2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by: Eric Barton <eric@bartonsoftware.com>
7 * Author: Nic Henke <nic@cray.com>
8 * Author: James Shimek <jshimek@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
26 #ifndef _GNILND_GNILND_H_
27 #define _GNILND_GNILND_H_
29 #include <linux/module.h>
30 #include <linux/kernel.h>
32 #include <linux/string.h>
33 #include <linux/stat.h>
34 #include <linux/errno.h>
35 #ifdef HAVE_LINUX_KERNEL_LOCK
36 #include <linux/smp_lock.h>
38 #include <linux/unistd.h>
39 #include <linux/uio.h>
40 #include <linux/time.h>
41 #include <asm/timex.h>
43 #include <asm/system.h>
44 #include <asm/uaccess.h>
47 #include <linux/init.h>
49 #include <linux/file.h>
50 #include <linux/stat.h>
51 #include <linux/list.h>
52 #include <linux/kmod.h>
53 #include <linux/sysctl.h>
54 #include <linux/kthread.h>
55 #include <linux/nmi.h>
59 #include <linux/nmi.h>
61 #define DEBUG_SUBSYSTEM S_LND
63 #include <libcfs/linux/kp30.h>
64 #include <libcfs/libcfs.h>
65 #include <lnet/lnet.h>
66 #include <lnet/lib-lnet.h>
67 #include <lnet/lnet-sysctl.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)
110 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
111 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
112 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
113 #define GNILND_COMPUTE 1 /* compute image */
115 #define GNILND_SCHED_THREADS 3 /* default # of kgnilnd_scheduler threads */
116 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
117 #define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
118 #define GNILND_COMPUTE 0 /* service image */
121 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
122 #define GNILND_EXTRA_BITS 1
123 /* maximum number of conns & bits for cqid in the SMSG event data */
124 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
125 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
126 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
127 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
128 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
130 /* need sane upper bound to limit copy overhead */
131 #define GNILND_MAX_IMMEDIATE (64<<10)
133 /* payload size to add to the base mailbox size
134 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
135 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
136 * the calculation return from that function.*/
137 #define GNILND_MBOX_PAYLOAD \
138 (GNILND_MAX_MSG_SIZE * \
139 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
141 /* timeout -> deadman timer for kgni mdd holds */
142 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
144 /* timeout for failing sends in t is in jiffies*/
145 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
147 /* time when to release from purgatory in the reaper thread in jiffies */
148 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
150 /* Macro for finding last_rx 2 datapoints are compared
151 * and the most recent one in jiffies is returned.
153 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
154 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
156 /************************************************************************
157 * Enum, flag and tag data
159 #define GNILND_INIT_NOTHING 0
160 #define GNILND_INIT_DATA 1
161 #define GNILND_INIT_ALL 2
163 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
164 #define GNILND_BUF_NONE 0 /* buffer type not set */
165 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
166 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
167 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
168 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
169 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
170 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
172 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
173 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
174 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
175 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
176 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
178 /* stash above max CQID to avoid any collision */
179 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
180 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
182 /* kgn_msg_t::gnm_type */
183 #define GNILND_MSG_NONE 0x00 /* illegal message */
184 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
185 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
186 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
187 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
188 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
189 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
190 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
191 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
192 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
193 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
194 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
195 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
196 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
197 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
198 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
199 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
200 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
202 /* defines for gnc_*scheduled states */
203 #define GNILND_CONN_IDLE 0
204 #define GNILND_CONN_SCHED 1
205 #define GNILND_CONN_WANTS_SCHED 2
206 #define GNILND_CONN_PROCESS 3
208 #define GNILND_DEV_IDLE 0
209 #define GNILND_DEV_IRQ 1
210 #define GNILND_DEV_LOOP 2
212 #define GNILND_DGRAM_IDLE 0
213 #define GNILND_DGRAM_SCHED 1
214 #define GNILND_DGRAM_PROCESS 2
216 #define GNILND_PEER_IDLE 0
217 #define GNILND_PEER_CONNECT 1
218 #define GNILND_PEER_POSTING 2
219 #define GNILND_PEER_POSTED 3
220 #define GNILND_PEER_NEEDS_DEATH 4
221 #define GNILND_PEER_KILL 5
223 /* for gnc_close_recvd */
224 #define GNILND_CLOSE_RX 1
225 #define GNILND_CLOSE_INJECT1 2
226 #define GNILND_CLOSE_INJECT2 3
227 #define GNILND_CLOSE_EARLY 4
229 /* defines for why quiesce trigger set */
230 #define GNILND_QUIESCE_IDLE 0
231 #define GNILND_QUIESCE_ADMIN 1
232 #define GNILND_QUIESCE_RESET 2
233 #define GNILND_QUIESCE_HW_QUIESCE 3
235 #define GNILND_PEER_CLEAN 0
236 #define GNILND_PEER_PERSISTING 1
238 #define GNILND_DEL_CONN 0
239 #define GNILND_DEL_PEER 1
240 #define GNILND_CLEAR_PURGATORY 2
242 #define GNILND_RCA_NODE_UP 0
243 #define GNILND_RCA_NODE_DOWN 1
244 #define GNILND_RCA_NODE_UNKNOWN 2
246 /* defines for reverse RDMA states */
247 #define GNILND_REVERSE_NONE 0
248 #define GNILND_REVERSE_GET 1
249 #define GNILND_REVERSE_PUT 2
250 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
252 typedef enum kgn_fmablk_state {
253 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
254 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
255 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
256 GNILND_FMABLK_FREED, /* after free */
257 } kgn_fmablk_state_t;
259 typedef enum kgn_tx_list_state {
260 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
261 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
262 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
263 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
264 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
265 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
266 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
267 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
268 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
269 GNILND_TX_FREED /* TX is free! */
270 } kgn_tx_list_state_t;
272 typedef enum kgn_conn_state {
273 /* don't start @ 0 - prevent memset(0) badness */
274 GNILND_CONN_DUMMY = 0,
276 GNILND_CONN_CONNECTING,
277 GNILND_CONN_ESTABLISHED,
281 GNILND_CONN_DESTROY_EP
284 /* changing these requires a change to GNILND_CONNREQ_VERSION and
285 * will result in dropped packets instead of NAKs. Adding to this is
286 * acceptable without changing the CONNREQ_VERSION, but code should
287 * be ready to handle NAKs on version mismatch */
288 typedef enum kgn_connreq_type {
289 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
290 GNILND_CONNREQ_NAK, /* NO soup for you! */
291 GNILND_CONNREQ_CLOSE, /* we should see other people */
292 } kgn_connreq_type_t;
294 typedef enum kgn_dgram_state {
295 /* don't use 0 to avoid thinking a memset of zero is valid data */
296 GNILND_DGRAM_USED = 1,
297 GNILND_DGRAM_POSTING,
299 GNILND_DGRAM_PROCESSING,
300 GNILND_DGRAM_CANCELED,
304 typedef enum kgn_dgram_type {
305 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
306 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
307 GNILND_DGRAM_NAK, /* NO soup for you! */
308 GNILND_DGRAM_CLOSE, /* we should see other people */
311 /************************************************************************
312 * Wire message structs. These are sent in sender's byte order
313 * (i.e. receiver checks magic and flips if required).
