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 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
902 * and the line of the calling function to allow us to debug problematic
903 * schedule calls in the future without the programmer having to mark
904 * the location manually.
906 #define kgnilnd_schedule_conn(conn) \
907 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 0);
909 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
910 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld, 0);
912 #define kgnilnd_schedule_conn_nolock(conn) \
913 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 1);
916 /* Macro wrapper for _kgnilnd_schedule_delay_conn. This will allow us to store
917 * extra data if we need to.
919 #define kgnilnd_schedule_delay_conn(conn) \
920 _kgnilnd_schedule_delay_conn(conn);
923 kgnilnd_thread_fini(void)
925 atomic_dec(&kgnilnd_data.kgn_nthreads);
928 static inline void kgnilnd_gl_mutex_lock(struct mutex *lock)
930 if (kgnilnd_data.kgn_enable_gl_mutex)
934 static inline void kgnilnd_gl_mutex_unlock(struct mutex *lock)
936 if (kgnilnd_data.kgn_enable_gl_mutex)
940 static inline void kgnilnd_conn_mutex_lock(struct mutex *lock)
942 if (!kgnilnd_data.kgn_enable_gl_mutex)
946 static inline void kgnilnd_conn_mutex_unlock(struct mutex *lock)
948 if (!kgnilnd_data.kgn_enable_gl_mutex)
952 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
953 * parts of the code to avoid a scheduler trip when the mutex is held
955 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
956 * has been acquired successfully, and 0 on contention.
958 * NOTE: this function follows the spin_trylock() convention, so
959 * it is negated to the down_trylock() return values! Be careful
960 * about this when converting semaphore users to mutexes.
962 * This function must not be used in interrupt context. The
963 * mutex must be released by the same task that acquired it.
965 static inline int __kgnilnd_mutex_trylock(struct mutex *lock)
968 unsigned long timeout;
970 LASSERT(!in_interrupt());
972 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
974 ret = mutex_trylock(lock);
981 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
983 if (!kgnilnd_data.kgn_enable_gl_mutex)
986 return __kgnilnd_mutex_trylock(lock);
989 static inline int kgnilnd_trylock(struct mutex *cq_lock,
990 struct mutex *c_lock)
992 if (kgnilnd_data.kgn_enable_gl_mutex)
993 return __kgnilnd_mutex_trylock(cq_lock);
995 return __kgnilnd_mutex_trylock(c_lock);
998 static inline void *kgnilnd_vzalloc(int size)
1001 if (*kgnilnd_tunables.kgn_vzalloc_noretry)
1002 ret = __ll_vmalloc(size, __GFP_HIGHMEM | GFP_NOIO | __GFP_ZERO |
1005 ret = __ll_vmalloc(size, __GFP_HIGHMEM | GFP_NOIO | __GFP_ZERO);
1007 LIBCFS_ALLOC_POST(ret, size);
1011 static inline void kgnilnd_vfree(void *ptr, int size)
1013 libcfs_kmem_dec(ptr, size);
1017 /* as of kernel version 4.2, set_mb is replaced with smp_store_mb */
1019 #define set_mb smp_store_mb
1022 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
1025 _kgnilnd_debug_msg(kgn_msg_t *msg,
1026 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1028 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
1030 CFS_CHECK_STACK(msgdata, mask, cdls); \
1032 if (((mask) & D_CANTMASK) != 0 || \
1033 ((libcfs_debug & (mask)) != 0 && \
1034 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1035 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
1038 /* for most callers (level is a constant) this is resolved at compile time */
1039 #define GNIDBG_MSG(level, msg, fmt, args...) \
1041 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1042 static struct cfs_debug_limit_state cdls; \
1043 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1044 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1045 "$$ "fmt" from %s ", ## args, \
1046 libcfs_nid2str((msg)->gnm_srcnid)); \
1048 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1049 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1050 "$$ "fmt" from %s ", ## args, \
1051 libcfs_nid2str((msg)->gnm_srcnid)); \
1055 /* user puts 'to nid' in msg for us */
1056 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
1058 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1059 static struct cfs_debug_limit_state cdls; \
1060 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1061 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1062 "$$ "fmt" ", ## args); \
1064 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1065 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1066 "$$ "fmt" ", ## args); \
1071 _kgnilnd_debug_conn(kgn_conn_t *conn,
1072 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1074 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
1076 CFS_CHECK_STACK(msgdata, mask, cdls); \
1078 if (((mask) & D_CANTMASK) != 0 || \
1079 ((libcfs_debug & (mask)) != 0 && \
1080 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1081 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
1084 /* for most callers (level is a constant) this is resolved at compile time */
1085 #define GNIDBG_CONN(level, conn, fmt, args...) \
1087 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1088 static struct cfs_debug_limit_state cdls; \
1089 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1090 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
1091 "$$ "fmt" ", ## args); \
