2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Copyright (c) 2013, 2014, Intel Corporation.
8 * Derived from work by: Eric Barton <eric@bartonsoftware.com>
9 * Author: Nic Henke <nic@cray.com>
10 * Author: James Shimek <jshimek@cray.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #ifndef _GNILND_GNILND_H_
29 #define _GNILND_GNILND_H_
31 #include <linux/module.h>
32 #include <linux/kernel.h>
34 #include <linux/string.h>
35 #include <linux/stat.h>
36 #include <linux/errno.h>
37 #ifdef HAVE_LINUX_KERNEL_LOCK
38 #include <linux/smp_lock.h>
40 #include <linux/unistd.h>
41 #include <linux/uio.h>
42 #include <linux/time.h>
43 #include <asm/timex.h>
45 #include <asm/uaccess.h>
48 #include <linux/init.h>
50 #include <linux/file.h>
51 #include <linux/stat.h>
52 #include <linux/list.h>
53 #include <linux/kmod.h>
54 #include <linux/sysctl.h>
55 #include <linux/kthread.h>
56 #include <linux/nmi.h>
60 #include <linux/nmi.h>
62 #define DEBUG_SUBSYSTEM S_LND
64 #include <libcfs/linux/kp30.h>
65 #include <libcfs/libcfs.h>
66 #include <lnet/lnet.h>
67 #include <lnet/lib-lnet.h>
70 #include "gnilnd_version.h"
73 /* tunables determined at compile time */
74 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
75 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
76 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
77 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
78 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
79 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
80 #define GNILND_SCHED_TIMEOUT 1
81 #define GNILND_DGRAM_TIMEOUT 2
82 #define GNILND_FAST_MAPPING_TRY \
83 *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
84 #define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
86 /* map failure timeout */
87 #define GNILND_MAP_TIMEOUT \
88 (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
89 *kgnilnd_tunables.kgn_timeout))
91 /* reaper thread wakup interval */
92 #define GNILND_REAPER_THREAD_WAKE 1
93 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
94 #define GNILND_REAPER_NCHECKS 4
97 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
98 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
99 #define GNILND_COOKIE 0xa3579 /* cookie used by along with ptag by GNI */
100 #define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
101 /* checksum values */
102 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
103 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
104 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
105 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
107 /* tune down some COMPUTE options as they won't see the same number of connections and
108 * don't need the throughput of multiple threads by default */
109 #if defined(CONFIG_CRAY_COMPUTE)
111 #define GNILND_SCHED_THREADS 2 /* default # of kgnilnd_scheduler threads */
113 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
115 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
116 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
117 #define GNILND_COMPUTE 1 /* compute image */
118 #define GNILND_FAST_RECONNECT 1 /* Fast Reconnect option */
120 #define GNILND_SCHED_THREADS 3 /* default # of kgnilnd_scheduler threads */
121 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
122 #define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
123 #define GNILND_COMPUTE 0 /* service image */
124 #define GNILND_FAST_RECONNECT 0 /* Fast Reconnect option */
127 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
128 #define GNILND_EXTRA_BITS 1
129 /* maximum number of conns & bits for cqid in the SMSG event data */
130 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
131 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
132 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
133 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
134 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
136 /* need sane upper bound to limit copy overhead */
137 #define GNILND_MAX_IMMEDIATE (64<<10)
139 /* Max number of connections to keep in purgatory per peer */
140 #define GNILND_PURGATORY_MAX 5
142 /* payload size to add to the base mailbox size
143 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
144 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
145 * the calculation return from that function.*/
146 #define GNILND_MBOX_PAYLOAD \
147 (GNILND_MAX_MSG_SIZE * \
148 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
150 /* timeout -> deadman timer for kgni mdd holds */
151 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
153 /* timeout for failing sends in t is in jiffies*/
154 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
156 /* time when to release from purgatory in the reaper thread in jiffies */
157 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
159 /* Macro for finding last_rx 2 datapoints are compared
160 * and the most recent one in jiffies is returned.
162 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
163 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
165 /************************************************************************
166 * Enum, flag and tag data
168 #define GNILND_INIT_NOTHING 0
169 #define GNILND_INIT_DATA 1
170 #define GNILND_INIT_ALL 2
172 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
173 #define GNILND_BUF_NONE 0 /* buffer type not set */
174 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
175 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
176 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
177 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
178 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
179 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
181 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
182 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
183 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
184 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
185 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
187 /* stash above max CQID to avoid any collision */
188 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
189 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
191 /* kgn_msg_t::gnm_type */
192 #define GNILND_MSG_NONE 0x00 /* illegal message */
193 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
194 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
195 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
196 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
197 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
198 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
199 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
200 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
201 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
202 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
203 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
204 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
205 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
206 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
207 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
208 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
209 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
211 /* defines for gnc_*scheduled states */
212 #define GNILND_CONN_IDLE 0
213 #define GNILND_CONN_SCHED 1
214 #define GNILND_CONN_WANTS_SCHED 2
215 #define GNILND_CONN_PROCESS 3
217 #define GNILND_DEV_IDLE 0
218 #define GNILND_DEV_IRQ 1
219 #define GNILND_DEV_LOOP 2
221 #define GNILND_DGRAM_IDLE 0
222 #define GNILND_DGRAM_SCHED 1
223 #define GNILND_DGRAM_PROCESS 2
225 #define GNILND_PEER_IDLE 0
226 #define GNILND_PEER_CONNECT 1
227 #define GNILND_PEER_POSTING 2
228 #define GNILND_PEER_POSTED 3
229 #define GNILND_PEER_NEEDS_DEATH 4
230 #define GNILND_PEER_KILL 5
232 /* for gnc_close_recvd */
233 #define GNILND_CLOSE_RX 1
234 #define GNILND_CLOSE_INJECT1 2
235 #define GNILND_CLOSE_INJECT2 3
236 #define GNILND_CLOSE_EARLY 4
238 /* defines for why quiesce trigger set */
239 #define GNILND_QUIESCE_IDLE 0
240 #define GNILND_QUIESCE_ADMIN 1
241 #define GNILND_QUIESCE_RESET 2
242 #define GNILND_QUIESCE_HW_QUIESCE 3
244 #define GNILND_PEER_CLEAN 0
245 #define GNILND_PEER_PERSISTING 1
247 #define GNILND_DEL_CONN 0
248 #define GNILND_DEL_PEER 1
249 #define GNILND_CLEAR_PURGATORY 2
251 #define GNILND_RCA_NODE_UP 0
252 #define GNILND_RCA_NODE_DOWN 1
253 #define GNILND_RCA_NODE_UNKNOWN 2
255 /* defines for reverse RDMA states */
256 #define GNILND_REVERSE_NONE 0
257 #define GNILND_REVERSE_GET 1
258 #define GNILND_REVERSE_PUT 2
259 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
261 typedef enum kgn_fmablk_state {
262 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
263 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
264 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
265 GNILND_FMABLK_FREED, /* after free */
266 } kgn_fmablk_state_t;
268 typedef enum kgn_tx_list_state {
269 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
270 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
271 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
272 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
273 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
274 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
275 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
276 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
277 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
278 GNILND_TX_FREED /* TX is free! */
279 } kgn_tx_list_state_t;
281 typedef enum kgn_conn_state {
282 /* don't start @ 0 - prevent memset(0) badness */
283 GNILND_CONN_DUMMY = 0,
285 GNILND_CONN_CONNECTING,
286 GNILND_CONN_ESTABLISHED,
290 GNILND_CONN_DESTROY_EP
293 /* changing these requires a change to GNILND_CONNREQ_VERSION and
294 * will result in dropped packets instead of NAKs. Adding to this is
295 * acceptable without changing the CONNREQ_VERSION, but code should
296 * be ready to handle NAKs on version mismatch */
297 typedef enum kgn_connreq_type {
298 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
299 GNILND_CONNREQ_NAK, /* NO soup for you! */
300 GNILND_CONNREQ_CLOSE, /* we should see other people */
301 } kgn_connreq_type_t;
303 typedef enum kgn_dgram_state {
304 /* don't use 0 to avoid thinking a memset of zero is valid data */
305 GNILND_DGRAM_USED = 1,
306 GNILND_DGRAM_POSTING,
308 GNILND_DGRAM_PROCESSING,
309 GNILND_DGRAM_CANCELED,
313 typedef enum kgn_dgram_type {
314 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
315 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
316 GNILND_DGRAM_NAK, /* NO soup for you! */
317 GNILND_DGRAM_CLOSE, /* we should see other people */
320 /************************************************************************
321 * Wire message structs. These are sent in sender's byte order
322 * (i.e. receiver checks magic and flips if required).