316 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
317 #define GNILND_DGRAM_MAGIC 0x0DDBA11
319 /* kgn_msg_t - FMA/SMSG wire struct
321 * - added checksum to FMA
322 * moved seq before paylod
323 * WIRE_ATTR added for alignment
325 * added gnm_payload_len for FMA payload size
327 * added gncm_retval to completion, allowing return code transmission
330 * changed how CQID and TX ids are assigned
332 * added retval on CLOSE
334 * added payload checksumming
336 * reworked checksumming a bit, changed payload checksums
338 #define GNILND_MSG_VERSION 8
339 /* kgn_connreq_t connection request datagram wire struct
344 #define GNILND_CONNREQ_VERSION 2
346 typedef struct kgn_gniparams {
347 __u32 gnpr_host_id; /* ph. host ID of the NIC */
348 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
349 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
350 } WIRE_ATTR kgn_gniparams_t;
352 typedef struct kgn_nak_data {
353 __s32 gnnd_errno; /* errno reason for NAK */
355 } WIRE_ATTR kgn_nak_data_t;
357 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
358 * without breaking the ability for us to properly NAK someone */
359 typedef struct kgn_connreq { /* connection request/response */
360 __u32 gncr_magic; /* I'm an gnilnd connreq */
361 __u32 gncr_cksum; /* checksum (0 == disabled) */
362 __u16 gncr_type; /* REQ, NAK, etc */
363 __u16 gncr_version; /* this is my version number */
364 __u32 gncr_timeout; /* sender's timeout */
365 __u64 gncr_srcnid; /* sender's NID */
366 __u64 gncr_dstnid; /* who sender expects to listen */
367 __u64 gncr_peerstamp; /* sender's instance stamp */
368 __u64 gncr_connstamp; /* sender's connection stamp */
370 /* everything before this needs to stay static, adding after should
371 * result in a change to GNILND_CONNREQ_VERSION */
374 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
375 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
377 } WIRE_ATTR kgn_connreq_t;
380 gni_mem_handle_t gnrd_key;
383 } WIRE_ATTR kgn_rdma_desc_t;
386 lnet_hdr_t gnim_hdr; /* LNet header */
387 /* LNet payload is in FMA "Message Data" */
388 } WIRE_ATTR kgn_immediate_msg_t;
391 lnet_hdr_t gnprm_hdr; /* LNet header */
392 __u64 gnprm_cookie; /* opaque completion cookie */
393 } WIRE_ATTR kgn_putreq_msg_t;
396 __u64 gnpam_src_cookie; /* reflected completion cookie */
397 __u64 gnpam_dst_cookie; /* opaque completion cookie */
398 __u16 gnpam_payload_cksum; /* checksum for get msg */
399 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
400 } WIRE_ATTR kgn_putack_msg_t;
403 lnet_hdr_t gngm_hdr; /* LNet header */
404 __u64 gngm_cookie; /* opaque completion cookie */
405 __u16 gngm_payload_cksum; /* checksum for put msg */
406 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
407 } WIRE_ATTR kgn_get_msg_t;
410 int gncm_retval; /* error on NAK, size on REQ */
411 __u64 gncm_cookie; /* reflected completion cookie */
412 } WIRE_ATTR kgn_completion_msg_t;
414 typedef struct { /* NB must fit in FMA "Prefix" */
415 __u32 gnm_magic; /* I'm an gni message */
416 __u16 gnm_version; /* this is my version number */
417 __u16 gnm_type; /* msg type */
418 __u64 gnm_srcnid; /* sender's NID */
419 __u64 gnm_connstamp; /* sender's connection stamp */
420 __u32 gnm_seq; /* incrementing sequence number */
421 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
422 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
423 __u32 gnm_payload_len; /* size of the FMA payload sent */
425 kgn_immediate_msg_t immediate;
426 kgn_putreq_msg_t putreq;
427 kgn_putack_msg_t putack;
429 kgn_completion_msg_t completion;
431 } WIRE_ATTR kgn_msg_t;
433 /************************************************************************
434 * runtime tunable data
437 typedef struct kgn_tunables {
438 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
439 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
440 int *kgn_credits; /* # concurrent sends */
441 int *kgn_fma_cq_size; /* # entries in receive CQ */
442 int *kgn_peer_credits; /* # LNet peer credits */
443 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
444 int *kgn_timeout; /* comms timeout (seconds) */
445 int *kgn_max_immediate; /* immediate payload breakpoint */
446 int *kgn_checksum; /* checksum data */
447 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
448 int *kgn_bte_dlvr_mode; /* BTE delivery mode mask */
449 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
450 int *kgn_ptag; /* PTAG for cdm_create */
451 int *kgn_max_retransmits; /* max number of FMA retransmits */
452 int *kgn_nwildcard; /* # wildcard per net to post */
453 int *kgn_nice; /* nice value for kgnilnd threads */
454 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
455 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
456 int *kgn_peer_hash_size; /* size of kgn_peers */
457 int *kgn_peer_health; /* enable/disable peer health */
458 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
459 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
460 int *kgn_mbox_per_block; /* mailboxes per fmablk */
461 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
462 int *kgn_mbox_credits; /* max credits per fma */
463 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
464 int *kgn_net_hash_size; /* size of kgn_net_ht */
465 int *kgn_hardware_timeout; /* max time for a message to get across the network */
466 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
467 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
468 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
469 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
470 int *kgn_reverse_rdma; /* Reverse RDMA setting */
471 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
472 cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
476 typedef struct kgn_mbox_info {
477 lnet_nid_t mbx_prev_nid;
478 lnet_nid_t mbx_prev_purg_nid;
479 unsigned long mbx_create_conn_memset;
480 unsigned long mbx_add_purgatory;
481 unsigned long mbx_detach_of_purgatory;
482 unsigned long mbx_release_from_purgatory;
483 unsigned long mbx_release_purg_active_dgram;
485 int mbx_nallocs_total;
488 typedef struct kgn_fma_memblock {
489 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
490 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
491 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
492 int gnm_num_mboxs; /* total mboxes allocated */
493 int gnm_avail_mboxs; /* number of available mailboxes in the block */
494 int gnm_held_mboxs; /* number of purgatory held mailboxes */
495 int gnm_mbox_size; /* size of the single mailbox */
496 int gnm_next_avail_mbox; /* next available mailbox */
497 long gnm_max_timeout; /* max timeout for possible purgatory hold */
498 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
499 void *gnm_block; /* pointer to mem. block */
500 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
501 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
502 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
503 } kgn_fma_memblock_t;
505 typedef struct kgn_device {
506 gni_nic_handle_t gnd_handle; /* device handle */
507 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
508 gni_err_handle_t gnd_err_handle; /* device error handle */
509 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
510 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
511 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
512 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
513 struct mutex gnd_cq_mutex; /* CQ access serialization */
514 __u32 gnd_host_id; /* ph. host ID of the NIC */
515 int gnd_id; /* device id, also index in kgn_devices */
516 __u32 gnd_nid; /* ph host ID translated to NID */
517 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
518 struct semaphore gnd_fmablk_sem; /* semaphore for FMA block memory alloc/free */
519 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
520 atomic_t gnd_nfmablk; /* # of fmablk live */
521 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
