2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Copyright (c) 2014, 2016, 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 #define DEBUG_SUBSYSTEM S_LND
33 #include <linux/module.h>
34 #include <linux/kernel.h>
36 #include <linux/string.h>
37 #include <linux/stat.h>
38 #include <linux/errno.h>
39 #ifdef HAVE_LINUX_KERNEL_LOCK
40 #include <linux/smp_lock.h>
42 #include <linux/unistd.h>
43 #include <linux/uio.h>
44 #include <linux/time.h>
45 #include <asm/timex.h>
47 #include <asm/uaccess.h>
50 #include <linux/init.h>
52 #include <linux/file.h>
53 #include <linux/stat.h>
54 #include <linux/list.h>
55 #include <linux/kmod.h>
56 #include <linux/sysctl.h>
57 #include <linux/kthread.h>
58 #include <linux/nmi.h>
62 #include <linux/nmi.h>
64 #include <lnet/lib-lnet.h>
68 static inline time_t cfs_duration_sec(long duration_jiffies)
70 return jiffies_to_msecs(duration_jiffies) / MSEC_PER_SEC;
74 #define GNILND_MBOX_SIZE KMALLOC_MAX_SIZE
76 #define GNILND_SHIFT_HIGH ((MAX_ORDER + PAGE_SHIFT - 1) <= 25 ? \
77 (MAX_ORDER + PAGE_SHIFT - 1) : 25)
78 #define GNILND_SHIFT_MAX GNILND_SHIFT_HIGH
79 #define GNILND_MBOX_SIZE (1UL << GNILND_SHIFT_MAX)
83 /* tunables determined at compile time */
84 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
85 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
86 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
87 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
88 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
89 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
90 #define GNILND_SCHED_TIMEOUT 1
91 #define GNILND_DGRAM_TIMEOUT 2
92 #define GNILND_FAST_MAPPING_TRY \
93 *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
94 #define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
96 /* map failure timeout */
97 #define GNILND_MAP_TIMEOUT \
98 (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
99 *kgnilnd_tunables.kgn_timeout))
101 /* Should we use the no_retry flag with vzalloc */
102 #define GNILND_VZALLOC_RETRY 0
104 /* reaper thread wakup interval */
105 #define GNILND_REAPER_THREAD_WAKE 1
106 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
107 #define GNILND_REAPER_NCHECKS 4
109 /* fixed constants */
110 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
111 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
112 #define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
113 /* checksum values */
114 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
115 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
116 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
117 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
119 /* tune down some COMPUTE options as they won't see the same number of connections and
120 * don't need the throughput of multiple threads by default */
121 #if defined(CONFIG_CRAY_COMPUTE)
123 #define GNILND_SCHED_THREADS 2 /* default # of kgnilnd_scheduler threads */
125 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
127 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
128 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
129 #define GNILND_COMPUTE 1 /* compute image */
130 #define GNILND_FAST_RECONNECT 1 /* Fast Reconnect option */
131 #define GNILND_DEFAULT_CREDITS 64 /* Default number of simultaneous transmits */
133 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
134 #define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
135 #define GNILND_COMPUTE 0 /* service image */
136 #define GNILND_FAST_RECONNECT 0 /* Fast Reconnect option */
137 #define GNILND_DEFAULT_CREDITS 256 /* Default number of simultaneous transmits */
140 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
141 #define GNILND_EXTRA_BITS 1
142 /* maximum number of conns & bits for cqid in the SMSG event data */
143 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
144 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
145 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
146 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
147 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
149 /* need sane upper bound to limit copy overhead */
150 #define GNILND_MAX_IMMEDIATE (64<<10)
151 /* allow for 4M transfers over gni. Note 2.5M used by DVS */
152 #define GNILND_MAX_IOV 1024
154 /* Max number of connections to keep in purgatory per peer */
155 #define GNILND_PURGATORY_MAX 5
156 /* Closing, don't put in purgatory */
157 #define GNILND_NOPURG 222
159 /* payload size to add to the base mailbox size
160 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
161 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
162 * the calculation return from that function.*/
163 #define GNILND_MBOX_PAYLOAD \
164 (GNILND_MAX_MSG_SIZE * \
165 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
167 /* timeout -> deadman timer for kgni mdd holds */
168 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
170 /* timeout for failing sends in t is in jiffies*/
171 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
173 /* time when to release from purgatory in the reaper thread in jiffies */
174 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
176 /* Macro for finding last_rx 2 datapoints are compared
177 * and the most recent one in jiffies is returned.
179 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
180 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
182 /* fmablk registration failures timeout before failing node */
183 #define GNILND_REGFAILTO_DISABLE -1
185 /************************************************************************
186 * Enum, flag and tag data
188 #define GNILND_INIT_NOTHING 0
189 #define GNILND_INIT_DATA 1
190 #define GNILND_INIT_ALL 2
192 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
193 #define GNILND_BUF_NONE 0 /* buffer type not set */
194 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
195 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
196 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
197 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
199 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
200 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
201 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
202 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
203 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
205 /* stash above max CQID to avoid any collision */
206 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
207 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
209 /* kgn_msg_t::gnm_type */
210 #define GNILND_MSG_NONE 0x00 /* illegal message */
211 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
212 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
213 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
214 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
215 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
216 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
217 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
218 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
219 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
220 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
221 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
222 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
223 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
224 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
225 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
226 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
227 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
229 /* defines for gnc_*scheduled states */
230 #define GNILND_CONN_IDLE 0
231 #define GNILND_CONN_SCHED 1
232 #define GNILND_CONN_WANTS_SCHED 2
233 #define GNILND_CONN_PROCESS 3
235 #define GNILND_DEV_IDLE 0
236 #define GNILND_DEV_IRQ 1
237 #define GNILND_DEV_LOOP 2
239 #define GNILND_DGRAM_IDLE 0
240 #define GNILND_DGRAM_SCHED 1
241 #define GNILND_DGRAM_PROCESS 2
243 #define GNILND_PEER_IDLE 0
244 #define GNILND_PEER_CONNECT 1
245 #define GNILND_PEER_POSTING 2
246 #define GNILND_PEER_POSTED 3
247 #define GNILND_PEER_NEEDS_DEATH 4
248 #define GNILND_PEER_KILL 5
250 /* for gnc_close_recvd */
251 #define GNILND_CLOSE_RX 1
252 #define GNILND_CLOSE_INJECT1 2
253 #define GNILND_CLOSE_INJECT2 3
254 #define GNILND_CLOSE_EARLY 4
256 /* defines for why quiesce trigger set */
257 #define GNILND_QUIESCE_IDLE 0
258 #define GNILND_QUIESCE_ADMIN 1
259 #define GNILND_QUIESCE_RESET 2
260 #define GNILND_QUIESCE_HW_QUIESCE 3
262 #define GNILND_PEER_CLEAN 0
263 #define GNILND_PEER_PERSISTING 1
265 #define GNILND_DEL_CONN 0
266 #define GNILND_DEL_PEER 1
267 #define GNILND_CLEAR_PURGATORY 2
269 #define GNILND_PEER_UP 0
270 #define GNILND_PEER_DOWN 1
271 #define GNILND_PEER_TIMED_OUT 2
272 #define GNILND_PEER_UNKNOWN 3
274 /* defines for reverse RDMA states */
275 #define GNILND_REVERSE_NONE 0
276 #define GNILND_REVERSE_GET 1
277 #define GNILND_REVERSE_PUT 2
278 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
280 typedef enum kgn_fmablk_state {
281 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
282 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
283 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
284 GNILND_FMABLK_FREED, /* after free */
285 } kgn_fmablk_state_t;
287 typedef enum kgn_tx_list_state {
288 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
289 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
290 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
291 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
292 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
293 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
294 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
295 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
296 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
297 GNILND_TX_FREED /* TX is free! */
298 } kgn_tx_list_state_t;
300 typedef enum kgn_conn_state {
301 /* don't start @ 0 - prevent memset(0) badness */
302 GNILND_CONN_DUMMY = 0,
304 GNILND_CONN_CONNECTING,
305 GNILND_CONN_ESTABLISHED,
309 GNILND_CONN_DESTROY_EP
312 /* changing these requires a change to GNILND_CONNREQ_VERSION and
313 * will result in dropped packets instead of NAKs. Adding to this is
314 * acceptable without changing the CONNREQ_VERSION, but code should
315 * be ready to handle NAKs on version mismatch */
316 typedef enum kgn_connreq_type {
317 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
318 GNILND_CONNREQ_NAK, /* NO soup for you! */
319 GNILND_CONNREQ_CLOSE, /* we should see other people */
320 } kgn_connreq_type_t;
322 typedef enum kgn_dgram_state {
323 /* don't use 0 to avoid thinking a memset of zero is valid data */
324 GNILND_DGRAM_USED = 1,
325 GNILND_DGRAM_POSTING,
327 GNILND_DGRAM_PROCESSING,
328 GNILND_DGRAM_CANCELED,
332 typedef enum kgn_dgram_type {
333 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
334 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
335 GNILND_DGRAM_NAK, /* NO soup for you! */
336 GNILND_DGRAM_CLOSE, /* we should see other people */
339 /************************************************************************
340 * Wire message structs. These are sent in sender's byte order
341 * (i.e. receiver checks magic and flips if required).