325 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
326 #define GNILND_DGRAM_MAGIC 0x0DDBA11
328 /* kgn_msg_t - FMA/SMSG wire struct
330 * - added checksum to FMA
331 * moved seq before paylod
332 * WIRE_ATTR added for alignment
334 * added gnm_payload_len for FMA payload size
336 * added gncm_retval to completion, allowing return code transmission
339 * changed how CQID and TX ids are assigned
341 * added retval on CLOSE
343 * added payload checksumming
345 * reworked checksumming a bit, changed payload checksums
347 #define GNILND_MSG_VERSION 8
348 /* kgn_connreq_t connection request datagram wire struct
353 #define GNILND_CONNREQ_VERSION 2
355 typedef struct kgn_gniparams {
356 __u32 gnpr_host_id; /* ph. host ID of the NIC */
357 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
358 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
359 } WIRE_ATTR kgn_gniparams_t;
361 typedef struct kgn_nak_data {
362 __s32 gnnd_errno; /* errno reason for NAK */
364 } WIRE_ATTR kgn_nak_data_t;
366 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
367 * without breaking the ability for us to properly NAK someone */
368 typedef struct kgn_connreq { /* connection request/response */
369 __u32 gncr_magic; /* I'm an gnilnd connreq */
370 __u32 gncr_cksum; /* checksum (0 == disabled) */
371 __u16 gncr_type; /* REQ, NAK, etc */
372 __u16 gncr_version; /* this is my version number */
373 __u32 gncr_timeout; /* sender's timeout */
374 __u64 gncr_srcnid; /* sender's NID */
375 __u64 gncr_dstnid; /* who sender expects to listen */
376 __u64 gncr_peerstamp; /* sender's instance stamp */
377 __u64 gncr_connstamp; /* sender's connection stamp */
379 /* everything before this needs to stay static, adding after should
380 * result in a change to GNILND_CONNREQ_VERSION */
383 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
384 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
386 } WIRE_ATTR kgn_connreq_t;
389 gni_mem_handle_t gnrd_key;
392 } WIRE_ATTR kgn_rdma_desc_t;
395 lnet_hdr_t gnim_hdr; /* LNet header */
396 /* LNet payload is in FMA "Message Data" */
397 } WIRE_ATTR kgn_immediate_msg_t;
400 lnet_hdr_t gnprm_hdr; /* LNet header */
401 __u64 gnprm_cookie; /* opaque completion cookie */
402 } WIRE_ATTR kgn_putreq_msg_t;
405 __u64 gnpam_src_cookie; /* reflected completion cookie */
406 __u64 gnpam_dst_cookie; /* opaque completion cookie */
407 __u16 gnpam_payload_cksum; /* checksum for get msg */
408 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
409 } WIRE_ATTR kgn_putack_msg_t;
412 lnet_hdr_t gngm_hdr; /* LNet header */
413 __u64 gngm_cookie; /* opaque completion cookie */
414 __u16 gngm_payload_cksum; /* checksum for put msg */
415 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
416 } WIRE_ATTR kgn_get_msg_t;
419 int gncm_retval; /* error on NAK, size on REQ */
420 __u64 gncm_cookie; /* reflected completion cookie */
421 } WIRE_ATTR kgn_completion_msg_t;
423 typedef struct { /* NB must fit in FMA "Prefix" */
424 __u32 gnm_magic; /* I'm an gni message */
425 __u16 gnm_version; /* this is my version number */
426 __u16 gnm_type; /* msg type */
427 __u64 gnm_srcnid; /* sender's NID */
428 __u64 gnm_connstamp; /* sender's connection stamp */
429 __u32 gnm_seq; /* incrementing sequence number */
430 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
431 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
432 __u32 gnm_payload_len; /* size of the FMA payload sent */
434 kgn_immediate_msg_t immediate;
435 kgn_putreq_msg_t putreq;
436 kgn_putack_msg_t putack;
438 kgn_completion_msg_t completion;
440 } WIRE_ATTR kgn_msg_t;
442 /************************************************************************
443 * runtime tunable data
446 typedef struct kgn_tunables {
447 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
448 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
449 int *kgn_credits; /* # concurrent sends */
450 int *kgn_fma_cq_size; /* # entries in receive CQ */
451 int *kgn_peer_credits; /* # LNet peer credits */
452 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
453 int *kgn_timeout; /* comms timeout (seconds) */
454 int *kgn_max_immediate; /* immediate payload breakpoint */
455 int *kgn_checksum; /* checksum data */
456 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
457 int *kgn_bte_dlvr_mode; /* BTE delivery mode mask */
458 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
459 int *kgn_ptag; /* PTAG for cdm_create */
460 int *kgn_max_retransmits; /* max number of FMA retransmits */
461 int *kgn_nwildcard; /* # wildcard per net to post */
462 int *kgn_nice; /* nice value for kgnilnd threads */
463 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
464 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
465 int *kgn_peer_hash_size; /* size of kgn_peers */
466 int *kgn_peer_health; /* enable/disable peer health */
467 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
468 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
469 int *kgn_mbox_per_block; /* mailboxes per fmablk */
470 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
471 int *kgn_mbox_credits; /* max credits per fma */
472 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
473 int *kgn_net_hash_size; /* size of kgn_net_ht */
474 int *kgn_hardware_timeout; /* max time for a message to get across the network */
475 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
476 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
477 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
478 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
479 int *kgn_reverse_rdma; /* Reverse RDMA setting */
480 int *kgn_eager_credits; /* allocated eager buffers */
481 int *kgn_fast_reconn; /* fast reconnection on conn timeout */
482 int *kgn_efault_lbug; /* LBUG on receiving an EFAULT */
483 int *kgn_max_purgatory; /* # conns/peer to keep in purgatory */
484 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
485 cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
489 typedef struct kgn_mbox_info {
490 lnet_nid_t mbx_prev_nid;
491 lnet_nid_t mbx_prev_purg_nid;
492 unsigned long mbx_create_conn_memset;
493 unsigned long mbx_add_purgatory;
494 unsigned long mbx_detach_of_purgatory;
495 unsigned long mbx_release_from_purgatory;