522 atomic_t gnd_neps; /* # EP allocated to conns */
523 short gnd_ready; /* stuff to do in scheduler thread */
524 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
525 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
526 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
527 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
528 struct list_head gnd_connd_peers; /* peers waiting for a connection */
529 spinlock_t gnd_connd_lock; /* serialise connd_peers */
530 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
531 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
532 int gnd_dgram_ready; /* dgrams need movin' */
533 struct list_head *gnd_dgrams; /* nid hash to dgrams */
534 atomic_t gnd_ndgrams; /* # dgrams extant */
535 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post on device */
536 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
537 struct list_head gnd_map_list; /* list of all mapped regions */
538 int gnd_map_version; /* version flag for map list */
539 struct timer_list gnd_map_timer; /* wakey-wakey */
540 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
541 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
542 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
543 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
544 __u32 gnd_map_nphys; /* # TX phys mappings */
545 __u32 gnd_map_physnop; /* # TX phys pages mapped */
546 __u32 gnd_map_nvirt; /* # TX virt mappings */
547 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
548 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
549 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
550 int gnd_map_attempt; /* last map attempt # */
551 unsigned long gnd_last_map; /* map timeout base */
552 struct list_head gnd_rdmaq; /* RDMA to be sent */
553 spinlock_t gnd_rdmaq_lock; /* play nice with others */
554 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
555 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
556 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
557 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
558 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
559 atomic_t gnd_short_ntx; /* TX stats: short messages */
560 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
561 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
562 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
563 atomic_t gnd_short_nrx; /* RX stats: short messages */
564 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
565 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
566 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
567 atomic_t gnd_fast_try; /* # of times fast send tried */
568 atomic_t gnd_fast_ok; /* # of times fast send ok */
569 atomic_t gnd_fast_block; /* # of times fast send blocked */
570 unsigned long gnd_mutex_delay;
571 atomic_t gnd_n_yield;
572 atomic_t gnd_n_schedule;
573 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
574 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
577 typedef struct kgn_net {
578 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
579 kgn_device_t *gnn_dev; /* device for this net */
580 lnet_ni_t *gnn_ni; /* network interface instance */
581 atomic_t gnn_refcount; /* # current references */
582 int gnn_shutdown; /* lnd_shutdown set */
583 __u16 gnn_netnum; /* stash netnum for quicker lookup */
586 static inline lnet_nid_t
587 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
589 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
592 static inline lnet_nid_t
593 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
595 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
598 /* The code for this is a bit ugly - but really this just boils down to a __u64
599 * that can have various parts accessed separately.
601 * The lower 32 bits is the ID
602 * we give to SMSG for our completion event - it needs to be globally unique across
603 * all TX currently in flight. We separate that out into the CQID so that we can
604 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
605 * the actual TX out of the per-connection gnc_tx_ref_table.
607 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
608 * has a unique value we can send with RDMA setup messages to ensure the completion for
609 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
610 * reuse is separated.
613 typedef struct kgn_tx_ev_id {
615 __u64 txe_cookie; /* are you my mommy ? */
617 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
619 __u32 txe_smsg_id; /* ID for SMSG CQ event */
620 /* N.B: Never ever ever ever use the bit shifts directly,
621 * you are just asking for a world of pain and are at the
622 * mercy of the compiler layouts */
624 __u32 txe_cqid :GNILND_CQID_NBITS;
625 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
632 typedef struct kgn_dgram {
633 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
634 kgn_dgram_state_t gndg_state; /* state of this dgram */
635 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
636 __u32 gndg_magic; /* saftey word */
637 unsigned long gndg_post_time; /* time when we posted */
638 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
639 kgn_connreq_t gndg_conn_out; /* connreq from local node */
640 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
643 typedef struct kgn_tx { /* message descriptor */
644 struct list_head tx_list; /* TX queues - peer, conn, rdma */
645 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
646 struct list_head *tx_list_p; /* pointer to current list */
647 struct kgn_conn *tx_conn; /* owning conn */
648 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
649 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
650 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
651 struct list_head tx_map_list; /* list entry on device map list */
652 unsigned int tx_nob; /* # bytes of payload */
653 int tx_buftype; /* payload buffer type */
654 int tx_phys_npages; /* # physical pages */
655 gni_mem_handle_t tx_map_key; /* mapping key */
656 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
657 gni_mem_segment_t *tx_phys; /* page descriptors */
658 kgn_msg_t tx_msg; /* FMA message buffer */
659 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
660 __u8 tx_state; /* state of the descriptor */
661 int tx_retrans; /* retrans count of RDMA */
662 int tx_rc; /* if we need to stash the ret code until we see completion */
663 void *tx_buffer; /* source/sink buffer */
664 void *tx_buffer_copy; /* pointer to page aligned buffer */
665 unsigned int tx_nob_rdma; /* nob actually rdma */
666 unsigned int tx_offset; /* offset of data into copied buffer */
668 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
669 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
672 /* we only use one or the other */
674 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
675 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
679 typedef struct kgn_conn {
680 kgn_device_t *gnc_device; /* which device */
681 struct kgn_peer *gnc_peer; /* owning peer */
682 int gnc_magic; /* magic value cleared before free */
683 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
684 struct list_head gnc_hashlist; /* stash in connection hash table */
685 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
686 struct list_head gnc_fmaq; /* txs queued for FMA */
687 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
688 __u64 gnc_peerstamp; /* peer's unique stamp */
689 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
690 __u64 gnc_my_connstamp; /* my unique connection stamp */
691 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
692 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
693 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
694 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
695 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
696 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
697 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
698 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
