2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Derived from work by: Eric Barton <eric@bartonsoftware.com>
7 * Author: Nic Henke <nic@cray.com>
8 * Author: James Shimek <jshimek@cray.com>
10 * This file is part of Lustre, http://www.lustre.org.
12 * Lustre is free software; you can redistribute it and/or
13 * modify it under the terms of version 2 of the GNU General Public
14 * License as published by the Free Software Foundation.
16 * Lustre is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with Lustre; if not, write to the Free Software
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
26 #ifndef _GNILND_GNILND_H_
27 #define _GNILND_GNILND_H_
29 #include <linux/module.h>
30 #include <linux/kernel.h>
32 #include <linux/string.h>
33 #include <linux/stat.h>
34 #include <linux/errno.h>
35 #ifdef HAVE_LINUX_KERNEL_LOCK
36 #include <linux/smp_lock.h>
38 #include <linux/unistd.h>
39 #include <linux/uio.h>
40 #include <linux/time.h>
41 #include <asm/timex.h>
43 #include <asm/system.h>
44 #include <asm/uaccess.h>
47 #include <linux/init.h>
49 #include <linux/file.h>
50 #include <linux/stat.h>
51 #include <linux/list.h>
52 #include <linux/kmod.h>
53 #include <linux/sysctl.h>
54 #include <linux/kthread.h>
55 #include <linux/nmi.h>
59 #include <linux/nmi.h>
61 #define DEBUG_SUBSYSTEM S_LND
63 #include <libcfs/linux/kp30.h>
64 #include <libcfs/libcfs.h>
65 #include <lnet/lnet.h>
66 #include <lnet/lib-lnet.h>
69 #include "gnilnd_version.h"
72 /* tunables determined at compile time */
73 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
74 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
75 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
76 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
77 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
78 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
79 #define GNILND_SCHED_TIMEOUT 1
80 #define GNILND_DGRAM_TIMEOUT 2
81 #define GNILND_FAST_MAPPING_TRY \
82 *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
83 #define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
85 /* map failure timeout */
86 #define GNILND_MAP_TIMEOUT \
87 (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
88 *kgnilnd_tunables.kgn_timeout))
90 /* reaper thread wakup interval */
91 #define GNILND_REAPER_THREAD_WAKE 1
92 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
93 #define GNILND_REAPER_NCHECKS 4
96 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
97 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
98 #define GNILND_COOKIE 0xa3579 /* cookie used by along with ptag by GNI */
99 #define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
100 /* checksum values */
101 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
102 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
103 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
104 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
106 /* tune down some COMPUTE options as they won't see the same number of connections and
107 * don't need the throughput of multiple threads by default */
108 #if defined(CONFIG_CRAY_COMPUTE)
109 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
110 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
111 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
112 #define GNILND_COMPUTE 1 /* compute image */
114 #define GNILND_SCHED_THREADS 3 /* default # of kgnilnd_scheduler threads */
115 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
116 #define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
117 #define GNILND_COMPUTE 0 /* service image */
120 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
121 #define GNILND_EXTRA_BITS 1
122 /* maximum number of conns & bits for cqid in the SMSG event data */
123 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
124 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
125 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
126 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
127 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
129 /* need sane upper bound to limit copy overhead */
130 #define GNILND_MAX_IMMEDIATE (64<<10)
132 /* payload size to add to the base mailbox size
133 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
134 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
135 * the calculation return from that function.*/
136 #define GNILND_MBOX_PAYLOAD \
137 (GNILND_MAX_MSG_SIZE * \
138 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
140 /* timeout -> deadman timer for kgni mdd holds */
141 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
143 /* timeout for failing sends in t is in jiffies*/
144 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
146 /* time when to release from purgatory in the reaper thread in jiffies */
147 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
149 /* Macro for finding last_rx 2 datapoints are compared
150 * and the most recent one in jiffies is returned.
152 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
153 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
155 /************************************************************************
156 * Enum, flag and tag data
158 #define GNILND_INIT_NOTHING 0
159 #define GNILND_INIT_DATA 1
160 #define GNILND_INIT_ALL 2
162 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
163 #define GNILND_BUF_NONE 0 /* buffer type not set */
164 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
165 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
166 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
167 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
168 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
169 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
171 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
172 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
173 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
174 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
175 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
177 /* stash above max CQID to avoid any collision */
178 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
179 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
181 /* kgn_msg_t::gnm_type */
182 #define GNILND_MSG_NONE 0x00 /* illegal message */
183 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
184 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
185 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
186 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
187 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
188 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
189 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
190 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
191 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
192 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
193 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
194 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
195 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
196 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
197 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
198 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
199 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
201 /* defines for gnc_*scheduled states */
202 #define GNILND_CONN_IDLE 0
203 #define GNILND_CONN_SCHED 1
204 #define GNILND_CONN_WANTS_SCHED 2
205 #define GNILND_CONN_PROCESS 3
207 #define GNILND_DEV_IDLE 0
208 #define GNILND_DEV_IRQ 1
209 #define GNILND_DEV_LOOP 2
