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 	u8 index;
80 	u8 last_index;
81 };
82 
83 enum tid_rdma_req_state {
84 	TID_REQUEST_INACTIVE = 0,
85 	TID_REQUEST_INIT,
86 	TID_REQUEST_INIT_RESEND,
87 	TID_REQUEST_ACTIVE,
88 	TID_REQUEST_RESEND,
89 	TID_REQUEST_RESEND_ACTIVE,
90 	TID_REQUEST_QUEUED,
91 	TID_REQUEST_SYNC,
92 	TID_REQUEST_RNR_NAK,
93 	TID_REQUEST_COMPLETE,
94 };
95 
96 struct tid_rdma_request {
97 	struct rvt_qp *qp;
98 	struct hfi1_ctxtdata *rcd;
99 	union {
100 		struct rvt_swqe *swqe;
101 		struct rvt_ack_entry *ack;
102 	} e;
103 
104 	struct tid_rdma_flow *flows;	/* array of tid flows */
105 	struct rvt_sge_state ss; /* SGE state for TID RDMA requests */
106 	u16 n_flows;		/* size of the flow buffer window */
107 	u16 setup_head;		/* flow index we are setting up */
108 	u16 clear_tail;		/* flow index we are clearing */
109 	u16 flow_idx;		/* flow index most recently set up */
110 	u16 acked_tail;
111 
112 	u32 seg_len;
113 	u32 total_len;
114 	u32 r_ack_psn;          /* next expected ack PSN */
115 	u32 r_flow_psn;         /* IB PSN of next segment start */
116 	u32 r_last_acked;       /* IB PSN of last ACK'ed packet */
117 	u32 s_next_psn;		/* IB PSN of next segment start for read */
118 
119 	u32 total_segs;		/* segments required to complete a request */
120 	u32 cur_seg;		/* index of current segment */
121 	u32 comp_seg;           /* index of last completed segment */
122 	u32 ack_seg;            /* index of last ack'ed segment */
123 	u32 alloc_seg;          /* index of next segment to be allocated */
124 	u32 isge;		/* index of "current" sge */
125 	u32 ack_pending;        /* num acks pending for this request */
126 
127 	enum tid_rdma_req_state state;
128 };
129 
130 /*
131  * When header suppression is used, PSNs associated with a "flow" are
132  * relevant (and not the PSNs maintained by verbs). Track per-flow
133  * PSNs here for a TID RDMA segment.
134  *
135  */
136 struct flow_state {
137 	u32 flags;
138 	u32 resp_ib_psn;     /* The IB PSN of the response for this flow */
139 	u32 generation;      /* generation of flow */
140 	u32 spsn;            /* starting PSN in TID space */
141 	u32 lpsn;            /* last PSN in TID space */
142 	u32 r_next_psn;      /* next PSN to be received (in TID space) */
143 
144 	/* For tid rdma read */
145 	u32 ib_spsn;         /* starting PSN in Verbs space */
146 	u32 ib_lpsn;         /* last PSn in Verbs space */
147 };
148 
149 struct tid_rdma_pageset {
150 	dma_addr_t addr : 48; /* Only needed for the first page */
151 	u8 idx: 8;
152 	u8 count : 7;
153 	u8 mapped: 1;
154 };
155 
156 /**
157  * kern_tid_node - used for managing TID's in TID groups
158  *
159  * @grp_idx: rcd relative index to tid_group
160  * @map: grp->map captured prior to programming this TID group in HW
161  * @cnt: Only @cnt of available group entries are actually programmed
162  */
163 struct kern_tid_node {
164 	struct tid_group *grp;
165 	u8 map;
166 	u8 cnt;
167 };
168 
169 /* Overall info for a TID RDMA segment */
170 struct tid_rdma_flow {
171 	/*
172 	 * While a TID RDMA segment is being transferred, it uses a QP number
173 	 * from the "KDETH section of QP numbers" (which is different from the
174 	 * QP number that originated the request). Bits 11-15 of these QP
175 	 * numbers identify the "TID flow" for the segment.