344 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
345 #define GNILND_DGRAM_MAGIC 0x0DDBA11
347 /* kgn_msg_t - FMA/SMSG wire struct
349 * - added checksum to FMA
350 * moved seq before paylod
351 * __packed added for alignment
353 * added gnm_payload_len for FMA payload size
355 * added gncm_retval to completion, allowing return code transmission
358 * changed how CQID and TX ids are assigned
360 * added retval on CLOSE
362 * added payload checksumming
364 * reworked checksumming a bit, changed payload checksums
366 #define GNILND_MSG_VERSION 8
367 /* kgn_connreq_t connection request datagram wire struct
372 #define GNILND_CONNREQ_VERSION 2
374 typedef struct kgn_gniparams {
375 __u32 gnpr_host_id; /* ph. host ID of the NIC */
376 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
377 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
378 } __packed kgn_gniparams_t;
380 typedef struct kgn_nak_data {
381 __s32 gnnd_errno; /* errno reason for NAK */
383 } __packed kgn_nak_data_t;
385 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
386 * without breaking the ability for us to properly NAK someone */
387 typedef struct kgn_connreq { /* connection request/response */
388 __u32 gncr_magic; /* I'm an gnilnd connreq */
389 __u32 gncr_cksum; /* checksum (0 == disabled) */
390 __u16 gncr_type; /* REQ, NAK, etc */
391 __u16 gncr_version; /* this is my version number */
392 __u32 gncr_timeout; /* sender's timeout */
393 __u64 gncr_srcnid; /* sender's NID */
394 __u64 gncr_dstnid; /* who sender expects to listen */
395 __u64 gncr_peerstamp; /* sender's instance stamp */
396 __u64 gncr_connstamp; /* sender's connection stamp */
398 /* everything before this needs to stay static, adding after should
399 * result in a change to GNILND_CONNREQ_VERSION */
402 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
403 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
405 } __packed kgn_connreq_t;
408 gni_mem_handle_t gnrd_key;
411 } __packed kgn_rdma_desc_t;
414 struct lnet_hdr_nid4 gnim_hdr; /* LNet header */
415 /* LNet payload is in FMA "Message Data" */
416 } __packed kgn_immediate_msg_t;
419 struct lnet_hdr_nid4 gnprm_hdr; /* LNet header */
420 __u64 gnprm_cookie; /* opaque completion cookie */
421 } __packed kgn_putreq_msg_t;
424 __u64 gnpam_src_cookie; /* reflected completion cookie */
425 __u64 gnpam_dst_cookie; /* opaque completion cookie */
426 __u16 gnpam_payload_cksum; /* checksum for get msg */
427 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
428 } __packed kgn_putack_msg_t;
431 struct lnet_hdr_nid4 gngm_hdr; /* LNet header */
432 __u64 gngm_cookie; /* opaque completion cookie */
433 __u16 gngm_payload_cksum; /* checksum for put msg */
434 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
435 } __packed kgn_get_msg_t;
438 int gncm_retval; /* error on NAK, size on REQ */
439 __u64 gncm_cookie; /* reflected completion cookie */
440 } __packed kgn_completion_msg_t;
442 typedef struct { /* NB must fit in FMA "Prefix" */
443 __u32 gnm_magic; /* I'm an gni message */
444 __u16 gnm_version; /* this is my version number */
445 __u16 gnm_type; /* msg type */
446 __u64 gnm_srcnid; /* sender's NID */
447 __u64 gnm_connstamp; /* sender's connection stamp */
448 __u32 gnm_seq; /* incrementing sequence number */
449 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
450 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
451 __u32 gnm_payload_len; /* size of the FMA payload sent */
453 kgn_immediate_msg_t immediate;
454 kgn_putreq_msg_t putreq;
455 kgn_putack_msg_t putack;
457 kgn_completion_msg_t completion;
459 } __packed kgn_msg_t;
461 /************************************************************************
462 * runtime tunable data
465 typedef struct kgn_tunables {
466 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
467 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
468 int *kgn_credits; /* # concurrent sends */
469 int *kgn_fma_cq_size; /* # entries in receive CQ */
470 int *kgn_peer_credits; /* # LNet peer credits */
471 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
472 int *kgn_timeout; /* comms timeout (seconds) */
473 int *kgn_max_immediate; /* immediate payload breakpoint */
474 int *kgn_checksum; /* checksum data */
475 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
476 int *kgn_bte_put_dlvr_mode; /* BTE Put delivery mode */
477 int *kgn_bte_get_dlvr_mode; /* BTE Get delivery mode */
478 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
479 int *kgn_ptag; /* PTAG for cdm_create */
480 int *kgn_pkey; /* PKEY for cdm_create */
481 int *kgn_max_retransmits; /* max number of FMA retransmits before entering delay list */
482 int *kgn_nwildcard; /* # wildcard per net to post */
483 int *kgn_nice; /* nice value for kgnilnd threads */
484 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
485 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
486 int *kgn_peer_hash_size; /* size of kgn_peers */
487 int *kgn_peer_health; /* enable/disable peer health */
488 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
489 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
490 int *kgn_mbox_per_block; /* mailboxes per fmablk */
491 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
492 int *kgn_mbox_credits; /* max credits per fma */
493 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
494 int *kgn_net_hash_size; /* size of kgn_net_ht */
495 int *kgn_hardware_timeout; /* max time for a message to get across the network */
496 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
497 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
498 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
499 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
500 int *kgn_reverse_rdma; /* Reverse RDMA setting */
501 int *kgn_eager_credits; /* allocated eager buffers */
502 int *kgn_fast_reconn; /* fast reconnection on conn timeout */
503 int *kgn_efault_lbug; /* LBUG on receiving an EFAULT */
504 int *kgn_max_purgatory; /* # conns/peer to keep in purgatory */
505 int *kgn_reg_fail_timeout; /* registration failure timeout */
506 int *kgn_thread_affinity; /* bind scheduler threads to cpus */
507 int *kgn_to_reconn_disable;/* disable reconnect after timeout */
508 int *kgn_thread_safe; /* use thread safe kgni API */
509 int *kgn_vzalloc_noretry; /* Should we pass the noretry flag */
512 typedef struct kgn_mbox_info {
513 lnet_nid_t mbx_prev_nid;
514 lnet_nid_t mbx_prev_purg_nid;
515 unsigned long mbx_create_conn_memset;