211 #define GNILND_DGRAM_IDLE 0
212 #define GNILND_DGRAM_SCHED 1
213 #define GNILND_DGRAM_PROCESS 2
215 #define GNILND_PEER_IDLE 0
216 #define GNILND_PEER_CONNECT 1
217 #define GNILND_PEER_POSTING 2
218 #define GNILND_PEER_POSTED 3
219 #define GNILND_PEER_NEEDS_DEATH 4
220 #define GNILND_PEER_KILL 5
222 /* for gnc_close_recvd */
223 #define GNILND_CLOSE_RX 1
224 #define GNILND_CLOSE_INJECT1 2
225 #define GNILND_CLOSE_INJECT2 3
226 #define GNILND_CLOSE_EARLY 4
228 /* defines for why quiesce trigger set */
229 #define GNILND_QUIESCE_IDLE 0
230 #define GNILND_QUIESCE_ADMIN 1
231 #define GNILND_QUIESCE_RESET 2
232 #define GNILND_QUIESCE_HW_QUIESCE 3
234 #define GNILND_PEER_CLEAN 0
235 #define GNILND_PEER_PERSISTING 1
237 #define GNILND_DEL_CONN 0
238 #define GNILND_DEL_PEER 1
239 #define GNILND_CLEAR_PURGATORY 2
241 #define GNILND_RCA_NODE_UP 0
242 #define GNILND_RCA_NODE_DOWN 1
243 #define GNILND_RCA_NODE_UNKNOWN 2
245 /* defines for reverse RDMA states */
246 #define GNILND_REVERSE_NONE 0
247 #define GNILND_REVERSE_GET 1
248 #define GNILND_REVERSE_PUT 2
249 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
251 typedef enum kgn_fmablk_state {
252 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
253 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
254 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
255 GNILND_FMABLK_FREED, /* after free */
256 } kgn_fmablk_state_t;
258 typedef enum kgn_tx_list_state {
259 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
260 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
261 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
262 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
263 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
264 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
265 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
266 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
267 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
268 GNILND_TX_FREED /* TX is free! */
269 } kgn_tx_list_state_t;
271 typedef enum kgn_conn_state {
272 /* don't start @ 0 - prevent memset(0) badness */
273 GNILND_CONN_DUMMY = 0,
275 GNILND_CONN_CONNECTING,
276 GNILND_CONN_ESTABLISHED,
280 GNILND_CONN_DESTROY_EP
283 /* changing these requires a change to GNILND_CONNREQ_VERSION and
284 * will result in dropped packets instead of NAKs. Adding to this is
285 * acceptable without changing the CONNREQ_VERSION, but code should
286 * be ready to handle NAKs on version mismatch */
287 typedef enum kgn_connreq_type {
288 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
289 GNILND_CONNREQ_NAK, /* NO soup for you! */
290 GNILND_CONNREQ_CLOSE, /* we should see other people */
291 } kgn_connreq_type_t;
293 typedef enum kgn_dgram_state {
294 /* don't use 0 to avoid thinking a memset of zero is valid data */
295 GNILND_DGRAM_USED = 1,
296 GNILND_DGRAM_POSTING,
298 GNILND_DGRAM_PROCESSING,
299 GNILND_DGRAM_CANCELED,
303 typedef enum kgn_dgram_type {
304 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
305 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
306 GNILND_DGRAM_NAK, /* NO soup for you! */
307 GNILND_DGRAM_CLOSE, /* we should see other people */
310 /************************************************************************
311 * Wire message structs. These are sent in sender's byte order
312 * (i.e. receiver checks magic and flips if required).
315 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
316 #define GNILND_DGRAM_MAGIC 0x0DDBA11
318 /* kgn_msg_t - FMA/SMSG wire struct
320 * - added checksum to FMA
321 * moved seq before paylod
322 * WIRE_ATTR added for alignment
324 * added gnm_payload_len for FMA payload size
326 * added gncm_retval to completion, allowing return code transmission
329 * changed how CQID and TX ids are assigned
331 * added retval on CLOSE
333 * added payload checksumming
335 * reworked checksumming a bit, changed payload checksums
337 #define GNILND_MSG_VERSION 8
338 /* kgn_connreq_t connection request datagram wire struct
343 #define GNILND_CONNREQ_VERSION 2
345 typedef struct kgn_gniparams {
346 __u32 gnpr_host_id; /* ph. host ID of the NIC */
347 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
348 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
349 } WIRE_ATTR kgn_gniparams_t;
351 typedef struct kgn_nak_data {
352 __s32 gnnd_errno; /* errno reason for NAK */
354 } WIRE_ATTR kgn_nak_data_t;
356 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
357 * without breaking the ability for us to properly NAK someone */
358 typedef struct kgn_connreq { /* connection request/response */
359 __u32 gncr_magic; /* I'm an gnilnd connreq */
360 __u32 gncr_cksum; /* checksum (0 == disabled) */
361 __u16 gncr_type; /* REQ, NAK, etc */
362 __u16 gncr_version; /* this is my version number */
363 __u32 gncr_timeout; /* sender's timeout */
364 __u64 gncr_srcnid; /* sender's NID */
365 __u64 gncr_dstnid; /* who sender expects to listen */
366 __u64 gncr_peerstamp; /* sender's instance stamp */
367 __u64 gncr_connstamp; /* sender's connection stamp */
369 /* everything before this needs to stay static, adding after should
370 * result in a change to GNILND_CONNREQ_VERSION */
373 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
374 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
376 } WIRE_ATTR kgn_connreq_t;
379 gni_mem_handle_t gnrd_key;
382 } WIRE_ATTR kgn_rdma_desc_t;
385 lnet_hdr_t gnim_hdr; /* LNet header */
386 /* LNet payload is in FMA "Message Data" */
387 } WIRE_ATTR kgn_immediate_msg_t;
390 lnet_hdr_t gnprm_hdr; /* LNet header */
391 __u64 gnprm_cookie; /* opaque completion cookie */
392 } WIRE_ATTR kgn_putreq_msg_t;
395 __u64 gnpam_src_cookie; /* reflected completion cookie */
396 __u64 gnpam_dst_cookie; /* opaque completion cookie */
397 __u16 gnpam_payload_cksum; /* checksum for get msg */
398 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
399 } WIRE_ATTR kgn_putack_msg_t;
402 lnet_hdr_t gngm_hdr; /* LNet header */
403 __u64 gngm_cookie; /* opaque completion cookie */
404 __u16 gngm_payload_cksum; /* checksum for put msg */
405 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
406 } WIRE_ATTR kgn_get_msg_t;
409 int gncm_retval; /* error on NAK, size on REQ */
410 __u64 gncm_cookie; /* reflected completion cookie */
411 } WIRE_ATTR kgn_completion_msg_t;
413 typedef struct { /* NB must fit in FMA "Prefix" */
414 __u32 gnm_magic; /* I'm an gni message */
415 __u16 gnm_version; /* this is my version number */
416 __u16 gnm_type; /* msg type */
417 __u64 gnm_srcnid; /* sender's NID */
418 __u64 gnm_connstamp; /* sender's connection stamp */
419 __u32 gnm_seq; /* incrementing sequence number */
420 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
421 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
422 __u32 gnm_payload_len; /* size of the FMA payload sent */
424 kgn_immediate_msg_t immediate;
425 kgn_putreq_msg_t putreq;
426 kgn_putack_msg_t putack;
428 kgn_completion_msg_t completion;
430 } WIRE_ATTR kgn_msg_t;
432 /************************************************************************
433 * runtime tunable data
436 typedef struct kgn_tunables {
437 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
438 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
439 int *kgn_credits; /* # concurrent sends */
440 int *kgn_fma_cq_size; /* # entries in receive CQ */
441 int *kgn_peer_credits; /* # LNet peer credits */
442 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
443 int *kgn_timeout; /* comms timeout (seconds) */
444 int *kgn_max_immediate; /* immediate payload breakpoint */
