1 /*
2  * Copyright (c) 2012 Intel Corporation.  All rights reserved.
3  * Copyright (c) 2006 - 2012 QLogic Corporation. 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 <linux/completion.h>
45 #include <rdma/ib_pack.h>
46 #include <rdma/ib_user_verbs.h>
47 
48 struct qib_ctxtdata;
49 struct qib_pportdata;
50 struct qib_devdata;
51 struct qib_verbs_txreq;
52 
53 #define QIB_MAX_RDMA_ATOMIC     16
54 #define QIB_GUIDS_PER_PORT	5
55 
56 #define QPN_MAX                 (1 << 24)
57 #define QPNMAP_ENTRIES          (QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
58 
59 /*
60  * Increment this value if any changes that break userspace ABI
61  * compatibility are made.
62  */
63 #define QIB_UVERBS_ABI_VERSION       2
64 
65 /*
66  * Define an ib_cq_notify value that is not valid so we know when CQ
67  * notifications are armed.
68  */
69 #define IB_CQ_NONE      (IB_CQ_NEXT_COMP + 1)
70 
71 #define IB_SEQ_NAK	(3 << 29)
72 
73 /* AETH NAK opcode values */
74 #define IB_RNR_NAK                      0x20
75 #define IB_NAK_PSN_ERROR                0x60
76 #define IB_NAK_INVALID_REQUEST          0x61
77 #define IB_NAK_REMOTE_ACCESS_ERROR      0x62
78 #define IB_NAK_REMOTE_OPERATIONAL_ERROR 0x63
79 #define IB_NAK_INVALID_RD_REQUEST       0x64
80 
81 /* Flags for checking QP state (see ib_qib_state_ops[]) */
82 #define QIB_POST_SEND_OK                0x01
83 #define QIB_POST_RECV_OK                0x02
84 #define QIB_PROCESS_RECV_OK             0x04
85 #define QIB_PROCESS_SEND_OK             0x08
86 #define QIB_PROCESS_NEXT_SEND_OK        0x10
87 #define QIB_FLUSH_SEND			0x20
88 #define QIB_FLUSH_RECV			0x40
89 #define QIB_PROCESS_OR_FLUSH_SEND \
90 	(QIB_PROCESS_SEND_OK | QIB_FLUSH_SEND)
91 
92 /* IB Performance Manager status values */
93 #define IB_PMA_SAMPLE_STATUS_DONE       0x00
94 #define IB_PMA_SAMPLE_STATUS_STARTED    0x01
95 #define IB_PMA_SAMPLE_STATUS_RUNNING    0x02
96 
97 /* Mandatory IB performance counter select values. */
98 #define IB_PMA_PORT_XMIT_DATA   cpu_to_be16(0x0001)
99 #define IB_PMA_PORT_RCV_DATA    cpu_to_be16(0x0002)
100 #define IB_PMA_PORT_XMIT_PKTS   cpu_to_be16(0x0003)
101 #define IB_PMA_PORT_RCV_PKTS    cpu_to_be16(0x0004)
102 #define IB_PMA_PORT_XMIT_WAIT   cpu_to_be16(0x0005)
103 
104 #define QIB_VENDOR_IPG		cpu_to_be16(0xFFA0)
105 
106 #define IB_BTH_REQ_ACK		(1 << 31)
107 #define IB_BTH_SOLICITED	(1 << 23)
108 #define IB_BTH_MIG_REQ		(1 << 22)
109 
110 /* XXX Should be defined in ib_verbs.h enum ib_port_cap_flags */
111 #define IB_PORT_OTHER_LOCAL_CHANGES_SUP (1 << 26)
112 
113 #define IB_GRH_VERSION		6
114 #define IB_GRH_VERSION_MASK	0xF
115 #define IB_GRH_VERSION_SHIFT	28
116 #define IB_GRH_TCLASS_MASK	0xFF
117 #define IB_GRH_TCLASS_SHIFT	20
118 #define IB_GRH_FLOW_MASK	0xFFFFF
119 #define IB_GRH_FLOW_SHIFT	0
120 #define IB_GRH_NEXT_HDR		0x1B
121 
122 #define IB_DEFAULT_GID_PREFIX	cpu_to_be64(0xfe80000000000000ULL)
123 
124 /* Values for set/get portinfo VLCap OperationalVLs */
125 #define IB_VL_VL0       1
126 #define IB_VL_VL0_1     2
127 #define IB_VL_VL0_3     3
128 #define IB_VL_VL0_7     4
129 #define IB_VL_VL0_14    5
130 
131 static inline int qib_num_vls(int vls)
132 {
133 	switch (vls) {
134 	default:
135 	case IB_VL_VL0:
136 		return 1;
137 	case IB_VL_VL0_1:
138 		return 2;
139 	case IB_VL_VL0_3:
140 		return 4;
141 	case IB_VL_VL0_7:
142 		return 8;
143 	case IB_VL_VL0_14:
144 		return 15;
145 	}
146 }
147 
148 struct ib_reth {
149 	__be64 vaddr;
150 	__be32 rkey;
151 	__be32 length;
152 } __attribute__ ((packed));
153 
154 struct ib_atomic_eth {
155 	__be32 vaddr[2];        /* unaligned so access as 2 32-bit words */
156 	__be32 rkey;
157 	__be64 swap_data;
158 	__be64 compare_data;
159 } __attribute__ ((packed));
160 
161 struct qib_other_headers {
162 	__be32 bth[3];
163 	union {
164 		struct {
165 			__be32 deth[2];
166 			__be32 imm_data;
167 		} ud;
168 		struct {
169 			struct ib_reth reth;
170 			__be32 imm_data;
171 		} rc;
172 		struct {
173 			__be32 aeth;
174 			__be32 atomic_ack_eth[2];
175 		} at;
176 		__be32 imm_data;
177 		__be32 aeth;
178 		struct ib_atomic_eth atomic_eth;
179 	} u;
180 } __attribute__ ((packed));
181 
182 /*
183  * Note that UD packets with a GRH header are 8+40+12+8 = 68 bytes
184  * long (72 w/ imm_data).  Only the first 56 bytes of the IB header
185  * will be in the eager header buffer.  The remaining 12 or 16 bytes
186  * are in the data buffer.
