1 | /* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */ |
2 | /* |
3 | * Copyright(c) 2016 - 2020 Intel Corporation. |
4 | */ |
5 | |
6 | #ifndef DEF_RDMAVT_INCQP_H |
7 | #define DEF_RDMAVT_INCQP_H |
8 | |
9 | #include <rdma/rdma_vt.h> |
10 | #include <rdma/ib_pack.h> |
11 | #include <rdma/ib_verbs.h> |
12 | #include <rdma/rdmavt_cq.h> |
13 | #include <rdma/rvt-abi.h> |
14 | /* |
15 | * Atomic bit definitions for r_aflags. |
16 | */ |
17 | #define RVT_R_WRID_VALID 0 |
18 | #define RVT_R_REWIND_SGE 1 |
19 | |
20 | /* |
21 | * Bit definitions for r_flags. |
22 | */ |
23 | #define RVT_R_REUSE_SGE 0x01 |
24 | #define RVT_R_RDMAR_SEQ 0x02 |
25 | #define RVT_R_RSP_NAK 0x04 |
26 | #define RVT_R_RSP_SEND 0x08 |
27 | #define RVT_R_COMM_EST 0x10 |
28 | |
29 | /* |
30 | * If a packet's QP[23:16] bits match this value, then it is |
31 | * a PSM packet and the hardware will expect a KDETH header |
32 | * following the BTH. |
33 | */ |
34 | #define RVT_KDETH_QP_PREFIX 0x80 |
35 | #define RVT_KDETH_QP_SUFFIX 0xffff |
36 | #define RVT_KDETH_QP_PREFIX_MASK 0x00ff0000 |
37 | #define RVT_KDETH_QP_PREFIX_SHIFT 16 |
38 | #define RVT_KDETH_QP_BASE (u32)(RVT_KDETH_QP_PREFIX << \ |
39 | RVT_KDETH_QP_PREFIX_SHIFT) |
40 | #define RVT_KDETH_QP_MAX (u32)(RVT_KDETH_QP_BASE + RVT_KDETH_QP_SUFFIX) |
41 | |
42 | /* |
43 | * If a packet's LNH == BTH and DEST QPN[23:16] in the BTH match this |
44 | * prefix value, then it is an AIP packet with a DETH containing the entropy |
45 | * value in byte 4 following the BTH. |
46 | */ |
47 | #define RVT_AIP_QP_PREFIX 0x81 |
48 | #define RVT_AIP_QP_SUFFIX 0xffff |
49 | #define RVT_AIP_QP_PREFIX_MASK 0x00ff0000 |
50 | #define RVT_AIP_QP_PREFIX_SHIFT 16 |
51 | #define RVT_AIP_QP_BASE (u32)(RVT_AIP_QP_PREFIX << \ |
52 | RVT_AIP_QP_PREFIX_SHIFT) |
53 | #define RVT_AIP_QPN_MAX BIT(RVT_AIP_QP_PREFIX_SHIFT) |
54 | #define RVT_AIP_QP_MAX (u32)(RVT_AIP_QP_BASE + RVT_AIP_QPN_MAX - 1) |
55 | |
56 | /* |
57 | * Bit definitions for s_flags. |
58 | * |
59 | * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled |
60 | * RVT_S_BUSY - send tasklet is processing the QP |
61 | * RVT_S_TIMER - the RC retry timer is active |
62 | * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics |
63 | * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs |
64 | * before processing the next SWQE |
65 | * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete |
66 | * before processing the next SWQE |
67 | * RVT_S_WAIT_RNR - waiting for RNR timeout |
68 | * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE |
69 | * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating |
70 | * next send completion entry not via send DMA |
71 | * RVT_S_WAIT_PIO - waiting for a send buffer to be available |
72 | * RVT_S_WAIT_TX - waiting for a struct verbs_txreq to be available |
73 | * RVT_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available |
74 | * RVT_S_WAIT_KMEM - waiting for kernel memory to be available |
75 | * RVT_S_WAIT_PSN - waiting for a packet to exit the send DMA queue |
76 | * RVT_S_WAIT_ACK - waiting for an ACK packet before sending more requests |
77 | * RVT_S_SEND_ONE - send one packet, request ACK, then wait for ACK |
78 | * RVT_S_ECN - a BECN was queued to the send engine |
79 | * RVT_S_MAX_BIT_MASK - The max bit that can be used by rdmavt |
80 | */ |
81 | #define RVT_S_SIGNAL_REQ_WR 0x0001 |
82 | #define RVT_S_BUSY 0x0002 |
83 | #define RVT_S_TIMER 0x0004 |
84 | #define RVT_S_RESP_PENDING 0x0008 |
85 | #define RVT_S_ACK_PENDING 0x0010 |
86 | #define RVT_S_WAIT_FENCE 0x0020 |
87 | #define RVT_S_WAIT_RDMAR 0x0040 |
88 | #define RVT_S_WAIT_RNR 0x0080 |
89 | #define RVT_S_WAIT_SSN_CREDIT 0x0100 |
90 | #define RVT_S_WAIT_DMA 0x0200 |
91 | #define RVT_S_WAIT_PIO 0x0400 |
92 | #define RVT_S_WAIT_TX 0x0800 |
93 | #define RVT_S_WAIT_DMA_DESC 0x1000 |
94 | #define RVT_S_WAIT_KMEM 0x2000 |
95 | #define RVT_S_WAIT_PSN 0x4000 |
96 | #define RVT_S_WAIT_ACK 0x8000 |
97 | #define RVT_S_SEND_ONE 0x10000 |
98 | #define RVT_S_UNLIMITED_CREDIT 0x20000 |
99 | #define RVT_S_ECN 0x40000 |
100 | #define RVT_S_MAX_BIT_MASK 0x800000 |
101 | |
102 | /* |
103 | * Drivers should use s_flags starting with bit 31 down to the bit next to |
104 | * RVT_S_MAX_BIT_MASK |
105 | */ |
106 | |
107 | /* |
108 | * Wait flags that would prevent any packet type from being sent. |
109 | */ |
110 | #define RVT_S_ANY_WAIT_IO \ |
