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 #include <linux/module.h>
32 #include <linux/kernel.h>
34 #include <linux/string.h>
35 #include <linux/stat.h>
36 #include <linux/errno.h>
37 #ifdef HAVE_LINUX_KERNEL_LOCK
38 #include <linux/smp_lock.h>
40 #include <linux/unistd.h>
41 #include <linux/uio.h>
42 #include <linux/time.h>
43 #include <asm/timex.h>
45 #include <asm/system.h>
46 #include <asm/uaccess.h>
49 #include <linux/init.h>
51 #include <linux/file.h>
52 #include <linux/stat.h>
53 #include <linux/list.h>
54 #include <linux/kmod.h>
55 #include <linux/sysctl.h>
56 #include <linux/kthread.h>
57 #include <linux/nmi.h>
61 #include <linux/nmi.h>
63 #define DEBUG_SUBSYSTEM S_LND
65 #include <libcfs/linux/kp30.h>
66 #include <libcfs/libcfs.h>
67 #include <lnet/lnet.h>
68 #include <lnet/lib-lnet.h>
71 #include "gnilnd_version.h"
74 /* tunables determined at compile time */
75 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
76 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
77 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
78 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
79 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
80 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
81 #define GNILND_SCHED_TIMEOUT 1
82 #define GNILND_DGRAM_TIMEOUT 2
83 #define GNILND_FAST_MAPPING_TRY \
84 *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
85 #define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
87 /* map failure timeout */
88 #define GNILND_MAP_TIMEOUT \
89 (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
90 *kgnilnd_tunables.kgn_timeout))
92 /* reaper thread wakup interval */
93 #define GNILND_REAPER_THREAD_WAKE 1
94 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
95 #define GNILND_REAPER_NCHECKS 4
98 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
99 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
100 #define GNILND_COOKIE 0xa3579 /* cookie used by along with ptag by GNI */
101 #define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
102 /* checksum values */
103 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
104 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
105 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
106 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
108 /* tune down some COMPUTE options as they won't see the same number of connections and
109 * don't need the throughput of multiple threads by default */
110 #if defined(CONFIG_CRAY_COMPUTE)
112 #define GNILND_SCHED_THREADS 2 /* default # of kgnilnd_scheduler threads */
114 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
116 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
117 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
118 #define GNILND_COMPUTE 1 /* compute image */
120 #define GNILND_SCHED_THREADS 3 /* default # of kgnilnd_scheduler threads */
121 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
122 #define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
123 #define GNILND_COMPUTE 0 /* service image */
126 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
127 #define GNILND_EXTRA_BITS 1
128 /* maximum number of conns & bits for cqid in the SMSG event data */
129 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
130 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
131 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
132 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
133 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
135 /* need sane upper bound to limit copy overhead */
136 #define GNILND_MAX_IMMEDIATE (64<<10)
138 /* payload size to add to the base mailbox size
139 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
140 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
141 * the calculation return from that function.*/
142 #define GNILND_MBOX_PAYLOAD \
143 (GNILND_MAX_MSG_SIZE * \
144 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
146 /* timeout -> deadman timer for kgni mdd holds */
147 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
149 /* timeout for failing sends in t is in jiffies*/
150 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
152 /* time when to release from purgatory in the reaper thread in jiffies */
153 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
155 /* Macro for finding last_rx 2 datapoints are compared
156 * and the most recent one in jiffies is returned.
158 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
159 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
161 /************************************************************************
162 * Enum, flag and tag data
164 #define GNILND_INIT_NOTHING 0
165 #define GNILND_INIT_DATA 1
166 #define GNILND_INIT_ALL 2
168 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
169 #define GNILND_BUF_NONE 0 /* buffer type not set */
170 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
171 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
172 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
173 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
174 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
175 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
177 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
178 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
179 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
180 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
181 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
183 /* stash above max CQID to avoid any collision */
184 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
185 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
187 /* kgn_msg_t::gnm_type */
188 #define GNILND_MSG_NONE 0x00 /* illegal message */
189 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
190 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
191 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
192 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
193 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
194 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
195 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
196 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
197 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
198 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
199 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
200 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
201 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
202 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
203 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
204 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
205 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
207 /* defines for gnc_*scheduled states */
208 #define GNILND_CONN_IDLE 0
209 #define GNILND_CONN_SCHED 1
210 #define GNILND_CONN_WANTS_SCHED 2
211 #define GNILND_CONN_PROCESS 3
213 #define GNILND_DEV_IDLE 0
214 #define GNILND_DEV_IRQ 1
215 #define GNILND_DEV_LOOP 2
217 #define GNILND_DGRAM_IDLE 0
218 #define GNILND_DGRAM_SCHED 1
219 #define GNILND_DGRAM_PROCESS 2
221 #define GNILND_PEER_IDLE 0
222 #define GNILND_PEER_CONNECT 1
223 #define GNILND_PEER_POSTING 2
224 #define GNILND_PEER_POSTED 3
225 #define GNILND_PEER_NEEDS_DEATH 4
226 #define GNILND_PEER_KILL 5
228 /* for gnc_close_recvd */
229 #define GNILND_CLOSE_RX 1
230 #define GNILND_CLOSE_INJECT1 2
231 #define GNILND_CLOSE_INJECT2 3
232 #define GNILND_CLOSE_EARLY 4
234 /* defines for why quiesce trigger set */
235 #define GNILND_QUIESCE_IDLE 0
236 #define GNILND_QUIESCE_ADMIN 1
237 #define GNILND_QUIESCE_RESET 2
238 #define GNILND_QUIESCE_HW_QUIESCE 3
240 #define GNILND_PEER_CLEAN 0
241 #define GNILND_PEER_PERSISTING 1
243 #define GNILND_DEL_CONN 0
244 #define GNILND_DEL_PEER 1
245 #define GNILND_CLEAR_PURGATORY 2
247 #define GNILND_RCA_NODE_UP 0
248 #define GNILND_RCA_NODE_DOWN 1
249 #define GNILND_RCA_NODE_UNKNOWN 2
251 /* defines for reverse RDMA states */
252 #define GNILND_REVERSE_NONE 0
253 #define GNILND_REVERSE_GET 1
254 #define GNILND_REVERSE_PUT 2
255 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
257 typedef enum kgn_fmablk_state {
258 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
259 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
260 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
261 GNILND_FMABLK_FREED, /* after free */
262 } kgn_fmablk_state_t;
264 typedef enum kgn_tx_list_state {
265 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
266 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
267 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
268 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
269 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
270 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
271 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
272 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
273 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
274 GNILND_TX_FREED /* TX is free! */
275 } kgn_tx_list_state_t;
277 typedef enum kgn_conn_state {
278 /* don't start @ 0 - prevent memset(0) badness */
279 GNILND_CONN_DUMMY = 0,
281 GNILND_CONN_CONNECTING,
282 GNILND_CONN_ESTABLISHED,
286 GNILND_CONN_DESTROY_EP
289 /* changing these requires a change to GNILND_CONNREQ_VERSION and
290 * will result in dropped packets instead of NAKs. Adding to this is
291 * acceptable without changing the CONNREQ_VERSION, but code should
292 * be ready to handle NAKs on version mismatch */
293 typedef enum kgn_connreq_type {
294 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
295 GNILND_CONNREQ_NAK, /* NO soup for you! */
296 GNILND_CONNREQ_CLOSE, /* we should see other people */
297 } kgn_connreq_type_t;
299 typedef enum kgn_dgram_state {
300 /* don't use 0 to avoid thinking a memset of zero is valid data */
301 GNILND_DGRAM_USED = 1,
302 GNILND_DGRAM_POSTING,
304 GNILND_DGRAM_PROCESSING,
305 GNILND_DGRAM_CANCELED,
309 typedef enum kgn_dgram_type {
310 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
311 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
312 GNILND_DGRAM_NAK, /* NO soup for you! */
313 GNILND_DGRAM_CLOSE, /* we should see other people */
316 /************************************************************************
317 * Wire message structs. These are sent in sender's byte order
318 * (i.e. receiver checks magic and flips if required).