187  */
188 struct qib_ib_header {
189 	__be16 lrh[4];
190 	union {
191 		struct {
192 			struct ib_grh grh;
193 			struct qib_other_headers oth;
194 		} l;
195 		struct qib_other_headers oth;
196 	} u;
197 } __attribute__ ((packed));
198 
199 struct qib_pio_header {
200 	__le32 pbc[2];
201 	struct qib_ib_header hdr;
202 } __attribute__ ((packed));
203 
204 /*
205  * There is one struct qib_mcast for each multicast GID.
206  * All attached QPs are then stored as a list of
207  * struct qib_mcast_qp.
208  */
209 struct qib_mcast_qp {
210 	struct list_head list;
211 	struct qib_qp *qp;
212 };
213 
214 struct qib_mcast {
215 	struct rb_node rb_node;
216 	union ib_gid mgid;
217 	struct list_head qp_list;
218 	wait_queue_head_t wait;
219 	atomic_t refcount;
220 	int n_attached;
221 };
222 
223 /* Protection domain */
224 struct qib_pd {
225 	struct ib_pd ibpd;
226 	int user;               /* non-zero if created from user space */
227 };
228 
229 /* Address Handle */
230 struct qib_ah {
231 	struct ib_ah ibah;
232 	struct ib_ah_attr attr;
233 	atomic_t refcount;
234 };
235 
236 /*
237  * This structure is used by qib_mmap() to validate an offset
238  * when an mmap() request is made.  The vm_area_struct then uses
239  * this as its vm_private_data.
240  */
241 struct qib_mmap_info {
242 	struct list_head pending_mmaps;
243 	struct ib_ucontext *context;
244 	void *obj;
245 	__u64 offset;
246 	struct kref ref;
247 	unsigned size;
248 };
249 
250 /*
251  * This structure is used to contain the head pointer, tail pointer,
252  * and completion queue entries as a single memory allocation so
253  * it can be mmap'ed into user space.
254  */
255 struct qib_cq_wc {
256 	u32 head;               /* index of next entry to fill */
257 	u32 tail;               /* index of next ib_poll_cq() entry */
258 	union {
259 		/* these are actually size ibcq.cqe + 1 */
260 		struct ib_uverbs_wc uqueue[0];
261 		struct ib_wc kqueue[0];
262 	};
263 };
264 
265 /*
266  * The completion queue structure.
267  */
268 struct qib_cq {
269 	struct ib_cq ibcq;
270 	struct work_struct comptask;
271 	spinlock_t lock; /* protect changes in this struct */
272 	u8 notify;
273 	u8 triggered;
274 	struct qib_cq_wc *queue;
275 	struct qib_mmap_info *ip;
276 };
277 
278 /*
279  * A segment is a linear region of low physical memory.
280  * XXX Maybe we should use phys addr here and kmap()/kunmap().
281  * Used by the verbs layer.
282  */
283 struct qib_seg {
284 	void *vaddr;
285 	size_t length;
286 };
287 
288 /* The number of qib_segs that fit in a page. */
289 #define QIB_SEGSZ     (PAGE_SIZE / sizeof(struct qib_seg))
290 
291 struct qib_segarray {
292 	struct qib_seg segs[QIB_SEGSZ];
293 };
294 
295 struct qib_mregion {
296 	struct ib_pd *pd;       /* shares refcnt of ibmr.pd */
297 	u64 user_base;          /* User's address for this region */
298 	u64 iova;               /* IB start address of this region */
299 	size_t length;
300 	u32 lkey;
301 	u32 offset;             /* offset (bytes) to start of region */
302 	int access_flags;
303 	u32 max_segs;           /* number of qib_segs in all the arrays */
304 	u32 mapsz;              /* size of the map array */
305 	u8  page_shift;         /* 0 - non unform/non powerof2 sizes */
306 	u8  lkey_published;     /* in global table */
307 	struct completion comp; /* complete when refcount goes to zero */
308 	struct rcu_head list;
309 	atomic_t refcount;
310 	struct qib_segarray *map[0];    /* the segments */
311 };
312 
313 /*
314  * These keep track of the copy progress within a memory region.
315  * Used by the verbs layer.
316  */
317 struct qib_sge {
318 	struct qib_mregion *mr;
319 	void *vaddr;            /* kernel virtual address of segment */
320 	u32 sge_length;         /* length of the SGE */
321 	u32 length;             /* remaining length of the segment */
322 	u16 m;                  /* current index: mr->map[m] */
323 	u16 n;                  /* current index: mr->map[m]->segs[n] */
324 };
325 
326 /* Memory region */
327 struct qib_mr {
328 	struct ib_mr ibmr;
329 	struct ib_umem *umem;
330 	struct qib_mregion mr;  /* must be last */
331 };
332 
333 /*
334  * Send work request queue entry.
335  * The size of the sg_list is determined when the QP is created and stored
336  * in qp->s_max_sge.
337  */
338 struct qib_swqe {
339 	struct ib_send_wr wr;   /* don't use wr.sg_list */
340 	u32 psn;                /* first packet sequence number */
341 	u32 lpsn;               /* last packet sequence number */
342 	u32 ssn;                /* send sequence number */
343 	u32 length;             /* total length of data in sg_list */
344 	struct qib_sge sg_list[0];
345 };
346 
347 /*
348  * Receive work request queue entry.