445 int *kgn_checksum; /* checksum data */
446 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
447 int *kgn_bte_dlvr_mode; /* BTE delivery mode mask */
448 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
449 int *kgn_ptag; /* PTAG for cdm_create */
450 int *kgn_max_retransmits; /* max number of FMA retransmits */
451 int *kgn_nwildcard; /* # wildcard per net to post */
452 int *kgn_nice; /* nice value for kgnilnd threads */
453 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
454 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
455 int *kgn_peer_hash_size; /* size of kgn_peers */
456 int *kgn_peer_health; /* enable/disable peer health */
457 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
458 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
459 int *kgn_mbox_per_block; /* mailboxes per fmablk */
460 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
461 int *kgn_mbox_credits; /* max credits per fma */
462 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
463 int *kgn_net_hash_size; /* size of kgn_net_ht */
464 int *kgn_hardware_timeout; /* max time for a message to get across the network */
465 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
466 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
467 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
468 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
469 int *kgn_reverse_rdma; /* Reverse RDMA setting */
470 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
471 cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
475 typedef struct kgn_mbox_info {
476 lnet_nid_t mbx_prev_nid;
477 lnet_nid_t mbx_prev_purg_nid;
478 unsigned long mbx_create_conn_memset;
479 unsigned long mbx_add_purgatory;
480 unsigned long mbx_detach_of_purgatory;
481 unsigned long mbx_release_from_purgatory;
482 unsigned long mbx_release_purg_active_dgram;
484 int mbx_nallocs_total;
487 typedef struct kgn_fma_memblock {
488 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
489 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
490 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
491 int gnm_num_mboxs; /* total mboxes allocated */
492 int gnm_avail_mboxs; /* number of available mailboxes in the block */
493 int gnm_held_mboxs; /* number of purgatory held mailboxes */
494 int gnm_mbox_size; /* size of the single mailbox */
495 int gnm_next_avail_mbox; /* next available mailbox */
496 long gnm_max_timeout; /* max timeout for possible purgatory hold */
497 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
498 void *gnm_block; /* pointer to mem. block */
499 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
500 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
501 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
502 } kgn_fma_memblock_t;
504 typedef struct kgn_device {
505 gni_nic_handle_t gnd_handle; /* device handle */
506 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
507 gni_err_handle_t gnd_err_handle; /* device error handle */
508 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
509 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
510 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
511 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
512 struct mutex gnd_cq_mutex; /* CQ access serialization */
513 __u32 gnd_host_id; /* ph. host ID of the NIC */
514 int gnd_id; /* device id, also index in kgn_devices */
515 __u32 gnd_nid; /* ph host ID translated to NID */
516 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
517 struct semaphore gnd_fmablk_sem; /* semaphore for FMA block memory alloc/free */
518 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
519 atomic_t gnd_nfmablk; /* # of fmablk live */
520 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
521 atomic_t gnd_neps; /* # EP allocated to conns */
522 short gnd_ready; /* stuff to do in scheduler thread */
523 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
524 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
525 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
526 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
527 struct list_head gnd_connd_peers; /* peers waiting for a connection */
528 spinlock_t gnd_connd_lock; /* serialise connd_peers */
529 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
530 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
531 int gnd_dgram_ready; /* dgrams need movin' */
532 struct list_head *gnd_dgrams; /* nid hash to dgrams */
533 atomic_t gnd_ndgrams; /* # dgrams extant */
534 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post on device */
535 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
536 struct list_head gnd_map_list; /* list of all mapped regions */
537 int gnd_map_version; /* version flag for map list */
538 struct timer_list gnd_map_timer; /* wakey-wakey */
539 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
540 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
541 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
542 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
543 __u32 gnd_map_nphys; /* # TX phys mappings */
544 __u32 gnd_map_physnop; /* # TX phys pages mapped */
545 __u32 gnd_map_nvirt; /* # TX virt mappings */
546 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
547 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
548 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
549 int gnd_map_attempt; /* last map attempt # */
550 unsigned long gnd_last_map; /* map timeout base */
551 struct list_head gnd_rdmaq; /* RDMA to be sent */
552 spinlock_t gnd_rdmaq_lock; /* play nice with others */
553 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
554 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
555 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
556 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
557 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
558 atomic_t gnd_short_ntx; /* TX stats: short messages */
559 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
560 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
561 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
562 atomic_t gnd_short_nrx; /* RX stats: short messages */
563 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
564 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
565 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
566 atomic_t gnd_fast_try; /* # of times fast send tried */
567 atomic_t gnd_fast_ok; /* # of times fast send ok */
568 atomic_t gnd_fast_block; /* # of times fast send blocked */
569 unsigned long gnd_mutex_delay;
570 atomic_t gnd_n_yield;
571 atomic_t gnd_n_schedule;
572 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
573 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
576 typedef struct kgn_net {
577 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
578 kgn_device_t *gnn_dev; /* device for this net */
579 lnet_ni_t *gnn_ni; /* network interface instance */
580 atomic_t gnn_refcount; /* # current references */
581 int gnn_shutdown; /* lnd_shutdown set */
582 __u16 gnn_netnum; /* stash netnum for quicker lookup */
585 static inline lnet_nid_t
586 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
588 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
591 static inline lnet_nid_t
592 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
594 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
597 /* The code for this is a bit ugly - but really this just boils down to a __u64
598 * that can have various parts accessed separately.
