2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Copyright (c) 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 #ifdef HAVE_COMPAT_RDMA
32 #include <linux/compat-2.6.h>
34 #include <linux/module.h>
35 #include <linux/kernel.h>
37 #include <linux/string.h>
38 #include <linux/stat.h>
39 #include <linux/errno.h>
40 #ifdef HAVE_LINUX_KERNEL_LOCK
41 #include <linux/smp_lock.h>
43 #include <linux/unistd.h>
44 #include <linux/uio.h>
45 #include <linux/time.h>
46 #include <asm/timex.h>
48 #include <asm/uaccess.h>
51 #include <linux/init.h>
53 #include <linux/file.h>
54 #include <linux/stat.h>
55 #include <linux/list.h>
56 #include <linux/kmod.h>
57 #include <linux/sysctl.h>
58 #include <linux/kthread.h>
59 #include <linux/nmi.h>
63 #include <linux/nmi.h>
65 #define DEBUG_SUBSYSTEM S_LND
67 #include <libcfs/libcfs.h>
68 #include <lnet/lnet.h>
69 #include <lnet/lib-lnet.h>
72 #include "gnilnd_version.h"
75 /* tunables determined at compile time */
76 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
77 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
78 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
79 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
80 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
81 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
82 #define GNILND_SCHED_TIMEOUT 1
83 #define GNILND_DGRAM_TIMEOUT 2
84 #define GNILND_FAST_MAPPING_TRY \
85 *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
86 #define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
88 /* map failure timeout */
89 #define GNILND_MAP_TIMEOUT \
90 (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
91 *kgnilnd_tunables.kgn_timeout))
93 /* reaper thread wakup interval */
94 #define GNILND_REAPER_THREAD_WAKE 1
95 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
96 #define GNILND_REAPER_NCHECKS 4
99 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
100 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
101 #define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
102 /* checksum values */
103 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
104 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
105 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
106 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
108 /* tune down some COMPUTE options as they won't see the same number of connections and
109 * don't need the throughput of multiple threads by default */
110 #if defined(CONFIG_CRAY_COMPUTE)
112 #define GNILND_SCHED_THREADS 2 /* default # of kgnilnd_scheduler threads */
114 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
116 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
117 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
118 #define GNILND_COMPUTE 1 /* compute image */
119 #define GNILND_FAST_RECONNECT 1 /* Fast Reconnect option */
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
141 /* Closing, don't put in purgatory */
142 #define GNILND_NOPURG 222
144 /* payload size to add to the base mailbox size
145 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
146 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
147 * the calculation return from that function.*/
148 #define GNILND_MBOX_PAYLOAD \
149 (GNILND_MAX_MSG_SIZE * \
150 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
152 /* timeout -> deadman timer for kgni mdd holds */
153 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
155 /* timeout for failing sends in t is in jiffies*/
156 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
158 /* time when to release from purgatory in the reaper thread in jiffies */
159 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
161 /* Macro for finding last_rx 2 datapoints are compared
162 * and the most recent one in jiffies is returned.
164 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
165 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
167 /* fmablk registration failures timeout before failing node */
168 #define GNILND_REGFAILTO_DISABLE -1
170 /************************************************************************
171 * Enum, flag and tag data
173 #define GNILND_INIT_NOTHING 0
174 #define GNILND_INIT_DATA 1
175 #define GNILND_INIT_ALL 2
177 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
178 #define GNILND_BUF_NONE 0 /* buffer type not set */
179 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
180 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
181 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
182 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
183 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
184 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
186 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
187 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
188 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
189 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
190 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
192 /* stash above max CQID to avoid any collision */
193 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
194 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
196 /* kgn_msg_t::gnm_type */
197 #define GNILND_MSG_NONE 0x00 /* illegal message */
198 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
199 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
200 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
201 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
202 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
203 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
204 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
205 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
206 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
207 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
208 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
209 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
210 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
211 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
212 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
213 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
214 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
216 /* defines for gnc_*scheduled states */
217 #define GNILND_CONN_IDLE 0
218 #define GNILND_CONN_SCHED 1
219 #define GNILND_CONN_WANTS_SCHED 2
220 #define GNILND_CONN_PROCESS 3
222 #define GNILND_DEV_IDLE 0
223 #define GNILND_DEV_IRQ 1
224 #define GNILND_DEV_LOOP 2
226 #define GNILND_DGRAM_IDLE 0
227 #define GNILND_DGRAM_SCHED 1
228 #define GNILND_DGRAM_PROCESS 2
230 #define GNILND_PEER_IDLE 0
231 #define GNILND_PEER_CONNECT 1
232 #define GNILND_PEER_POSTING 2
233 #define GNILND_PEER_POSTED 3
234 #define GNILND_PEER_NEEDS_DEATH 4
235 #define GNILND_PEER_KILL 5
237 /* for gnc_close_recvd */
238 #define GNILND_CLOSE_RX 1
239 #define GNILND_CLOSE_INJECT1 2
240 #define GNILND_CLOSE_INJECT2 3
241 #define GNILND_CLOSE_EARLY 4
243 /* defines for why quiesce trigger set */
244 #define GNILND_QUIESCE_IDLE 0
245 #define GNILND_QUIESCE_ADMIN 1
246 #define GNILND_QUIESCE_RESET 2
247 #define GNILND_QUIESCE_HW_QUIESCE 3
249 #define GNILND_PEER_CLEAN 0
250 #define GNILND_PEER_PERSISTING 1
252 #define GNILND_DEL_CONN 0
253 #define GNILND_DEL_PEER 1
254 #define GNILND_CLEAR_PURGATORY 2
256 #define GNILND_RCA_NODE_UP 0
257 #define GNILND_RCA_NODE_DOWN 1
258 #define GNILND_RCA_NODE_UNKNOWN 2
260 /* defines for reverse RDMA states */
261 #define GNILND_REVERSE_NONE 0
262 #define GNILND_REVERSE_GET 1
263 #define GNILND_REVERSE_PUT 2
264 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
266 typedef enum kgn_fmablk_state {
267 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
268 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
269 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
270 GNILND_FMABLK_FREED, /* after free */
271 } kgn_fmablk_state_t;
273 typedef enum kgn_tx_list_state {
274 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
275 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
276 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
277 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
278 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
279 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
280 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
281 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
282 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
283 GNILND_TX_FREED /* TX is free! */
284 } kgn_tx_list_state_t;
286 typedef enum kgn_conn_state {
287 /* don't start @ 0 - prevent memset(0) badness */
288 GNILND_CONN_DUMMY = 0,
290 GNILND_CONN_CONNECTING,
291 GNILND_CONN_ESTABLISHED,
295 GNILND_CONN_DESTROY_EP
298 /* changing these requires a change to GNILND_CONNREQ_VERSION and
299 * will result in dropped packets instead of NAKs. Adding to this is
300 * acceptable without changing the CONNREQ_VERSION, but code should
301 * be ready to handle NAKs on version mismatch */
302 typedef enum kgn_connreq_type {
303 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
304 GNILND_CONNREQ_NAK, /* NO soup for you! */
305 GNILND_CONNREQ_CLOSE, /* we should see other people */
306 } kgn_connreq_type_t;
308 typedef enum kgn_dgram_state {
309 /* don't use 0 to avoid thinking a memset of zero is valid data */
310 GNILND_DGRAM_USED = 1,
311 GNILND_DGRAM_POSTING,
313 GNILND_DGRAM_PROCESSING,
314 GNILND_DGRAM_CANCELED,
318 typedef enum kgn_dgram_type {
319 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
320 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
321 GNILND_DGRAM_NAK, /* NO soup for you! */
322 GNILND_DGRAM_CLOSE, /* we should see other people */
325 /************************************************************************
326 * Wire message structs. These are sent in sender's byte order
327 * (i.e. receiver checks magic and flips if required).