516 unsigned long mbx_add_purgatory;
517 unsigned long mbx_detach_of_purgatory;
518 unsigned long mbx_release_from_purgatory;
519 unsigned long mbx_release_purg_active_dgram;
521 int mbx_nallocs_total;
524 typedef struct kgn_fma_memblock {
525 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
526 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
527 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
528 int gnm_num_mboxs; /* total mboxes allocated */
529 int gnm_avail_mboxs; /* number of available mailboxes in the block */
530 int gnm_held_mboxs; /* number of purgatory held mailboxes */
531 int gnm_mbox_size; /* size of the single mailbox */
532 int gnm_next_avail_mbox; /* next available mailbox */
533 long gnm_max_timeout; /* max timeout for possible purgatory hold */
534 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
535 void *gnm_block; /* pointer to mem. block */
536 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
537 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
538 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
539 } kgn_fma_memblock_t;
541 typedef struct kgn_device {
542 gni_nic_handle_t gnd_handle; /* device handle */
543 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
544 gni_err_handle_t gnd_err_handle; /* device error handle */
545 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
546 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
547 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
548 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
549 struct mutex gnd_cq_mutex; /* CQ access serialization */
550 __u32 gnd_host_id; /* ph. host ID of the NIC */
551 int gnd_id; /* device id, also index in kgn_devices */
552 __u32 gnd_nid; /* ph host ID translated to NID */
553 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
554 struct mutex gnd_fmablk_mutex; /* mutex for FMA block memory alloc/free */
555 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
556 atomic_t gnd_nfmablk; /* # of fmablk live */
557 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
558 atomic_t gnd_neps; /* # EP allocated to conns */
559 short gnd_ready; /* stuff to do in scheduler thread */
560 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
561 struct list_head gnd_delay_conns; /* connections in need of dla/or smsg credits */
562 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
563 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
564 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
565 struct list_head gnd_connd_peers; /* peers waiting for a connection */
566 spinlock_t gnd_connd_lock; /* serialise connd_peers */
567 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
568 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
569 int gnd_dgram_ready; /* dgrams need movin' */
570 struct list_head *gnd_dgrams; /* nid hash to dgrams */
571 atomic_t gnd_ndgrams; /* # dgrams extant */
572 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post*/
573 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
574 struct list_head gnd_map_list; /* list of all mapped regions */
575 int gnd_map_version; /* version flag for map list */
576 struct timer_list gnd_map_timer; /* wakey-wakey */
577 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
578 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
579 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
580 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
581 __u32 gnd_map_nphys; /* # TX phys mappings */
582 __u32 gnd_map_physnop; /* # TX phys pages mapped */
583 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
584 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
585 int gnd_map_attempt; /* last map attempt # */
586 unsigned long gnd_last_map; /* map timeout base */
587 struct list_head gnd_rdmaq; /* RDMA to be sent */
588 spinlock_t gnd_rdmaq_lock; /* play nice with others */
589 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
590 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
591 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
592 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
593 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
594 atomic_t gnd_short_ntx; /* TX stats: short messages */
595 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
596 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
597 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
598 atomic_t gnd_short_nrx; /* RX stats: short messages */
599 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
600 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
601 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
602 atomic_t gnd_fast_try; /* # of times fast send tried */
603 atomic_t gnd_fast_ok; /* # of times fast send ok */
604 atomic_t gnd_fast_block; /* # of times fast send blocked */
605 unsigned long gnd_mutex_delay;
606 atomic_t gnd_n_yield;
607 atomic_t gnd_n_schedule;
608 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
609 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
610 void *gnd_smdd_hold_buf; /* buffer to keep smdd */
611 gni_mem_handle_t gnd_smdd_hold_hndl; /* buffer mem handle */
614 typedef struct kgn_net {
615 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
616 kgn_device_t *gnn_dev; /* device for this net */
617 struct lnet_ni *gnn_ni; /* network interface instance */
618 atomic_t gnn_refcount; /* # current references */
619 int gnn_shutdown; /* lnd_shutdown set */
620 __u16 gnn_netnum; /* stash netnum for quicker lookup */
623 static inline lnet_nid_t
624 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
626 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
629 static inline lnet_nid_t
630 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
632 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
635 /* The code for this is a bit ugly - but really this just boils down to a __u64
636 * that can have various parts accessed separately.
638 * The lower 32 bits is the ID
639 * we give to SMSG for our completion event - it needs to be globally unique across
640 * all TX currently in flight. We separate that out into the CQID so that we can
641 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
642 * the actual TX out of the per-connection gnc_tx_ref_table.
644 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
645 * has a unique value we can send with RDMA setup messages to ensure the completion for
646 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
647 * reuse is separated.
