2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Copyright (c) 2013, 2014, Intel Corporation.
8 * Derived from work by: Eric Barton <eric@bartonsoftware.com>
9 * Author: Nic Henke <nic@cray.com>
10 * Author: James Shimek <jshimek@cray.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #ifndef _GNILND_GNILND_H_
29 #define _GNILND_GNILND_H_
31 #ifdef HAVE_COMPAT_RDMA
32 #include <linux/compat-2.6.h>
34 #include <linux/module.h>
35 #include <linux/kernel.h>
37 #include <linux/string.h>
38 #include <linux/stat.h>
39 #include <linux/errno.h>
40 #ifdef HAVE_LINUX_KERNEL_LOCK
41 #include <linux/smp_lock.h>
43 #include <linux/unistd.h>
44 #include <linux/uio.h>
45 #include <linux/time.h>
46 #include <asm/timex.h>
48 #include <asm/uaccess.h>
51 #include <linux/init.h>
53 #include <linux/file.h>
54 #include <linux/stat.h>
55 #include <linux/list.h>
56 #include <linux/kmod.h>
57 #include <linux/sysctl.h>
58 #include <linux/kthread.h>
59 #include <linux/nmi.h>
63 #include <linux/nmi.h>
65 #define DEBUG_SUBSYSTEM S_LND
67 #include <libcfs/linux/kp30.h>
68 #include <libcfs/libcfs.h>
69 #include <lnet/lnet.h>
70 #include <lnet/lib-lnet.h>
73 #include "gnilnd_version.h"
76 /* tunables determined at compile time */
77 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
78 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
79 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
80 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
81 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
82 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
83 #define GNILND_SCHED_TIMEOUT 1
84 #define GNILND_DGRAM_TIMEOUT 2
85 #define GNILND_FAST_MAPPING_TRY \
86 *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
87 #define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
89 /* map failure timeout */
90 #define GNILND_MAP_TIMEOUT \
91 (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
92 *kgnilnd_tunables.kgn_timeout))
94 /* reaper thread wakup interval */
95 #define GNILND_REAPER_THREAD_WAKE 1
96 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
97 #define GNILND_REAPER_NCHECKS 4
100 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
101 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
102 #define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
103 /* checksum values */
104 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
105 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
106 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
107 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
109 /* tune down some COMPUTE options as they won't see the same number of connections and
110 * don't need the throughput of multiple threads by default */
111 #if defined(CONFIG_CRAY_COMPUTE)
113 #define GNILND_SCHED_THREADS 2 /* default # of kgnilnd_scheduler threads */
115 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
117 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
118 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
119 #define GNILND_COMPUTE 1 /* compute image */
120 #define GNILND_FAST_RECONNECT 1 /* Fast Reconnect option */
122 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
123 #define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
124 #define GNILND_COMPUTE 0 /* service image */
125 #define GNILND_FAST_RECONNECT 0 /* Fast Reconnect option */
128 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
129 #define GNILND_EXTRA_BITS 1
130 /* maximum number of conns & bits for cqid in the SMSG event data */
131 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
132 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
133 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
134 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
135 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
137 /* need sane upper bound to limit copy overhead */
138 #define GNILND_MAX_IMMEDIATE (64<<10)
140 /* Max number of connections to keep in purgatory per peer */
141 #define GNILND_PURGATORY_MAX 5
142 /* Closing, don't put in purgatory */
143 #define GNILND_NOPURG 222
145 /* payload size to add to the base mailbox size
146 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
147 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
148 * the calculation return from that function.*/
149 #define GNILND_MBOX_PAYLOAD \
150 (GNILND_MAX_MSG_SIZE * \
151 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
153 /* timeout -> deadman timer for kgni mdd holds */
154 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
156 /* timeout for failing sends in t is in jiffies*/
157 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
159 /* time when to release from purgatory in the reaper thread in jiffies */
160 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
162 /* Macro for finding last_rx 2 datapoints are compared
163 * and the most recent one in jiffies is returned.
165 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
166 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
168 /************************************************************************
169 * Enum, flag and tag data
171 #define GNILND_INIT_NOTHING 0
172 #define GNILND_INIT_DATA 1
173 #define GNILND_INIT_ALL 2
175 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
176 #define GNILND_BUF_NONE 0 /* buffer type not set */
177 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
178 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
179 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
180 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
181 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
182 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
184 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
185 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
186 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
187 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
188 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
190 /* stash above max CQID to avoid any collision */
191 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
192 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
194 /* kgn_msg_t::gnm_type */
195 #define GNILND_MSG_NONE 0x00 /* illegal message */
196 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
197 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
198 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
199 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
200 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
201 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
202 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
203 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
204 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
205 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
206 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
207 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
208 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
209 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
210 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
211 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
212 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
214 /* defines for gnc_*scheduled states */
215 #define GNILND_CONN_IDLE 0
216 #define GNILND_CONN_SCHED 1
217 #define GNILND_CONN_WANTS_SCHED 2
218 #define GNILND_CONN_PROCESS 3
220 #define GNILND_DEV_IDLE 0
221 #define GNILND_DEV_IRQ 1
222 #define GNILND_DEV_LOOP 2
224 #define GNILND_DGRAM_IDLE 0
225 #define GNILND_DGRAM_SCHED 1
226 #define GNILND_DGRAM_PROCESS 2
228 #define GNILND_PEER_IDLE 0
229 #define GNILND_PEER_CONNECT 1
230 #define GNILND_PEER_POSTING 2
231 #define GNILND_PEER_POSTED 3
232 #define GNILND_PEER_NEEDS_DEATH 4
233 #define GNILND_PEER_KILL 5
235 /* for gnc_close_recvd */
236 #define GNILND_CLOSE_RX 1
237 #define GNILND_CLOSE_INJECT1 2
238 #define GNILND_CLOSE_INJECT2 3
239 #define GNILND_CLOSE_EARLY 4
241 /* defines for why quiesce trigger set */
242 #define GNILND_QUIESCE_IDLE 0
243 #define GNILND_QUIESCE_ADMIN 1
244 #define GNILND_QUIESCE_RESET 2
245 #define GNILND_QUIESCE_HW_QUIESCE 3
247 #define GNILND_PEER_CLEAN 0
248 #define GNILND_PEER_PERSISTING 1
250 #define GNILND_DEL_CONN 0
251 #define GNILND_DEL_PEER 1
252 #define GNILND_CLEAR_PURGATORY 2
254 #define GNILND_RCA_NODE_UP 0
255 #define GNILND_RCA_NODE_DOWN 1
256 #define GNILND_RCA_NODE_UNKNOWN 2
258 /* defines for reverse RDMA states */
259 #define GNILND_REVERSE_NONE 0
260 #define GNILND_REVERSE_GET 1
261 #define GNILND_REVERSE_PUT 2
262 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
264 typedef enum kgn_fmablk_state {
265 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
266 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
267 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
268 GNILND_FMABLK_FREED, /* after free */
269 } kgn_fmablk_state_t;
271 typedef enum kgn_tx_list_state {
272 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
273 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
274 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
275 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
276 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
277 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
278 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
279 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
280 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
281 GNILND_TX_FREED /* TX is free! */
282 } kgn_tx_list_state_t;
284 typedef enum kgn_conn_state {
285 /* don't start @ 0 - prevent memset(0) badness */
286 GNILND_CONN_DUMMY = 0,
288 GNILND_CONN_CONNECTING,
289 GNILND_CONN_ESTABLISHED,
293 GNILND_CONN_DESTROY_EP
296 /* changing these requires a change to GNILND_CONNREQ_VERSION and
297 * will result in dropped packets instead of NAKs. Adding to this is
298 * acceptable without changing the CONNREQ_VERSION, but code should
299 * be ready to handle NAKs on version mismatch */
300 typedef enum kgn_connreq_type {
301 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
302 GNILND_CONNREQ_NAK, /* NO soup for you! */
303 GNILND_CONNREQ_CLOSE, /* we should see other people */
304 } kgn_connreq_type_t;
306 typedef enum kgn_dgram_state {
307 /* don't use 0 to avoid thinking a memset of zero is valid data */
308 GNILND_DGRAM_USED = 1,
309 GNILND_DGRAM_POSTING,
311 GNILND_DGRAM_PROCESSING,
312 GNILND_DGRAM_CANCELED,
316 typedef enum kgn_dgram_type {
317 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
318 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
319 GNILND_DGRAM_NAK, /* NO soup for you! */
320 GNILND_DGRAM_CLOSE, /* we should see other people */
323 /************************************************************************
324 * Wire message structs. These are sent in sender's byte order
325 * (i.e. receiver checks magic and flips if required).