349  * The size of the sg_list is determined when the QP (or SRQ) is created
350  * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
351  */
352 struct qib_rwqe {
353 	u64 wr_id;
354 	u8 num_sge;
355 	struct ib_sge sg_list[0];
356 };
357 
358 /*
359  * This structure is used to contain the head pointer, tail pointer,
360  * and receive work queue entries as a single memory allocation so
361  * it can be mmap'ed into user space.
362  * Note that the wq array elements are variable size so you can't
363  * just index into the array to get the N'th element;
364  * use get_rwqe_ptr() instead.
365  */
366 struct qib_rwq {
367 	u32 head;               /* new work requests posted to the head */
368 	u32 tail;               /* receives pull requests from here. */
369 	struct qib_rwqe wq[0];
370 };
371 
372 struct qib_rq {
373 	struct qib_rwq *wq;
374 	u32 size;               /* size of RWQE array */
375 	u8 max_sge;
376 	spinlock_t lock /* protect changes in this struct */
377 		____cacheline_aligned_in_smp;
378 };
379 
380 struct qib_srq {
381 	struct ib_srq ibsrq;
382 	struct qib_rq rq;
383 	struct qib_mmap_info *ip;
384 	/* send signal when number of RWQEs < limit */
385 	u32 limit;
386 };
387 
388 struct qib_sge_state {
389 	struct qib_sge *sg_list;      /* next SGE to be used if any */
390 	struct qib_sge sge;   /* progress state for the current SGE */
391 	u32 total_len;
392 	u8 num_sge;
393 };
394 
395 /*
396  * This structure holds the information that the send tasklet needs
397  * to send a RDMA read response or atomic operation.
398  */
399 struct qib_ack_entry {
400 	u8 opcode;
401 	u8 sent;
402 	u32 psn;
403 	u32 lpsn;
404 	union {
405 		struct qib_sge rdma_sge;
406 		u64 atomic_data;
407 	};
408 };
409 
410 /*
411  * Variables prefixed with s_ are for the requester (sender).
412  * Variables prefixed with r_ are for the responder (receiver).
413  * Variables prefixed with ack_ are for responder replies.
414  *
415  * Common variables are protected by both r_rq.lock and s_lock in that order
416  * which only happens in modify_qp() or changing the QP 'state'.
417  */
418 struct qib_qp {
419 	struct ib_qp ibqp;
420 	/* read mostly fields above and below */
421 	struct ib_ah_attr remote_ah_attr;
422 	struct ib_ah_attr alt_ah_attr;
423 	struct qib_qp __rcu *next;            /* link list for QPN hash table */
424 	struct qib_swqe *s_wq;  /* send work queue */
425 	struct qib_mmap_info *ip;
426 	struct qib_ib_header *s_hdr;     /* next packet header to send */
427 	unsigned long timeout_jiffies;  /* computed from timeout */
428 
429 	enum ib_mtu path_mtu;
430 	u32 remote_qpn;
431 	u32 pmtu;		/* decoded from path_mtu */
432 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
433 	u32 s_size;             /* send work queue size */
434 	u32 s_rnr_timeout;      /* number of milliseconds for RNR timeout */
435 
436 	u8 state;               /* QP state */
437 	u8 qp_access_flags;
438 	u8 alt_timeout;         /* Alternate path timeout for this QP */
439 	u8 timeout;             /* Timeout for this QP */
440 	u8 s_srate;
441 	u8 s_mig_state;
442 	u8 port_num;
443 	u8 s_pkey_index;        /* PKEY index to use */
444 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
445 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
446 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
447 	u8 s_retry_cnt;         /* number of times to retry */
448 	u8 s_rnr_retry_cnt;
449 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
450 	u8 s_max_sge;           /* size of s_wq->sg_list */
451 	u8 s_draining;
452 
453 	/* start of read/write fields */
454 
455 	atomic_t refcount ____cacheline_aligned_in_smp;
456 	wait_queue_head_t wait;
457 
458 
459 	struct qib_ack_entry s_ack_queue[QIB_MAX_RDMA_ATOMIC + 1]
460 		____cacheline_aligned_in_smp;
461 	struct qib_sge_state s_rdma_read_sge;
462 
463 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
464 	unsigned long r_aflags;
465 	u64 r_wr_id;            /* ID for current receive WQE */
466 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
467 	u32 r_len;              /* total length of r_sge */
468 	u32 r_rcv_len;          /* receive data len processed */
469 	u32 r_psn;              /* expected rcv packet sequence number */
470 	u32 r_msn;              /* message sequence number */
471 
472 	u8 r_state;             /* opcode of last packet received */
473 	u8 r_flags;
474 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
475 
476 	struct list_head rspwait;       /* link for waititing to respond */
477 
478 	struct qib_sge_state r_sge;     /* current receive data */
479 	struct qib_rq r_rq;             /* receive work queue */
480 
481 	spinlock_t s_lock ____cacheline_aligned_in_smp;
482 	struct qib_sge_state *s_cur_sge;
483 	u32 s_flags;
484 	struct qib_verbs_txreq *s_tx;
485 	struct qib_swqe *s_wqe;
486 	struct qib_sge_state s_sge;     /* current send request data */
487 	struct qib_mregion *s_rdma_mr;
488 	atomic_t s_dma_busy;
489 	u32 s_cur_size;         /* size of send packet in bytes */
490 	u32 s_len;              /* total length of s_sge */
491 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
492 	u32 s_next_psn;         /* PSN for next request */
493 	u32 s_last_psn;         /* last response PSN processed */
494 	u32 s_sending_psn;      /* lowest PSN that is being sent */
495 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
496 	u32 s_psn;              /* current packet sequence number */
497 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
498 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
499 	u32 s_head;             /* new entries added here */
500 	u32 s_tail;             /* next entry to process */
501 	u32 s_cur;              /* current work queue entry */
502 	u32 s_acked;            /* last un-ACK'ed entry */
503 	u32 s_last;             /* last completed entry */
504 	u32 s_ssn;              /* SSN of tail entry */
505 	u32 s_lsn;              /* limit sequence number (credit) */
506 	u16 s_hdrwords;         /* size of s_hdr in 32 bit words */
507 	u16 s_rdma_ack_cnt;
508 	u8 s_state;             /* opcode of last packet sent */
509 	u8 s_ack_state;         /* opcode of packet to ACK */
510 	u8 s_nak_state;         /* non-zero if NAK is pending */
511 	u8 r_nak_state;         /* non-zero if NAK is pending */
512 	u8 s_retry;             /* requester retry counter */
513 	u8 s_rnr_retry;         /* requester RNR retry counter */
514 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
515 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
516 
517 	struct qib_sge_state s_ack_rdma_sge;
518 	struct timer_list s_timer;
519 	struct list_head iowait;        /* link for wait PIO buf */
520 
521 	struct work_struct s_work;
522 
523 	wait_queue_head_t wait_dma;
524 
525 	struct qib_sge r_sg_list[0] /* verified SGEs */
526 		____cacheline_aligned_in_smp;
527 };
528 
529 /*
530  * Atomic bit definitions for r_aflags.