650 typedef struct kgn_tx_ev_id {
652 __u64 txe_cookie; /* are you my mommy ? */
654 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
656 __u32 txe_smsg_id; /* ID for SMSG CQ event */
657 /* N.B: Never ever ever ever use the bit shifts directly,
658 * you are just asking for a world of pain and are at the
659 * mercy of the compiler layouts */
661 __u32 txe_cqid :GNILND_CQID_NBITS;
662 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
669 typedef struct kgn_dgram {
670 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
671 kgn_dgram_state_t gndg_state; /* state of this dgram */
672 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
673 __u32 gndg_magic; /* saftey word */
674 unsigned long gndg_post_time; /* time when we posted */
675 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
676 kgn_connreq_t gndg_conn_out; /* connreq from local node */
677 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
680 typedef struct kgn_tx { /* message descriptor */
681 struct list_head tx_list; /* TX queues - peer, conn, rdma */
682 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
683 struct list_head *tx_list_p; /* pointer to current list */
684 struct kgn_conn *tx_conn; /* owning conn */
685 struct lnet_msg *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
686 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
687 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
688 struct list_head tx_map_list; /* list entry on device map list */
689 unsigned int tx_nob; /* # bytes of payload */
690 int tx_buftype; /* payload buffer type */
691 int tx_phys_npages; /* # physical pages */
692 gni_mem_handle_t tx_map_key; /* mapping key */
693 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
694 gni_mem_segment_t *tx_phys; /* page descriptors */
695 kgn_msg_t tx_msg; /* FMA message buffer */
696 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
697 __u8 tx_state; /* state of the descriptor */
698 int tx_retrans; /* retrans count of RDMA */
699 int tx_rc; /* if we need to stash the ret code until we see completion */
700 void *tx_buffer; /* source/sink buffer */
701 void *tx_buffer_copy; /* pointer to page aligned buffer */
702 unsigned int tx_nob_rdma; /* nob actually rdma */
703 unsigned int tx_offset; /* offset of data into copied buffer */
705 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
706 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
709 /* we only use one or the other */
711 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
712 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
716 typedef struct kgn_conn {
717 kgn_device_t *gnc_device; /* which device */
718 struct kgn_peer *gnc_peer; /* owning peer */
719 int gnc_magic; /* magic value cleared before free */
720 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
721 struct list_head gnc_hashlist; /* stash in connection hash table */
722 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
723 struct list_head gnc_fmaq; /* txs queued for FMA */
724 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
725 struct list_head gnc_delaylist; /* If on this list schedule anytime we get interrupted */
726 __u64 gnc_peerstamp; /* peer's unique stamp */
727 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
728 __u64 gnc_my_connstamp; /* my unique connection stamp */
729 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
730 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
731 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
732 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
733 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
734 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
735 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
736 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
737 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
738 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
739 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
740 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
741 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
742 __u32 gnc_cqid; /* my completion callback id (non-unique) */
743 atomic_t gnc_tx_seq; /* tx msg sequence number */
744 atomic_t gnc_rx_seq; /* rx msg sequence number */
745 struct mutex gnc_smsg_mutex; /* tx smsg sequence serialization */
746 struct mutex gnc_rdma_mutex; /* tx rdma sequence serialization */
747 __u64 gnc_tx_retrans; /* # retrans on SMSG */
748 atomic_t gnc_nlive_fma; /* # live FMA */
749 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
750 atomic_t gnc_nlive_rdma; /* # live RDMA */
751 short gnc_close_sent; /* I've sent CLOSE */
752 short gnc_close_recvd; /* I've received CLOSE */
753 short gnc_in_purgatory; /* in the sin bin */
754 int gnc_error; /* errno when conn being closed due to error */
755 int gnc_peer_error; /* errno peer sent us on CLOSE */
756 kgn_conn_state_t gnc_state; /* connection state */
757 int gnc_scheduled; /* being attented to */
758 char gnc_sched_caller[30]; /* what function last called schedule */
759 int gnc_sched_line; /* what line # last called schedule */
760 atomic_t gnc_refcount; /* # users */
761 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
762 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
763 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
764 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
765 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
766 unsigned long gnc_tx_bits[(GNILND_MAX_MSG_ID/8)/sizeof(unsigned long)]; /* bit table for tx id */
767 int gnc_next_tx; /* next tx to use in tx_ref_table */
768 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
769 int gnc_mbox_id; /* id of mbox in fma_blk */
770 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
771 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
772 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
773 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
774 void *remote_mbox_addr; /* save remote mbox address */
777 typedef struct kgn_mdd_purgatory {
778 gni_mem_handle_t gmp_map_key; /* mapping key */
779 struct list_head gmp_list; /* entry point for purgatory list */
780 } kgn_mdd_purgatory_t;
782 typedef struct kgn_peer {
783 struct list_head gnp_list; /* stash on global peer list */
784 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
785 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
786 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
787 kgn_net_t *gnp_net; /* net instance for this peer */
788 lnet_nid_t gnp_nid; /* who's on the other end(s) */
789 atomic_t gnp_refcount; /* # users */
790 __u32 gnp_host_id; /* ph. host ID of the peer */
791 short gnp_connecting; /* connection forming */
792 short gnp_pending_unlink; /* need last conn close to trigger unlink */
793 int gnp_last_errno; /* last error conn saw */
794 time64_t gnp_last_alive; /* last time I had valid comms */
795 int gnp_last_dgram_errno; /* last error dgrams saw */
796 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
797 unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
798 unsigned long gnp_reconnect_interval; /* exponential backoff */
799 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
800 int gnp_state; /* up/down/timedout */
801 unsigned long gnp_down_event_time; /* time peer down */
802 unsigned long gnp_up_event_time; /* time peer back up */
805 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
806 * like lnet_parse. It allows a single pointer to let us get enough
807 * information in _recv and friends */
808 typedef struct kgn_rx {
809 kgn_conn_t *grx_conn; /* connection */
810 kgn_msg_t *grx_msg; /* message */
811 struct lnet_msg *grx_lntmsg; /* lnet msg for this rx (eager only) */
812 int grx_eager; /* if eager, we copied msg to somewhere */
813 struct timespec64 grx_received; /* time this msg received */
816 typedef struct kgn_data {
817 int kgn_init; /* initialisation state */
818 int kgn_shutdown; /* shut down? */
819 int kgn_wc_kill; /* Should I repost the WC */
820 atomic_t kgn_nthreads; /* # live threads */
821 int kgn_nresets; /* number of stack resets */
822 int kgn_in_reset; /* are we in stack reset ? */
824 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
826 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
827 int kgn_ndevs; /* # devices */
829 int kgn_ruhroh_running; /* ruhroh thread is running */
830 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
831 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
832 int kgn_quiesce_trigger; /* should we quiesce ? */
833 atomic_t kgn_nquiesce; /* how many quiesced ? */
834 struct mutex kgn_quiesce_mutex; /* serialize ruhroh task, startup and shutdown */
835 int kgn_needs_reset; /* we need stack reset */
837 /* These next three members implement communication from gnilnd into
838 * the ruhroh task. To ensure correct operation of the task, code that
839 * writes into them must use memory barriers to ensure that the changes
840 * are visible to other cores in the order the members appear below. */
841 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
842 int kgn_bump_info_rdy; /* we have info needed to bump */
843 int kgn_needs_pause; /* we need to pause for network quiesce */
845 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
846 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
848 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
849 struct list_head *kgn_peers; /* hash table of all my known peers */
850 atomic_t kgn_npeers; /* # peers extant */
851 int kgn_peer_version; /* version flag for peer tables */
853 struct list_head *kgn_conns; /* conns hashed by cqid */
854 atomic_t kgn_nconns; /* # connections extant */
855 atomic_t kgn_neager_allocs; /* # of eager allocations */
856 __u64 kgn_peerstamp; /* when I started up */
857 __u64 kgn_connstamp; /* conn stamp generator */
858 int kgn_conn_version; /* version flag for conn tables */
859 int kgn_next_cqid; /* cqid generator */
861 long kgn_new_min_timeout; /* minimum timeout on any new conn */
862 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
863 spinlock_t kgn_reaper_lock; /* serialise */
865 struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
866 struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
867 struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
868 atomic_t kgn_ntx; /* # tx in use */
869 struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
871 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
872 __u64 kgn_cksum_npages; /* # pages alloc'd for checksumming */
873 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
874 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
876 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
877 long kgn_rdmaq_override; /* bytes per second override */
879 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
881 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
882 atomic_t kgn_npending_conns; /* # of conns with pending closes */
883 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
884 unsigned long kgn_last_scheduled; /* last time schedule was called */
885 unsigned long kgn_last_condresched; /* last time cond_resched was called */
886 atomic_t kgn_rev_offset; /* # of REV rdma w/misaligned offsets */
887 atomic_t kgn_rev_length; /* # of REV rdma have misaligned len */
888 atomic_t kgn_rev_copy_buff; /* # of REV rdma buffer copies */
889 unsigned long free_pages_limit; /* # of free pages reserve from fma block allocations */
890 int kgn_enable_gl_mutex; /* kgni api mtx enable */
893 extern kgn_data_t kgnilnd_data;
894 extern kgn_tunables_t kgnilnd_tunables;
896 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
897 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
898 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held);
899 extern int _kgnilnd_schedule_delay_conn(kgn_conn_t *conn);
901 static inline int kgnilnd_timeout(void)
903 return *kgnilnd_tunables.kgn_timeout ?