1093 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1094 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
1095 "$$ "fmt" ", ## args); \
1100 _kgnilnd_debug_tx(kgn_tx_t *tx,
1101 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1103 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
1105 CFS_CHECK_STACK(msgdata, mask, cdls); \
1107 if (((mask) & D_CANTMASK) != 0 || \
1108 ((libcfs_debug & (mask)) != 0 && \
1109 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1110 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1113 /* for most callers (level is a constant) this is resolved at compile time */
1114 #define GNIDBG_TX(level, tx, fmt, args...) \
1116 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1117 static struct cfs_debug_limit_state cdls; \
1118 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1119 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1120 "$$ "fmt" ", ## args); \
1122 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1123 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1124 "$$ "fmt" ", ## args); \
1128 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1130 if (unlikely(!(cond))) { \
1131 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1136 #define GNILND_IS_QUIESCED \
1137 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1138 atomic_read(&kgnilnd_data.kgn_nthreads))
1140 #define KGNILND_SPIN_QUIESCE \
1142 /* E.T phone home */ \
1143 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1144 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1145 while (kgnilnd_data.kgn_quiesce_trigger) { \
1146 msleep_interruptible(MSEC_PER_SEC); \
1148 /* Mom, my homework is done */ \
1149 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1150 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1153 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1154 #ifndef LIBCFS_DEBUG
1155 #error "this code uses actions inside LASSERT for ref counting"
1158 #define kgnilnd_admin_addref(atomic) \
1160 int val = atomic_inc_return(&atomic); \
1161 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1162 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1165 #define kgnilnd_admin_decref(atomic) \
1167 int val = atomic_dec_return(&atomic); \
1168 LASSERTF(val >= 0, #atomic " refcount %d\n", val); \
1169 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1171 wake_up_var(&kgnilnd_data); \
1174 #define kgnilnd_net_addref(net) \
1176 int val = atomic_inc_return(&net->gnn_refcount); \
1177 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1178 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1179 libcfs_nidstr(&net->gnn_ni->ni_nid), val); \
1182 #define kgnilnd_net_decref(net) \
1184 int val = atomic_dec_return(&net->gnn_refcount); \
1185 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1186 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1187 libcfs_nidstr(&net->gnn_ni->ni_nid), val); \
1190 #define kgnilnd_peer_addref(peer) \
1192 int val = atomic_inc_return(&peer->gnp_refcount); \
1193 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1194 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1195 libcfs_nid2str(peer->gnp_nid), val); \
1198 #define kgnilnd_peer_decref(peer) \
1200 int val = atomic_dec_return(&peer->gnp_refcount); \
1201 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1202 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1203 libcfs_nid2str(peer->gnp_nid), val); \
1205 kgnilnd_destroy_peer(peer); \
1208 #define kgnilnd_conn_addref(conn) \
1213 val = atomic_inc_return(&conn->gnc_refcount); \
1214 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1215 "conn %p refc %d to %s\n", \
1218 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1220 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1222 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1227 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1228 * through the scheduler thread to get the EP destroyed. This avoids some
1229 * messy semaphore business and allows us to reuse the connd_list and existing
1230 * linkage and avoid creating extra lists just for destroying EPs */
1232 /* Safety Disclaimer:
1233 * Q: If we decrement the refcount and then check it again, is it possible that
1234 * another caller could have passed through this macro concurrently? If so,
1235 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1237 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1238 * once we are refcount <= 2. It hinges around gnc_state and membership of
1239 * gnc_hashlist. There are two ways to find a connection - either ask for
1240 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1241 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1244 * - #1 from create (kgnilnd_create_conn)
1245 * - #2 for EP (kgnilnd_create_conn)
1246 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1247 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1249 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1250 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1252 * There could be more from TX descriptors during the lifetime of a live
1255 * If we nuke the conn before finish_connect, we won't have parallel paths
1256 * because nobody besides the dgram handler for the single outstanding
1257 * dgram can find the connection as it isn't in any searchable tables yet.