330 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
331 #define GNILND_DGRAM_MAGIC 0x0DDBA11
333 /* kgn_msg_t - FMA/SMSG wire struct
335 * - added checksum to FMA
336 * moved seq before paylod
337 * WIRE_ATTR added for alignment
339 * added gnm_payload_len for FMA payload size
341 * added gncm_retval to completion, allowing return code transmission
344 * changed how CQID and TX ids are assigned
346 * added retval on CLOSE
348 * added payload checksumming
350 * reworked checksumming a bit, changed payload checksums
352 #define GNILND_MSG_VERSION 8
353 /* kgn_connreq_t connection request datagram wire struct
358 #define GNILND_CONNREQ_VERSION 2
360 typedef struct kgn_gniparams {
361 __u32 gnpr_host_id; /* ph. host ID of the NIC */
362 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
363 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
364 } WIRE_ATTR kgn_gniparams_t;
366 typedef struct kgn_nak_data {
367 __s32 gnnd_errno; /* errno reason for NAK */
369 } WIRE_ATTR kgn_nak_data_t;
371 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
372 * without breaking the ability for us to properly NAK someone */
373 typedef struct kgn_connreq { /* connection request/response */
374 __u32 gncr_magic; /* I'm an gnilnd connreq */
375 __u32 gncr_cksum; /* checksum (0 == disabled) */
376 __u16 gncr_type; /* REQ, NAK, etc */
377 __u16 gncr_version; /* this is my version number */
378 __u32 gncr_timeout; /* sender's timeout */
379 __u64 gncr_srcnid; /* sender's NID */
380 __u64 gncr_dstnid; /* who sender expects to listen */
381 __u64 gncr_peerstamp; /* sender's instance stamp */
382 __u64 gncr_connstamp; /* sender's connection stamp */
384 /* everything before this needs to stay static, adding after should
385 * result in a change to GNILND_CONNREQ_VERSION */
388 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
389 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
391 } WIRE_ATTR kgn_connreq_t;
394 gni_mem_handle_t gnrd_key;
397 } WIRE_ATTR kgn_rdma_desc_t;
400 lnet_hdr_t gnim_hdr; /* LNet header */
401 /* LNet payload is in FMA "Message Data" */
402 } WIRE_ATTR kgn_immediate_msg_t;
405 lnet_hdr_t gnprm_hdr; /* LNet header */
406 __u64 gnprm_cookie; /* opaque completion cookie */
407 } WIRE_ATTR kgn_putreq_msg_t;
410 __u64 gnpam_src_cookie; /* reflected completion cookie */
411 __u64 gnpam_dst_cookie; /* opaque completion cookie */
412 __u16 gnpam_payload_cksum; /* checksum for get msg */
413 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
414 } WIRE_ATTR kgn_putack_msg_t;
417 lnet_hdr_t gngm_hdr; /* LNet header */
418 __u64 gngm_cookie; /* opaque completion cookie */
419 __u16 gngm_payload_cksum; /* checksum for put msg */
420 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
421 } WIRE_ATTR kgn_get_msg_t;
424 int gncm_retval; /* error on NAK, size on REQ */
425 __u64 gncm_cookie; /* reflected completion cookie */
426 } WIRE_ATTR kgn_completion_msg_t;
428 typedef struct { /* NB must fit in FMA "Prefix" */
429 __u32 gnm_magic; /* I'm an gni message */
430 __u16 gnm_version; /* this is my version number */
431 __u16 gnm_type; /* msg type */
432 __u64 gnm_srcnid; /* sender's NID */
433 __u64 gnm_connstamp; /* sender's connection stamp */
434 __u32 gnm_seq; /* incrementing sequence number */
435 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
436 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
437 __u32 gnm_payload_len; /* size of the FMA payload sent */
439 kgn_immediate_msg_t immediate;
440 kgn_putreq_msg_t putreq;
441 kgn_putack_msg_t putack;
443 kgn_completion_msg_t completion;
445 } WIRE_ATTR kgn_msg_t;
447 /************************************************************************
448 * runtime tunable data
451 typedef struct kgn_tunables {
452 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
453 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
454 int *kgn_credits; /* # concurrent sends */
455 int *kgn_fma_cq_size; /* # entries in receive CQ */
456 int *kgn_peer_credits; /* # LNet peer credits */
457 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
458 int *kgn_timeout; /* comms timeout (seconds) */
459 int *kgn_max_immediate; /* immediate payload breakpoint */
460 int *kgn_checksum; /* checksum data */
461 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
462 int *kgn_bte_put_dlvr_mode; /* BTE Put delivery mode */
463 int *kgn_bte_get_dlvr_mode; /* BTE Get delivery mode */
464 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
465 int *kgn_ptag; /* PTAG for cdm_create */
466 int *kgn_pkey; /* PKEY for cdm_create */
467 int *kgn_max_retransmits; /* max number of FMA retransmits */
468 int *kgn_nwildcard; /* # wildcard per net to post */
469 int *kgn_nice; /* nice value for kgnilnd threads */
470 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
471 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
472 int *kgn_peer_hash_size; /* size of kgn_peers */
473 int *kgn_peer_health; /* enable/disable peer health */
474 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
475 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
476 int *kgn_mbox_per_block; /* mailboxes per fmablk */
477 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
478 int *kgn_mbox_credits; /* max credits per fma */
479 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
480 int *kgn_net_hash_size; /* size of kgn_net_ht */
481 int *kgn_hardware_timeout; /* max time for a message to get across the network */
482 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
483 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
484 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
485 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
486 int *kgn_reverse_rdma; /* Reverse RDMA setting */
487 int *kgn_eager_credits; /* allocated eager buffers */
488 int *kgn_fast_reconn; /* fast reconnection on conn timeout */
489 int *kgn_efault_lbug; /* LBUG on receiving an EFAULT */
490 int *kgn_max_purgatory; /* # conns/peer to keep in purgatory */
491 int *kgn_reg_fail_timeout; /* registration failure timeout */
492 int *kgn_thread_affinity; /* bind scheduler threads to cpus */
493 int *kgn_thread_safe; /* use thread safe kgni API */
496 typedef struct kgn_mbox_info {
497 lnet_nid_t mbx_prev_nid;
498 lnet_nid_t mbx_prev_purg_nid;
499 unsigned long mbx_create_conn_memset;
500 unsigned long mbx_add_purgatory;
501 unsigned long mbx_detach_of_purgatory;
502 unsigned long mbx_release_from_purgatory;
503 unsigned long mbx_release_purg_active_dgram;
505 int mbx_nallocs_total;
508 typedef struct kgn_fma_memblock {
509 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
510 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
511 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
512 int gnm_num_mboxs; /* total mboxes allocated */
513 int gnm_avail_mboxs; /* number of available mailboxes in the block */
514 int gnm_held_mboxs; /* number of purgatory held mailboxes */
515 int gnm_mbox_size; /* size of the single mailbox */
516 int gnm_next_avail_mbox; /* next available mailbox */
517 long gnm_max_timeout; /* max timeout for possible purgatory hold */
518 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
519 void *gnm_block; /* pointer to mem. block */
520 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
521 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
522 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
523 } kgn_fma_memblock_t;
525 typedef struct kgn_device {
526 gni_nic_handle_t gnd_handle; /* device handle */
527 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
528 gni_err_handle_t gnd_err_handle; /* device error handle */
529 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
530 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
531 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
532 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
533 struct mutex gnd_cq_mutex; /* CQ access serialization */
534 __u32 gnd_host_id; /* ph. host ID of the NIC */
535 int gnd_id; /* device id, also index in kgn_devices */
536 __u32 gnd_nid; /* ph host ID translated to NID */
537 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
538 struct mutex gnd_fmablk_mutex; /* mutex for FMA block memory alloc/free */