496 unsigned long mbx_release_purg_active_dgram;
498 int mbx_nallocs_total;
501 typedef struct kgn_fma_memblock {
502 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
503 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
504 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
505 int gnm_num_mboxs; /* total mboxes allocated */
506 int gnm_avail_mboxs; /* number of available mailboxes in the block */
507 int gnm_held_mboxs; /* number of purgatory held mailboxes */
508 int gnm_mbox_size; /* size of the single mailbox */
509 int gnm_next_avail_mbox; /* next available mailbox */
510 long gnm_max_timeout; /* max timeout for possible purgatory hold */
511 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
512 void *gnm_block; /* pointer to mem. block */
513 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
514 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
515 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
516 } kgn_fma_memblock_t;
518 typedef struct kgn_device {
519 gni_nic_handle_t gnd_handle; /* device handle */
520 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
521 gni_err_handle_t gnd_err_handle; /* device error handle */
522 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
523 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
524 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
525 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
526 struct mutex gnd_cq_mutex; /* CQ access serialization */
527 __u32 gnd_host_id; /* ph. host ID of the NIC */
528 int gnd_id; /* device id, also index in kgn_devices */
529 __u32 gnd_nid; /* ph host ID translated to NID */
530 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
531 struct mutex gnd_fmablk_mutex; /* mutex for FMA block memory alloc/free */
532 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
533 atomic_t gnd_nfmablk; /* # of fmablk live */
534 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
535 atomic_t gnd_neps; /* # EP allocated to conns */
536 short gnd_ready; /* stuff to do in scheduler thread */
537 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
538 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
539 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
540 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
541 struct list_head gnd_connd_peers; /* peers waiting for a connection */
542 spinlock_t gnd_connd_lock; /* serialise connd_peers */
543 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
544 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
545 int gnd_dgram_ready; /* dgrams need movin' */
546 struct list_head *gnd_dgrams; /* nid hash to dgrams */
547 atomic_t gnd_ndgrams; /* # dgrams extant */
548 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post*/
549 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
550 struct list_head gnd_map_list; /* list of all mapped regions */
551 int gnd_map_version; /* version flag for map list */
552 struct timer_list gnd_map_timer; /* wakey-wakey */
553 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
554 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
555 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
556 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
557 __u32 gnd_map_nphys; /* # TX phys mappings */
558 __u32 gnd_map_physnop; /* # TX phys pages mapped */
559 __u32 gnd_map_nvirt; /* # TX virt mappings */
560 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
561 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
562 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
563 int gnd_map_attempt; /* last map attempt # */
564 unsigned long gnd_last_map; /* map timeout base */
565 struct list_head gnd_rdmaq; /* RDMA to be sent */
566 spinlock_t gnd_rdmaq_lock; /* play nice with others */
567 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
568 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
569 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
570 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
571 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
572 atomic_t gnd_short_ntx; /* TX stats: short messages */
573 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
574 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
575 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
576 atomic_t gnd_short_nrx; /* RX stats: short messages */
577 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
578 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
579 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
580 atomic_t gnd_fast_try; /* # of times fast send tried */
581 atomic_t gnd_fast_ok; /* # of times fast send ok */
582 atomic_t gnd_fast_block; /* # of times fast send blocked */
583 unsigned long gnd_mutex_delay;
584 atomic_t gnd_n_yield;
585 atomic_t gnd_n_schedule;
586 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
587 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
590 typedef struct kgn_net {
591 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
592 kgn_device_t *gnn_dev; /* device for this net */
593 lnet_ni_t *gnn_ni; /* network interface instance */
594 atomic_t gnn_refcount; /* # current references */
595 int gnn_shutdown; /* lnd_shutdown set */
596 __u16 gnn_netnum; /* stash netnum for quicker lookup */
599 static inline lnet_nid_t
600 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
602 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
605 static inline lnet_nid_t
606 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
608 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
611 /* The code for this is a bit ugly - but really this just boils down to a __u64
612 * that can have various parts accessed separately.
614 * The lower 32 bits is the ID
615 * we give to SMSG for our completion event - it needs to be globally unique across
616 * all TX currently in flight. We separate that out into the CQID so that we can
617 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
618 * the actual TX out of the per-connection gnc_tx_ref_table.
620 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
621 * has a unique value we can send with RDMA setup messages to ensure the completion for
622 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
623 * reuse is separated.
