1 /* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */
2 /*
3  * Copyright(c) 2018 Intel Corporation.
4  *
5  */
6 #ifndef HFI1_TID_RDMA_H
7 #define HFI1_TID_RDMA_H
8 
9 #include <linux/circ_buf.h>
10 #include "common.h"
11 
12 /* Add a convenience helper */
13 #define CIRC_ADD(val, add, size) (((val) + (add)) & ((size) - 1))
14 #define CIRC_NEXT(val, size) CIRC_ADD(val, 1, size)
15 #define CIRC_PREV(val, size) CIRC_ADD(val, -1, size)
16 
17 #define TID_RDMA_MIN_SEGMENT_SIZE       BIT(18)   /* 256 KiB (for now) */
18 #define TID_RDMA_MAX_SEGMENT_SIZE       BIT(18)   /* 256 KiB (for now) */
19 #define TID_RDMA_MAX_PAGES              (BIT(18) >> PAGE_SHIFT)
20 
21 /*
22  * Bit definitions for priv->s_flags.
23  * These bit flags overload the bit flags defined for the QP's s_flags.
24  * Due to the fact that these bit fields are used only for the QP priv
25  * s_flags, there are no collisions.
26  *
27  * HFI1_S_TID_WAIT_INTERLCK - QP is waiting for requester interlock
28  * HFI1_R_TID_WAIT_INTERLCK - QP is waiting for responder interlock
29  */
30 #define HFI1_S_TID_BUSY_SET       BIT(0)
31 /* BIT(1) reserved for RVT_S_BUSY. */
32 #define HFI1_R_TID_RSC_TIMER      BIT(2)
33 /* BIT(3) reserved for RVT_S_RESP_PENDING. */
34 /* BIT(4) reserved for RVT_S_ACK_PENDING. */
35 #define HFI1_S_TID_WAIT_INTERLCK  BIT(5)
36 #define HFI1_R_TID_WAIT_INTERLCK  BIT(6)
37 /* BIT(7) - BIT(15) reserved for RVT_S_WAIT_*. */
38 /* BIT(16) reserved for RVT_S_SEND_ONE */
39 #define HFI1_S_TID_RETRY_TIMER    BIT(17)
40 /* BIT(18) reserved for RVT_S_ECN. */
41 #define HFI1_R_TID_SW_PSN         BIT(19)
42 /* BIT(26) reserved for HFI1_S_WAIT_HALT */
43 /* BIT(27) reserved for HFI1_S_WAIT_TID_RESP */
44 /* BIT(28) reserved for HFI1_S_WAIT_TID_SPACE */
45 
46 /*
47  * Unlike regular IB RDMA VERBS, which do not require an entry
48  * in the s_ack_queue, TID RDMA WRITE requests do because they
49  * generate responses.
50  * Therefore, the s_ack_queue needs to be extended by a certain
51  * amount. The key point is that the queue needs to be extended
52  * without letting the "user" know so they user doesn't end up
53  * using these extra entries.
54  */
55 #define HFI1_TID_RDMA_WRITE_CNT 8
56 
57 struct tid_rdma_params {
58 	struct rcu_head rcu_head;
59 	u32 qp;
60 	u32 max_len;
61 	u16 jkey;
62 	u8 max_read;
63 	u8 max_write;
64 	u8 timeout;
65 	u8 urg;
66 	u8 version;
67 };
68 
69 struct tid_rdma_qp_params {
70 	struct work_struct trigger_work;
71 	struct tid_rdma_params local;
72 	struct tid_rdma_params __rcu *remote;
73 };
74 
75 /* Track state for each hardware flow */
76 struct tid_flow_state {
77 	u32 generation;
78 	u32 psn;
79 	u32 r_next_psn;      /* next PSN to be received (in TID space) */
80 	u8 index;
81 	u8 last_index;
82 	u8 flags;
83 };
84 
85 enum tid_rdma_req_state {
86 	TID_REQUEST_INACTIVE = 0,
87 	TID_REQUEST_INIT,
88 	TID_REQUEST_INIT_RESEND,
89 	TID_REQUEST_ACTIVE,
90 	TID_REQUEST_RESEND,
91 	TID_REQUEST_RESEND_ACTIVE,
92 	TID_REQUEST_QUEUED,
93 	TID_REQUEST_SYNC,
94 	TID_REQUEST_RNR_NAK,
95 	TID_REQUEST_COMPLETE,
96 };
97 
98 struct tid_rdma_request {
99 	struct rvt_qp *qp;
100 	struct hfi1_ctxtdata *rcd;
101 	union {
102 		struct rvt_swqe *swqe;
103 		struct rvt_ack_entry *ack;
104 	} e;
105 
106 	struct tid_rdma_flow *flows;	/* array of tid flows */
107 	struct rvt_sge_state ss; /* SGE state for TID RDMA requests */
108 	u16 n_flows;		/* size of the flow buffer window */
109 	u16 setup_head;		/* flow index we are setting up */
110 	u16 clear_tail;		/* flow index we are clearing */
111 	u16 flow_idx;		/* flow index most recently set up */
112 	u16 acked_tail;
113 
114 	u32 seg_len;
115 	u32 total_len;
116 	u32 r_ack_psn;          /* next expected ack PSN */
117 	u32 r_flow_psn;         /* IB PSN of next segment start */
118 	u32 r_last_acked;       /* IB PSN of last ACK'ed packet */