328 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
329 #define GNILND_DGRAM_MAGIC 0x0DDBA11
331 /* kgn_msg_t - FMA/SMSG wire struct
333 * - added checksum to FMA
334 * moved seq before paylod
335 * WIRE_ATTR added for alignment
337 * added gnm_payload_len for FMA payload size
339 * added gncm_retval to completion, allowing return code transmission
342 * changed how CQID and TX ids are assigned
344 * added retval on CLOSE
346 * added payload checksumming
348 * reworked checksumming a bit, changed payload checksums
350 #define GNILND_MSG_VERSION 8
351 /* kgn_connreq_t connection request datagram wire struct
356 #define GNILND_CONNREQ_VERSION 2
358 typedef struct kgn_gniparams {
359 __u32 gnpr_host_id; /* ph. host ID of the NIC */
360 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
361 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
362 } WIRE_ATTR kgn_gniparams_t;
364 typedef struct kgn_nak_data {
365 __s32 gnnd_errno; /* errno reason for NAK */
367 } WIRE_ATTR kgn_nak_data_t;
369 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
370 * without breaking the ability for us to properly NAK someone */
371 typedef struct kgn_connreq { /* connection request/response */
372 __u32 gncr_magic; /* I'm an gnilnd connreq */
373 __u32 gncr_cksum; /* checksum (0 == disabled) */
374 __u16 gncr_type; /* REQ, NAK, etc */
375 __u16 gncr_version; /* this is my version number */
376 __u32 gncr_timeout; /* sender's timeout */
377 __u64 gncr_srcnid; /* sender's NID */
378 __u64 gncr_dstnid; /* who sender expects to listen */
379 __u64 gncr_peerstamp; /* sender's instance stamp */
380 __u64 gncr_connstamp; /* sender's connection stamp */
382 /* everything before this needs to stay static, adding after should
383 * result in a change to GNILND_CONNREQ_VERSION */
386 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
387 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
389 } WIRE_ATTR kgn_connreq_t;
392 gni_mem_handle_t gnrd_key;
395 } WIRE_ATTR kgn_rdma_desc_t;
398 lnet_hdr_t gnim_hdr; /* LNet header */
399 /* LNet payload is in FMA "Message Data" */
400 } WIRE_ATTR kgn_immediate_msg_t;
403 lnet_hdr_t gnprm_hdr; /* LNet header */
404 __u64 gnprm_cookie; /* opaque completion cookie */
405 } WIRE_ATTR kgn_putreq_msg_t;
408 __u64 gnpam_src_cookie; /* reflected completion cookie */
409 __u64 gnpam_dst_cookie; /* opaque completion cookie */
410 __u16 gnpam_payload_cksum; /* checksum for get msg */
411 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
412 } WIRE_ATTR kgn_putack_msg_t;
415 lnet_hdr_t gngm_hdr; /* LNet header */
416 __u64 gngm_cookie; /* opaque completion cookie */
417 __u16 gngm_payload_cksum; /* checksum for put msg */
418 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
419 } WIRE_ATTR kgn_get_msg_t;
422 int gncm_retval; /* error on NAK, size on REQ */
423 __u64 gncm_cookie; /* reflected completion cookie */
424 } WIRE_ATTR kgn_completion_msg_t;
426 typedef struct { /* NB must fit in FMA "Prefix" */
427 __u32 gnm_magic; /* I'm an gni message */
428 __u16 gnm_version; /* this is my version number */
429 __u16 gnm_type; /* msg type */
430 __u64 gnm_srcnid; /* sender's NID */
431 __u64 gnm_connstamp; /* sender's connection stamp */
432 __u32 gnm_seq; /* incrementing sequence number */
433 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
434 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
435 __u32 gnm_payload_len; /* size of the FMA payload sent */
437 kgn_immediate_msg_t immediate;
438 kgn_putreq_msg_t putreq;
439 kgn_putack_msg_t putack;
441 kgn_completion_msg_t completion;
443 } WIRE_ATTR kgn_msg_t;
445 /************************************************************************
446 * runtime tunable data
449 typedef struct kgn_tunables {
450 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
451 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
452 int *kgn_credits; /* # concurrent sends */
453 int *kgn_fma_cq_size; /* # entries in receive CQ */
454 int *kgn_peer_credits; /* # LNet peer credits */
455 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
456 int *kgn_timeout; /* comms timeout (seconds) */
457 int *kgn_max_immediate; /* immediate payload breakpoint */
458 int *kgn_checksum; /* checksum data */
459 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
460 int *kgn_bte_dlvr_mode; /* BTE delivery mode mask */
461 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
462 int *kgn_ptag; /* PTAG for cdm_create */
463 int *kgn_pkey; /* PKEY for cdm_create */
464 int *kgn_max_retransmits; /* max number of FMA retransmits */
465 int *kgn_nwildcard; /* # wildcard per net to post */
466 int *kgn_nice; /* nice value for kgnilnd threads */
467 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
468 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
469 int *kgn_peer_hash_size; /* size of kgn_peers */
470 int *kgn_peer_health; /* enable/disable peer health */
471 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
472 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
473 int *kgn_mbox_per_block; /* mailboxes per fmablk */
474 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
475 int *kgn_mbox_credits; /* max credits per fma */
476 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
477 int *kgn_net_hash_size; /* size of kgn_net_ht */
478 int *kgn_hardware_timeout; /* max time for a message to get across the network */
479 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
480 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
481 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
482 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
483 int *kgn_reverse_rdma; /* Reverse RDMA setting */
484 int *kgn_eager_credits; /* allocated eager buffers */
485 int *kgn_fast_reconn; /* fast reconnection on conn timeout */
486 int *kgn_efault_lbug; /* LBUG on receiving an EFAULT */
487 int *kgn_max_purgatory; /* # conns/peer to keep in purgatory */
488 int *kgn_thread_affinity; /* bind scheduler threads to cpus */
489 int *kgn_thread_safe; /* use thread safe kgni API */
490 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
491 struct ctl_table_header *kgn_sysctl; /* sysctl interface */
495 typedef struct kgn_mbox_info {
496 lnet_nid_t mbx_prev_nid;
497 lnet_nid_t mbx_prev_purg_nid;
498 unsigned long mbx_create_conn_memset;
499 unsigned long mbx_add_purgatory;
500 unsigned long mbx_detach_of_purgatory;
