1 /*
2 * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation.
3 * All rights reserved.
4 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5 *
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
11 *
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
14 * conditions are met:
15 *
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
18 * disclaimer.
19 *
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
24 *
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * SOFTWARE.
33 */
34
35 #ifndef QIB_VERBS_H
36 #define QIB_VERBS_H
37
38 #include <linux/types.h>
39 #include <linux/spinlock.h>
40 #include <linux/kernel.h>
41 #include <linux/interrupt.h>
42 #include <linux/kref.h>
43 #include <linux/workqueue.h>
44 #include <rdma/ib_pack.h>
45 #include <rdma/ib_user_verbs.h>
46
47 struct qib_ctxtdata;
48 struct qib_pportdata;
49 struct qib_devdata;
50 struct qib_verbs_txreq;
51
52 #define QIB_MAX_RDMA_ATOMIC 16
53 #define QIB_GUIDS_PER_PORT 5
54
55 #define QPN_MAX (1 << 24)
56 #define QPNMAP_ENTRIES (QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
57
58 /*
59 * Increment this value if any changes that break userspace ABI
60 * compatibility are made.
61 */
62 #define QIB_UVERBS_ABI_VERSION 2
63
64 /*
65 * Define an ib_cq_notify value that is not valid so we know when CQ
66 * notifications are armed.
67 */
68 #define IB_CQ_NONE (IB_CQ_NEXT_COMP + 1)
69
70 #define IB_SEQ_NAK (3 << 29)
71
72 /* AETH NAK opcode values */
73 #define IB_RNR_NAK 0x20
74 #define IB_NAK_PSN_ERROR 0x60
75 #define IB_NAK_INVALID_REQUEST 0x61
76 #define IB_NAK_REMOTE_ACCESS_ERROR 0x62
77 #define IB_NAK_REMOTE_OPERATIONAL_ERROR 0x63
78 #define IB_NAK_INVALID_RD_REQUEST 0x64
79
80 /* Flags for checking QP state (see ib_qib_state_ops[]) */
81 #define QIB_POST_SEND_OK 0x01
82 #define QIB_POST_RECV_OK 0x02
83 #define QIB_PROCESS_RECV_OK 0x04
84 #define QIB_PROCESS_SEND_OK 0x08
85 #define QIB_PROCESS_NEXT_SEND_OK 0x10
86 #define QIB_FLUSH_SEND 0x20
87 #define QIB_FLUSH_RECV 0x40
88 #define QIB_PROCESS_OR_FLUSH_SEND \
89 (QIB_PROCESS_SEND_OK | QIB_FLUSH_SEND)
90
91 /* IB Performance Manager status values */
92 #define IB_PMA_SAMPLE_STATUS_DONE 0x00
93 #define IB_PMA_SAMPLE_STATUS_STARTED 0x01
94 #define IB_PMA_SAMPLE_STATUS_RUNNING 0x02
95
96 /* Mandatory IB performance counter select values. */
97 #define IB_PMA_PORT_XMIT_DATA cpu_to_be16(0x0001)
98 #define IB_PMA_PORT_RCV_DATA cpu_to_be16(0x0002)
99 #define IB_PMA_PORT_XMIT_PKTS cpu_to_be16(0x0003)
100 #define IB_PMA_PORT_RCV_PKTS cpu_to_be16(0x0004)
101 #define IB_PMA_PORT_XMIT_WAIT cpu_to_be16(0x0005)
102
103 #define QIB_VENDOR_IPG cpu_to_be16(0xFFA0)
104
105 #define IB_BTH_REQ_ACK (1 << 31)
106 #define IB_BTH_SOLICITED (1 << 23)
107 #define IB_BTH_MIG_REQ (1 << 22)
108
109 /* XXX Should be defined in ib_verbs.h enum ib_port_cap_flags */
110 #define IB_PORT_OTHER_LOCAL_CHANGES_SUP (1 << 26)
111
112 #define IB_GRH_VERSION 6
113 #define IB_GRH_VERSION_MASK 0xF
114 #define IB_GRH_VERSION_SHIFT 28
115 #define IB_GRH_TCLASS_MASK 0xFF
116 #define IB_GRH_TCLASS_SHIFT 20
117 #define IB_GRH_FLOW_MASK 0xFFFFF
118 #define IB_GRH_FLOW_SHIFT 0
119 #define IB_GRH_NEXT_HDR 0x1B
120
121 #define IB_DEFAULT_GID_PREFIX cpu_to_be64(0xfe80000000000000ULL)
122
123 /* Values for set/get portinfo VLCap OperationalVLs */
124 #define IB_VL_VL0 1
125 #define IB_VL_VL0_1 2
126 #define IB_VL_VL0_3 3
127 #define IB_VL_VL0_7 4
128 #define IB_VL_VL0_14 5
129
qib_num_vls(int vls)130 static inline int qib_num_vls(int vls)
131 {
132 switch (vls) {
133 default:
134 case IB_VL_VL0:
135 return 1;
136 case IB_VL_VL0_1:
137 return 2;
138 case IB_VL_VL0_3:
139 return 4;
140 case IB_VL_VL0_7:
141 return 8;
142 case IB_VL_VL0_14:
143 return 15;
144 }
145 }
146
147 struct ib_reth {
148 __be64 vaddr;
149 __be32 rkey;
150 __be32 length;
151 } __attribute__ ((packed));
152
153 struct ib_atomic_eth {
154 __be32 vaddr[2]; /* unaligned so access as 2 32-bit words */
155 __be32 rkey;
156 __be64 swap_data;
157 __be64 compare_data;
158 } __attribute__ ((packed));
159
160 struct qib_other_headers {
161 __be32 bth[3];
162 union {
163 struct {
164 __be32 deth[2];
165 __be32 imm_data;
166 } ud;
167 struct {
168 struct ib_reth reth;
169 __be32 imm_data;
170 } rc;
171 struct {
172 __be32 aeth;
173 __be32 atomic_ack_eth[2];
174 } at;
175 __be32 imm_data;
176 __be32 aeth;
177 struct ib_atomic_eth atomic_eth;
178 } u;
179 } __attribute__ ((packed));
180
181 /*
182 * Note that UD packets with a GRH header are 8+40+12+8 = 68 bytes
183 * long (72 w/ imm_data). Only the first 56 bytes of the IB header
184 * will be in the eager header buffer. The remaining 12 or 16 bytes
185 * are in the data buffer.