904 *kgnilnd_tunables.kgn_timeout :
905 lnet_get_lnd_timeout();
908 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
909 * and the line of the calling function to allow us to debug problematic
910 * schedule calls in the future without the programmer having to mark
911 * the location manually.
913 #define kgnilnd_schedule_conn(conn) \
914 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 0);
916 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
917 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld, 0);
919 #define kgnilnd_schedule_conn_nolock(conn) \
920 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 1);
923 /* Macro wrapper for _kgnilnd_schedule_delay_conn. This will allow us to store
924 * extra data if we need to.
926 #define kgnilnd_schedule_delay_conn(conn) \
927 _kgnilnd_schedule_delay_conn(conn);
930 kgnilnd_thread_fini(void)
932 atomic_dec(&kgnilnd_data.kgn_nthreads);
935 static inline void kgnilnd_gl_mutex_lock(struct mutex *lock)
937 if (kgnilnd_data.kgn_enable_gl_mutex)
941 static inline void kgnilnd_gl_mutex_unlock(struct mutex *lock)
943 if (kgnilnd_data.kgn_enable_gl_mutex)
947 static inline void kgnilnd_conn_mutex_lock(struct mutex *lock)
949 if (!kgnilnd_data.kgn_enable_gl_mutex)
953 static inline void kgnilnd_conn_mutex_unlock(struct mutex *lock)
955 if (!kgnilnd_data.kgn_enable_gl_mutex)
959 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
960 * parts of the code to avoid a scheduler trip when the mutex is held
962 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
963 * has been acquired successfully, and 0 on contention.
965 * NOTE: this function follows the spin_trylock() convention, so
966 * it is negated to the down_trylock() return values! Be careful
967 * about this when converting semaphore users to mutexes.
969 * This function must not be used in interrupt context. The
970 * mutex must be released by the same task that acquired it.
972 static inline int __kgnilnd_mutex_trylock(struct mutex *lock)
975 unsigned long timeout;
977 LASSERT(!in_interrupt());
979 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
981 ret = mutex_trylock(lock);
988 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
990 if (!kgnilnd_data.kgn_enable_gl_mutex)
993 return __kgnilnd_mutex_trylock(lock);
996 static inline int kgnilnd_trylock(struct mutex *cq_lock,
997 struct mutex *c_lock)
999 if (kgnilnd_data.kgn_enable_gl_mutex)
1000 return __kgnilnd_mutex_trylock(cq_lock);
1002 return __kgnilnd_mutex_trylock(c_lock);
1005 static inline void *kgnilnd_vzalloc(int size)
1008 if (*kgnilnd_tunables.kgn_vzalloc_noretry)
1009 ret = __ll_vmalloc(size, __GFP_HIGHMEM | GFP_NOIO | __GFP_ZERO |
1012 ret = __ll_vmalloc(size, __GFP_HIGHMEM | GFP_NOIO | __GFP_ZERO);
1014 LIBCFS_ALLOC_POST(ret, size, "alloc");
1018 static inline void kgnilnd_vfree(void *ptr, int size)
1020 LIBCFS_FREE_PRE(ptr, size, "vfree");
1024 /* as of kernel version 4.2, set_mb is replaced with smp_store_mb */
1026 #define set_mb smp_store_mb
1029 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
1032 _kgnilnd_debug_msg(kgn_msg_t *msg,
1033 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1035 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
1037 CFS_CHECK_STACK(msgdata, mask, cdls); \
1039 if (((mask) & D_CANTMASK) != 0 || \
1040 ((libcfs_debug & (mask)) != 0 && \
1041 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1042 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
1045 /* for most callers (level is a constant) this is resolved at compile time */
1046 #define GNIDBG_MSG(level, msg, fmt, args...) \
1048 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1049 static struct cfs_debug_limit_state cdls; \
1050 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1051 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1052 "$$ "fmt" from %s ", ## args, \
1053 libcfs_nid2str((msg)->gnm_srcnid)); \
1055 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1056 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1057 "$$ "fmt" from %s ", ## args, \
1058 libcfs_nid2str((msg)->gnm_srcnid)); \
1062 /* user puts 'to nid' in msg for us */
1063 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
1065 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1066 static struct cfs_debug_limit_state cdls; \
1067 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1068 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1069 "$$ "fmt" ", ## args); \
1071 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1072 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1073 "$$ "fmt" ", ## args); \
1078 _kgnilnd_debug_conn(kgn_conn_t *conn,
1079 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1081 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
1083 CFS_CHECK_STACK(msgdata, mask, cdls); \
1085 if (((mask) & D_CANTMASK) != 0 || \
1086 ((libcfs_debug & (mask)) != 0 && \
1087 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1088 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
1091 /* for most callers (level is a constant) this is resolved at compile time */
1092 #define GNIDBG_CONN(level, conn, fmt, args...) \
1094 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1095 static struct cfs_debug_limit_state cdls; \
1096 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1097 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
1098 "$$ "fmt" ", ## args); \
1100 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1101 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
1102 "$$ "fmt" ", ## args); \
1107 _kgnilnd_debug_tx(kgn_tx_t *tx,
1108 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1110 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
1112 CFS_CHECK_STACK(msgdata, mask, cdls); \
1114 if (((mask) & D_CANTMASK) != 0 || \
1115 ((libcfs_debug & (mask)) != 0 && \
1116 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1117 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1120 /* for most callers (level is a constant) this is resolved at compile time */
1121 #define GNIDBG_TX(level, tx, fmt, args...) \
1123 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1124 static struct cfs_debug_limit_state cdls; \
1125 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1126 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1127 "$$ "fmt" ", ## args); \
1129 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1130 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1131 "$$ "fmt" ", ## args); \
1135 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1137 if (unlikely(!(cond))) { \
1138 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1143 #define GNILND_IS_QUIESCED \
1144 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1145 atomic_read(&kgnilnd_data.kgn_nthreads))
1147 #define KGNILND_SPIN_QUIESCE \
1149 /* E.T phone home */ \
1150 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1151 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1152 while (kgnilnd_data.kgn_quiesce_trigger) { \
1153 msleep_interruptible(MSEC_PER_SEC); \
1155 /* Mom, my homework is done */ \
1156 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1157 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1160 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1161 #ifndef LIBCFS_DEBUG
1162 #error "this code uses actions inside LASSERT for ref counting"
1165 #define kgnilnd_admin_addref(atomic) \
1167 int val = atomic_inc_return(&atomic); \
1168 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1169 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1172 #define kgnilnd_admin_decref(atomic) \
1174 int val = atomic_dec_return(&atomic); \
1175 LASSERTF(val >= 0, #atomic " refcount %d\n", val); \
1176 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1178 wake_up_var(&kgnilnd_data); \
1181 #define kgnilnd_net_addref(net) \
1183 int val = atomic_inc_return(&net->gnn_refcount); \
1184 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1185 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1186 libcfs_nidstr(&net->gnn_ni->ni_nid), val); \
1189 #define kgnilnd_net_decref(net) \
1191 int val = atomic_dec_return(&net->gnn_refcount); \
1192 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1193 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1194 libcfs_nidstr(&net->gnn_ni->ni_nid), val); \
1197 #define kgnilnd_peer_addref(peer) \
1199 int val = atomic_inc_return(&peer->gnp_refcount); \
1200 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1201 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1202 libcfs_nid2str(peer->gnp_nid), val); \
1205 #define kgnilnd_peer_decref(peer) \
1207 int val = atomic_dec_return(&peer->gnp_refcount); \
1208 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1209 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1210 libcfs_nid2str(peer->gnp_nid), val); \
1212 kgnilnd_destroy_peer(peer); \
1215 #define kgnilnd_conn_addref(conn) \
1220 val = atomic_inc_return(&conn->gnc_refcount); \
1221 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1222 "conn %p refc %d to %s\n", \
1225 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1227 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1229 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1234 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1235 * through the scheduler thread to get the EP destroyed. This avoids some
1236 * messy semaphore business and allows us to reuse the connd_list and existing
1237 * linkage and avoid creating extra lists just for destroying EPs */
1239 /* Safety Disclaimer:
1240 * Q: If we decrement the refcount and then check it again, is it possible that
1241 * another caller could have passed through this macro concurrently? If so,
1242 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1244 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1245 * once we are refcount <= 2. It hinges around gnc_state and membership of
1246 * gnc_hashlist. There are two ways to find a connection - either ask for
1247 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1248 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1251 * - #1 from create (kgnilnd_create_conn)
1252 * - #2 for EP (kgnilnd_create_conn)
1253 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1254 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1256 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1257 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1259 * There could be more from TX descriptors during the lifetime of a live
1262 * If we nuke the conn before finish_connect, we won't have parallel paths
1263 * because nobody besides the dgram handler for the single outstanding
1264 * dgram can find the connection as it isn't in any searchable tables yet.
