event/octeontx2: use WFE while waiting for head
[dpdk.git] / drivers / event / octeontx2 / otx2_worker.h
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(C) 2019 Marvell International Ltd.
3  */
4
5 #ifndef __OTX2_WORKER_H__
6 #define __OTX2_WORKER_H__
7
8 #include <rte_common.h>
9 #include <rte_branch_prediction.h>
10
11 #include <otx2_common.h>
12 #include "otx2_evdev.h"
13
14 /* SSO Operations */
15
16 static __rte_always_inline uint16_t
17 otx2_ssogws_get_work(struct otx2_ssogws *ws, struct rte_event *ev,
18                      const uint32_t flags, const void * const lookup_mem)
19 {
20         union otx2_sso_event event;
21         uint64_t tstamp_ptr;
22         uint64_t get_work1;
23         uint64_t mbuf;
24
25         otx2_write64(BIT_ULL(16) | /* wait for work. */
26                      1, /* Use Mask set 0. */
27                      ws->getwrk_op);
28
29         if (flags & NIX_RX_OFFLOAD_PTYPE_F)
30                 rte_prefetch_non_temporal(lookup_mem);
31 #ifdef RTE_ARCH_ARM64
32         asm volatile(
33                         "               ldr %[tag], [%[tag_loc]]        \n"
34                         "               ldr %[wqp], [%[wqp_loc]]        \n"
35                         "               tbz %[tag], 63, done%=          \n"
36                         "               sevl                            \n"
37                         "rty%=:         wfe                             \n"
38                         "               ldr %[tag], [%[tag_loc]]        \n"
39                         "               ldr %[wqp], [%[wqp_loc]]        \n"
40                         "               tbnz %[tag], 63, rty%=          \n"
41                         "done%=:        dmb ld                          \n"
42                         "               prfm pldl1keep, [%[wqp], #8]    \n"
43                         "               sub %[mbuf], %[wqp], #0x80      \n"
44                         "               prfm pldl1keep, [%[mbuf]]       \n"
45                         : [tag] "=&r" (event.get_work0),
46                           [wqp] "=&r" (get_work1),
47                           [mbuf] "=&r" (mbuf)
48                         : [tag_loc] "r" (ws->tag_op),
49                           [wqp_loc] "r" (ws->wqp_op)
50                         );
51 #else
52         event.get_work0 = otx2_read64(ws->tag_op);
53         while ((BIT_ULL(63)) & event.get_work0)
54                 event.get_work0 = otx2_read64(ws->tag_op);
55
56         get_work1 = otx2_read64(ws->wqp_op);
57         rte_prefetch0((const void *)get_work1);
58         mbuf = (uint64_t)((char *)get_work1 - sizeof(struct rte_mbuf));
59         rte_prefetch0((const void *)mbuf);
60 #endif
61
62         event.get_work0 = (event.get_work0 & (0x3ull << 32)) << 6 |
63                 (event.get_work0 & (0x3FFull << 36)) << 4 |
64                 (event.get_work0 & 0xffffffff);
65         ws->cur_tt = event.sched_type;
66         ws->cur_grp = event.queue_id;
67
68         if (event.sched_type != SSO_TT_EMPTY &&
69             event.event_type == RTE_EVENT_TYPE_ETHDEV) {
70                 otx2_wqe_to_mbuf(get_work1, mbuf, event.sub_event_type,
71                                  (uint32_t) event.get_work0, flags, lookup_mem);
72                 /* Extracting tstamp, if PTP enabled*/
73                 tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)get_work1)
74                                              + OTX2_SSO_WQE_SG_PTR);
75                 otx2_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, ws->tstamp,
76                                         flags, (uint64_t *)tstamp_ptr);
77                 get_work1 = mbuf;
78         }
79
80         ev->event = event.get_work0;
81         ev->u64 = get_work1;
82
83         return !!get_work1;
84 }
85
86 /* Used in cleaning up workslot. */
87 static __rte_always_inline uint16_t
88 otx2_ssogws_get_work_empty(struct otx2_ssogws *ws, struct rte_event *ev,
89                            const uint32_t flags)
90 {
91         union otx2_sso_event event;
92         uint64_t tstamp_ptr;
93         uint64_t get_work1;
94         uint64_t mbuf;
95
96 #ifdef RTE_ARCH_ARM64
97         asm volatile(
98                         "               ldr %[tag], [%[tag_loc]]        \n"