186 */
187 struct qib_ib_header {
188 __be16 lrh[4];
189 union {
190 struct {
191 struct ib_grh grh;
192 struct qib_other_headers oth;
193 } l;
194 struct qib_other_headers oth;
195 } u;
196 } __attribute__ ((packed));
197
198 struct qib_pio_header {
199 __le32 pbc[2];
200 struct qib_ib_header hdr;
201 } __attribute__ ((packed));
202
203 /*
204 * There is one struct qib_mcast for each multicast GID.
205 * All attached QPs are then stored as a list of
206 * struct qib_mcast_qp.
207 */
208 struct qib_mcast_qp {
209 struct list_head list;
210 struct qib_qp *qp;
211 };
212
213 struct qib_mcast {
214 struct rb_node rb_node;
215 union ib_gid mgid;
216 struct list_head qp_list;
217 wait_queue_head_t wait;
218 atomic_t refcount;
219 int n_attached;
220 };
221
222 /* Protection domain */
223 struct qib_pd {
224 struct ib_pd ibpd;
225 int user; /* non-zero if created from user space */
226 };
227
228 /* Address Handle */
229 struct qib_ah {
230 struct ib_ah ibah;
231 struct ib_ah_attr attr;
232 atomic_t refcount;
233 };
234
235 /*
236 * This structure is used by qib_mmap() to validate an offset
237 * when an mmap() request is made. The vm_area_struct then uses
238 * this as its vm_private_data.
239 */
240 struct qib_mmap_info {
241 struct list_head pending_mmaps;
242 struct ib_ucontext *context;
243 void *obj;
244 __u64 offset;
245 struct kref ref;
246 unsigned size;
247 };
248
249 /*
250 * This structure is used to contain the head pointer, tail pointer,
251 * and completion queue entries as a single memory allocation so
252 * it can be mmap'ed into user space.
253 */
254 struct qib_cq_wc {
255 u32 head; /* index of next entry to fill */
256 u32 tail; /* index of next ib_poll_cq() entry */
257 union {
258 /* these are actually size ibcq.cqe + 1 */
259 struct ib_uverbs_wc uqueue[0];
260 struct ib_wc kqueue[0];
261 };
262 };
263
264 /*
265 * The completion queue structure.
266 */
267 struct qib_cq {
268 struct ib_cq ibcq;
269 struct work_struct comptask;
270 spinlock_t lock; /* protect changes in this struct */
271 u8 notify;
272 u8 triggered;
273 struct qib_cq_wc *queue;
274 struct qib_mmap_info *ip;
275 };
276
277 /*
278 * A segment is a linear region of low physical memory.
279 * XXX Maybe we should use phys addr here and kmap()/kunmap().
280 * Used by the verbs layer.
281 */
282 struct qib_seg {
283 void *vaddr;
284 size_t length;
285 };
286
287 /* The number of qib_segs that fit in a page. */
288 #define QIB_SEGSZ (PAGE_SIZE / sizeof(struct qib_seg))
289
290 struct qib_segarray {
291 struct qib_seg segs[QIB_SEGSZ];
292 };
293
294 struct qib_mregion {
295 struct ib_pd *pd; /* shares refcnt of ibmr.pd */
296 u64 user_base; /* User's address for this region */
297 u64 iova; /* IB start address of this region */
298 size_t length;
299 u32 lkey;
300 u32 offset; /* offset (bytes) to start of region */
301 int access_flags;
302 u32 max_segs; /* number of qib_segs in all the arrays */
303 u32 mapsz; /* size of the map array */
304 u8 page_shift; /* 0 - non unform/non powerof2 sizes */
305 atomic_t refcount;
306 struct qib_segarray *map[0]; /* the segments */
307 };
308
309 /*
310 * These keep track of the copy progress within a memory region.
311 * Used by the verbs layer.
312 */
313 struct qib_sge {
314 struct qib_mregion *mr;
315 void *vaddr; /* kernel virtual address of segment */
316 u32 sge_length; /* length of the SGE */
317 u32 length; /* remaining length of the segment */
318 u16 m; /* current index: mr->map[m] */
319 u16 n; /* current index: mr->map[m]->segs[n] */
320 };
321
322 /* Memory region */
323 struct qib_mr {
324 struct ib_mr ibmr;
325 struct ib_umem *umem;
326 struct qib_mregion mr; /* must be last */
327 };
328
329 /*
330 * Send work request queue entry.
331 * The size of the sg_list is determined when the QP is created and stored
332 * in qp->s_max_sge.