1266 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1267 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1268 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1269 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1270 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1271 * sent. Now the gnc_state == CLOSED, so we hit
1272 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1273 * - we've nuked them from the peer and CQ id tables, so we own them and
1274 * are guaranteed serial access - hence the complete lack of conn list
1275 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1276 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1277 * kgnilnd_process_conns. Then the next scheduler loop would call
1278 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1279 * kgnilnd_process_conns.
1281 * Clearly, we are totally safe. Clearly.
1284 #define kgnilnd_conn_decref(conn) \
1289 val = atomic_dec_return(&conn->gnc_refcount); \
1290 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1293 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1295 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1297 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1302 (conn->gnc_ephandle != NULL) && \
1303 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1304 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1305 kgnilnd_schedule_conn(conn); \
1306 } else if (val == 0) { \
1307 kgnilnd_destroy_conn(conn); \
1311 static inline struct list_head *
1312 kgnilnd_nid2peerlist(lnet_nid_t nid)
1314 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1316 RETURN(&kgnilnd_data.kgn_peers[hash]);
1319 static inline struct list_head *
1320 kgnilnd_netnum2netlist(__u16 netnum)
1322 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1324 RETURN(&kgnilnd_data.kgn_nets[hash]);
1328 kgnilnd_peer_active(kgn_peer_t *peer)
1330 /* Am I in the peer hash table? */
1331 return (!list_empty(&peer->gnp_list));
1334 /* need write_lock on kgn_peer_conn_lock */
1336 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1338 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1339 peer, libcfs_nid2str(peer->gnp_nid),
1340 !list_empty(&peer->gnp_conns),
1341 !list_empty(&peer->gnp_tx_queue));
1343 /* kgn_peer_conn_lock protects us from conflict with
1344 * kgnilnd_peer_notify and gnp_persistent */
1345 RETURN ((list_empty(&peer->gnp_conns)) &&
1346 (list_empty(&peer->gnp_tx_queue)));
1349 /* returns positive if error was for a clean shutdown of conn */
1351 kgnilnd_conn_clean_errno(int errno)
1353 /* - ESHUTDOWN - LND is unloading
1354 * - EUCLEAN - admin requested via "lctl del_peer"
1355 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1356 * - ENOTRECOVERABLE - stack reset
1357 * - EISCONN - cleared via "lctl push"
1358 * not doing ESTALE - that isn't clean */
1359 RETURN ((errno == 0) ||
1360 (errno == -ESHUTDOWN) ||
1361 (errno == -EUCLEAN) ||
1362 (errno == -ENETRESET) ||
1363 (errno == -EISCONN) ||
1364 (errno == -ENOTRECOVERABLE));
1367 /* returns positive if error results in purgatory hold */
1369 kgnilnd_check_purgatory_errno(int errno)
1371 /* We don't want to save the purgatory lists these cases:
1372 * - EUCLEAN - admin requested via "lctl del_peer"
1373 * - ESHUTDOWN - LND is unloading
1375 RETURN ((errno != -ESHUTDOWN) &&
1376 (errno != -EUCLEAN));
1380 /* returns positive if a purgatory hold is needed */
1382 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1386 if (conn->gnc_peer) {
1387 loopback = conn->gnc_peer->gnp_nid ==
1388 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
1390 /* short circuit - a conn that didn't complete
1391 * setup never needs a purgatory hold */
1394 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1395 conn, conn->gnc_peer ?
1396 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1398 loopback, conn->gnc_close_recvd);
1400 /* we only use a purgatory hold if we've not received the CLOSE msg
1401 * from our peer - without that message, we can't know the state of
1402 * the other end of this connection and must put it into purgatory
1403 * to prevent reuse and corruption.
1404 * The theory is that a TX error can be communicated in all other cases
1406 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1407 kgnilnd_check_purgatory_errno(conn->gnc_error));
1410 static inline const char *
1411 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1413 static inline struct list_head *
1414 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1415 kgn_tx_list_state_t to_state)
1418 case GNILND_TX_PEERQ:
1419 return &peer->gnp_tx_queue;
1420 case GNILND_TX_FMAQ:
1421 return &conn->gnc_fmaq;
1422 case GNILND_TX_LIVE_FMAQ:
1423 case GNILND_TX_LIVE_RDMAQ:
1424 case GNILND_TX_DYING:
1426 case GNILND_TX_MAPQ:
1427 return &conn->gnc_device->gnd_map_tx;
1428 case GNILND_TX_RDMAQ:
1429 return &conn->gnc_device->gnd_rdmaq;
1431 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1432 CERROR("invalid state requested: %s\n",
1433 kgnilnd_tx_state2str(to_state));
1439 /* should hold tx, conn or peer lock when calling */
1441 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1442 kgn_conn_t *conn, kgn_tx_list_state_t state,
1445 struct list_head *list = NULL;
1447 /* make sure we have a sane TX state to start */
1448 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1449 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1450 list_empty(&tx->tx_list),
1451 "bad state with tx_list %s",
1452 list_empty(&tx->tx_list) ? "empty" : "not empty");
1454 /* WTF - you are already on that state buttmunch */
1455 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1456 "already at %s", kgnilnd_tx_state2str(state));
1458 /* get proper list from the state requested */
1459 list = kgnilnd_tx_state2list(peer, conn, state);
1463 case GNILND_TX_PEERQ:
1464 kgnilnd_peer_addref(peer);
1466 case GNILND_TX_ALLOCD:
1467 /* no refs needed */
1469 case GNILND_TX_FMAQ:
1470 kgnilnd_conn_addref(conn);
1472 case GNILND_TX_MAPQ:
1473 atomic_inc(&conn->gnc_device->gnd_nq_map);
1474 kgnilnd_conn_addref(conn);
1476 case GNILND_TX_LIVE_FMAQ:
1477 atomic_inc(&conn->gnc_nlive_fma);
1478 kgnilnd_conn_addref(conn);
1480 case GNILND_TX_LIVE_RDMAQ:
1481 atomic_inc(&conn->gnc_nlive_rdma);
1482 kgnilnd_conn_addref(conn);
1484 case GNILND_TX_RDMAQ:
1485 atomic_inc(&conn->gnc_nq_rdma);
1486 kgnilnd_conn_addref(conn);
1488 case GNILND_TX_DYING:
1489 kgnilnd_conn_addref(conn);
1492 CERROR("invalid state requested: %s\n",
1493 kgnilnd_tx_state2str(state));
1498 /* if this changes, change kgnilnd_alloc_tx */
1499 tx->tx_list_state = state;
1501 /* some states don't have lists - we track them in the per conn
1502 * TX table instead. Waste not, want not! */
1504 tx->tx_list_p = list;
1506 list_add_tail(&tx->tx_list, list);
1508 list_add(&tx->tx_list, list);
1510 /* set dummy list_p to make book keeping happy and let debugging
1511 * be a hair easier */
1512 tx->tx_list_p = (void *)state;
1515 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1516 kgnilnd_tx_state2str(state), list);
1520 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1521 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1523 /* These is only 1 "off-list" state */
1524 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1525 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1527 /* new_state == ALLOCD means we are deallocating this tx,
1528 * so make sure it was on a valid list to start with */
1529 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1530 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1531 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1532 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1535 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1538 switch (tx->tx_list_state) {