111 | (RVT_S_WAIT_PIO | RVT_S_WAIT_TX | \ |
112 | RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM) |
113 | |
114 | /* |
115 | * Wait flags that would prevent send work requests from making progress. |
116 | */ |
117 | #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \ |
118 | RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \ |
119 | RVT_S_WAIT_PSN | RVT_S_WAIT_ACK) |
120 | |
121 | #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND) |
122 | |
123 | /* Number of bits to pay attention to in the opcode for checking qp type */ |
124 | #define RVT_OPCODE_QP_MASK 0xE0 |
125 | |
126 | /* Flags for checking QP state (see ib_rvt_state_ops[]) */ |
127 | #define RVT_POST_SEND_OK 0x01 |
128 | #define RVT_POST_RECV_OK 0x02 |
129 | #define RVT_PROCESS_RECV_OK 0x04 |
130 | #define RVT_PROCESS_SEND_OK 0x08 |
131 | #define RVT_PROCESS_NEXT_SEND_OK 0x10 |
132 | #define RVT_FLUSH_SEND 0x20 |
133 | #define RVT_FLUSH_RECV 0x40 |
134 | #define RVT_PROCESS_OR_FLUSH_SEND \ |
135 | (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND) |
136 | #define RVT_SEND_OR_FLUSH_OR_RECV_OK \ |
137 | (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND | RVT_PROCESS_RECV_OK) |
138 | |
139 | /* |
140 | * Internal send flags |
141 | */ |
142 | #define RVT_SEND_RESERVE_USED IB_SEND_RESERVED_START |
143 | #define RVT_SEND_COMPLETION_ONLY (IB_SEND_RESERVED_START << 1) |
144 | |
145 | /** |
146 | * rvt_ud_wr - IB UD work plus AH cache |
147 | * @wr: valid IB work request |
148 | * @attr: pointer to an allocated AH attribute |
149 | * |
150 | * Special case the UD WR so we can keep track of the AH attributes. |
151 | * |
152 | * NOTE: This data structure is stricly ordered wr then attr. I.e the attr |
153 | * MUST come after wr. The ib_ud_wr is sized and copied in rvt_post_one_wr. |
154 | * The copy assumes that wr is first. |
155 | */ |
156 | struct rvt_ud_wr { |
157 | struct ib_ud_wr wr; |
158 | struct rdma_ah_attr *attr; |
159 | }; |
160 | |
161 | /* |
162 | * Send work request queue entry. |
163 | * The size of the sg_list is determined when the QP is created and stored |
164 | * in qp->s_max_sge. |
165 | */ |
166 | struct rvt_swqe { |
167 | union { |
168 | struct ib_send_wr wr; /* don't use wr.sg_list */ |
169 | struct rvt_ud_wr ud_wr; |
170 | struct ib_reg_wr reg_wr; |
171 | struct ib_rdma_wr rdma_wr; |
172 | struct ib_atomic_wr atomic_wr; |
173 | }; |
174 | u32 psn; /* first packet sequence number */ |
175 | u32 lpsn; /* last packet sequence number */ |
176 | u32 ssn; /* send sequence number */ |
177 | u32 length; /* total length of data in sg_list */ |
178 | void *priv; /* driver dependent field */ |
179 | struct rvt_sge sg_list[]; |
180 | }; |
181 | |
182 | /** |
183 | * struct rvt_krwq - kernel struct receive work request |
184 | * @p_lock: lock to protect producer of the kernel buffer |
185 | * @head: index of next entry to fill |
186 | * @c_lock:lock to protect consumer of the kernel buffer |
187 | * @tail: index of next entry to pull |
188 | * @count: count is aproximate of total receive enteries posted |
189 | * @rvt_rwqe: struct of receive work request queue entry |
190 | * |
191 | * This structure is used to contain the head pointer, |
192 | * tail pointer and receive work queue entries for kernel |
193 | * mode user. |
194 | */ |
195 | struct rvt_krwq { |
196 | spinlock_t p_lock; /* protect producer */ |
197 | u32 head; /* new work requests posted to the head */ |
198 | |
199 | /* protect consumer */ |
200 | spinlock_t c_lock ____cacheline_aligned_in_smp; |
201 | u32 tail; /* receives pull requests from here. */ |
202 | u32 count; /* approx count of receive entries posted */ |
203 | struct rvt_rwqe *curr_wq; |
204 | struct rvt_rwqe wq[]; |
205 | }; |
206 | |
207 | /* |
208 | * rvt_get_swqe_ah - Return the pointer to the struct rvt_ah |
209 | * @swqe: valid Send WQE |
210 | * |
211 | */ |
212 | static inline struct rvt_ah *rvt_get_swqe_ah(struct rvt_swqe *swqe) |
213 | { |
214 | return ibah_to_rvtah(ibah: swqe->ud_wr.wr.ah); |
215 | } |
216 | |
217 | /** |
218 | * rvt_get_swqe_ah_attr - Return the cached ah attribute information |
219 | * @swqe: valid Send WQE |
220 | * |
221 | */ |
222 | static inline struct rdma_ah_attr *rvt_get_swqe_ah_attr(struct rvt_swqe *swqe) |
223 | { |
224 | return swqe->ud_wr.attr; |
225 | } |
226 | |
227 | /** |