333 */
334 struct qib_swqe {
335 struct ib_send_wr wr; /* don't use wr.sg_list */
336 u32 psn; /* first packet sequence number */
337 u32 lpsn; /* last packet sequence number */
338 u32 ssn; /* send sequence number */
339 u32 length; /* total length of data in sg_list */
340 struct qib_sge sg_list[0];
341 };
342
343 /*
344 * Receive work request queue entry.
345 * The size of the sg_list is determined when the QP (or SRQ) is created
346 * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
347 */
348 struct qib_rwqe {
349 u64 wr_id;
350 u8 num_sge;
351 struct ib_sge sg_list[0];
352 };
353
354 /*
355 * This structure is used to contain the head pointer, tail pointer,
356 * and receive work queue entries as a single memory allocation so
357 * it can be mmap'ed into user space.
358 * Note that the wq array elements are variable size so you can't
359 * just index into the array to get the N'th element;
360 * use get_rwqe_ptr() instead.
361 */
362 struct qib_rwq {
363 u32 head; /* new work requests posted to the head */
364 u32 tail; /* receives pull requests from here. */
365 struct qib_rwqe wq[0];
366 };
367
368 struct qib_rq {
369 struct qib_rwq *wq;
370 spinlock_t lock; /* protect changes in this struct */
371 u32 size; /* size of RWQE array */
372 u8 max_sge;
373 };
374
375 struct qib_srq {
376 struct ib_srq ibsrq;
377 struct qib_rq rq;
378 struct qib_mmap_info *ip;
379 /* send signal when number of RWQEs < limit */
380 u32 limit;
381 };
382
383 struct qib_sge_state {
384 struct qib_sge *sg_list; /* next SGE to be used if any */
385 struct qib_sge sge; /* progress state for the current SGE */
386 u32 total_len;
387 u8 num_sge;
388 };
389
390 /*
391 * This structure holds the information that the send tasklet needs
392 * to send a RDMA read response or atomic operation.
393 */
394 struct qib_ack_entry {
395 u8 opcode;
396 u8 sent;
397 u32 psn;
398 u32 lpsn;
399 union {
400 struct qib_sge rdma_sge;
401 u64 atomic_data;
402 };
403 };
404
405 /*
406 * Variables prefixed with s_ are for the requester (sender).
407 * Variables prefixed with r_ are for the responder (receiver).
408 * Variables prefixed with ack_ are for responder replies.
409 *
410 * Common variables are protected by both r_rq.lock and s_lock in that order
411 * which only happens in modify_qp() or changing the QP 'state'.
412 */
413 struct qib_qp {
414 struct ib_qp ibqp;
415 struct qib_qp *next; /* link list for QPN hash table */
416 struct qib_qp *timer_next; /* link list for qib_ib_timer() */
417 struct list_head iowait; /* link for wait PIO buf */
418 struct list_head rspwait; /* link for waititing to respond */
419 struct ib_ah_attr remote_ah_attr;
420 struct ib_ah_attr alt_ah_attr;
421 struct qib_ib_header s_hdr; /* next packet header to send */
422 atomic_t refcount;
423 wait_queue_head_t wait;
424 wait_queue_head_t wait_dma;
425 struct timer_list s_timer;
426 struct work_struct s_work;
427 struct qib_mmap_info *ip;
428 struct qib_sge_state *s_cur_sge;
429 struct qib_verbs_txreq *s_tx;
430 struct qib_mregion *s_rdma_mr;
431 struct qib_sge_state s_sge; /* current send request data */
432 struct qib_ack_entry s_ack_queue[QIB_MAX_RDMA_ATOMIC + 1];
433 struct qib_sge_state s_ack_rdma_sge;
434 struct qib_sge_state s_rdma_read_sge;
435 struct qib_sge_state r_sge; /* current receive data */
436 spinlock_t r_lock; /* used for APM */
437 spinlock_t s_lock;
438 atomic_t s_dma_busy;
439 u32 s_flags;
440 u32 s_cur_size; /* size of send packet in bytes */
441 u32 s_len; /* total length of s_sge */
442 u32 s_rdma_read_len; /* total length of s_rdma_read_sge */
443 u32 s_next_psn; /* PSN for next request */
444 u32 s_last_psn; /* last response PSN processed */
445 u32 s_sending_psn; /* lowest PSN that is being sent */
446 u32 s_sending_hpsn; /* highest PSN that is being sent */
447 u32 s_psn; /* current packet sequence number */
448 u32 s_ack_rdma_psn; /* PSN for sending RDMA read responses */
449 u32 s_ack_psn; /* PSN for acking sends and RDMA writes */
450 u32 s_rnr_timeout; /* number of milliseconds for RNR timeout */
451 u32 r_ack_psn; /* PSN for next ACK or atomic ACK */
452 u64 r_wr_id; /* ID for current receive WQE */
453 unsigned long r_aflags;
454 u32 r_len; /* total length of r_sge */
455 u32 r_rcv_len; /* receive data len processed */
456 u32 r_psn; /* expected rcv packet sequence number */
457 u32 r_msn; /* message sequence number */