699 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
700 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
701 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
702 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
703 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
704 __u32 gnc_cqid; /* my completion callback id (non-unique) */
705 __u32 gnc_tx_seq; /* tx msg sequence number */
706 __u32 gnc_rx_seq; /* rx msg sequence number */
707 __u64 gnc_tx_retrans; /* # retrans on SMSG */
708 atomic_t gnc_nlive_fma; /* # live FMA */
709 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
710 atomic_t gnc_nlive_rdma; /* # live RDMA */
711 short gnc_close_sent; /* I've sent CLOSE */
712 short gnc_close_recvd; /* I've received CLOSE */
713 short gnc_in_purgatory; /* in the sin bin */
714 int gnc_error; /* errno when conn being closed due to error */
715 int gnc_peer_error; /* errno peer sent us on CLOSE */
716 kgn_conn_state_t gnc_state; /* connection state */
717 int gnc_scheduled; /* being attented to */
718 char gnc_sched_caller[30]; /* what function last called schedule */
719 int gnc_sched_line; /* what line # last called schedule */
720 atomic_t gnc_refcount; /* # users */
721 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
722 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
723 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
724 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
725 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
726 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
727 int gnc_next_tx; /* next tx to use in tx_ref_table */
728 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
729 int gnc_mbox_id; /* id of mbox in fma_blk */
730 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
731 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
732 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
733 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
734 void *remote_mbox_addr; /* save remote mbox address */
737 typedef struct kgn_mdd_purgatory {
738 gni_mem_handle_t gmp_map_key; /* mapping key */
739 struct list_head gmp_list; /* entry point for purgatory list */
740 } kgn_mdd_purgatory_t;
742 typedef struct kgn_peer {
743 struct list_head gnp_list; /* stash on global peer list */
744 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
745 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
746 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
747 kgn_net_t *gnp_net; /* net instance for this peer */
748 lnet_nid_t gnp_nid; /* who's on the other end(s) */
749 atomic_t gnp_refcount; /* # users */
750 __u32 gnp_host_id; /* ph. host ID of the peer */
751 short gnp_connecting; /* connection forming */
752 short gnp_pending_unlink; /* need last conn close to trigger unlink */
753 int gnp_last_errno; /* last error conn saw */
754 unsigned long gnp_last_alive; /* last time I had valid comms */
755 int gnp_last_dgram_errno; /* last error dgrams saw */
756 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
757 unsigned long gnp_reconnect_time; /* CURRENT_SECONDS when reconnect OK */
758 unsigned long gnp_reconnect_interval; /* exponential backoff */
759 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
760 int gnp_down; /* rca says peer down */
761 unsigned long gnp_down_event_time; /* time peer down */
762 unsigned long gnp_up_event_time; /* time peer back up */
765 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
766 * like lnet_parse. It allows a single pointer to let us get enough
767 * information in _recv and friends */
768 typedef struct kgn_rx {
769 kgn_conn_t *grx_conn; /* connection */
770 kgn_msg_t *grx_msg; /* message */
771 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
772 int grx_eager; /* if eager, we copied msg to somewhere */
773 struct timespec grx_received; /* time this msg received */
776 typedef struct kgn_data {
777 int kgn_init; /* initialisation state */
778 int kgn_shutdown; /* shut down? */
779 int kgn_wc_kill; /* Should I repost the WC */
780 atomic_t kgn_nthreads; /* # live threads */
781 int kgn_nresets; /* number of stack resets */
782 int kgn_in_reset; /* are we in stack reset ? */
784 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
786 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
787 int kgn_ndevs; /* # devices */
789 int kgn_ruhroh_running; /* ruhroh thread is running */
790 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
791 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
792 int kgn_quiesce_trigger; /* should we quiesce ? */
793 atomic_t kgn_nquiesce; /* how many quiesced ? */
794 struct semaphore kgn_quiesce_sem; /* serialize ruhroh task, startup and shutdown */
795 int kgn_needs_reset; /* we need stack reset */
797 /* These next three members implement communication from gnilnd into
798 * the ruhroh task. To ensure correct operation of the task, code that
799 * writes into them must use memory barriers to ensure that the changes
800 * are visible to other cores in the order the members appear below. */
801 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
802 int kgn_bump_info_rdy; /* we have info needed to bump */
803 int kgn_needs_pause; /* we need to pause for network quiesce */
805 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
806 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
808 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
809 struct list_head *kgn_peers; /* hash table of all my known peers */
810 atomic_t kgn_npeers; /* # peers extant */
811 int kgn_peer_version; /* version flag for peer tables */
813 struct list_head *kgn_conns; /* conns hashed by cqid */
814 atomic_t kgn_nconns; /* # connections extant */
815 __u64 kgn_peerstamp; /* when I started up */
816 __u64 kgn_connstamp; /* conn stamp generator */
817 int kgn_conn_version; /* version flag for conn tables */
818 int kgn_next_cqid; /* cqid generator */
820 long kgn_new_min_timeout; /* minimum timeout on any new conn */
821 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
822 spinlock_t kgn_reaper_lock; /* serialise */
824 cfs_mem_cache_t *kgn_rx_cache; /* rx descriptor space */
825 cfs_mem_cache_t *kgn_tx_cache; /* tx descriptor memory */
826 cfs_mem_cache_t *kgn_tx_phys_cache; /* tx phys descriptor memory */
827 atomic_t kgn_ntx; /* # tx in use */
828 cfs_mem_cache_t *kgn_dgram_cache; /* outgoing datagrams */
830 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
831 __u64 kgn_cksum_npages; /* Number of pages allocated for checksumming */
832 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
833 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
835 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
836 long kgn_rdmaq_override; /* bytes per second override */
838 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
840 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
841 atomic_t kgn_npending_conns; /* # of conns with pending closes */
842 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
843 unsigned long kgn_last_scheduled; /* last time schedule was called in a sched thread */
844 unsigned long kgn_last_condresched; /* last time cond_resched was called in a sched thread */
845 atomic_t kgn_rev_offset; /* number of time REV rdma have been misaligned offsets */
846 atomic_t kgn_rev_length; /* Number of times REV rdma have been misaligned lengths */
847 atomic_t kgn_rev_copy_buff; /* Number of times REV rdma have had to make a copy buffer */
850 extern kgn_data_t kgnilnd_data;
851 extern kgn_tunables_t kgnilnd_tunables;
853 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
854 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
855 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
857 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
858 * and the line of the calling function to allow us to debug problematic
859 * schedule calls in the future without the programmer having to mark
860 * the location manually.