539 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
540 atomic_t gnd_nfmablk; /* # of fmablk live */
541 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
542 atomic_t gnd_neps; /* # EP allocated to conns */
543 short gnd_ready; /* stuff to do in scheduler thread */
544 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
545 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
546 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
547 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
548 struct list_head gnd_connd_peers; /* peers waiting for a connection */
549 spinlock_t gnd_connd_lock; /* serialise connd_peers */
550 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
551 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
552 int gnd_dgram_ready; /* dgrams need movin' */
553 struct list_head *gnd_dgrams; /* nid hash to dgrams */
554 atomic_t gnd_ndgrams; /* # dgrams extant */
555 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post*/
556 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
557 struct list_head gnd_map_list; /* list of all mapped regions */
558 int gnd_map_version; /* version flag for map list */
559 struct timer_list gnd_map_timer; /* wakey-wakey */
560 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
561 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
562 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
563 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
564 __u32 gnd_map_nphys; /* # TX phys mappings */
565 __u32 gnd_map_physnop; /* # TX phys pages mapped */
566 __u32 gnd_map_nvirt; /* # TX virt mappings */
567 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
568 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
569 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
570 int gnd_map_attempt; /* last map attempt # */
571 unsigned long gnd_last_map; /* map timeout base */
572 struct list_head gnd_rdmaq; /* RDMA to be sent */
573 spinlock_t gnd_rdmaq_lock; /* play nice with others */
574 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
575 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
576 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
577 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
578 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
579 atomic_t gnd_short_ntx; /* TX stats: short messages */
580 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
581 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
582 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
583 atomic_t gnd_short_nrx; /* RX stats: short messages */
584 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
585 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
586 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
587 atomic_t gnd_fast_try; /* # of times fast send tried */
588 atomic_t gnd_fast_ok; /* # of times fast send ok */
589 atomic_t gnd_fast_block; /* # of times fast send blocked */
590 unsigned long gnd_mutex_delay;
591 atomic_t gnd_n_yield;
592 atomic_t gnd_n_schedule;
593 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
594 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
595 void *gnd_smdd_hold_buf; /* buffer to keep smdd */
596 gni_mem_handle_t gnd_smdd_hold_hndl; /* buffer mem handle */
599 typedef struct kgn_net {
600 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
601 kgn_device_t *gnn_dev; /* device for this net */
602 lnet_ni_t *gnn_ni; /* network interface instance */
603 atomic_t gnn_refcount; /* # current references */
604 int gnn_shutdown; /* lnd_shutdown set */
605 __u16 gnn_netnum; /* stash netnum for quicker lookup */
608 static inline lnet_nid_t
609 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
611 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
614 static inline lnet_nid_t
615 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
617 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
620 /* The code for this is a bit ugly - but really this just boils down to a __u64
621 * that can have various parts accessed separately.
623 * The lower 32 bits is the ID
624 * we give to SMSG for our completion event - it needs to be globally unique across
625 * all TX currently in flight. We separate that out into the CQID so that we can
626 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
627 * the actual TX out of the per-connection gnc_tx_ref_table.
629 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
630 * has a unique value we can send with RDMA setup messages to ensure the completion for
631 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
632 * reuse is separated.
635 typedef struct kgn_tx_ev_id {
637 __u64 txe_cookie; /* are you my mommy ? */
639 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
641 __u32 txe_smsg_id; /* ID for SMSG CQ event */
642 /* N.B: Never ever ever ever use the bit shifts directly,
643 * you are just asking for a world of pain and are at the
644 * mercy of the compiler layouts */
646 __u32 txe_cqid :GNILND_CQID_NBITS;
647 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
654 typedef struct kgn_dgram {
655 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
656 kgn_dgram_state_t gndg_state; /* state of this dgram */
657 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
658 __u32 gndg_magic; /* saftey word */
659 unsigned long gndg_post_time; /* time when we posted */
660 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
661 kgn_connreq_t gndg_conn_out; /* connreq from local node */
662 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
665 typedef struct kgn_tx { /* message descriptor */
666 struct list_head tx_list; /* TX queues - peer, conn, rdma */
667 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
668 struct list_head *tx_list_p; /* pointer to current list */
669 struct kgn_conn *tx_conn; /* owning conn */
670 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
671 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
672 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
673 struct list_head tx_map_list; /* list entry on device map list */
674 unsigned int tx_nob; /* # bytes of payload */
675 int tx_buftype; /* payload buffer type */
676 int tx_phys_npages; /* # physical pages */
677 gni_mem_handle_t tx_map_key; /* mapping key */
678 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
679 gni_mem_segment_t *tx_phys; /* page descriptors */
680 kgn_msg_t tx_msg; /* FMA message buffer */
681 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
682 __u8 tx_state; /* state of the descriptor */
683 int tx_retrans; /* retrans count of RDMA */
684 int tx_rc; /* if we need to stash the ret code until we see completion */
685 void *tx_buffer; /* source/sink buffer */
686 void *tx_buffer_copy; /* pointer to page aligned buffer */
687 unsigned int tx_nob_rdma; /* nob actually rdma */
688 unsigned int tx_offset; /* offset of data into copied buffer */
690 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
691 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
694 /* we only use one or the other */
696 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
697 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
701 typedef struct kgn_conn {
702 kgn_device_t *gnc_device; /* which device */
703 struct kgn_peer *gnc_peer; /* owning peer */
704 int gnc_magic; /* magic value cleared before free */
705 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
706 struct list_head gnc_hashlist; /* stash in connection hash table */
707 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
708 struct list_head gnc_fmaq; /* txs queued for FMA */
709 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
710 __u64 gnc_peerstamp; /* peer's unique stamp */
711 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
712 __u64 gnc_my_connstamp; /* my unique connection stamp */
713 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
714 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
715 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
716 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
717 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
718 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
719 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
720 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
