]>
git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - net/smc/smc_tx.c
1 // SPDX-License-Identifier: GPL-2.0
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
7 * Copy user space data into send buffer, if send buffer space available.
9 * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available.
11 * Copyright IBM Corp. 2016
13 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
16 #include <linux/net.h>
17 #include <linux/rcupdate.h>
18 #include <linux/workqueue.h>
19 #include <linux/sched/signal.h>
27 #include "smc_close.h"
31 #define SMC_TX_WORK_DELAY 0
32 #define SMC_TX_CORK_DELAY (HZ >> 2) /* 250 ms */
34 /***************************** sndbuf producer *******************************/
36 /* callback implementation for sk.sk_write_space()
37 * to wakeup sndbuf producers that blocked with smc_tx_wait().
38 * called under sk_socket lock.
40 static void smc_tx_write_space(struct sock
*sk
)
42 struct socket
*sock
= sk
->sk_socket
;
43 struct smc_sock
*smc
= smc_sk(sk
);
46 /* similar to sk_stream_write_space */
47 if (atomic_read(&smc
->conn
.sndbuf_space
) && sock
) {
48 clear_bit(SOCK_NOSPACE
, &sock
->flags
);
50 wq
= rcu_dereference(sk
->sk_wq
);
51 if (skwq_has_sleeper(wq
))
52 wake_up_interruptible_poll(&wq
->wait
,
53 EPOLLOUT
| EPOLLWRNORM
|
55 if (wq
&& wq
->fasync_list
&& !(sk
->sk_shutdown
& SEND_SHUTDOWN
))
56 sock_wake_async(wq
, SOCK_WAKE_SPACE
, POLL_OUT
);
61 /* Wakeup sndbuf producers that blocked with smc_tx_wait().
62 * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
64 void smc_tx_sndbuf_nonfull(struct smc_sock
*smc
)
66 if (smc
->sk
.sk_socket
&&
67 test_bit(SOCK_NOSPACE
, &smc
->sk
.sk_socket
->flags
))
68 smc
->sk
.sk_write_space(&smc
->sk
);
71 /* blocks sndbuf producer until at least one byte of free space available
72 * or urgent Byte was consumed
74 static int smc_tx_wait(struct smc_sock
*smc
, int flags
)
76 DEFINE_WAIT_FUNC(wait
, woken_wake_function
);
77 struct smc_connection
*conn
= &smc
->conn
;
78 struct sock
*sk
= &smc
->sk
;
83 /* similar to sk_stream_wait_memory */
84 timeo
= sock_sndtimeo(sk
, flags
& MSG_DONTWAIT
);
85 noblock
= timeo
? false : true;
86 add_wait_queue(sk_sleep(sk
), &wait
);
88 sk_set_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
90 (sk
->sk_shutdown
& SEND_SHUTDOWN
) ||
91 conn
->local_tx_ctrl
.conn_state_flags
.peer_done_writing
) {
95 if (smc_cdc_rxed_any_close(conn
)) {
101 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
105 if (signal_pending(current
)) {
106 rc
= sock_intr_errno(timeo
);
109 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
110 if (atomic_read(&conn
->sndbuf_space
) && !conn
->urg_tx_pend
)
111 break; /* at least 1 byte of free & no urgent data */
112 set_bit(SOCK_NOSPACE
, &sk
->sk_socket
->flags
);
113 sk_wait_event(sk
, &timeo
,
115 (sk
->sk_shutdown
& SEND_SHUTDOWN
) ||
116 smc_cdc_rxed_any_close(conn
) ||
117 (atomic_read(&conn
->sndbuf_space
) &&
121 remove_wait_queue(sk_sleep(sk
), &wait
);
125 static bool smc_tx_is_corked(struct smc_sock
*smc
)
127 struct tcp_sock
*tp
= tcp_sk(smc
->clcsock
->sk
);
129 return (tp
->nonagle
& TCP_NAGLE_CORK
) ? true : false;
132 /* sndbuf producer: main API called by socket layer.