862 #define kgnilnd_schedule_conn(conn) \
863 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
865 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
866 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
869 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
871 struct task_struct *thrd = kthread_run(fn, arg, "%s_%02d", name, id);
873 return PTR_ERR(thrd);
875 atomic_inc(&kgnilnd_data.kgn_nthreads);
880 kgnilnd_thread_fini(void)
882 atomic_dec(&kgnilnd_data.kgn_nthreads);
885 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
886 * parts of the code to avoid a scheduler trip when the mutex is held
888 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
889 * has been acquired successfully, and 0 on contention.
891 * NOTE: this function follows the spin_trylock() convention, so
892 * it is negated to the down_trylock() return values! Be careful
893 * about this when converting semaphore users to mutexes.
895 * This function must not be used in interrupt context. The
896 * mutex must be released by the same task that acquired it.
898 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
901 unsigned long timeout;
903 LASSERT(!in_interrupt());
905 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
907 ret = mutex_trylock(lock);
914 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
917 _kgnilnd_debug_msg(kgn_msg_t *msg,
918 struct libcfs_debug_msg_data *data, const char *fmt, ... );
920 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
922 CFS_CHECK_STACK(msgdata, mask, cdls); \
924 if (((mask) & D_CANTMASK) != 0 || \
925 ((libcfs_debug & (mask)) != 0 && \
926 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
927 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
930 /* for most callers (level is a constant) this is resolved at compile time */
931 #define GNIDBG_MSG(level, msg, fmt, args...) \
933 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
934 static cfs_debug_limit_state_t cdls; \
935 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
936 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
937 "$$ "fmt" from %s ", ## args, \
938 libcfs_nid2str((msg)->gnm_srcnid)); \
940 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
941 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
942 "$$ "fmt" from %s ", ## args, \
943 libcfs_nid2str((msg)->gnm_srcnid)); \
947 /* user puts 'to nid' in msg for us */
948 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
950 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
951 static cfs_debug_limit_state_t cdls; \
952 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
953 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
954 "$$ "fmt" ", ## args); \
956 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
957 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
958 "$$ "fmt" ", ## args); \
963 _kgnilnd_debug_conn(kgn_conn_t *conn,
964 struct libcfs_debug_msg_data *data, const char *fmt, ... );
966 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
968 CFS_CHECK_STACK(msgdata, mask, cdls); \
970 if (((mask) & D_CANTMASK) != 0 || \
971 ((libcfs_debug & (mask)) != 0 && \
972 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
973 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
976 /* for most callers (level is a constant) this is resolved at compile time */
977 #define GNIDBG_CONN(level, conn, fmt, args...) \
979 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
980 static cfs_debug_limit_state_t cdls; \
981 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
982 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
983 "$$ "fmt" ", ## args); \
985 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
986 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
987 "$$ "fmt" ", ## args); \
992 _kgnilnd_debug_tx(kgn_tx_t *tx,
993 struct libcfs_debug_msg_data *data, const char *fmt, ... );
995 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
997 CFS_CHECK_STACK(msgdata, mask, cdls); \
999 if (((mask) & D_CANTMASK) != 0 || \
1000 ((libcfs_debug & (mask)) != 0 && \
1001 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1002 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1005 /* for most callers (level is a constant) this is resolved at compile time */
1006 #define GNIDBG_TX(level, tx, fmt, args...) \
1008 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1009 static cfs_debug_limit_state_t cdls; \
1010 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1011 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1012 "$$ "fmt" ", ## args); \
1014 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1015 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1016 "$$ "fmt" ", ## args); \
1020 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1022 if (unlikely(!(cond))) { \
1023 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1028 #define GNILND_IS_QUIESCED \
1029 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1030 atomic_read(&kgnilnd_data.kgn_nthreads))
1032 #define KGNILND_SPIN_QUIESCE \
1034 /* E.T phone home */ \
1035 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1036 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1037 while (kgnilnd_data.kgn_quiesce_trigger) { \
1038 set_current_state(TASK_INTERRUPTIBLE); \
1039 cfs_schedule_timeout_and_set_state(TASK_INTERRUPTIBLE, \
1040 cfs_time_seconds(1)); \
1042 /* Mom, my homework is done */ \
1043 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1044 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1047 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1048 #ifndef LIBCFS_DEBUG
1049 #error "this code uses actions inside LASSERT for ref counting"
1052 #define kgnilnd_admin_addref(atomic) \
1054 int val = atomic_inc_return(&atomic); \
1055 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1056 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1059 #define kgnilnd_admin_decref(atomic) \
1061 int val = atomic_dec_return(&atomic); \
1062 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
1063 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1066 #define kgnilnd_net_addref(net) \
1068 int val = atomic_inc_return(&net->gnn_refcount); \
1069 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1070 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1071 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1074 #define kgnilnd_net_decref(net) \
1076 int val = atomic_dec_return(&net->gnn_refcount); \
1077 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1078 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1079 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1082 #define kgnilnd_peer_addref(peer) \
1084 int val = atomic_inc_return(&peer->gnp_refcount); \
1085 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1086 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1087 libcfs_nid2str(peer->gnp_nid), val); \
1090 #define kgnilnd_peer_decref(peer) \
1092 int val = atomic_dec_return(&peer->gnp_refcount); \
1093 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1094 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1095 libcfs_nid2str(peer->gnp_nid), val); \
1097 kgnilnd_destroy_peer(peer); \
1100 #define kgnilnd_conn_addref(conn) \
1105 val = atomic_inc_return(&conn->gnc_refcount); \
1106 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1107 "conn %p refc %d to %s\n", \
1110 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1112 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1114 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1119 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1120 * through the scheduler thread to get the EP destroyed. This avoids some
1121 * messy semaphore business and allows us to reuse the connd_list and existing
1122 * linkage and avoid creating extra lists just for destroying EPs */
1124 /* Safety Disclaimer:
1125 * Q: If we decrement the refcount and then check it again, is it possible that
1126 * another caller could have passed through this macro concurrently? If so,
1127 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1129 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1130 * once we are refcount <= 2. It hinges around gnc_state and membership of
1131 * gnc_hashlist. There are two ways to find a connection - either ask for
1132 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1133 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1136 * - #1 from create (kgnilnd_create_conn)
1137 * - #2 for EP (kgnilnd_create_conn)
1138 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1139 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1141 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1142 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1144 * There could be more from TX descriptors during the lifetime of a live
1147 * If we nuke the conn before finish_connect, we won't have parallel paths
1148 * because nobody besides the dgram handler for the single outstanding
1149 * dgram can find the connection as it isn't in any searchable tables yet.
