Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. | |
3 | * | |
4 | * Copyright (c) 2011, 2012, Intel Corporation. | |
5 | * | |
6 | * Author: Zach Brown <zab@zabbo.net> | |
7 | * Author: Peter J. Braam <braam@clusterfs.com> | |
8 | * Author: Phil Schwan <phil@clusterfs.com> | |
9 | * Author: Eric Barton <eric@bartonsoftware.com> | |
10 | * | |
11 | * This file is part of Lustre, http://www.lustre.org | |
12 | * | |
13 | * Portals is free software; you can redistribute it and/or | |
14 | * modify it under the terms of version 2 of the GNU General Public | |
15 | * License as published by the Free Software Foundation. | |
16 | * | |
17 | * Portals is distributed in the hope that it will be useful, | |
18 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
19 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
20 | * GNU General Public License for more details. | |
21 | * | |
d7e09d03 PT |
22 | */ |
23 | ||
12c41f00 JH |
24 | #ifndef _SOCKLND_SOCKLND_H_ |
25 | #define _SOCKLND_SOCKLND_H_ | |
26 | ||
d7e09d03 PT |
27 | #define DEBUG_PORTAL_ALLOC |
28 | #define DEBUG_SUBSYSTEM S_LND | |
29 | ||
12c41f00 JH |
30 | #include <linux/crc32.h> |
31 | #include <linux/errno.h> | |
32 | #include <linux/if.h> | |
33 | #include <linux/init.h> | |
34 | #include <linux/kernel.h> | |
35 | #include <linux/kmod.h> | |
36 | #include <linux/list.h> | |
37 | #include <linux/mm.h> | |
38 | #include <linux/module.h> | |
39 | #include <linux/stat.h> | |
40 | #include <linux/string.h> | |
41 | #include <linux/syscalls.h> | |
42 | #include <linux/sysctl.h> | |
43 | #include <linux/uio.h> | |
44 | #include <linux/unistd.h> | |
a82de7ea | 45 | #include <asm/irq.h> |
12c41f00 JH |
46 | #include <net/sock.h> |
47 | #include <net/tcp.h> | |
d7e09d03 | 48 | |
9fdaf8c0 GKH |
49 | #include "../../../include/linux/libcfs/libcfs.h" |
50 | #include "../../../include/linux/lnet/lnet.h" | |
51 | #include "../../../include/linux/lnet/lib-lnet.h" | |
52 | #include "../../../include/linux/lnet/socklnd.h" | |
d7e09d03 | 53 | |
12c41f00 JH |
54 | /* assume one thread for each connection type */ |
55 | #define SOCKNAL_NSCHEDS 3 | |
56 | #define SOCKNAL_NSCHEDS_HIGH (SOCKNAL_NSCHEDS << 1) | |
57 | ||
97d10d0a MS |
58 | #define SOCKNAL_PEER_HASH_SIZE 101 /* # peer lists */ |
59 | #define SOCKNAL_RESCHED 100 /* # scheduler loops before reschedule */ | |
60 | #define SOCKNAL_INSANITY_RECONN 5000 /* connd is trying on reconn infinitely */ | |
61 | #define SOCKNAL_ENOMEM_RETRY CFS_TICK /* jiffies between retries */ | |
d7e09d03 | 62 | |
97d10d0a MS |
63 | #define SOCKNAL_SINGLE_FRAG_TX 0 /* disable multi-fragment sends */ |
64 | #define SOCKNAL_SINGLE_FRAG_RX 0 /* disable multi-fragment receives */ | |
d7e09d03 | 65 | |
97d10d0a | 66 | #define SOCKNAL_VERSION_DEBUG 0 /* enable protocol version debugging */ |
d7e09d03 | 67 | |
4420cfd3 JS |
68 | /* |
69 | * risk kmap deadlock on multi-frag I/O (backs off to single-frag if disabled). | |
70 | * no risk if we're not running on a CONFIG_HIGHMEM platform. | |
71 | */ | |
d7e09d03 PT |
72 | #ifdef CONFIG_HIGHMEM |
73 | # define SOCKNAL_RISK_KMAP_DEADLOCK 0 | |
74 | #else | |
75 | # define SOCKNAL_RISK_KMAP_DEADLOCK 1 | |
76 | #endif | |
77 | ||
78 | struct ksock_sched_info; | |
79 | ||
ff13fd40 | 80 | struct ksock_sched { /* per scheduler state */ |
97d10d0a MS |
81 | spinlock_t kss_lock; /* serialise */ |
82 | struct list_head kss_rx_conns; /* conn waiting to be read */ | |
83 | struct list_head kss_tx_conns; /* conn waiting to be written */ | |
84 | struct list_head kss_zombie_noop_txs; /* zombie noop tx list */ | |
85 | wait_queue_head_t kss_waitq; /* where scheduler sleeps */ | |
86 | int kss_nconns; /* # connections assigned to | |
87 | * this scheduler */ | |
88 | struct ksock_sched_info *kss_info; /* owner of it */ | |
ff13fd40 | 89 | }; |
d7e09d03 PT |
90 | |
91 | struct ksock_sched_info { | |
97d10d0a MS |
92 | int ksi_nthreads_max; /* max allowed threads */ |
93 | int ksi_nthreads; /* number of threads */ | |
94 | int ksi_cpt; /* CPT id */ | |
ff13fd40 | 95 | struct ksock_sched *ksi_scheds; /* array of schedulers */ |