119 	u32 s_next_psn;		/* IB PSN of next segment start for read */
120 
121 	u32 total_segs;		/* segments required to complete a request */
122 	u32 cur_seg;		/* index of current segment */
123 	u32 comp_seg;           /* index of last completed segment */
124 	u32 ack_seg;            /* index of last ack'ed segment */
125 	u32 alloc_seg;          /* index of next segment to be allocated */
126 	u32 isge;		/* index of "current" sge */
127 	u32 ack_pending;        /* num acks pending for this request */
128 
129 	enum tid_rdma_req_state state;
130 };
131 
132 /*
133  * When header suppression is used, PSNs associated with a "flow" are
134  * relevant (and not the PSNs maintained by verbs). Track per-flow
135  * PSNs here for a TID RDMA segment.
136  *
137  */
138 struct flow_state {
139 	u32 flags;
140 	u32 resp_ib_psn;     /* The IB PSN of the response for this flow */
141 	u32 generation;      /* generation of flow */
142 	u32 spsn;            /* starting PSN in TID space */
143 	u32 lpsn;            /* last PSN in TID space */
144 	u32 r_next_psn;      /* next PSN to be received (in TID space) */
145 
146 	/* For tid rdma read */
147 	u32 ib_spsn;         /* starting PSN in Verbs space */
148 	u32 ib_lpsn;         /* last PSn in Verbs space */
149 };
150 
151 struct tid_rdma_pageset {
152 	dma_addr_t addr : 48; /* Only needed for the first page */
153 	u8 idx: 8;
154 	u8 count : 7;
155 	u8 mapped: 1;
156 };
157 
158 /**
159  * kern_tid_node - used for managing TID's in TID groups
160  *
161  * @grp_idx: rcd relative index to tid_group
162  * @map: grp->map captured prior to programming this TID group in HW
163  * @cnt: Only @cnt of available group entries are actually programmed
164  */
165 struct kern_tid_node {
166 	struct tid_group *grp;
167 	u8 map;
168 	u8 cnt;
169 };
170 
171 /* Overall info for a TID RDMA segment */
172 struct tid_rdma_flow {
173 	/*
174 	 * While a TID RDMA segment is being transferred, it uses a QP number
175 	 * from the "KDETH section of QP numbers" (which is different from the
176 	 * QP number that originated the request). Bits 11-15 of these QP
177 	 * numbers identify the "TID flow" for the segment.
178 	 */
179 	struct flow_state flow_state;
180 	struct tid_rdma_request *req;
181 	u32 tid_qpn;
182 	u32 tid_offset;
183 	u32 length;
184 	u32 sent;
185 	u8 tnode_cnt;
186 	u8 tidcnt;
187 	u8 tid_idx;
188 	u8 idx;
189 	u8 npagesets;
190 	u8 npkts;
191 	u8 pkt;
192 	u8 resync_npkts;
193 	struct kern_tid_node tnode[TID_RDMA_MAX_PAGES];
194 	struct tid_rdma_pageset pagesets[TID_RDMA_MAX_PAGES];
195 	u32 tid_entry[TID_RDMA_MAX_PAGES];
196 };
197 
198 enum tid_rnr_nak_state {
199 	TID_RNR_NAK_INIT = 0,
200 	TID_RNR_NAK_SEND,
201 	TID_RNR_NAK_SENT,
202 };
203 
204 bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data);
205 bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data);
206 bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data);
207 void tid_rdma_conn_error(struct rvt_qp *qp);
208 void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p);
209 
210 int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit);
211 int hfi1_kern_exp_rcv_setup(struct tid_rdma_request *req,
212 			    struct rvt_sge_state *ss, bool *last);
213 int hfi1_kern_exp_rcv_clear(struct tid_rdma_request *req);
214 void hfi1_kern_exp_rcv_clear_all(struct tid_rdma_request *req);
215 void __trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe);
216 
217 /**
218  * trdma_clean_swqe - clean flows for swqe if large send queue
219  * @qp: the qp
220  * @wqe: the send wqe
221  */
222 static inline void trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
223 {
224 	if (!wqe->priv)
225 		return;
226 	__trdma_clean_swqe(qp, wqe);
227 }
228 
229 void hfi1_kern_read_tid_flow_free(struct rvt_qp *qp);
230 