1151 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1152 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1153 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1154 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1155 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1156 * sent. Now the gnc_state == CLOSED, so we hit
1157 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1158 * - we've nuked them from the peer and CQ id tables, so we own them and
1159 * are guaranteed serial access - hence the complete lack of conn list
1160 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1161 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1162 * kgnilnd_process_conns. Then the next scheduler loop would call
1163 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1164 * kgnilnd_process_conns.
1166 * Clearly, we are totally safe. Clearly.
1169 #define kgnilnd_conn_decref(conn) \
1174 val = atomic_dec_return(&conn->gnc_refcount); \
1175 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1178 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1180 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1182 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1187 (conn->gnc_ephandle != NULL) && \
1188 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1189 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1190 kgnilnd_schedule_conn(conn); \
1191 } else if (val == 0) { \
1192 kgnilnd_destroy_conn(conn); \
1196 static inline struct list_head *
1197 kgnilnd_nid2peerlist(lnet_nid_t nid)
1199 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1201 RETURN(&kgnilnd_data.kgn_peers[hash]);
1204 static inline struct list_head *
1205 kgnilnd_netnum2netlist(__u16 netnum)
1207 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1209 RETURN(&kgnilnd_data.kgn_nets[hash]);
1213 kgnilnd_peer_active(kgn_peer_t *peer)
1215 /* Am I in the peer hash table? */
1216 return (!list_empty(&peer->gnp_list));
1219 /* need write_lock on kgn_peer_conn_lock */
1221 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1223 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1224 peer, libcfs_nid2str(peer->gnp_nid),
1225 !list_empty(&peer->gnp_conns),
1226 !list_empty(&peer->gnp_tx_queue));
1228 /* kgn_peer_conn_lock protects us from conflict with
1229 * kgnilnd_peer_notify and gnp_persistent */
1230 RETURN ((list_empty(&peer->gnp_conns)) &&
1231 (list_empty(&peer->gnp_tx_queue)));
1234 /* returns positive if error was for a clean shutdown of conn */
1236 kgnilnd_conn_clean_errno(int errno)
1238 /* - ESHUTDOWN - LND is unloading
1239 * - EUCLEAN - admin requested via "lctl del_peer"
1240 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1241 * - ENOTRECOVERABLE - stack reset
1242 * - EISCONN - cleared via "lctl push"
1243 * not doing ESTALE - that isn't clean */
1244 RETURN ((errno == 0) ||
1245 (errno == -ESHUTDOWN) ||
1246 (errno == -EUCLEAN) ||
1247 (errno == -ENETRESET) ||
1248 (errno == -EISCONN) ||
1249 (errno == -ENOTRECOVERABLE));
1252 /* returns positive if error results in purgatory hold */
1254 kgnilnd_check_purgatory_errno(int errno)
1256 /* We don't want to save the purgatory lists these cases:
1257 * - EUCLEAN - admin requested via "lctl del_peer"
1258 * - ESHUTDOWN - LND is unloading
1260 RETURN ((errno != -ESHUTDOWN) &&
1261 (errno != -EUCLEAN));
1265 /* returns positive if a purgatory hold is needed */
1267 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1271 if (conn->gnc_peer) {
1272 loopback = conn->gnc_peer->gnp_nid ==
1273 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1275 /* short circuit - a conn that didn't complete
1276 * setup never needs a purgatory hold */
1279 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1280 conn, conn->gnc_peer ?
1281 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1283 loopback, conn->gnc_close_recvd);
1285 /* we only use a purgatory hold if we've not received the CLOSE msg
1286 * from our peer - without that message, we can't know the state of
1287 * the other end of this connection and must put it into purgatory
1288 * to prevent reuse and corruption.
1289 * The theory is that a TX error can be communicated in all other cases
1291 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1292 kgnilnd_check_purgatory_errno(conn->gnc_error));
1295 static inline const char *
1296 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1298 static inline struct list_head *
1299 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1300 kgn_tx_list_state_t to_state)
1303 case GNILND_TX_PEERQ:
1304 return &peer->gnp_tx_queue;
1305 case GNILND_TX_FMAQ:
1306 return &conn->gnc_fmaq;
1307 case GNILND_TX_LIVE_FMAQ:
1308 case GNILND_TX_LIVE_RDMAQ:
1309 case GNILND_TX_DYING:
1311 case GNILND_TX_MAPQ:
1312 return &conn->gnc_device->gnd_map_tx;
1313 case GNILND_TX_RDMAQ:
1314 return &conn->gnc_device->gnd_rdmaq;
1316 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1317 CERROR("invalid state requested: %s\n",
1318 kgnilnd_tx_state2str(to_state));
1324 /* should hold tx, conn or peer lock when calling */
1326 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1327 kgn_conn_t *conn, kgn_tx_list_state_t state,
1330 struct list_head *list = NULL;
1332 /* make sure we have a sane TX state to start */
1333 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1334 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1335 list_empty(&tx->tx_list),
1336 "bad state with tx_list %s",
1337 list_empty(&tx->tx_list) ? "empty" : "not empty");
1339 /* WTF - you are already on that state buttmunch */
1340 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1341 "already at %s", kgnilnd_tx_state2str(state));
1343 /* get proper list from the state requested */
1344 list = kgnilnd_tx_state2list(peer, conn, state);
1348 case GNILND_TX_PEERQ:
1349 kgnilnd_peer_addref(peer);
1351 case GNILND_TX_ALLOCD:
1352 /* no refs needed */
1354 case GNILND_TX_FMAQ:
1355 kgnilnd_conn_addref(conn);
1357 case GNILND_TX_MAPQ:
1358 atomic_inc(&conn->gnc_device->gnd_nq_map);
1359 kgnilnd_conn_addref(conn);
1361 case GNILND_TX_LIVE_FMAQ:
1362 atomic_inc(&conn->gnc_nlive_fma);
1363 kgnilnd_conn_addref(conn);
1365 case GNILND_TX_LIVE_RDMAQ:
1366 atomic_inc(&conn->gnc_nlive_rdma);
1367 kgnilnd_conn_addref(conn);
1369 case GNILND_TX_RDMAQ:
1370 atomic_inc(&conn->gnc_nq_rdma);
1371 kgnilnd_conn_addref(conn);
1373 case GNILND_TX_DYING:
1374 kgnilnd_conn_addref(conn);
1377 CERROR("invalid state requested: %s\n",
1378 kgnilnd_tx_state2str(state));
1383 /* if this changes, change kgnilnd_alloc_tx */
1384 tx->tx_list_state = state;
1386 /* some states don't have lists - we track them in the per conn
1387 * TX table instead. Waste not, want not! */
1389 tx->tx_list_p = list;
1391 list_add_tail(&tx->tx_list, list);
1393 list_add(&tx->tx_list, list);
1395 /* set dummy list_p to make book keeping happy and let debugging
1396 * be a hair easier */
1397 tx->tx_list_p = (void *)state;
1400 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1401 kgnilnd_tx_state2str(state), list);
1405 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1406 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1408 /* These is only 1 "off-list" state */
1409 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1410 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1412 /* new_state == ALLOCD means we are deallocating this tx,