228 | * rvt_get_swqe_remote_qpn - Access the remote QPN value |
229 | * @swqe: valid Send WQE |
230 | * |
231 | */ |
232 | static inline u32 rvt_get_swqe_remote_qpn(struct rvt_swqe *swqe) |
233 | { |
234 | return swqe->ud_wr.wr.remote_qpn; |
235 | } |
236 | |
237 | /** |
238 | * rvt_get_swqe_remote_qkey - Acces the remote qkey value |
239 | * @swqe: valid Send WQE |
240 | * |
241 | */ |
242 | static inline u32 rvt_get_swqe_remote_qkey(struct rvt_swqe *swqe) |
243 | { |
244 | return swqe->ud_wr.wr.remote_qkey; |
245 | } |
246 | |
247 | /** |
248 | * rvt_get_swqe_pkey_index - Access the pkey index |
249 | * @swqe: valid Send WQE |
250 | * |
251 | */ |
252 | static inline u16 rvt_get_swqe_pkey_index(struct rvt_swqe *swqe) |
253 | { |
254 | return swqe->ud_wr.wr.pkey_index; |
255 | } |
256 | |
257 | struct rvt_rq { |
258 | struct rvt_rwq *wq; |
259 | struct rvt_krwq *kwq; |
260 | u32 size; /* size of RWQE array */ |
261 | u8 max_sge; |
262 | /* protect changes in this struct */ |
263 | spinlock_t lock ____cacheline_aligned_in_smp; |
264 | }; |
265 | |
266 | /** |
267 | * rvt_get_rq_count - count numbers of request work queue entries |
268 | * in circular buffer |
269 | * @rq: data structure for request queue entry |
270 | * @head: head indices of the circular buffer |
271 | * @tail: tail indices of the circular buffer |
272 | * |
273 | * Return - total number of entries in the Receive Queue |
274 | */ |
275 | |
276 | static inline u32 rvt_get_rq_count(struct rvt_rq *rq, u32 head, u32 tail) |
277 | { |
278 | u32 count = head - tail; |
279 | |
280 | if ((s32)count < 0) |
281 | count += rq->size; |
282 | return count; |
283 | } |
284 | |
285 | /* |
286 | * This structure holds the information that the send tasklet needs |
287 | * to send a RDMA read response or atomic operation. |
288 | */ |
289 | struct rvt_ack_entry { |
290 | struct rvt_sge rdma_sge; |
291 | u64 atomic_data; |
292 | u32 psn; |
293 | u32 lpsn; |
294 | u8 opcode; |
295 | u8 sent; |
296 | void *priv; |
297 | }; |
298 | |
299 | #define RC_QP_SCALING_INTERVAL 5 |
300 | |
301 | #define RVT_OPERATION_PRIV 0x00000001 |
302 | #define RVT_OPERATION_ATOMIC 0x00000002 |
303 | #define RVT_OPERATION_ATOMIC_SGE 0x00000004 |
304 | #define RVT_OPERATION_LOCAL 0x00000008 |
305 | #define RVT_OPERATION_USE_RESERVE 0x00000010 |
306 | #define RVT_OPERATION_IGN_RNR_CNT 0x00000020 |
307 | |
308 | #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1) |
309 | |
310 | /** |
311 | * rvt_operation_params - op table entry |
312 | * @length - the length to copy into the swqe entry |
313 | * @qpt_support - a bit mask indicating QP type support |
314 | * @flags - RVT_OPERATION flags (see above) |
315 | * |
316 | * This supports table driven post send so that |
317 | * the driver can have differing an potentially |
318 | * different sets of operations. |
319 | * |
320 | **/ |
321 | |
322 | struct rvt_operation_params { |
323 | size_t length; |
324 | u32 qpt_support; |
325 | u32 flags; |
326 | }; |
327 | |
328 | /* |
329 | * Common variables are protected by both r_rq.lock and s_lock in that order |
330 | * which only happens in modify_qp() or changing the QP 'state'. |
331 | */ |
332 | struct rvt_qp { |
333 | struct ib_qp ibqp; |
334 | void *priv; /* Driver private data */ |
335 | /* read mostly fields above and below */ |
336 | struct rdma_ah_attr remote_ah_attr; |
337 | struct rdma_ah_attr alt_ah_attr; |
338 | struct rvt_qp __rcu *next; /* link list for QPN hash table */ |
339 | struct rvt_swqe *s_wq; /* send work queue */ |
340 | struct rvt_mmap_info *ip; |
341 | |
342 | unsigned long timeout_jiffies; /* computed from timeout */ |
343 | |
344 | int srate_mbps; /* s_srate (below) converted to Mbit/s */ |
345 | pid_t pid; /* pid for user mode QPs */ |
346 | u32 remote_qpn; |
347 | u32 qkey; /* QKEY for this QP (for UD or RD) */ |
348 | u32 s_size; /* send work queue size */ |
349 | |
350 | u16 pmtu; /* decoded from path_mtu */ |
351 | u8 log_pmtu; /* shift for pmtu */ |
352 | u8 state; /* QP state */ |
353 | u8 allowed_ops; /* high order bits of allowed opcodes */ |
354 | u8 qp_access_flags; |
355 | u8 alt_timeout; /* Alternate path timeout for this QP */ |
356 | u8 timeout; /* Timeout for this QP */ |
357 | u8 s_srate; |
358 | u8 s_mig_state; |
359 | u8 port_num; |
360 | u8 s_pkey_index; /* PKEY index to use */ |
361 | u8 s_alt_pkey_index; /* Alternate path PKEY index to use */ |