600 * The lower 32 bits is the ID
601 * we give to SMSG for our completion event - it needs to be globally unique across
602 * all TX currently in flight. We separate that out into the CQID so that we can
603 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
604 * the actual TX out of the per-connection gnc_tx_ref_table.
606 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
607 * has a unique value we can send with RDMA setup messages to ensure the completion for
608 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
609 * reuse is separated.
612 typedef struct kgn_tx_ev_id {
614 __u64 txe_cookie; /* are you my mommy ? */
616 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
618 __u32 txe_smsg_id; /* ID for SMSG CQ event */
619 /* N.B: Never ever ever ever use the bit shifts directly,
620 * you are just asking for a world of pain and are at the
621 * mercy of the compiler layouts */
623 __u32 txe_cqid :GNILND_CQID_NBITS;
624 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
631 typedef struct kgn_dgram {
632 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
633 kgn_dgram_state_t gndg_state; /* state of this dgram */
634 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
635 __u32 gndg_magic; /* saftey word */
636 unsigned long gndg_post_time; /* time when we posted */
637 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
638 kgn_connreq_t gndg_conn_out; /* connreq from local node */
639 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
642 typedef struct kgn_tx { /* message descriptor */
643 struct list_head tx_list; /* TX queues - peer, conn, rdma */
644 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
645 struct list_head *tx_list_p; /* pointer to current list */
646 struct kgn_conn *tx_conn; /* owning conn */
647 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
648 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
649 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
650 struct list_head tx_map_list; /* list entry on device map list */
651 unsigned int tx_nob; /* # bytes of payload */
652 int tx_buftype; /* payload buffer type */
653 int tx_phys_npages; /* # physical pages */
654 gni_mem_handle_t tx_map_key; /* mapping key */
655 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
656 gni_mem_segment_t *tx_phys; /* page descriptors */
657 kgn_msg_t tx_msg; /* FMA message buffer */
658 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
659 __u8 tx_state; /* state of the descriptor */
660 int tx_retrans; /* retrans count of RDMA */
661 int tx_rc; /* if we need to stash the ret code until we see completion */
662 void *tx_buffer; /* source/sink buffer */
663 void *tx_buffer_copy; /* pointer to page aligned buffer */
664 unsigned int tx_nob_rdma; /* nob actually rdma */
665 unsigned int tx_offset; /* offset of data into copied buffer */
667 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
668 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
671 /* we only use one or the other */
673 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
674 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
678 typedef struct kgn_conn {
679 kgn_device_t *gnc_device; /* which device */
680 struct kgn_peer *gnc_peer; /* owning peer */
681 int gnc_magic; /* magic value cleared before free */
682 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
683 struct list_head gnc_hashlist; /* stash in connection hash table */
684 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
685 struct list_head gnc_fmaq; /* txs queued for FMA */
686 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
687 __u64 gnc_peerstamp; /* peer's unique stamp */
688 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
689 __u64 gnc_my_connstamp; /* my unique connection stamp */
690 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
691 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
692 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
693 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
694 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
695 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
696 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
697 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
698 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
699 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
700 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
701 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
702 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
703 __u32 gnc_cqid; /* my completion callback id (non-unique) */
704 __u32 gnc_tx_seq; /* tx msg sequence number */
705 __u32 gnc_rx_seq; /* rx msg sequence number */
706 __u64 gnc_tx_retrans; /* # retrans on SMSG */
707 atomic_t gnc_nlive_fma; /* # live FMA */
708 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
709 atomic_t gnc_nlive_rdma; /* # live RDMA */
710 short gnc_close_sent; /* I've sent CLOSE */
711 short gnc_close_recvd; /* I've received CLOSE */
712 short gnc_in_purgatory; /* in the sin bin */
713 int gnc_error; /* errno when conn being closed due to error */
714 int gnc_peer_error; /* errno peer sent us on CLOSE */
715 kgn_conn_state_t gnc_state; /* connection state */
716 int gnc_scheduled; /* being attented to */
717 char gnc_sched_caller[30]; /* what function last called schedule */
718 int gnc_sched_line; /* what line # last called schedule */
719 atomic_t gnc_refcount; /* # users */
720 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
721 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
722 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
723 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
724 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
725 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
726 int gnc_next_tx; /* next tx to use in tx_ref_table */
727 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
728 int gnc_mbox_id; /* id of mbox in fma_blk */
729 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
730 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
731 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
732 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
733 void *remote_mbox_addr; /* save remote mbox address */
736 typedef struct kgn_mdd_purgatory {
737 gni_mem_handle_t gmp_map_key; /* mapping key */
738 struct list_head gmp_list; /* entry point for purgatory list */
739 } kgn_mdd_purgatory_t;
741 typedef struct kgn_peer {
742 struct list_head gnp_list; /* stash on global peer list */
743 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
744 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
745 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
746 kgn_net_t *gnp_net; /* net instance for this peer */
747 lnet_nid_t gnp_nid; /* who's on the other end(s) */
748 atomic_t gnp_refcount; /* # users */
749 __u32 gnp_host_id; /* ph. host ID of the peer */
750 short gnp_connecting; /* connection forming */
751 short gnp_pending_unlink; /* need last conn close to trigger unlink */
752 int gnp_last_errno; /* last error conn saw */
753 unsigned long gnp_last_alive; /* last time I had valid comms */
754 int gnp_last_dgram_errno; /* last error dgrams saw */
755 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
756 unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
757 unsigned long gnp_reconnect_interval; /* exponential backoff */
758 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
759 int gnp_down; /* rca says peer down */
760 unsigned long gnp_down_event_time; /* time peer down */
761 unsigned long gnp_up_event_time; /* time peer back up */
764 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
765 * like lnet_parse. It allows a single pointer to let us get enough
766 * information in _recv and friends */
767 typedef struct kgn_rx {
768 kgn_conn_t *grx_conn; /* connection */
769 kgn_msg_t *grx_msg; /* message */
770 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
771 int grx_eager; /* if eager, we copied msg to somewhere */
772 struct timespec grx_received; /* time this msg received */
775 typedef struct kgn_data {
776 int kgn_init; /* initialisation state */
777 int kgn_shutdown; /* shut down? */
778 int kgn_wc_kill; /* Should I repost the WC */
779 atomic_t kgn_nthreads; /* # live threads */