501 unsigned long mbx_release_from_purgatory;
502 unsigned long mbx_release_purg_active_dgram;
504 int mbx_nallocs_total;
507 typedef struct kgn_fma_memblock {
508 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
509 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
510 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
511 int gnm_num_mboxs; /* total mboxes allocated */
512 int gnm_avail_mboxs; /* number of available mailboxes in the block */
513 int gnm_held_mboxs; /* number of purgatory held mailboxes */
514 int gnm_mbox_size; /* size of the single mailbox */
515 int gnm_next_avail_mbox; /* next available mailbox */
516 long gnm_max_timeout; /* max timeout for possible purgatory hold */
517 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
518 void *gnm_block; /* pointer to mem. block */
519 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
520 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
521 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
522 } kgn_fma_memblock_t;
524 typedef struct kgn_device {
525 gni_nic_handle_t gnd_handle; /* device handle */
526 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
527 gni_err_handle_t gnd_err_handle; /* device error handle */
528 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
529 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
530 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
531 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
532 struct mutex gnd_cq_mutex; /* CQ access serialization */
533 __u32 gnd_host_id; /* ph. host ID of the NIC */
534 int gnd_id; /* device id, also index in kgn_devices */
535 __u32 gnd_nid; /* ph host ID translated to NID */
536 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
537 struct mutex gnd_fmablk_mutex; /* mutex for FMA block memory alloc/free */
538 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
539 atomic_t gnd_nfmablk; /* # of fmablk live */
540 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
541 atomic_t gnd_neps; /* # EP allocated to conns */
542 short gnd_ready; /* stuff to do in scheduler thread */
543 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
544 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
545 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
546 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
547 struct list_head gnd_connd_peers; /* peers waiting for a connection */
548 spinlock_t gnd_connd_lock; /* serialise connd_peers */
549 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
550 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
551 int gnd_dgram_ready; /* dgrams need movin' */
552 struct list_head *gnd_dgrams; /* nid hash to dgrams */
553 atomic_t gnd_ndgrams; /* # dgrams extant */
554 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post*/
555 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
556 struct list_head gnd_map_list; /* list of all mapped regions */
557 int gnd_map_version; /* version flag for map list */
558 struct timer_list gnd_map_timer; /* wakey-wakey */
559 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
560 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
561 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
562 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
563 __u32 gnd_map_nphys; /* # TX phys mappings */
564 __u32 gnd_map_physnop; /* # TX phys pages mapped */
565 __u32 gnd_map_nvirt; /* # TX virt mappings */
566 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
567 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
568 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
569 int gnd_map_attempt; /* last map attempt # */
570 unsigned long gnd_last_map; /* map timeout base */
571 struct list_head gnd_rdmaq; /* RDMA to be sent */
572 spinlock_t gnd_rdmaq_lock; /* play nice with others */
573 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
574 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
575 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
576 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
577 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
578 atomic_t gnd_short_ntx; /* TX stats: short messages */
579 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
580 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
581 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
582 atomic_t gnd_short_nrx; /* RX stats: short messages */
583 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
584 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
585 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
586 atomic_t gnd_fast_try; /* # of times fast send tried */
587 atomic_t gnd_fast_ok; /* # of times fast send ok */
588 atomic_t gnd_fast_block; /* # of times fast send blocked */
589 unsigned long gnd_mutex_delay;
590 atomic_t gnd_n_yield;
591 atomic_t gnd_n_schedule;
592 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
593 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
594 void *gnd_smdd_hold_buf; /* buffer to keep smdd */
595 gni_mem_handle_t gnd_smdd_hold_hndl; /* buffer mem handle */
598 typedef struct kgn_net {
599 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
600 kgn_device_t *gnn_dev; /* device for this net */
601 lnet_ni_t *gnn_ni; /* network interface instance */
602 atomic_t gnn_refcount; /* # current references */
603 int gnn_shutdown; /* lnd_shutdown set */
604 __u16 gnn_netnum; /* stash netnum for quicker lookup */
607 static inline lnet_nid_t
608 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
610 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
613 static inline lnet_nid_t
614 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
616 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
619 /* The code for this is a bit ugly - but really this just boils down to a __u64
620 * that can have various parts accessed separately.
622 * The lower 32 bits is the ID
623 * we give to SMSG for our completion event - it needs to be globally unique across
624 * all TX currently in flight. We separate that out into the CQID so that we can
625 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
626 * the actual TX out of the per-connection gnc_tx_ref_table.
628 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
629 * has a unique value we can send with RDMA setup messages to ensure the completion for
630 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
631 * reuse is separated.
