stats_perctx *s = get_cpu_ptr(dd->tx_opstats);
703 
704 	inc_opstats(plen * 4, &s->stats[ps->opcode]);
705 	put_cpu_ptr(s);
706 #endif
707 }
708 
709 /*
710  * Build the number of DMA descriptors needed to send length bytes of data.
711  *
712  * NOTE: DMA mapping is held in the tx until completed in the ring or
713  *       the tx desc is freed without having been submitted to the ring
714  *
715  * This routine ensures all the helper routine calls succeed.
716  */
717 /* New API */
718 static int build_verbs_tx_desc(
719 	struct sdma_engine *sde,
720 	u32 length,
721 	struct verbs_txreq *tx,
722 	struct hfi1_ahg_info *ahg_info,
723 	u64 pbc)
724 {
725 	int ret = 0;
726 	struct hfi1_sdma_header *phdr = &tx->phdr;
727 	u16 hdrbytes = (tx->hdr_dwords + sizeof(pbc) / 4) << 2;
728 	u8 extra_bytes = 0;
729 
730 	if (tx->phdr.hdr.hdr_type) {
731 		/*
732 		 * hdrbytes accounts for PBC. Need to subtract 8 bytes
733 		 * before calculating padding.
734 		 */
735 		extra_bytes = hfi1_get_16b_padding(hdrbytes - 8, length) +
736 			      (SIZE_OF_CRC << 2) + SIZE_OF_LT;
737 	}
738 	if (!ahg_info->ahgcount) {
739 		ret = sdma_txinit_ahg(
740 			&tx->txreq,
741 			ahg_info->tx_flags,
742 			hdrbytes + length +
743 			extra_bytes,
744 			ahg_info->ahgidx,
745 			0,
746 			NULL,
747 			0,
748 			verbs_sdma_complete);
749 		if (ret)
750 			goto bail_txadd;
751 		phdr->pbc = cpu_to_le64(pbc);
752 		ret = sdma_txadd_kvaddr(
753 			sde->dd,
754 			&tx->txreq,
755 			phdr,
756 			hdrbytes);
757 		if (ret)
758 			goto bail_txadd;
759 	} else {
760 		ret = sdma_txinit_ahg(
761 			&tx->txreq,
762 			ahg_info->tx_flags,
763 			length,
764 			ahg_info->ahgidx,
765 			ahg_info->ahgcount,
766 			ahg_info->ahgdesc,
767 			hdrbytes,
768 			verbs_sdma_complete);
769 		if (ret)
770 			goto bail_txadd;
771 	}
772 	/* add the ulp payload - if any. tx->ss can be NULL for acks */
773 	if (tx->ss) {
774 		ret = build_verbs_ulp_payload(sde, length, tx);
775 		if (ret)
776 			goto bail_txadd;
777 	}
778 
779 	/* add icrc, lt byte, and padding to flit */
780 	if (extra_bytes)
781 		ret = sdma_txadd_daddr(sde->dd, &tx->txreq,
782 				       sde->dd->sdma_pad_phys, extra_bytes);
783 
784 bail_txadd:
785 	return ret;
786 }
787 
788 static u64 update_hcrc(u8 opcode, u64 pbc)
789 {
790 	if ((opcode & IB_OPCODE_TID_RDMA) == IB_OPCODE_TID_RDMA) {
791 		pbc &= ~PBC_INSERT_HCRC_SMASK;
792 		pbc |= (u64)PBC_IHCRC_LKDETH << PBC_INSERT_HCRC_SHIFT;
793 	}
794 	return pbc;
795 }
796 
797 int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
798 			u64 pbc)
799 {
800 	struct hfi1_qp_priv *priv = qp->priv;
801 	struct hfi1_ahg_info *ahg_info = priv->s_ahg;
802 	u32 hdrwords = ps->s_txreq->hdr_dwords;
803 	u32 len = ps->s_txreq->s_cur_size;
804 	u32 plen;
805 	struct hfi1_ibdev *dev = ps->dev;
806 	struct hfi1_pportdata *ppd = ps->ppd;
807 	struct verbs_txreq *tx;
808 	u8 sc5 = priv->s_sc;
809 	int ret;
810 	u32 dwords;
811 
812 	if (ps->s_txreq->phdr.hdr.hdr_type) {
813 		u8 extra_bytes = hfi1_get_16b_padding((hdrwords << 2), len);
814 
815 		dwords = (len + extra_bytes + (SIZE_OF_CRC << 2) +
816 			  SIZE_OF_LT) >> 2;
817 	} else {
818 		dwords = (len + 3) >> 2;
819 	}
820 	plen = hdrwords + dwords + sizeof(pbc) / 4;
821 
822 	tx = ps->s_txreq;
823 	if (!sdma_txreq_built(&tx->txreq)) {
824 		if (likely(pbc == 0)) {
825 			u32 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
826 
827 			/* No vl15 here */
828 			/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
829 			if (ps->s_txreq->phdr.hdr.hdr_type)
830 				pbc |= PBC_PACKET_BYPASS |
831 				       PBC_INSERT_BYPASS_ICRC;
832 			else
833 				pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
834 
835 			pbc = create_pbc(ppd,
836 					 pbc,
837 					 qp->srate_mbps,
838 					 vl,
839 					 plen);
840 
841 			if (unlikely(hfi1_dbg_should_fault_tx(qp, ps->opcode)))
842 				pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
843 			else
844 				/* Update HCRC based on packet opcode */
845 				pbc = update_hcrc(ps->opcode, pbc);
846 		}
847 		tx->wqe = qp->s_wqe;
848 		ret = build_verbs_tx_desc(tx->sde, len, tx, ahg_info, pbc);
849 		if (unlikely(ret))
850 			goto bail_build;
851 	}
852 	ret =  sdma_send_txreq(tx->sde, ps->wait, &tx->txreq, ps->pkts_sent);
853 	if (unlikely(ret < 0)) {
854 		if (ret == -ECOMM)
855 			goto bail_ecomm;
856 		return ret;
857 	}
858 
859 	update_tx_opstats(qp, ps, plen);
860 	trace_sdma_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
861 				&ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
862 	return ret;
863 
864 bail_ecomm:
865 	/* The current one got "sent" */
866 	return 0;
867 bail_build:
868 	ret = wait_kmem(dev, qp, ps);
869 	if (!ret) {
870 		/* free txreq - bad state */
871 		hfi1_put_txreq(ps->s_txreq);
872 		ps->s_txreq = NULL;
873 	}
874 	return ret;
875 }
876 
877 /*
878  * If we are now in the error state, return zero to flush the
879  * send work request.
880  */
881 static int pio_wait(struct rvt_qp *qp,
882 		    struct send_context *sc,
883 		    struct hfi1_pkt_state *ps,
884 		    u32 flag)
885 {
886 	struct hfi1_qp_priv *priv = qp->priv;
887 	struct hfi1_devdata *dd = sc->dd;
888 	unsigned long flags;
889 	int ret = 0;
890 
891 	/*
892 	 * Note that as soon as want_buffer() is called and
893 	 * possibly before it returns, sc_piobufavail()
894 	 * could be called. Therefore, put QP on the I/O wait list before
895 	 * enabling the PIO avail interrupt.