626 typedef struct kgn_tx_ev_id {
628 __u64 txe_cookie; /* are you my mommy ? */
630 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
632 __u32 txe_smsg_id; /* ID for SMSG CQ event */
633 /* N.B: Never ever ever ever use the bit shifts directly,
634 * you are just asking for a world of pain and are at the
635 * mercy of the compiler layouts */
637 __u32 txe_cqid :GNILND_CQID_NBITS;
638 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
645 typedef struct kgn_dgram {
646 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
647 kgn_dgram_state_t gndg_state; /* state of this dgram */
648 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
649 __u32 gndg_magic; /* saftey word */
650 unsigned long gndg_post_time; /* time when we posted */
651 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
652 kgn_connreq_t gndg_conn_out; /* connreq from local node */
653 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
656 typedef struct kgn_tx { /* message descriptor */
657 struct list_head tx_list; /* TX queues - peer, conn, rdma */
658 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
659 struct list_head *tx_list_p; /* pointer to current list */
660 struct kgn_conn *tx_conn; /* owning conn */
661 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
662 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
663 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
664 struct list_head tx_map_list; /* list entry on device map list */
665 unsigned int tx_nob; /* # bytes of payload */
666 int tx_buftype; /* payload buffer type */
667 int tx_phys_npages; /* # physical pages */
668 gni_mem_handle_t tx_map_key; /* mapping key */
669 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
670 gni_mem_segment_t *tx_phys; /* page descriptors */
671 kgn_msg_t tx_msg; /* FMA message buffer */
672 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
673 __u8 tx_state; /* state of the descriptor */
674 int tx_retrans; /* retrans count of RDMA */
675 int tx_rc; /* if we need to stash the ret code until we see completion */
676 void *tx_buffer; /* source/sink buffer */
677 void *tx_buffer_copy; /* pointer to page aligned buffer */
678 unsigned int tx_nob_rdma; /* nob actually rdma */
679 unsigned int tx_offset; /* offset of data into copied buffer */
681 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
682 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
685 /* we only use one or the other */
687 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
688 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
692 typedef struct kgn_conn {
693 kgn_device_t *gnc_device; /* which device */
694 struct kgn_peer *gnc_peer; /* owning peer */
695 int gnc_magic; /* magic value cleared before free */
696 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
697 struct list_head gnc_hashlist; /* stash in connection hash table */
698 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
699 struct list_head gnc_fmaq; /* txs queued for FMA */
700 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
701 __u64 gnc_peerstamp; /* peer's unique stamp */
702 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
703 __u64 gnc_my_connstamp; /* my unique connection stamp */
704 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
705 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
706 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
707 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
708 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
709 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
710 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
711 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
712 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
713 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
714 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
715 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
716 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
717 __u32 gnc_cqid; /* my completion callback id (non-unique) */
718 __u32 gnc_tx_seq; /* tx msg sequence number */
719 __u32 gnc_rx_seq; /* rx msg sequence number */
720 __u64 gnc_tx_retrans; /* # retrans on SMSG */
721 atomic_t gnc_nlive_fma; /* # live FMA */
722 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
723 atomic_t gnc_nlive_rdma; /* # live RDMA */
724 short gnc_close_sent; /* I've sent CLOSE */
725 short gnc_close_recvd; /* I've received CLOSE */
726 short gnc_in_purgatory; /* in the sin bin */
727 int gnc_error; /* errno when conn being closed due to error */
728 int gnc_peer_error; /* errno peer sent us on CLOSE */
729 kgn_conn_state_t gnc_state; /* connection state */
730 int gnc_scheduled; /* being attented to */
731 char gnc_sched_caller[30]; /* what function last called schedule */
732 int gnc_sched_line; /* what line # last called schedule */
733 atomic_t gnc_refcount; /* # users */
734 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
735 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
736 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
737 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
738 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
739 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
740 int gnc_next_tx; /* next tx to use in tx_ref_table */
741 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
742 int gnc_mbox_id; /* id of mbox in fma_blk */
743 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
744 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
745 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
746 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
747 void *remote_mbox_addr; /* save remote mbox address */
750 typedef struct kgn_mdd_purgatory {
751 gni_mem_handle_t gmp_map_key; /* mapping key */
752 struct list_head gmp_list; /* entry point for purgatory list */
753 } kgn_mdd_purgatory_t;
755 typedef struct kgn_peer {
756 struct list_head gnp_list; /* stash on global peer list */
757 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
758 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
759 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
760 kgn_net_t *gnp_net; /* net instance for this peer */
761 lnet_nid_t gnp_nid; /* who's on the other end(s) */
762 atomic_t gnp_refcount; /* # users */
763 __u32 gnp_host_id; /* ph. host ID of the peer */
764 short gnp_connecting; /* connection forming */
765 short gnp_pending_unlink; /* need last conn close to trigger unlink */
766 int gnp_last_errno; /* last error conn saw */
767 unsigned long gnp_last_alive; /* last time I had valid comms */
768 int gnp_last_dgram_errno; /* last error dgrams saw */
769 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
770 unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
771 unsigned long gnp_reconnect_interval; /* exponential backoff */
772 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
773 int gnp_down; /* rca says peer down */
774 unsigned long gnp_down_event_time; /* time peer down */
775 unsigned long gnp_up_event_time; /* time peer back up */
778 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
779 * like lnet_parse. It allows a single pointer to let us get enough
780 * information in _recv and friends */
781 typedef struct kgn_rx {
782 kgn_conn_t *grx_conn; /* connection */
783 kgn_msg_t *grx_msg; /* message */
784 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
785 int grx_eager; /* if eager, we copied msg to somewhere */
786 struct timespec grx_received; /* time this msg received */
789 typedef struct kgn_data {
790 int kgn_init; /* initialisation state */
791 int kgn_shutdown; /* shut down? */
792 int kgn_wc_kill; /* Should I repost the WC */
793 atomic_t kgn_nthreads; /* # live threads */
794 int kgn_nresets; /* number of stack resets */
795 int kgn_in_reset; /* are we in stack reset ? */
797 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
799 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
800 int kgn_ndevs; /* # devices */
802 int kgn_ruhroh_running; /* ruhroh thread is running */
803 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
804 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
805 int kgn_quiesce_trigger; /* should we quiesce ? */
806 atomic_t kgn_nquiesce; /* how many quiesced ? */
807 struct mutex kgn_quiesce_mutex; /* serialize ruhroh task, startup and shutdown */
808 int kgn_needs_reset; /* we need stack reset */
810 /* These next three members implement communication from gnilnd into
811 * the ruhroh task. To ensure correct operation of the task, code that
812 * writes into them must use memory barriers to ensure that the changes
813 * are visible to other cores in the order the members appear below. */
814 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
815 int kgn_bump_info_rdy; /* we have info needed to bump */
816 int kgn_needs_pause; /* we need to pause for network quiesce */
818 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
819 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
821 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
822 struct list_head *kgn_peers; /* hash table of all my known peers */
823 atomic_t kgn_npeers; /* # peers extant */
824 int kgn_peer_version; /* version flag for peer tables */
826 struct list_head *kgn_conns; /* conns hashed by cqid */
827 atomic_t kgn_nconns; /* # connections extant */
828 atomic_t kgn_neager_allocs; /* # of eager allocations */
829 __u64 kgn_peerstamp; /* when I started up */
830 __u64 kgn_connstamp; /* conn stamp generator */
831 int kgn_conn_version; /* version flag for conn tables */
832 int kgn_next_cqid; /* cqid generator */
834 long kgn_new_min_timeout; /* minimum timeout on any new conn */
835 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
836 spinlock_t kgn_reaper_lock; /* serialise */
838 struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
839 struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
840 struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
841 atomic_t kgn_ntx; /* # tx in use */
842 struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
844 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
845 __u64 kgn_cksum_npages; /* # pages alloc'd for checksumming */
846 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
847 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
849 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
850 long kgn_rdmaq_override; /* bytes per second override */
852 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
854 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
855 atomic_t kgn_npending_conns; /* # of conns with pending closes */
856 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
857 unsigned long kgn_last_scheduled; /* last time schedule was called */
858 unsigned long kgn_last_condresched; /* last time cond_resched was called */
859 atomic_t kgn_rev_offset; /* # of REV rdma w/misaligned offsets */
860 atomic_t kgn_rev_length; /* # of REV rdma have misaligned len */
861 atomic_t kgn_rev_copy_buff; /* # of REV rdma buffer copies */
862 struct socket *kgn_sock; /* for Apollo */
863 unsigned long free_pages_limit; /* # of free pages reserve from fma block allocations */
866 extern kgn_data_t kgnilnd_data;
867 extern kgn_tunables_t kgnilnd_tunables;
869 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
870 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
871 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
873 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
874 * and the line of the calling function to allow us to debug problematic
875 * schedule calls in the future without the programmer having to mark
876 * the location manually.