634 typedef struct kgn_tx_ev_id {
636 __u64 txe_cookie; /* are you my mommy ? */
638 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
640 __u32 txe_smsg_id; /* ID for SMSG CQ event */
641 /* N.B: Never ever ever ever use the bit shifts directly,
642 * you are just asking for a world of pain and are at the
643 * mercy of the compiler layouts */
645 __u32 txe_cqid :GNILND_CQID_NBITS;
646 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
653 typedef struct kgn_dgram {
654 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
655 kgn_dgram_state_t gndg_state; /* state of this dgram */
656 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
657 __u32 gndg_magic; /* saftey word */
658 unsigned long gndg_post_time; /* time when we posted */
659 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
660 kgn_connreq_t gndg_conn_out; /* connreq from local node */
661 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
664 typedef struct kgn_tx { /* message descriptor */
665 struct list_head tx_list; /* TX queues - peer, conn, rdma */
666 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
667 struct list_head *tx_list_p; /* pointer to current list */
668 struct kgn_conn *tx_conn; /* owning conn */
669 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
670 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
671 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
672 struct list_head tx_map_list; /* list entry on device map list */
673 unsigned int tx_nob; /* # bytes of payload */
674 int tx_buftype; /* payload buffer type */
675 int tx_phys_npages; /* # physical pages */
676 gni_mem_handle_t tx_map_key; /* mapping key */
677 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
678 gni_mem_segment_t *tx_phys; /* page descriptors */
679 kgn_msg_t tx_msg; /* FMA message buffer */
680 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
681 __u8 tx_state; /* state of the descriptor */
682 int tx_retrans; /* retrans count of RDMA */
683 int tx_rc; /* if we need to stash the ret code until we see completion */
684 void *tx_buffer; /* source/sink buffer */
685 void *tx_buffer_copy; /* pointer to page aligned buffer */
686 unsigned int tx_nob_rdma; /* nob actually rdma */
687 unsigned int tx_offset; /* offset of data into copied buffer */
689 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
690 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
693 /* we only use one or the other */
695 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
696 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
700 typedef struct kgn_conn {
701 kgn_device_t *gnc_device; /* which device */
702 struct kgn_peer *gnc_peer; /* owning peer */
703 int gnc_magic; /* magic value cleared before free */
704 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
705 struct list_head gnc_hashlist; /* stash in connection hash table */
706 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
707 struct list_head gnc_fmaq; /* txs queued for FMA */
708 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
709 __u64 gnc_peerstamp; /* peer's unique stamp */
710 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
711 __u64 gnc_my_connstamp; /* my unique connection stamp */
712 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
713 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
714 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
715 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
716 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
717 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
718 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
719 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
720 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
721 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
722 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
723 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
724 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
725 __u32 gnc_cqid; /* my completion callback id (non-unique) */
726 atomic_t gnc_tx_seq; /* tx msg sequence number */
727 atomic_t gnc_rx_seq; /* rx msg sequence number */
728 struct mutex gnc_smsg_mutex; /* tx smsg sequence serialization */
729 struct mutex gnc_rdma_mutex; /* tx rdma sequence serialization */
730 __u64 gnc_tx_retrans; /* # retrans on SMSG */
731 atomic_t gnc_nlive_fma; /* # live FMA */
732 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
733 atomic_t gnc_nlive_rdma; /* # live RDMA */
734 short gnc_close_sent; /* I've sent CLOSE */
735 short gnc_close_recvd; /* I've received CLOSE */
736 short gnc_in_purgatory; /* in the sin bin */
737 int gnc_error; /* errno when conn being closed due to error */
738 int gnc_peer_error; /* errno peer sent us on CLOSE */
739 kgn_conn_state_t gnc_state; /* connection state */
740 int gnc_scheduled; /* being attented to */
741 char gnc_sched_caller[30]; /* what function last called schedule */
742 int gnc_sched_line; /* what line # last called schedule */
743 atomic_t gnc_refcount; /* # users */
744 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
745 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
746 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
747 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
748 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
749 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
750 int gnc_next_tx; /* next tx to use in tx_ref_table */
751 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
752 int gnc_mbox_id; /* id of mbox in fma_blk */
753 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
754 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
755 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
756 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
757 void *remote_mbox_addr; /* save remote mbox address */
760 typedef struct kgn_mdd_purgatory {
761 gni_mem_handle_t gmp_map_key; /* mapping key */
762 struct list_head gmp_list; /* entry point for purgatory list */
763 } kgn_mdd_purgatory_t;
765 typedef struct kgn_peer {
766 struct list_head gnp_list; /* stash on global peer list */
767 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
768 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
769 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
770 kgn_net_t *gnp_net; /* net instance for this peer */
771 lnet_nid_t gnp_nid; /* who's on the other end(s) */
772 atomic_t gnp_refcount; /* # users */
773 __u32 gnp_host_id; /* ph. host ID of the peer */
774 short gnp_connecting; /* connection forming */
775 short gnp_pending_unlink; /* need last conn close to trigger unlink */
776 int gnp_last_errno; /* last error conn saw */
777 unsigned long gnp_last_alive; /* last time I had valid comms */
778 int gnp_last_dgram_errno; /* last error dgrams saw */
779 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
780 unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
781 unsigned long gnp_reconnect_interval; /* exponential backoff */
782 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
783 int gnp_down; /* rca says peer down */
784 unsigned long gnp_down_event_time; /* time peer down */
785 unsigned long gnp_up_event_time; /* time peer back up */
788 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
789 * like lnet_parse. It allows a single pointer to let us get enough
790 * information in _recv and friends */
791 typedef struct kgn_rx {
792 kgn_conn_t *grx_conn; /* connection */
793 kgn_msg_t *grx_msg; /* message */
794 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
795 int grx_eager; /* if eager, we copied msg to somewhere */
796 struct timespec grx_received; /* time this msg received */
799 typedef struct kgn_data {
800 int kgn_init; /* initialisation state */
801 int kgn_shutdown; /* shut down? */
802 int kgn_wc_kill; /* Should I repost the WC */
803 atomic_t kgn_nthreads; /* # live threads */
804 int kgn_nresets; /* number of stack resets */
805 int kgn_in_reset; /* are we in stack reset ? */
807 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
809 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
810 int kgn_ndevs; /* # devices */
812 int kgn_ruhroh_running; /* ruhroh thread is running */
813 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
814 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
815 int kgn_quiesce_trigger; /* should we quiesce ? */
816 atomic_t kgn_nquiesce; /* how many quiesced ? */
817 struct mutex kgn_quiesce_mutex; /* serialize ruhroh task, startup and shutdown */
818 int kgn_needs_reset; /* we need stack reset */
820 /* These next three members implement communication from gnilnd into
821 * the ruhroh task. To ensure correct operation of the task, code that
822 * writes into them must use memory barriers to ensure that the changes
823 * are visible to other cores in the order the members appear below. */
824 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
825 int kgn_bump_info_rdy; /* we have info needed to bump */
826 int kgn_needs_pause; /* we need to pause for network quiesce */
828 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
829 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
831 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
832 struct list_head *kgn_peers; /* hash table of all my known peers */
833 atomic_t kgn_npeers; /* # peers extant */
834 int kgn_peer_version; /* version flag for peer tables */
836 struct list_head *kgn_conns; /* conns hashed by cqid */
837 atomic_t kgn_nconns; /* # connections extant */
838 atomic_t kgn_neager_allocs; /* # of eager allocations */
839 __u64 kgn_peerstamp; /* when I started up */
840 __u64 kgn_connstamp; /* conn stamp generator */
841 int kgn_conn_version; /* version flag for conn tables */
842 int kgn_next_cqid; /* cqid generator */
844 long kgn_new_min_timeout; /* minimum timeout on any new conn */
845 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
846 spinlock_t kgn_reaper_lock; /* serialise */
848 struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
849 struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
850 struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
851 atomic_t kgn_ntx; /* # tx in use */
852 struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
854 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
855 __u64 kgn_cksum_npages; /* # pages alloc'd for checksumming */
856 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
857 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
859 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
860 long kgn_rdmaq_override; /* bytes per second override */
862 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
864 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
865 atomic_t kgn_npending_conns; /* # of conns with pending closes */
866 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
867 unsigned long kgn_last_scheduled; /* last time schedule was called */
868 unsigned long kgn_last_condresched; /* last time cond_resched was called */
869 atomic_t kgn_rev_offset; /* # of REV rdma w/misaligned offsets */
870 atomic_t kgn_rev_length; /* # of REV rdma have misaligned len */
871 atomic_t kgn_rev_copy_buff; /* # of REV rdma buffer copies */
872 struct socket *kgn_sock; /* for Apollo */
873 unsigned long free_pages_limit; /* # of free pages reserve from fma block allocations */
874 int kgn_enable_gl_mutex; /* kgni api mtx enable */
877 extern kgn_data_t kgnilnd_data;
878 extern kgn_tunables_t kgnilnd_tunables;
880 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
881 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
882 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
884 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
885 * and the line of the calling function to allow us to debug problematic
886 * schedule calls in the future without the programmer having to mark
887 * the location manually.