133 * called under sock lock.
135 int smc_tx_sendmsg(struct smc_sock
*smc
, struct msghdr
*msg
, size_t len
)
137 size_t copylen
, send_done
= 0, send_remaining
= len
;
138 size_t chunk_len
, chunk_off
, chunk_len_sum
;
139 struct smc_connection
*conn
= &smc
->conn
;
140 union smc_host_cursor prep
;
141 struct sock
*sk
= &smc
->sk
;
147 /* This should be in poll */
148 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE
, sk
);
150 if (sk
->sk_err
|| (sk
->sk_shutdown
& SEND_SHUTDOWN
)) {
155 while (msg_data_left(msg
)) {
156 if (sk
->sk_state
== SMC_INIT
)
158 if (smc
->sk
.sk_shutdown
& SEND_SHUTDOWN
||
159 (smc
->sk
.sk_err
== ECONNABORTED
) ||
160 conn
->local_tx_ctrl
.conn_state_flags
.peer_conn_abort
)
162 if (smc_cdc_rxed_any_close(conn
))
163 return send_done
?: -ECONNRESET
;
165 if (msg
->msg_flags
& MSG_OOB
)
166 conn
->local_tx_ctrl
.prod_flags
.urg_data_pending
= 1;
168 if (!atomic_read(&conn
->sndbuf_space
) || conn
->urg_tx_pend
) {
171 rc
= smc_tx_wait(smc
, msg
->msg_flags
);
177 /* initialize variables for 1st iteration of subsequent loop */
178 /* could be just 1 byte, even after smc_tx_wait above */
179 writespace
= atomic_read(&conn
->sndbuf_space
);
180 /* not more than what user space asked for */
181 copylen
= min_t(size_t, send_remaining
, writespace
);
182 /* determine start of sndbuf */
183 sndbuf_base
= conn
->sndbuf_desc
->cpu_addr
;
184 smc_curs_copy(&prep
, &conn
->tx_curs_prep
, conn
);
185 tx_cnt_prep
= prep
.count
;
186 /* determine chunks where to write into sndbuf */
187 /* either unwrapped case, or 1st chunk of wrapped case */
188 chunk_len
= min_t(size_t, copylen
, conn
->sndbuf_desc
->len
-
190 chunk_len_sum
= chunk_len
;
191 chunk_off
= tx_cnt_prep
;
192 smc_sndbuf_sync_sg_for_cpu(conn
);
193 for (chunk
= 0; chunk
< 2; chunk
++) {
194 rc
= memcpy_from_msg(sndbuf_base
+ chunk_off
,
197 smc_sndbuf_sync_sg_for_device(conn
);
202 send_done
+= chunk_len
;
203 send_remaining
-= chunk_len
;
205 if (chunk_len_sum
== copylen
)
206 break; /* either on 1st or 2nd iteration */
207 /* prepare next (== 2nd) iteration */
208 chunk_len
= copylen
- chunk_len
; /* remainder */
209 chunk_len_sum
+= chunk_len
;
210 chunk_off
= 0; /* modulo offset in send ring buffer */
212 smc_sndbuf_sync_sg_for_device(conn
);
214 smc_curs_add(conn
->sndbuf_desc
->len
, &prep
, copylen
);
215 smc_curs_copy(&conn
->tx_curs_prep
, &prep
, conn
);
216 /* increased in send tasklet smc_cdc_tx_handler() */
217 smp_mb__before_atomic();
218 atomic_sub(copylen
, &conn
->sndbuf_space
);
219 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
220 smp_mb__after_atomic();
221 /* since we just produced more new data into sndbuf,
222 * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
224 if ((msg
->msg_flags
& MSG_OOB
) && !send_remaining
)
225 conn
->urg_tx_pend
= true;
226 if ((msg
->msg_flags
& MSG_MORE
|| smc_tx_is_corked(smc
)) &&
227 (atomic_read(&conn