176 	 */
177 	struct flow_state flow_state;
178 	struct tid_rdma_request *req;
179 	u32 tid_qpn;
180 	u32 tid_offset;
181 	u32 length;
182 	u32 sent;
183 	u8 tnode_cnt;
184 	u8 tidcnt;
185 	u8 tid_idx;
186 	u8 idx;
187 	u8 npagesets;
188 	u8 npkts;
189 	u8 pkt;
190 	u8 resync_npkts;
191 	struct kern_tid_node tnode[TID_RDMA_MAX_PAGES];
192 	struct tid_rdma_pageset pagesets[TID_RDMA_MAX_PAGES];
193 	u32 tid_entry[TID_RDMA_MAX_PAGES];
194 };
195 
196 enum tid_rnr_nak_state {
197 	TID_RNR_NAK_INIT = 0,
198 	TID_RNR_NAK_SEND,
199 	TID_RNR_NAK_SENT,
200 };
201 
202 bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data);
203 bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data);
204 bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data);
205 void tid_rdma_conn_error(struct rvt_qp *qp);
206 void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p);
207 
208 int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit);
209 int hfi1_kern_exp_rcv_setup(struct tid_rdma_request *req,
210 			    struct rvt_sge_state *ss, bool *last);
211 int hfi1_kern_exp_rcv_clear(struct tid_rdma_request *req);
212 void hfi1_kern_exp_rcv_clear_all(struct tid_rdma_request *req);
213 void __trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe);
214 
215 /**
216  * trdma_clean_swqe - clean flows for swqe if large send queue
217  * @qp: the qp
218  * @wqe: the send wqe
219  */
220 static inline void trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
221 {
222 	if (!wqe->priv)
223 		return;
224 	__trdma_clean_swqe(qp, wqe);
225 }
226 
227 void hfi1_kern_read_tid_flow_free(struct rvt_qp *qp);
228 
229 int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp,
230 		      struct ib_qp_init_attr *init_attr);
231 void hfi1_qp_priv_tid_free(struct rvt_dev_info *rdi, struct rvt_qp *qp);
232 
233 void hfi1_tid_rdma_flush_wait(struct rvt_qp *qp);
234 
235 int hfi1_kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp);
236 void hfi1_kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp);
237 void hfi1_kern_init_ctxt_generations(struct hfi1_ctxtdata *rcd);
238 
239 struct cntr_entry;
240 u64 hfi1_access_sw_tid_wait(const struct cntr_entry *entry,
241 			    void *context, int vl, int mode, u64 data);
242 
243 u32 hfi1_build_tid_rdma_read_packet(struct rvt_swqe *wqe,
244 				    struct ib_other_headers *ohdr,
245 				    u32 *bth1, u32 *bth2, u32 *len);
246 u32 hfi1_build_tid_rdma_read_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
247 				 struct ib_other_headers *ohdr, u32 *bth1,
248 				 u32 *bth2, u32 *len);
249 void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet);
250 u32 hfi1_build_tid_rdma_read_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
251 				  struct ib_other_headers *ohdr, u32 *bth0,
252 				  u32 *bth1, u32 *bth2, u32 *len, bool *last);
253 void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet);
254 bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
255 			      struct hfi1_pportdata *ppd,
256 			      struct hfi1_packet *packet);
257 void hfi1_tid_rdma_restart_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
258 			       u32 *bth2);
259 void hfi1_qp_kern_exp_rcv_clear_all(struct rvt_qp *qp);
260 bool hfi1_tid_rdma_wqe_interlock(struct rvt_qp *qp, struct rvt_swqe *wqe);
261 
262 void setup_tid_rdma_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe);
263 static inline void hfi1_setup_tid_rdma_wqe(struct rvt_qp *qp,
264 					   struct rvt_swqe *wqe)
265 {
266 	if (wqe->priv &&
267 	    (wqe->wr.opcode == IB_WR_RDMA_READ ||
268 	     wqe->wr.opcode == IB_WR_RDMA_WRITE) &&
269 	    wqe->length >= TID_RDMA_MIN_SEGMENT_SIZE)
270 		setup_tid_rdma_wqe(qp, wqe);
271 }
272 
273 u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
274 				  struct ib_other_headers *ohdr,
275 				  u32 *bth1, u32 *bth2, u32 *len);
276 
277 void hfi1_compute_tid_rdma_flow_wt(void);
278 
279 void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet);
280 
281 u32 hfi1_build_tid_rdma_write_resp(struct rvt_qp *qp, struct rvt_ack_entry *e,
282 				   struct ib_other_headers *ohdr, u32 *bth1,
283 				   u32 bth2, u32 *len,
284 				   struct rvt_sge_state **ss);
285 
286 void hfi1_del_tid_reap_timer(struct rvt_qp *qp);
287 
288 void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet);
289 
290 bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe,
291 				struct ib_other_headers *ohdr,
292 				u32 *bth1, u32 *bth2, u32 *len);
293 
294 void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet);
295 
296 u32 hfi1_build_tid_rdma_write_ack(struct rvt_qp *qp, struct rvt_ack_entry *e,
297 				  struct ib_other_headers *ohdr, u16 iflow,
298 				  u32 *bth1, u32 *bth2);
299 
300 void hfi1_rc_rcv_tid_rdma_ack(struct hfi1_packet *packet);
301 
302 void hfi1_add_tid_retry_timer(struct rvt_qp *qp);
303 void hfi1_del_tid_retry_timer(struct rvt_qp *qp);
304 
305 u32 hfi1_build_tid_rdma_resync(struct rvt_qp *qp, struct rvt_swqe *wqe,
306 			       struct ib_other_headers *ohdr, u32 *bth1,
307 			       u32 *bth2, u16 fidx);
308 
309 void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet);
310 
311 struct hfi1_pkt_state;
312 int hfi1_make_tid_rdma_pkt(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
313 
314 void _hfi1_do_tid_send(struct work_struct *work);
315 
316 bool hfi1_schedule_tid_send(struct rvt_qp *qp);
317 
318 bool hfi1_tid_rdma_ack_interlock(struct rvt_qp *qp, struct rvt_ack_entry *e);
319 
320 #endif /* HFI1_TID_RDMA_H */
321