889 #define kgnilnd_schedule_conn(conn) \
890 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
892 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
893 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
896 kgnilnd_thread_fini(void)
898 atomic_dec(&kgnilnd_data.kgn_nthreads);
901 static inline void kgnilnd_gl_mutex_lock(struct mutex *lock)
903 if (kgnilnd_data.kgn_enable_gl_mutex)
907 static inline void kgnilnd_gl_mutex_unlock(struct mutex *lock)
909 if (kgnilnd_data.kgn_enable_gl_mutex)
913 static inline void kgnilnd_conn_mutex_lock(struct mutex *lock)
915 if (!kgnilnd_data.kgn_enable_gl_mutex)
919 static inline void kgnilnd_conn_mutex_unlock(struct mutex *lock)
921 if (!kgnilnd_data.kgn_enable_gl_mutex)
925 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
926 * parts of the code to avoid a scheduler trip when the mutex is held
928 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
929 * has been acquired successfully, and 0 on contention.
931 * NOTE: this function follows the spin_trylock() convention, so
932 * it is negated to the down_trylock() return values! Be careful
933 * about this when converting semaphore users to mutexes.
935 * This function must not be used in interrupt context. The
936 * mutex must be released by the same task that acquired it.
938 static inline int __kgnilnd_mutex_trylock(struct mutex *lock)
941 unsigned long timeout;
943 LASSERT(!in_interrupt());
945 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
947 ret = mutex_trylock(lock);
954 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
956 if (!kgnilnd_data.kgn_enable_gl_mutex)
959 return __kgnilnd_mutex_trylock(lock);
962 static inline int kgnilnd_trylock(struct mutex *cq_lock,
963 struct mutex *c_lock)
965 if (kgnilnd_data.kgn_enable_gl_mutex)
966 return __kgnilnd_mutex_trylock(cq_lock);
968 return __kgnilnd_mutex_trylock(c_lock);
971 static inline void *kgnilnd_vzalloc(int size)
973 void *ret = __vmalloc(size, __GFP_HIGHMEM | GFP_NOFS | __GFP_ZERO,
975 LIBCFS_ALLOC_POST(ret, size);
979 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
982 _kgnilnd_debug_msg(kgn_msg_t *msg,
983 struct libcfs_debug_msg_data *data, const char *fmt, ... );
985 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
987 CFS_CHECK_STACK(msgdata, mask, cdls); \
989 if (((mask) & D_CANTMASK) != 0 || \
990 ((libcfs_debug & (mask)) != 0 && \
991 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
992 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
995 /* for most callers (level is a constant) this is resolved at compile time */
996 #define GNIDBG_MSG(level, msg, fmt, args...) \
998 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
999 static cfs_debug_limit_state_t cdls; \
1000 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1001 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1002 "$$ "fmt" from %s ", ## args, \
1003 libcfs_nid2str((msg)->gnm_srcnid)); \
1005 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1006 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1007 "$$ "fmt" from %s ", ## args, \
1008 libcfs_nid2str((msg)->gnm_srcnid)); \
1012 /* user puts 'to nid' in msg for us */
1013 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
1015 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1016 static cfs_debug_limit_state_t cdls; \
1017 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1018 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1019 "$$ "fmt" ", ## args); \
1021 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1022 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1023 "$$ "fmt" ", ## args); \
1028 _kgnilnd_debug_conn(kgn_conn_t *conn,
1029 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1031 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
1033 CFS_CHECK_STACK(msgdata, mask, cdls); \
1035 if (((mask) & D_CANTMASK) != 0 || \
1036 ((libcfs_debug & (mask)) != 0 && \
1037 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1038 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
1041 /* for most callers (level is a constant) this is resolved at compile time */
1042 #define GNIDBG_CONN(level, conn, fmt, args...) \
1044 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1045 static cfs_debug_limit_state_t cdls; \
1046 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1047 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
1048 "$$ "fmt" ", ## args); \
1050 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1051 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
1052 "$$ "fmt" ", ## args); \
1057 _kgnilnd_debug_tx(kgn_tx_t *tx,
1058 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1060 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
1062 CFS_CHECK_STACK(msgdata, mask, cdls); \
1064 if (((mask) & D_CANTMASK) != 0 || \
1065 ((libcfs_debug & (mask)) != 0 && \
1066 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1067 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1070 /* for most callers (level is a constant) this is resolved at compile time */
1071 #define GNIDBG_TX(level, tx, fmt, args...) \
1073 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1074 static cfs_debug_limit_state_t cdls; \
1075 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1076 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1077 "$$ "fmt" ", ## args); \
1079 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1080 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1081 "$$ "fmt" ", ## args); \
1085 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1087 if (unlikely(!(cond))) { \
1088 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1093 #define GNILND_IS_QUIESCED \
1094 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1095 atomic_read(&kgnilnd_data.kgn_nthreads))
1097 #define KGNILND_SPIN_QUIESCE \
1099 /* E.T phone home */ \
1100 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1101 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1102 while (kgnilnd_data.kgn_quiesce_trigger) { \
1103 msleep_interruptible(MSEC_PER_SEC); \
1105 /* Mom, my homework is done */ \
1106 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1107 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1110 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1111 #ifndef LIBCFS_DEBUG
1112 #error "this code uses actions inside LASSERT for ref counting"
1115 #define kgnilnd_admin_addref(atomic) \
1117 int val = atomic_inc_return(&atomic); \
1118 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1119 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1122 #define kgnilnd_admin_decref(atomic) \
1124 int val = atomic_dec_return(&atomic); \
1125 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
1126 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1129 #define kgnilnd_net_addref(net) \
1131 int val = atomic_inc_return(&net->gnn_refcount); \
1132 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1133 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1134 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1137 #define kgnilnd_net_decref(net) \
1139 int val = atomic_dec_return(&net->gnn_refcount); \
1140 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1141 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1142 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1145 #define kgnilnd_peer_addref(peer) \
1147 int val = atomic_inc_return(&peer->gnp_refcount); \
1148 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1149 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1150 libcfs_nid2str(peer->gnp_nid), val); \
1153 #define kgnilnd_peer_decref(peer) \
1155 int val = atomic_dec_return(&peer->gnp_refcount); \
1156 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1157 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1158 libcfs_nid2str(peer->gnp_nid), val); \
1160 kgnilnd_destroy_peer(peer); \
1163 #define kgnilnd_conn_addref(conn) \
1168 val = atomic_inc_return(&conn->gnc_refcount); \
1169 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1170 "conn %p refc %d to %s\n", \
1173 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1175 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1177 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1182 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1183 * through the scheduler thread to get the EP destroyed. This avoids some
1184 * messy semaphore business and allows us to reuse the connd_list and existing
1185 * linkage and avoid creating extra lists just for destroying EPs */
1187 /* Safety Disclaimer:
1188 * Q: If we decrement the refcount and then check it again, is it possible that
1189 * another caller could have passed through this macro concurrently? If so,
1190 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1192 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1193 * once we are refcount <= 2. It hinges around gnc_state and membership of
1194 * gnc_hashlist. There are two ways to find a connection - either ask for
1195 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1196 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1199 * - #1 from create (kgnilnd_create_conn)
1200 * - #2 for EP (kgnilnd_create_conn)
1201 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1202 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1204 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1205 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1207 * There could be more from TX descriptors during the lifetime of a live
1210 * If we nuke the conn before finish_connect, we won't have parallel paths
1211 * because nobody besides the dgram handler for the single outstanding
1212 * dgram can find the connection as it isn't in any searchable tables yet.