321 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
322 #define GNILND_DGRAM_MAGIC 0x0DDBA11
324 /* kgn_msg_t - FMA/SMSG wire struct
326 * - added checksum to FMA
327 * moved seq before paylod
328 * WIRE_ATTR added for alignment
330 * added gnm_payload_len for FMA payload size
332 * added gncm_retval to completion, allowing return code transmission
335 * changed how CQID and TX ids are assigned
337 * added retval on CLOSE
339 * added payload checksumming
341 * reworked checksumming a bit, changed payload checksums
343 #define GNILND_MSG_VERSION 8
344 /* kgn_connreq_t connection request datagram wire struct
349 #define GNILND_CONNREQ_VERSION 2
351 typedef struct kgn_gniparams {
352 __u32 gnpr_host_id; /* ph. host ID of the NIC */
353 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
354 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
355 } WIRE_ATTR kgn_gniparams_t;
357 typedef struct kgn_nak_data {
358 __s32 gnnd_errno; /* errno reason for NAK */
360 } WIRE_ATTR kgn_nak_data_t;
362 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
363 * without breaking the ability for us to properly NAK someone */
364 typedef struct kgn_connreq { /* connection request/response */
365 __u32 gncr_magic; /* I'm an gnilnd connreq */
366 __u32 gncr_cksum; /* checksum (0 == disabled) */
367 __u16 gncr_type; /* REQ, NAK, etc */
368 __u16 gncr_version; /* this is my version number */
369 __u32 gncr_timeout; /* sender's timeout */
370 __u64 gncr_srcnid; /* sender's NID */
371 __u64 gncr_dstnid; /* who sender expects to listen */
372 __u64 gncr_peerstamp; /* sender's instance stamp */
373 __u64 gncr_connstamp; /* sender's connection stamp */
375 /* everything before this needs to stay static, adding after should
376 * result in a change to GNILND_CONNREQ_VERSION */
379 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
380 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
382 } WIRE_ATTR kgn_connreq_t;
385 gni_mem_handle_t gnrd_key;
388 } WIRE_ATTR kgn_rdma_desc_t;
391 lnet_hdr_t gnim_hdr; /* LNet header */
392 /* LNet payload is in FMA "Message Data" */
393 } WIRE_ATTR kgn_immediate_msg_t;
396 lnet_hdr_t gnprm_hdr; /* LNet header */
397 __u64 gnprm_cookie; /* opaque completion cookie */
398 } WIRE_ATTR kgn_putreq_msg_t;
401 __u64 gnpam_src_cookie; /* reflected completion cookie */
402 __u64 gnpam_dst_cookie; /* opaque completion cookie */
403 __u16 gnpam_payload_cksum; /* checksum for get msg */
404 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
405 } WIRE_ATTR kgn_putack_msg_t;
408 lnet_hdr_t gngm_hdr; /* LNet header */
409 __u64 gngm_cookie; /* opaque completion cookie */
410 __u16 gngm_payload_cksum; /* checksum for put msg */
411 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
412 } WIRE_ATTR kgn_get_msg_t;
415 int gncm_retval; /* error on NAK, size on REQ */
416 __u64 gncm_cookie; /* reflected completion cookie */
417 } WIRE_ATTR kgn_completion_msg_t;
419 typedef struct { /* NB must fit in FMA "Prefix" */
420 __u32 gnm_magic; /* I'm an gni message */
421 __u16 gnm_version; /* this is my version number */
422 __u16 gnm_type; /* msg type */
423 __u64 gnm_srcnid; /* sender's NID */
424 __u64 gnm_connstamp; /* sender's connection stamp */
425 __u32 gnm_seq; /* incrementing sequence number */
426 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
427 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
428 __u32 gnm_payload_len; /* size of the FMA payload sent */
430 kgn_immediate_msg_t immediate;
431 kgn_putreq_msg_t putreq;
432 kgn_putack_msg_t putack;
434 kgn_completion_msg_t completion;
436 } WIRE_ATTR kgn_msg_t;
438 /************************************************************************
439 * runtime tunable data
442 typedef struct kgn_tunables {
443 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
444 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
445 int *kgn_credits; /* # concurrent sends */
446 int *kgn_fma_cq_size; /* # entries in receive CQ */
447 int *kgn_peer_credits; /* # LNet peer credits */
448 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
449 int *kgn_timeout; /* comms timeout (seconds) */
450 int *kgn_max_immediate; /* immediate payload breakpoint */
451 int *kgn_checksum; /* checksum data */
452 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
453 int *kgn_bte_dlvr_mode; /* BTE delivery mode mask */
454 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
455 int *kgn_ptag; /* PTAG for cdm_create */
456 int *kgn_max_retransmits; /* max number of FMA retransmits */
457 int *kgn_nwildcard; /* # wildcard per net to post */
458 int *kgn_nice; /* nice value for kgnilnd threads */
459 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
460 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
461 int *kgn_peer_hash_size; /* size of kgn_peers */
462 int *kgn_peer_health; /* enable/disable peer health */
463 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
464 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
465 int *kgn_mbox_per_block; /* mailboxes per fmablk */
466 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
467 int *kgn_mbox_credits; /* max credits per fma */
468 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
469 int *kgn_net_hash_size; /* size of kgn_net_ht */
470 int *kgn_hardware_timeout; /* max time for a message to get across the network */
471 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
472 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
473 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
474 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
475 int *kgn_reverse_rdma; /* Reverse RDMA setting */
476 int *kgn_eager_credits; /* allocated eager buffers */
477 int *kgn_efault_lbug; /* Should we LBUG on receiving an EFAULT */
478 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
479 cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
483 typedef struct kgn_mbox_info {
484 lnet_nid_t mbx_prev_nid;
485 lnet_nid_t mbx_prev_purg_nid;
486 unsigned long mbx_create_conn_memset;
487 unsigned long mbx_add_purgatory;
488 unsigned long mbx_detach_of_purgatory;
489 unsigned long mbx_release_from_purgatory;
490 unsigned long mbx_release_purg_active_dgram;
492 int mbx_nallocs_total;