878 #define kgnilnd_schedule_conn(conn) \
879 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
881 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
882 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
885 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
887 struct task_struct *thrd = kthread_run(fn, arg, "%s_%02d", name, id);
889 return PTR_ERR(thrd);
891 atomic_inc(&kgnilnd_data.kgn_nthreads);
896 kgnilnd_thread_fini(void)
898 atomic_dec(&kgnilnd_data.kgn_nthreads);
901 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
902 * parts of the code to avoid a scheduler trip when the mutex is held
904 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
905 * has been acquired successfully, and 0 on contention.
907 * NOTE: this function follows the spin_trylock() convention, so
908 * it is negated to the down_trylock() return values! Be careful
909 * about this when converting semaphore users to mutexes.
911 * This function must not be used in interrupt context. The
912 * mutex must be released by the same task that acquired it.
914 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
917 unsigned long timeout;
919 LASSERT(!in_interrupt());
921 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
923 ret = mutex_trylock(lock);
930 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
933 _kgnilnd_debug_msg(kgn_msg_t *msg,
934 struct libcfs_debug_msg_data *data, const char *fmt, ... );
936 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
938 CFS_CHECK_STACK(msgdata, mask, cdls); \
940 if (((mask) & D_CANTMASK) != 0 || \
941 ((libcfs_debug & (mask)) != 0 && \
942 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
943 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
946 /* for most callers (level is a constant) this is resolved at compile time */
947 #define GNIDBG_MSG(level, msg, fmt, args...) \
949 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
950 static cfs_debug_limit_state_t cdls; \
951 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
952 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
953 "$$ "fmt" from %s ", ## args, \
954 libcfs_nid2str((msg)->gnm_srcnid)); \
956 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
957 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
958 "$$ "fmt" from %s ", ## args, \
959 libcfs_nid2str((msg)->gnm_srcnid)); \
963 /* user puts 'to nid' in msg for us */
964 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
966 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
967 static cfs_debug_limit_state_t cdls; \
968 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
969 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
970 "$$ "fmt" ", ## args); \
972 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
973 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
974 "$$ "fmt" ", ## args); \
979 _kgnilnd_debug_conn(kgn_conn_t *conn,
980 struct libcfs_debug_msg_data *data, const char *fmt, ... );
982 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
984 CFS_CHECK_STACK(msgdata, mask, cdls); \
986 if (((mask) & D_CANTMASK) != 0 || \
987 ((libcfs_debug & (mask)) != 0 && \
988 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
989 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
992 /* for most callers (level is a constant) this is resolved at compile time */
993 #define GNIDBG_CONN(level, conn, fmt, args...) \
995 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
996 static cfs_debug_limit_state_t cdls; \
997 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
998 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
999 "$$ "fmt" ", ## args); \
1001 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1002 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
1003 "$$ "fmt" ", ## args); \
1008 _kgnilnd_debug_tx(kgn_tx_t *tx,
1009 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1011 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
1013 CFS_CHECK_STACK(msgdata, mask, cdls); \
1015 if (((mask) & D_CANTMASK) != 0 || \
1016 ((libcfs_debug & (mask)) != 0 && \
1017 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1018 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1021 /* for most callers (level is a constant) this is resolved at compile time */
1022 #define GNIDBG_TX(level, tx, fmt, args...) \
1024 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1025 static cfs_debug_limit_state_t cdls; \
1026 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1027 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1028 "$$ "fmt" ", ## args); \
1030 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1031 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1032 "$$ "fmt" ", ## args); \
1036 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1038 if (unlikely(!(cond))) { \
1039 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1044 #define GNILND_IS_QUIESCED \
1045 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1046 atomic_read(&kgnilnd_data.kgn_nthreads))
1048 #define KGNILND_SPIN_QUIESCE \
1050 /* E.T phone home */ \
1051 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1052 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1053 while (kgnilnd_data.kgn_quiesce_trigger) { \
1054 msleep_interruptible(MSEC_PER_SEC); \
1056 /* Mom, my homework is done */ \
1057 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1058 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1061 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1062 #ifndef LIBCFS_DEBUG
1063 #error "this code uses actions inside LASSERT for ref counting"
1066 #define kgnilnd_admin_addref(atomic) \
1068 int val = atomic_inc_return(&atomic); \
1069 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1070 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1073 #define kgnilnd_admin_decref(atomic) \
1075 int val = atomic_dec_return(&atomic); \
1076 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
1077 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1080 #define kgnilnd_net_addref(net) \
1082 int val = atomic_inc_return(&net->gnn_refcount); \
1083 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1084 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1085 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1088 #define kgnilnd_net_decref(net) \
1090 int val = atomic_dec_return(&net->gnn_refcount); \
1091 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1092 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1093 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1096 #define kgnilnd_peer_addref(peer) \
1098 int val = atomic_inc_return(&peer->gnp_refcount); \
1099 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1100 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1101 libcfs_nid2str(peer->gnp_nid), val); \
1104 #define kgnilnd_peer_decref(peer) \
1106 int val = atomic_dec_return(&peer->gnp_refcount); \
1107 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1108 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1109 libcfs_nid2str(peer->gnp_nid), val); \
1111 kgnilnd_destroy_peer(peer); \
1114 #define kgnilnd_conn_addref(conn) \
1119 val = atomic_inc_return(&conn->gnc_refcount); \
1120 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1121 "conn %p refc %d to %s\n", \
1124 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1126 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1128 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1133 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1134 * through the scheduler thread to get the EP destroyed. This avoids some
1135 * messy semaphore business and allows us to reuse the connd_list and existing
1136 * linkage and avoid creating extra lists just for destroying EPs */
1138 /* Safety Disclaimer:
1139 * Q: If we decrement the refcount and then check it again, is it possible that
1140 * another caller could have passed through this macro concurrently? If so,
1141 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1143 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1144 * once we are refcount <= 2. It hinges around gnc_state and membership of
1145 * gnc_hashlist. There are two ways to find a connection - either ask for
1146 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1147 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1150 * - #1 from create (kgnilnd_create_conn)
1151 * - #2 for EP (kgnilnd_create_conn)
1152 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1153 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1155 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1156 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1158 * There could be more from TX descriptors during the lifetime of a live
1161 * If we nuke the conn before finish_connect, we won't have parallel paths
1162 * because nobody besides the dgram handler for the single outstanding
1163 * dgram can find the connection as it isn't in any searchable tables yet.