1539 case GNILND_TX_PEERQ:
1540 kgnilnd_peer_decref(peer);
1542 case GNILND_TX_FREED:
1543 case GNILND_TX_IDLE:
1544 case GNILND_TX_ALLOCD:
1545 /* no refs needed */
1547 case GNILND_TX_DYING:
1548 kgnilnd_conn_decref(conn);
1550 case GNILND_TX_FMAQ:
1551 kgnilnd_conn_decref(conn);
1553 case GNILND_TX_MAPQ:
1554 atomic_dec(&conn->gnc_device->gnd_nq_map);
1555 kgnilnd_conn_decref(conn);
1557 case GNILND_TX_LIVE_FMAQ:
1558 atomic_dec(&conn->gnc_nlive_fma);
1559 kgnilnd_conn_decref(conn);
1561 case GNILND_TX_LIVE_RDMAQ:
1562 atomic_dec(&conn->gnc_nlive_rdma);
1563 kgnilnd_conn_decref(conn);
1565 case GNILND_TX_RDMAQ:
1566 atomic_dec(&conn->gnc_nq_rdma);
1567 kgnilnd_conn_decref(conn);
1568 /* don't need to assert on default, already did in set */
1571 /* for ALLOCD, this might already be true, but no harm doing it again */
1572 list_del_init(&tx->tx_list);
1573 tx->tx_list_p = NULL;
1574 tx->tx_list_state = new_state;
1578 kgnilnd_tx_mapped(kgn_tx_t *tx)
1580 return tx->tx_buftype == GNILND_BUF_PHYS_MAPPED;
1583 static inline struct list_head *
1584 kgnilnd_cqid2connlist(__u32 cqid)
1586 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1588 return (&kgnilnd_data.kgn_conns [hash]);
1591 static inline kgn_conn_t *
1592 kgnilnd_cqid2conn_locked(__u32 cqid)
1594 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1595 struct list_head *tmp;
1598 list_for_each(tmp, conns) {
1599 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1601 if (conn->gnc_cqid == cqid)
1608 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1610 kgnilnd_get_cqid_locked(void)
1616 cqid = kgnilnd_data.kgn_next_cqid++;
1617 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1621 kgnilnd_data.kgn_next_cqid = 1;
1624 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1630 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1632 kgn_tx_t *tx = NULL;
1633 kgn_conn_t *conn = NULL;
1635 /* set to NULL so any early return is an error */
1639 LASSERTF((ev_id->txe_idx > 0) &&
1640 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1641 "bogus txe_idx %d >= %d\n",
1642 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1644 LASSERTF((ev_id->txe_cqid > 0) &&
1645 (ev_id->txe_cqid < GNILND_MAX_CQID),
1646 "bogus txe_cqid %d >= %d\n",
1647 ev_id->txe_cqid, GNILND_MAX_CQID);
1649 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1650 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1653 /* Conn was destroyed? */
1654 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1655 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1658 /* just insurance */
1659 kgnilnd_conn_addref(conn);
1660 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1661 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1663 /* we know this is safe - as the TX won't be reused until AFTER
1664 * the conn is unlinked from the cqid hash, so we can use the TX
1665 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1667 spin_lock(&conn->gnc_tx_lock);
1668 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1669 spin_unlock(&conn->gnc_tx_lock);
1671 /* We could have a tx that was cleared out by other forces
1672 * lctl disconnect or del_peer. */
1674 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1675 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1676 kgnilnd_conn_decref(conn);
1680 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1681 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1682 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1684 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1685 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1686 "found tx %p txe_idx %d",
1687 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1688 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1690 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1692 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1697 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1700 /* set_normalized_timepsec isn't exported from the kernel, so
1701 * we need to do the same thing inline */
1702 static inline struct timespec
1703 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1709 sec = lhs.tv_sec - rhs.tv_sec;
1710 nsec = lhs.tv_nsec - rhs.tv_nsec;
1712 while (nsec >= NSEC_PER_SEC) {
1713 nsec -= NSEC_PER_SEC;
1717 nsec += NSEC_PER_SEC;
1726 kgnilnd_count_list(struct list_head *q)
1728 struct list_head *e;
1731 list_for_each(e, q) {
1738 /* kgnilnd_find_net adds a reference to the net it finds
1739 * this is so the net will not be removed before the calling function
1740 * has time to use the data returned. This reference needs to be released
1741 * by the calling function once it has finished using the returned net
1745 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1750 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1756 list_for_each_entry(net,
1757 kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))),
1759 if (!net->gnn_shutdown &&
1760 LNET_NID_NET(&net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1761 kgnilnd_net_addref(net);
1762 up_read(&kgnilnd_data.kgn_net_rw_sem);
1768 up_read(&kgnilnd_data.kgn_net_rw_sem);
1773 #ifdef CONFIG_DEBUG_SLAB
1774 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1776 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1779 int kgnilnd_dev_init(kgn_device_t *dev);
1780 void kgnilnd_dev_fini(kgn_device_t *dev);
1781 int kgnilnd_startup(struct lnet_ni *ni);
1782 void kgnilnd_shutdown(struct lnet_ni *ni);
1783 int kgnilnd_base_startup(void);
1784 void kgnilnd_base_shutdown(void);
1786 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1787 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1788 void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
1789 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1791 int kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg);
1792 int kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1793 int kgnilnd_eager_recv(struct lnet_ni *ni, void *private,
1794 struct lnet_msg *lntmsg, void **new_private);
1795 int kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1796 int delayed, unsigned int niov,
1797 struct bio_vec *kiov,
1798 unsigned int offset, unsigned int mlen, unsigned int rlen);
1800 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
1801 unsigned int offset, unsigned int nob, int dump_blob);
1803 /* purgatory functions */
1804 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1805 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1806 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1807 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1809 void kgnilnd_update_reaper_timeout(long timeout);
1810 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1811 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1812 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1813 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1814 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1815 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held);
1816 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1818 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1819 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
1820 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1821 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1823 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1824 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1825 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1826 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1827 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1828 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net,