495 typedef struct kgn_fma_memblock {
496 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
497 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
498 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
499 int gnm_num_mboxs; /* total mboxes allocated */
500 int gnm_avail_mboxs; /* number of available mailboxes in the block */
501 int gnm_held_mboxs; /* number of purgatory held mailboxes */
502 int gnm_mbox_size; /* size of the single mailbox */
503 int gnm_next_avail_mbox; /* next available mailbox */
504 long gnm_max_timeout; /* max timeout for possible purgatory hold */
505 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
506 void *gnm_block; /* pointer to mem. block */
507 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
508 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
509 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
510 } kgn_fma_memblock_t;
512 typedef struct kgn_device {
513 gni_nic_handle_t gnd_handle; /* device handle */
514 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
515 gni_err_handle_t gnd_err_handle; /* device error handle */
516 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
517 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
518 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
519 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
520 struct mutex gnd_cq_mutex; /* CQ access serialization */
521 __u32 gnd_host_id; /* ph. host ID of the NIC */
522 int gnd_id; /* device id, also index in kgn_devices */
523 __u32 gnd_nid; /* ph host ID translated to NID */
524 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
525 struct mutex gnd_fmablk_mutex; /* mutex for FMA block memory alloc/free */
526 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
527 atomic_t gnd_nfmablk; /* # of fmablk live */
528 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
529 atomic_t gnd_neps; /* # EP allocated to conns */
530 short gnd_ready; /* stuff to do in scheduler thread */
531 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
532 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
533 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
534 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
535 struct list_head gnd_connd_peers; /* peers waiting for a connection */
536 spinlock_t gnd_connd_lock; /* serialise connd_peers */
537 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
538 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
539 int gnd_dgram_ready; /* dgrams need movin' */
540 struct list_head *gnd_dgrams; /* nid hash to dgrams */
541 atomic_t gnd_ndgrams; /* # dgrams extant */
542 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post on device */
543 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
544 struct list_head gnd_map_list; /* list of all mapped regions */
545 int gnd_map_version; /* version flag for map list */
546 struct timer_list gnd_map_timer; /* wakey-wakey */
547 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
548 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
549 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
550 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
551 __u32 gnd_map_nphys; /* # TX phys mappings */
552 __u32 gnd_map_physnop; /* # TX phys pages mapped */
553 __u32 gnd_map_nvirt; /* # TX virt mappings */
554 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
555 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
556 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
557 int gnd_map_attempt; /* last map attempt # */
558 unsigned long gnd_last_map; /* map timeout base */
559 struct list_head gnd_rdmaq; /* RDMA to be sent */
560 spinlock_t gnd_rdmaq_lock; /* play nice with others */
561 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
562 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
563 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
564 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
565 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
566 atomic_t gnd_short_ntx; /* TX stats: short messages */
567 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
568 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
569 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
570 atomic_t gnd_short_nrx; /* RX stats: short messages */
571 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
572 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
573 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
574 atomic_t gnd_fast_try; /* # of times fast send tried */
575 atomic_t gnd_fast_ok; /* # of times fast send ok */
576 atomic_t gnd_fast_block; /* # of times fast send blocked */
577 unsigned long gnd_mutex_delay;
578 atomic_t gnd_n_yield;
579 atomic_t gnd_n_schedule;
580 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
581 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
584 typedef struct kgn_net {
585 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
586 kgn_device_t *gnn_dev; /* device for this net */
587 lnet_ni_t *gnn_ni; /* network interface instance */
588 atomic_t gnn_refcount; /* # current references */
589 int gnn_shutdown; /* lnd_shutdown set */
590 __u16 gnn_netnum; /* stash netnum for quicker lookup */
593 static inline lnet_nid_t
594 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
596 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
599 static inline lnet_nid_t
600 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
602 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
605 /* The code for this is a bit ugly - but really this just boils down to a __u64
606 * that can have various parts accessed separately.
608 * The lower 32 bits is the ID
609 * we give to SMSG for our completion event - it needs to be globally unique across
610 * all TX currently in flight. We separate that out into the CQID so that we can
611 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
612 * the actual TX out of the per-connection gnc_tx_ref_table.
614 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
615 * has a unique value we can send with RDMA setup messages to ensure the completion for
616 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
617 * reuse is separated.