1165 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1166 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1167 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1168 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1169 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1170 * sent. Now the gnc_state == CLOSED, so we hit
1171 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1172 * - we've nuked them from the peer and CQ id tables, so we own them and
1173 * are guaranteed serial access - hence the complete lack of conn list
1174 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1175 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1176 * kgnilnd_process_conns. Then the next scheduler loop would call
1177 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1178 * kgnilnd_process_conns.
1180 * Clearly, we are totally safe. Clearly.
1183 #define kgnilnd_conn_decref(conn) \
1188 val = atomic_dec_return(&conn->gnc_refcount); \
1189 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1192 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1194 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1196 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1201 (conn->gnc_ephandle != NULL) && \
1202 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1203 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1204 kgnilnd_schedule_conn(conn); \
1205 } else if (val == 0) { \
1206 kgnilnd_destroy_conn(conn); \
1210 static inline struct list_head *
1211 kgnilnd_nid2peerlist(lnet_nid_t nid)
1213 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1215 RETURN(&kgnilnd_data.kgn_peers[hash]);
1218 static inline struct list_head *
1219 kgnilnd_netnum2netlist(__u16 netnum)
1221 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1223 RETURN(&kgnilnd_data.kgn_nets[hash]);
1227 kgnilnd_peer_active(kgn_peer_t *peer)
1229 /* Am I in the peer hash table? */
1230 return (!list_empty(&peer->gnp_list));
1233 /* need write_lock on kgn_peer_conn_lock */
1235 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1237 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1238 peer, libcfs_nid2str(peer->gnp_nid),
1239 !list_empty(&peer->gnp_conns),
1240 !list_empty(&peer->gnp_tx_queue));
1242 /* kgn_peer_conn_lock protects us from conflict with
1243 * kgnilnd_peer_notify and gnp_persistent */
1244 RETURN ((list_empty(&peer->gnp_conns)) &&
1245 (list_empty(&peer->gnp_tx_queue)));
1248 /* returns positive if error was for a clean shutdown of conn */
1250 kgnilnd_conn_clean_errno(int errno)
1252 /* - ESHUTDOWN - LND is unloading
1253 * - EUCLEAN - admin requested via "lctl del_peer"
1254 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1255 * - ENOTRECOVERABLE - stack reset
1256 * - EISCONN - cleared via "lctl push"
1257 * not doing ESTALE - that isn't clean */
1258 RETURN ((errno == 0) ||
1259 (errno == -ESHUTDOWN) ||
1260 (errno == -EUCLEAN) ||
1261 (errno == -ENETRESET) ||
1262 (errno == -EISCONN) ||
1263 (errno == -ENOTRECOVERABLE));
1266 /* returns positive if error results in purgatory hold */
1268 kgnilnd_check_purgatory_errno(int errno)
1270 /* We don't want to save the purgatory lists these cases:
1271 * - EUCLEAN - admin requested via "lctl del_peer"
1272 * - ESHUTDOWN - LND is unloading
1274 RETURN ((errno != -ESHUTDOWN) &&
1275 (errno != -EUCLEAN));
1279 /* returns positive if a purgatory hold is needed */
1281 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1285 if (conn->gnc_peer) {
1286 loopback = conn->gnc_peer->gnp_nid ==
1287 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1289 /* short circuit - a conn that didn't complete
1290 * setup never needs a purgatory hold */
1293 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1294 conn, conn->gnc_peer ?
1295 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1297 loopback, conn->gnc_close_recvd);
1299 /* we only use a purgatory hold if we've not received the CLOSE msg
1300 * from our peer - without that message, we can't know the state of
1301 * the other end of this connection and must put it into purgatory
1302 * to prevent reuse and corruption.
1303 * The theory is that a TX error can be communicated in all other cases
1305 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1306 kgnilnd_check_purgatory_errno(conn->gnc_error));
1309 static inline const char *
1310 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1312 static inline struct list_head *
1313 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1314 kgn_tx_list_state_t to_state)
1317 case GNILND_TX_PEERQ:
1318 return &peer->gnp_tx_queue;
1319 case GNILND_TX_FMAQ:
1320 return &conn->gnc_fmaq;
1321 case GNILND_TX_LIVE_FMAQ:
1322 case GNILND_TX_LIVE_RDMAQ:
1323 case GNILND_TX_DYING:
1325 case GNILND_TX_MAPQ:
1326 return &conn->gnc_device->gnd_map_tx;
1327 case GNILND_TX_RDMAQ:
1328 return &conn->gnc_device->gnd_rdmaq;
1330 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1331 CERROR("invalid state requested: %s\n",
1332 kgnilnd_tx_state2str(to_state));
1338 /* should hold tx, conn or peer lock when calling */
1340 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1341 kgn_conn_t *conn, kgn_tx_list_state_t state,
1344 struct list_head *list = NULL;
1346 /* make sure we have a sane TX state to start */
1347 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1348 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1349 list_empty(&tx->tx_list),
1350 "bad state with tx_list %s",
1351 list_empty(&tx->tx_list) ? "empty" : "not empty");
1353 /* WTF - you are already on that state buttmunch */
1354 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1355 "already at %s", kgnilnd_tx_state2str(state));
1357 /* get proper list from the state requested */
1358 list = kgnilnd_tx_state2list(peer, conn, state);
1362 case GNILND_TX_PEERQ:
1363 kgnilnd_peer_addref(peer);
1365 case GNILND_TX_ALLOCD:
1366 /* no refs needed */
1368 case GNILND_TX_FMAQ:
1369 kgnilnd_conn_addref(conn);
1371 case GNILND_TX_MAPQ:
1372 atomic_inc(&conn->gnc_device->gnd_nq_map);
1373 kgnilnd_conn_addref(conn);
1375 case GNILND_TX_LIVE_FMAQ:
1376 atomic_inc(&conn->gnc_nlive_fma);
1377 kgnilnd_conn_addref(conn);
1379 case GNILND_TX_LIVE_RDMAQ:
1380 atomic_inc(&conn->gnc_nlive_rdma);
1381 kgnilnd_conn_addref(conn);
1383 case GNILND_TX_RDMAQ:
1384 atomic_inc(&conn->gnc_nq_rdma);
1385 kgnilnd_conn_addref(conn);
1387 case GNILND_TX_DYING:
1388 kgnilnd_conn_addref(conn);
1391 CERROR("invalid state requested: %s\n",
1392 kgnilnd_tx_state2str(state));
1397 /* if this changes, change kgnilnd_alloc_tx */
1398 tx->tx_list_state = state;
1400 /* some states don't have lists - we track them in the per conn
1401 * TX table instead. Waste not, want not! */
1403 tx->tx_list_p = list;
1405 list_add_tail(&tx->tx_list, list);
1407 list_add(&tx->tx_list, list);
1409 /* set dummy list_p to make book keeping happy and let debugging
1410 * be a hair easier */
1411 tx->tx_list_p = (void *)state;
1414 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1415 kgnilnd_tx_state2str(state), list);
1419 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1420 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1422 /* These is only 1 "off-list" state */
1423 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1424 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1426 /* new_state == ALLOCD means we are deallocating this tx,