780 int kgn_nresets; /* number of stack resets */
781 int kgn_in_reset; /* are we in stack reset ? */
783 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
785 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
786 int kgn_ndevs; /* # devices */
788 int kgn_ruhroh_running; /* ruhroh thread is running */
789 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
790 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
791 int kgn_quiesce_trigger; /* should we quiesce ? */
792 atomic_t kgn_nquiesce; /* how many quiesced ? */
793 struct semaphore kgn_quiesce_sem; /* serialize ruhroh task, startup and shutdown */
794 int kgn_needs_reset; /* we need stack reset */
796 /* These next three members implement communication from gnilnd into
797 * the ruhroh task. To ensure correct operation of the task, code that
798 * writes into them must use memory barriers to ensure that the changes
799 * are visible to other cores in the order the members appear below. */
800 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
801 int kgn_bump_info_rdy; /* we have info needed to bump */
802 int kgn_needs_pause; /* we need to pause for network quiesce */
804 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
805 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
807 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
808 struct list_head *kgn_peers; /* hash table of all my known peers */
809 atomic_t kgn_npeers; /* # peers extant */
810 int kgn_peer_version; /* version flag for peer tables */
812 struct list_head *kgn_conns; /* conns hashed by cqid */
813 atomic_t kgn_nconns; /* # connections extant */
814 __u64 kgn_peerstamp; /* when I started up */
815 __u64 kgn_connstamp; /* conn stamp generator */
816 int kgn_conn_version; /* version flag for conn tables */
817 int kgn_next_cqid; /* cqid generator */
819 long kgn_new_min_timeout; /* minimum timeout on any new conn */
820 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
821 spinlock_t kgn_reaper_lock; /* serialise */
823 struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
824 struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
825 struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
826 atomic_t kgn_ntx; /* # tx in use */
827 struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
829 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
830 __u64 kgn_cksum_npages; /* Number of pages allocated for checksumming */
831 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
832 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
834 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
835 long kgn_rdmaq_override; /* bytes per second override */
837 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
839 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
840 atomic_t kgn_npending_conns; /* # of conns with pending closes */
841 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
842 unsigned long kgn_last_scheduled; /* last time schedule was called in a sched thread */
843 unsigned long kgn_last_condresched; /* last time cond_resched was called in a sched thread */
844 atomic_t kgn_rev_offset; /* number of time REV rdma have been misaligned offsets */
845 atomic_t kgn_rev_length; /* Number of times REV rdma have been misaligned lengths */
846 atomic_t kgn_rev_copy_buff; /* Number of times REV rdma have had to make a copy buffer */
849 extern kgn_data_t kgnilnd_data;
850 extern kgn_tunables_t kgnilnd_tunables;
852 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
853 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
854 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
856 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
857 * and the line of the calling function to allow us to debug problematic
858 * schedule calls in the future without the programmer having to mark
859 * the location manually.
861 #define kgnilnd_schedule_conn(conn) \
862 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
864 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
865 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
868 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
870 struct task_struct *thrd = kthread_run(fn, arg, "%s_%02d", name, id);
872 return PTR_ERR(thrd);
874 atomic_inc(&kgnilnd_data.kgn_nthreads);
879 kgnilnd_thread_fini(void)
881 atomic_dec(&kgnilnd_data.kgn_nthreads);
884 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
885 * parts of the code to avoid a scheduler trip when the mutex is held
887 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
888 * has been acquired successfully, and 0 on contention.
890 * NOTE: this function follows the spin_trylock() convention, so
891 * it is negated to the down_trylock() return values! Be careful
892 * about this when converting semaphore users to mutexes.
894 * This function must not be used in interrupt context. The
895 * mutex must be released by the same task that acquired it.
897 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
900 unsigned long timeout;
902 LASSERT(!in_interrupt());
904 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
906 ret = mutex_trylock(lock);
913 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
916 _kgnilnd_debug_msg(kgn_msg_t *msg,
917 struct libcfs_debug_msg_data *data, const char *fmt, ... );
919 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
921 CFS_CHECK_STACK(msgdata, mask, cdls); \
923 if (((mask) & D_CANTMASK) != 0 || \
924 ((libcfs_debug & (mask)) != 0 && \
925 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
926 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
929 /* for most callers (level is a constant) this is resolved at compile time */
930 #define GNIDBG_MSG(level, msg, fmt, args...) \
932 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
933 static cfs_debug_limit_state_t cdls; \
934 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
935 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
936 "$$ "fmt" from %s ", ## args, \
937 libcfs_nid2str((msg)->gnm_srcnid)); \
939 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
940 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
941 "$$ "fmt" from %s ", ## args, \
942 libcfs_nid2str((msg)->gnm_srcnid)); \
946 /* user puts 'to nid' in msg for us */
947 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
949 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
950 static cfs_debug_limit_state_t cdls; \
951 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
952 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
953 "$$ "fmt" ", ## args); \
955 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
956 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
957 "$$ "fmt" ", ## args); \
962 _kgnilnd_debug_conn(kgn_conn_t *conn,
963 struct libcfs_debug_msg_data *data, const char *fmt, ... );
965 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
967 CFS_CHECK_STACK(msgdata, mask, cdls); \
969 if (((mask) & D_CANTMASK) != 0 || \
970 ((libcfs_debug & (mask)) != 0 && \
971 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
972 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
975 /* for most callers (level is a constant) this is resolved at compile time */
976 #define GNIDBG_CONN(level, conn, fmt, args...) \
978 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
979 static cfs_debug_limit_state_t cdls; \
980 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
981 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
982 "$$ "fmt" ", ## args); \
984 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
985 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
986 "$$ "fmt" ", ## args); \
991 _kgnilnd_debug_tx(kgn_tx_t *tx,
992 struct libcfs_debug_msg_data *data, const char *fmt, ... );
994 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
996 CFS_CHECK_STACK(msgdata, mask, cdls); \
998 if (((mask) & D_CANTMASK) != 0 || \
999 ((libcfs_debug & (mask)) != 0 && \
1000 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1001 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1004 /* for most callers (level is a constant) this is resolved at compile time */
1005 #define GNIDBG_TX(level, tx, fmt, args...) \
1007 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1008 static cfs_debug_limit_state_t cdls; \