362 | u8 r_max_rd_atomic; /* max number of RDMA read/atomic to receive */ |
363 | u8 s_max_rd_atomic; /* max number of RDMA read/atomic to send */ |
364 | u8 s_retry_cnt; /* number of times to retry */ |
365 | u8 s_rnr_retry_cnt; |
366 | u8 r_min_rnr_timer; /* retry timeout value for RNR NAKs */ |
367 | u8 s_max_sge; /* size of s_wq->sg_list */ |
368 | u8 s_draining; |
369 | |
370 | /* start of read/write fields */ |
371 | atomic_t refcount ____cacheline_aligned_in_smp; |
372 | wait_queue_head_t wait; |
373 | |
374 | struct rvt_ack_entry *s_ack_queue; |
375 | struct rvt_sge_state s_rdma_read_sge; |
376 | |
377 | spinlock_t r_lock ____cacheline_aligned_in_smp; /* used for APM */ |
378 | u32 r_psn; /* expected rcv packet sequence number */ |
379 | unsigned long r_aflags; |
380 | u64 r_wr_id; /* ID for current receive WQE */ |
381 | u32 r_ack_psn; /* PSN for next ACK or atomic ACK */ |
382 | u32 r_len; /* total length of r_sge */ |
383 | u32 r_rcv_len; /* receive data len processed */ |
384 | u32 r_msn; /* message sequence number */ |
385 | |
386 | u8 r_state; /* opcode of last packet received */ |
387 | u8 r_flags; |
388 | u8 r_head_ack_queue; /* index into s_ack_queue[] */ |
389 | u8 r_adefered; /* defered ack count */ |
390 | |
391 | struct list_head rspwait; /* link for waiting to respond */ |
392 | |
393 | struct rvt_sge_state r_sge; /* current receive data */ |
394 | struct rvt_rq r_rq; /* receive work queue */ |
395 | |
396 | /* post send line */ |
397 | spinlock_t s_hlock ____cacheline_aligned_in_smp; |
398 | u32 s_head; /* new entries added here */ |
399 | u32 s_next_psn; /* PSN for next request */ |
400 | u32 s_avail; /* number of entries avail */ |
401 | u32 s_ssn; /* SSN of tail entry */ |
402 | atomic_t s_reserved_used; /* reserved entries in use */ |
403 | |
404 | spinlock_t s_lock ____cacheline_aligned_in_smp; |
405 | u32 s_flags; |
406 | struct rvt_sge_state *s_cur_sge; |
407 | struct rvt_swqe *s_wqe; |
408 | struct rvt_sge_state s_sge; /* current send request data */ |
409 | struct rvt_mregion *s_rdma_mr; |
410 | u32 s_len; /* total length of s_sge */ |
411 | u32 s_rdma_read_len; /* total length of s_rdma_read_sge */ |
412 | u32 s_last_psn; /* last response PSN processed */ |
413 | u32 s_sending_psn; /* lowest PSN that is being sent */ |
414 | u32 s_sending_hpsn; /* highest PSN that is being sent */ |
415 | u32 s_psn; /* current packet sequence number */ |
416 | u32 s_ack_rdma_psn; /* PSN for sending RDMA read responses */ |
417 | u32 s_ack_psn; /* PSN for acking sends and RDMA writes */ |
418 | u32 s_tail; /* next entry to process */ |
419 | u32 s_cur; /* current work queue entry */ |
420 | u32 s_acked; /* last un-ACK'ed entry */ |
421 | u32 s_last; /* last completed entry */ |
422 | u32 s_lsn; /* limit sequence number (credit) */ |
423 | u32 s_ahgpsn; /* set to the psn in the copy of the header */ |
424 | u16 s_cur_size; /* size of send packet in bytes */ |
425 | u16 s_rdma_ack_cnt; |
426 | u8 s_hdrwords; /* size of s_hdr in 32 bit words */ |
427 | s8 s_ahgidx; |
428 | u8 s_state; /* opcode of last packet sent */ |
429 | u8 s_ack_state; /* opcode of packet to ACK */ |
430 | u8 s_nak_state; /* non-zero if NAK is pending */ |
431 | u8 r_nak_state; /* non-zero if NAK is pending */ |
432 | u8 s_retry; /* requester retry counter */ |
433 | u8 s_rnr_retry; /* requester RNR retry counter */ |
434 | u8 s_num_rd_atomic; /* number of RDMA read/atomic pending */ |
435 | u8 s_tail_ack_queue; /* index into s_ack_queue[] */ |
436 | u8 s_acked_ack_queue; /* index into s_ack_queue[] */ |
437 | |
438 | struct rvt_sge_state s_ack_rdma_sge; |
439 | struct timer_list s_timer; |
440 | struct hrtimer s_rnr_timer; |
441 | |
442 | atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */ |
443 | |
444 | /* |
445 | * This sge list MUST be last. Do not add anything below here. |
446 | */ |
447 | struct rvt_sge *r_sg_list /* verified SGEs */ |
448 | ____cacheline_aligned_in_smp; |
449 | }; |
450 | |
451 | struct rvt_srq { |
452 | struct ib_srq ibsrq; |
453 | struct rvt_rq rq; |
454 | struct rvt_mmap_info *ip; |
455 | /* send signal when number of RWQEs < limit */ |
456 | u32 limit; |
457 | }; |
458 | |
459 | static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq) |
460 | { |
461 | return container_of(ibsrq, struct rvt_srq, ibsrq); |
462 | } |
463 | |