721 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
722 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
723 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
724 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
725 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
726 __u32 gnc_cqid; /* my completion callback id (non-unique) */
727 atomic_t gnc_tx_seq; /* tx msg sequence number */
728 atomic_t gnc_rx_seq; /* rx msg sequence number */
729 struct mutex gnc_smsg_mutex; /* tx smsg sequence serialization */
730 struct mutex gnc_rdma_mutex; /* tx rdma sequence serialization */
731 __u64 gnc_tx_retrans; /* # retrans on SMSG */
732 atomic_t gnc_nlive_fma; /* # live FMA */
733 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
734 atomic_t gnc_nlive_rdma; /* # live RDMA */
735 short gnc_close_sent; /* I've sent CLOSE */
736 short gnc_close_recvd; /* I've received CLOSE */
737 short gnc_in_purgatory; /* in the sin bin */
738 int gnc_error; /* errno when conn being closed due to error */
739 int gnc_peer_error; /* errno peer sent us on CLOSE */
740 kgn_conn_state_t gnc_state; /* connection state */
741 int gnc_scheduled; /* being attented to */
742 char gnc_sched_caller[30]; /* what function last called schedule */
743 int gnc_sched_line; /* what line # last called schedule */
744 atomic_t gnc_refcount; /* # users */
745 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
746 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
747 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
748 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
749 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
750 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
751 int gnc_next_tx; /* next tx to use in tx_ref_table */
752 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
753 int gnc_mbox_id; /* id of mbox in fma_blk */
754 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
755 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
756 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
757 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
758 void *remote_mbox_addr; /* save remote mbox address */
761 typedef struct kgn_mdd_purgatory {
762 gni_mem_handle_t gmp_map_key; /* mapping key */
763 struct list_head gmp_list; /* entry point for purgatory list */
764 } kgn_mdd_purgatory_t;
766 typedef struct kgn_peer {
767 struct list_head gnp_list; /* stash on global peer list */
768 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
769 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
770 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
771 kgn_net_t *gnp_net; /* net instance for this peer */
772 lnet_nid_t gnp_nid; /* who's on the other end(s) */
773 atomic_t gnp_refcount; /* # users */
774 __u32 gnp_host_id; /* ph. host ID of the peer */
775 short gnp_connecting; /* connection forming */
776 short gnp_pending_unlink; /* need last conn close to trigger unlink */
777 int gnp_last_errno; /* last error conn saw */
778 unsigned long gnp_last_alive; /* last time I had valid comms */
779 int gnp_last_dgram_errno; /* last error dgrams saw */
780 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
781 unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
782 unsigned long gnp_reconnect_interval; /* exponential backoff */
783 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
784 int gnp_down; /* rca says peer down */
785 unsigned long gnp_down_event_time; /* time peer down */
786 unsigned long gnp_up_event_time; /* time peer back up */
789 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
790 * like lnet_parse. It allows a single pointer to let us get enough
791 * information in _recv and friends */
792 typedef struct kgn_rx {
793 kgn_conn_t *grx_conn; /* connection */
794 kgn_msg_t *grx_msg; /* message */
795 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
796 int grx_eager; /* if eager, we copied msg to somewhere */
797 struct timespec grx_received; /* time this msg received */
800 typedef struct kgn_data {
801 int kgn_init; /* initialisation state */
802 int kgn_shutdown; /* shut down? */
803 int kgn_wc_kill; /* Should I repost the WC */
804 atomic_t kgn_nthreads; /* # live threads */
805 int kgn_nresets; /* number of stack resets */
806 int kgn_in_reset; /* are we in stack reset ? */
808 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
810 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
811 int kgn_ndevs; /* # devices */
813 int kgn_ruhroh_running; /* ruhroh thread is running */
814 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
815 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
816 int kgn_quiesce_trigger; /* should we quiesce ? */
817 atomic_t kgn_nquiesce; /* how many quiesced ? */
818 struct mutex kgn_quiesce_mutex; /* serialize ruhroh task, startup and shutdown */
819 int kgn_needs_reset; /* we need stack reset */
821 /* These next three members implement communication from gnilnd into
822 * the ruhroh task. To ensure correct operation of the task, code that
823 * writes into them must use memory barriers to ensure that the changes
824 * are visible to other cores in the order the members appear below. */
825 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
826 int kgn_bump_info_rdy; /* we have info needed to bump */
827 int kgn_needs_pause; /* we need to pause for network quiesce */
829 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
830 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
832 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
833 struct list_head *kgn_peers; /* hash table of all my known peers */
834 atomic_t kgn_npeers; /* # peers extant */
835 int kgn_peer_version; /* version flag for peer tables */
837 struct list_head *kgn_conns; /* conns hashed by cqid */
838 atomic_t kgn_nconns; /* # connections extant */
839 atomic_t kgn_neager_allocs; /* # of eager allocations */
840 __u64 kgn_peerstamp; /* when I started up */
841 __u64 kgn_connstamp; /* conn stamp generator */
842 int kgn_conn_version; /* version flag for conn tables */
843 int kgn_next_cqid; /* cqid generator */
845 long kgn_new_min_timeout; /* minimum timeout on any new conn */
846 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
847 spinlock_t kgn_reaper_lock; /* serialise */
849 struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
850 struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
851 struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
852 atomic_t kgn_ntx; /* # tx in use */
853 struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
855 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
856 __u64 kgn_cksum_npages; /* # pages alloc'd for checksumming */
857 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
858 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
860 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
861 long kgn_rdmaq_override; /* bytes per second override */
863 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
865 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
866 atomic_t kgn_npending_conns; /* # of conns with pending closes */
867 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
868 unsigned long kgn_last_scheduled; /* last time schedule was called */
869 unsigned long kgn_last_condresched; /* last time cond_resched was called */
870 atomic_t kgn_rev_offset; /* # of REV rdma w/misaligned offsets */
871 atomic_t kgn_rev_length; /* # of REV rdma have misaligned len */
872 atomic_t kgn_rev_copy_buff; /* # of REV rdma buffer copies */
873 struct socket *kgn_sock; /* for Apollo */
874 unsigned long free_pages_limit; /* # of free pages reserve from fma block allocations */
875 int kgn_enable_gl_mutex; /* kgni api mtx enable */
878 extern kgn_data_t kgnilnd_data;
879 extern kgn_tunables_t kgnilnd_tunables;
881 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
882 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
883 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
885 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
886 * and the line of the calling function to allow us to debug problematic
887 * schedule calls in the future without the programmer having to mark
888 * the location manually.