458 u16 s_hdrwords; /* size of s_hdr in 32 bit words */
459 u16 s_rdma_ack_cnt;
460 u8 state; /* QP state */
461 u8 s_state; /* opcode of last packet sent */
462 u8 s_ack_state; /* opcode of packet to ACK */
463 u8 s_nak_state; /* non-zero if NAK is pending */
464 u8 r_state; /* opcode of last packet received */
465 u8 r_nak_state; /* non-zero if NAK is pending */
466 u8 r_min_rnr_timer; /* retry timeout value for RNR NAKs */
467 u8 r_flags;
468 u8 r_max_rd_atomic; /* max number of RDMA read/atomic to receive */
469 u8 r_head_ack_queue; /* index into s_ack_queue[] */
470 u8 qp_access_flags;
471 u8 s_max_sge; /* size of s_wq->sg_list */
472 u8 s_retry_cnt; /* number of times to retry */
473 u8 s_rnr_retry_cnt;
474 u8 s_retry; /* requester retry counter */
475 u8 s_rnr_retry; /* requester RNR retry counter */
476 u8 s_pkey_index; /* PKEY index to use */
477 u8 s_alt_pkey_index; /* Alternate path PKEY index to use */
478 u8 s_max_rd_atomic; /* max number of RDMA read/atomic to send */
479 u8 s_num_rd_atomic; /* number of RDMA read/atomic pending */
480 u8 s_tail_ack_queue; /* index into s_ack_queue[] */
481 u8 s_srate;
482 u8 s_draining;
483 u8 s_mig_state;
484 u8 timeout; /* Timeout for this QP */
485 u8 alt_timeout; /* Alternate path timeout for this QP */
486 u8 port_num;
487 enum ib_mtu path_mtu;
488 u32 remote_qpn;
489 u32 qkey; /* QKEY for this QP (for UD or RD) */
490 u32 s_size; /* send work queue size */
491 u32 s_head; /* new entries added here */
492 u32 s_tail; /* next entry to process */
493 u32 s_cur; /* current work queue entry */
494 u32 s_acked; /* last un-ACK'ed entry */
495 u32 s_last; /* last completed entry */
496 u32 s_ssn; /* SSN of tail entry */
497 u32 s_lsn; /* limit sequence number (credit) */
498 struct qib_swqe *s_wq; /* send work queue */
499 struct qib_swqe *s_wqe;
500 struct qib_rq r_rq; /* receive work queue */
501 struct qib_sge r_sg_list[0]; /* verified SGEs */
502 };
503
504 /*
505 * Atomic bit definitions for r_aflags.
506 */
507 #define QIB_R_WRID_VALID 0
508 #define QIB_R_REWIND_SGE 1
509
510 /*
511 * Bit definitions for r_flags.
512 */
513 #define QIB_R_REUSE_SGE 0x01
514 #define QIB_R_RDMAR_SEQ 0x02
515 #define QIB_R_RSP_NAK 0x04
516 #define QIB_R_RSP_SEND 0x08
517 #define QIB_R_COMM_EST 0x10
518
519 /*
520 * Bit definitions for s_flags.
521 *
522 * QIB_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
523 * QIB_S_BUSY - send tasklet is processing the QP
524 * QIB_S_TIMER - the RC retry timer is active
525 * QIB_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
526 * QIB_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
527 * before processing the next SWQE
528 * QIB_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
529 * before processing the next SWQE
530 * QIB_S_WAIT_RNR - waiting for RNR timeout
531 * QIB_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
532 * QIB_S_WAIT_DMA - waiting for send DMA queue to drain before generating
533 * next send completion entry not via send DMA
534 * QIB_S_WAIT_PIO - waiting for a send buffer to be available
535 * QIB_S_WAIT_TX - waiting for a struct qib_verbs_txreq to be available
536 * QIB_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
537 * QIB_S_WAIT_KMEM - waiting for kernel memory to be available
538 * QIB_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
539 * QIB_S_WAIT_ACK - waiting for an ACK packet before sending more requests
540 * QIB_S_SEND_ONE - send one packet, request ACK, then wait for ACK
541 */
542 #define QIB_S_SIGNAL_REQ_WR 0x0001
543 #define QIB_S_BUSY 0x0002
544 #define QIB_S_TIMER 0x0004
545 #define QIB_S_RESP_PENDING 0x0008
546 #define QIB_S_ACK_PENDING 0x0010
547 #define QIB_S_WAIT_FENCE 0x0020
548 #define QIB_S_WAIT_RDMAR 0x0040
549 #define QIB_S_WAIT_RNR 0x0080
550 #define QIB_S_WAIT_SSN_CREDIT 0x0100
551 #define QIB_S_WAIT_DMA 0x0200
552 #define QIB_S_WAIT_PIO 0x0400
553 #define QIB_S_WAIT_TX 0x0800
554 #define QIB_S_WAIT_DMA_DESC 0x1000
555 #define QIB_S_WAIT_KMEM 0x2000
556 #define QIB_S_WAIT_PSN 0x4000
557 #define QIB_S_WAIT_ACK 0x8000
558 #define QIB_S_SEND_ONE 0x10000
559 #define QIB_S_UNLIMITED_CREDIT 0x20000
560
561 /*
562 * Wait flags that would prevent any packet type from being sent.
563 */
564 #define QIB_S_ANY_WAIT_IO (QIB_S_WAIT_PIO | QIB_S_WAIT_TX | \
565 QIB_S_WAIT_DMA_DESC | QIB_S_WAIT_KMEM)
566
567 /*
568 * Wait flags that would prevent send work requests from making progress.