531  */
532 #define QIB_R_WRID_VALID        0
533 #define QIB_R_REWIND_SGE        1
534 
535 /*
536  * Bit definitions for r_flags.
537  */
538 #define QIB_R_REUSE_SGE 0x01
539 #define QIB_R_RDMAR_SEQ 0x02
540 #define QIB_R_RSP_NAK   0x04
541 #define QIB_R_RSP_SEND  0x08
542 #define QIB_R_COMM_EST  0x10
543 
544 /*
545  * Bit definitions for s_flags.
546  *
547  * QIB_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
548  * QIB_S_BUSY - send tasklet is processing the QP
549  * QIB_S_TIMER - the RC retry timer is active
550  * QIB_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
551  * QIB_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
552  *                         before processing the next SWQE
553  * QIB_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
554  *                         before processing the next SWQE
555  * QIB_S_WAIT_RNR - waiting for RNR timeout
556  * QIB_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
557  * QIB_S_WAIT_DMA - waiting for send DMA queue to drain before generating
558  *                  next send completion entry not via send DMA
559  * QIB_S_WAIT_PIO - waiting for a send buffer to be available
560  * QIB_S_WAIT_TX - waiting for a struct qib_verbs_txreq to be available
561  * QIB_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
562  * QIB_S_WAIT_KMEM - waiting for kernel memory to be available
563  * QIB_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
564  * QIB_S_WAIT_ACK - waiting for an ACK packet before sending more requests
565  * QIB_S_SEND_ONE - send one packet, request ACK, then wait for ACK
566  */
567 #define QIB_S_SIGNAL_REQ_WR	0x0001
568 #define QIB_S_BUSY		0x0002
569 #define QIB_S_TIMER		0x0004
570 #define QIB_S_RESP_PENDING	0x0008
571 #define QIB_S_ACK_PENDING	0x0010
572 #define QIB_S_WAIT_FENCE	0x0020
573 #define QIB_S_WAIT_RDMAR	0x0040
574 #define QIB_S_WAIT_RNR		0x0080
575 #define QIB_S_WAIT_SSN_CREDIT	0x0100
576 #define QIB_S_WAIT_DMA		0x0200
577 #define QIB_S_WAIT_PIO		0x0400
578 #define QIB_S_WAIT_TX		0x0800
579 #define QIB_S_WAIT_DMA_DESC	0x1000
580 #define QIB_S_WAIT_KMEM		0x2000
581 #define QIB_S_WAIT_PSN		0x4000
582 #define QIB_S_WAIT_ACK		0x8000
583 #define QIB_S_SEND_ONE		0x10000
584 #define QIB_S_UNLIMITED_CREDIT	0x20000
585 
586 /*
587  * Wait flags that would prevent any packet type from being sent.
588  */
589 #define QIB_S_ANY_WAIT_IO (QIB_S_WAIT_PIO | QIB_S_WAIT_TX | \
590 	QIB_S_WAIT_DMA_DESC | QIB_S_WAIT_KMEM)
591 
592 /*
593  * Wait flags that would prevent send work requests from making progress.
594  */
595 #define QIB_S_ANY_WAIT_SEND (QIB_S_WAIT_FENCE | QIB_S_WAIT_RDMAR | \
596 	QIB_S_WAIT_RNR | QIB_S_WAIT_SSN_CREDIT | QIB_S_WAIT_DMA | \
597 	QIB_S_WAIT_PSN | QIB_S_WAIT_ACK)
598 
599 #define QIB_S_ANY_WAIT (QIB_S_ANY_WAIT_IO | QIB_S_ANY_WAIT_SEND)
600 
601 #define QIB_PSN_CREDIT  16
602 
603 /*
604  * Since struct qib_swqe is not a fixed size, we can't simply index into
605  * struct qib_qp.s_wq.  This function does the array index computation.
606  */
607 static inline struct qib_swqe *get_swqe_ptr(struct qib_qp *qp,
608 					      unsigned n)
609 {
610 	return (struct qib_swqe *)((char *)qp->s_wq +
611 				     (sizeof(struct qib_swqe) +
612 				      qp->s_max_sge *
613 				      sizeof(struct qib_sge)) * n);
614 }
615 
616 /*
617  * Since struct qib_rwqe is not a fixed size, we can't simply index into
618  * struct qib_rwq.wq.  This function does the array index computation.