1009 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1010 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1011 "$$ "fmt" ", ## args); \
1013 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1014 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1015 "$$ "fmt" ", ## args); \
1019 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1021 if (unlikely(!(cond))) { \
1022 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1027 #define GNILND_IS_QUIESCED \
1028 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1029 atomic_read(&kgnilnd_data.kgn_nthreads))
1031 #define KGNILND_SPIN_QUIESCE \
1033 /* E.T phone home */ \
1034 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1035 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1036 while (kgnilnd_data.kgn_quiesce_trigger) { \
1037 set_current_state(TASK_INTERRUPTIBLE); \
1038 schedule_timeout(HZ); \
1040 /* Mom, my homework is done */ \
1041 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1042 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1045 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1046 #ifndef LIBCFS_DEBUG
1047 #error "this code uses actions inside LASSERT for ref counting"
1050 #define kgnilnd_admin_addref(atomic) \
1052 int val = atomic_inc_return(&atomic); \
1053 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1054 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1057 #define kgnilnd_admin_decref(atomic) \
1059 int val = atomic_dec_return(&atomic); \
1060 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
1061 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1064 #define kgnilnd_net_addref(net) \
1066 int val = atomic_inc_return(&net->gnn_refcount); \
1067 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1068 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1069 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1072 #define kgnilnd_net_decref(net) \
1074 int val = atomic_dec_return(&net->gnn_refcount); \
1075 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1076 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1077 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1080 #define kgnilnd_peer_addref(peer) \
1082 int val = atomic_inc_return(&peer->gnp_refcount); \
1083 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1084 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1085 libcfs_nid2str(peer->gnp_nid), val); \
1088 #define kgnilnd_peer_decref(peer) \
1090 int val = atomic_dec_return(&peer->gnp_refcount); \
1091 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1092 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1093 libcfs_nid2str(peer->gnp_nid), val); \
1095 kgnilnd_destroy_peer(peer); \
1098 #define kgnilnd_conn_addref(conn) \
1103 val = atomic_inc_return(&conn->gnc_refcount); \
1104 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1105 "conn %p refc %d to %s\n", \
1108 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1110 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1112 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1117 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1118 * through the scheduler thread to get the EP destroyed. This avoids some
1119 * messy semaphore business and allows us to reuse the connd_list and existing
1120 * linkage and avoid creating extra lists just for destroying EPs */
1122 /* Safety Disclaimer:
1123 * Q: If we decrement the refcount and then check it again, is it possible that
1124 * another caller could have passed through this macro concurrently? If so,
1125 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1127 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1128 * once we are refcount <= 2. It hinges around gnc_state and membership of
1129 * gnc_hashlist. There are two ways to find a connection - either ask for
1130 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1131 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1134 * - #1 from create (kgnilnd_create_conn)
1135 * - #2 for EP (kgnilnd_create_conn)
1136 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1137 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1139 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1140 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1142 * There could be more from TX descriptors during the lifetime of a live
1145 * If we nuke the conn before finish_connect, we won't have parallel paths
1146 * because nobody besides the dgram handler for the single outstanding
1147 * dgram can find the connection as it isn't in any searchable tables yet.
1149 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1150 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1151 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1152 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1153 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1154 * sent. Now the gnc_state == CLOSED, so we hit
1155 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1156 * - we've nuked them from the peer and CQ id tables, so we own them and
1157 * are guaranteed serial access - hence the complete lack of conn list
1158 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1159 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1160 * kgnilnd_process_conns. Then the next scheduler loop would call
1161 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1162 * kgnilnd_process_conns.
1164 * Clearly, we are totally safe. Clearly.
1167 #define kgnilnd_conn_decref(conn) \
1172 val = atomic_dec_return(&conn->gnc_refcount); \
1173 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1176 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1178 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1180 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1185 (conn->gnc_ephandle != NULL) && \
1186 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1187 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1188 kgnilnd_schedule_conn(conn); \
1189 } else if (val == 0) { \
1190 kgnilnd_destroy_conn(conn); \
1194 static inline struct list_head *
1195 kgnilnd_nid2peerlist(lnet_nid_t nid)
1197 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1199 RETURN(&kgnilnd_data.kgn_peers[hash]);
1202 static inline struct list_head *
1203 kgnilnd_netnum2netlist(__u16 netnum)
1205 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1207 RETURN(&kgnilnd_data.kgn_nets[hash]);
1211 kgnilnd_peer_active(kgn_peer_t *peer)
1213 /* Am I in the peer hash table? */
1214 return (!list_empty(&peer->gnp_list));
1217 /* need write_lock on kgn_peer_conn_lock */
1219 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1221 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1222 peer, libcfs_nid2str(peer->gnp_nid),
1223 !list_empty(&peer->gnp_conns),
1224 !list_empty(&peer->gnp_tx_queue));
1226 /* kgn_peer_conn_lock protects us from conflict with
1227 * kgnilnd_peer_notify and gnp_persistent */
1228 RETURN ((list_empty(&peer->gnp_conns)) &&
1229 (list_empty(&peer->gnp_tx_queue)));
1232 /* returns positive if error was for a clean shutdown of conn */
1234 kgnilnd_conn_clean_errno(int errno)
1236 /* - ESHUTDOWN - LND is unloading
1237 * - EUCLEAN - admin requested via "lctl del_peer"
1238 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1239 * - ENOTRECOVERABLE - stack reset
1240 * - EISCONN - cleared via "lctl push"
1241 * not doing ESTALE - that isn't clean */
1242 RETURN ((errno == 0) ||
1243 (errno == -ESHUTDOWN) ||
1244 (errno == -EUCLEAN) ||
1245 (errno == -ENETRESET) ||
1246 (errno == -EISCONN) ||
1247 (errno == -ENOTRECOVERABLE));
1250 /* returns positive if error results in purgatory hold */
1252 kgnilnd_check_purgatory_errno(int errno)
1254 /* We don't want to save the purgatory lists these cases:
1255 * - EUCLEAN - admin requested via "lctl del_peer"
1256 * - ESHUTDOWN - LND is unloading
1258 RETURN ((errno != -ESHUTDOWN) &&
1259 (errno != -EUCLEAN));
1263 /* returns positive if a purgatory hold is needed */
1265 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1269 if (conn->gnc_peer) {
1270 loopback = conn->gnc_peer->gnp_nid ==
1271 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1273 /* short circuit - a conn that didn't complete
1274 * setup never needs a purgatory hold */
1277 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1278 conn, conn->gnc_peer ?