896 	 */
897 	spin_lock_irqsave(&qp->s_lock, flags);
898 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
899 		write_seqlock(&sc->waitlock);
900 		list_add_tail(&ps->s_txreq->txreq.list,
901 			      &ps->wait->tx_head);
902 		if (list_empty(&priv->s_iowait.list)) {
903 			struct hfi1_ibdev *dev = &dd->verbs_dev;
904 			int was_empty;
905 
906 			dev->n_piowait += !!(flag & RVT_S_WAIT_PIO);
907 			dev->n_piodrain += !!(flag & HFI1_S_WAIT_PIO_DRAIN);
908 			qp->s_flags |= flag;
909 			was_empty = list_empty(&sc->piowait);
910 			iowait_get_priority(&priv->s_iowait);
911 			iowait_queue(ps->pkts_sent, &priv->s_iowait,
912 				     &sc->piowait);
913 			priv->s_iowait.lock = &sc->waitlock;
914 			trace_hfi1_qpsleep(qp, RVT_S_WAIT_PIO);
915 			rvt_get_qp(qp);
916 			/* counting: only call wantpiobuf_intr if first user */
917 			if (was_empty)
918 				hfi1_sc_wantpiobuf_intr(sc, 1);
919 		}
920 		write_sequnlock(&sc->waitlock);
921 		hfi1_qp_unbusy(qp, ps->wait);
922 		ret = -EBUSY;
923 	}
924 	spin_unlock_irqrestore(&qp->s_lock, flags);
925 	return ret;
926 }
927 
928 static void verbs_pio_complete(void *arg, int code)
929 {
930 	struct rvt_qp *qp = (struct rvt_qp *)arg;
931 	struct hfi1_qp_priv *priv = qp->priv;
932 
933 	if (iowait_pio_dec(&priv->s_iowait))
934 		iowait_drain_wakeup(&priv->s_iowait);
935 }
936 
937 int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
938 			u64 pbc)
939 {
940 	struct hfi1_qp_priv *priv = qp->priv;
941 	u32 hdrwords = ps->s_txreq->hdr_dwords;
942 	struct rvt_sge_state *ss = ps->s_txreq->ss;
943 	u32 len = ps->s_txreq->s_cur_size;
944 	u32 dwords;
945 	u32 plen;
946 	struct hfi1_pportdata *ppd = ps->ppd;
947 	u32 *hdr;
948 	u8 sc5;
949 	unsigned long flags = 0;
950 	struct send_context *sc;
951 	struct pio_buf *pbuf;
952 	int wc_status = IB_WC_SUCCESS;
953 	int ret = 0;
954 	pio_release_cb cb = NULL;
955 	u8 extra_bytes = 0;
956 
957 	if (ps->s_txreq->phdr.hdr.hdr_type) {
958 		u8 pad_size = hfi1_get_16b_padding((hdrwords << 2), len);
959 
960 		extra_bytes = pad_size + (SIZE_OF_CRC << 2) + SIZE_OF_LT;
961 		dwords = (len + extra_bytes) >> 2;
962 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.opah;
963 	} else {
964 		dwords = (len + 3) >> 2;
965 		hdr = (u32 *)&ps->s_txreq->phdr.hdr.ibh;
966 	}
967 	plen = hdrwords + dwords + sizeof(pbc) / 4;
968 
969 	/* only RC/UC use complete */
970 	switch (qp->ibqp.qp_type) {
971 	case IB_QPT_RC:
972 	case IB_QPT_UC:
973 		cb = verbs_pio_complete;
974 		break;
975 	default:
976 		break;
977 	}
978 
979 	/* vl15 special case taken care of in ud.c */
980 	sc5 = priv->s_sc;
981 	sc = ps->s_txreq->psc;
982 
983 	if (likely(pbc == 0)) {
984 		u8 vl = sc_to_vlt(dd_from_ibdev(qp->ibqp.device), sc5);
985 
986 		/* set PBC_DC_INFO bit (aka SC[4]) in pbc */
987 		if (ps->s_txreq->phdr.hdr.hdr_type)
988 			pbc |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
989 		else
990 			pbc |= (ib_is_sc5(sc5) << PBC_DC_INFO_SHIFT);
991 
992 		pbc = create_pbc(ppd, pbc, qp->srate_mbps, vl, plen);
993 		if (unlikely(hfi1_dbg_should_fault_tx(qp, ps->opcode)))
994 			pbc = hfi1_fault_tx(qp, ps->opcode, pbc);
995 		else
996 			/* Update HCRC based on packet opcode */
997 			pbc = update_hcrc(ps->opcode, pbc);
998 	}
999 	if (cb)
1000 		iowait_pio_inc(&priv->s_iowait);
1001 	pbuf = sc_buffer_alloc(sc, plen, cb, qp);
1002 	if (IS_ERR_OR_NULL(pbuf)) {
1003 		if (cb)
1004 			verbs_pio_complete(qp, 0);
1005 		if (IS_ERR(pbuf)) {
1006 			/*
1007 			 * If we have filled the PIO buffers to capacity and are
1008 			 * not in an active state this request is not going to
1009 			 * go out to so just complete it with an error or else a
1010 			 * ULP or the core may be stuck waiting.
1011 			 */
1012 			hfi1_cdbg(
1013 				PIO,
1014 				"alloc failed. state not active, completing");
1015 			wc_status = IB_WC_GENERAL_ERR;
1016 			goto pio_bail;
1017 		} else {
1018 			/*
1019 			 * This is a normal occurrence. The PIO buffs are full
1020 			 * up but we are still happily sending, well we could be
1021 			 * so lets continue to queue the request.
1022 			 */
1023 			hfi1_cdbg(PIO, "alloc failed. state active, queuing");
1024 			ret = pio_wait(qp, sc, ps, RVT_S_WAIT_PIO);
1025 			if (!ret)
1026 				/* txreq not queued - free */
1027 				goto bail;
1028 			/* tx consumed in wait */
1029 			return ret;
1030 		}
1031 	}
1032 
1033 	if (dwords == 0) {
1034 		pio_copy(ppd->dd, pbuf, pbc, hdr, hdrwords);
1035 	} else {
1036 		seg_pio_copy_start(pbuf, pbc,
1037 				   hdr, hdrwords * 4);
1038 		if (ss) {
1039 			while (len) {
1040 				void *addr = ss->sge.vaddr;
1041 				u32 slen = rvt_get_sge_length(&ss->sge, len);
1042 
1043 				rvt_update_sge(ss, slen, false);
1044 				seg_pio_copy_mid(pbuf, addr, slen);
1045 				len -= slen;
1046 			}
1047 		}
1048 		/* add icrc, lt byte, and padding to flit */
1049 		if (extra_bytes)
1050 			seg_pio_copy_mid(pbuf, ppd->dd->sdma_pad_dma,
1051 					 extra_bytes);
1052 
1053 		seg_pio_copy_end(pbuf);
1054 	}
1055 
1056 	update_tx_opstats(qp, ps, plen);
1057 	trace_pio_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
1058 			       &ps->s_txreq->phdr.hdr, ib_is_sc5(sc5));
1059 
1060 pio_bail:
1061 	spin_lock_irqsave(&qp->s_lock, flags);
1062 	if (qp->s_wqe) {
1063 		rvt_send_complete(qp, qp->s_wqe, wc_status);
1064 	} else if (qp->ibqp.qp_type == IB_QPT_RC) {
1065 		if (unlikely(wc_status == IB_WC_GENERAL_ERR))
1066 			hfi1_rc_verbs_aborted(qp, &ps->s_txreq->phdr.hdr);
1067 		hfi1_rc_send_complete(qp, &ps->s_txreq->phdr.hdr);
1068 	}
1069 	spin_unlock_irqrestore(&qp->s_lock, flags);
1070 
1071 	ret = 0;
1072 
1073 bail:
1074 	hfi1_put_txreq(ps->s_txreq);
1075 	return ret;
1076 }
1077 
1078 /*
1079  * egress_pkey_matches_entry - return 1 if the pkey matches ent (ent
1080  * being an entry from the partition key table), return 0
1081  * otherwise. Use the matching criteria for egress partition keys
1082  * specified in the OPAv1 spec., section 9.1l.7.