1214 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1215 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1216 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1217 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1218 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1219 * sent. Now the gnc_state == CLOSED, so we hit
1220 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1221 * - we've nuked them from the peer and CQ id tables, so we own them and
1222 * are guaranteed serial access - hence the complete lack of conn list
1223 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1224 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1225 * kgnilnd_process_conns. Then the next scheduler loop would call
1226 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1227 * kgnilnd_process_conns.
1229 * Clearly, we are totally safe. Clearly.
1232 #define kgnilnd_conn_decref(conn) \
1237 val = atomic_dec_return(&conn->gnc_refcount); \
1238 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1241 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1243 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1245 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1250 (conn->gnc_ephandle != NULL) && \
1251 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1252 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1253 kgnilnd_schedule_conn(conn); \
1254 } else if (val == 0) { \
1255 kgnilnd_destroy_conn(conn); \
1259 static inline struct list_head *
1260 kgnilnd_nid2peerlist(lnet_nid_t nid)
1262 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1264 RETURN(&kgnilnd_data.kgn_peers[hash]);
1267 static inline struct list_head *
1268 kgnilnd_netnum2netlist(__u16 netnum)
1270 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1272 RETURN(&kgnilnd_data.kgn_nets[hash]);
1276 kgnilnd_peer_active(kgn_peer_t *peer)
1278 /* Am I in the peer hash table? */
1279 return (!list_empty(&peer->gnp_list));
1282 /* need write_lock on kgn_peer_conn_lock */
1284 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1286 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1287 peer, libcfs_nid2str(peer->gnp_nid),
1288 !list_empty(&peer->gnp_conns),
1289 !list_empty(&peer->gnp_tx_queue));
1291 /* kgn_peer_conn_lock protects us from conflict with
1292 * kgnilnd_peer_notify and gnp_persistent */
1293 RETURN ((list_empty(&peer->gnp_conns)) &&
1294 (list_empty(&peer->gnp_tx_queue)));
1297 /* returns positive if error was for a clean shutdown of conn */
1299 kgnilnd_conn_clean_errno(int errno)
1301 /* - ESHUTDOWN - LND is unloading
1302 * - EUCLEAN - admin requested via "lctl del_peer"
1303 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1304 * - ENOTRECOVERABLE - stack reset
1305 * - EISCONN - cleared via "lctl push"
1306 * not doing ESTALE - that isn't clean */
1307 RETURN ((errno == 0) ||
1308 (errno == -ESHUTDOWN) ||
1309 (errno == -EUCLEAN) ||
1310 (errno == -ENETRESET) ||
1311 (errno == -EISCONN) ||
1312 (errno == -ENOTRECOVERABLE));
1315 /* returns positive if error results in purgatory hold */
1317 kgnilnd_check_purgatory_errno(int errno)
1319 /* We don't want to save the purgatory lists these cases:
1320 * - EUCLEAN - admin requested via "lctl del_peer"
1321 * - ESHUTDOWN - LND is unloading
1323 RETURN ((errno != -ESHUTDOWN) &&
1324 (errno != -EUCLEAN));
1328 /* returns positive if a purgatory hold is needed */
1330 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1334 if (conn->gnc_peer) {
1335 loopback = conn->gnc_peer->gnp_nid ==
1336 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1338 /* short circuit - a conn that didn't complete
1339 * setup never needs a purgatory hold */
1342 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1343 conn, conn->gnc_peer ?
1344 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1346 loopback, conn->gnc_close_recvd);
1348 /* we only use a purgatory hold if we've not received the CLOSE msg
1349 * from our peer - without that message, we can't know the state of
1350 * the other end of this connection and must put it into purgatory
1351 * to prevent reuse and corruption.
1352 * The theory is that a TX error can be communicated in all other cases
1354 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1355 kgnilnd_check_purgatory_errno(conn->gnc_error));
1358 static inline const char *
1359 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1361 static inline struct list_head *
1362 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1363 kgn_tx_list_state_t to_state)
1366 case GNILND_TX_PEERQ:
1367 return &peer->gnp_tx_queue;
1368 case GNILND_TX_FMAQ:
1369 return &conn->gnc_fmaq;
1370 case GNILND_TX_LIVE_FMAQ:
1371 case GNILND_TX_LIVE_RDMAQ:
1372 case GNILND_TX_DYING:
1374 case GNILND_TX_MAPQ:
1375 return &conn->gnc_device->gnd_map_tx;
1376 case GNILND_TX_RDMAQ:
1377 return &conn->gnc_device->gnd_rdmaq;
1379 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1380 CERROR("invalid state requested: %s\n",
1381 kgnilnd_tx_state2str(to_state));
1387 /* should hold tx, conn or peer lock when calling */
1389 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1390 kgn_conn_t *conn, kgn_tx_list_state_t state,
1393 struct list_head *list = NULL;
1395 /* make sure we have a sane TX state to start */
1396 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1397 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1398 list_empty(&tx->tx_list),
1399 "bad state with tx_list %s",
1400 list_empty(&tx->tx_list) ? "empty" : "not empty");
1402 /* WTF - you are already on that state buttmunch */
1403 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1404 "already at %s", kgnilnd_tx_state2str(state));
1406 /* get proper list from the state requested */
1407 list = kgnilnd_tx_state2list(peer, conn, state);
1411 case GNILND_TX_PEERQ:
1412 kgnilnd_peer_addref(peer);
1414 case GNILND_TX_ALLOCD:
1415 /* no refs needed */
1417 case GNILND_TX_FMAQ:
1418 kgnilnd_conn_addref(conn);
1420 case GNILND_TX_MAPQ:
1421 atomic_inc(&conn->gnc_device->gnd_nq_map);
1422 kgnilnd_conn_addref(conn);
1424 case GNILND_TX_LIVE_FMAQ:
1425 atomic_inc(&conn->gnc_nlive_fma);
1426 kgnilnd_conn_addref(conn);
1428 case GNILND_TX_LIVE_RDMAQ:
1429 atomic_inc(&conn->gnc_nlive_rdma);
1430 kgnilnd_conn_addref(conn);
1432 case GNILND_TX_RDMAQ:
1433 atomic_inc(&conn->gnc_nq_rdma);
1434 kgnilnd_conn_addref(conn);
1436 case GNILND_TX_DYING:
1437 kgnilnd_conn_addref(conn);
1440 CERROR("invalid state requested: %s\n",
1441 kgnilnd_tx_state2str(state));
1446 /* if this changes, change kgnilnd_alloc_tx */
1447 tx->tx_list_state = state;
1449 /* some states don't have lists - we track them in the per conn
1450 * TX table instead. Waste not, want not! */
1452 tx->tx_list_p = list;
1454 list_add_tail(&tx->tx_list, list);
1456 list_add(&tx->tx_list, list);
1458 /* set dummy list_p to make book keeping happy and let debugging
1459 * be a hair easier */
1460 tx->tx_list_p = (void *)state;
1463 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1464 kgnilnd_tx_state2str(state), list);
1468 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1469 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1471 /* These is only 1 "off-list" state */
1472 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1473 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1475 /* new_state == ALLOCD means we are deallocating this tx,
1476 * so make sure it was on a valid list to start with */
1477 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1478 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1479 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1480 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1483 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1486 switch (tx->tx_list_state) {