->sndbuf_space
) >
228 (conn
->sndbuf_desc
->len
>> 1)))
229 /* for a corked socket defer the RDMA writes if there
230 * is still sufficient sndbuf_space available
232 schedule_delayed_work(&conn
->tx_work
,
235 smc_tx_sndbuf_nonempty(conn
);
236 } /* while (msg_data_left(msg)) */
241 rc
= sk_stream_error(sk
, msg
->msg_flags
, rc
);
242 /* make sure we wake any epoll edge trigger waiter */
243 if (unlikely(rc
== -EAGAIN
))
244 sk
->sk_write_space(sk
);
248 /***************************** sndbuf consumer *******************************/
250 /* sndbuf consumer: actual data transfer of one target chunk with ISM write */
251 int smcd_tx_ism_write(struct smc_connection
*conn
, void *data
, size_t len
,
252 u32 offset
, int signal
)
254 struct smc_ism_position pos
;
257 memset(&pos
, 0, sizeof(pos
));
258 pos
.token
= conn
->peer_token
;
259 pos
.index
= conn
->peer_rmbe_idx
;
260 pos
.offset
= conn
->tx_off
+ offset
;
262 rc
= smc_ism_write(conn
->lgr
->smcd
, &pos
, data
, len
);
264 conn
->local_tx_ctrl
.conn_state_flags
.peer_conn_abort
= 1;
268 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
269 static int smc_tx_rdma_write(struct smc_connection
*conn
, int peer_rmbe_offset
,
270 int num_sges
, struct ib_rdma_wr
*rdma_wr
)
272 struct smc_link_group
*lgr
= conn
->lgr
;
273 struct smc_link
*link
;
276 link
= &lgr
->lnk
[SMC_SINGLE_LINK
];
277 rdma_wr
->wr
.wr_id
= smc_wr_tx_get_next_wr_id(link
);
278 rdma_wr
->wr
.num_sge
= num_sges
;
279 rdma_wr
->remote_addr
=
280 lgr
->rtokens
[conn
->rtoken_idx
][SMC_SINGLE_LINK
].dma_addr
+
281 /* RMBE within RMB */
283 /* offset within RMBE */
285 rdma_wr
->rkey
= lgr
->rtokens
[conn
->rtoken_idx
][SMC_SINGLE_LINK
].rkey
;
286 rc
= ib_post_send(link
->roce_qp
, &rdma_wr
->wr
, NULL
);
288 conn
->local_tx_ctrl
.conn_state_flags
.peer_conn_abort
= 1;
289 smc_lgr_terminate(lgr
);
294 /* sndbuf consumer */
295 static inline void smc_tx_advance_cursors(struct smc_connection
*conn
,
296 union smc_host_cursor
*prod
,
297 union smc_host_cursor
*sent
,
300 smc_curs_add(conn
->peer_rmbe_size
, prod
, len
);
301 /* increased in recv tasklet smc_cdc_msg_rcv() */
302 smp_mb__before_atomic();
303 /* data in flight reduces usable snd_wnd */
304 atomic_sub(len
, &conn
->peer_rmbe_space
);
305 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
306 smp_mb__after_atomic();
307 smc_curs_add(conn
->sndbuf_desc
->len
, sent
, len
);
310 /* SMC-R helper for smc_tx_rdma_writes() */
311 static int smcr_tx_rdma_writes(struct smc_connection
*conn
, size_t len
,
312 size_t src_off
, size_t src_len
,
313 size_t dst_off
, size_t dst_len
,
314 struct smc_rdma_wr
*wr_rdma_buf
)
316 dma_addr_t dma_addr
=
317 sg_dma_address(conn
->sndbuf_desc
->sgt
[SMC_SINGLE_LINK
].sgl
);
318 int src_len_sum
= src_len
, dst_len_sum
= dst_len
;
319 int sent_count
= src_off
;
320 int srcchunk
, dstchunk
;
324 for (dstchunk
= 0; dstchunk
< 2; dstchunk
++) {