1083  */
1084 static inline int egress_pkey_matches_entry(u16 pkey, u16 ent)
1085 {
1086 	u16 mkey = pkey & PKEY_LOW_15_MASK;
1087 	u16 mentry = ent & PKEY_LOW_15_MASK;
1088 
1089 	if (mkey == mentry) {
1090 		/*
1091 		 * If pkey[15] is set (full partition member),
1092 		 * is bit 15 in the corresponding table element
1093 		 * clear (limited member)?
1094 		 */
1095 		if (pkey & PKEY_MEMBER_MASK)
1096 			return !!(ent & PKEY_MEMBER_MASK);
1097 		return 1;
1098 	}
1099 	return 0;
1100 }
1101 
1102 /**
1103  * egress_pkey_check - check P_KEY of a packet
1104  * @ppd:  Physical IB port data
1105  * @slid: SLID for packet
1106  * @pkey: PKEY for header
1107  * @sc5:  SC for packet
1108  * @s_pkey_index: It will be used for look up optimization for kernel contexts
1109  * only. If it is negative value, then it means user contexts is calling this
1110  * function.
1111  *
1112  * It checks if hdr's pkey is valid.
1113  *
1114  * Return: 0 on success, otherwise, 1
1115  */
1116 int egress_pkey_check(struct hfi1_pportdata *ppd, u32 slid, u16 pkey,
1117 		      u8 sc5, int8_t s_pkey_index)
1118 {
1119 	struct hfi1_devdata *dd;
1120 	int i;
1121 	int is_user_ctxt_mechanism = (s_pkey_index < 0);
1122 
1123 	if (!(ppd->part_enforce & HFI1_PART_ENFORCE_OUT))
1124 		return 0;
1125 
1126 	/* If SC15, pkey[0:14] must be 0x7fff */
1127 	if ((sc5 == 0xf) && ((pkey & PKEY_LOW_15_MASK) != PKEY_LOW_15_MASK))
1128 		goto bad;
1129 
1130 	/* Is the pkey = 0x0, or 0x8000? */
1131 	if ((pkey & PKEY_LOW_15_MASK) == 0)
1132 		goto bad;
1133 
1134 	/*
1135 	 * For the kernel contexts only, if a qp is passed into the function,
1136 	 * the most likely matching pkey has index qp->s_pkey_index
1137 	 */
1138 	if (!is_user_ctxt_mechanism &&
1139 	    egress_pkey_matches_entry(pkey, ppd->pkeys[s_pkey_index])) {
1140 		return 0;
1141 	}
1142 
1143 	for (i = 0; i < MAX_PKEY_VALUES; i++) {
1144 		if (egress_pkey_matches_entry(pkey, ppd->pkeys[i]))
1145 			return 0;
1146 	}
1147 bad:
1148 	/*
1149 	 * For the user-context mechanism, the P_KEY check would only happen
1150 	 * once per SDMA request, not once per packet.  Therefore, there's no
1151 	 * need to increment the counter for the user-context mechanism.
1152 	 */
1153 	if (!is_user_ctxt_mechanism) {
1154 		incr_cntr64(&ppd->port_xmit_constraint_errors);
1155 		dd = ppd->dd;
1156 		if (!(dd->err_info_xmit_constraint.status &
1157 		      OPA_EI_STATUS_SMASK)) {
1158 			dd->err_info_xmit_constraint.status |=
1159 				OPA_EI_STATUS_SMASK;
1160 			dd->err_info_xmit_constraint.slid = slid;
1161 			dd->err_info_xmit_constraint.pkey = pkey;
1162 		}
1163 	}
1164 	return 1;
1165 }
1166 
1167 /*
1168  * get_send_routine - choose an egress routine
1169  *
1170  * Choose an egress routine based on QP type
1171  * and size
1172  */
1173 static inline send_routine get_send_routine(struct rvt_qp *qp,
1174 					    struct hfi1_pkt_state *ps)
1175 {
1176 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1177 	struct hfi1_qp_priv *priv = qp->priv;
1178 	struct verbs_txreq *tx = ps->s_txreq;
1179 
1180 	if (unlikely(!(dd->flags & HFI1_HAS_SEND_DMA)))
1181 		return dd->process_pio_send;
1182 	switch (qp->ibqp.qp_type) {
1183 	case IB_QPT_SMI:
1184 		return dd->process_pio_send;
1185 	case IB_QPT_GSI:
1186 	case IB_QPT_UD:
1187 		break;
1188 	case IB_QPT_UC:
1189 	case IB_QPT_RC:
1190 		priv->s_running_pkt_size =
1191 			(tx->s_cur_size + priv->s_running_pkt_size) / 2;
1192 		if (piothreshold &&
1193 		    priv->s_running_pkt_size <= min(piothreshold, qp->pmtu) &&
1194 		    (BIT(ps->opcode & OPMASK) & pio_opmask[ps->opcode >> 5]) &&
1195 		    iowait_sdma_pending(&priv->s_iowait) == 0 &&
1196 		    !sdma_txreq_built(&tx->txreq))
1197 			return dd->process_pio_send;
1198 		break;
1199 	default:
1200 		break;
1201 	}
1202 	return dd->process_dma_send;
1203 }
1204 
1205 /**
1206  * hfi1_verbs_send - send a packet
1207  * @qp: the QP to send on
1208  * @ps: the state of the packet to send
1209  *
1210  * Return zero if packet is sent or queued OK.
1211  * Return non-zero and clear qp->s_flags RVT_S_BUSY otherwise.