231 int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp,
232 		      struct ib_qp_init_attr *init_attr);
233 void hfi1_qp_priv_tid_free(struct rvt_dev_info *rdi, struct rvt_qp *qp);
234 
235 void hfi1_tid_rdma_flush_wait(struct rvt_qp *qp);
236 
237 int hfi1_kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp);
238 void hfi1_kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp);
239 void hfi1_kern_init_ctxt_generations(struct hfi1_ctxtdata *rcd);
240 
241 struct cntr_entry;
242 u64 hfi1_access_sw_tid_wait(const struct cntr_entry *entry,
243 			    void *context, int vl, int mode, u64 data);
244 
245 u32 hfi1_build_tid_rdma_read_packet(struct rvt_swqe *wqe,
246 				    struct ib_other_headers *ohdr,
247 				    u32 *bth1, u32 *bth2, u32 *len);
248 u32 hfi1_build_tid_rdma_read_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
249 				 struct ib_other_headers *ohdr, u32 *bth1,
250 				 u32 *bth2, u32 *len);
251 void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet);
252 u32 hfi1_build_tid_rdma_read_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
253 				  struct ib_other_headers *ohdr, u32 *bth0,
254 				  u32 *bth1, u32 *bth2, u32 *len, bool *last);
255 void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet);
256 bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
257 			      struct hfi1_pportdata *ppd,
258 			      struct hfi1_packet *packet);
259 void hfi1_tid_rdma_restart_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
260 			       u32 *bth2);
261 void hfi1_qp_kern_exp_rcv_clear_all(struct rvt_qp *qp);
262 bool hfi1_tid_rdma_wqe_interlock(struct rvt_qp *qp, struct rvt_swqe *wqe);
263 
264 void setup_tid_rdma_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe);
265 static inline void hfi1_setup_tid_rdma_wqe(struct rvt_qp *qp,
266 					   struct rvt_swqe *wqe)
267 {
268 	if (wqe->priv &&
269 	    (wqe->wr.opcode == IB_WR_RDMA_READ ||
270 	     wqe->wr.opcode == IB_WR_RDMA_WRITE) &&
271 	    wqe->length >= TID_RDMA_MIN_SEGMENT_SIZE)
272 		setup_tid_rdma_wqe(qp, wqe);
273 }
274 
275 u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
276 				  struct ib_other_headers *ohdr,
277 				  u32 *bth1, u32 *bth2, u32 *len);
278 
279 void hfi1_compute_tid_rdma_flow_wt(void);
280 
281 void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet);
282 
283 u32 hfi1_build_tid_rdma_write_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
284 				   struct ib_other_headers *ohdr, u32 *bth1,
285 				   u32 bth2, u32 *len,
286 				   struct rvt_sge_state **ss);
287 
288 void hfi1_del_tid_reap_timer(struct rvt_qp *qp);
289 
290 void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet);
291 
292 bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe,
293 				struct ib_other_headers *ohdr,
294 				u32 *bth1, u32 *bth2, u32 *len);
295 
296 void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet);
297 
298 u32 hfi1_build_tid_rdma_write_ack(struct rvt_qp *qp, struct rvt_ack_entry *e,
299 				  struct ib_other_headers *ohdr, u16 iflow,
300 				  u32 *bth1, u32 *bth2);
301 
302 void hfi1_rc_rcv_tid_rdma_ack(struct hfi1_packet *packet);
303 
304 void hfi1_add_tid_retry_timer(struct rvt_qp *qp);
305 void hfi1_del_tid_retry_timer(struct rvt_qp *qp);
306 
307 u32 hfi1_build_tid_rdma_resync(struct rvt_qp *qp, struct rvt_swqe *wqe,
308 			       struct ib_other_headers *ohdr, u32 *bth1,
309 			       u32 *bth2, u16 fidx);
310 
311 void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet);
312 
313 struct hfi1_pkt_state;
314 int hfi1_make_tid_rdma_pkt(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
315 
316 void _hfi1_do_tid_send(struct work_struct *work);
317 
318 bool hfi1_schedule_tid_send(struct rvt_qp *qp);
319 
320 bool hfi1_tid_rdma_ack_interlock(struct rvt_qp *qp, struct rvt_ack_entry *e);
321 
322 #endif /* HFI1_TID_RDMA_H */
323