1259 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1260 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1261 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1262 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1263 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1264 * sent. Now the gnc_state == CLOSED, so we hit
1265 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1266 * - we've nuked them from the peer and CQ id tables, so we own them and
1267 * are guaranteed serial access - hence the complete lack of conn list
1268 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1269 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1270 * kgnilnd_process_conns. Then the next scheduler loop would call
1271 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1272 * kgnilnd_process_conns.
1274 * Clearly, we are totally safe. Clearly.
1277 #define kgnilnd_conn_decref(conn) \
1282 val = atomic_dec_return(&conn->gnc_refcount); \
1283 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1286 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1288 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1290 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1295 (conn->gnc_ephandle != NULL) && \
1296 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1297 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1298 kgnilnd_schedule_conn(conn); \
1299 } else if (val == 0) { \
1300 kgnilnd_destroy_conn(conn); \
1304 static inline struct list_head *
1305 kgnilnd_nid2peerlist(lnet_nid_t nid)
1307 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1309 RETURN(&kgnilnd_data.kgn_peers[hash]);
1312 static inline struct list_head *
1313 kgnilnd_netnum2netlist(__u16 netnum)
1315 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1317 RETURN(&kgnilnd_data.kgn_nets[hash]);
1321 kgnilnd_peer_active(kgn_peer_t *peer)
1323 /* Am I in the peer hash table? */
1324 return (!list_empty(&peer->gnp_list));
1327 /* need write_lock on kgn_peer_conn_lock */
1329 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1331 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1332 peer, libcfs_nid2str(peer->gnp_nid),
1333 !list_empty(&peer->gnp_conns),
1334 !list_empty(&peer->gnp_tx_queue));
1336 /* kgn_peer_conn_lock protects us from conflict with
1337 * kgnilnd_peer_notify and gnp_persistent */
1338 RETURN ((list_empty(&peer->gnp_conns)) &&
1339 (list_empty(&peer->gnp_tx_queue)));
1342 /* returns positive if error was for a clean shutdown of conn */
1344 kgnilnd_conn_clean_errno(int errno)
1346 /* - ESHUTDOWN - LND is unloading
1347 * - EUCLEAN - admin requested via "lctl del_peer"
1348 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1349 * - ENOTRECOVERABLE - stack reset
1350 * - EISCONN - cleared via "lctl push"
1351 * not doing ESTALE - that isn't clean */
1352 RETURN ((errno == 0) ||
1353 (errno == -ESHUTDOWN) ||
1354 (errno == -EUCLEAN) ||
1355 (errno == -ENETRESET) ||
1356 (errno == -EISCONN) ||
1357 (errno == -ENOTRECOVERABLE));
1360 /* returns positive if error results in purgatory hold */
1362 kgnilnd_check_purgatory_errno(int errno)
1364 /* We don't want to save the purgatory lists these cases:
1365 * - EUCLEAN - admin requested via "lctl del_peer"
1366 * - ESHUTDOWN - LND is unloading
1368 RETURN ((errno != -ESHUTDOWN) &&
1369 (errno != -EUCLEAN));
1373 /* returns positive if a purgatory hold is needed */
1375 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1379 if (conn->gnc_peer) {
1380 loopback = conn->gnc_peer->gnp_nid ==
1381 lnet_nid_to_nid4(&conn->gnc_peer->gnp_net->gnn_ni->ni_nid);
1383 /* short circuit - a conn that didn't complete
1384 * setup never needs a purgatory hold */
1387 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1388 conn, conn->gnc_peer ?
1389 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1391 loopback, conn->gnc_close_recvd);
1393 /* we only use a purgatory hold if we've not received the CLOSE msg
1394 * from our peer - without that message, we can't know the state of
1395 * the other end of this connection and must put it into purgatory
1396 * to prevent reuse and corruption.