1212  */
1213 int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
1214 {
1215 	struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
1216 	struct hfi1_qp_priv *priv = qp->priv;
1217 	struct ib_other_headers *ohdr = NULL;
1218 	send_routine sr;
1219 	int ret;
1220 	u16 pkey;
1221 	u32 slid;
1222 	u8 l4 = 0;
1223 
1224 	/* locate the pkey within the headers */
1225 	if (ps->s_txreq->phdr.hdr.hdr_type) {
1226 		struct hfi1_16b_header *hdr = &ps->s_txreq->phdr.hdr.opah;
1227 
1228 		l4 = hfi1_16B_get_l4(hdr);
1229 		if (l4 == OPA_16B_L4_IB_LOCAL)
1230 			ohdr = &hdr->u.oth;
1231 		else if (l4 == OPA_16B_L4_IB_GLOBAL)
1232 			ohdr = &hdr->u.l.oth;
1233 
1234 		slid = hfi1_16B_get_slid(hdr);
1235 		pkey = hfi1_16B_get_pkey(hdr);
1236 	} else {
1237 		struct ib_header *hdr = &ps->s_txreq->phdr.hdr.ibh;
1238 		u8 lnh = ib_get_lnh(hdr);
1239 
1240 		if (lnh == HFI1_LRH_GRH)
1241 			ohdr = &hdr->u.l.oth;
1242 		else
1243 			ohdr = &hdr->u.oth;
1244 		slid = ib_get_slid(hdr);
1245 		pkey = ib_bth_get_pkey(ohdr);
1246 	}
1247 
1248 	if (likely(l4 != OPA_16B_L4_FM))
1249 		ps->opcode = ib_bth_get_opcode(ohdr);
1250 	else
1251 		ps->opcode = IB_OPCODE_UD_SEND_ONLY;
1252 
1253 	sr = get_send_routine(qp, ps);
1254 	ret = egress_pkey_check(dd->pport, slid, pkey,
1255 				priv->s_sc, qp->s_pkey_index);
1256 	if (unlikely(ret)) {
1257 		/*
1258 		 * The value we are returning here does not get propagated to
1259 		 * the verbs caller. Thus we need to complete the request with
1260 		 * error otherwise the caller could be sitting waiting on the
1261 		 * completion event. Only do this for PIO. SDMA has its own
1262 		 * mechanism for handling the errors. So for SDMA we can just
1263 		 * return.
1264 		 */
1265 		if (sr == dd->process_pio_send) {
1266 			unsigned long flags;
1267 
1268 			hfi1_cdbg(PIO, "%s() Failed. Completing with err",
1269 				  __func__);
1270 			spin_lock_irqsave(&qp->s_lock, flags);
1271 			rvt_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
1272 			spin_unlock_irqrestore(&qp->s_lock, flags);
1273 		}
1274 		return -EINVAL;
1275 	}
1276 	if (sr == dd->process_dma_send && iowait_pio_pending(&priv->s_iowait))
1277 		return pio_wait(qp,
1278 				ps->s_txreq->psc,
1279 				ps,
1280 				HFI1_S_WAIT_PIO_DRAIN);
1281 	return sr(qp, ps, 0);
1282 }
1283 
1284 /**
1285  * hfi1_fill_device_attr - Fill in rvt dev info device attributes.
1286  * @dd: the device data structure
1287  */
1288 static void hfi1_fill_device_attr(struct hfi1_devdata *dd)
1289 {
1290 	struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
1291 	u32 ver = dd->dc8051_ver;
1292 
1293 	memset(&rdi->dparms.props, 0, sizeof(rdi->dparms.props));
1294 
1295 	rdi->dparms.props.fw_ver = ((u64)(dc8051_ver_maj(ver)) << 32) |
1296 		((u64)(dc8051_ver_min(ver)) << 16) |
1297 		(u64)dc8051_ver_patch(ver);
1298 
1299 	rdi->dparms.props.device_cap_flags = IB_DEVICE_BAD_PKEY_CNTR |
1300 			IB_DEVICE_BAD_QKEY_CNTR | IB_DEVICE_SHUTDOWN_PORT |
1301 			IB_DEVICE_SYS_IMAGE_GUID | IB_DEVICE_RC_RNR_NAK_GEN |
1302 			IB_DEVICE_PORT_ACTIVE_EVENT | IB_DEVICE_SRQ_RESIZE |
1303 			IB_DEVICE_MEM_MGT_EXTENSIONS;
1304 	rdi->dparms.props.kernel_cap_flags = IBK_RDMA_NETDEV_OPA;
1305 	rdi->dparms.props.page_size_cap = PAGE_SIZE;
1306 	rdi->dparms.props.vendor_id = dd->oui1 << 16 | dd->oui2 << 8 | dd->oui3;
1307 	rdi->dparms.props.vendor_part_id = dd->pcidev->device;
1308 	rdi->dparms.props.hw_ver = dd->minrev;
1309 	rdi->dparms.props.sys_image_guid = ib_hfi1_sys_image_guid;
1310 	rdi->dparms.props.max_mr_size = U64_MAX;
1311 	rdi->dparms.props.max_fast_reg_page_list_len = UINT_MAX;
1312 	rdi->dparms.props.max_qp = hfi1_max_qps;
1313 	rdi->dparms.props.max_qp_wr =
1314 		(hfi1_max_qp_wrs >= HFI1_QP_WQE_INVALID ?
1315 		 HFI1_QP_WQE_INVALID - 1 : hfi1_max_qp_wrs);
1316 	rdi->dparms.props.max_send_sge = hfi1_max_sges;
1317 	rdi->dparms.props.max_recv_sge = hfi1_max_sges;
1318 	rdi->dparms.props.max_sge_rd = hfi1_max_sges;
1319 	rdi->dparms.props.max_cq = hfi1_max_cqs;
1320 	rdi->dparms.props.max_ah = hfi1_max_ahs;
1321 	rdi->dparms.props.max_cqe = hfi1_max_cqes;
1322 	rdi->dparms.props.max_pd = hfi1_max_pds;
1323 	rdi->dparms.props.max_qp_rd_atom = HFI1_MAX_RDMA_ATOMIC;
1324 	rdi->dparms.props.max_qp_init_rd_atom = 255;
1325 	rdi->dparms.props.max_srq = hfi1_max_srqs;
1326 	rdi->dparms.props.max_srq_wr = hfi1_max_srq_wrs;
1327 	rdi->dparms.props.max_srq_sge = hfi1_max_srq_sges;
1328 	rdi->dparms.props.atomic_cap = IB_ATOMIC_GLOB;
1329 	rdi->dparms.props.max_pkeys = hfi1_get_npkeys(dd);
1330 	rdi->dparms.props.max_mcast_grp = hfi1_max_mcast_grps;
1331 	rdi->dparms.props.max_mcast_qp_attach = hfi1_max_mcast_qp_attached;
1332 	rdi->dparms.props.max_total_mcast_qp_attach =
1333 					rdi->dparms.props.max_mcast_qp_attach *
1334 					rdi->dparms.props.max_mcast_grp;
1335 }
1336 
1337 static inline u16 opa_speed_to_ib(u16 in)
1338 {
1339 	u16 out = 0;
1340 
1341 	if (in & OPA_LINK_SPEED_25G)
1342 		out |= IB_SPEED_EDR;
1343 	if (in & OPA_LINK_SPEED_12_5G)
1344 		out |= IB_SPEED_FDR;
1345 
1346 	return out;
1347 }
1348 
1349 /*
1350  * Convert a single OPA link width (no multiple flags) to an IB value.
1351  * A zero OPA link width means link down, which means the IB width value
1352  * is a don't care.