619  */
620 static inline struct qib_rwqe *get_rwqe_ptr(struct qib_rq *rq, unsigned n)
621 {
622 	return (struct qib_rwqe *)
623 		((char *) rq->wq->wq +
624 		 (sizeof(struct qib_rwqe) +
625 		  rq->max_sge * sizeof(struct ib_sge)) * n);
626 }
627 
628 /*
629  * QPN-map pages start out as NULL, they get allocated upon
630  * first use and are never deallocated. This way,
631  * large bitmaps are not allocated unless large numbers of QPs are used.
632  */
633 struct qpn_map {
634 	void *page;
635 };
636 
637 struct qib_qpn_table {
638 	spinlock_t lock; /* protect changes in this struct */
639 	unsigned flags;         /* flags for QP0/1 allocated for each port */
640 	u32 last;               /* last QP number allocated */
641 	u32 nmaps;              /* size of the map table */
642 	u16 limit;
643 	u16 mask;
644 	/* bit map of free QP numbers other than 0/1 */
645 	struct qpn_map map[QPNMAP_ENTRIES];
646 };
647 
648 struct qib_lkey_table {
649 	spinlock_t lock; /* protect changes in this struct */
650 	u32 next;               /* next unused index (speeds search) */
651 	u32 gen;                /* generation count */
652 	u32 max;                /* size of the table */
653 	struct qib_mregion __rcu **table;
654 };
655 
656 struct qib_opcode_stats {
657 	u64 n_packets;          /* number of packets */
658 	u64 n_bytes;            /* total number of bytes */
659 };
660 
661 struct qib_ibport {
662 	struct qib_qp __rcu *qp0;
663 	struct qib_qp __rcu *qp1;
664 	struct ib_mad_agent *send_agent;	/* agent for SMI (traps) */
665 	struct qib_ah *sm_ah;
666 	struct qib_ah *smi_ah;
667 	struct rb_root mcast_tree;
668 	spinlock_t lock;		/* protect changes in this struct */
669 
670 	/* non-zero when timer is set */
671 	unsigned long mkey_lease_timeout;
672 	unsigned long trap_timeout;
673 	__be64 gid_prefix;      /* in network order */
674 	__be64 mkey;
675 	__be64 guids[QIB_GUIDS_PER_PORT	- 1];	/* writable GUIDs */
676 	u64 tid;		/* TID for traps */
677 	u64 n_unicast_xmit;     /* total unicast packets sent */
678 	u64 n_unicast_rcv;      /* total unicast packets received */
679 	u64 n_multicast_xmit;   /* total multicast packets sent */
680 	u64 n_multicast_rcv;    /* total multicast packets received */
681 	u64 z_symbol_error_counter;             /* starting count for PMA */
682 	u64 z_link_error_recovery_counter;      /* starting count for PMA */
683 	u64 z_link_downed_counter;              /* starting count for PMA */
684 	u64 z_port_rcv_errors;                  /* starting count for PMA */
685 	u64 z_port_rcv_remphys_errors;          /* starting count for PMA */
686 	u64 z_port_xmit_discards;               /* starting count for PMA */
687 	u64 z_port_xmit_data;                   /* starting count for PMA */
688 	u64 z_port_rcv_data;                    /* starting count for PMA */
689 	u64 z_port_xmit_packets;                /* starting count for PMA */
690 	u64 z_port_rcv_packets;                 /* starting count for PMA */
691 	u32 z_local_link_integrity_errors;      /* starting count for PMA */
692 	u32 z_excessive_buffer_overrun_errors;  /* starting count for PMA */
693 	u32 z_vl15_dropped;                     /* starting count for PMA */
694 	u32 n_rc_resends;
695 	u32 n_rc_acks;
696 	u32 n_rc_qacks;
697 	u32 n_rc_delayed_comp;
698 	u32 n_seq_naks;
699 	u32 n_rdma_seq;
700 	u32 n_rnr_naks;
701 	u32 n_other_naks;
702 	u32 n_loop_pkts;
703 	u32 n_pkt_drops;
704 	u32 n_vl15_dropped;
705 	u32 n_rc_timeouts;
706 	u32 n_dmawait;
707 	u32 n_unaligned;
708 	u32 n_rc_dupreq;
709 	u32 n_rc_seqnak;
710 	u32 port_cap_flags;
711 	u32 pma_sample_start;
712 	u32 pma_sample_interval;
713 	__be16 pma_counter_select[5];
714 	u16 pma_tag;
715 	u16 pkey_violations;
716 	u16 qkey_violations;
717 	u16 mkey_violations;
718 	u16 mkey_lease_period;
719 	u16 sm_lid;
720 	u16 repress_traps;
721 	u8 sm_sl;
722 	u8 mkeyprot;
723 	u8 subnet_timeout;
724 	u8 vl_high_limit;
725 	u8 sl_to_vl[16];
726 
727 	struct qib_opcode_stats opstats[128];
728 };
729 
730 
731 struct qib_ibdev {
732 	struct ib_device ibdev;
733 	struct list_head pending_mmaps;
734 	spinlock_t mmap_offset_lock; /* protect mmap_offset */
735 	u32 mmap_offset;
736 	struct qib_mregion __rcu *dma_mr;
737 
738 	/* QP numbers are shared by all IB ports */
739 	struct qib_qpn_table qpn_table;