569 */
570 #define QIB_S_ANY_WAIT_SEND (QIB_S_WAIT_FENCE | QIB_S_WAIT_RDMAR | \
571 QIB_S_WAIT_RNR | QIB_S_WAIT_SSN_CREDIT | QIB_S_WAIT_DMA | \
572 QIB_S_WAIT_PSN | QIB_S_WAIT_ACK)
573
574 #define QIB_S_ANY_WAIT (QIB_S_ANY_WAIT_IO | QIB_S_ANY_WAIT_SEND)
575
576 #define QIB_PSN_CREDIT 16
577
578 /*
579 * Since struct qib_swqe is not a fixed size, we can't simply index into
580 * struct qib_qp.s_wq. This function does the array index computation.
581 */
get_swqe_ptr(struct qib_qp * qp,unsigned n)582 static inline struct qib_swqe *get_swqe_ptr(struct qib_qp *qp,
583 unsigned n)
584 {
585 return (struct qib_swqe *)((char *)qp->s_wq +
586 (sizeof(struct qib_swqe) +
587 qp->s_max_sge *
588 sizeof(struct qib_sge)) * n);
589 }
590
591 /*
592 * Since struct qib_rwqe is not a fixed size, we can't simply index into
593 * struct qib_rwq.wq. This function does the array index computation.
594 */
get_rwqe_ptr(struct qib_rq * rq,unsigned n)595 static inline struct qib_rwqe *get_rwqe_ptr(struct qib_rq *rq, unsigned n)
596 {
597 return (struct qib_rwqe *)
598 ((char *) rq->wq->wq +
599 (sizeof(struct qib_rwqe) +
600 rq->max_sge * sizeof(struct ib_sge)) * n);
601 }
602
603 /*
604 * QPN-map pages start out as NULL, they get allocated upon
605 * first use and are never deallocated. This way,
606 * large bitmaps are not allocated unless large numbers of QPs are used.
607 */
608 struct qpn_map {
609 void *page;
610 };
611
612 struct qib_qpn_table {
613 spinlock_t lock; /* protect changes in this struct */
614 unsigned flags; /* flags for QP0/1 allocated for each port */
615 u32 last; /* last QP number allocated */
616 u32 nmaps; /* size of the map table */
617 u16 limit;
618 u16 mask;
619 /* bit map of free QP numbers other than 0/1 */
620 struct qpn_map map[QPNMAP_ENTRIES];
621 };
622
623 struct qib_lkey_table {
624 spinlock_t lock; /* protect changes in this struct */
625 u32 next; /* next unused index (speeds search) */
626 u32 gen; /* generation count */
627 u32 max; /* size of the table */
628 struct qib_mregion **table;
629 };
630
631 struct qib_opcode_stats {
632 u64 n_packets; /* number of packets */
633 u64 n_bytes; /* total number of bytes */
634 };
635
636 struct qib_ibport {
637 struct qib_qp *qp0;
638 struct qib_qp *qp1;
639 struct ib_mad_agent *send_agent; /* agent for SMI (traps) */
640 struct qib_ah *sm_ah;
641 struct qib_ah *smi_ah;
642 struct rb_root mcast_tree;
643 spinlock_t lock; /* protect changes in this struct */
644
645 /* non-zero when timer is set */
646 unsigned long mkey_lease_timeout;
647 unsigned long trap_timeout;
648 __be64 gid_prefix; /* in network order */
649 __be64 mkey;
650 __be64 guids[QIB_GUIDS_PER_PORT - 1]; /* writable GUIDs */
651 u64 tid; /* TID for traps */
652 u64 n_unicast_xmit; /* total unicast packets sent */
653 u64 n_unicast_rcv; /* total unicast packets received */
654 u64 n_multicast_xmit; /* total multicast packets sent */
655 u64 n_multicast_rcv; /* total multicast packets received */
656 u64 z_symbol_error_counter; /* starting count for PMA */
657 u64 z_link_error_recovery_counter; /* starting count for PMA */
658 u64 z_link_downed_counter; /* starting count for PMA */
659 u64 z_port_rcv_errors; /* starting count for PMA */
660 u64 z_port_rcv_remphys_errors; /* starting count for PMA */
661 u64 z_port_xmit_discards; /* starting count for PMA */
662 u64 z_port_xmit_data; /* starting count for PMA */
663 u64 z_port_rcv_data; /* starting count for PMA */
664 u64 z_port_xmit_packets; /* starting count for PMA */
665 u64 z_port_rcv_packets; /* starting count for PMA */
666 u32 z_local_link_integrity_errors; /* starting count for PMA */
667 u32 z_excessive_buffer_overrun_errors; /* starting count for PMA */
668 u32 z_vl15_dropped; /* starting count for PMA */
669 u32 n_rc_resends;
670 u32 n_rc_acks;
671 u32 n_rc_qacks;
672 u32 n_rc_delayed_comp;
673 u32 n_seq_naks;
674 u32 n_rdma_seq;
675 u32 n_rnr_naks;
676 u32 n_other_naks;
677 u32 n_loop_pkts;
678 u32 n_pkt_drops;
679 u32 n_vl15_dropped;
680 u32 n_rc_timeouts;
681 u32 n_dmawait;
682 u32 n_unaligned;
683 u32 n_rc_dupreq;
684 u32 n_rc_seqnak;
685 u32 port_cap_flags;
686 u32 pma_sample_start;
687 u32 pma_sample_interval;
688 __be16 pma_counter_select[5];
689 u16 pma_tag;
690 u16 pkey_violations;
691 u16 qkey_violations;