620 typedef struct kgn_tx_ev_id {
622 __u64 txe_cookie; /* are you my mommy ? */
624 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
626 __u32 txe_smsg_id; /* ID for SMSG CQ event */
627 /* N.B: Never ever ever ever use the bit shifts directly,
628 * you are just asking for a world of pain and are at the
629 * mercy of the compiler layouts */
631 __u32 txe_cqid :GNILND_CQID_NBITS;
632 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
639 typedef struct kgn_dgram {
640 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
641 kgn_dgram_state_t gndg_state; /* state of this dgram */
642 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
643 __u32 gndg_magic; /* saftey word */
644 unsigned long gndg_post_time; /* time when we posted */
645 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
646 kgn_connreq_t gndg_conn_out; /* connreq from local node */
647 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
650 typedef struct kgn_tx { /* message descriptor */
651 struct list_head tx_list; /* TX queues - peer, conn, rdma */
652 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
653 struct list_head *tx_list_p; /* pointer to current list */
654 struct kgn_conn *tx_conn; /* owning conn */
655 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
656 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
657 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
658 struct list_head tx_map_list; /* list entry on device map list */
659 unsigned int tx_nob; /* # bytes of payload */
660 int tx_buftype; /* payload buffer type */
661 int tx_phys_npages; /* # physical pages */
662 gni_mem_handle_t tx_map_key; /* mapping key */
663 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
664 gni_mem_segment_t *tx_phys; /* page descriptors */
665 kgn_msg_t tx_msg; /* FMA message buffer */
666 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
667 __u8 tx_state; /* state of the descriptor */
668 int tx_retrans; /* retrans count of RDMA */
669 int tx_rc; /* if we need to stash the ret code until we see completion */
670 void *tx_buffer; /* source/sink buffer */
671 void *tx_buffer_copy; /* pointer to page aligned buffer */
672 unsigned int tx_nob_rdma; /* nob actually rdma */
673 unsigned int tx_offset; /* offset of data into copied buffer */
675 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
676 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
679 /* we only use one or the other */
681 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
682 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
686 typedef struct kgn_conn {
687 kgn_device_t *gnc_device; /* which device */
688 struct kgn_peer *gnc_peer; /* owning peer */
689 int gnc_magic; /* magic value cleared before free */
690 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
691 struct list_head gnc_hashlist; /* stash in connection hash table */
692 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
693 struct list_head gnc_fmaq; /* txs queued for FMA */
694 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
695 __u64 gnc_peerstamp; /* peer's unique stamp */
696 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
697 __u64 gnc_my_connstamp; /* my unique connection stamp */
698 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
699 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
700 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
701 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
702 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
703 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
704 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
705 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
706 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
707 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
708 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
709 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
710 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
711 __u32 gnc_cqid; /* my completion callback id (non-unique) */
712 __u32 gnc_tx_seq; /* tx msg sequence number */
713 __u32 gnc_rx_seq; /* rx msg sequence number */
714 __u64 gnc_tx_retrans; /* # retrans on SMSG */
715 atomic_t gnc_nlive_fma; /* # live FMA */
716 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
717 atomic_t gnc_nlive_rdma; /* # live RDMA */
718 short gnc_close_sent; /* I've sent CLOSE */
719 short gnc_close_recvd; /* I've received CLOSE */
720 short gnc_in_purgatory; /* in the sin bin */
721 int gnc_error; /* errno when conn being closed due to error */
722 int gnc_peer_error; /* errno peer sent us on CLOSE */
723 kgn_conn_state_t gnc_state; /* connection state */
724 int gnc_scheduled; /* being attented to */
725 char gnc_sched_caller[30]; /* what function last called schedule */
726 int gnc_sched_line; /* what line # last called schedule */
727 atomic_t gnc_refcount; /* # users */
728 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
729 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
730 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
731 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
732 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
733 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
734 int gnc_next_tx; /* next tx to use in tx_ref_table */
735 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
736 int gnc_mbox_id; /* id of mbox in fma_blk */
737 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
738 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
739 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
740 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
741 void *remote_mbox_addr; /* save remote mbox address */
744 typedef struct kgn_mdd_purgatory {
745 gni_mem_handle_t gmp_map_key; /* mapping key */
746 struct list_head gmp_list; /* entry point for purgatory list */
747 } kgn_mdd_purgatory_t;
749 typedef struct kgn_peer {
750 struct list_head gnp_list; /* stash on global peer list */
751 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
752 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
753 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
754 kgn_net_t *gnp_net; /* net instance for this peer */
755 lnet_nid_t gnp_nid; /* who's on the other end(s) */
756 atomic_t gnp_refcount; /* # users */
757 __u32 gnp_host_id; /* ph. host ID of the peer */
758 short gnp_connecting; /* connection forming */
759 short gnp_pending_unlink; /* need last conn close to trigger unlink */
760 int gnp_last_errno; /* last error conn saw */
761 unsigned long gnp_last_alive; /* last time I had valid comms */
762 int gnp_last_dgram_errno; /* last error dgrams saw */
763 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
764 unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
765 unsigned long gnp_reconnect_interval; /* exponential backoff */
766 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
767 int gnp_down; /* rca says peer down */
768 unsigned long gnp_down_event_time; /* time peer down */
769 unsigned long gnp_up_event_time; /* time peer back up */
772 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
773 * like lnet_parse. It allows a single pointer to let us get enough
774 * information in _recv and friends */
775 typedef struct kgn_rx {
776 kgn_conn_t *grx_conn; /* connection */
777 kgn_msg_t *grx_msg; /* message */
778 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
779 int grx_eager; /* if eager, we copied msg to somewhere */
780 struct timespec grx_received; /* time this msg received */
783 typedef struct kgn_data {
784 int kgn_init; /* initialisation state */
785 int kgn_shutdown; /* shut down? */
786 int kgn_wc_kill; /* Should I repost the WC */
787 atomic_t kgn_nthreads; /* # live threads */
788 int kgn_nresets; /* number of stack resets */
789 int kgn_in_reset; /* are we in stack reset ? */
791 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
793 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
794 int kgn_ndevs; /* # devices */
796 int kgn_ruhroh_running; /* ruhroh thread is running */
797 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
798 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
799 int kgn_quiesce_trigger; /* should we quiesce ? */
800 atomic_t kgn_nquiesce; /* how many quiesced ? */
801 struct mutex kgn_quiesce_mutex; /* serialize ruhroh task, startup and shutdown */
802 int kgn_needs_reset; /* we need stack reset */
804 /* These next three members implement communication from gnilnd into
805 * the ruhroh task. To ensure correct operation of the task, code that
806 * writes into them must use memory barriers to ensure that the changes
807 * are visible to other cores in the order the members appear below. */
808 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
809 int kgn_bump_info_rdy; /* we have info needed to bump */
810 int kgn_needs_pause; /* we need to pause for network quiesce */
812 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
813 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
815 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
816 struct list_head *kgn_peers; /* hash table of all my known peers */
817 atomic_t kgn_npeers; /* # peers extant */
818 int kgn_peer_version; /* version flag for peer tables */
820 struct list_head *kgn_conns; /* conns hashed by cqid */
821 atomic_t kgn_nconns; /* # connections extant */
822 atomic_t kgn_neager_allocs; /* # of eager allocations */
823 __u64 kgn_peerstamp; /* when I started up */
824 __u64 kgn_connstamp; /* conn stamp generator */
825 int kgn_conn_version; /* version flag for conn tables */
826 int kgn_next_cqid; /* cqid generator */
828 long kgn_new_min_timeout; /* minimum timeout on any new conn */
829 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
830 spinlock_t kgn_reaper_lock; /* serialise */
832 struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
833 struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
834 struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
835 atomic_t kgn_ntx; /* # tx in use */
836 struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
838 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
839 __u64 kgn_cksum_npages; /* Number of pages allocated for checksumming */
840 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
841 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
843 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
844 long kgn_rdmaq_override; /* bytes per second override */
846 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
848 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
849 atomic_t kgn_npending_conns; /* # of conns with pending closes */
850 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
851 unsigned long kgn_last_scheduled; /* last time schedule was called in a sched thread */
852 unsigned long kgn_last_condresched; /* last time cond_resched was called in a sched thread */
853 atomic_t kgn_rev_offset; /* number of time REV rdma have been misaligned offsets */
854 atomic_t kgn_rev_length; /* Number of times REV rdma have been misaligned lengths */
855 atomic_t kgn_rev_copy_buff; /* Number of times REV rdma have had to make a copy buffer */
856 struct socket *kgn_sock; /* for Apollo */
859 extern kgn_data_t kgnilnd_data;
860 extern kgn_tunables_t kgnilnd_tunables;
862 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
863 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
864 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
866 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
867 * and the line of the calling function to allow us to debug problematic
868 * schedule calls in the future without the programmer having to mark
869 * the location manually.