464 | static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp) |
465 | { |
466 | return container_of(ibqp, struct rvt_qp, ibqp); |
467 | } |
468 | |
469 | #define RVT_QPN_MAX BIT(24) |
470 | #define RVT_QPNMAP_ENTRIES (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE) |
471 | #define RVT_BITS_PER_PAGE (PAGE_SIZE * BITS_PER_BYTE) |
472 | #define RVT_BITS_PER_PAGE_MASK (RVT_BITS_PER_PAGE - 1) |
473 | #define RVT_QPN_MASK IB_QPN_MASK |
474 | |
475 | /* |
476 | * QPN-map pages start out as NULL, they get allocated upon |
477 | * first use and are never deallocated. This way, |
478 | * large bitmaps are not allocated unless large numbers of QPs are used. |
479 | */ |
480 | struct rvt_qpn_map { |
481 | void *page; |
482 | }; |
483 | |
484 | struct rvt_qpn_table { |
485 | spinlock_t lock; /* protect changes to the qp table */ |
486 | unsigned flags; /* flags for QP0/1 allocated for each port */ |
487 | u32 last; /* last QP number allocated */ |
488 | u32 nmaps; /* size of the map table */ |
489 | u16 limit; |
490 | u8 incr; |
491 | /* bit map of free QP numbers other than 0/1 */ |
492 | struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES]; |
493 | }; |
494 | |
495 | struct rvt_qp_ibdev { |
496 | u32 qp_table_size; |
497 | u32 qp_table_bits; |
498 | struct rvt_qp __rcu **qp_table; |
499 | spinlock_t qpt_lock; /* qptable lock */ |
500 | struct rvt_qpn_table qpn_table; |
501 | }; |
502 | |
503 | /* |
504 | * There is one struct rvt_mcast for each multicast GID. |
505 | * All attached QPs are then stored as a list of |
506 | * struct rvt_mcast_qp. |
507 | */ |
508 | struct rvt_mcast_qp { |
509 | struct list_head list; |
510 | struct rvt_qp *qp; |
511 | }; |
512 | |
513 | struct rvt_mcast_addr { |
514 | union ib_gid mgid; |
515 | u16 lid; |
516 | }; |
517 | |
518 | struct rvt_mcast { |
519 | struct rb_node rb_node; |
520 | struct rvt_mcast_addr mcast_addr; |
521 | struct list_head qp_list; |
522 | wait_queue_head_t wait; |
523 | atomic_t refcount; |
524 | int n_attached; |
525 | }; |
526 | |
527 | /* |
528 | * Since struct rvt_swqe is not a fixed size, we can't simply index into |
529 | * struct rvt_qp.s_wq. This function does the array index computation. |
530 | */ |
531 | static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp, |
532 | unsigned n) |
533 | { |
534 | return (struct rvt_swqe *)((char *)qp->s_wq + |
535 | (sizeof(struct rvt_swqe) + |
536 | qp->s_max_sge * |
537 | sizeof(struct rvt_sge)) * n); |
538 | } |
539 | |
540 | /* |
541 | * Since struct rvt_rwqe is not a fixed size, we can't simply index into |
542 | * struct rvt_rwq.wq. This function does the array index computation. |
543 | */ |
544 | static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n) |
545 | { |
546 | return (struct rvt_rwqe *) |
547 | ((char *)rq->kwq->curr_wq + |
548 | (sizeof(struct rvt_rwqe) + |
549 | rq->max_sge * sizeof(struct ib_sge)) * n); |
550 | } |
551 | |
552 | /** |
553 | * rvt_is_user_qp - return if this is user mode QP |
554 | * @qp - the target QP |
555 | */ |
556 | static inline bool rvt_is_user_qp(struct rvt_qp *qp) |
557 | { |
558 | return !!qp->pid; |
559 | } |
560 | |
561 | /** |
562 | * rvt_get_qp - get a QP reference |
563 | * @qp - the QP to hold |
564 | */ |
565 | static inline void rvt_get_qp(struct rvt_qp *qp) |
566 | { |
567 | atomic_inc(v: &qp->refcount); |
568 | } |
569 | |
570 | /** |
571 | * rvt_put_qp - release a QP reference |
572 | * @qp - the QP to release |
573 | */ |
574 | static inline void rvt_put_qp(struct rvt_qp *qp) |
575 | { |
576 | if (qp && atomic_dec_and_test(v: &qp->refcount)) |
577 | wake_up(&qp->wait); |
578 | } |
579 | |
580 | /** |
581 | * rvt_put_swqe - drop mr refs held by swqe |
582 | * @wqe - the send wqe |
583 | * |
584 | * This drops any mr references held by the swqe |
585 | */ |
586 | static inline void rvt_put_swqe(struct rvt_swqe *wqe) |
587 | { |
588 | int i; |
589 | |
590 | for (i = 0; i < wqe->wr.num_sge; i++) { |
591 | struct rvt_sge *sge = &wqe->sg_list[i]; |
592 | |
593 | rvt_put_mr(mr: sge->mr); |
594 | } |
595 | } |
596 | |
597 | /** |
598 | * rvt_qp_wqe_reserve - reserve operation |
599 | * @qp - the rvt qp |
600 | * @wqe - the send wqe |
601 | * |
602 | * This routine used in post send to record |
603 | * a wqe relative reserved operation use. |
604 | */ |
605 | static inline void rvt_qp_wqe_reserve( |
606 | struct rvt_qp *qp, |