890 #define kgnilnd_schedule_conn(conn) \
891 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
893 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
894 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
897 kgnilnd_thread_fini(void)
899 atomic_dec(&kgnilnd_data.kgn_nthreads);
902 static inline void kgnilnd_gl_mutex_lock(struct mutex *lock)
904 if (kgnilnd_data.kgn_enable_gl_mutex)
908 static inline void kgnilnd_gl_mutex_unlock(struct mutex *lock)
910 if (kgnilnd_data.kgn_enable_gl_mutex)
914 static inline void kgnilnd_conn_mutex_lock(struct mutex *lock)
916 if (!kgnilnd_data.kgn_enable_gl_mutex)
920 static inline void kgnilnd_conn_mutex_unlock(struct mutex *lock)
922 if (!kgnilnd_data.kgn_enable_gl_mutex)
926 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
927 * parts of the code to avoid a scheduler trip when the mutex is held
929 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
930 * has been acquired successfully, and 0 on contention.
932 * NOTE: this function follows the spin_trylock() convention, so
933 * it is negated to the down_trylock() return values! Be careful
934 * about this when converting semaphore users to mutexes.
936 * This function must not be used in interrupt context. The
937 * mutex must be released by the same task that acquired it.
939 static inline int __kgnilnd_mutex_trylock(struct mutex *lock)
942 unsigned long timeout;
944 LASSERT(!in_interrupt());
946 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
948 ret = mutex_trylock(lock);
955 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
957 if (!kgnilnd_data.kgn_enable_gl_mutex)
960 return __kgnilnd_mutex_trylock(lock);
963 static inline int kgnilnd_trylock(struct mutex *cq_lock,
964 struct mutex *c_lock)
966 if (kgnilnd_data.kgn_enable_gl_mutex)
967 return __kgnilnd_mutex_trylock(cq_lock);
969 return __kgnilnd_mutex_trylock(c_lock);
972 static inline void *kgnilnd_vzalloc(int size)
974 void *ret = __vmalloc(size, __GFP_HIGHMEM | GFP_NOIO | __GFP_NORETRY | __GFP_ZERO,
976 LIBCFS_ALLOC_POST(ret, size);
980 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
983 _kgnilnd_debug_msg(kgn_msg_t *msg,
984 struct libcfs_debug_msg_data *data, const char *fmt, ... );
986 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
988 CFS_CHECK_STACK(msgdata, mask, cdls); \
990 if (((mask) & D_CANTMASK) != 0 || \
991 ((libcfs_debug & (mask)) != 0 && \
992 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
993 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
996 /* for most callers (level is a constant) this is resolved at compile time */
997 #define GNIDBG_MSG(level, msg, fmt, args...) \
999 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1000 static struct cfs_debug_limit_state cdls; \
1001 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1002 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1003 "$$ "fmt" from %s ", ## args, \
1004 libcfs_nid2str((msg)->gnm_srcnid)); \
1006 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1007 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1008 "$$ "fmt" from %s ", ## args, \
1009 libcfs_nid2str((msg)->gnm_srcnid)); \
1013 /* user puts 'to nid' in msg for us */
1014 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
1016 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1017 static struct cfs_debug_limit_state cdls; \
1018 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1019 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1020 "$$ "fmt" ", ## args); \
1022 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1023 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1024 "$$ "fmt" ", ## args); \
1029 _kgnilnd_debug_conn(kgn_conn_t *conn,
1030 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1032 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
1034 CFS_CHECK_STACK(msgdata, mask, cdls); \
1036 if (((mask) & D_CANTMASK) != 0 || \
1037 ((libcfs_debug & (mask)) != 0 && \
1038 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1039 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
1042 /* for most callers (level is a constant) this is resolved at compile time */
1043 #define GNIDBG_CONN(level, conn, fmt, args...) \
1045 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1046 static struct cfs_debug_limit_state cdls; \
1047 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1048 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
1049 "$$ "fmt" ", ## args); \
1051 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1052 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
1053 "$$ "fmt" ", ## args); \
1058 _kgnilnd_debug_tx(kgn_tx_t *tx,
1059 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1061 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
1063 CFS_CHECK_STACK(msgdata, mask, cdls); \
1065 if (((mask) & D_CANTMASK) != 0 || \
1066 ((libcfs_debug & (mask)) != 0 && \
1067 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1068 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1071 /* for most callers (level is a constant) this is resolved at compile time */
1072 #define GNIDBG_TX(level, tx, fmt, args...) \
1074 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1075 static struct cfs_debug_limit_state cdls; \
1076 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1077 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1078 "$$ "fmt" ", ## args); \
1080 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1081 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1082 "$$ "fmt" ", ## args); \
1086 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1088 if (unlikely(!(cond))) { \
1089 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1094 #define GNILND_IS_QUIESCED \
1095 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1096 atomic_read(&kgnilnd_data.kgn_nthreads))
1098 #define KGNILND_SPIN_QUIESCE \
1100 /* E.T phone home */ \
1101 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1102 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1103 while (kgnilnd_data.kgn_quiesce_trigger) { \
1104 msleep_interruptible(MSEC_PER_SEC); \
1106 /* Mom, my homework is done */ \
1107 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1108 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1111 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1112 #ifndef LIBCFS_DEBUG
1113 #error "this code uses actions inside LASSERT for ref counting"
1116 #define kgnilnd_admin_addref(atomic) \
1118 int val = atomic_inc_return(&atomic); \
1119 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1120 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1123 #define kgnilnd_admin_decref(atomic) \
1125 int val = atomic_dec_return(&atomic); \
1126 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
1127 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1130 #define kgnilnd_net_addref(net) \
1132 int val = atomic_inc_return(&net->gnn_refcount); \
1133 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1134 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1135 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1138 #define kgnilnd_net_decref(net) \
1140 int val = atomic_dec_return(&net->gnn_refcount); \
1141 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1142 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1143 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1146 #define kgnilnd_peer_addref(peer) \
1148 int val = atomic_inc_return(&peer->gnp_refcount); \
1149 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1150 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1151 libcfs_nid2str(peer->gnp_nid), val); \
1154 #define kgnilnd_peer_decref(peer) \
1156 int val = atomic_dec_return(&peer->gnp_refcount); \
1157 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1158 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1159 libcfs_nid2str(peer->gnp_nid), val); \
1161 kgnilnd_destroy_peer(peer); \
1164 #define kgnilnd_conn_addref(conn) \
1169 val = atomic_inc_return(&conn->gnc_refcount); \
1170 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1171 "conn %p refc %d to %s\n", \
1174 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1176 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1178 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1183 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1184 * through the scheduler thread to get the EP destroyed. This avoids some
1185 * messy semaphore business and allows us to reuse the connd_list and existing
1186 * linkage and avoid creating extra lists just for destroying EPs */
1188 /* Safety Disclaimer:
1189 * Q: If we decrement the refcount and then check it again, is it possible that
1190 * another caller could have passed through this macro concurrently? If so,
1191 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1193 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1194 * once we are refcount <= 2. It hinges around gnc_state and membership of
1195 * gnc_hashlist. There are two ways to find a connection - either ask for
1196 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1197 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1200 * - #1 from create (kgnilnd_create_conn)
1201 * - #2 for EP (kgnilnd_create_conn)
1202 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1203 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1205 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1206 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1208 * There could be more from TX descriptors during the lifetime of a live
1211 * If we nuke the conn before finish_connect, we won't have parallel paths
1212 * because nobody besides the dgram handler for the single outstanding
1213 * dgram can find the connection as it isn't in any searchable tables yet.