1413 * so make sure it was on a valid list to start with */
1414 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1415 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1416 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1417 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1420 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1423 switch (tx->tx_list_state) {
1424 case GNILND_TX_PEERQ:
1425 kgnilnd_peer_decref(peer);
1427 case GNILND_TX_FREED:
1428 case GNILND_TX_IDLE:
1429 case GNILND_TX_ALLOCD:
1430 /* no refs needed */
1432 case GNILND_TX_DYING:
1433 kgnilnd_conn_decref(conn);
1435 case GNILND_TX_FMAQ:
1436 kgnilnd_conn_decref(conn);
1438 case GNILND_TX_MAPQ:
1439 atomic_dec(&conn->gnc_device->gnd_nq_map);
1440 kgnilnd_conn_decref(conn);
1442 case GNILND_TX_LIVE_FMAQ:
1443 atomic_dec(&conn->gnc_nlive_fma);
1444 kgnilnd_conn_decref(conn);
1446 case GNILND_TX_LIVE_RDMAQ:
1447 atomic_dec(&conn->gnc_nlive_rdma);
1448 kgnilnd_conn_decref(conn);
1450 case GNILND_TX_RDMAQ:
1451 atomic_dec(&conn->gnc_nq_rdma);
1452 kgnilnd_conn_decref(conn);
1453 /* don't need to assert on default, already did in set */
1456 /* for ALLOCD, this might already be true, but no harm doing it again */
1457 list_del_init(&tx->tx_list);
1458 tx->tx_list_p = NULL;
1459 tx->tx_list_state = new_state;
1463 kgnilnd_tx_mapped(kgn_tx_t *tx)
1465 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1466 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1469 static inline struct list_head *
1470 kgnilnd_cqid2connlist(__u32 cqid)
1472 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1474 return (&kgnilnd_data.kgn_conns [hash]);
1477 static inline kgn_conn_t *
1478 kgnilnd_cqid2conn_locked(__u32 cqid)
1480 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1481 struct list_head *tmp;
1484 list_for_each(tmp, conns) {
1485 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1487 if (conn->gnc_cqid == cqid)
1494 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1496 kgnilnd_get_cqid_locked(void)
1502 cqid = kgnilnd_data.kgn_next_cqid++;
1503 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1507 kgnilnd_data.kgn_next_cqid = 1;
1510 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1516 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1518 kgn_tx_t *tx = NULL;
1519 kgn_conn_t *conn = NULL;
1521 /* set to NULL so any early return is an error */
1525 LASSERTF((ev_id->txe_idx > 0) &&
1526 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1527 "bogus txe_idx %d >= %d\n",
1528 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1530 LASSERTF((ev_id->txe_cqid > 0) &&
1531 (ev_id->txe_cqid < GNILND_MAX_CQID),
1532 "bogus txe_cqid %d >= %d\n",
1533 ev_id->txe_cqid, GNILND_MAX_CQID);
1535 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1536 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1539 /* Conn was destroyed? */
1540 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1541 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1544 /* just insurance */
1545 kgnilnd_conn_addref(conn);
1546 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1547 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1549 /* we know this is safe - as the TX won't be reused until AFTER
1550 * the conn is unlinked from the cqid hash, so we can use the TX
1551 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1553 spin_lock(&conn->gnc_tx_lock);
1554 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1555 spin_unlock(&conn->gnc_tx_lock);
1557 /* We could have a tx that was cleared out by other forces
1558 * lctl disconnect or del_peer. */
1560 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1561 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1562 kgnilnd_conn_decref(conn);
1566 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1567 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1568 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1570 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1571 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1572 "found tx %p txe_idx %d",
1573 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1574 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1576 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1578 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1583 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1586 /* set_normalized_timepsec isn't exported from the kernel, so
1587 * we need to do the same thing inline */
1588 static inline struct timespec
1589 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1595 sec = lhs.tv_sec - rhs.tv_sec;
1596 nsec = lhs.tv_nsec - rhs.tv_nsec;
1598 while (nsec >= NSEC_PER_SEC) {
1599 nsec -= NSEC_PER_SEC;
1603 nsec += NSEC_PER_SEC;
1612 kgnilnd_count_list(struct list_head *q)
1614 struct list_head *e;
1617 list_for_each(e, q) {
1624 /* kgnilnd_find_net adds a reference to the net it finds
1625 * this is so the net will not be removed before the calling function
1626 * has time to use the data returned. This reference needs to be released
1627 * by the calling function once it has finished using the returned net
1631 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1636 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1642 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1643 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1644 kgnilnd_net_addref(net);
1645 up_read(&kgnilnd_data.kgn_net_rw_sem);
1651 up_read(&kgnilnd_data.kgn_net_rw_sem);
1656 #ifdef CONFIG_DEBUG_SLAB
1657 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1659 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1662 int kgnilnd_dev_init(kgn_device_t *dev);
1663 void kgnilnd_dev_fini(kgn_device_t *dev);
1664 int kgnilnd_startup(lnet_ni_t *ni);
1665 void kgnilnd_shutdown(lnet_ni_t *ni);
1666 int kgnilnd_base_startup(void);
1667 void kgnilnd_base_shutdown(void);
1669 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1670 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1671 void kgnilnd_unmap_phys_fmablk(kgn_device_t *device);
1672 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1674 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1675 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1676 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1677 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1678 lnet_msg_t *lntmsg, void **new_private);
1679 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1680 int delayed, unsigned int niov,
1681 struct iovec *iov, lnet_kiov_t *kiov,
1682 unsigned int offset, unsigned int mlen, unsigned int rlen);
1684 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1686 /* purgatory functions */
1687 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1688 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1689 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1690 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1692 void kgnilnd_update_reaper_timeout(long timeout);
1693 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1694 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1695 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1696 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1697 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1698 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