d7e09d03 PT |
96 | }; |
97 | ||
97d10d0a MS |
98 | #define KSOCK_CPT_SHIFT 16 |
99 | #define KSOCK_THREAD_ID(cpt, sid) (((cpt) << KSOCK_CPT_SHIFT) | (sid)) | |
100 | #define KSOCK_THREAD_CPT(id) ((id) >> KSOCK_CPT_SHIFT) | |
101 | #define KSOCK_THREAD_SID(id) ((id) & ((1UL << KSOCK_CPT_SHIFT) - 1)) | |
d7e09d03 | 102 | |
ff13fd40 | 103 | struct ksock_interface { /* in-use interface */ |
d7e09d03 PT |
104 | __u32 ksni_ipaddr; /* interface's IP address */ |
105 | __u32 ksni_netmask; /* interface's network mask */ | |
106 | int ksni_nroutes; /* # routes using (active) */ | |
107 | int ksni_npeers; /* # peers using (passive) */ | |
108 | char ksni_name[IFNAMSIZ]; /* interface name */ | |
ff13fd40 | 109 | }; |
d7e09d03 | 110 | |
ff13fd40 | 111 | struct ksock_tunables { |
97d10d0a MS |
112 | int *ksnd_timeout; /* "stuck" socket timeout |
113 | * (seconds) */ | |
114 | int *ksnd_nscheds; /* # scheduler threads in each | |
115 | * pool while starting */ | |
116 | int *ksnd_nconnds; /* # connection daemons */ | |
117 | int *ksnd_nconnds_max; /* max # connection daemons */ | |
118 | int *ksnd_min_reconnectms; /* first connection retry after | |
119 | * (ms)... */ | |
120 | int *ksnd_max_reconnectms; /* ...exponentially increasing to | |
121 | * this */ | |
122 | int *ksnd_eager_ack; /* make TCP ack eagerly? */ | |
123 | int *ksnd_typed_conns; /* drive sockets by type? */ | |
124 | int *ksnd_min_bulk; /* smallest "large" message */ | |
125 | int *ksnd_tx_buffer_size; /* socket tx buffer size */ | |
126 | int *ksnd_rx_buffer_size; /* socket rx buffer size */ | |
127 | int *ksnd_nagle; /* enable NAGLE? */ | |
128 | int *ksnd_round_robin; /* round robin for multiple | |
129 | * interfaces */ | |
130 | int *ksnd_keepalive; /* # secs for sending keepalive | |
131 | * NOOP */ | |
132 | int *ksnd_keepalive_idle; /* # idle secs before 1st probe | |
133 | */ | |
134 | int *ksnd_keepalive_count; /* # probes */ | |
135 | int *ksnd_keepalive_intvl; /* time between probes */ | |
136 | int *ksnd_credits; /* # concurrent sends */ | |
137 | int *ksnd_peertxcredits; /* # concurrent sends to 1 peer | |
138 | */ | |
139 | int *ksnd_peerrtrcredits; /* # per-peer router buffer | |
140 | * credits */ | |
141 | int *ksnd_peertimeout; /* seconds to consider peer dead | |
142 | */ | |
143 | int *ksnd_enable_csum; /* enable check sum */ | |
144 | int *ksnd_inject_csum_error; /* set non-zero to inject | |
145 | * checksum error */ | |
146 | int *ksnd_nonblk_zcack; /* always send zc-ack on | |
147 | * non-blocking connection */ | |
148 | unsigned int *ksnd_zc_min_payload; /* minimum zero copy payload | |
149 | * size */ | |
150 | int *ksnd_zc_recv; /* enable ZC receive (for | |
151 | * Chelsio TOE) */ | |
152 | int *ksnd_zc_recv_min_nfrags; /* minimum # of fragments to | |
153 | * enable ZC receive */ | |
ff13fd40 | 154 | }; |
d7e09d03 | 155 | |
ff13fd40 | 156 | struct ksock_net { |
d7e09d03 PT |
157 | __u64 ksnn_incarnation; /* my epoch */ |
158 | spinlock_t ksnn_lock; /* serialise */ | |
159 | struct list_head ksnn_list; /* chain on global list */ | |
160 | int ksnn_npeers; /* # peers */ | |
161 | int ksnn_shutdown; /* shutting down? */ | |
162 | int ksnn_ninterfaces; /* IP interfaces */ | |
ff13fd40 JS |
163 | struct ksock_interface ksnn_interfaces[LNET_MAX_INTERFACES]; |
164 | }; | |
d7e09d03 PT |
165 | |
166 | /** connd timeout */ | |
167 | #define SOCKNAL_CONND_TIMEOUT 120 | |
168 | /** reserved thread for accepting & creating new connd */ | |
169 | #define SOCKNAL_CONND_RESV 1 | |
170 | ||
ff13fd40 | 171 | struct ksock_nal_data { |
97d10d0a MS |
172 | int ksnd_init; /* initialisation state |
173 | */ | |
174 | int ksnd_nnets; /* # networks set up */ | |
175 | struct list_head ksnd_nets; /* list of nets */ | |
176 | rwlock_t ksnd_global_lock; /* stabilize peer/conn | |
177 | * ops */ | |
178 | struct list_head *ksnd_peers; /* hash table of all my | |
179 | * known peers */ | |
180 | int ksnd_peer_hash_size; /* size of ksnd_peers */ | |
181 | ||
182 | int ksnd_nthreads; /* # live threads */ | |