1215 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1216 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1217 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1218 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1219 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1220 * sent. Now the gnc_state == CLOSED, so we hit
1221 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1222 * - we've nuked them from the peer and CQ id tables, so we own them and
1223 * are guaranteed serial access - hence the complete lack of conn list
1224 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1225 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1226 * kgnilnd_process_conns. Then the next scheduler loop would call
1227 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1228 * kgnilnd_process_conns.
1230 * Clearly, we are totally safe. Clearly.
1233 #define kgnilnd_conn_decref(conn) \
1238 val = atomic_dec_return(&conn->gnc_refcount); \
1239 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1242 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1244 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1246 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1251 (conn->gnc_ephandle != NULL) && \
1252 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1253 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1254 kgnilnd_schedule_conn(conn); \
1255 } else if (val == 0) { \
1256 kgnilnd_destroy_conn(conn); \
1260 static inline struct list_head *
1261 kgnilnd_nid2peerlist(lnet_nid_t nid)
1263 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1265 RETURN(&kgnilnd_data.kgn_peers[hash]);
1268 static inline struct list_head *
1269 kgnilnd_netnum2netlist(__u16 netnum)
1271 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1273 RETURN(&kgnilnd_data.kgn_nets[hash]);
1277 kgnilnd_peer_active(kgn_peer_t *peer)
1279 /* Am I in the peer hash table? */
1280 return (!list_empty(&peer->gnp_list));
1283 /* need write_lock on kgn_peer_conn_lock */
1285 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1287 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1288 peer, libcfs_nid2str(peer->gnp_nid),
1289 !list_empty(&peer->gnp_conns),
1290 !list_empty(&peer->gnp_tx_queue));
1292 /* kgn_peer_conn_lock protects us from conflict with
1293 * kgnilnd_peer_notify and gnp_persistent */
1294 RETURN ((list_empty(&peer->gnp_conns)) &&
1295 (list_empty(&peer->gnp_tx_queue)));
1298 /* returns positive if error was for a clean shutdown of conn */
1300 kgnilnd_conn_clean_errno(int errno)
1302 /* - ESHUTDOWN - LND is unloading
1303 * - EUCLEAN - admin requested via "lctl del_peer"
1304 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1305 * - ENOTRECOVERABLE - stack reset
1306 * - EISCONN - cleared via "lctl push"
1307 * not doing ESTALE - that isn't clean */
1308 RETURN ((errno == 0) ||
1309 (errno == -ESHUTDOWN) ||
1310 (errno == -EUCLEAN) ||
1311 (errno == -ENETRESET) ||
1312 (errno == -EISCONN) ||
1313 (errno == -ENOTRECOVERABLE));
1316 /* returns positive if error results in purgatory hold */
1318 kgnilnd_check_purgatory_errno(int errno)
1320 /* We don't want to save the purgatory lists these cases:
1321 * - EUCLEAN - admin requested via "lctl del_peer"
1322 * - ESHUTDOWN - LND is unloading
1324 RETURN ((errno != -ESHUTDOWN) &&
1325 (errno != -EUCLEAN));
1329 /* returns positive if a purgatory hold is needed */
1331 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1335 if (conn->gnc_peer) {
1336 loopback = conn->gnc_peer->gnp_nid ==
1337 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1339 /* short circuit - a conn that didn't complete
1340 * setup never needs a purgatory hold */
1343 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1344 conn, conn->gnc_peer ?
1345 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1347 loopback, conn->gnc_close_recvd);
1349 /* we only use a purgatory hold if we've not received the CLOSE msg
1350 * from our peer - without that message, we can't know the state of
1351 * the other end of this connection and must put it into purgatory
1352 * to prevent reuse and corruption.
1353 * The theory is that a TX error can be communicated in all other cases
1355 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1356 kgnilnd_check_purgatory_errno(conn->gnc_error));
1359 static inline const char *
1360 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1362 static inline struct list_head *
1363 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1364 kgn_tx_list_state_t to_state)
1367 case GNILND_TX_PEERQ:
1368 return &peer->gnp_tx_queue;
1369 case GNILND_TX_FMAQ:
1370 return &conn->gnc_fmaq;
1371 case GNILND_TX_LIVE_FMAQ:
1372 case GNILND_TX_LIVE_RDMAQ:
1373 case GNILND_TX_DYING:
1375 case GNILND_TX_MAPQ:
1376 return &conn->gnc_device->gnd_map_tx;
1377 case GNILND_TX_RDMAQ:
1378 return &conn->gnc_device->gnd_rdmaq;
1380 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1381 CERROR("invalid state requested: %s\n",
1382 kgnilnd_tx_state2str(to_state));
1388 /* should hold tx, conn or peer lock when calling */
1390 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1391 kgn_conn_t *conn, kgn_tx_list_state_t state,
1394 struct list_head *list = NULL;
1396 /* make sure we have a sane TX state to start */
1397 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1398 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1399 list_empty(&tx->tx_list),
1400 "bad state with tx_list %s",
1401 list_empty(&tx->tx_list) ? "empty" : "not empty");
1403 /* WTF - you are already on that state buttmunch */
1404 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1405 "already at %s", kgnilnd_tx_state2str(state));
1407 /* get proper list from the state requested */
1408 list = kgnilnd_tx_state2list(peer, conn, state);
1412 case GNILND_TX_PEERQ:
1413 kgnilnd_peer_addref(peer);
1415 case GNILND_TX_ALLOCD:
1416 /* no refs needed */
1418 case GNILND_TX_FMAQ:
1419 kgnilnd_conn_addref(conn);
1421 case GNILND_TX_MAPQ:
1422 atomic_inc(&conn->gnc_device->gnd_nq_map);
1423 kgnilnd_conn_addref(conn);
1425 case GNILND_TX_LIVE_FMAQ:
1426 atomic_inc(&conn->gnc_nlive_fma);
1427 kgnilnd_conn_addref(conn);
1429 case GNILND_TX_LIVE_RDMAQ:
1430 atomic_inc(&conn->gnc_nlive_rdma);
1431 kgnilnd_conn_addref(conn);
1433 case GNILND_TX_RDMAQ:
1434 atomic_inc(&conn->gnc_nq_rdma);
1435 kgnilnd_conn_addref(conn);
1437 case GNILND_TX_DYING:
1438 kgnilnd_conn_addref(conn);
1441 CERROR("invalid state requested: %s\n",
1442 kgnilnd_tx_state2str(state));
1447 /* if this changes, change kgnilnd_alloc_tx */
1448 tx->tx_list_state = state;
1450 /* some states don't have lists - we track them in the per conn
1451 * TX table instead. Waste not, want not! */
1453 tx->tx_list_p = list;
1455 list_add_tail(&tx->tx_list, list);
1457 list_add(&tx->tx_list, list);
1459 /* set dummy list_p to make book keeping happy and let debugging
1460 * be a hair easier */
1461 tx->tx_list_p = (void *)state;
1464 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1465 kgnilnd_tx_state2str(state), list);
1469 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1470 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1472 /* These is only 1 "off-list" state */
1473 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1474 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1476 /* new_state == ALLOCD means we are deallocating this tx,
1477 * so make sure it was on a valid list to start with */
1478 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1479 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1480 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1481 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1484 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1487 switch (tx->tx_list_state) {