1397 * The theory is that a TX error can be communicated in all other cases
1399 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1400 kgnilnd_check_purgatory_errno(conn->gnc_error));
1403 static inline const char *
1404 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1406 static inline struct list_head *
1407 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1408 kgn_tx_list_state_t to_state)
1411 case GNILND_TX_PEERQ:
1412 return &peer->gnp_tx_queue;
1413 case GNILND_TX_FMAQ:
1414 return &conn->gnc_fmaq;
1415 case GNILND_TX_LIVE_FMAQ:
1416 case GNILND_TX_LIVE_RDMAQ:
1417 case GNILND_TX_DYING:
1419 case GNILND_TX_MAPQ:
1420 return &conn->gnc_device->gnd_map_tx;
1421 case GNILND_TX_RDMAQ:
1422 return &conn->gnc_device->gnd_rdmaq;
1424 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1425 CERROR("invalid state requested: %s\n",
1426 kgnilnd_tx_state2str(to_state));
1432 /* should hold tx, conn or peer lock when calling */
1434 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1435 kgn_conn_t *conn, kgn_tx_list_state_t state,
1438 struct list_head *list = NULL;
1440 /* make sure we have a sane TX state to start */
1441 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1442 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1443 list_empty(&tx->tx_list),
1444 "bad state with tx_list %s",
1445 list_empty(&tx->tx_list) ? "empty" : "not empty");
1447 /* WTF - you are already on that state buttmunch */
1448 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1449 "already at %s", kgnilnd_tx_state2str(state));
1451 /* get proper list from the state requested */
1452 list = kgnilnd_tx_state2list(peer, conn, state);
1456 case GNILND_TX_PEERQ:
1457 kgnilnd_peer_addref(peer);
1459 case GNILND_TX_ALLOCD:
1460 /* no refs needed */
1462 case GNILND_TX_FMAQ:
1463 kgnilnd_conn_addref(conn);
1465 case GNILND_TX_MAPQ:
1466 atomic_inc(&conn->gnc_device->gnd_nq_map);
1467 kgnilnd_conn_addref(conn);
1469 case GNILND_TX_LIVE_FMAQ:
1470 atomic_inc(&conn->gnc_nlive_fma);
1471 kgnilnd_conn_addref(conn);
1473 case GNILND_TX_LIVE_RDMAQ:
1474 atomic_inc(&conn->gnc_nlive_rdma);
1475 kgnilnd_conn_addref(conn);
1477 case GNILND_TX_RDMAQ:
1478 atomic_inc(&conn->gnc_nq_rdma);
1479 kgnilnd_conn_addref(conn);
1481 case GNILND_TX_DYING:
1482 kgnilnd_conn_addref(conn);
1485 CERROR("invalid state requested: %s\n",
1486 kgnilnd_tx_state2str(state));
1491 /* if this changes, change kgnilnd_alloc_tx */
1492 tx->tx_list_state = state;
1494 /* some states don't have lists - we track them in the per conn
1495 * TX table instead. Waste not, want not! */
1497 tx->tx_list_p = list;
1499 list_add_tail(&tx->tx_list, list);
1501 list_add(&tx->tx_list, list);
1503 /* set dummy list_p to make book keeping happy and let debugging
1504 * be a hair easier */
1505 tx->tx_list_p = (void *)state;
1508 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1509 kgnilnd_tx_state2str(state), list);
1513 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1514 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1516 /* These is only 1 "off-list" state */
1517 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1518 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1520 /* new_state == ALLOCD means we are deallocating this tx,
1521 * so make sure it was on a valid list to start with */
1522 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1523 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1524 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1525 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1528 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1531 switch (tx->tx_list_state) {
1532 case GNILND_TX_PEERQ:
1533 kgnilnd_peer_decref(peer);
1535 case GNILND_TX_FREED:
1536 case GNILND_TX_IDLE:
1537 case GNILND_TX_ALLOCD:
1538 /* no refs needed */
1540 case GNILND_TX_DYING:
1541 kgnilnd_conn_decref(conn);
1543 case GNILND_TX_FMAQ:
1544 kgnilnd_conn_decref(conn);
1546 case GNILND_TX_MAPQ:
1547 atomic_dec(&conn->gnc_device->gnd_nq_map);
1548 kgnilnd_conn_decref(conn);
1550 case GNILND_TX_LIVE_FMAQ:
1551 atomic_dec(&conn->gnc_nlive_fma);
1552 kgnilnd_conn_decref(conn);
1554 case GNILND_TX_LIVE_RDMAQ:
1555 atomic_dec(&conn->gnc_nlive_rdma);
1556 kgnilnd_conn_decref(conn);
1558 case GNILND_TX_RDMAQ:
1559 atomic_dec(&conn->gnc_nq_rdma);
1560 kgnilnd_conn_decref(conn);
1561 /* don't need to assert on default, already did in set */
1564 /* for ALLOCD, this might already be true, but no harm doing it again */
1565 list_del_init(&tx->tx_list);
1566 tx->tx_list_p = NULL;
1567 tx->tx_list_state = new_state;
1571 kgnilnd_tx_mapped(kgn_tx_t *tx)
1573 return tx->tx_buftype == GNILND_BUF_PHYS_MAPPED;
1576 static inline struct list_head *
1577 kgnilnd_cqid2connlist(__u32 cqid)