607 | struct rvt_swqe *wqe) |
608 | { |
609 | atomic_inc(v: &qp->s_reserved_used); |
610 | } |
611 | |
612 | /** |
613 | * rvt_qp_wqe_unreserve - clean reserved operation |
614 | * @qp - the rvt qp |
615 | * @flags - send wqe flags |
616 | * |
617 | * This decrements the reserve use count. |
618 | * |
619 | * This call MUST precede the change to |
620 | * s_last to insure that post send sees a stable |
621 | * s_avail. |
622 | * |
623 | * An smp_mp__after_atomic() is used to insure |
624 | * the compiler does not juggle the order of the s_last |
625 | * ring index and the decrementing of s_reserved_used. |
626 | */ |
627 | static inline void rvt_qp_wqe_unreserve(struct rvt_qp *qp, int flags) |
628 | { |
629 | if (unlikely(flags & RVT_SEND_RESERVE_USED)) { |
630 | atomic_dec(v: &qp->s_reserved_used); |
631 | /* insure no compiler re-order up to s_last change */ |
632 | smp_mb__after_atomic(); |
633 | } |
634 | } |
635 | |
636 | extern const enum ib_wc_opcode ib_rvt_wc_opcode[]; |
637 | |
638 | /* |
639 | * Compare the lower 24 bits of the msn values. |
640 | * Returns an integer <, ==, or > than zero. |
641 | */ |
642 | static inline int rvt_cmp_msn(u32 a, u32 b) |
643 | { |
644 | return (((int)a) - ((int)b)) << 8; |
645 | } |
646 | |
647 | __be32 rvt_compute_aeth(struct rvt_qp *qp); |
648 | |
649 | void rvt_get_credit(struct rvt_qp *qp, u32 aeth); |
650 | |
651 | u32 rvt_restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, u32 len); |
652 | |
653 | /** |
654 | * rvt_div_round_up_mtu - round up divide |
655 | * @qp - the qp pair |
656 | * @len - the length |
657 | * |
658 | * Perform a shift based mtu round up divide |
659 | */ |
660 | static inline u32 rvt_div_round_up_mtu(struct rvt_qp *qp, u32 len) |
661 | { |
662 | return (len + qp->pmtu - 1) >> qp->log_pmtu; |
663 | } |
664 | |
665 | /** |
666 | * @qp - the qp pair |
667 | * @len - the length |
668 | * |
669 | * Perform a shift based mtu divide |
670 | */ |
671 | static inline u32 rvt_div_mtu(struct rvt_qp *qp, u32 len) |
672 | { |
673 | return len >> qp->log_pmtu; |
674 | } |
675 | |
676 | /** |
677 | * rvt_timeout_to_jiffies - Convert a ULP timeout input into jiffies |
678 | * @timeout - timeout input(0 - 31). |
679 | * |
680 | * Return a timeout value in jiffies. |
681 | */ |
682 | static inline unsigned long rvt_timeout_to_jiffies(u8 timeout) |
683 | { |
684 | if (timeout > 31) |
685 | timeout = 31; |
686 | |
687 | return usecs_to_jiffies(u: 1U << timeout) * 4096UL / 1000UL; |
688 | } |
689 | |
690 | /** |
691 | * rvt_lookup_qpn - return the QP with the given QPN |
692 | * @ibp: the ibport |
693 | * @qpn: the QP number to look up |
694 | * |
695 | * The caller must hold the rcu_read_lock(), and keep the lock until |
696 | * the returned qp is no longer in use. |
697 | */ |
698 | static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi, |
699 | struct rvt_ibport *rvp, |
700 | u32 qpn) __must_hold(RCU) |
701 | { |
702 | struct rvt_qp *qp = NULL; |
703 | |
704 | if (unlikely(qpn <= 1)) { |
705 | qp = rcu_dereference(rvp->qp[qpn]); |
706 | } else { |
707 | u32 n = hash_32(val: qpn, bits: rdi->qp_dev->qp_table_bits); |
708 | |
709 | for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp; |
710 | qp = rcu_dereference(qp->next)) |
711 | if (qp->ibqp.qp_num == qpn) |
712 | break; |
713 | } |
714 | return qp; |
715 | } |
716 | |
717 | /** |
718 | * rvt_mod_retry_timer - mod a retry timer |
719 | * @qp - the QP |
720 | * @shift - timeout shift to wait for multiple packets |
721 | * Modify a potentially already running retry timer |
722 | */ |
723 | static inline void rvt_mod_retry_timer_ext(struct rvt_qp *qp, u8 shift) |
724 | { |
725 | struct ib_qp *ibqp = &qp->ibqp; |
726 | struct rvt_dev_info *rdi = ib_to_rvt(ibdev: ibqp->device); |
727 | |
728 | lockdep_assert_held(&qp->s_lock); |
729 | qp->s_flags |= RVT_S_TIMER; |
730 | /* 4.096 usec. * (1 << qp->timeout) */ |
731 | mod_timer(timer: &qp->s_timer, expires: jiffies + rdi->busy_jiffies + |
732 | (qp->timeout_jiffies << shift)); |
733 | } |
734 | |
735 | static inline void rvt_mod_retry_timer(struct rvt_qp *qp) |
736 | { |
737 | return rvt_mod_retry_timer_ext(qp, shift: 0); |
738 | } |
739 | |
740 | /** |
741 | * rvt_put_qp_swqe - drop refs held by swqe |
742 | * @qp: the send qp |
743 | * @wqe: the send wqe |
744 | * |
745 | * This drops any references held by the swqe |