326 wr_rdma_buf
->wr_tx_rdma
[dstchunk
].wr
.sg_list
;
329 for (srcchunk
= 0; srcchunk
< 2; srcchunk
++) {
330 sge
[srcchunk
].addr
= dma_addr
+ src_off
;
331 sge
[srcchunk
].length
= src_len
;
335 if (src_off
>= conn
->sndbuf_desc
->len
)
336 src_off
-= conn
->sndbuf_desc
->len
;
337 /* modulo in send ring */
338 if (src_len_sum
== dst_len
)
339 break; /* either on 1st or 2nd iteration */
340 /* prepare next (== 2nd) iteration */
341 src_len
= dst_len
- src_len
; /* remainder */
342 src_len_sum
+= src_len
;
344 rc
= smc_tx_rdma_write(conn
, dst_off
, num_sges
,
345 &wr_rdma_buf
->wr_tx_rdma
[dstchunk
]);
348 if (dst_len_sum
== len
)
349 break; /* either on 1st or 2nd iteration */
350 /* prepare next (== 2nd) iteration */
351 dst_off
= 0; /* modulo offset in RMBE ring buffer */
352 dst_len
= len
- dst_len
; /* remainder */
353 dst_len_sum
+= dst_len
;
354 src_len
= min_t(int, dst_len
, conn
->sndbuf_desc
->len
-
356 src_len_sum
= src_len
;
361 /* SMC-D helper for smc_tx_rdma_writes() */
362 static int smcd_tx_rdma_writes(struct smc_connection
*conn
, size_t len
,
363 size_t src_off
, size_t src_len
,
364 size_t dst_off
, size_t dst_len
)
366 int src_len_sum
= src_len
, dst_len_sum
= dst_len
;
367 int srcchunk
, dstchunk
;
370 for (dstchunk
= 0; dstchunk
< 2; dstchunk
++) {
371 for (srcchunk
= 0; srcchunk
< 2; srcchunk
++) {
372 void *data
= conn
->sndbuf_desc
->cpu_addr
+ src_off
;
374 rc
= smcd_tx_ism_write(conn
, data
, src_len
, dst_off
+
375 sizeof(struct smcd_cdc_msg
), 0);
380 if (src_off
>= conn
->sndbuf_desc
->len
)
381 src_off
-= conn
->sndbuf_desc
->len
;
382 /* modulo in send ring */
383 if (src_len_sum
== dst_len
)
384 break; /* either on 1st or 2nd iteration */
385 /* prepare next (== 2nd) iteration */
386 src_len
= dst_len
- src_len
; /* remainder */
387 src_len_sum
+= src_len
;
389 if (dst_len_sum
== len
)
390 break; /* either on 1st or 2nd iteration */
391 /* prepare next (== 2nd) iteration */
392 dst_off
= 0; /* modulo offset in RMBE ring buffer */
393 dst_len
= len
- dst_len
; /* remainder */
394 dst_len_sum
+= dst_len
;
395 src_len
= min_t(int, dst_len
, conn
->sndbuf_desc
->len
- src_off
);
396 src_len_sum
= src_len
;
401 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
402 * usable snd_wnd as max transmit
404 static int smc_tx_rdma_writes(struct smc_connection
*conn
,
405 struct smc_rdma_wr
*wr_rdma_buf
)
407 size_t len
, src_len
, dst_off
, dst_len
; /* current chunk values */
408 union smc_host_cursor sent
, prep
, prod
, cons
;
409 struct smc_cdc_producer_flags
*pflags
;
410 int to_send
, rmbespace
;
414 smc_curs_copy(&sent
, &conn
->tx_curs_sent
, conn
);
415 smc_curs_copy(&prep
, &conn
->tx_curs_prep
, conn
);
416 /* cf. wmem_alloc - (snd_max - snd_una) */
417 to_send
= smc_curs_diff(conn
->sndbuf_desc
->len
, &sent
, &prep
);
421 /* destination: RMBE */
423 rmbespace
= atomic_read(&conn
->peer_rmbe_space
);
426 smc_curs_copy(&prod