692 u16 mkey_violations;
693 u16 mkey_lease_period;
694 u16 sm_lid;
695 u16 repress_traps;
696 u8 sm_sl;
697 u8 mkeyprot;
698 u8 subnet_timeout;
699 u8 vl_high_limit;
700 u8 sl_to_vl[16];
701
702 struct qib_opcode_stats opstats[128];
703 };
704
705 struct qib_ibdev {
706 struct ib_device ibdev;
707 struct list_head pending_mmaps;
708 spinlock_t mmap_offset_lock; /* protect mmap_offset */
709 u32 mmap_offset;
710 struct qib_mregion *dma_mr;
711
712 /* QP numbers are shared by all IB ports */
713 struct qib_qpn_table qpn_table;
714 struct qib_lkey_table lk_table;
715 struct list_head piowait; /* list for wait PIO buf */
716 struct list_head dmawait; /* list for wait DMA */
717 struct list_head txwait; /* list for wait qib_verbs_txreq */
718 struct list_head memwait; /* list for wait kernel memory */
719 struct list_head txreq_free;
720 struct timer_list mem_timer;
721 struct qib_qp **qp_table;
722 struct qib_pio_header *pio_hdrs;
723 dma_addr_t pio_hdrs_phys;
724 /* list of QPs waiting for RNR timer */
725 spinlock_t pending_lock; /* protect wait lists, PMA counters, etc. */
726 unsigned qp_table_size; /* size of the hash table */
727 spinlock_t qpt_lock;
728
729 u32 n_piowait;
730 u32 n_txwait;
731
732 u32 n_pds_allocated; /* number of PDs allocated for device */
733 spinlock_t n_pds_lock;
734 u32 n_ahs_allocated; /* number of AHs allocated for device */
735 spinlock_t n_ahs_lock;
736 u32 n_cqs_allocated; /* number of CQs allocated for device */
737 spinlock_t n_cqs_lock;
738 u32 n_qps_allocated; /* number of QPs allocated for device */
739 spinlock_t n_qps_lock;
740 u32 n_srqs_allocated; /* number of SRQs allocated for device */
741 spinlock_t n_srqs_lock;
742 u32 n_mcast_grps_allocated; /* number of mcast groups allocated */
743 spinlock_t n_mcast_grps_lock;
744 };
745
746 struct qib_verbs_counters {
747 u64 symbol_error_counter;
748 u64 link_error_recovery_counter;
749 u64 link_downed_counter;
750 u64 port_rcv_errors;
751 u64 port_rcv_remphys_errors;
752 u64 port_xmit_discards;
753 u64 port_xmit_data;
754 u64 port_rcv_data;
755 u64 port_xmit_packets;
756 u64 port_rcv_packets;
757 u32 local_link_integrity_errors;
758 u32 excessive_buffer_overrun_errors;
759 u32 vl15_dropped;
760 };
761
to_imr(struct ib_mr * ibmr)762 static inline struct qib_mr *to_imr(struct ib_mr *ibmr)
763 {
764 return container_of(ibmr, struct qib_mr, ibmr);
765 }
766
to_ipd(struct ib_pd * ibpd)767 static inline struct qib_pd *to_ipd(struct ib_pd *ibpd)
768 {
769 return container_of(ibpd, struct qib_pd, ibpd);
770 }
771
to_iah(struct ib_ah * ibah)772 static inline struct qib_ah *to_iah(struct ib_ah *ibah)
773 {
774 return container_of(ibah, struct qib_ah, ibah);
775 }
776
to_icq(struct ib_cq * ibcq)777 static inline struct qib_cq *to_icq(struct ib_cq *ibcq)
778 {
779 return container_of(ibcq, struct qib_cq, ibcq);
780 }
781
to_isrq(struct ib_srq * ibsrq)782 static inline struct qib_srq *to_isrq(struct ib_srq *ibsrq)
783 {
784 return container_of(ibsrq, struct qib_srq, ibsrq);
785 }
786
to_iqp(struct ib_qp * ibqp)787 static inline struct qib_qp *to_iqp(struct ib_qp *ibqp)
788 {
789 return container_of(ibqp, struct qib_qp, ibqp);
790 }
791
to_idev(struct ib_device * ibdev)792 static inline struct qib_ibdev *to_idev(struct ib_device *ibdev)
793 {
794 return container_of(ibdev, struct qib_ibdev, ibdev);
795 }
796
797 /*
798 * Send if not busy or waiting for I/O and either
799 * a RC response is pending or we can process send work requests.
800 */
qib_send_ok(struct qib_qp * qp)801 static inline int qib_send_ok(struct qib_qp *qp)
802 {
803 return !(qp->s_flags & (QIB_S_BUSY | QIB_S_ANY_WAIT_IO)) &&
804 (qp->s_hdrwords || (qp->s_flags & QIB_S_RESP_PENDING) ||
805 !(qp->s_flags & QIB_S_ANY_WAIT_SEND));
806 }
807
808 extern struct workqueue_struct *qib_cq_wq;
809
810 /*
811 * This must be called with s_lock held.
812 */
qib_schedule_send(struct qib_qp * qp)813 static inline void qib_schedule_send(struct qib_qp *qp)
814 {
815 if (qib_send_ok(qp))
816 queue_work(ib_wq, &qp->s_work);
817 }
818
qib_pkey_ok(u16 pkey1,u16 pkey2)819 static inline int qib_pkey_ok(u16 pkey1, u16 pkey2)
820 {
821 u16 p1 = pkey1 & 0x7FFF;
822 u16 p2 = pkey2 & 0x7FFF;
823
824 /*
825 * Low 15 bits must be non-zero and match, and
826 * one of the two must be a full member.