746 | */ |
747 | static inline void rvt_put_qp_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe) |
748 | { |
749 | rvt_put_swqe(wqe); |
750 | if (qp->allowed_ops == IB_OPCODE_UD) |
751 | rdma_destroy_ah_attr(ah_attr: wqe->ud_wr.attr); |
752 | } |
753 | |
754 | /** |
755 | * rvt_qp_sqwe_incr - increment ring index |
756 | * @qp: the qp |
757 | * @val: the starting value |
758 | * |
759 | * Return: the new value wrapping as appropriate |
760 | */ |
761 | static inline u32 |
762 | rvt_qp_swqe_incr(struct rvt_qp *qp, u32 val) |
763 | { |
764 | if (++val >= qp->s_size) |
765 | val = 0; |
766 | return val; |
767 | } |
768 | |
769 | int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err); |
770 | |
771 | /** |
772 | * rvt_recv_cq - add a new entry to completion queue |
773 | * by receive queue |
774 | * @qp: receive queue |
775 | * @wc: work completion entry to add |
776 | * @solicited: true if @entry is solicited |
777 | * |
778 | * This is wrapper function for rvt_enter_cq function call by |
779 | * receive queue. If rvt_cq_enter return false, it means cq is |
780 | * full and the qp is put into error state. |
781 | */ |
782 | static inline void rvt_recv_cq(struct rvt_qp *qp, struct ib_wc *wc, |
783 | bool solicited) |
784 | { |
785 | struct rvt_cq *cq = ibcq_to_rvtcq(ibcq: qp->ibqp.recv_cq); |
786 | |
787 | if (unlikely(!rvt_cq_enter(cq, wc, solicited))) |
788 | rvt_error_qp(qp, err: IB_WC_LOC_QP_OP_ERR); |
789 | } |
790 | |
791 | /** |
792 | * rvt_send_cq - add a new entry to completion queue |
793 | * by send queue |
794 | * @qp: send queue |
795 | * @wc: work completion entry to add |
796 | * @solicited: true if @entry is solicited |
797 | * |
798 | * This is wrapper function for rvt_enter_cq function call by |
799 | * send queue. If rvt_cq_enter return false, it means cq is |
800 | * full and the qp is put into error state. |
801 | */ |
802 | static inline void rvt_send_cq(struct rvt_qp *qp, struct ib_wc *wc, |
803 | bool solicited) |
804 | { |
805 | struct rvt_cq *cq = ibcq_to_rvtcq(ibcq: qp->ibqp.send_cq); |
806 | |
807 | if (unlikely(!rvt_cq_enter(cq, wc, solicited))) |
808 | rvt_error_qp(qp, err: IB_WC_LOC_QP_OP_ERR); |
809 | } |
810 | |
811 | /** |
812 | * rvt_qp_complete_swqe - insert send completion |
813 | * @qp - the qp |
814 | * @wqe - the send wqe |
815 | * @opcode - wc operation (driver dependent) |
816 | * @status - completion status |
817 | * |
818 | * Update the s_last information, and then insert a send |
819 | * completion into the completion |
820 | * queue if the qp indicates it should be done. |
821 | * |
822 | * See IBTA 10.7.3.1 for info on completion |
823 | * control. |
824 | * |
825 | * Return: new last |
826 | */ |
827 | static inline u32 |
828 | rvt_qp_complete_swqe(struct rvt_qp *qp, |
829 | struct rvt_swqe *wqe, |
830 | enum ib_wc_opcode opcode, |
831 | enum ib_wc_status status) |
832 | { |
833 | bool need_completion; |
834 | u64 wr_id; |
835 | u32 byte_len, last; |
836 | int flags = wqe->wr.send_flags; |
837 | |
838 | rvt_qp_wqe_unreserve(qp, flags); |
839 | rvt_put_qp_swqe(qp, wqe); |
840 | |
841 | need_completion = |
842 | !(flags & RVT_SEND_RESERVE_USED) && |
843 | (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) || |
844 | (flags & IB_SEND_SIGNALED) || |
845 | status != IB_WC_SUCCESS); |
846 | if (need_completion) { |
847 | wr_id = wqe->wr.wr_id; |
848 | byte_len = wqe->length; |
849 | /* above fields required before writing s_last */ |
850 | } |
851 | last = rvt_qp_swqe_incr(qp, val: qp->s_last); |
852 | /* see rvt_qp_is_avail() */ |
853 | smp_store_release(&qp->s_last, last); |
854 | if (need_completion) { |
855 | struct ib_wc w = { |
856 | .wr_id = wr_id, |
857 | .status = status, |
858 | .opcode = opcode, |
859 | .qp = &qp->ibqp, |
860 | .byte_len = byte_len, |
861 | }; |
862 | rvt_send_cq(qp, wc: &w, solicited: status != IB_WC_SUCCESS); |
863 | } |
864 | return last; |
865 | } |
866 | |
867 | extern const int ib_rvt_state_ops[]; |
868 | |
869 | struct rvt_dev_info; |
870 | int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only); |
871 | void rvt_comm_est(struct rvt_qp *qp); |
872 | void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err); |
873 | unsigned long rvt_rnr_tbl_to_usec(u32 index); |
874 | enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t); |
875 | void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth); |