183 | int ksnd_shuttingdown; /* tell threads to exit | |
184 | */ | |
185 | struct ksock_sched_info **ksnd_sched_info; /* schedulers info */ | |
186 | ||
187 | atomic_t ksnd_nactive_txs; /* #active txs */ | |
188 | ||
189 | struct list_head ksnd_deathrow_conns; /* conns to close: | |
190 | * reaper_lock*/ | |
191 | struct list_head ksnd_zombie_conns; /* conns to free: | |
192 | * reaper_lock */ | |
193 | struct list_head ksnd_enomem_conns; /* conns to retry: | |
194 | * reaper_lock*/ | |
195 | wait_queue_head_t ksnd_reaper_waitq; /* reaper sleeps here */ | |
196 | unsigned long ksnd_reaper_waketime; /* when reaper will wake | |
197 | */ | |
198 | spinlock_t ksnd_reaper_lock; /* serialise */ | |
199 | ||
200 | int ksnd_enomem_tx; /* test ENOMEM sender */ | |
201 | int ksnd_stall_tx; /* test sluggish sender | |
202 | */ | |
203 | int ksnd_stall_rx; /* test sluggish | |
204 | * receiver */ | |
205 | ||
206 | struct list_head ksnd_connd_connreqs; /* incoming connection | |
207 | * requests */ | |
208 | struct list_head ksnd_connd_routes; /* routes waiting to be | |
209 | * connected */ | |
210 | wait_queue_head_t ksnd_connd_waitq; /* connds sleep here */ | |
211 | int ksnd_connd_connecting; /* # connds connecting | |
212 | */ | |
74ad578f | 213 | time64_t ksnd_connd_failed_stamp;/* time stamp of the |
97d10d0a MS |
214 | * last failed |
215 | * connecting attempt */ | |
74ad578f | 216 | time64_t ksnd_connd_starting_stamp;/* time stamp of the |
97d10d0a MS |
217 | * last starting connd |
218 | */ | |
74ad578f | 219 | unsigned ksnd_connd_starting; /* # starting connd */ |
97d10d0a MS |
220 | unsigned ksnd_connd_running; /* # running connd */ |
221 | spinlock_t ksnd_connd_lock; /* serialise */ | |
222 | ||
223 | struct list_head ksnd_idle_noop_txs; /* list head for freed | |
224 | * noop tx */ | |
225 | spinlock_t ksnd_tx_lock; /* serialise, g_lock | |
226 | * unsafe */ | |
d7e09d03 | 227 | |
ff13fd40 | 228 | }; |
d7e09d03 | 229 | |
97d10d0a MS |
230 | #define SOCKNAL_INIT_NOTHING 0 |
231 | #define SOCKNAL_INIT_DATA 1 | |
232 | #define SOCKNAL_INIT_ALL 2 | |
d7e09d03 | 233 | |
4420cfd3 JS |
234 | /* |
235 | * A packet just assembled for transmission is represented by 1 or more | |
d7e09d03 PT |
236 | * struct iovec fragments (the first frag contains the portals header), |
237 | * followed by 0 or more lnet_kiov_t fragments. | |
238 | * | |
239 | * On the receive side, initially 1 struct iovec fragment is posted for | |
240 | * receive (the header). Once the header has been received, the payload is | |
241 | * received into either struct iovec or lnet_kiov_t fragments, depending on | |
4420cfd3 JS |
242 | * what the header matched or whether the message needs forwarding. |
243 | */ | |
97d10d0a MS |
244 | struct ksock_conn; /* forward ref */ |
245 | struct ksock_peer; /* forward ref */ | |
246 | struct ksock_route; /* forward ref */ | |
247 | struct ksock_proto; /* forward ref */ | |
d7e09d03 | 248 | |
ff13fd40 | 249 | struct ksock_tx { /* transmit packet */ |
97d10d0a MS |
250 | struct list_head tx_list; /* queue on conn for transmission etc |
251 | */ | |
252 | struct list_head tx_zc_list; /* queue on peer for ZC request */ | |
253 | atomic_t tx_refcount; /* tx reference count */ | |
254 | int tx_nob; /* # packet bytes */ | |
255 | int tx_resid; /* residual bytes */ | |
256 | int tx_niov; /* # packet iovec frags */ | |
257 | struct kvec *tx_iov; /* packet iovec frags */ | |
258 | int tx_nkiov; /* # packet page frags */ | |
259 | unsigned short tx_zc_aborted; /* aborted ZC request */ | |
260 | unsigned short tx_zc_capable:1; /* payload is large enough for ZC */ | |
261 | unsigned short tx_zc_checked:1; /* Have I checked if I should ZC? */ | |
262 | unsigned short tx_nonblk:1; /* it's a non-blocking ACK */ | |
263 | lnet_kiov_t *tx_kiov; /* packet page frags */ | |
264 | struct ksock_conn *tx_conn; /* owning conn */ | |
265 | lnet_msg_t *tx_lnetmsg; /* lnet message for lnet_finalize() | |
266 | */ | |
267 | unsigned long tx_deadline; /* when (in jiffies) tx times out */ | |