871 #define kgnilnd_schedule_conn(conn) \
872 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
874 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
875 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
878 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
880 struct task_struct *thrd = kthread_run(fn, arg, "%s_%02d", name, id);
882 return PTR_ERR(thrd);
884 atomic_inc(&kgnilnd_data.kgn_nthreads);
889 kgnilnd_thread_fini(void)
891 atomic_dec(&kgnilnd_data.kgn_nthreads);
894 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
895 * parts of the code to avoid a scheduler trip when the mutex is held
897 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
898 * has been acquired successfully, and 0 on contention.
900 * NOTE: this function follows the spin_trylock() convention, so
901 * it is negated to the down_trylock() return values! Be careful
902 * about this when converting semaphore users to mutexes.
904 * This function must not be used in interrupt context. The
905 * mutex must be released by the same task that acquired it.
907 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
910 unsigned long timeout;
912 LASSERT(!in_interrupt());
914 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
916 ret = mutex_trylock(lock);
923 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
926 _kgnilnd_debug_msg(kgn_msg_t *msg,
927 struct libcfs_debug_msg_data *data, const char *fmt, ... );
929 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
931 CFS_CHECK_STACK(msgdata, mask, cdls); \
933 if (((mask) & D_CANTMASK) != 0 || \
934 ((libcfs_debug & (mask)) != 0 && \
935 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
936 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
939 /* for most callers (level is a constant) this is resolved at compile time */
940 #define GNIDBG_MSG(level, msg, fmt, args...) \
942 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
943 static cfs_debug_limit_state_t cdls; \
944 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
945 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
946 "$$ "fmt" from %s ", ## args, \
947 libcfs_nid2str((msg)->gnm_srcnid)); \
949 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
950 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
951 "$$ "fmt" from %s ", ## args, \
952 libcfs_nid2str((msg)->gnm_srcnid)); \
956 /* user puts 'to nid' in msg for us */
957 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
959 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
960 static cfs_debug_limit_state_t cdls; \
961 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
962 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
963 "$$ "fmt" ", ## args); \
965 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
966 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
967 "$$ "fmt" ", ## args); \
972 _kgnilnd_debug_conn(kgn_conn_t *conn,
973 struct libcfs_debug_msg_data *data, const char *fmt, ... );
975 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
977 CFS_CHECK_STACK(msgdata, mask, cdls); \
979 if (((mask) & D_CANTMASK) != 0 || \
980 ((libcfs_debug & (mask)) != 0 && \
981 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
982 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
985 /* for most callers (level is a constant) this is resolved at compile time */
986 #define GNIDBG_CONN(level, conn, fmt, args...) \
988 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
989 static cfs_debug_limit_state_t cdls; \
990 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
991 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
992 "$$ "fmt" ", ## args); \
994 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
995 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
996 "$$ "fmt" ", ## args); \
1001 _kgnilnd_debug_tx(kgn_tx_t *tx,
1002 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1004 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
1006 CFS_CHECK_STACK(msgdata, mask, cdls); \
1008 if (((mask) & D_CANTMASK) != 0 || \
1009 ((libcfs_debug & (mask)) != 0 && \
1010 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1011 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1014 /* for most callers (level is a constant) this is resolved at compile time */
1015 #define GNIDBG_TX(level, tx, fmt, args...) \
1017 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1018 static cfs_debug_limit_state_t cdls; \
1019 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1020 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1021 "$$ "fmt" ", ## args); \
1023 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1024 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1025 "$$ "fmt" ", ## args); \
1029 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1031 if (unlikely(!(cond))) { \
1032 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1037 #define GNILND_IS_QUIESCED \
1038 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1039 atomic_read(&kgnilnd_data.kgn_nthreads))
1041 #define KGNILND_SPIN_QUIESCE \
1043 /* E.T phone home */ \
1044 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1045 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1046 while (kgnilnd_data.kgn_quiesce_trigger) { \
1047 msleep_interruptible(MSEC_PER_SEC); \
1049 /* Mom, my homework is done */ \
1050 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1051 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1054 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1055 #ifndef LIBCFS_DEBUG
1056 #error "this code uses actions inside LASSERT for ref counting"
1059 #define kgnilnd_admin_addref(atomic) \
1061 int val = atomic_inc_return(&atomic); \
1062 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1063 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1066 #define kgnilnd_admin_decref(atomic) \
1068 int val = atomic_dec_return(&atomic); \
1069 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
1070 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1073 #define kgnilnd_net_addref(net) \
1075 int val = atomic_inc_return(&net->gnn_refcount); \
1076 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1077 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1078 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1081 #define kgnilnd_net_decref(net) \
1083 int val = atomic_dec_return(&net->gnn_refcount); \
1084 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1085 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1086 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1089 #define kgnilnd_peer_addref(peer) \
1091 int val = atomic_inc_return(&peer->gnp_refcount); \
1092 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1093 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1094 libcfs_nid2str(peer->gnp_nid), val); \
1097 #define kgnilnd_peer_decref(peer) \
1099 int val = atomic_dec_return(&peer->gnp_refcount); \
1100 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1101 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1102 libcfs_nid2str(peer->gnp_nid), val); \
1104 kgnilnd_destroy_peer(peer); \
1107 #define kgnilnd_conn_addref(conn) \
1112 val = atomic_inc_return(&conn->gnc_refcount); \
1113 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1114 "conn %p refc %d to %s\n", \
1117 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1119 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1121 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1126 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1127 * through the scheduler thread to get the EP destroyed. This avoids some
1128 * messy semaphore business and allows us to reuse the connd_list and existing
1129 * linkage and avoid creating extra lists just for destroying EPs */
1131 /* Safety Disclaimer:
1132 * Q: If we decrement the refcount and then check it again, is it possible that
1133 * another caller could have passed through this macro concurrently? If so,
1134 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1136 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1137 * once we are refcount <= 2. It hinges around gnc_state and membership of
1138 * gnc_hashlist. There are two ways to find a connection - either ask for
1139 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1140 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1143 * - #1 from create (kgnilnd_create_conn)
1144 * - #2 for EP (kgnilnd_create_conn)
1145 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1146 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1148 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1149 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1151 * There could be more from TX descriptors during the lifetime of a live
1154 * If we nuke the conn before finish_connect, we won't have parallel paths
1155 * because nobody besides the dgram handler for the single outstanding
1156 * dgram can find the connection as it isn't in any searchable tables yet.