1427 * so make sure it was on a valid list to start with */
1428 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1429 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1430 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1431 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1434 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1437 switch (tx->tx_list_state) {
1438 case GNILND_TX_PEERQ:
1439 kgnilnd_peer_decref(peer);
1441 case GNILND_TX_FREED:
1442 case GNILND_TX_IDLE:
1443 case GNILND_TX_ALLOCD:
1444 /* no refs needed */
1446 case GNILND_TX_DYING:
1447 kgnilnd_conn_decref(conn);
1449 case GNILND_TX_FMAQ:
1450 kgnilnd_conn_decref(conn);
1452 case GNILND_TX_MAPQ:
1453 atomic_dec(&conn->gnc_device->gnd_nq_map);
1454 kgnilnd_conn_decref(conn);
1456 case GNILND_TX_LIVE_FMAQ:
1457 atomic_dec(&conn->gnc_nlive_fma);
1458 kgnilnd_conn_decref(conn);
1460 case GNILND_TX_LIVE_RDMAQ:
1461 atomic_dec(&conn->gnc_nlive_rdma);
1462 kgnilnd_conn_decref(conn);
1464 case GNILND_TX_RDMAQ:
1465 atomic_dec(&conn->gnc_nq_rdma);
1466 kgnilnd_conn_decref(conn);
1467 /* don't need to assert on default, already did in set */
1470 /* for ALLOCD, this might already be true, but no harm doing it again */
1471 list_del_init(&tx->tx_list);
1472 tx->tx_list_p = NULL;
1473 tx->tx_list_state = new_state;
1477 kgnilnd_tx_mapped(kgn_tx_t *tx)
1479 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1480 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1483 static inline struct list_head *
1484 kgnilnd_cqid2connlist(__u32 cqid)
1486 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1488 return (&kgnilnd_data.kgn_conns [hash]);
1491 static inline kgn_conn_t *
1492 kgnilnd_cqid2conn_locked(__u32 cqid)
1494 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1495 struct list_head *tmp;
1498 list_for_each(tmp, conns) {
1499 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1501 if (conn->gnc_cqid == cqid)
1508 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1510 kgnilnd_get_cqid_locked(void)
1516 cqid = kgnilnd_data.kgn_next_cqid++;
1517 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1521 kgnilnd_data.kgn_next_cqid = 1;
1524 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1530 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1532 kgn_tx_t *tx = NULL;
1533 kgn_conn_t *conn = NULL;
1535 /* set to NULL so any early return is an error */
1539 LASSERTF((ev_id->txe_idx > 0) &&
1540 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1541 "bogus txe_idx %d >= %d\n",
1542 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1544 LASSERTF((ev_id->txe_cqid > 0) &&
1545 (ev_id->txe_cqid < GNILND_MAX_CQID),
1546 "bogus txe_cqid %d >= %d\n",
1547 ev_id->txe_cqid, GNILND_MAX_CQID);
1549 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1550 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1553 /* Conn was destroyed? */
1554 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1555 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1558 /* just insurance */
1559 kgnilnd_conn_addref(conn);
1560 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1561 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1563 /* we know this is safe - as the TX won't be reused until AFTER
1564 * the conn is unlinked from the cqid hash, so we can use the TX
1565 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1567 spin_lock(&conn->gnc_tx_lock);
1568 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1569 spin_unlock(&conn->gnc_tx_lock);
1571 /* We could have a tx that was cleared out by other forces
1572 * lctl disconnect or del_peer. */
1574 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1575 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1576 kgnilnd_conn_decref(conn);
1580 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1581 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1582 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1584 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1585 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1586 "found tx %p txe_idx %d",
1587 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1588 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1590 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1592 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1597 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1600 /* set_normalized_timepsec isn't exported from the kernel, so
1601 * we need to do the same thing inline */
1602 static inline struct timespec
1603 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1609 sec = lhs.tv_sec - rhs.tv_sec;
1610 nsec = lhs.tv_nsec - rhs.tv_nsec;
1612 while (nsec >= NSEC_PER_SEC) {
1613 nsec -= NSEC_PER_SEC;
1617 nsec += NSEC_PER_SEC;
1626 kgnilnd_count_list(struct list_head *q)
1628 struct list_head *e;
1631 list_for_each(e, q) {
1638 /* kgnilnd_find_net adds a reference to the net it finds
1639 * this is so the net will not be removed before the calling function
1640 * has time to use the data returned. This reference needs to be released
1641 * by the calling function once it has finished using the returned net
1645 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1650 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1656 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1657 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1658 kgnilnd_net_addref(net);
1659 up_read(&kgnilnd_data.kgn_net_rw_sem);
1665 up_read(&kgnilnd_data.kgn_net_rw_sem);
1670 #ifdef CONFIG_DEBUG_SLAB
1671 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1673 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1676 int kgnilnd_dev_init(kgn_device_t *dev);
1677 void kgnilnd_dev_fini(kgn_device_t *dev);
1678 int kgnilnd_startup(lnet_ni_t *ni);
1679 void kgnilnd_shutdown(lnet_ni_t *ni);
1680 int kgnilnd_base_startup(void);
1681 void kgnilnd_base_shutdown(void);
1683 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1684 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1685 void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
1686 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1688 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1689 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1690 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1691 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1692 lnet_msg_t *lntmsg, void **new_private);
1693 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1694 int delayed, unsigned int niov,
1695 struct iovec *iov, lnet_kiov_t *kiov,
1696 unsigned int offset, unsigned int mlen, unsigned int rlen);
1698 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1700 /* purgatory functions */
1701 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1702 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1703 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1704 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1706 void kgnilnd_update_reaper_timeout(long timeout);
1707 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1708 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1709 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1710 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1711 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1712 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