268 | ksock_msg_t tx_msg; /* socklnd message buffer */ | |
269 | int tx_desc_size; /* size of this descriptor */ | |
d7e09d03 PT |
270 | union { |
271 | struct { | |
97d10d0a MS |
272 | struct kvec iov; /* virt hdr */ |
273 | lnet_kiov_t kiov[0]; /* paged payload */ | |
274 | } paged; | |
d7e09d03 | 275 | struct { |
97d10d0a MS |
276 | struct kvec iov[1]; /* virt hdr + payload */ |
277 | } virt; | |
278 | } tx_frags; | |
ff13fd40 | 279 | }; |
d7e09d03 | 280 | |
ff13fd40 | 281 | #define KSOCK_NOOP_TX_SIZE (offsetof(struct ksock_tx, tx_frags.paged.kiov[0])) |
d7e09d03 | 282 | |
ff13fd40 | 283 | /* network zero copy callback descriptor embedded in struct ksock_tx */ |
d7e09d03 | 284 | |
4420cfd3 JS |
285 | /* |
286 | * space for the rx frag descriptors; we either read a single contiguous | |
287 | * header, or up to LNET_MAX_IOV frags of payload of either type. | |
288 | */ | |
ff13fd40 | 289 | union ksock_rxiovspace { |
f351bad2 | 290 | struct kvec iov[LNET_MAX_IOV]; |
d7e09d03 | 291 | lnet_kiov_t kiov[LNET_MAX_IOV]; |
ff13fd40 | 292 | }; |
d7e09d03 | 293 | |
97d10d0a MS |
294 | #define SOCKNAL_RX_KSM_HEADER 1 /* reading ksock message header */ |
295 | #define SOCKNAL_RX_LNET_HEADER 2 /* reading lnet message header */ | |
296 | #define SOCKNAL_RX_PARSE 3 /* Calling lnet_parse() */ | |
297 | #define SOCKNAL_RX_PARSE_WAIT 4 /* waiting to be told to read the body */ | |
298 | #define SOCKNAL_RX_LNET_PAYLOAD 5 /* reading lnet payload (to deliver here) */ | |
299 | #define SOCKNAL_RX_SLOP 6 /* skipping body */ | |
d7e09d03 | 300 | |
ff13fd40 | 301 | struct ksock_conn { |
97d10d0a MS |
302 | struct ksock_peer *ksnc_peer; /* owning peer */ |
303 | struct ksock_route *ksnc_route; /* owning route */ | |
304 | struct list_head ksnc_list; /* stash on peer's conn list */ | |
305 | struct socket *ksnc_sock; /* actual socket */ | |
306 | void *ksnc_saved_data_ready; /* socket's original | |
307 | * data_ready() callback */ | |
308 | void *ksnc_saved_write_space; /* socket's original | |
309 | * write_space() callback */ | |
310 | atomic_t ksnc_conn_refcount;/* conn refcount */ | |
311 | atomic_t ksnc_sock_refcount;/* sock refcount */ | |
ff13fd40 JS |
312 | struct ksock_sched *ksnc_scheduler; /* who schedules this connection |
313 | */ | |
97d10d0a MS |
314 | __u32 ksnc_myipaddr; /* my IP */ |
315 | __u32 ksnc_ipaddr; /* peer's IP */ | |
316 | int ksnc_port; /* peer's port */ | |
317 | signed int ksnc_type:3; /* type of connection, should be | |
318 | * signed value */ | |
319 | unsigned int ksnc_closing:1; /* being shut down */ | |
320 | unsigned int ksnc_flip:1; /* flip or not, only for V2.x */ | |
321 | unsigned int ksnc_zc_capable:1; /* enable to ZC */ | |
322 | struct ksock_proto *ksnc_proto; /* protocol for the connection */ | |
d7e09d03 PT |
323 | |
324 | /* reader */ | |
97d10d0a MS |
325 | struct list_head ksnc_rx_list; /* where I enq waiting input or a |
326 | * forwarding descriptor */ | |
327 | unsigned long ksnc_rx_deadline; /* when (in jiffies) receive times | |
328 | * out */ | |
329 | __u8 ksnc_rx_started; /* started receiving a message */ | |
330 | __u8 ksnc_rx_ready; /* data ready to read */ | |
331 | __u8 ksnc_rx_scheduled; /* being progressed */ | |
332 | __u8 ksnc_rx_state; /* what is being read */ | |
333 | int ksnc_rx_nob_left; /* # bytes to next hdr/body */ | |
334 | int ksnc_rx_nob_wanted;/* bytes actually wanted */ | |
335 | int ksnc_rx_niov; /* # iovec frags */ | |
336 | struct kvec *ksnc_rx_iov; /* the iovec frags */ | |
337 | int ksnc_rx_nkiov; /* # page frags */ | |
338 | lnet_kiov_t *ksnc_rx_kiov; /* the page frags */ | |
ff13fd40 | 339 | union ksock_rxiovspace ksnc_rx_iov_space; /* space for frag descriptors */ |
97d10d0a MS |
340 | __u32 ksnc_rx_csum; /* partial checksum for incoming |
341 | * data */ | |
342 | void *ksnc_cookie; /* rx lnet_finalize passthru arg | |
343 | */ | |
344 | ksock_msg_t ksnc_msg; /* incoming message buffer: | |
345 | * V2.x message takes the | |
346 | * whole struct | |
347 | * V1.x message is a bare | |
348 | * lnet_hdr_t, it's stored in | |
349 | * ksnc_msg.ksm_u.lnetmsg */ | |