740 	struct qib_lkey_table lk_table;
741 	struct list_head piowait;       /* list for wait PIO buf */
742 	struct list_head dmawait;	/* list for wait DMA */
743 	struct list_head txwait;        /* list for wait qib_verbs_txreq */
744 	struct list_head memwait;       /* list for wait kernel memory */
745 	struct list_head txreq_free;
746 	struct timer_list mem_timer;
747 	struct qib_qp __rcu **qp_table;
748 	struct qib_pio_header *pio_hdrs;
749 	dma_addr_t pio_hdrs_phys;
750 	/* list of QPs waiting for RNR timer */
751 	spinlock_t pending_lock; /* protect wait lists, PMA counters, etc. */
752 	u32 qp_table_size; /* size of the hash table */
753 	u32 qp_rnd; /* random bytes for hash */
754 	spinlock_t qpt_lock;
755 
756 	u32 n_piowait;
757 	u32 n_txwait;
758 
759 	u32 n_pds_allocated;    /* number of PDs allocated for device */
760 	spinlock_t n_pds_lock;
761 	u32 n_ahs_allocated;    /* number of AHs allocated for device */
762 	spinlock_t n_ahs_lock;
763 	u32 n_cqs_allocated;    /* number of CQs allocated for device */
764 	spinlock_t n_cqs_lock;
765 	u32 n_qps_allocated;    /* number of QPs allocated for device */
766 	spinlock_t n_qps_lock;
767 	u32 n_srqs_allocated;   /* number of SRQs allocated for device */
768 	spinlock_t n_srqs_lock;
769 	u32 n_mcast_grps_allocated; /* number of mcast groups allocated */
770 	spinlock_t n_mcast_grps_lock;
771 };
772 
773 struct qib_verbs_counters {
774 	u64 symbol_error_counter;
775 	u64 link_error_recovery_counter;
776 	u64 link_downed_counter;
777 	u64 port_rcv_errors;
778 	u64 port_rcv_remphys_errors;
779 	u64 port_xmit_discards;
780 	u64 port_xmit_data;
781 	u64 port_rcv_data;
782 	u64 port_xmit_packets;
783 	u64 port_rcv_packets;
784 	u32 local_link_integrity_errors;
785 	u32 excessive_buffer_overrun_errors;
786 	u32 vl15_dropped;
787 };
788 
789 static inline struct qib_mr *to_imr(struct ib_mr *ibmr)
790 {
791 	return container_of(ibmr, struct qib_mr, ibmr);
792 }
793 
794 static inline struct qib_pd *to_ipd(struct ib_pd *ibpd)
795 {
796 	return container_of(ibpd, struct qib_pd, ibpd);
797 }
798 
799 static inline struct qib_ah *to_iah(struct ib_ah *ibah)
800 {
801 	return container_of(ibah, struct qib_ah, ibah);
802 }
803 
804 static inline struct qib_cq *to_icq(struct ib_cq *ibcq)
805 {
806 	return container_of(ibcq, struct qib_cq, ibcq);
807 }
808 
809 static inline struct qib_srq *to_isrq(struct ib_srq *ibsrq)
810 {
811 	return container_of(ibsrq, struct qib_srq, ibsrq);
812 }
813 
814 static inline struct qib_qp *to_iqp(struct ib_qp *ibqp)
815 {
816 	return container_of(ibqp, struct qib_qp, ibqp);
817 }
818 
819 static inline struct qib_ibdev *to_idev(struct ib_device *ibdev)
820 {
821 	return container_of(ibdev, struct qib_ibdev, ibdev);
822 }
823 
824 /*
825  * Send if not busy or waiting for I/O and either
826  * a RC response is pending or we can process send work requests.
827  */
828 static inline int qib_send_ok(struct qib_qp *qp)
829 {
830 	return !(qp->s_flags & (QIB_S_BUSY | QIB_S_ANY_WAIT_IO)) &&
831 		(qp->s_hdrwords || (qp->s_flags & QIB_S_RESP_PENDING) ||
832 		 !(qp->s_flags & QIB_S_ANY_WAIT_SEND));
833 }
834 
835 extern struct workqueue_struct *qib_cq_wq;
836 
837 /*
838  * This must be called with s_lock held.
839  */
840 void qib_schedule_send(struct qib_qp *qp);
841 
842 static inline int qib_pkey_ok(u16 pkey1, u16 pkey2)
843 {
844 	u16 p1 = pkey1 & 0x7FFF;
845 	u16 p2 = pkey2 & 0x7FFF;
846 
847 	/*
848 	 * Low 15 bits must be non-zero and match, and
849 	 * one of the two must be a full member.
850 	 */
851 	return p1 && p1 == p2 && ((__s16)pkey1 < 0 || (__s16)pkey2 < 0);
852 }
853 
854 void qib_bad_pqkey(struct qib_ibport *ibp, __be16 trap_num, u32 key, u32 sl,
855 		   u32 qp1, u32 qp2, __be16 lid1, __be16 lid2);
856 void qib_cap_mask_chg(struct qib_ibport *ibp);
857 void qib_sys_guid_chg(struct qib_ibport *ibp);
858 void qib_node_desc_chg(struct qib_ibport *ibp);
859 int qib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
860 		    struct ib_wc *in_wc, struct ib_grh *in_grh,
861 		    struct ib_mad *in_mad, struct ib_mad *out_mad);
862 int qib_create_agents(struct qib_ibdev *dev);
863 void qib_free_agents(struct qib_ibdev *dev);
864 
865 /*
866  * Compare the lower 24 bits of the two values.
867  * Returns an integer <, ==, or > than zero.