1487 case GNILND_TX_PEERQ:
1488 kgnilnd_peer_decref(peer);
1490 case GNILND_TX_FREED:
1491 case GNILND_TX_IDLE:
1492 case GNILND_TX_ALLOCD:
1493 /* no refs needed */
1495 case GNILND_TX_DYING:
1496 kgnilnd_conn_decref(conn);
1498 case GNILND_TX_FMAQ:
1499 kgnilnd_conn_decref(conn);
1501 case GNILND_TX_MAPQ:
1502 atomic_dec(&conn->gnc_device->gnd_nq_map);
1503 kgnilnd_conn_decref(conn);
1505 case GNILND_TX_LIVE_FMAQ:
1506 atomic_dec(&conn->gnc_nlive_fma);
1507 kgnilnd_conn_decref(conn);
1509 case GNILND_TX_LIVE_RDMAQ:
1510 atomic_dec(&conn->gnc_nlive_rdma);
1511 kgnilnd_conn_decref(conn);
1513 case GNILND_TX_RDMAQ:
1514 atomic_dec(&conn->gnc_nq_rdma);
1515 kgnilnd_conn_decref(conn);
1516 /* don't need to assert on default, already did in set */
1519 /* for ALLOCD, this might already be true, but no harm doing it again */
1520 list_del_init(&tx->tx_list);
1521 tx->tx_list_p = NULL;
1522 tx->tx_list_state = new_state;
1526 kgnilnd_tx_mapped(kgn_tx_t *tx)
1528 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1529 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1532 static inline struct list_head *
1533 kgnilnd_cqid2connlist(__u32 cqid)
1535 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1537 return (&kgnilnd_data.kgn_conns [hash]);
1540 static inline kgn_conn_t *
1541 kgnilnd_cqid2conn_locked(__u32 cqid)
1543 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1544 struct list_head *tmp;
1547 list_for_each(tmp, conns) {
1548 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1550 if (conn->gnc_cqid == cqid)
1557 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1559 kgnilnd_get_cqid_locked(void)
1565 cqid = kgnilnd_data.kgn_next_cqid++;
1566 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1570 kgnilnd_data.kgn_next_cqid = 1;
1573 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1579 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1581 kgn_tx_t *tx = NULL;
1582 kgn_conn_t *conn = NULL;
1584 /* set to NULL so any early return is an error */
1588 LASSERTF((ev_id->txe_idx > 0) &&
1589 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1590 "bogus txe_idx %d >= %d\n",
1591 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1593 LASSERTF((ev_id->txe_cqid > 0) &&
1594 (ev_id->txe_cqid < GNILND_MAX_CQID),
1595 "bogus txe_cqid %d >= %d\n",
1596 ev_id->txe_cqid, GNILND_MAX_CQID);
1598 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1599 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1602 /* Conn was destroyed? */
1603 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1604 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1607 /* just insurance */
1608 kgnilnd_conn_addref(conn);
1609 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1610 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1612 /* we know this is safe - as the TX won't be reused until AFTER
1613 * the conn is unlinked from the cqid hash, so we can use the TX
1614 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1616 spin_lock(&conn->gnc_tx_lock);
1617 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1618 spin_unlock(&conn->gnc_tx_lock);
1620 /* We could have a tx that was cleared out by other forces
1621 * lctl disconnect or del_peer. */
1623 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1624 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1625 kgnilnd_conn_decref(conn);
1629 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1630 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1631 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1633 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1634 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1635 "found tx %p txe_idx %d",
1636 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1637 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1639 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1641 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1646 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1649 /* set_normalized_timepsec isn't exported from the kernel, so
1650 * we need to do the same thing inline */
1651 static inline struct timespec
1652 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1658 sec = lhs.tv_sec - rhs.tv_sec;
1659 nsec = lhs.tv_nsec - rhs.tv_nsec;
1661 while (nsec >= NSEC_PER_SEC) {
1662 nsec -= NSEC_PER_SEC;
1666 nsec += NSEC_PER_SEC;
1675 kgnilnd_count_list(struct list_head *q)
1677 struct list_head *e;
1680 list_for_each(e, q) {
1687 /* kgnilnd_find_net adds a reference to the net it finds
1688 * this is so the net will not be removed before the calling function
1689 * has time to use the data returned. This reference needs to be released
1690 * by the calling function once it has finished using the returned net
1694 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1699 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1705 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1706 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1707 kgnilnd_net_addref(net);
1708 up_read(&kgnilnd_data.kgn_net_rw_sem);
1714 up_read(&kgnilnd_data.kgn_net_rw_sem);
1719 #ifdef CONFIG_DEBUG_SLAB
1720 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1722 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1725 int kgnilnd_dev_init(kgn_device_t *dev);
1726 void kgnilnd_dev_fini(kgn_device_t *dev);
1727 int kgnilnd_startup(lnet_ni_t *ni);
1728 void kgnilnd_shutdown(lnet_ni_t *ni);
1729 int kgnilnd_base_startup(void);
1730 void kgnilnd_base_shutdown(void);
1732 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1733 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1734 void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
1735 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1737 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1738 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1739 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1740 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1741 lnet_msg_t *lntmsg, void **new_private);
1742 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1743 int delayed, unsigned int niov,
1744 struct iovec *iov, lnet_kiov_t *kiov,
1745 unsigned int offset, unsigned int mlen, unsigned int rlen);
1747 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1749 /* purgatory functions */
1750 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1751 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1752 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1753 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1755 void kgnilnd_update_reaper_timeout(long timeout);
1756 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1757 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1758 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1759 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1760 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1761 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
1762 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1764 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1765 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
1766 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1767 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1769 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1770 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1771 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1772 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1773 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1774 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1775 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1776 void kgnilnd_consume_rx(kgn_rx_t *rx);