1158 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1159 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1160 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1161 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1162 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1163 * sent. Now the gnc_state == CLOSED, so we hit
1164 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1165 * - we've nuked them from the peer and CQ id tables, so we own them and
1166 * are guaranteed serial access - hence the complete lack of conn list
1167 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1168 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1169 * kgnilnd_process_conns. Then the next scheduler loop would call
1170 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1171 * kgnilnd_process_conns.
1173 * Clearly, we are totally safe. Clearly.
1176 #define kgnilnd_conn_decref(conn) \
1181 val = atomic_dec_return(&conn->gnc_refcount); \
1182 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1185 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1187 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1189 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1194 (conn->gnc_ephandle != NULL) && \
1195 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1196 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1197 kgnilnd_schedule_conn(conn); \
1198 } else if (val == 0) { \
1199 kgnilnd_destroy_conn(conn); \
1203 static inline struct list_head *
1204 kgnilnd_nid2peerlist(lnet_nid_t nid)
1206 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1208 RETURN(&kgnilnd_data.kgn_peers[hash]);
1211 static inline struct list_head *
1212 kgnilnd_netnum2netlist(__u16 netnum)
1214 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1216 RETURN(&kgnilnd_data.kgn_nets[hash]);
1220 kgnilnd_peer_active(kgn_peer_t *peer)
1222 /* Am I in the peer hash table? */
1223 return (!list_empty(&peer->gnp_list));
1226 /* need write_lock on kgn_peer_conn_lock */
1228 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1230 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1231 peer, libcfs_nid2str(peer->gnp_nid),
1232 !list_empty(&peer->gnp_conns),
1233 !list_empty(&peer->gnp_tx_queue));
1235 /* kgn_peer_conn_lock protects us from conflict with
1236 * kgnilnd_peer_notify and gnp_persistent */
1237 RETURN ((list_empty(&peer->gnp_conns)) &&
1238 (list_empty(&peer->gnp_tx_queue)));
1241 /* returns positive if error was for a clean shutdown of conn */
1243 kgnilnd_conn_clean_errno(int errno)
1245 /* - ESHUTDOWN - LND is unloading
1246 * - EUCLEAN - admin requested via "lctl del_peer"
1247 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1248 * - ENOTRECOVERABLE - stack reset
1249 * - EISCONN - cleared via "lctl push"
1250 * not doing ESTALE - that isn't clean */
1251 RETURN ((errno == 0) ||
1252 (errno == -ESHUTDOWN) ||
1253 (errno == -EUCLEAN) ||
1254 (errno == -ENETRESET) ||
1255 (errno == -EISCONN) ||
1256 (errno == -ENOTRECOVERABLE));
1259 /* returns positive if error results in purgatory hold */
1261 kgnilnd_check_purgatory_errno(int errno)
1263 /* We don't want to save the purgatory lists these cases:
1264 * - EUCLEAN - admin requested via "lctl del_peer"
1265 * - ESHUTDOWN - LND is unloading
1267 RETURN ((errno != -ESHUTDOWN) &&
1268 (errno != -EUCLEAN));
1272 /* returns positive if a purgatory hold is needed */
1274 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1278 if (conn->gnc_peer) {
1279 loopback = conn->gnc_peer->gnp_nid ==
1280 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1282 /* short circuit - a conn that didn't complete
1283 * setup never needs a purgatory hold */
1286 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1287 conn, conn->gnc_peer ?
1288 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1290 loopback, conn->gnc_close_recvd);
1292 /* we only use a purgatory hold if we've not received the CLOSE msg
1293 * from our peer - without that message, we can't know the state of
1294 * the other end of this connection and must put it into purgatory
1295 * to prevent reuse and corruption.
1296 * The theory is that a TX error can be communicated in all other cases
1298 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1299 kgnilnd_check_purgatory_errno(conn->gnc_error));
1302 static inline const char *
1303 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1305 static inline struct list_head *
1306 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1307 kgn_tx_list_state_t to_state)
1310 case GNILND_TX_PEERQ:
1311 return &peer->gnp_tx_queue;
1312 case GNILND_TX_FMAQ:
1313 return &conn->gnc_fmaq;
1314 case GNILND_TX_LIVE_FMAQ:
1315 case GNILND_TX_LIVE_RDMAQ:
1316 case GNILND_TX_DYING:
1318 case GNILND_TX_MAPQ:
1319 return &conn->gnc_device->gnd_map_tx;
1320 case GNILND_TX_RDMAQ:
1321 return &conn->gnc_device->gnd_rdmaq;
1323 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1324 CERROR("invalid state requested: %s\n",
1325 kgnilnd_tx_state2str(to_state));
1331 /* should hold tx, conn or peer lock when calling */
1333 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1334 kgn_conn_t *conn, kgn_tx_list_state_t state,
1337 struct list_head *list = NULL;
1339 /* make sure we have a sane TX state to start */
1340 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1341 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1342 list_empty(&tx->tx_list),
1343 "bad state with tx_list %s",
1344 list_empty(&tx->tx_list) ? "empty" : "not empty");
1346 /* WTF - you are already on that state buttmunch */
1347 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1348 "already at %s", kgnilnd_tx_state2str(state));
1350 /* get proper list from the state requested */
1351 list = kgnilnd_tx_state2list(peer, conn, state);
1355 case GNILND_TX_PEERQ:
1356 kgnilnd_peer_addref(peer);
1358 case GNILND_TX_ALLOCD:
1359 /* no refs needed */
1361 case GNILND_TX_FMAQ:
1362 kgnilnd_conn_addref(conn);
1364 case GNILND_TX_MAPQ:
1365 atomic_inc(&conn->gnc_device->gnd_nq_map);
1366 kgnilnd_conn_addref(conn);
1368 case GNILND_TX_LIVE_FMAQ:
1369 atomic_inc(&conn->gnc_nlive_fma);
1370 kgnilnd_conn_addref(conn);
1372 case GNILND_TX_LIVE_RDMAQ:
1373 atomic_inc(&conn->gnc_nlive_rdma);
1374 kgnilnd_conn_addref(conn);
1376 case GNILND_TX_RDMAQ:
1377 atomic_inc(&conn->gnc_nq_rdma);
1378 kgnilnd_conn_addref(conn);
1380 case GNILND_TX_DYING:
1381 kgnilnd_conn_addref(conn);
1384 CERROR("invalid state requested: %s\n",
1385 kgnilnd_tx_state2str(state));
1390 /* if this changes, change kgnilnd_alloc_tx */
1391 tx->tx_list_state = state;
1393 /* some states don't have lists - we track them in the per conn
1394 * TX table instead. Waste not, want not! */
1396 tx->tx_list_p = list;
1398 list_add_tail(&tx->tx_list, list);
1400 list_add(&tx->tx_list, list);
1402 /* set dummy list_p to make book keeping happy and let debugging
1403 * be a hair easier */
1404 tx->tx_list_p = (void *)state;
1407 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1408 kgnilnd_tx_state2str(state), list);
1412 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1413 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1415 /* These is only 1 "off-list" state */
1416 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1417 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1419 /* new_state == ALLOCD means we are deallocating this tx,