d7e09d03 PT |
350 | |
351 | /* WRITER */ | |
97d10d0a MS |
352 | struct list_head ksnc_tx_list; /* where I enq waiting for output |
353 | * space */ | |
354 | struct list_head ksnc_tx_queue; /* packets waiting to be sent */ | |
ff13fd40 | 355 | struct ksock_tx *ksnc_tx_carrier; /* next TX that can carry a LNet |
97d10d0a MS |
356 | * message or ZC-ACK */ |
357 | unsigned long ksnc_tx_deadline; /* when (in jiffies) tx times out | |
358 | */ | |
359 | int ksnc_tx_bufnob; /* send buffer marker */ | |
360 | atomic_t ksnc_tx_nob; /* # bytes queued */ | |
361 | int ksnc_tx_ready; /* write space */ | |
362 | int ksnc_tx_scheduled; /* being progressed */ | |
363 | unsigned long ksnc_tx_last_post; /* time stamp of the last posted | |
364 | * TX */ | |
ff13fd40 | 365 | }; |
d7e09d03 | 366 | |
ff13fd40 | 367 | struct ksock_route { |
97d10d0a MS |
368 | struct list_head ksnr_list; /* chain on peer route list */ |
369 | struct list_head ksnr_connd_list; /* chain on ksnr_connd_routes */ | |
370 | struct ksock_peer *ksnr_peer; /* owning peer */ | |
371 | atomic_t ksnr_refcount; /* # users */ | |
372 | unsigned long ksnr_timeout; /* when (in jiffies) reconnection | |
373 | * can happen next */ | |
374 | long ksnr_retry_interval; /* how long between retries */ | |
375 | __u32 ksnr_myipaddr; /* my IP */ | |
376 | __u32 ksnr_ipaddr; /* IP address to connect to */ | |
377 | int ksnr_port; /* port to connect to */ | |
378 | unsigned int ksnr_scheduled:1; /* scheduled for attention */ | |
379 | unsigned int ksnr_connecting:1; /* connection establishment in | |
380 | * progress */ | |
381 | unsigned int ksnr_connected:4; /* connections established by | |
382 | * type */ | |
383 | unsigned int ksnr_deleted:1; /* been removed from peer? */ | |
384 | unsigned int ksnr_share_count; /* created explicitly? */ | |
385 | int ksnr_conn_count; /* # conns established by this | |
386 | * route */ | |
ff13fd40 | 387 | }; |
d7e09d03 | 388 | |
97d10d0a | 389 | #define SOCKNAL_KEEPALIVE_PING 1 /* cookie for keepalive ping */ |
d7e09d03 | 390 | |
ff13fd40 | 391 | struct ksock_peer { |
97d10d0a MS |
392 | struct list_head ksnp_list; /* stash on global peer list */ |
393 | unsigned long ksnp_last_alive; /* when (in jiffies) I was last | |
394 | * alive */ | |
395 | lnet_process_id_t ksnp_id; /* who's on the other end(s) */ | |
396 | atomic_t ksnp_refcount; /* # users */ | |
397 | int ksnp_sharecount; /* lconf usage counter */ | |
398 | int ksnp_closing; /* being closed */ | |
399 | int ksnp_accepting; /* # passive connections pending | |
400 | */ | |
401 | int ksnp_error; /* errno on closing last conn */ | |
402 | __u64 ksnp_zc_next_cookie; /* ZC completion cookie */ | |
403 | __u64 ksnp_incarnation; /* latest known peer incarnation | |
404 | */ | |
405 | struct ksock_proto *ksnp_proto; /* latest known peer protocol */ | |
406 | struct list_head ksnp_conns; /* all active connections */ | |
407 | struct list_head ksnp_routes; /* routes */ | |
408 | struct list_head ksnp_tx_queue; /* waiting packets */ | |
409 | spinlock_t ksnp_lock; /* serialize, g_lock unsafe */ | |
410 | struct list_head ksnp_zc_req_list; /* zero copy requests wait for | |
411 | * ACK */ | |
412 | unsigned long ksnp_send_keepalive; /* time to send keepalive */ | |
413 | lnet_ni_t *ksnp_ni; /* which network */ | |
414 | int ksnp_n_passive_ips; /* # of... */ | |
415 | ||
416 | /* preferred local interfaces */ | |
417 | __u32 ksnp_passive_ips[LNET_MAX_INTERFACES]; | |
ff13fd40 | 418 | }; |
d7e09d03 | 419 | |
ff13fd40 | 420 | struct ksock_connreq { |
97d10d0a MS |
421 | struct list_head ksncr_list; /* stash on ksnd_connd_connreqs */ |
422 | lnet_ni_t *ksncr_ni; /* chosen NI */ | |
423 | struct socket *ksncr_sock; /* accepted socket */ | |
ff13fd40 | 424 | }; |
d7e09d03 | 425 | |
ff13fd40 JS |
426 | extern struct ksock_nal_data ksocknal_data; |
427 | extern struct ksock_tunables ksocknal_tunables; | |
d7e09d03 | 428 | |
97d10d0a MS |
429 | #define SOCKNAL_MATCH_NO 0 /* TX can't match type of connection */ |
430 | #define SOCKNAL_MATCH_YES 1 /* TX matches type of connection */ | |