1353  */
1354 static inline u16 opa_width_to_ib(u16 in)
1355 {
1356 	switch (in) {
1357 	case OPA_LINK_WIDTH_1X:
1358 	/* map 2x and 3x to 1x as they don't exist in IB */
1359 	case OPA_LINK_WIDTH_2X:
1360 	case OPA_LINK_WIDTH_3X:
1361 		return IB_WIDTH_1X;
1362 	default: /* link down or unknown, return our largest width */
1363 	case OPA_LINK_WIDTH_4X:
1364 		return IB_WIDTH_4X;
1365 	}
1366 }
1367 
1368 static int query_port(struct rvt_dev_info *rdi, u32 port_num,
1369 		      struct ib_port_attr *props)
1370 {
1371 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1372 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1373 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1374 	u32 lid = ppd->lid;
1375 
1376 	/* props being zeroed by the caller, avoid zeroing it here */
1377 	props->lid = lid ? lid : 0;
1378 	props->lmc = ppd->lmc;
1379 	/* OPA logical states match IB logical states */
1380 	props->state = driver_lstate(ppd);
1381 	props->phys_state = driver_pstate(ppd);
1382 	props->gid_tbl_len = HFI1_GUIDS_PER_PORT;
1383 	props->active_width = (u8)opa_width_to_ib(ppd->link_width_active);
1384 	/* see rate_show() in ib core/sysfs.c */
1385 	props->active_speed = opa_speed_to_ib(ppd->link_speed_active);
1386 	props->max_vl_num = ppd->vls_supported;
1387 
1388 	/* Once we are a "first class" citizen and have added the OPA MTUs to
1389 	 * the core we can advertise the larger MTU enum to the ULPs, for now
1390 	 * advertise only 4K.
1391 	 *
1392 	 * Those applications which are either OPA aware or pass the MTU enum
1393 	 * from the Path Records to us will get the new 8k MTU.  Those that
1394 	 * attempt to process the MTU enum may fail in various ways.
1395 	 */
1396 	props->max_mtu = mtu_to_enum((!valid_ib_mtu(hfi1_max_mtu) ?
1397 				      4096 : hfi1_max_mtu), IB_MTU_4096);
1398 	props->active_mtu = !valid_ib_mtu(ppd->ibmtu) ? props->max_mtu :
1399 		mtu_to_enum(ppd->ibmtu, IB_MTU_4096);
1400 	props->phys_mtu = hfi1_max_mtu;
1401 
1402 	return 0;
1403 }
1404 
1405 static int modify_device(struct ib_device *device,
1406 			 int device_modify_mask,
1407 			 struct ib_device_modify *device_modify)
1408 {
1409 	struct hfi1_devdata *dd = dd_from_ibdev(device);
1410 	unsigned i;
1411 	int ret;
1412 
1413 	if (device_modify_mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
1414 				   IB_DEVICE_MODIFY_NODE_DESC)) {
1415 		ret = -EOPNOTSUPP;
1416 		goto bail;
1417 	}
1418 
1419 	if (device_modify_mask & IB_DEVICE_MODIFY_NODE_DESC) {
1420 		memcpy(device->node_desc, device_modify->node_desc,
1421 		       IB_DEVICE_NODE_DESC_MAX);
1422 		for (i = 0; i < dd->num_pports; i++) {
1423 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1424 
1425 			hfi1_node_desc_chg(ibp);
1426 		}
1427 	}
1428 
1429 	if (device_modify_mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
1430 		ib_hfi1_sys_image_guid =
1431 			cpu_to_be64(device_modify->sys_image_guid);
1432 		for (i = 0; i < dd->num_pports; i++) {
1433 			struct hfi1_ibport *ibp = &dd->pport[i].ibport_data;
1434 
1435 			hfi1_sys_guid_chg(ibp);
1436 		}
1437 	}
1438 
1439 	ret = 0;
1440 
1441 bail:
1442 	return ret;
1443 }
1444 
1445 static int shut_down_port(struct rvt_dev_info *rdi, u32 port_num)
1446 {
1447 	struct hfi1_ibdev *verbs_dev = dev_from_rdi(rdi);
1448 	struct hfi1_devdata *dd = dd_from_dev(verbs_dev);
1449 	struct hfi1_pportdata *ppd = &dd->pport[port_num - 1];
1450 	int ret;
1451 
1452 	set_link_down_reason(ppd, OPA_LINKDOWN_REASON_UNKNOWN, 0,
1453 			     OPA_LINKDOWN_REASON_UNKNOWN);
1454 	ret = set_link_state(ppd, HLS_DN_DOWNDEF);
1455 	return ret;
1456 }
1457 
1458 static int hfi1_get_guid_be(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
1459 			    int guid_index, __be64 *guid)
1460 {
1461 	struct hfi1_ibport *ibp = container_of(rvp, struct hfi1_ibport, rvp);
1462 
1463 	if (guid_index >= HFI1_GUIDS_PER_PORT)
1464 		return -EINVAL;
1465 
1466 	*guid = get_sguid(ibp, guid_index);
1467 	return 0;
1468 }
1469 
1470 /*
1471  * convert ah port,sl to sc
1472  */
1473 u8 ah_to_sc(struct ib_device *ibdev, struct rdma_ah_attr *ah)
1474 {
1475 	struct hfi1_ibport *ibp = to_iport(ibdev, rdma_ah_get_port_num(ah));
1476 
1477 	return ibp->sl_to_sc[rdma_ah_get_sl(ah)];
1478 }
1479 
1480 static int hfi1_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr)
1481 {
1482 	struct hfi1_ibport *ibp;
1483 	struct hfi1_pportdata *ppd;
1484 	struct hfi1_devdata *dd;
1485 	u8 sc5;
1486 	u8 sl;
1487 
1488 	if (hfi1_check_mcast(rdma_ah_get_dlid(ah_attr)) &&
1489 	    !(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
1490 		return -EINVAL;
1491 
1492 	/* test the mapping for validity */
1493 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1494 	ppd = ppd_from_ibp(ibp);
1495 	dd = dd_from_ppd(ppd);
1496 
1497 	sl = rdma_ah_get_sl(ah_attr);
1498 	if (sl >= ARRAY_SIZE(ibp->sl_to_sc))
1499 		return -EINVAL;
1500 	sl = array_index_nospec(sl, ARRAY_SIZE(ibp->sl_to_sc));
1501 
1502 	sc5 = ibp->sl_to_sc[sl];
1503 	if (sc_to_vlt(dd, sc5) > num_vls && sc_to_vlt(dd, sc5) != 0xf)
1504 		return -EINVAL;
1505 	return 0;
1506 }
1507 
1508 static void hfi1_notify_new_ah(struct ib_device *ibdev,
1509 			       struct rdma_ah_attr *ah_attr,
1510 			       struct rvt_ah *ah)
1511 {
1512 	struct hfi1_ibport *ibp;
1513 	struct hfi1_pportdata *ppd;
1514 	struct hfi1_devdata *dd;
1515 	u8 sc5;
1516 	struct rdma_ah_attr *attr = &ah->attr;
1517 
1518 	/*
1519 	 * Do not trust reading anything from rvt_ah at this point as it is not
1520 	 * done being setup. We can however modify things which we need to set.