827 */
828 return p1 && p1 == p2 && ((__s16)pkey1 < 0 || (__s16)pkey2 < 0);
829 }
830
831 void qib_bad_pqkey(struct qib_ibport *ibp, __be16 trap_num, u32 key, u32 sl,
832 u32 qp1, u32 qp2, __be16 lid1, __be16 lid2);
833 void qib_cap_mask_chg(struct qib_ibport *ibp);
834 void qib_sys_guid_chg(struct qib_ibport *ibp);
835 void qib_node_desc_chg(struct qib_ibport *ibp);
836 int qib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
837 struct ib_wc *in_wc, struct ib_grh *in_grh,
838 struct ib_mad *in_mad, struct ib_mad *out_mad);
839 int qib_create_agents(struct qib_ibdev *dev);
840 void qib_free_agents(struct qib_ibdev *dev);
841
842 /*
843 * Compare the lower 24 bits of the two values.
844 * Returns an integer <, ==, or > than zero.
845 */
qib_cmp24(u32 a,u32 b)846 static inline int qib_cmp24(u32 a, u32 b)
847 {
848 return (((int) a) - ((int) b)) << 8;
849 }
850
851 struct qib_mcast *qib_mcast_find(struct qib_ibport *ibp, union ib_gid *mgid);
852
853 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
854 u64 *rwords, u64 *spkts, u64 *rpkts,
855 u64 *xmit_wait);
856
857 int qib_get_counters(struct qib_pportdata *ppd,
858 struct qib_verbs_counters *cntrs);
859
860 int qib_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
861
862 int qib_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
863
864 int qib_mcast_tree_empty(struct qib_ibport *ibp);
865
866 __be32 qib_compute_aeth(struct qib_qp *qp);
867
868 struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn);
869
870 struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
871 struct ib_qp_init_attr *init_attr,
872 struct ib_udata *udata);
873
874 int qib_destroy_qp(struct ib_qp *ibqp);
875
876 int qib_error_qp(struct qib_qp *qp, enum ib_wc_status err);
877
878 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
879 int attr_mask, struct ib_udata *udata);
880
881 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
882 int attr_mask, struct ib_qp_init_attr *init_attr);
883
884 unsigned qib_free_all_qps(struct qib_devdata *dd);
885
886 void qib_init_qpn_table(struct qib_devdata *dd, struct qib_qpn_table *qpt);
887
888 void qib_free_qpn_table(struct qib_qpn_table *qpt);
889
890 void qib_get_credit(struct qib_qp *qp, u32 aeth);
891
892 unsigned qib_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult);
893
894 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail);
895
896 void qib_put_txreq(struct qib_verbs_txreq *tx);
897
898 int qib_verbs_send(struct qib_qp *qp, struct qib_ib_header *hdr,
899 u32 hdrwords, struct qib_sge_state *ss, u32 len);
900
901 void qib_copy_sge(struct qib_sge_state *ss, void *data, u32 length,
902 int release);
903
904 void qib_skip_sge(struct qib_sge_state *ss, u32 length, int release);
905
906 void qib_uc_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
907 int has_grh, void *data, u32 tlen, struct qib_qp *qp);
908
909 void qib_rc_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
910 int has_grh, void *data, u32 tlen, struct qib_qp *qp);
911
912 int qib_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr);
913
914 void qib_rc_rnr_retry(unsigned long arg);
915
916 void qib_rc_send_complete(struct qib_qp *qp, struct qib_ib_header *hdr);
917
918 void qib_rc_error(struct qib_qp *qp, enum ib_wc_status err);
919
920 int qib_post_ud_send(struct qib_qp *qp, struct ib_send_wr *wr);
921
922 void qib_ud_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
923 int has_grh, void *data, u32 tlen, struct qib_qp *qp);
924
925 int qib_alloc_lkey(struct qib_lkey_table *rkt, struct qib_mregion *mr);
926
927 int qib_free_lkey(struct qib_ibdev *dev, struct qib_mregion *mr);
928
929 int qib_lkey_ok(struct qib_lkey_table *rkt, struct qib_pd *pd,
930 struct qib_sge *isge, struct ib_sge *sge, int acc);
931
932 int qib_rkey_ok(struct qib_qp *qp, struct qib_sge *sge,
933 u32 len, u64 vaddr, u32 rkey, int acc);
934
935 int qib_post_srq_receive(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
936 struct ib_recv_wr **bad_wr);
937
938 struct ib_srq *qib_create_srq(struct ib_pd *ibpd,
939 struct ib_srq_init_attr *srq_init_attr,
940 struct ib_udata *udata);
941
942 int qib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
943 enum ib_srq_attr_mask attr_mask,
944 struct ib_udata *udata);
945
946 int qib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr);
947
948 int qib_destroy_srq(struct ib_srq *ibsrq);
949
950 void qib_cq_enter(struct qib_cq *cq, struct ib_wc *entry, int sig);
951
952 int qib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry);
953