431 | #define SOCKNAL_MATCH_MAY 2 /* TX can be sent on the connection, but not | |
432 | * preferred */ | |
d7e09d03 | 433 | |
ff13fd40 | 434 | struct ksock_proto { |
97d10d0a MS |
435 | /* version number of protocol */ |
436 | int pro_version; | |
437 | ||
438 | /* handshake function */ | |
ff13fd40 | 439 | int (*pro_send_hello)(struct ksock_conn *, ksock_hello_msg_t *); |
97d10d0a MS |
440 | |
441 | /* handshake function */ | |
ff13fd40 | 442 | int (*pro_recv_hello)(struct ksock_conn *, ksock_hello_msg_t *, int); |
97d10d0a MS |
443 | |
444 | /* message pack */ | |
ff13fd40 | 445 | void (*pro_pack)(struct ksock_tx *); |
97d10d0a MS |
446 | |
447 | /* message unpack */ | |
448 | void (*pro_unpack)(ksock_msg_t *); | |
449 | ||
450 | /* queue tx on the connection */ | |
ff13fd40 | 451 | struct ksock_tx *(*pro_queue_tx_msg)(struct ksock_conn *, struct ksock_tx *); |
97d10d0a MS |
452 | |
453 | /* queue ZC ack on the connection */ | |
ff13fd40 | 454 | int (*pro_queue_tx_zcack)(struct ksock_conn *, struct ksock_tx *, __u64); |
97d10d0a MS |
455 | |
456 | /* handle ZC request */ | |
ff13fd40 | 457 | int (*pro_handle_zcreq)(struct ksock_conn *, __u64, int); |
97d10d0a MS |
458 | |
459 | /* handle ZC ACK */ | |
ff13fd40 | 460 | int (*pro_handle_zcack)(struct ksock_conn *, __u64, __u64); |
97d10d0a | 461 | |
4420cfd3 JS |
462 | /* |
463 | * msg type matches the connection type: | |
97d10d0a MS |
464 | * return value: |
465 | * return MATCH_NO : no | |
466 | * return MATCH_YES : matching type | |
4420cfd3 JS |
467 | * return MATCH_MAY : can be backup |
468 | */ | |
ff13fd40 JS |
469 | int (*pro_match_tx)(struct ksock_conn *, struct ksock_tx *, int); |
470 | }; | |
d7e09d03 | 471 | |
ff13fd40 JS |
472 | extern struct ksock_proto ksocknal_protocol_v1x; |
473 | extern struct ksock_proto ksocknal_protocol_v2x; | |
474 | extern struct ksock_proto ksocknal_protocol_v3x; | |
d7e09d03 | 475 | |
97d10d0a MS |
476 | #define KSOCK_PROTO_V1_MAJOR LNET_PROTO_TCP_VERSION_MAJOR |
477 | #define KSOCK_PROTO_V1_MINOR LNET_PROTO_TCP_VERSION_MINOR | |
478 | #define KSOCK_PROTO_V1 KSOCK_PROTO_V1_MAJOR | |
d7e09d03 PT |
479 | |
480 | #ifndef CPU_MASK_NONE | |
481 | #define CPU_MASK_NONE 0UL | |
482 | #endif | |
483 | ||
12c41f00 JH |
484 | static inline __u32 ksocknal_csum(__u32 crc, unsigned char const *p, size_t len) |
485 | { | |
486 | #if 1 | |
487 | return crc32_le(crc, p, len); | |
488 | #else | |
489 | while (len-- > 0) | |
490 | crc = ((crc + 0x100) & ~0xff) | ((crc + *p++) & 0xff) ; | |
491 | return crc; | |
492 | #endif | |
493 | } | |
494 | ||
d7e09d03 PT |
495 | static inline int |
496 | ksocknal_route_mask(void) | |
497 | { | |
498 | if (!*ksocknal_tunables.ksnd_typed_conns) | |
499 | return (1 << SOCKLND_CONN_ANY); | |
500 | ||
501 | return ((1 << SOCKLND_CONN_CONTROL) | | |
502 | (1 << SOCKLND_CONN_BULK_IN) | | |
503 | (1 << SOCKLND_CONN_BULK_OUT)); | |
504 | } | |
505 | ||
506 | static inline struct list_head * | |
d9dfb48f | 507 | ksocknal_nid2peerlist(lnet_nid_t nid) |
d7e09d03 PT |
508 | { |
509 | unsigned int hash = ((unsigned int)nid) % ksocknal_data.ksnd_peer_hash_size; | |
510 | ||
71397095 | 511 | return &ksocknal_data.ksnd_peers[hash]; |
d7e09d03 PT |
512 | } |
513 | ||
514 | static inline void | |
ff13fd40 | 515 | ksocknal_conn_addref(struct ksock_conn *conn) |
d7e09d03 | 516 | { |
d9dfb48f | 517 | LASSERT(atomic_read(&conn->ksnc_conn_refcount) > 0); |
d7e09d03 PT |
518 | atomic_inc(&conn->ksnc_conn_refcount); |
519 | } | |
520 | ||
ff13fd40 JS |
521 | void ksocknal_queue_zombie_conn(struct ksock_conn *conn); |
522 | void ksocknal_finalize_zcreq(struct ksock_conn *conn); | |
d7e09d03 PT |
523 | |
524 | static inline void | |
ff13fd40 | 525 | ksocknal_conn_decref(struct ksock_conn *conn) |
d7e09d03 | 526 | { |
d9dfb48f | 527 | LASSERT(atomic_read(&conn->ksnc_conn_refcount) > 0); |
d7e09d03 PT |
528 | if (atomic_dec_and_test(&conn->ksnc_conn_refcount)) |
529 | ksocknal_queue_zombie_conn(conn); | |
530 | } | |
531 | ||
532 | static inline int | |
ff13fd40 | 533 | ksocknal_connsock_addref(struct ksock_conn *conn) |