876 | void rvt_del_timers_sync(struct rvt_qp *qp); |
877 | void rvt_stop_rc_timers(struct rvt_qp *qp); |
878 | void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift); |
879 | static inline void rvt_add_retry_timer(struct rvt_qp *qp) |
880 | { |
881 | rvt_add_retry_timer_ext(qp, shift: 0); |
882 | } |
883 | |
884 | void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss, |
885 | void *data, u32 length, |
886 | bool release, bool copy_last); |
887 | void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe, |
888 | enum ib_wc_status status); |
889 | void rvt_ruc_loopback(struct rvt_qp *qp); |
890 | |
891 | /** |
892 | * struct rvt_qp_iter - the iterator for QPs |
893 | * @qp - the current QP |
894 | * |
895 | * This structure defines the current iterator |
896 | * state for sequenced access to all QPs relative |
897 | * to an rvt_dev_info. |
898 | */ |
899 | struct rvt_qp_iter { |
900 | struct rvt_qp *qp; |
901 | /* private: backpointer */ |
902 | struct rvt_dev_info *rdi; |
903 | /* private: callback routine */ |
904 | void (*cb)(struct rvt_qp *qp, u64 v); |
905 | /* private: for arg to callback routine */ |
906 | u64 v; |
907 | /* private: number of SMI,GSI QPs for device */ |
908 | int specials; |
909 | /* private: current iterator index */ |
910 | int n; |
911 | }; |
912 | |
913 | /** |
914 | * ib_cq_tail - Return tail index of cq buffer |
915 | * @send_cq - The cq for send |
916 | * |
917 | * This is called in qp_iter_print to get tail |
918 | * of cq buffer. |
919 | */ |
920 | static inline u32 ib_cq_tail(struct ib_cq *send_cq) |
921 | { |
922 | struct rvt_cq *cq = ibcq_to_rvtcq(ibcq: send_cq); |
923 | |
924 | return ibcq_to_rvtcq(ibcq: send_cq)->ip ? |
925 | RDMA_READ_UAPI_ATOMIC(cq->queue->tail) : |
926 | ibcq_to_rvtcq(ibcq: send_cq)->kqueue->tail; |
927 | } |
928 | |
929 | /** |
930 | * ib_cq_head - Return head index of cq buffer |
931 | * @send_cq - The cq for send |
932 | * |
933 | * This is called in qp_iter_print to get head |
934 | * of cq buffer. |
935 | */ |
936 | static inline u32 ib_cq_head(struct ib_cq *send_cq) |
937 | { |
938 | struct rvt_cq *cq = ibcq_to_rvtcq(ibcq: send_cq); |
939 | |
940 | return ibcq_to_rvtcq(ibcq: send_cq)->ip ? |
941 | RDMA_READ_UAPI_ATOMIC(cq->queue->head) : |
942 | ibcq_to_rvtcq(ibcq: send_cq)->kqueue->head; |
943 | } |
944 | |
945 | /** |
946 | * rvt_free_rq - free memory allocated for rvt_rq struct |
947 | * @rvt_rq: request queue data structure |
948 | * |
949 | * This function should only be called if the rvt_mmap_info() |
950 | * has not succeeded. |
951 | */ |
952 | static inline void rvt_free_rq(struct rvt_rq *rq) |
953 | { |
954 | kvfree(addr: rq->kwq); |
955 | rq->kwq = NULL; |
956 | vfree(addr: rq->wq); |
957 | rq->wq = NULL; |
958 | } |
959 | |
960 | /** |
961 | * rvt_to_iport - Get the ibport pointer |
962 | * @qp: the qp pointer |
963 | * |
964 | * This function returns the ibport pointer from the qp pointer. |
965 | */ |
966 | static inline struct rvt_ibport *rvt_to_iport(struct rvt_qp *qp) |
967 | { |
968 | struct rvt_dev_info *rdi = ib_to_rvt(ibdev: qp->ibqp.device); |
969 | |
970 | return rdi->ports[qp->port_num - 1]; |
971 | } |
972 | |
973 | /** |
974 | * rvt_rc_credit_avail - Check if there are enough RC credits for the request |
975 | * @qp: the qp |
976 | * @wqe: the request |
977 | * |
978 | * This function returns false when there are not enough credits for the given |
979 | * request and true otherwise. |
980 | */ |
981 | static inline bool rvt_rc_credit_avail(struct rvt_qp *qp, struct rvt_swqe *wqe) |
982 | { |
983 | lockdep_assert_held(&qp->s_lock); |
984 | if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) && |
985 | rvt_cmp_msn(a: wqe->ssn, b: qp->s_lsn + 1) > 0) { |
986 | struct rvt_ibport *rvp = rvt_to_iport(qp); |
987 | |
988 | qp->s_flags |= RVT_S_WAIT_SSN_CREDIT; |
989 | rvp->n_rc_crwaits++; |
990 | return false; |
991 | } |
992 | return true; |
993 | } |
994 | |
995 | struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi, |
996 | u64 v, |
997 | void (*cb)(struct rvt_qp *qp, u64 v)); |
998 | int rvt_qp_iter_next(struct rvt_qp_iter *iter); |
999 | void rvt_qp_iter(struct rvt_dev_info *rdi, |
1000 | u64 v, |
1001 | void (*cb)(struct rvt_qp *qp, u64 v)); |
1002 | void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey); |
1003 | #endif /* DEF_RDMAVT_INCQP_H */ |
1004 | |