1579 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1581 return (&kgnilnd_data.kgn_conns [hash]);
1584 static inline kgn_conn_t *
1585 kgnilnd_cqid2conn_locked(__u32 cqid)
1587 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1588 struct list_head *tmp;
1591 list_for_each(tmp, conns) {
1592 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1594 if (conn->gnc_cqid == cqid)
1601 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1603 kgnilnd_get_cqid_locked(void)
1609 cqid = kgnilnd_data.kgn_next_cqid++;
1610 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1614 kgnilnd_data.kgn_next_cqid = 1;
1617 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1623 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1625 kgn_tx_t *tx = NULL;
1626 kgn_conn_t *conn = NULL;
1628 /* set to NULL so any early return is an error */
1632 LASSERTF((ev_id->txe_idx > 0) &&
1633 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1634 "bogus txe_idx %d >= %d\n",
1635 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1637 LASSERTF((ev_id->txe_cqid > 0) &&
1638 (ev_id->txe_cqid < GNILND_MAX_CQID),
1639 "bogus txe_cqid %d >= %d\n",
1640 ev_id->txe_cqid, GNILND_MAX_CQID);
1642 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1643 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1646 /* Conn was destroyed? */
1647 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1648 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1651 /* just insurance */
1652 kgnilnd_conn_addref(conn);
1653 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1654 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1656 /* we know this is safe - as the TX won't be reused until AFTER
1657 * the conn is unlinked from the cqid hash, so we can use the TX
1658 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1660 spin_lock(&conn->gnc_tx_lock);
1661 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1662 spin_unlock(&conn->gnc_tx_lock);
1664 /* We could have a tx that was cleared out by other forces
1665 * lctl disconnect or del_peer. */
1667 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1668 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1669 kgnilnd_conn_decref(conn);
1673 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1674 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1675 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1677 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1678 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1679 "found tx %p txe_idx %d",
1680 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1681 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1683 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1685 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1690 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1693 /* set_normalized_timepsec isn't exported from the kernel, so
1694 * we need to do the same thing inline */
1695 static inline struct timespec
1696 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1702 sec = lhs.tv_sec - rhs.tv_sec;
1703 nsec = lhs.tv_nsec - rhs.tv_nsec;
1705 while (nsec >= NSEC_PER_SEC) {
1706 nsec -= NSEC_PER_SEC;
1710 nsec += NSEC_PER_SEC;
1719 kgnilnd_count_list(struct list_head *q)
1721 struct list_head *e;
1724 list_for_each(e, q) {
1731 /* kgnilnd_find_net adds a reference to the net it finds
1732 * this is so the net will not be removed before the calling function
1733 * has time to use the data returned. This reference needs to be released
1734 * by the calling function once it has finished using the returned net
1738 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1743 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1749 list_for_each_entry(net,
1750 kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))),
1752 if (!net->gnn_shutdown &&
1753 LNET_NID_NET(&net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1754 kgnilnd_net_addref(net);
1755 up_read(&kgnilnd_data.kgn_net_rw_sem);
1761 up_read(&kgnilnd_data.kgn_net_rw_sem);
1766 #ifdef CONFIG_DEBUG_SLAB
1767 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1769 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1772 int kgnilnd_dev_init(kgn_device_t *dev);
1773 void kgnilnd_dev_fini(kgn_device_t *dev);
1774 int kgnilnd_startup(struct lnet_ni *ni);
1775 void kgnilnd_shutdown(struct lnet_ni *ni);
1776 int kgnilnd_base_startup(void);
1777 void kgnilnd_base_shutdown(void);
1779 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1780 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1781 void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
1782 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1784 int kgnilnd_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg);