99                         "               ldr %[wqp], [%[wqp_loc]]        \n"
100                         "               tbz %[tag], 63, done%=          \n"
101                         "               sevl                            \n"
102                         "rty%=:         wfe                             \n"
103                         "               ldr %[tag], [%[tag_loc]]        \n"
104                         "               ldr %[wqp], [%[wqp_loc]]        \n"
105                         "               tbnz %[tag], 63, rty%=          \n"
106                         "done%=:        dmb ld                          \n"
107                         "               prfm pldl1keep, [%[wqp], #8]    \n"
108                         "               sub %[mbuf], %[wqp], #0x80      \n"
109                         "               prfm pldl1keep, [%[mbuf]]       \n"
110                         : [tag] "=&r" (event.get_work0),
111                           [wqp] "=&r" (get_work1),
112                           [mbuf] "=&r" (mbuf)
113                         : [tag_loc] "r" (ws->tag_op),
114                           [wqp_loc] "r" (ws->wqp_op)
115                         );
116 #else
117         event.get_work0 = otx2_read64(ws->tag_op);
118         while ((BIT_ULL(63)) & event.get_work0)
119                 event.get_work0 = otx2_read64(ws->tag_op);
120
121         get_work1 = otx2_read64(ws->wqp_op);
122         rte_prefetch_non_temporal((const void *)get_work1);
123         mbuf = (uint64_t)((char *)get_work1 - sizeof(struct rte_mbuf));
124         rte_prefetch_non_temporal((const void *)mbuf);
125 #endif
126
127         event.get_work0 = (event.get_work0 & (0x3ull << 32)) << 6 |
128                 (event.get_work0 & (0x3FFull << 36)) << 4 |
129                 (event.get_work0 & 0xffffffff);
130         ws->cur_tt = event.sched_type;
131         ws->cur_grp = event.queue_id;
132
133         if (event.sched_type != SSO_TT_EMPTY &&
134             event.event_type == RTE_EVENT_TYPE_ETHDEV) {
135                 otx2_wqe_to_mbuf(get_work1, mbuf, event.sub_event_type,
136                                  (uint32_t) event.get_work0, flags, NULL);
137                 /* Extracting tstamp, if PTP enabled*/
138                 tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)get_work1)
139                                              + OTX2_SSO_WQE_SG_PTR);
140                 otx2_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, ws->tstamp,
141                                         flags, (uint64_t *)tstamp_ptr);
142                 get_work1 = mbuf;
143         }
144
145         ev->event = event.get_work0;
146         ev->u64 = get_work1;
147
148         return !!get_work1;
149 }
150
151 static __rte_always_inline void
152 otx2_ssogws_add_work(struct otx2_ssogws *ws, const uint64_t event_ptr,
153                      const uint32_t tag, const uint8_t new_tt,
154                      const uint16_t grp)
155 {
156         uint64_t add_work0;
157
158         add_work0 = tag | ((uint64_t)(new_tt) << 32);
159         otx2_store_pair(add_work0, event_ptr, ws->grps_base[grp]);
160 }
161
162 static __rte_always_inline void
163 otx2_ssogws_swtag_desched(struct otx2_ssogws *ws, uint32_t tag, uint8_t new_tt,
164                           uint16_t grp)
165 {
166         uint64_t val;
167
168         val = tag | ((uint64_t)(new_tt & 0x3) << 32) | ((uint64_t)grp << 34);
169         otx2_write64(val, ws->swtag_desched_op);
170 }
171
172 static __rte_always_inline void
173 otx2_ssogws_swtag_norm(struct otx2_ssogws *ws, uint32_t tag, uint8_t new_tt)
174 {
175         uint64_t val;
176
177         val = tag | ((uint64_t)(new_tt & 0x3) << 32);
178         otx2_write64(val, ws->swtag_norm_op);
179 }
180
181 static __rte_always_inline void
182 otx2_ssogws_swtag_untag(struct otx2_ssogws *ws)
183 {
184         otx2_write64(0, OTX2_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
185                      SSOW_LF_GWS_OP_SWTAG_UNTAG);
186         ws->cur_tt = SSO_SYNC_UNTAGGED;
187 }
188
189 static __rte_always_inline void
190 otx2_ssogws_swtag_flush(struct otx2_ssogws *ws)
191 {
192         otx2_write64(0, OTX2_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
193                      SSOW_LF_GWS_OP_SWTAG_FLUSH);
194         ws->cur_tt = SSO_SYNC_EMPTY;