1521 	 */
1522 
1523 	ibp = to_iport(ibdev, rdma_ah_get_port_num(ah_attr));
1524 	ppd = ppd_from_ibp(ibp);
1525 	sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)];
1526 	hfi1_update_ah_attr(ibdev, attr);
1527 	hfi1_make_opa_lid(attr);
1528 	dd = dd_from_ppd(ppd);
1529 	ah->vl = sc_to_vlt(dd, sc5);
1530 	if (ah->vl < num_vls || ah->vl == 15)
1531 		ah->log_pmtu = ilog2(dd->vld[ah->vl].mtu);
1532 }
1533 
1534 /**
1535  * hfi1_get_npkeys - return the size of the PKEY table for context 0
1536  * @dd: the hfi1_ib device
1537  */
1538 unsigned hfi1_get_npkeys(struct hfi1_devdata *dd)
1539 {
1540 	return ARRAY_SIZE(dd->pport[0].pkeys);
1541 }
1542 
1543 static void init_ibport(struct hfi1_pportdata *ppd)
1544 {
1545 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1546 	size_t sz = ARRAY_SIZE(ibp->sl_to_sc);
1547 	int i;
1548 
1549 	for (i = 0; i < sz; i++) {
1550 		ibp->sl_to_sc[i] = i;
1551 		ibp->sc_to_sl[i] = i;
1552 	}
1553 
1554 	for (i = 0; i < RVT_MAX_TRAP_LISTS ; i++)
1555 		INIT_LIST_HEAD(&ibp->rvp.trap_lists[i].list);
1556 	timer_setup(&ibp->rvp.trap_timer, hfi1_handle_trap_timer, 0);
1557 
1558 	spin_lock_init(&ibp->rvp.lock);
1559 	/* Set the prefix to the default value (see ch. 4.1.1) */
1560 	ibp->rvp.gid_prefix = IB_DEFAULT_GID_PREFIX;
1561 	ibp->rvp.sm_lid = 0;
1562 	/*
1563 	 * Below should only set bits defined in OPA PortInfo.CapabilityMask
1564 	 * and PortInfo.CapabilityMask3
1565 	 */
1566 	ibp->rvp.port_cap_flags = IB_PORT_AUTO_MIGR_SUP |
1567 		IB_PORT_CAP_MASK_NOTICE_SUP;
1568 	ibp->rvp.port_cap3_flags = OPA_CAP_MASK3_IsSharedSpaceSupported;
1569 	ibp->rvp.pma_counter_select[0] = IB_PMA_PORT_XMIT_DATA;
1570 	ibp->rvp.pma_counter_select[1] = IB_PMA_PORT_RCV_DATA;
1571 	ibp->rvp.pma_counter_select[2] = IB_PMA_PORT_XMIT_PKTS;
1572 	ibp->rvp.pma_counter_select[3] = IB_PMA_PORT_RCV_PKTS;
1573 	ibp->rvp.pma_counter_select[4] = IB_PMA_PORT_XMIT_WAIT;
1574 
1575 	RCU_INIT_POINTER(ibp->rvp.qp[0], NULL);
1576 	RCU_INIT_POINTER(ibp->rvp.qp[1], NULL);
1577 }
1578 
1579 static void hfi1_get_dev_fw_str(struct ib_device *ibdev, char *str)
1580 {
1581 	struct rvt_dev_info *rdi = ib_to_rvt(ibdev);
1582 	struct hfi1_ibdev *dev = dev_from_rdi(rdi);
1583 	u32 ver = dd_from_dev(dev)->dc8051_ver;
1584 
1585 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%u.%u.%u", dc8051_ver_maj(ver),
1586 		 dc8051_ver_min(ver), dc8051_ver_patch(ver));
1587 }
1588 
1589 static const char * const driver_cntr_names[] = {
1590 	/* must be element 0*/
1591 	"DRIVER_KernIntr",
1592 	"DRIVER_ErrorIntr",
1593 	"DRIVER_Tx_Errs",
1594 	"DRIVER_Rcv_Errs",
1595 	"DRIVER_HW_Errs",
1596 	"DRIVER_NoPIOBufs",
1597 	"DRIVER_CtxtsOpen",
1598 	"DRIVER_RcvLen_Errs",
1599 	"DRIVER_EgrBufFull",
1600 	"DRIVER_EgrHdrFull"
1601 };
1602 
1603 static DEFINE_MUTEX(cntr_names_lock); /* protects the *_cntr_names bufers */
1604 static struct rdma_stat_desc *dev_cntr_descs;
1605 static struct rdma_stat_desc *port_cntr_descs;
1606 int num_driver_cntrs = ARRAY_SIZE(driver_cntr_names);
1607 static int num_dev_cntrs;
1608 static int num_port_cntrs;
1609 static int cntr_names_initialized;
1610 
1611 /*
1612  * Convert a list of names separated by '\n' into an array of NULL terminated
1613  * strings. Optionally some entries can be reserved in the array to hold extra
1614  * external strings.
1615  */
1616 static int init_cntr_names(const char *names_in, const size_t names_len,
1617 			   int num_extra_names, int *num_cntrs,
1618 			   struct rdma_stat_desc **cntr_descs)
1619 {
1620 	struct rdma_stat_desc *q;
1621 	char *names_out, *p;
1622 	int i, n;
1623 
1624 	n = 0;
1625 	for (i = 0; i < names_len; i++)
1626 		if (names_in[i] == '\n')
1627 			n++;
1628 
1629 	names_out =
1630 		kzalloc((n + num_extra_names) * sizeof(*q) + names_len,
1631 			GFP_KERNEL);
1632 	if (!names_out) {
1633 		*num_cntrs = 0;
1634 		*cntr_descs = NULL;
1635 		return -ENOMEM;
1636 	}
1637 
1638 	p = names_out + (n + num_extra_names) * sizeof(*q);
1639 	memcpy(p, names_in, names_len);
1640 
1641 	q = (struct rdma_stat_desc *)names_out;
1642 	for (i = 0; i < n; i++) {
1643 		q[i].name = p;
1644 		p = strchr(p, '\n');
1645 		*p++ = '\0';
1646 	}
1647 
1648 	*num_cntrs = n;
1649 	*cntr_descs = (struct rdma_stat_desc *)names_out;
1650 	return 0;
1651 }
1652 
1653 static int init_counters(struct ib_device *ibdev)
1654 {
1655 	struct hfi1_devdata *dd = dd_from_ibdev(ibdev);
1656 	int i, err = 0;
1657 
1658 	mutex_lock(&cntr_names_lock);
1659 	if (cntr_names_initialized)
1660 		goto out_unlock;
1661 
1662 	err = init_cntr_names(dd->cntrnames, dd->cntrnameslen, num_driver_cntrs,
1663 			      &num_dev_cntrs, &dev_cntr_descs);
1664 	if (err)
1665 		goto out_unlock;
1666 
1667 	for (i = 0; i < num_driver_cntrs; i++)
1668 		dev_cntr_descs[num_dev_cntrs + i].name = driver_cntr_names[i];
1669 
1670 	err = init_cntr_names(dd->portcntrnames, dd->portcntrnameslen, 0,