1829 struct lnet_processid *target);
1830 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1831 void kgnilnd_consume_rx(kgn_rx_t *rx);
1833 void kgnilnd_schedule_device(kgn_device_t *dev);
1834 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1835 void kgnilnd_schedule_device_timer(cfs_timer_cb_arg_t data);
1836 void kgnilnd_schedule_device_timer_rd(cfs_timer_cb_arg_t data);
1838 int kgnilnd_reaper(void *arg);
1839 int kgnilnd_scheduler(void *arg);
1840 int kgnilnd_dgram_mover(void *arg);
1841 int kgnilnd_rca(void *arg);
1842 int kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id);
1844 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1845 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1846 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1847 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1848 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1849 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1850 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1851 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1852 void kgnilnd_peer_alive(kgn_peer_t *peer);
1853 void kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive);
1854 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1855 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1856 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1857 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1859 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1860 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1861 void kgnilnd_wakeup_rca_thread(void);
1862 int kgnilnd_start_rca_thread(void);
1863 int kgnilnd_get_node_state(__u32 nid);
1865 int kgnilnd_tunables_init(void);
1866 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1868 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1869 void kgnilnd_pause_threads(void);
1870 int kgnilnd_hw_in_quiesce(void);
1871 int kgnilnd_check_hw_quiesce(void);
1872 void kgnilnd_quiesce_wait(char *reason);
1873 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1874 int kgnilnd_ruhroh_thread(void *arg);
1875 void kgnilnd_reset_stack(void);
1876 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1878 void kgnilnd_insert_sysctl(void);
1879 void kgnilnd_remove_sysctl(void);
1880 void kgnilnd_proc_init(void);
1881 void kgnilnd_proc_fini(void);
1884 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1886 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1887 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1888 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
1890 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1891 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1892 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1893 int kgnilnd_cancel_dgrams(kgn_device_t *dev);
1894 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1896 int kgnilnd_dgram_waitq(void *arg);
1898 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1900 /* struct2str functions - we don't use a default: case to cause the compile
1901 * to fail if there is a missing case. This allows us to hide these down here
1902 * out of the way but ensure we'll catch any updates to the enum/types
1905 #define DO_TYPE(x) case x: return #x;
1906 static inline const char *
1907 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1909 /* Only want single char string for this */
1911 case GNILND_FMABLK_IDLE:
1913 case GNILND_FMABLK_PHYS:
1915 case GNILND_FMABLK_VIRT:
1917 case GNILND_FMABLK_FREED:
1920 return "<unknown state>";
1923 static inline const char *
1924 kgnilnd_msgtype2str(int type)
1927 DO_TYPE(GNILND_MSG_NONE);
1928 DO_TYPE(GNILND_MSG_NOOP);
1929 DO_TYPE(GNILND_MSG_IMMEDIATE);
1930 DO_TYPE(GNILND_MSG_PUT_REQ);
1931 DO_TYPE(GNILND_MSG_PUT_NAK);
1932 DO_TYPE(GNILND_MSG_PUT_ACK);
1933 DO_TYPE(GNILND_MSG_PUT_DONE);
1934 DO_TYPE(GNILND_MSG_GET_REQ);
1935 DO_TYPE(GNILND_MSG_GET_NAK);
1936 DO_TYPE(GNILND_MSG_GET_DONE);
1937 DO_TYPE(GNILND_MSG_CLOSE);
1938 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1939 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1940 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1941 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1942 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1943 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1944 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1946 return "<unknown msg type>";
1949 static inline const char *
1950 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1953 DO_TYPE(GNILND_TX_IDLE);
1954 DO_TYPE(GNILND_TX_ALLOCD);
1955 DO_TYPE(GNILND_TX_PEERQ);
1956 DO_TYPE(GNILND_TX_MAPQ);
1957 DO_TYPE(GNILND_TX_FMAQ);
1958 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1959 DO_TYPE(GNILND_TX_RDMAQ);
1960 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1961 DO_TYPE(GNILND_TX_DYING);
1962 DO_TYPE(GNILND_TX_FREED);
1964 return "<unknown state>";
1967 static inline const char *
1968 kgnilnd_conn_state2str(kgn_conn_t *conn)
1970 kgn_conn_state_t state = conn->gnc_state;
1972 DO_TYPE(GNILND_CONN_DUMMY);
1973 DO_TYPE(GNILND_CONN_LISTEN);
1974 DO_TYPE(GNILND_CONN_CONNECTING);
1975 DO_TYPE(GNILND_CONN_ESTABLISHED);
1976 DO_TYPE(GNILND_CONN_CLOSING);
1977 DO_TYPE(GNILND_CONN_CLOSED);
1978 DO_TYPE(GNILND_CONN_DONE);
1979 DO_TYPE(GNILND_CONN_DESTROY_EP);
1984 static inline const char *
1985 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1987 kgn_connreq_type_t type = connreq->gncr_type;
1990 DO_TYPE(GNILND_CONNREQ_REQ);
1991 DO_TYPE(GNILND_CONNREQ_NAK);
1992 DO_TYPE(GNILND_CONNREQ_CLOSE);
1997 static inline const char *
1998 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
2000 kgn_dgram_state_t state = dgram->gndg_state;
2003 DO_TYPE(GNILND_DGRAM_USED);
2004 DO_TYPE(GNILND_DGRAM_POSTING);
2005 DO_TYPE(GNILND_DGRAM_POSTED);
2006 DO_TYPE(GNILND_DGRAM_PROCESSING);
2007 DO_TYPE(GNILND_DGRAM_DONE);
2008 DO_TYPE(GNILND_DGRAM_CANCELED);
2013 static inline const char *
2014 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
2016 kgn_dgram_type_t type = dgram->gndg_type;
2019 DO_TYPE(GNILND_DGRAM_REQ);
2020 DO_TYPE(GNILND_DGRAM_WC_REQ);
2021 DO_TYPE(GNILND_DGRAM_NAK);
2022 DO_TYPE(GNILND_DGRAM_CLOSE);
2027 static inline const char *
2028 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
2031 DO_TYPE(GNILND_DGRAM_REQ);
2032 DO_TYPE(GNILND_DGRAM_WC_REQ);
2033 DO_TYPE(GNILND_DGRAM_NAK);
2034 DO_TYPE(GNILND_DGRAM_CLOSE);
2041 /* pulls in tunables per platform and adds in nid/nic conversion
2042 * if RCA wasn't available at build time */
2043 #include "gnilnd_hss_ops.h"
2044 /* API wrapper functions - include late to pick up all of the other defines */
2045 #include "gnilnd_api_wrap.h"
2047 #if defined(CONFIG_CRAY_GEMINI)
2048 #include "gnilnd_gemini.h"
2049 #elif defined(CONFIG_CRAY_ARIES)
2050 #include "gnilnd_aries.h"
2052 #error "Undefined Network Hardware Type"
2055 extern uint32_t kgni_driver_version;
2058 kgnilnd_check_kgni_version(void)
2062 kgnilnd_data.kgn_enable_gl_mutex = 1;
2063 kdv = symbol_get(kgni_driver_version);
2065 LCONSOLE_INFO("Not using thread safe locking -"
2066 " no symbol kgni_driver_version\n");
2070 /* Thread-safe kgni implemented in minor ver 0x44/45, code rev 0xb9 */
2071 if (*kdv < GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9)) {
2072 symbol_put(kgni_driver_version);
2073 LCONSOLE_INFO("Not using thread safe locking, gni version 0x%x,"
2074 " need >= 0x%x\n", *kdv,
2075 GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9));
2079 symbol_put(kgni_driver_version);
2081 if (!*kgnilnd_tunables.kgn_thread_safe) {
2085 /* Use thread-safe locking */
2086 kgnilnd_data.kgn_enable_gl_mutex = 0;
2089 #endif /* _GNILND_GNILND_H_ */