195 }
196
197 static __rte_always_inline void
198 otx2_ssogws_desched(struct otx2_ssogws *ws)
199 {
200         otx2_write64(0, OTX2_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
201                      SSOW_LF_GWS_OP_DESCHED);
202 }
203
204 static __rte_always_inline void
205 otx2_ssogws_swtag_wait(struct otx2_ssogws *ws)
206 {
207 #ifdef RTE_ARCH_ARM64
208         uint64_t swtp;
209
210         asm volatile (
211                         "       ldr %[swtb], [%[swtp_loc]]      \n"
212                         "       cbz %[swtb], done%=             \n"
213                         "       sevl                            \n"
214                         "rty%=: wfe                             \n"
215                         "       ldr %[swtb], [%[swtp_loc]]      \n"
216                         "       cbnz %[swtb], rty%=             \n"
217                         "done%=:                                \n"
218                         : [swtb] "=&r" (swtp)
219                         : [swtp_loc] "r" (ws->swtp_op)
220                         );
221 #else
222         /* Wait for the SWTAG/SWTAG_FULL operation */
223         while (otx2_read64(ws->swtp_op))
224                 ;
225 #endif
226 }
227
228 static __rte_always_inline void
229 otx2_ssogws_head_wait(struct otx2_ssogws *ws)
230 {
231 #ifdef RTE_ARCH_ARM64
232         uint64_t tag;
233
234         asm volatile (
235                         "       ldr %[tag], [%[tag_op]]         \n"
236                         "       tbnz %[tag], 35, done%=         \n"
237                         "       sevl                            \n"
238                         "rty%=: wfe                             \n"
239                         "       ldr %[tag], [%[tag_op]]         \n"
240                         "       tbz %[tag], 35, rty%=           \n"
241                         "done%=:                                \n"
242                         : [tag] "=&r" (tag)
243                         : [tag_op] "r" (ws->tag_op)
244                         );
245 #else
246         /* Wait for the HEAD to be set */
247         while (!(otx2_read64(ws->tag_op) & BIT_ULL(35)))
248                 ;
249 #endif
250 }
251
252 static __rte_always_inline void
253 otx2_ssogws_order(struct otx2_ssogws *ws, const uint8_t wait_flag)
254 {
255         if (wait_flag)
256                 otx2_ssogws_head_wait(ws);
257
258         rte_cio_wmb();
259 }
260
261 static __rte_always_inline const struct otx2_eth_txq *
262 otx2_ssogws_xtract_meta(struct rte_mbuf *m)
263 {
264         return rte_eth_devices[m->port].data->tx_queues[
265                         rte_event_eth_tx_adapter_txq_get(m)];
266 }
267
268 static __rte_always_inline void
269 otx2_ssogws_prepare_pkt(const struct otx2_eth_txq *txq, struct rte_mbuf *m,
270                         uint64_t *cmd, const uint32_t flags)
271 {
272         otx2_lmt_mov(cmd, txq->cmd, otx2_nix_tx_ext_subs(flags));
273         otx2_nix_xmit_prepare(m, cmd, flags);
274 }
275
276 static __rte_always_inline uint16_t
277 otx2_ssogws_event_tx(struct otx2_ssogws *ws, struct rte_event ev[],
278                      uint64_t *cmd, const uint32_t flags)
279 {
280         struct rte_mbuf *m = ev[0].mbuf;
281         const struct otx2_eth_txq *txq = otx2_ssogws_xtract_meta(m);
282
283         /* Perform header writes before barrier for TSO */
284         otx2_nix_xmit_prepare_tso(m, flags);
285         otx2_ssogws_order(ws, !ev->sched_type);
286         otx2_ssogws_prepare_pkt(txq, m, cmd, flags);
287
288         if (flags & NIX_TX_MULTI_SEG_F) {
289                 const uint16_t segdw = otx2_nix_prepare_mseg(m, cmd, flags);
290                 otx2_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
291                                              m->ol_flags, segdw, flags);
292                 otx2_nix_xmit_mseg_one(cmd, txq->lmt_addr, txq->io_addr, segdw);
293         } else {
294                 /* Passing no of segdw as 4: HDR + EXT + SG + SMEM */
295                 otx2_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
296                                              m->ol_flags, 4, flags);
297                 otx2_nix_xmit_one(cmd, txq->lmt_addr, txq->io_addr, flags);
298         }
299
300         return 1;
301 }
302
303 #endif