1671 			      &num_port_cntrs, &port_cntr_descs);
1672 	if (err) {
1673 		kfree(dev_cntr_descs);
1674 		dev_cntr_descs = NULL;
1675 		goto out_unlock;
1676 	}
1677 	cntr_names_initialized = 1;
1678 
1679 out_unlock:
1680 	mutex_unlock(&cntr_names_lock);
1681 	return err;
1682 }
1683 
1684 static struct rdma_hw_stats *hfi1_alloc_hw_device_stats(struct ib_device *ibdev)
1685 {
1686 	if (init_counters(ibdev))
1687 		return NULL;
1688 	return rdma_alloc_hw_stats_struct(dev_cntr_descs,
1689 					  num_dev_cntrs + num_driver_cntrs,
1690 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
1691 }
1692 
1693 static struct rdma_hw_stats *hfi_alloc_hw_port_stats(struct ib_device *ibdev,
1694 						     u32 port_num)
1695 {
1696 	if (init_counters(ibdev))
1697 		return NULL;
1698 	return rdma_alloc_hw_stats_struct(port_cntr_descs, num_port_cntrs,
1699 					  RDMA_HW_STATS_DEFAULT_LIFESPAN);
1700 }
1701 
1702 static u64 hfi1_sps_ints(void)
1703 {
1704 	unsigned long index, flags;
1705 	struct hfi1_devdata *dd;
1706 	u64 sps_ints = 0;
1707 
1708 	xa_lock_irqsave(&hfi1_dev_table, flags);
1709 	xa_for_each(&hfi1_dev_table, index, dd) {
1710 		sps_ints += get_all_cpu_total(dd->int_counter);
1711 	}
1712 	xa_unlock_irqrestore(&hfi1_dev_table, flags);
1713 	return sps_ints;
1714 }
1715 
1716 static int get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1717 			u32 port, int index)
1718 {
1719 	u64 *values;
1720 	int count;
1721 
1722 	if (!port) {
1723 		u64 *stats = (u64 *)&hfi1_stats;
1724 		int i;
1725 
1726 		hfi1_read_cntrs(dd_from_ibdev(ibdev), NULL, &values);
1727 		values[num_dev_cntrs] = hfi1_sps_ints();
1728 		for (i = 1; i < num_driver_cntrs; i++)
1729 			values[num_dev_cntrs + i] = stats[i];
1730 		count = num_dev_cntrs + num_driver_cntrs;
1731 	} else {
1732 		struct hfi1_ibport *ibp = to_iport(ibdev, port);
1733 
1734 		hfi1_read_portcntrs(ppd_from_ibp(ibp), NULL, &values);
1735 		count = num_port_cntrs;
1736 	}
1737 
1738 	memcpy(stats->value, values, count * sizeof(u64));
1739 	return count;
1740 }
1741 
1742 static const struct ib_device_ops hfi1_dev_ops = {
1743 	.owner = THIS_MODULE,
1744 	.driver_id = RDMA_DRIVER_HFI1,
1745 
1746 	.alloc_hw_device_stats = hfi1_alloc_hw_device_stats,
1747 	.alloc_hw_port_stats = hfi_alloc_hw_port_stats,
1748 	.alloc_rdma_netdev = hfi1_vnic_alloc_rn,
1749 	.device_group = &ib_hfi1_attr_group,
1750 	.get_dev_fw_str = hfi1_get_dev_fw_str,
1751 	.get_hw_stats = get_hw_stats,
1752 	.modify_device = modify_device,
1753 	.port_groups = hfi1_attr_port_groups,
1754 	/* keep process mad in the driver */
1755 	.process_mad = hfi1_process_mad,
1756 	.rdma_netdev_get_params = hfi1_ipoib_rn_get_params,
1757 };
1758 
1759 /**
1760  * hfi1_register_ib_device - register our device with the infiniband core
1761  * @dd: the device data structure
1762  * Return 0 if successful, errno if unsuccessful.
1763  */
1764 int hfi1_register_ib_device(struct hfi1_devdata *dd)
1765 {
1766 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1767 	struct ib_device *ibdev = &dev->rdi.ibdev;
1768 	struct hfi1_pportdata *ppd = dd->pport;
1769 	struct hfi1_ibport *ibp = &ppd->ibport_data;
1770 	unsigned i;
1771 	int ret;
1772 
1773 	for (i = 0; i < dd->num_pports; i++)
1774 		init_ibport(ppd + i);
1775 
1776 	/* Only need to initialize non-zero fields. */
1777 
1778 	timer_setup(&dev->mem_timer, mem_timer, 0);
1779 
1780 	seqlock_init(&dev->iowait_lock);
1781 	seqlock_init(&dev->txwait_lock);
1782 	INIT_LIST_HEAD(&dev->txwait);
1783 	INIT_LIST_HEAD(&dev->memwait);
1784 
1785 	ret = verbs_txreq_init(dev);
1786 	if (ret)
1787 		goto err_verbs_txreq;
1788 
1789 	/* Use first-port GUID as node guid */
1790 	ibdev->node_guid = get_sguid(ibp, HFI1_PORT_GUID_INDEX);
1791 
1792 	/*
1793 	 * The system image GUID is supposed to be the same for all
1794 	 * HFIs in a single system but since there can be other
1795 	 * device types in the system, we can't be sure this is unique.
1796 	 */
1797 	if (!ib_hfi1_sys_image_guid)
1798 		ib_hfi1_sys_image_guid = ibdev->node_guid;
1799 	ibdev->phys_port_cnt = dd->num_pports;
1800 	ibdev->dev.parent = &dd->pcidev->dev;
1801 
1802 	ib_set_device_ops(ibdev, &hfi1_dev_ops);
1803 
1804 	strlcpy(ibdev->node_desc, init_utsname()->nodename,
1805 		sizeof(ibdev->node_desc));
1806 
1807 	/*
1808 	 * Fill in rvt info object.
1809 	 */
1810 	dd->verbs_dev.rdi.driver_f.get_pci_dev = get_pci_dev;
1811 	dd->verbs_dev.rdi.driver_f.check_ah = hfi1_check_ah;
1812 	dd->verbs_dev.rdi.driver_f.notify_new_ah = hfi1_notify_new_ah;
1813 	dd->verbs_dev.rdi.driver_f.get_guid_be = hfi1_get_guid_be;
1814 	dd->verbs_dev.rdi.driver_f.query_port_state = query_port;
1815 	dd->verbs_dev.rdi.driver_f.shut_down_port = shut_down_port;
1816 	dd->verbs_dev.rdi.driver_f.cap_mask_chg = hfi1_cap_mask_chg;
1817 	/*
1818 	 * Fill in rvt info device attributes.