1420 * so make sure it was on a valid list to start with */
1421 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1422 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1423 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1424 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1427 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1430 switch (tx->tx_list_state) {
1431 case GNILND_TX_PEERQ:
1432 kgnilnd_peer_decref(peer);
1434 case GNILND_TX_FREED:
1435 case GNILND_TX_IDLE:
1436 case GNILND_TX_ALLOCD:
1437 /* no refs needed */
1439 case GNILND_TX_DYING:
1440 kgnilnd_conn_decref(conn);
1442 case GNILND_TX_FMAQ:
1443 kgnilnd_conn_decref(conn);
1445 case GNILND_TX_MAPQ:
1446 atomic_dec(&conn->gnc_device->gnd_nq_map);
1447 kgnilnd_conn_decref(conn);
1449 case GNILND_TX_LIVE_FMAQ:
1450 atomic_dec(&conn->gnc_nlive_fma);
1451 kgnilnd_conn_decref(conn);
1453 case GNILND_TX_LIVE_RDMAQ:
1454 atomic_dec(&conn->gnc_nlive_rdma);
1455 kgnilnd_conn_decref(conn);
1457 case GNILND_TX_RDMAQ:
1458 atomic_dec(&conn->gnc_nq_rdma);
1459 kgnilnd_conn_decref(conn);
1460 /* don't need to assert on default, already did in set */
1463 /* for ALLOCD, this might already be true, but no harm doing it again */
1464 list_del_init(&tx->tx_list);
1465 tx->tx_list_p = NULL;
1466 tx->tx_list_state = new_state;
1470 kgnilnd_tx_mapped(kgn_tx_t *tx)
1472 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1473 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1476 static inline struct list_head *
1477 kgnilnd_cqid2connlist(__u32 cqid)
1479 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1481 return (&kgnilnd_data.kgn_conns [hash]);
1484 static inline kgn_conn_t *
1485 kgnilnd_cqid2conn_locked(__u32 cqid)
1487 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1488 struct list_head *tmp;
1491 list_for_each(tmp, conns) {
1492 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1494 if (conn->gnc_cqid == cqid)
1501 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1503 kgnilnd_get_cqid_locked(void)
1509 cqid = kgnilnd_data.kgn_next_cqid++;
1510 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1514 kgnilnd_data.kgn_next_cqid = 1;
1517 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1523 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1525 kgn_tx_t *tx = NULL;
1526 kgn_conn_t *conn = NULL;
1528 /* set to NULL so any early return is an error */
1532 LASSERTF((ev_id->txe_idx > 0) &&
1533 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1534 "bogus txe_idx %d >= %d\n",
1535 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1537 LASSERTF((ev_id->txe_cqid > 0) &&
1538 (ev_id->txe_cqid < GNILND_MAX_CQID),
1539 "bogus txe_cqid %d >= %d\n",
1540 ev_id->txe_cqid, GNILND_MAX_CQID);
1542 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1543 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1546 /* Conn was destroyed? */
1547 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1548 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1551 /* just insurance */
1552 kgnilnd_conn_addref(conn);
1553 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1554 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1556 /* we know this is safe - as the TX won't be reused until AFTER
1557 * the conn is unlinked from the cqid hash, so we can use the TX
1558 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1560 spin_lock(&conn->gnc_tx_lock);
1561 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1562 spin_unlock(&conn->gnc_tx_lock);
1564 /* We could have a tx that was cleared out by other forces
1565 * lctl disconnect or del_peer. */
1567 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1568 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1569 kgnilnd_conn_decref(conn);
1573 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1574 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1575 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1577 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1578 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1579 "found tx %p txe_idx %d",
1580 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1581 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1583 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1585 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1590 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1593 /* set_normalized_timepsec isn't exported from the kernel, so
1594 * we need to do the same thing inline */
1595 static inline struct timespec
1596 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1602 sec = lhs.tv_sec - rhs.tv_sec;
1603 nsec = lhs.tv_nsec - rhs.tv_nsec;
1605 while (nsec >= NSEC_PER_SEC) {
1606 nsec -= NSEC_PER_SEC;
1610 nsec += NSEC_PER_SEC;
1619 kgnilnd_count_list(struct list_head *q)
1621 struct list_head *e;
1624 list_for_each(e, q) {
1631 /* kgnilnd_find_net adds a reference to the net it finds
1632 * this is so the net will not be removed before the calling function
1633 * has time to use the data returned. This reference needs to be released
1634 * by the calling function once it has finished using the returned net
1638 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1643 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1649 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1650 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1651 kgnilnd_net_addref(net);
1652 up_read(&kgnilnd_data.kgn_net_rw_sem);
1658 up_read(&kgnilnd_data.kgn_net_rw_sem);
1663 #ifdef CONFIG_DEBUG_SLAB
1664 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1666 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1669 int kgnilnd_dev_init(kgn_device_t *dev);
1670 void kgnilnd_dev_fini(kgn_device_t *dev);
1671 int kgnilnd_startup(lnet_ni_t *ni);
1672 void kgnilnd_shutdown(lnet_ni_t *ni);
1673 int kgnilnd_base_startup(void);
1674 void kgnilnd_base_shutdown(void);
1676 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1677 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1678 void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
1679 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1681 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1682 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1683 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1684 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1685 lnet_msg_t *lntmsg, void **new_private);
1686 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1687 int delayed, unsigned int niov,
1688 struct iovec *iov, lnet_kiov_t *kiov,
1689 unsigned int offset, unsigned int mlen, unsigned int rlen);
1691 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1693 /* purgatory functions */
1694 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1695 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1696 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1697 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1699 void kgnilnd_update_reaper_timeout(long timeout);
1700 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1701 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1702 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1703 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1704 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1705 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