954 struct ib_cq *qib_create_cq(struct ib_device *ibdev, int entries,
955 int comp_vector, struct ib_ucontext *context,
956 struct ib_udata *udata);
957
958 int qib_destroy_cq(struct ib_cq *ibcq);
959
960 int qib_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags);
961
962 int qib_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata);
963
964 struct ib_mr *qib_get_dma_mr(struct ib_pd *pd, int acc);
965
966 struct ib_mr *qib_reg_phys_mr(struct ib_pd *pd,
967 struct ib_phys_buf *buffer_list,
968 int num_phys_buf, int acc, u64 *iova_start);
969
970 struct ib_mr *qib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
971 u64 virt_addr, int mr_access_flags,
972 struct ib_udata *udata);
973
974 int qib_dereg_mr(struct ib_mr *ibmr);
975
976 struct ib_mr *qib_alloc_fast_reg_mr(struct ib_pd *pd, int max_page_list_len);
977
978 struct ib_fast_reg_page_list *qib_alloc_fast_reg_page_list(
979 struct ib_device *ibdev, int page_list_len);
980
981 void qib_free_fast_reg_page_list(struct ib_fast_reg_page_list *pl);
982
983 int qib_fast_reg_mr(struct qib_qp *qp, struct ib_send_wr *wr);
984
985 struct ib_fmr *qib_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
986 struct ib_fmr_attr *fmr_attr);
987
988 int qib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
989 int list_len, u64 iova);
990
991 int qib_unmap_fmr(struct list_head *fmr_list);
992
993 int qib_dealloc_fmr(struct ib_fmr *ibfmr);
994
995 void qib_release_mmap_info(struct kref *ref);
996
997 struct qib_mmap_info *qib_create_mmap_info(struct qib_ibdev *dev, u32 size,
998 struct ib_ucontext *context,
999 void *obj);
1000
1001 void qib_update_mmap_info(struct qib_ibdev *dev, struct qib_mmap_info *ip,
1002 u32 size, void *obj);
1003
1004 int qib_mmap(struct ib_ucontext *context, struct vm_area_struct *vma);
1005
1006 int qib_get_rwqe(struct qib_qp *qp, int wr_id_only);
1007
1008 void qib_migrate_qp(struct qib_qp *qp);
1009
1010 int qib_ruc_check_hdr(struct qib_ibport *ibp, struct qib_ib_header *hdr,
1011 int has_grh, struct qib_qp *qp, u32 bth0);
1012
1013 u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr,
1014 struct ib_global_route *grh, u32 hwords, u32 nwords);
1015
1016 void qib_make_ruc_header(struct qib_qp *qp, struct qib_other_headers *ohdr,
1017 u32 bth0, u32 bth2);
1018
1019 void qib_do_send(struct work_struct *work);
1020
1021 void qib_send_complete(struct qib_qp *qp, struct qib_swqe *wqe,
1022 enum ib_wc_status status);
1023
1024 void qib_send_rc_ack(struct qib_qp *qp);
1025
1026 int qib_make_rc_req(struct qib_qp *qp);
1027
1028 int qib_make_uc_req(struct qib_qp *qp);
1029
1030 int qib_make_ud_req(struct qib_qp *qp);
1031
1032 int qib_register_ib_device(struct qib_devdata *);
1033
1034 void qib_unregister_ib_device(struct qib_devdata *);
1035
1036 void qib_ib_rcv(struct qib_ctxtdata *, void *, void *, u32);
1037
1038 void qib_ib_piobufavail(struct qib_devdata *);
1039
1040 unsigned qib_get_npkeys(struct qib_devdata *);
1041
1042 unsigned qib_get_pkey(struct qib_ibport *, unsigned);
1043
1044 extern const enum ib_wc_opcode ib_qib_wc_opcode[];
1045
1046 /*
1047 * Below HCA-independent IB PhysPortState values, returned
1048 * by the f_ibphys_portstate() routine.
1049 */
1050 #define IB_PHYSPORTSTATE_SLEEP 1
1051 #define IB_PHYSPORTSTATE_POLL 2
1052 #define IB_PHYSPORTSTATE_DISABLED 3
1053 #define IB_PHYSPORTSTATE_CFG_TRAIN 4
1054 #define IB_PHYSPORTSTATE_LINKUP 5
1055 #define IB_PHYSPORTSTATE_LINK_ERR_RECOVER 6
1056 #define IB_PHYSPORTSTATE_CFG_DEBOUNCE 8
1057 #define IB_PHYSPORTSTATE_CFG_IDLE 0xB
1058 #define IB_PHYSPORTSTATE_RECOVERY_RETRAIN 0xC
1059 #define IB_PHYSPORTSTATE_RECOVERY_WAITRMT 0xE
1060 #define IB_PHYSPORTSTATE_RECOVERY_IDLE 0xF
1061 #define IB_PHYSPORTSTATE_CFG_ENH 0x10
1062 #define IB_PHYSPORTSTATE_CFG_WAIT_ENH 0x13
1063
1064 extern const int ib_qib_state_ops[];
1065
1066 extern __be64 ib_qib_sys_image_guid; /* in network order */
1067
1068 extern unsigned int ib_qib_lkey_table_size;
1069
1070 extern unsigned int ib_qib_max_cqes;
1071
1072 extern unsigned int ib_qib_max_cqs;
1073
1074 extern unsigned int ib_qib_max_qp_wrs;
1075
1076 extern unsigned int ib_qib_max_qps;
1077
1078 extern unsigned int ib_qib_max_sges;
1079
1080 extern unsigned int ib_qib_max_mcast_grps;
1081
1082 extern unsigned int ib_qib_max_mcast_qp_attached;
1083
1084 extern unsigned int ib_qib_max_srqs;
1085
1086 extern unsigned int ib_qib_max_srq_sges;
1087
1088 extern unsigned int ib_qib_max_srq_wrs;
1089
1090 extern const u32 ib_qib_rnr_table[];
1091
1092 extern struct ib_dma_mapping_ops qib_dma_mapping_ops;
1093
1094 #endif /* QIB_VERBS_H */
1095