d7e09d03 | 534 | { |
97d10d0a | 535 | int rc = -ESHUTDOWN; |
d7e09d03 PT |
536 | |
537 | read_lock(&ksocknal_data.ksnd_global_lock); | |
538 | if (!conn->ksnc_closing) { | |
539 | LASSERT(atomic_read(&conn->ksnc_sock_refcount) > 0); | |
540 | atomic_inc(&conn->ksnc_sock_refcount); | |
541 | rc = 0; | |
542 | } | |
543 | read_unlock(&ksocknal_data.ksnd_global_lock); | |
544 | ||
71397095 | 545 | return rc; |
d7e09d03 PT |
546 | } |
547 | ||
548 | static inline void | |
ff13fd40 | 549 | ksocknal_connsock_decref(struct ksock_conn *conn) |
d7e09d03 | 550 | { |
d9dfb48f | 551 | LASSERT(atomic_read(&conn->ksnc_sock_refcount) > 0); |
d7e09d03 | 552 | if (atomic_dec_and_test(&conn->ksnc_sock_refcount)) { |
d9dfb48f | 553 | LASSERT(conn->ksnc_closing); |
e52fc91d | 554 | sock_release(conn->ksnc_sock); |
d7e09d03 PT |
555 | conn->ksnc_sock = NULL; |
556 | ksocknal_finalize_zcreq(conn); | |
557 | } | |
558 | } | |
559 | ||
560 | static inline void | |
ff13fd40 | 561 | ksocknal_tx_addref(struct ksock_tx *tx) |
d7e09d03 | 562 | { |
d9dfb48f | 563 | LASSERT(atomic_read(&tx->tx_refcount) > 0); |
d7e09d03 PT |
564 | atomic_inc(&tx->tx_refcount); |
565 | } | |
566 | ||
ff13fd40 JS |
567 | void ksocknal_tx_prep(struct ksock_conn *, struct ksock_tx *tx); |
568 | void ksocknal_tx_done(lnet_ni_t *ni, struct ksock_tx *tx); | |
d7e09d03 PT |
569 | |
570 | static inline void | |
ff13fd40 | 571 | ksocknal_tx_decref(struct ksock_tx *tx) |
d7e09d03 | 572 | { |
d9dfb48f | 573 | LASSERT(atomic_read(&tx->tx_refcount) > 0); |
d7e09d03 PT |
574 | if (atomic_dec_and_test(&tx->tx_refcount)) |
575 | ksocknal_tx_done(NULL, tx); | |
576 | } | |
577 | ||
578 | static inline void | |
ff13fd40 | 579 | ksocknal_route_addref(struct ksock_route *route) |
d7e09d03 | 580 | { |
d9dfb48f | 581 | LASSERT(atomic_read(&route->ksnr_refcount) > 0); |
d7e09d03 PT |
582 | atomic_inc(&route->ksnr_refcount); |
583 | } | |
584 | ||
ff13fd40 | 585 | void ksocknal_destroy_route(struct ksock_route *route); |
d7e09d03 PT |
586 | |
587 | static inline void | |
ff13fd40 | 588 | ksocknal_route_decref(struct ksock_route *route) |
d7e09d03 | 589 | { |
d9dfb48f | 590 | LASSERT(atomic_read(&route->ksnr_refcount) > 0); |
d7e09d03 | 591 | if (atomic_dec_and_test(&route->ksnr_refcount)) |
d9dfb48f | 592 | ksocknal_destroy_route(route); |
d7e09d03 PT |
593 | } |
594 | ||
595 | static inline void | |
ff13fd40 | 596 | ksocknal_peer_addref(struct ksock_peer *peer) |
d7e09d03 | 597 | { |
d9dfb48f | 598 | LASSERT(atomic_read(&peer->ksnp_refcount) > 0); |
d7e09d03 PT |
599 | atomic_inc(&peer->ksnp_refcount); |
600 | } | |
601 | ||
ff13fd40 | 602 | void ksocknal_destroy_peer(struct ksock_peer *peer); |
d7e09d03 PT |
603 | |
604 | static inline void | |
ff13fd40 | 605 | ksocknal_peer_decref(struct ksock_peer *peer) |
d7e09d03 | 606 | { |
d9dfb48f | 607 | LASSERT(atomic_read(&peer->ksnp_refcount) > 0); |
d7e09d03 | 608 | if (atomic_dec_and_test(&peer->ksnp_refcount)) |
d9dfb48f | 609 | ksocknal_destroy_peer(peer); |
d7e09d03 PT |
610 | } |
611 | ||
d9dfb48f AD |
612 | int ksocknal_startup(lnet_ni_t *ni); |
613 | void ksocknal_shutdown(lnet_ni_t *ni); | |
d7e09d03 | 614 | int ksocknal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg); |
d9dfb48f | 615 | int ksocknal_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg); |
d7e09d03 | 616 | int ksocknal_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, |
c1b7b8eb | 617 | int delayed, struct iov_iter *to, unsigned int rlen); |
e327dc88 | 618 | int ksocknal_accept(lnet_ni_t *ni, struct socket *sock); |
d7e09d03 | 619 | |
8150a97f | 620 | int ksocknal_add_peer(lnet_ni_t *ni, lnet_process_id_t id, __u32 ip, int port); |
ff13fd40 JS |
621 | struct ksock_peer *ksocknal_find_peer_locked(lnet_ni_t *ni, lnet_process_id_t id); |
622 | struct ksock_peer *ksocknal_find_peer(lnet_ni_t *ni, lnet_process_id_t id); | |
623 | void ksocknal_peer_failed(struct ksock_peer *peer); | |
624 | int ksocknal_create_conn(lnet_ni_t *ni, struct ksock_route *route, | |
8150a97f | 625 | struct socket *sock, int type); |