1279 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1281 loopback, conn->gnc_close_recvd);
1283 /* we only use a purgatory hold if we've not received the CLOSE msg
1284 * from our peer - without that message, we can't know the state of
1285 * the other end of this connection and must put it into purgatory
1286 * to prevent reuse and corruption.
1287 * The theory is that a TX error can be communicated in all other cases
1289 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1290 kgnilnd_check_purgatory_errno(conn->gnc_error));
1293 static inline const char *
1294 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1296 static inline struct list_head *
1297 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1298 kgn_tx_list_state_t to_state)
1301 case GNILND_TX_PEERQ:
1302 return &peer->gnp_tx_queue;
1303 case GNILND_TX_FMAQ:
1304 return &conn->gnc_fmaq;
1305 case GNILND_TX_LIVE_FMAQ:
1306 case GNILND_TX_LIVE_RDMAQ:
1307 case GNILND_TX_DYING:
1309 case GNILND_TX_MAPQ:
1310 return &conn->gnc_device->gnd_map_tx;
1311 case GNILND_TX_RDMAQ:
1312 return &conn->gnc_device->gnd_rdmaq;
1314 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1315 CERROR("invalid state requested: %s\n",
1316 kgnilnd_tx_state2str(to_state));
1322 /* should hold tx, conn or peer lock when calling */
1324 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1325 kgn_conn_t *conn, kgn_tx_list_state_t state,
1328 struct list_head *list = NULL;
1330 /* make sure we have a sane TX state to start */
1331 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1332 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1333 list_empty(&tx->tx_list),
1334 "bad state with tx_list %s",
1335 list_empty(&tx->tx_list) ? "empty" : "not empty");
1337 /* WTF - you are already on that state buttmunch */
1338 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1339 "already at %s", kgnilnd_tx_state2str(state));
1341 /* get proper list from the state requested */
1342 list = kgnilnd_tx_state2list(peer, conn, state);
1346 case GNILND_TX_PEERQ:
1347 kgnilnd_peer_addref(peer);
1349 case GNILND_TX_ALLOCD:
1350 /* no refs needed */
1352 case GNILND_TX_FMAQ:
1353 kgnilnd_conn_addref(conn);
1355 case GNILND_TX_MAPQ:
1356 atomic_inc(&conn->gnc_device->gnd_nq_map);
1357 kgnilnd_conn_addref(conn);
1359 case GNILND_TX_LIVE_FMAQ:
1360 atomic_inc(&conn->gnc_nlive_fma);
1361 kgnilnd_conn_addref(conn);
1363 case GNILND_TX_LIVE_RDMAQ:
1364 atomic_inc(&conn->gnc_nlive_rdma);
1365 kgnilnd_conn_addref(conn);
1367 case GNILND_TX_RDMAQ:
1368 atomic_inc(&conn->gnc_nq_rdma);
1369 kgnilnd_conn_addref(conn);
1371 case GNILND_TX_DYING:
1372 kgnilnd_conn_addref(conn);
1375 CERROR("invalid state requested: %s\n",
1376 kgnilnd_tx_state2str(state));
1381 /* if this changes, change kgnilnd_alloc_tx */
1382 tx->tx_list_state = state;
1384 /* some states don't have lists - we track them in the per conn
1385 * TX table instead. Waste not, want not! */
1387 tx->tx_list_p = list;
1389 list_add_tail(&tx->tx_list, list);
1391 list_add(&tx->tx_list, list);
1393 /* set dummy list_p to make book keeping happy and let debugging
1394 * be a hair easier */
1395 tx->tx_list_p = (void *)state;
1398 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1399 kgnilnd_tx_state2str(state), list);
1403 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1404 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1406 /* These is only 1 "off-list" state */
1407 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1408 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1410 /* new_state == ALLOCD means we are deallocating this tx,
1411 * so make sure it was on a valid list to start with */
1412 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1413 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1414 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1415 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1418 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1421 switch (tx->tx_list_state) {
1422 case GNILND_TX_PEERQ:
1423 kgnilnd_peer_decref(peer);
1425 case GNILND_TX_FREED:
1426 case GNILND_TX_IDLE:
1427 case GNILND_TX_ALLOCD:
1428 /* no refs needed */
1430 case GNILND_TX_DYING:
1431 kgnilnd_conn_decref(conn);
1433 case GNILND_TX_FMAQ:
1434 kgnilnd_conn_decref(conn);
1436 case GNILND_TX_MAPQ:
1437 atomic_dec(&conn->gnc_device->gnd_nq_map);
1438 kgnilnd_conn_decref(conn);
1440 case GNILND_TX_LIVE_FMAQ:
1441 atomic_dec(&conn->gnc_nlive_fma);
1442 kgnilnd_conn_decref(conn);
1444 case GNILND_TX_LIVE_RDMAQ:
1445 atomic_dec(&conn->gnc_nlive_rdma);
1446 kgnilnd_conn_decref(conn);
1448 case GNILND_TX_RDMAQ:
1449 atomic_dec(&conn->gnc_nq_rdma);
1450 kgnilnd_conn_decref(conn);
1451 /* don't need to assert on default, already did in set */
1454 /* for ALLOCD, this might already be true, but no harm doing it again */
1455 list_del_init(&tx->tx_list);
1456 tx->tx_list_p = NULL;
1457 tx->tx_list_state = new_state;
1461 kgnilnd_tx_mapped(kgn_tx_t *tx)
1463 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1464 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1467 static inline struct list_head *
1468 kgnilnd_cqid2connlist(__u32 cqid)
1470 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1472 return (&kgnilnd_data.kgn_conns [hash]);
1475 static inline kgn_conn_t *
1476 kgnilnd_cqid2conn_locked(__u32 cqid)
1478 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1479 struct list_head *tmp;
1482 list_for_each(tmp, conns) {
1483 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1485 if (conn->gnc_cqid == cqid)
1492 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1494 kgnilnd_get_cqid_locked(void)
1500 cqid = kgnilnd_data.kgn_next_cqid++;
1501 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1505 kgnilnd_data.kgn_next_cqid = 1;
1508 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1514 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1516 kgn_tx_t *tx = NULL;
1517 kgn_conn_t *conn = NULL;
1519 /* set to NULL so any early return is an error */
1523 LASSERTF((ev_id->txe_idx > 0) &&
1524 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1525 "bogus txe_idx %d >= %d\n",
1526 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1528 LASSERTF((ev_id->txe_cqid > 0) &&
1529 (ev_id->txe_cqid < GNILND_MAX_CQID),
1530 "bogus txe_cqid %d >= %d\n",
1531 ev_id->txe_cqid, GNILND_MAX_CQID);
1533 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1534 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1537 /* Conn was destroyed? */
1538 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1539 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1542 /* just insurance */
1543 kgnilnd_conn_addref(conn);
1544 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1545 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1547 /* we know this is safe - as the TX won't be reused until AFTER