1785 int kgnilnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1786 int kgnilnd_eager_recv(struct lnet_ni *ni, void *private,
1787 struct lnet_msg *lntmsg, void **new_private);
1788 int kgnilnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1789 int delayed, unsigned int niov,
1790 struct bio_vec *kiov,
1791 unsigned int offset, unsigned int mlen, unsigned int rlen);
1793 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, struct bio_vec *kiov,
1794 unsigned int offset, unsigned int nob, int dump_blob);
1796 /* purgatory functions */
1797 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1798 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1799 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1800 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1802 void kgnilnd_update_reaper_timeout(long timeout);
1803 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1804 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1805 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1806 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1807 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1808 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held);
1809 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1811 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1812 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
1813 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1814 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1816 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1817 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1818 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1819 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1820 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1821 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net,
1822 struct lnet_processid *target);
1823 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1824 void kgnilnd_consume_rx(kgn_rx_t *rx);
1826 void kgnilnd_schedule_device(kgn_device_t *dev);
1827 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1828 void kgnilnd_schedule_device_timer(cfs_timer_cb_arg_t data);
1829 void kgnilnd_schedule_device_timer_rd(cfs_timer_cb_arg_t data);
1831 int kgnilnd_reaper(void *arg);
1832 int kgnilnd_scheduler(void *arg);
1833 int kgnilnd_dgram_mover(void *arg);
1834 int kgnilnd_rca(void *arg);
1835 int kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id);
1837 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1838 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1839 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1840 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1841 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1842 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1843 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1844 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1845 void kgnilnd_peer_alive(kgn_peer_t *peer);
1846 void kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive);
1847 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1848 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1849 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1850 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1852 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1853 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1854 void kgnilnd_wakeup_rca_thread(void);
1855 int kgnilnd_start_rca_thread(void);
1856 int kgnilnd_get_node_state(__u32 nid);
1858 int kgnilnd_tunables_init(void);
1859 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1861 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1862 void kgnilnd_pause_threads(void);
1863 int kgnilnd_hw_in_quiesce(void);
1864 int kgnilnd_check_hw_quiesce(void);
1865 void kgnilnd_quiesce_wait(char *reason);
1866 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1867 int kgnilnd_ruhroh_thread(void *arg);
1868 void kgnilnd_reset_stack(void);
1869 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1871 void kgnilnd_insert_sysctl(void);
1872 void kgnilnd_remove_sysctl(void);
1873 void kgnilnd_proc_init(void);
1874 void kgnilnd_proc_fini(void);
1877 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1879 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1880 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1881 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
1883 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1884 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1885 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1886 int kgnilnd_cancel_dgrams(kgn_device_t *dev);
1887 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1889 int kgnilnd_dgram_waitq(void *arg);