ff13fd40 JS |
626 | void ksocknal_close_conn_locked(struct ksock_conn *conn, int why); |
627 | void ksocknal_terminate_conn(struct ksock_conn *conn); | |
628 | void ksocknal_destroy_conn(struct ksock_conn *conn); | |
629 | int ksocknal_close_peer_conns_locked(struct ksock_peer *peer, | |
8150a97f | 630 | __u32 ipaddr, int why); |
ff13fd40 | 631 | int ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why); |
8150a97f | 632 | int ksocknal_close_matching_conns(lnet_process_id_t id, __u32 ipaddr); |
ff13fd40 | 633 | struct ksock_conn *ksocknal_find_conn_locked(struct ksock_peer *peer, |
24c198e9 | 634 | struct ksock_tx *tx, int nonblk); |
8150a97f | 635 | |
ff13fd40 | 636 | int ksocknal_launch_packet(lnet_ni_t *ni, struct ksock_tx *tx, |
8150a97f | 637 | lnet_process_id_t id); |
ff13fd40 JS |
638 | struct ksock_tx *ksocknal_alloc_tx(int type, int size); |
639 | void ksocknal_free_tx(struct ksock_tx *tx); | |
640 | struct ksock_tx *ksocknal_alloc_tx_noop(__u64 cookie, int nonblk); | |
641 | void ksocknal_next_tx_carrier(struct ksock_conn *conn); | |
642 | void ksocknal_queue_tx_locked(struct ksock_tx *tx, struct ksock_conn *conn); | |
8150a97f JP |
643 | void ksocknal_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int error); |
644 | void ksocknal_notify(lnet_ni_t *ni, lnet_nid_t gw_nid, int alive); | |
645 | void ksocknal_query(struct lnet_ni *ni, lnet_nid_t nid, unsigned long *when); | |
646 | int ksocknal_thread_start(int (*fn)(void *arg), void *arg, char *name); | |
647 | void ksocknal_thread_fini(void); | |
ff13fd40 JS |
648 | void ksocknal_launch_all_connections_locked(struct ksock_peer *peer); |
649 | struct ksock_route *ksocknal_find_connectable_route_locked(struct ksock_peer *peer); | |
650 | struct ksock_route *ksocknal_find_connecting_route_locked(struct ksock_peer *peer); | |
651 | int ksocknal_new_packet(struct ksock_conn *conn, int skip); | |
8150a97f JP |
652 | int ksocknal_scheduler(void *arg); |
653 | int ksocknal_connd(void *arg); | |
654 | int ksocknal_reaper(void *arg); | |
ff13fd40 | 655 | int ksocknal_send_hello(lnet_ni_t *ni, struct ksock_conn *conn, |
8150a97f | 656 | lnet_nid_t peer_nid, ksock_hello_msg_t *hello); |
ff13fd40 | 657 | int ksocknal_recv_hello(lnet_ni_t *ni, struct ksock_conn *conn, |
8150a97f JP |
658 | ksock_hello_msg_t *hello, lnet_process_id_t *id, |
659 | __u64 *incarnation); | |
ff13fd40 JS |
660 | void ksocknal_read_callback(struct ksock_conn *conn); |
661 | void ksocknal_write_callback(struct ksock_conn *conn); | |
662 | ||
663 | int ksocknal_lib_zc_capable(struct ksock_conn *conn); | |
664 | void ksocknal_lib_save_callback(struct socket *sock, struct ksock_conn *conn); | |
665 | void ksocknal_lib_set_callback(struct socket *sock, struct ksock_conn *conn); | |
666 | void ksocknal_lib_reset_callback(struct socket *sock, struct ksock_conn *conn); | |
667 | void ksocknal_lib_push_conn(struct ksock_conn *conn); | |
668 | int ksocknal_lib_get_conn_addrs(struct ksock_conn *conn); | |
8150a97f | 669 | int ksocknal_lib_setup_sock(struct socket *so); |
ff13fd40 JS |
670 | int ksocknal_lib_send_iov(struct ksock_conn *conn, struct ksock_tx *tx); |
671 | int ksocknal_lib_send_kiov(struct ksock_conn *conn, struct ksock_tx *tx); | |
672 | void ksocknal_lib_eager_ack(struct ksock_conn *conn); | |
673 | int ksocknal_lib_recv_iov(struct ksock_conn *conn); | |
674 | int ksocknal_lib_recv_kiov(struct ksock_conn *conn); | |
675 | int ksocknal_lib_get_conn_tunables(struct ksock_conn *conn, int *txmem, | |
8150a97f JP |
676 | int *rxmem, int *nagle); |
677 | ||
ff13fd40 JS |
678 | void ksocknal_read_callback(struct ksock_conn *conn); |
679 | void ksocknal_write_callback(struct ksock_conn *conn); | |
ee98e442 | 680 | |
8150a97f JP |
681 | int ksocknal_tunables_init(void); |
682 | ||
ff13fd40 | 683 | void ksocknal_lib_csum_tx(struct ksock_tx *tx); |
8150a97f | 684 | |
ff13fd40 | 685 | int ksocknal_lib_memory_pressure(struct ksock_conn *conn); |
8150a97f | 686 | int ksocknal_lib_bind_thread_to_cpu(int id); |
12c41f00 JH |
687 | |
688 | #endif /* _SOCKLND_SOCKLND_H_ */ |