, &conn
->local_tx_ctrl
.prod
, conn
);
427 smc_curs_copy(&cons
, &conn
->local_rx_ctrl
.cons
, conn
);
429 /* if usable snd_wnd closes ask peer to advertise once it opens again */
430 pflags
= &conn
->local_tx_ctrl
.prod_flags
;
431 pflags
->write_blocked
= (to_send
>= rmbespace
);
432 /* cf. usable snd_wnd */
433 len
= min(to_send
, rmbespace
);
435 /* initialize variables for first iteration of subsequent nested loop */
436 dst_off
= prod
.count
;
437 if (prod
.wrap
== cons
.wrap
) {
438 /* the filled destination area is unwrapped,
439 * hence the available free destination space is wrapped
440 * and we need 2 destination chunks of sum len; start with 1st
441 * which is limited by what's available in sndbuf
443 dst_len
= min_t(size_t,
444 conn
->peer_rmbe_size
- prod
.count
, len
);
446 /* the filled destination area is wrapped,
447 * hence the available free destination space is unwrapped
448 * and we need a single destination chunk of entire len
452 /* dst_len determines the maximum src_len */
453 if (sent
.count
+ dst_len
<= conn
->sndbuf_desc
->len
) {
454 /* unwrapped src case: single chunk of entire dst_len */
457 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
458 src_len
= conn
->sndbuf_desc
->len
- sent
.count
;
461 if (conn
->lgr
->is_smcd
)
462 rc
= smcd_tx_rdma_writes(conn
, len
, sent
.count
, src_len
,
465 rc
= smcr_tx_rdma_writes(conn
, len
, sent
.count
, src_len
,
466 dst_off
, dst_len
, wr_rdma_buf
);
470 if (conn
->urg_tx_pend
&& len
== to_send
)
471 pflags
->urg_data_present
= 1;
472 smc_tx_advance_cursors(conn
, &prod
, &sent
, len
);
473 /* update connection's cursors with advanced local cursors */
474 smc_curs_copy(&conn
->local_tx_ctrl
.prod
, &prod
, conn
);
476 smc_curs_copy(&conn
->tx_curs_sent
, &sent
, conn
);/* src: local sndbuf */
481 /* Wakeup sndbuf consumers from any context (IRQ or process)
482 * since there is more data to transmit; usable snd_wnd as max transmit
484 static int smcr_tx_sndbuf_nonempty(struct smc_connection
*conn
)
486 struct smc_cdc_producer_flags
*pflags
= &conn
->local_tx_ctrl
.prod_flags
;
487 struct smc_rdma_wr
*wr_rdma_buf
;
488 struct smc_cdc_tx_pend
*pend
;
489 struct smc_wr_buf
*wr_buf
;
492 rc
= smc_cdc_get_free_slot(conn
, &wr_buf
, &wr_rdma_buf
, &pend
);
495 struct smc_sock
*smc
=
496 container_of(conn
, struct smc_sock
, conn
);
498 if (smc
->sk
.sk_err
== ECONNABORTED
)
499 return sock_error(&smc
->sk
);
501 if (conn
->alert_token_local
) /* connection healthy */
502 mod_delayed_work(system_wq
, &conn
->tx_work
,
508 spin_lock_bh(&conn
->send_lock
);
509 if (!pflags
->urg_data_present
) {
510 rc
= smc_tx_rdma_writes(conn
, wr_rdma_buf
);
512 smc_wr_tx_put_slot(&conn
->lgr
->lnk
[SMC_SINGLE_LINK
],
513 (struct smc_wr_tx_pend_priv
*)pend
);
518 rc
= smc_cdc_msg_send(conn
, wr_buf
, pend
);
519 if (!rc
&& pflags
->urg_data_present
) {
520 pflags
->urg_data_pending
= 0;