868  */
869 static inline int qib_cmp24(u32 a, u32 b)
870 {
871 	return (((int) a) - ((int) b)) << 8;
872 }
873 
874 struct qib_mcast *qib_mcast_find(struct qib_ibport *ibp, union ib_gid *mgid);
875 
876 int qib_snapshot_counters(struct qib_pportdata *ppd, u64 *swords,
877 			  u64 *rwords, u64 *spkts, u64 *rpkts,
878 			  u64 *xmit_wait);
879 
880 int qib_get_counters(struct qib_pportdata *ppd,
881 		     struct qib_verbs_counters *cntrs);
882 
883 int qib_multicast_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
884 
885 int qib_multicast_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid);
886 
887 int qib_mcast_tree_empty(struct qib_ibport *ibp);
888 
889 __be32 qib_compute_aeth(struct qib_qp *qp);
890 
891 struct qib_qp *qib_lookup_qpn(struct qib_ibport *ibp, u32 qpn);
892 
893 struct ib_qp *qib_create_qp(struct ib_pd *ibpd,
894 			    struct ib_qp_init_attr *init_attr,
895 			    struct ib_udata *udata);
896 
897 int qib_destroy_qp(struct ib_qp *ibqp);
898 
899 int qib_error_qp(struct qib_qp *qp, enum ib_wc_status err);
900 
901 int qib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
902 		  int attr_mask, struct ib_udata *udata);
903 
904 int qib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
905 		 int attr_mask, struct ib_qp_init_attr *init_attr);
906 
907 unsigned qib_free_all_qps(struct qib_devdata *dd);
908 
909 void qib_init_qpn_table(struct qib_devdata *dd, struct qib_qpn_table *qpt);
910 
911 void qib_free_qpn_table(struct qib_qpn_table *qpt);
912 
913 void qib_get_credit(struct qib_qp *qp, u32 aeth);
914 
915 unsigned qib_pkt_delay(u32 plen, u8 snd_mult, u8 rcv_mult);
916 
917 void qib_verbs_sdma_desc_avail(struct qib_pportdata *ppd, unsigned avail);
918 
919 void qib_put_txreq(struct qib_verbs_txreq *tx);
920 
921 int qib_verbs_send(struct qib_qp *qp, struct qib_ib_header *hdr,
922 		   u32 hdrwords, struct qib_sge_state *ss, u32 len);
923 
924 void qib_copy_sge(struct qib_sge_state *ss, void *data, u32 length,
925 		  int release);
926 
927 void qib_skip_sge(struct qib_sge_state *ss, u32 length, int release);
928 
929 void qib_uc_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
930 		int has_grh, void *data, u32 tlen, struct qib_qp *qp);
931 
932 void qib_rc_rcv(struct qib_ctxtdata *rcd, struct qib_ib_header *hdr,
933 		int has_grh, void *data, u32 tlen, struct qib_qp *qp);
934 
935 int qib_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr);
936 
937 struct ib_ah *qib_create_qp0_ah(struct qib_ibport *ibp, u16 dlid);
938 
939 void qib_rc_rnr_retry(unsigned long arg);
940 
941 void qib_rc_send_complete(struct qib_qp *qp, struct qib_ib_header *hdr);
942 
943 void qib_rc_error(struct qib_qp *qp, enum ib_wc_status err);
944 
945 int qib_post_ud_send(struct qib_qp *qp, struct ib_send_wr *wr);
946 
947 void qib_ud_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
948 		int has_grh, void *data, u32 tlen, struct qib_qp *qp);
949 
950 int qib_alloc_lkey(struct qib_mregion *mr, int dma_region);
951 
952 void qib_free_lkey(struct qib_mregion *mr);
953 
954 int qib_lkey_ok(struct qib_lkey_table *rkt, struct qib_pd *pd,
955 		struct qib_sge *isge, struct ib_sge *sge, int acc);
956 
957 int qib_rkey_ok(struct qib_qp *qp, struct qib_sge *sge,
958 		u32 len, u64 vaddr, u32 rkey, int acc);
959 
960 int qib_post_srq_receive(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
961 			 struct ib_recv_wr **bad_wr);
962 
963 struct ib_srq *qib_create_srq(struct ib_pd *ibpd,
964 			      struct ib_srq_init_attr *srq_init_attr,
965 			      struct ib_udata *udata);
966 
967 int qib_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
968 		   enum ib_srq_attr_mask attr_mask,
969 		   struct ib_udata *udata);
970 
971 int qib_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr);
972 
973 int qib_destroy_srq(struct ib_srq *ibsrq);
974 
975 void qib_cq_enter(struct qib_cq *cq, struct ib_wc *entry, int sig);
976 
977 int qib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry);
978 
979 struct ib_cq *qib_create_cq(struct ib_device *ibdev, int entries,
980 			    int comp_vector, struct ib_ucontext *context,
981 			    struct ib_udata *udata);
982 
983 int qib_destroy_cq(struct ib_cq *ibcq);
984 
985 int qib_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags);
986 
987 int qib_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata);
988 
989 struct ib_mr *qib_get_dma_mr(struct ib_pd *pd, int acc);
990 
991 struct ib_mr *qib_reg_phys_mr(struct ib_pd *pd,
992 			      struct ib_phys_buf *buffer_list,
993 			      int num_phys_buf, int acc, u64 *iova_start);
994 
995 struct ib_mr *qib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
996 			      u64 virt_addr, int mr_access_flags,