1778 void kgnilnd_schedule_device(kgn_device_t *dev);
1779 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1780 void kgnilnd_schedule_device_timer(unsigned long arg);
1782 int kgnilnd_reaper(void *arg);
1783 int kgnilnd_scheduler(void *arg);
1784 int kgnilnd_dgram_mover(void *arg);
1785 int kgnilnd_rca(void *arg);
1786 int kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id);
1788 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1789 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1790 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1791 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1792 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1793 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1794 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1795 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1796 void kgnilnd_peer_alive(kgn_peer_t *peer);
1797 void kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive);
1798 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1799 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1800 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1801 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1803 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1804 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1805 void kgnilnd_wakeup_rca_thread(void);
1806 int kgnilnd_start_rca_thread(void);
1807 int kgnilnd_get_node_state(__u32 nid);
1809 int kgnilnd_tunables_init(void);
1810 void kgnilnd_tunables_fini(void);
1811 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1813 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1814 void kgnilnd_pause_threads(void);
1815 int kgnilnd_hw_in_quiesce(void);
1816 int kgnilnd_check_hw_quiesce(void);
1817 void kgnilnd_quiesce_wait(char *reason);
1818 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1819 int kgnilnd_ruhroh_thread(void *arg);
1820 void kgnilnd_reset_stack(void);
1821 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1823 void kgnilnd_insert_sysctl(void);
1824 void kgnilnd_remove_sysctl(void);
1825 void kgnilnd_proc_init(void);
1826 void kgnilnd_proc_fini(void);
1829 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1831 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1832 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1833 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
1835 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1836 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1837 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1838 int kgnilnd_cancel_dgrams(kgn_device_t *dev);
1839 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1841 int kgnilnd_dgram_waitq(void *arg);
1843 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1845 /* struct2str functions - we don't use a default: case to cause the compile
1846 * to fail if there is a missing case. This allows us to hide these down here
1847 * out of the way but ensure we'll catch any updates to the enum/types
1850 #define DO_TYPE(x) case x: return #x;
1851 static inline const char *
1852 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1854 /* Only want single char string for this */
1856 case GNILND_FMABLK_IDLE:
1858 case GNILND_FMABLK_PHYS:
1860 case GNILND_FMABLK_VIRT:
1862 case GNILND_FMABLK_FREED:
1865 return "<unknown state>";
1868 static inline const char *
1869 kgnilnd_msgtype2str(int type)
1872 DO_TYPE(GNILND_MSG_NONE);
1873 DO_TYPE(GNILND_MSG_NOOP);
1874 DO_TYPE(GNILND_MSG_IMMEDIATE);
1875 DO_TYPE(GNILND_MSG_PUT_REQ);
1876 DO_TYPE(GNILND_MSG_PUT_NAK);
1877 DO_TYPE(GNILND_MSG_PUT_ACK);
1878 DO_TYPE(GNILND_MSG_PUT_DONE);
1879 DO_TYPE(GNILND_MSG_GET_REQ);
1880 DO_TYPE(GNILND_MSG_GET_NAK);
1881 DO_TYPE(GNILND_MSG_GET_DONE);
1882 DO_TYPE(GNILND_MSG_CLOSE);
1883 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1884 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1885 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1886 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1887 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1888 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1889 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1891 return "<unknown msg type>";
1894 static inline const char *
1895 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1898 DO_TYPE(GNILND_TX_IDLE);
1899 DO_TYPE(GNILND_TX_ALLOCD);
1900 DO_TYPE(GNILND_TX_PEERQ);
1901 DO_TYPE(GNILND_TX_MAPQ);
1902 DO_TYPE(GNILND_TX_FMAQ);
1903 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1904 DO_TYPE(GNILND_TX_RDMAQ);
1905 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1906 DO_TYPE(GNILND_TX_DYING);
1907 DO_TYPE(GNILND_TX_FREED);
1909 return "<unknown state>";
1912 static inline const char *
1913 kgnilnd_conn_state2str(kgn_conn_t *conn)
1915 kgn_conn_state_t state = conn->gnc_state;
1917 DO_TYPE(GNILND_CONN_DUMMY);
1918 DO_TYPE(GNILND_CONN_LISTEN);
1919 DO_TYPE(GNILND_CONN_CONNECTING);
1920 DO_TYPE(GNILND_CONN_ESTABLISHED);
1921 DO_TYPE(GNILND_CONN_CLOSING);
1922 DO_TYPE(GNILND_CONN_CLOSED);
1923 DO_TYPE(GNILND_CONN_DONE);
1924 DO_TYPE(GNILND_CONN_DESTROY_EP);
1929 static inline const char *
1930 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1932 kgn_connreq_type_t type = connreq->gncr_type;
1935 DO_TYPE(GNILND_CONNREQ_REQ);
1936 DO_TYPE(GNILND_CONNREQ_NAK);
1937 DO_TYPE(GNILND_CONNREQ_CLOSE);
1942 static inline const char *
1943 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1945 kgn_dgram_state_t state = dgram->gndg_state;
1948 DO_TYPE(GNILND_DGRAM_USED);
1949 DO_TYPE(GNILND_DGRAM_POSTING);
1950 DO_TYPE(GNILND_DGRAM_POSTED);
1951 DO_TYPE(GNILND_DGRAM_PROCESSING);
1952 DO_TYPE(GNILND_DGRAM_DONE);
1953 DO_TYPE(GNILND_DGRAM_CANCELED);
1958 static inline const char *
1959 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1961 kgn_dgram_type_t type = dgram->gndg_type;
1964 DO_TYPE(GNILND_DGRAM_REQ);
1965 DO_TYPE(GNILND_DGRAM_WC_REQ);
1966 DO_TYPE(GNILND_DGRAM_NAK);
1967 DO_TYPE(GNILND_DGRAM_CLOSE);
1972 static inline const char *
1973 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
1976 DO_TYPE(GNILND_DGRAM_REQ);
1977 DO_TYPE(GNILND_DGRAM_WC_REQ);
1978 DO_TYPE(GNILND_DGRAM_NAK);
1979 DO_TYPE(GNILND_DGRAM_CLOSE);
1986 /* pulls in tunables per platform and adds in nid/nic conversion
1987 * if RCA wasn't available at build time */
1988 #include "gnilnd_hss_ops.h"
1989 /* API wrapper functions - include late to pick up all of the other defines */
1990 #include "gnilnd_api_wrap.h"
1992 #if defined(CONFIG_CRAY_GEMINI)
1993 #include "gnilnd_gemini.h"
1994 #elif defined(CONFIG_CRAY_ARIES)
1995 #include "gnilnd_aries.h"
1997 #error "Undefined Network Hardware Type"
2000 extern uint32_t kgni_driver_version;
2003 kgnilnd_check_kgni_version(void)
2007 kgnilnd_data.kgn_enable_gl_mutex = 1;
2008 kdv = symbol_get(kgni_driver_version);
2010 LCONSOLE_INFO("Not using thread safe locking -"
2011 " no symbol kgni_driver_version\n");
2015 /* Thread-safe kgni implemented in minor ver 0x44/45, code rev 0xb9 */
2016 if (*kdv < GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9)) {
2017 symbol_put(kgni_driver_version);
2018 LCONSOLE_INFO("Not using thread safe locking, gni version 0x%x,"
2019 " need >= 0x%x\n", *kdv,
2020 GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9));
2024 symbol_put(kgni_driver_version);
2026 if (!*kgnilnd_tunables.kgn_thread_safe) {
2030 /* Use thread-safe locking */
2031 kgnilnd_data.kgn_enable_gl_mutex = 0;
2034 #endif /* _GNILND_GNILND_H_ */