1488 case GNILND_TX_PEERQ:
1489 kgnilnd_peer_decref(peer);
1491 case GNILND_TX_FREED:
1492 case GNILND_TX_IDLE:
1493 case GNILND_TX_ALLOCD:
1494 /* no refs needed */
1496 case GNILND_TX_DYING:
1497 kgnilnd_conn_decref(conn);
1499 case GNILND_TX_FMAQ:
1500 kgnilnd_conn_decref(conn);
1502 case GNILND_TX_MAPQ:
1503 atomic_dec(&conn->gnc_device->gnd_nq_map);
1504 kgnilnd_conn_decref(conn);
1506 case GNILND_TX_LIVE_FMAQ:
1507 atomic_dec(&conn->gnc_nlive_fma);
1508 kgnilnd_conn_decref(conn);
1510 case GNILND_TX_LIVE_RDMAQ:
1511 atomic_dec(&conn->gnc_nlive_rdma);
1512 kgnilnd_conn_decref(conn);
1514 case GNILND_TX_RDMAQ:
1515 atomic_dec(&conn->gnc_nq_rdma);
1516 kgnilnd_conn_decref(conn);
1517 /* don't need to assert on default, already did in set */
1520 /* for ALLOCD, this might already be true, but no harm doing it again */
1521 list_del_init(&tx->tx_list);
1522 tx->tx_list_p = NULL;
1523 tx->tx_list_state = new_state;
1527 kgnilnd_tx_mapped(kgn_tx_t *tx)
1529 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1530 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1533 static inline struct list_head *
1534 kgnilnd_cqid2connlist(__u32 cqid)
1536 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1538 return (&kgnilnd_data.kgn_conns [hash]);
1541 static inline kgn_conn_t *
1542 kgnilnd_cqid2conn_locked(__u32 cqid)
1544 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1545 struct list_head *tmp;
1548 list_for_each(tmp, conns) {
1549 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1551 if (conn->gnc_cqid == cqid)
1558 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1560 kgnilnd_get_cqid_locked(void)
1566 cqid = kgnilnd_data.kgn_next_cqid++;
1567 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1571 kgnilnd_data.kgn_next_cqid = 1;
1574 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1580 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1582 kgn_tx_t *tx = NULL;
1583 kgn_conn_t *conn = NULL;
1585 /* set to NULL so any early return is an error */
1589 LASSERTF((ev_id->txe_idx > 0) &&
1590 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1591 "bogus txe_idx %d >= %d\n",
1592 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1594 LASSERTF((ev_id->txe_cqid > 0) &&
1595 (ev_id->txe_cqid < GNILND_MAX_CQID),
1596 "bogus txe_cqid %d >= %d\n",
1597 ev_id->txe_cqid, GNILND_MAX_CQID);
1599 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1600 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1603 /* Conn was destroyed? */
1604 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1605 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1608 /* just insurance */
1609 kgnilnd_conn_addref(conn);
1610 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1611 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1613 /* we know this is safe - as the TX won't be reused until AFTER
1614 * the conn is unlinked from the cqid hash, so we can use the TX
1615 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1617 spin_lock(&conn->gnc_tx_lock);
1618 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1619 spin_unlock(&conn->gnc_tx_lock);
1621 /* We could have a tx that was cleared out by other forces
1622 * lctl disconnect or del_peer. */
1624 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1625 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1626 kgnilnd_conn_decref(conn);
1630 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1631 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1632 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1634 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1635 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1636 "found tx %p txe_idx %d",
1637 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1638 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1640 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1642 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1647 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1650 /* set_normalized_timepsec isn't exported from the kernel, so
1651 * we need to do the same thing inline */
1652 static inline struct timespec
1653 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1659 sec = lhs.tv_sec - rhs.tv_sec;
1660 nsec = lhs.tv_nsec - rhs.tv_nsec;
1662 while (nsec >= NSEC_PER_SEC) {
1663 nsec -= NSEC_PER_SEC;
1667 nsec += NSEC_PER_SEC;
1676 kgnilnd_count_list(struct list_head *q)
1678 struct list_head *e;
1681 list_for_each(e, q) {
1688 /* kgnilnd_find_net adds a reference to the net it finds
1689 * this is so the net will not be removed before the calling function
1690 * has time to use the data returned. This reference needs to be released
1691 * by the calling function once it has finished using the returned net
1695 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1700 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1706 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1707 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1708 kgnilnd_net_addref(net);
1709 up_read(&kgnilnd_data.kgn_net_rw_sem);
1715 up_read(&kgnilnd_data.kgn_net_rw_sem);
1720 #ifdef CONFIG_DEBUG_SLAB
1721 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1723 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1726 int kgnilnd_dev_init(kgn_device_t *dev);
1727 void kgnilnd_dev_fini(kgn_device_t *dev);
1728 int kgnilnd_startup(lnet_ni_t *ni);
1729 void kgnilnd_shutdown(lnet_ni_t *ni);
1730 int kgnilnd_base_startup(void);
1731 void kgnilnd_base_shutdown(void);
1733 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1734 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1735 void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
1736 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1738 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1739 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1740 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1741 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1742 lnet_msg_t *lntmsg, void **new_private);
1743 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1744 int delayed, unsigned int niov,
1745 struct kvec *iov, lnet_kiov_t *kiov,
1746 unsigned int offset, unsigned int mlen, unsigned int rlen);
1748 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1750 /* purgatory functions */
1751 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1752 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1753 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1754 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1756 void kgnilnd_update_reaper_timeout(long timeout);
1757 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1758 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1759 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1760 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1761 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1762 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
1763 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1765 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1766 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
1767 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1768 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1770 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1771 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1772 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1773 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1774 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1775 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1776 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1777 void kgnilnd_consume_rx(kgn_rx_t *rx);