1891 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1893 /* struct2str functions - we don't use a default: case to cause the compile
1894 * to fail if there is a missing case. This allows us to hide these down here
1895 * out of the way but ensure we'll catch any updates to the enum/types
1898 #define DO_TYPE(x) case x: return #x;
1899 static inline const char *
1900 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1902 /* Only want single char string for this */
1904 case GNILND_FMABLK_IDLE:
1906 case GNILND_FMABLK_PHYS:
1908 case GNILND_FMABLK_VIRT:
1910 case GNILND_FMABLK_FREED:
1913 return "<unknown state>";
1916 static inline const char *
1917 kgnilnd_msgtype2str(int type)
1920 DO_TYPE(GNILND_MSG_NONE);
1921 DO_TYPE(GNILND_MSG_NOOP);
1922 DO_TYPE(GNILND_MSG_IMMEDIATE);
1923 DO_TYPE(GNILND_MSG_PUT_REQ);
1924 DO_TYPE(GNILND_MSG_PUT_NAK);
1925 DO_TYPE(GNILND_MSG_PUT_ACK);
1926 DO_TYPE(GNILND_MSG_PUT_DONE);
1927 DO_TYPE(GNILND_MSG_GET_REQ);
1928 DO_TYPE(GNILND_MSG_GET_NAK);
1929 DO_TYPE(GNILND_MSG_GET_DONE);
1930 DO_TYPE(GNILND_MSG_CLOSE);
1931 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1932 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1933 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1934 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1935 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1936 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1937 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1939 return "<unknown msg type>";
1942 static inline const char *
1943 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1946 DO_TYPE(GNILND_TX_IDLE);
1947 DO_TYPE(GNILND_TX_ALLOCD);
1948 DO_TYPE(GNILND_TX_PEERQ);
1949 DO_TYPE(GNILND_TX_MAPQ);
1950 DO_TYPE(GNILND_TX_FMAQ);
1951 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1952 DO_TYPE(GNILND_TX_RDMAQ);
1953 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1954 DO_TYPE(GNILND_TX_DYING);
1955 DO_TYPE(GNILND_TX_FREED);
1957 return "<unknown state>";
1960 static inline const char *
1961 kgnilnd_conn_state2str(kgn_conn_t *conn)
1963 kgn_conn_state_t state = conn->gnc_state;
1965 DO_TYPE(GNILND_CONN_DUMMY);
1966 DO_TYPE(GNILND_CONN_LISTEN);
1967 DO_TYPE(GNILND_CONN_CONNECTING);
1968 DO_TYPE(GNILND_CONN_ESTABLISHED);
1969 DO_TYPE(GNILND_CONN_CLOSING);
1970 DO_TYPE(GNILND_CONN_CLOSED);
1971 DO_TYPE(GNILND_CONN_DONE);
1972 DO_TYPE(GNILND_CONN_DESTROY_EP);
1977 static inline const char *
1978 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1980 kgn_connreq_type_t type = connreq->gncr_type;
1983 DO_TYPE(GNILND_CONNREQ_REQ);
1984 DO_TYPE(GNILND_CONNREQ_NAK);
1985 DO_TYPE(GNILND_CONNREQ_CLOSE);
1990 static inline const char *
1991 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1993 kgn_dgram_state_t state = dgram->gndg_state;
1996 DO_TYPE(GNILND_DGRAM_USED);
1997 DO_TYPE(GNILND_DGRAM_POSTING);
1998 DO_TYPE(GNILND_DGRAM_POSTED);
1999 DO_TYPE(GNILND_DGRAM_PROCESSING);
2000 DO_TYPE(GNILND_DGRAM_DONE);
2001 DO_TYPE(GNILND_DGRAM_CANCELED);
2006 static inline const char *
2007 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
2009 kgn_dgram_type_t type = dgram->gndg_type;
2012 DO_TYPE(GNILND_DGRAM_REQ);
2013 DO_TYPE(GNILND_DGRAM_WC_REQ);
2014 DO_TYPE(GNILND_DGRAM_NAK);
2015 DO_TYPE(GNILND_DGRAM_CLOSE);
2020 static inline const char *
2021 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
2024 DO_TYPE(GNILND_DGRAM_REQ);
2025 DO_TYPE(GNILND_DGRAM_WC_REQ);
2026 DO_TYPE(GNILND_DGRAM_NAK);
2027 DO_TYPE(GNILND_DGRAM_CLOSE);
2034 /* pulls in tunables per platform and adds in nid/nic conversion
2035 * if RCA wasn't available at build time */
2036 #include "gnilnd_hss_ops.h"
2037 /* API wrapper functions - include late to pick up all of the other defines */
2038 #include "gnilnd_api_wrap.h"
2040 #if defined(CONFIG_CRAY_GEMINI)
2041 #include "gnilnd_gemini.h"
2042 #elif defined(CONFIG_CRAY_ARIES)
2043 #include "gnilnd_aries.h"
2045 #error "Undefined Network Hardware Type"
2048 extern uint32_t kgni_driver_version;
2051 kgnilnd_check_kgni_version(void)
2055 kgnilnd_data.kgn_enable_gl_mutex = 1;
2056 kdv = symbol_get(kgni_driver_version);
2058 LCONSOLE_INFO("Not using thread safe locking -"
2059 " no symbol kgni_driver_version\n");
2063 /* Thread-safe kgni implemented in minor ver 0x44/45, code rev 0xb9 */
2064 if (*kdv < GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9)) {
2065 symbol_put(kgni_driver_version);
2066 LCONSOLE_INFO("Not using thread safe locking, gni version 0x%x,"
2067 " need >= 0x%x\n", *kdv,
2068 GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9));
2072 symbol_put(kgni_driver_version);
2074 if (!*kgnilnd_tunables.kgn_thread_safe) {
2078 /* Use thread-safe locking */
2079 kgnilnd_data.kgn_enable_gl_mutex = 0;
2082 #endif /* _GNILND_GNILND_H_ */