1819 	 */
1820 	hfi1_fill_device_attr(dd);
1821 
1822 	/* queue pair */
1823 	dd->verbs_dev.rdi.dparms.qp_table_size = hfi1_qp_table_size;
1824 	dd->verbs_dev.rdi.dparms.qpn_start = 0;
1825 	dd->verbs_dev.rdi.dparms.qpn_inc = 1;
1826 	dd->verbs_dev.rdi.dparms.qos_shift = dd->qos_shift;
1827 	dd->verbs_dev.rdi.dparms.qpn_res_start = RVT_KDETH_QP_BASE;
1828 	dd->verbs_dev.rdi.dparms.qpn_res_end = RVT_AIP_QP_MAX;
1829 	dd->verbs_dev.rdi.dparms.max_rdma_atomic = HFI1_MAX_RDMA_ATOMIC;
1830 	dd->verbs_dev.rdi.dparms.psn_mask = PSN_MASK;
1831 	dd->verbs_dev.rdi.dparms.psn_shift = PSN_SHIFT;
1832 	dd->verbs_dev.rdi.dparms.psn_modify_mask = PSN_MODIFY_MASK;
1833 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_INTEL_OPA |
1834 						RDMA_CORE_CAP_OPA_AH;
1835 	dd->verbs_dev.rdi.dparms.max_mad_size = OPA_MGMT_MAD_SIZE;
1836 
1837 	dd->verbs_dev.rdi.driver_f.qp_priv_alloc = qp_priv_alloc;
1838 	dd->verbs_dev.rdi.driver_f.qp_priv_init = hfi1_qp_priv_init;
1839 	dd->verbs_dev.rdi.driver_f.qp_priv_free = qp_priv_free;
1840 	dd->verbs_dev.rdi.driver_f.free_all_qps = free_all_qps;
1841 	dd->verbs_dev.rdi.driver_f.notify_qp_reset = notify_qp_reset;
1842 	dd->verbs_dev.rdi.driver_f.do_send = hfi1_do_send_from_rvt;
1843 	dd->verbs_dev.rdi.driver_f.schedule_send = hfi1_schedule_send;
1844 	dd->verbs_dev.rdi.driver_f.schedule_send_no_lock = _hfi1_schedule_send;
1845 	dd->verbs_dev.rdi.driver_f.get_pmtu_from_attr = get_pmtu_from_attr;
1846 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1847 	dd->verbs_dev.rdi.driver_f.flush_qp_waiters = flush_qp_waiters;
1848 	dd->verbs_dev.rdi.driver_f.stop_send_queue = stop_send_queue;
1849 	dd->verbs_dev.rdi.driver_f.quiesce_qp = quiesce_qp;
1850 	dd->verbs_dev.rdi.driver_f.notify_error_qp = notify_error_qp;
1851 	dd->verbs_dev.rdi.driver_f.mtu_from_qp = mtu_from_qp;
1852 	dd->verbs_dev.rdi.driver_f.mtu_to_path_mtu = mtu_to_path_mtu;
1853 	dd->verbs_dev.rdi.driver_f.check_modify_qp = hfi1_check_modify_qp;
1854 	dd->verbs_dev.rdi.driver_f.modify_qp = hfi1_modify_qp;
1855 	dd->verbs_dev.rdi.driver_f.notify_restart_rc = hfi1_restart_rc;
1856 	dd->verbs_dev.rdi.driver_f.setup_wqe = hfi1_setup_wqe;
1857 	dd->verbs_dev.rdi.driver_f.comp_vect_cpu_lookup =
1858 						hfi1_comp_vect_mappings_lookup;
1859 
1860 	/* completeion queue */
1861 	dd->verbs_dev.rdi.ibdev.num_comp_vectors = dd->comp_vect_possible_cpus;
1862 	dd->verbs_dev.rdi.dparms.node = dd->node;
1863 
1864 	/* misc settings */
1865 	dd->verbs_dev.rdi.flags = 0; /* Let rdmavt handle it all */
1866 	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
1867 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
1868 	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
1869 	dd->verbs_dev.rdi.dparms.sge_copy_mode = sge_copy_mode;
1870 	dd->verbs_dev.rdi.dparms.wss_threshold = wss_threshold;
1871 	dd->verbs_dev.rdi.dparms.wss_clean_period = wss_clean_period;
1872 	dd->verbs_dev.rdi.dparms.reserved_operations = 1;
1873 	dd->verbs_dev.rdi.dparms.extra_rdma_atomic = HFI1_TID_RDMA_WRITE_CNT;
1874 
1875 	/* post send table */
1876 	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;
1877 
1878 	/* opcode translation table */
1879 	dd->verbs_dev.rdi.wc_opcode = ib_hfi1_wc_opcode;
1880 
1881 	ppd = dd->pport;
1882 	for (i = 0; i < dd->num_pports; i++, ppd++)
1883 		rvt_init_port(&dd->verbs_dev.rdi,
1884 			      &ppd->ibport_data.rvp,
1885 			      i,
1886 			      ppd->pkeys);
1887 
1888 	ret = rvt_register_device(&dd->verbs_dev.rdi);
1889 	if (ret)
1890 		goto err_verbs_txreq;
1891 
1892 	ret = hfi1_verbs_register_sysfs(dd);
1893 	if (ret)
1894 		goto err_class;
1895 
1896 	return ret;
1897 
1898 err_class:
1899 	rvt_unregister_device(&dd->verbs_dev.rdi);
1900 err_verbs_txreq:
1901 	verbs_txreq_exit(dev);
1902 	dd_dev_err(dd, "cannot register verbs: %d!\n", -ret);
1903 	return ret;
1904 }
1905 
1906 void hfi1_unregister_ib_device(struct hfi1_devdata *dd)
1907 {
1908 	struct hfi1_ibdev *dev = &dd->verbs_dev;
1909 
1910 	hfi1_verbs_unregister_sysfs(dd);
1911 
1912 	rvt_unregister_device(&dd->verbs_dev.rdi);
1913 
1914 	if (!list_empty(&dev->txwait))
1915 		dd_dev_err(dd, "txwait list not empty!\n");
1916 	if (!list_empty(&dev->memwait))
1917 		dd_dev_err(dd, "memwait list not empty!\n");
1918 
1919 	del_timer_sync(&dev->mem_timer);
1920 	verbs_txreq_exit(dev);
1921 
1922 	mutex_lock(&cntr_names_lock);
1923 	kfree(dev_cntr_descs);
1924 	kfree(port_cntr_descs);
1925 	dev_cntr_descs = NULL;
1926 	port_cntr_descs = NULL;
1927 	cntr_names_initialized = 0;
1928 	mutex_unlock(&cntr_names_lock);
1929 }
1930 
1931 void hfi1_cnp_rcv(struct hfi1_packet *packet)
1932 {
1933 	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1934 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1935 	struct ib_header *hdr = packet->hdr;
1936 	struct rvt_qp *qp = packet->qp;
1937 	u32 lqpn, rqpn = 0;
1938 	u16 rlid = 0;
1939 	u8 sl, sc5, svc_type;
1940 
1941 	switch (packet->qp->ibqp.qp_type) {
1942 	case IB_QPT_UC:
1943 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
1944 		rqpn = qp->remote_qpn;
1945 		svc_type = IB_CC_SVCTYPE_UC;
1946 		break;
1947 	case IB_QPT_RC:
1948 		rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
1949 		rqpn = qp->remote_qpn;
1950 		svc_type = IB_CC_SVCTYPE_RC;
1951 		break;
1952 	case IB_QPT_SMI:
1953 	case IB_QPT_GSI:
1954 	case IB_QPT_UD:
1955 		svc_type = IB_CC_SVCTYPE_UD;
1956 		break;
1957 	default:
1958 		ibp->rvp.n_pkt_drops++;
1959 		return;
1960 	}
1961 
1962 	sc5 = hfi1_9B_get_sc5(hdr, packet->rhf);
1963 	sl = ibp->sc_to_sl[sc5];
1964 	lqpn = qp->ibqp.qp_num;
1965 
1966 	process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
1967 }
1968