1713 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1715 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1716 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
1717 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1718 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1720 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1721 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1722 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1723 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1724 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1725 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1726 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1727 void kgnilnd_consume_rx(kgn_rx_t *rx);
1729 void kgnilnd_schedule_device(kgn_device_t *dev);
1730 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1731 void kgnilnd_schedule_device_timer(unsigned long arg);
1733 int kgnilnd_reaper(void *arg);
1734 int kgnilnd_scheduler(void *arg);
1735 int kgnilnd_dgram_mover(void *arg);
1736 int kgnilnd_rca(void *arg);
1738 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1739 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1740 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1741 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1742 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1743 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1744 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1745 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1746 void kgnilnd_peer_alive(kgn_peer_t *peer);
1747 void kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive);
1748 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1749 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1750 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1751 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1753 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1754 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1755 void kgnilnd_wakeup_rca_thread(void);
1756 int kgnilnd_start_rca_thread(void);
1757 int kgnilnd_get_node_state(__u32 nid);
1759 int kgnilnd_tunables_init(void);
1760 void kgnilnd_tunables_fini(void);
1761 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1763 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1764 void kgnilnd_pause_threads(void);
1765 int kgnilnd_hw_in_quiesce(void);
1766 int kgnilnd_check_hw_quiesce(void);
1767 void kgnilnd_quiesce_wait(char *reason);
1768 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1769 int kgnilnd_ruhroh_thread(void *arg);
1770 void kgnilnd_reset_stack(void);
1771 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1773 void kgnilnd_insert_sysctl(void);
1774 void kgnilnd_remove_sysctl(void);
1775 void kgnilnd_proc_init(void);
1776 void kgnilnd_proc_fini(void);
1779 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1781 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1782 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1783 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
1785 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1786 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1787 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1788 int kgnilnd_cancel_dgrams(kgn_device_t *dev);
1789 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1791 int kgnilnd_dgram_waitq(void *arg);
1793 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1795 /* struct2str functions - we don't use a default: case to cause the compile
1796 * to fail if there is a missing case. This allows us to hide these down here
1797 * out of the way but ensure we'll catch any updates to the enum/types
1800 #define DO_TYPE(x) case x: return #x;
1801 static inline const char *
1802 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1804 /* Only want single char string for this */
1806 case GNILND_FMABLK_IDLE:
1808 case GNILND_FMABLK_PHYS:
1810 case GNILND_FMABLK_VIRT:
1812 case GNILND_FMABLK_FREED:
1815 return "<unknown state>";
1818 static inline const char *
1819 kgnilnd_msgtype2str(int type)
1822 DO_TYPE(GNILND_MSG_NONE);
1823 DO_TYPE(GNILND_MSG_NOOP);
1824 DO_TYPE(GNILND_MSG_IMMEDIATE);
1825 DO_TYPE(GNILND_MSG_PUT_REQ);
1826 DO_TYPE(GNILND_MSG_PUT_NAK);
1827 DO_TYPE(GNILND_MSG_PUT_ACK);
1828 DO_TYPE(GNILND_MSG_PUT_DONE);
1829 DO_TYPE(GNILND_MSG_GET_REQ);
1830 DO_TYPE(GNILND_MSG_GET_NAK);
1831 DO_TYPE(GNILND_MSG_GET_DONE);
1832 DO_TYPE(GNILND_MSG_CLOSE);
1833 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1834 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1835 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1836 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1837 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1838 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1839 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1841 return "<unknown msg type>";
1844 static inline const char *
1845 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1848 DO_TYPE(GNILND_TX_IDLE);
1849 DO_TYPE(GNILND_TX_ALLOCD);
1850 DO_TYPE(GNILND_TX_PEERQ);
1851 DO_TYPE(GNILND_TX_MAPQ);
1852 DO_TYPE(GNILND_TX_FMAQ);
1853 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1854 DO_TYPE(GNILND_TX_RDMAQ);
1855 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1856 DO_TYPE(GNILND_TX_DYING);
1857 DO_TYPE(GNILND_TX_FREED);
1859 return "<unknown state>";
1862 static inline const char *
1863 kgnilnd_conn_state2str(kgn_conn_t *conn)
1865 kgn_conn_state_t state = conn->gnc_state;
1867 DO_TYPE(GNILND_CONN_DUMMY);
1868 DO_TYPE(GNILND_CONN_LISTEN);
1869 DO_TYPE(GNILND_CONN_CONNECTING);
1870 DO_TYPE(GNILND_CONN_ESTABLISHED);
1871 DO_TYPE(GNILND_CONN_CLOSING);
1872 DO_TYPE(GNILND_CONN_CLOSED);
1873 DO_TYPE(GNILND_CONN_DONE);
1874 DO_TYPE(GNILND_CONN_DESTROY_EP);
1879 static inline const char *
1880 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1882 kgn_connreq_type_t type = connreq->gncr_type;
1885 DO_TYPE(GNILND_CONNREQ_REQ);
1886 DO_TYPE(GNILND_CONNREQ_NAK);
1887 DO_TYPE(GNILND_CONNREQ_CLOSE);
1892 static inline const char *
1893 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1895 kgn_dgram_state_t state = dgram->gndg_state;
1898 DO_TYPE(GNILND_DGRAM_USED);
1899 DO_TYPE(GNILND_DGRAM_POSTING);
1900 DO_TYPE(GNILND_DGRAM_POSTED);
1901 DO_TYPE(GNILND_DGRAM_PROCESSING);
1902 DO_TYPE(GNILND_DGRAM_DONE);
1903 DO_TYPE(GNILND_DGRAM_CANCELED);
1908 static inline const char *
1909 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1911 kgn_dgram_type_t type = dgram->gndg_type;
1914 DO_TYPE(GNILND_DGRAM_REQ);
1915 DO_TYPE(GNILND_DGRAM_WC_REQ);
1916 DO_TYPE(GNILND_DGRAM_NAK);
1917 DO_TYPE(GNILND_DGRAM_CLOSE);
1922 static inline const char *
1923 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
1926 DO_TYPE(GNILND_DGRAM_REQ);
1927 DO_TYPE(GNILND_DGRAM_WC_REQ);
1928 DO_TYPE(GNILND_DGRAM_NAK);
1929 DO_TYPE(GNILND_DGRAM_CLOSE);
1936 /* API wrapper functions - include late to pick up all of the other defines */
1937 #include "gnilnd_api_wrap.h"
1939 /* pulls in tunables per platform and adds in nid/nic conversion
1940 * if RCA wasn't available at build time */
1941 #include "gnilnd_hss_ops.h"
1943 #if defined(CONFIG_CRAY_GEMINI)
1944 #include "gnilnd_gemini.h"
1945 #elif defined(CONFIG_CRAY_ARIES)
1946 #include "gnilnd_aries.h"
1948 #error "Undefined Network Hardware Type"
1951 #endif /* _GNILND_GNILND_H_ */