1706 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1708 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1709 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
1710 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1711 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1713 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1714 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1715 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1716 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1717 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1718 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1719 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1720 void kgnilnd_consume_rx(kgn_rx_t *rx);
1722 void kgnilnd_schedule_device(kgn_device_t *dev);
1723 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1724 void kgnilnd_schedule_device_timer(unsigned long arg);
1726 int kgnilnd_reaper(void *arg);
1727 int kgnilnd_scheduler(void *arg);
1728 int kgnilnd_dgram_mover(void *arg);
1729 int kgnilnd_rca(void *arg);
1731 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1732 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1733 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1734 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1735 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1736 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1737 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1738 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1739 void kgnilnd_peer_alive(kgn_peer_t *peer);
1740 void kgnilnd_peer_notify(kgn_peer_t *peer, int error);
1741 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1742 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1743 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1744 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1746 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1747 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1748 void kgnilnd_wakeup_rca_thread(void);
1749 int kgnilnd_start_rca_thread(void);
1750 int kgnilnd_get_node_state(__u32 nid);
1752 int kgnilnd_tunables_init(void);
1753 void kgnilnd_tunables_fini(void);
1754 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1756 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1757 void kgnilnd_pause_threads(void);
1758 int kgnilnd_hw_in_quiesce(void);
1759 int kgnilnd_check_hw_quiesce(void);
1760 void kgnilnd_quiesce_wait(char *reason);
1761 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1762 int kgnilnd_ruhroh_thread(void *arg);
1763 void kgnilnd_reset_stack(void);
1764 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1766 void kgnilnd_insert_sysctl(void);
1767 void kgnilnd_remove_sysctl(void);
1768 void kgnilnd_proc_init(void);
1769 void kgnilnd_proc_fini(void);
1772 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1774 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1775 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1776 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
1778 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1779 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1780 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1781 int kgnilnd_cancel_dgrams(kgn_device_t *dev);
1782 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1784 int kgnilnd_dgram_waitq(void *arg);
1786 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1788 /* struct2str functions - we don't use a default: case to cause the compile
1789 * to fail if there is a missing case. This allows us to hide these down here
1790 * out of the way but ensure we'll catch any updates to the enum/types
1793 #define DO_TYPE(x) case x: return #x;
1794 static inline const char *
1795 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1797 /* Only want single char string for this */
1799 case GNILND_FMABLK_IDLE:
1801 case GNILND_FMABLK_PHYS:
1803 case GNILND_FMABLK_VIRT:
1805 case GNILND_FMABLK_FREED:
1808 return "<unknown state>";
1811 static inline const char *
1812 kgnilnd_msgtype2str(int type)
1815 DO_TYPE(GNILND_MSG_NONE);
1816 DO_TYPE(GNILND_MSG_NOOP);
1817 DO_TYPE(GNILND_MSG_IMMEDIATE);
1818 DO_TYPE(GNILND_MSG_PUT_REQ);
1819 DO_TYPE(GNILND_MSG_PUT_NAK);
1820 DO_TYPE(GNILND_MSG_PUT_ACK);
1821 DO_TYPE(GNILND_MSG_PUT_DONE);
1822 DO_TYPE(GNILND_MSG_GET_REQ);
1823 DO_TYPE(GNILND_MSG_GET_NAK);
1824 DO_TYPE(GNILND_MSG_GET_DONE);
1825 DO_TYPE(GNILND_MSG_CLOSE);
1826 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1827 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1828 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1829 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1830 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1831 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1832 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1834 return "<unknown msg type>";
1837 static inline const char *
1838 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1841 DO_TYPE(GNILND_TX_IDLE);
1842 DO_TYPE(GNILND_TX_ALLOCD);
1843 DO_TYPE(GNILND_TX_PEERQ);
1844 DO_TYPE(GNILND_TX_MAPQ);
1845 DO_TYPE(GNILND_TX_FMAQ);
1846 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1847 DO_TYPE(GNILND_TX_RDMAQ);
1848 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1849 DO_TYPE(GNILND_TX_DYING);
1850 DO_TYPE(GNILND_TX_FREED);
1852 return "<unknown state>";
1855 static inline const char *
1856 kgnilnd_conn_state2str(kgn_conn_t *conn)
1858 kgn_conn_state_t state = conn->gnc_state;
1860 DO_TYPE(GNILND_CONN_DUMMY);
1861 DO_TYPE(GNILND_CONN_LISTEN);
1862 DO_TYPE(GNILND_CONN_CONNECTING);
1863 DO_TYPE(GNILND_CONN_ESTABLISHED);
1864 DO_TYPE(GNILND_CONN_CLOSING);
1865 DO_TYPE(GNILND_CONN_CLOSED);
1866 DO_TYPE(GNILND_CONN_DONE);
1867 DO_TYPE(GNILND_CONN_DESTROY_EP);
1872 static inline const char *
1873 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1875 kgn_connreq_type_t type = connreq->gncr_type;
1878 DO_TYPE(GNILND_CONNREQ_REQ);
1879 DO_TYPE(GNILND_CONNREQ_NAK);
1880 DO_TYPE(GNILND_CONNREQ_CLOSE);
1885 static inline const char *
1886 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1888 kgn_dgram_state_t state = dgram->gndg_state;
1891 DO_TYPE(GNILND_DGRAM_USED);
1892 DO_TYPE(GNILND_DGRAM_POSTING);
1893 DO_TYPE(GNILND_DGRAM_POSTED);
1894 DO_TYPE(GNILND_DGRAM_PROCESSING);
1895 DO_TYPE(GNILND_DGRAM_DONE);
1896 DO_TYPE(GNILND_DGRAM_CANCELED);
1901 static inline const char *
1902 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1904 kgn_dgram_type_t type = dgram->gndg_type;
1907 DO_TYPE(GNILND_DGRAM_REQ);
1908 DO_TYPE(GNILND_DGRAM_WC_REQ);
1909 DO_TYPE(GNILND_DGRAM_NAK);
1910 DO_TYPE(GNILND_DGRAM_CLOSE);
1915 static inline const char *
1916 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
1919 DO_TYPE(GNILND_DGRAM_REQ);
1920 DO_TYPE(GNILND_DGRAM_WC_REQ);
1921 DO_TYPE(GNILND_DGRAM_NAK);
1922 DO_TYPE(GNILND_DGRAM_CLOSE);
1929 /* API wrapper functions - include late to pick up all of the other defines */
1930 #include "gnilnd_api_wrap.h"
1932 /* pulls in tunables per platform and adds in nid/nic conversion
1933 * if RCA wasn't available at build time */
1934 #include "gnilnd_hss_ops.h"
1936 #if defined(CONFIG_CRAY_GEMINI)
1937 #include "gnilnd_gemini.h"
1938 #elif defined(CONFIG_CRAY_ARIES)
1939 #include "gnilnd_aries.h"
1941 #error "Undefined Network Hardware Type"
1944 #endif /* _GNILND_GNILND_H_ */