1779 void kgnilnd_schedule_device(kgn_device_t *dev);
1780 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1781 void kgnilnd_schedule_device_timer(unsigned long arg);
1783 int kgnilnd_reaper(void *arg);
1784 int kgnilnd_scheduler(void *arg);
1785 int kgnilnd_dgram_mover(void *arg);
1786 int kgnilnd_rca(void *arg);
1787 int kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id);
1789 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1790 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1791 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1792 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1793 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1794 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1795 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1796 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1797 void kgnilnd_peer_alive(kgn_peer_t *peer);
1798 void kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive);
1799 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1800 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1801 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1802 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1804 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1805 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1806 void kgnilnd_wakeup_rca_thread(void);
1807 int kgnilnd_start_rca_thread(void);
1808 int kgnilnd_get_node_state(__u32 nid);
1810 int kgnilnd_tunables_init(void);
1811 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1813 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1814 void kgnilnd_pause_threads(void);
1815 int kgnilnd_hw_in_quiesce(void);
1816 int kgnilnd_check_hw_quiesce(void);
1817 void kgnilnd_quiesce_wait(char *reason);
1818 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1819 int kgnilnd_ruhroh_thread(void *arg);
1820 void kgnilnd_reset_stack(void);
1821 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1823 void kgnilnd_insert_sysctl(void);
1824 void kgnilnd_remove_sysctl(void);
1825 void kgnilnd_proc_init(void);
1826 void kgnilnd_proc_fini(void);
1829 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1831 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1832 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1833 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
1835 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1836 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1837 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1838 int kgnilnd_cancel_dgrams(kgn_device_t *dev);
1839 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1841 int kgnilnd_dgram_waitq(void *arg);
1843 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1845 /* struct2str functions - we don't use a default: case to cause the compile
1846 * to fail if there is a missing case. This allows us to hide these down here
1847 * out of the way but ensure we'll catch any updates to the enum/types
1850 #define DO_TYPE(x) case x: return #x;
1851 static inline const char *
1852 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1854 /* Only want single char string for this */
1856 case GNILND_FMABLK_IDLE:
1858 case GNILND_FMABLK_PHYS:
1860 case GNILND_FMABLK_VIRT:
1862 case GNILND_FMABLK_FREED:
1865 return "<unknown state>";
1868 static inline const char *
1869 kgnilnd_msgtype2str(int type)
1872 DO_TYPE(GNILND_MSG_NONE);
1873 DO_TYPE(GNILND_MSG_NOOP);
1874 DO_TYPE(GNILND_MSG_IMMEDIATE);
1875 DO_TYPE(GNILND_MSG_PUT_REQ);
1876 DO_TYPE(GNILND_MSG_PUT_NAK);
1877 DO_TYPE(GNILND_MSG_PUT_ACK);
1878 DO_TYPE(GNILND_MSG_PUT_DONE);
1879 DO_TYPE(GNILND_MSG_GET_REQ);
1880 DO_TYPE(GNILND_MSG_GET_NAK);
1881 DO_TYPE(GNILND_MSG_GET_DONE);
1882 DO_TYPE(GNILND_MSG_CLOSE);
1883 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1884 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1885 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1886 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1887 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1888 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1889 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1891 return "<unknown msg type>";
1894 static inline const char *
1895 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1898 DO_TYPE(GNILND_TX_IDLE);
1899 DO_TYPE(GNILND_TX_ALLOCD);
1900 DO_TYPE(GNILND_TX_PEERQ);
1901 DO_TYPE(GNILND_TX_MAPQ);
1902 DO_TYPE(GNILND_TX_FMAQ);
1903 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1904 DO_TYPE(GNILND_TX_RDMAQ);
1905 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1906 DO_TYPE(GNILND_TX_DYING);
1907 DO_TYPE(GNILND_TX_FREED);
1909 return "<unknown state>";
1912 static inline const char *
1913 kgnilnd_conn_state2str(kgn_conn_t *conn)
1915 kgn_conn_state_t state = conn->gnc_state;
1917 DO_TYPE(GNILND_CONN_DUMMY);
1918 DO_TYPE(GNILND_CONN_LISTEN);
1919 DO_TYPE(GNILND_CONN_CONNECTING);
1920 DO_TYPE(GNILND_CONN_ESTABLISHED);
1921 DO_TYPE(GNILND_CONN_CLOSING);
1922 DO_TYPE(GNILND_CONN_CLOSED);
1923 DO_TYPE(GNILND_CONN_DONE);
1924 DO_TYPE(GNILND_CONN_DESTROY_EP);
1929 static inline const char *
1930 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1932 kgn_connreq_type_t type = connreq->gncr_type;
1935 DO_TYPE(GNILND_CONNREQ_REQ);
1936 DO_TYPE(GNILND_CONNREQ_NAK);
1937 DO_TYPE(GNILND_CONNREQ_CLOSE);
1942 static inline const char *
1943 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1945 kgn_dgram_state_t state = dgram->gndg_state;
1948 DO_TYPE(GNILND_DGRAM_USED);
1949 DO_TYPE(GNILND_DGRAM_POSTING);
1950 DO_TYPE(GNILND_DGRAM_POSTED);
1951 DO_TYPE(GNILND_DGRAM_PROCESSING);
1952 DO_TYPE(GNILND_DGRAM_DONE);
1953 DO_TYPE(GNILND_DGRAM_CANCELED);
1958 static inline const char *
1959 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1961 kgn_dgram_type_t type = dgram->gndg_type;
1964 DO_TYPE(GNILND_DGRAM_REQ);
1965 DO_TYPE(GNILND_DGRAM_WC_REQ);
1966 DO_TYPE(GNILND_DGRAM_NAK);
1967 DO_TYPE(GNILND_DGRAM_CLOSE);
1972 static inline const char *
1973 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
1976 DO_TYPE(GNILND_DGRAM_REQ);
1977 DO_TYPE(GNILND_DGRAM_WC_REQ);
1978 DO_TYPE(GNILND_DGRAM_NAK);
1979 DO_TYPE(GNILND_DGRAM_CLOSE);
1986 /* pulls in tunables per platform and adds in nid/nic conversion
1987 * if RCA wasn't available at build time */
1988 #include "gnilnd_hss_ops.h"
1989 /* API wrapper functions - include late to pick up all of the other defines */
1990 #include "gnilnd_api_wrap.h"
1992 #if defined(CONFIG_CRAY_GEMINI)
1993 #include "gnilnd_gemini.h"
1994 #elif defined(CONFIG_CRAY_ARIES)
1995 #include "gnilnd_aries.h"
1997 #error "Undefined Network Hardware Type"
2000 extern uint32_t kgni_driver_version;
2003 kgnilnd_check_kgni_version(void)
2007 kgnilnd_data.kgn_enable_gl_mutex = 1;
2008 kdv = symbol_get(kgni_driver_version);
2010 LCONSOLE_INFO("Not using thread safe locking -"
2011 " no symbol kgni_driver_version\n");
2015 /* Thread-safe kgni implemented in minor ver 0x44/45, code rev 0xb9 */
2016 if (*kdv < GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9)) {
2017 symbol_put(kgni_driver_version);
2018 LCONSOLE_INFO("Not using thread safe locking, gni version 0x%x,"
2019 " need >= 0x%x\n", *kdv,
2020 GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9));
2024 symbol_put(kgni_driver_version);
2026 if (!*kgnilnd_tunables.kgn_thread_safe) {
2030 /* Use thread-safe locking */
2031 kgnilnd_data.kgn_enable_gl_mutex = 0;
2034 #endif /* _GNILND_GNILND_H_ */