997 			      struct ib_udata *udata);
998 
999 int qib_dereg_mr(struct ib_mr *ibmr);
1000 
1001 struct ib_mr *qib_alloc_fast_reg_mr(struct ib_pd *pd, int max_page_list_len);
1002 
1003 struct ib_fast_reg_page_list *qib_alloc_fast_reg_page_list(
1004 				struct ib_device *ibdev, int page_list_len);
1005 
1006 void qib_free_fast_reg_page_list(struct ib_fast_reg_page_list *pl);
1007 
1008 int qib_fast_reg_mr(struct qib_qp *qp, struct ib_send_wr *wr);
1009 
1010 struct ib_fmr *qib_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1011 			     struct ib_fmr_attr *fmr_attr);
1012 
1013 int qib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
1014 		     int list_len, u64 iova);
1015 
1016 int qib_unmap_fmr(struct list_head *fmr_list);
1017 
1018 int qib_dealloc_fmr(struct ib_fmr *ibfmr);
1019 
1020 static inline void qib_get_mr(struct qib_mregion *mr)
1021 {
1022 	atomic_inc(&mr->refcount);
1023 }
1024 
1025 void mr_rcu_callback(struct rcu_head *list);
1026 
1027 static inline void qib_put_mr(struct qib_mregion *mr)
1028 {
1029 	if (unlikely(atomic_dec_and_test(&mr->refcount)))
1030 		call_rcu(&mr->list, mr_rcu_callback);
1031 }
1032 
1033 static inline void qib_put_ss(struct qib_sge_state *ss)
1034 {
1035 	while (ss->num_sge) {
1036 		qib_put_mr(ss->sge.mr);
1037 		if (--ss->num_sge)
1038 			ss->sge = *ss->sg_list++;
1039 	}
1040 }
1041 
1042 
1043 void qib_release_mmap_info(struct kref *ref);
1044 
1045 struct qib_mmap_info *qib_create_mmap_info(struct qib_ibdev *dev, u32 size,
1046 					   struct ib_ucontext *context,
1047 					   void *obj);
1048 
1049 void qib_update_mmap_info(struct qib_ibdev *dev, struct qib_mmap_info *ip,
1050 			  u32 size, void *obj);
1051 
1052 int qib_mmap(struct ib_ucontext *context, struct vm_area_struct *vma);
1053 
1054 int qib_get_rwqe(struct qib_qp *qp, int wr_id_only);
1055 
1056 void qib_migrate_qp(struct qib_qp *qp);
1057 
1058 int qib_ruc_check_hdr(struct qib_ibport *ibp, struct qib_ib_header *hdr,
1059 		      int has_grh, struct qib_qp *qp, u32 bth0);
1060 
1061 u32 qib_make_grh(struct qib_ibport *ibp, struct ib_grh *hdr,
1062 		 struct ib_global_route *grh, u32 hwords, u32 nwords);
1063 
1064 void qib_make_ruc_header(struct qib_qp *qp, struct qib_other_headers *ohdr,
1065 			 u32 bth0, u32 bth2);
1066 
1067 void qib_do_send(struct work_struct *work);
1068 
1069 void qib_send_complete(struct qib_qp *qp, struct qib_swqe *wqe,
1070 		       enum ib_wc_status status);
1071 
1072 void qib_send_rc_ack(struct qib_qp *qp);
1073 
1074 int qib_make_rc_req(struct qib_qp *qp);
1075 
1076 int qib_make_uc_req(struct qib_qp *qp);
1077 
1078 int qib_make_ud_req(struct qib_qp *qp);
1079 
1080 int qib_register_ib_device(struct qib_devdata *);
1081 
1082 void qib_unregister_ib_device(struct qib_devdata *);
1083 
1084 void qib_ib_rcv(struct qib_ctxtdata *, void *, void *, u32);
1085 
1086 void qib_ib_piobufavail(struct qib_devdata *);
1087 
1088 unsigned qib_get_npkeys(struct qib_devdata *);
1089 
1090 unsigned qib_get_pkey(struct qib_ibport *, unsigned);
1091 
1092 extern const enum ib_wc_opcode ib_qib_wc_opcode[];
1093 
1094 /*
1095  * Below  HCA-independent IB PhysPortState values, returned
1096  * by the f_ibphys_portstate() routine.
1097  */
1098 #define IB_PHYSPORTSTATE_SLEEP 1
1099 #define IB_PHYSPORTSTATE_POLL 2
1100 #define IB_PHYSPORTSTATE_DISABLED 3
1101 #define IB_PHYSPORTSTATE_CFG_TRAIN 4
1102 #define IB_PHYSPORTSTATE_LINKUP 5
1103 #define IB_PHYSPORTSTATE_LINK_ERR_RECOVER 6
1104 #define IB_PHYSPORTSTATE_CFG_DEBOUNCE 8
1105 #define IB_PHYSPORTSTATE_CFG_IDLE 0xB
1106 #define IB_PHYSPORTSTATE_RECOVERY_RETRAIN 0xC
1107 #define IB_PHYSPORTSTATE_RECOVERY_WAITRMT 0xE
1108 #define IB_PHYSPORTSTATE_RECOVERY_IDLE 0xF
1109 #define IB_PHYSPORTSTATE_CFG_ENH 0x10
1110 #define IB_PHYSPORTSTATE_CFG_WAIT_ENH 0x13
1111 
1112 extern const int ib_qib_state_ops[];
1113 
1114 extern __be64 ib_qib_sys_image_guid;    /* in network order */
1115 
1116 extern unsigned int ib_qib_lkey_table_size;
1117 
1118 extern unsigned int ib_qib_max_cqes;
1119 
1120 extern unsigned int ib_qib_max_cqs;
1121 
1122 extern unsigned int ib_qib_max_qp_wrs;
1123 
1124 extern unsigned int ib_qib_max_qps;
1125 
1126 extern unsigned int ib_qib_max_sges;
1127 
1128 extern unsigned int ib_qib_max_mcast_grps;
1129 
1130 extern unsigned int ib_qib_max_mcast_qp_attached;
1131 
1132 extern unsigned int ib_qib_max_srqs;
1133 
1134 extern unsigned int ib_qib_max_srq_sges;
1135 
1136 extern unsigned int ib_qib_max_srq_wrs;
1137 
1138 extern const u32 ib_qib_rnr_table[];
1139 
1140 extern struct ib_dma_mapping_ops qib_dma_mapping_ops;
1141 
1142 #endif                          /* QIB_VERBS_H */
1143