1699 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1701 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1702 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net);
1703 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1704 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1706 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1707 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1708 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1709 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1710 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1711 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1712 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1713 void kgnilnd_consume_rx(kgn_rx_t *rx);
1715 void kgnilnd_schedule_device(kgn_device_t *dev);
1716 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1717 void kgnilnd_schedule_device_timer(unsigned long arg);
1719 int kgnilnd_reaper(void *arg);
1720 int kgnilnd_scheduler(void *arg);
1721 int kgnilnd_dgram_mover(void *arg);
1722 int kgnilnd_rca(void *arg);
1724 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1725 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1726 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1727 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1728 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1729 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1730 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1731 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1732 void kgnilnd_peer_alive(kgn_peer_t *peer);
1733 void kgnilnd_peer_notify(kgn_peer_t *peer, int error);
1734 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1735 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1736 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1737 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1739 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1740 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1741 void kgnilnd_wakeup_rca_thread(void);
1742 int kgnilnd_start_rca_thread(void);
1744 int kgnilnd_tunables_init(void);
1745 void kgnilnd_tunables_fini(void);
1746 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1748 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1749 void kgnilnd_pause_threads(void);
1750 int kgnilnd_hw_in_quiesce(void);
1751 int kgnilnd_check_hw_quiesce(void);
1752 void kgnilnd_quiesce_wait(char *reason);
1753 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1754 int kgnilnd_ruhroh_thread(void *arg);
1755 void kgnilnd_reset_stack(void);
1756 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1758 void kgnilnd_insert_sysctl(void);
1759 void kgnilnd_remove_sysctl(void);
1760 void kgnilnd_proc_init(void);
1761 void kgnilnd_proc_fini(void);
1764 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1766 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1767 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1768 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram);
1770 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1771 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1772 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1773 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1775 int kgnilnd_dgram_waitq(void *arg);
1777 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1779 /* struct2str functions - we don't use a default: case to cause the compile
1780 * to fail if there is a missing case. This allows us to hide these down here
1781 * out of the way but ensure we'll catch any updates to the enum/types
1784 #define DO_TYPE(x) case x: return #x;
1785 static inline const char *
1786 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1788 /* Only want single char string for this */
1790 case GNILND_FMABLK_IDLE:
1792 case GNILND_FMABLK_PHYS:
1794 case GNILND_FMABLK_VIRT:
1796 case GNILND_FMABLK_FREED:
1799 return "<unknown state>";
1802 static inline const char *
1803 kgnilnd_msgtype2str(int type)
1806 DO_TYPE(GNILND_MSG_NONE);
1807 DO_TYPE(GNILND_MSG_NOOP);
1808 DO_TYPE(GNILND_MSG_IMMEDIATE);
1809 DO_TYPE(GNILND_MSG_PUT_REQ);
1810 DO_TYPE(GNILND_MSG_PUT_NAK);
1811 DO_TYPE(GNILND_MSG_PUT_ACK);
1812 DO_TYPE(GNILND_MSG_PUT_DONE);
1813 DO_TYPE(GNILND_MSG_GET_REQ);
1814 DO_TYPE(GNILND_MSG_GET_NAK);
1815 DO_TYPE(GNILND_MSG_GET_DONE);
1816 DO_TYPE(GNILND_MSG_CLOSE);
1817 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1818 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1819 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1820 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1821 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1822 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1823 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1825 return "<unknown msg type>";
1828 static inline const char *
1829 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1832 DO_TYPE(GNILND_TX_IDLE);
1833 DO_TYPE(GNILND_TX_ALLOCD);
1834 DO_TYPE(GNILND_TX_PEERQ);
1835 DO_TYPE(GNILND_TX_MAPQ);
1836 DO_TYPE(GNILND_TX_FMAQ);
1837 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1838 DO_TYPE(GNILND_TX_RDMAQ);
1839 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1840 DO_TYPE(GNILND_TX_DYING);
1841 DO_TYPE(GNILND_TX_FREED);
1843 return "<unknown state>";
1846 static inline const char *
1847 kgnilnd_conn_state2str(kgn_conn_t *conn)
1849 kgn_conn_state_t state = conn->gnc_state;
1851 DO_TYPE(GNILND_CONN_DUMMY);
1852 DO_TYPE(GNILND_CONN_LISTEN);
1853 DO_TYPE(GNILND_CONN_CONNECTING);
1854 DO_TYPE(GNILND_CONN_ESTABLISHED);
1855 DO_TYPE(GNILND_CONN_CLOSING);
1856 DO_TYPE(GNILND_CONN_CLOSED);
1857 DO_TYPE(GNILND_CONN_DONE);
1858 DO_TYPE(GNILND_CONN_DESTROY_EP);
1863 static inline const char *
1864 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1866 kgn_connreq_type_t type = connreq->gncr_type;
1869 DO_TYPE(GNILND_CONNREQ_REQ);
1870 DO_TYPE(GNILND_CONNREQ_NAK);
1871 DO_TYPE(GNILND_CONNREQ_CLOSE);
1876 static inline const char *
1877 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1879 kgn_dgram_state_t state = dgram->gndg_state;
1882 DO_TYPE(GNILND_DGRAM_USED);
1883 DO_TYPE(GNILND_DGRAM_POSTING);
1884 DO_TYPE(GNILND_DGRAM_POSTED);
1885 DO_TYPE(GNILND_DGRAM_PROCESSING);
1886 DO_TYPE(GNILND_DGRAM_DONE);
1887 DO_TYPE(GNILND_DGRAM_CANCELED);
1892 static inline const char *
1893 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1895 kgn_dgram_type_t type = dgram->gndg_type;
1898 DO_TYPE(GNILND_DGRAM_REQ);
1899 DO_TYPE(GNILND_DGRAM_WC_REQ);
1900 DO_TYPE(GNILND_DGRAM_NAK);
1901 DO_TYPE(GNILND_DGRAM_CLOSE);
1906 static inline const char *
1907 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
1910 DO_TYPE(GNILND_DGRAM_REQ);
1911 DO_TYPE(GNILND_DGRAM_WC_REQ);
1912 DO_TYPE(GNILND_DGRAM_NAK);
1913 DO_TYPE(GNILND_DGRAM_CLOSE);
1920 /* API wrapper functions - include late to pick up all of the other defines */
1921 #include "gnilnd_api_wrap.h"
1923 /* pulls in tunables per platform and adds in nid/nic conversion
1924 * if RCA wasn't available at build time */
1925 #include "gnilnd_hss_ops.h"
1927 #if defined(CONFIG_CRAY_GEMINI)
1928 #include "gnilnd_gemini.h"
1929 #elif defined(CONFIG_CRAY_ARIES)
1930 #include "gnilnd_aries.h"
1932 #error "Undefined Network Hardware Type"
1935 #endif /* _GNILND_GNILND_H_ */