521 pflags
->urg_data_present
= 0;
525 spin_unlock_bh(&conn
->send_lock
);
529 static int smcd_tx_sndbuf_nonempty(struct smc_connection
*conn
)
531 struct smc_cdc_producer_flags
*pflags
= &conn
->local_tx_ctrl
.prod_flags
;
534 spin_lock_bh(&conn
->send_lock
);
535 if (!pflags
->urg_data_present
)
536 rc
= smc_tx_rdma_writes(conn
, NULL
);
538 rc
= smcd_cdc_msg_send(conn
);
540 if (!rc
&& pflags
->urg_data_present
) {
541 pflags
->urg_data_pending
= 0;
542 pflags
->urg_data_present
= 0;
544 spin_unlock_bh(&conn
->send_lock
);
548 int smc_tx_sndbuf_nonempty(struct smc_connection
*conn
)
552 if (conn
->lgr
->is_smcd
)
553 rc
= smcd_tx_sndbuf_nonempty(conn
);
555 rc
= smcr_tx_sndbuf_nonempty(conn
);
558 /* trigger socket release if connection is closing */
559 struct smc_sock
*smc
= container_of(conn
, struct smc_sock
,
561 smc_close_wake_tx_prepared(smc
);
566 /* Wakeup sndbuf consumers from process context
567 * since there is more data to transmit
569 void smc_tx_work(struct work_struct
*work
)
571 struct smc_connection
*conn
= container_of(to_delayed_work(work
),
572 struct smc_connection
,
574 struct smc_sock
*smc
= container_of(conn
, struct smc_sock
, conn
);
578 if (smc
->sk
.sk_err
||
579 !conn
->alert_token_local
||
580 conn
->local_rx_ctrl
.conn_state_flags
.peer_conn_abort
)
583 rc
= smc_tx_sndbuf_nonempty(conn
);
584 if (!rc
&& conn
->local_rx_ctrl
.prod_flags
.write_blocked
&&
585 !atomic_read(&conn
->bytes_to_rcv
))
586 conn
->local_rx_ctrl
.prod_flags
.write_blocked
= 0;
589 release_sock(&smc
->sk
);
592 void smc_tx_consumer_update(struct smc_connection
*conn
, bool force
)
594 union smc_host_cursor cfed
, cons
, prod
;
595 int sender_free
= conn
->rmb_desc
->len
;
598 smc_curs_copy(&cons
, &conn
->local_tx_ctrl
.cons
, conn
);
599 smc_curs_copy(&cfed
, &conn
->rx_curs_confirmed
, conn
);
600 to_confirm
= smc_curs_diff(conn
->rmb_desc
->len
, &cfed
, &cons
);
601 if (to_confirm
> conn
->rmbe_update_limit
) {
602 smc_curs_copy(&prod
, &conn
->local_rx_ctrl
.prod
, conn
);
603 sender_free
= conn
->rmb_desc
->len
-
604 smc_curs_diff_large(conn
->rmb_desc
->len
,
608 if (conn
->local_rx_ctrl
.prod_flags
.cons_curs_upd_req
||
610 ((to_confirm
> conn
->rmbe_update_limit
) &&
611 ((sender_free
<= (conn
->rmb_desc
->len
/ 2)) ||
612 conn
->local_rx_ctrl
.prod_flags
.write_blocked
))) {
613 if ((smc_cdc_get_slot_and_msg_send(conn
) < 0) &&
614 conn
->alert_token_local
) { /* connection healthy */
615 schedule_delayed_work(&conn
->tx_work
,
620 if (conn
->local_rx_ctrl
.prod_flags
.write_blocked
&&
621 !atomic_read(&conn
->bytes_to_rcv
))
622 conn
->local_rx_ctrl
.prod_flags
.write_blocked
= 0;
625 /***************************** send initialize *******************************/
627 /* Initialize send properties on connection establishment. NB: not __init! */
628 void smc_tx_init(struct smc_sock
*smc
)
630 smc
->sk
.sk_write_space
= smc_tx_write_space
;