1548 * the conn is unlinked from the cqid hash, so we can use the TX
1549 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1551 spin_lock(&conn->gnc_tx_lock);
1552 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1553 spin_unlock(&conn->gnc_tx_lock);
1555 /* We could have a tx that was cleared out by other forces
1556 * lctl disconnect or del_peer. */
1558 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1559 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1560 kgnilnd_conn_decref(conn);
1564 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1565 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1566 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1568 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1569 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1570 "found tx %p txe_idx %d",
1571 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1572 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1574 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1576 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1581 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1584 /* set_normalized_timepsec isn't exported from the kernel, so
1585 * we need to do the same thing inline */
1586 static inline struct timespec
1587 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1593 sec = lhs.tv_sec - rhs.tv_sec;
1594 nsec = lhs.tv_nsec - rhs.tv_nsec;
1596 while (nsec >= NSEC_PER_SEC) {
1597 nsec -= NSEC_PER_SEC;
1601 nsec += NSEC_PER_SEC;
1610 kgnilnd_count_list(struct list_head *q)
1612 struct list_head *e;
1615 list_for_each(e, q) {
1622 /* kgnilnd_find_net adds a reference to the net it finds
1623 * this is so the net will not be removed before the calling function
1624 * has time to use the data returned. This reference needs to be released
1625 * by the calling function once it has finished using the returned net
1629 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1634 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1640 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1641 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1642 kgnilnd_net_addref(net);
1643 up_read(&kgnilnd_data.kgn_net_rw_sem);
1649 up_read(&kgnilnd_data.kgn_net_rw_sem);
1654 #ifdef CONFIG_DEBUG_SLAB
1655 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1657 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1660 int kgnilnd_dev_init(kgn_device_t *dev);
1661 void kgnilnd_dev_fini(kgn_device_t *dev);
1662 int kgnilnd_startup(lnet_ni_t *ni);
1663 void kgnilnd_shutdown(lnet_ni_t *ni);
1664 int kgnilnd_base_startup(void);
1665 void kgnilnd_base_shutdown(void);
1667 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1668 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1669 void kgnilnd_unmap_phys_fmablk(kgn_device_t *device);
1670 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1672 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1673 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1674 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1675 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1676 lnet_msg_t *lntmsg, void **new_private);
1677 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1678 int delayed, unsigned int niov,
1679 struct iovec *iov, lnet_kiov_t *kiov,
1680 unsigned int offset, unsigned int mlen, unsigned int rlen);
1682 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1684 /* purgatory functions */
1685 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1686 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1687 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1688 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1690 void kgnilnd_update_reaper_timeout(long timeout);
1691 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1692 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1693 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1694 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1695 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1696 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
1697 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1699 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1700 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net);
1701 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1702 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1704 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1705 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1706 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1707 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1708 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1709 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1710 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1711 void kgnilnd_consume_rx(kgn_rx_t *rx);
1713 void kgnilnd_schedule_device(kgn_device_t *dev);
1714 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1715 void kgnilnd_schedule_device_timer(unsigned long arg);
1717 int kgnilnd_reaper(void *arg);
1718 int kgnilnd_scheduler(void *arg);
1719 int kgnilnd_dgram_mover(void *arg);
1720 int kgnilnd_rca(void *arg);
1722 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1723 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1724 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1725 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1726 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1727 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1728 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1729 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1730 void kgnilnd_peer_alive(kgn_peer_t *peer);
1731 void kgnilnd_peer_notify(kgn_peer_t *peer, int error);
1732 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1733 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1734 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1735 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1737 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1738 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1739 void kgnilnd_wakeup_rca_thread(void);
1740 int kgnilnd_start_rca_thread(void);
1742 int kgnilnd_tunables_init(void);
1743 void kgnilnd_tunables_fini(void);
1744 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1746 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1747 void kgnilnd_pause_threads(void);
1748 int kgnilnd_hw_in_quiesce(void);
1749 int kgnilnd_check_hw_quiesce(void);
1750 void kgnilnd_quiesce_wait(char *reason);
1751 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1752 int kgnilnd_ruhroh_thread(void *arg);
1753 void kgnilnd_reset_stack(void);
1754 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1756 void kgnilnd_insert_sysctl(void);
1757 void kgnilnd_remove_sysctl(void);
1758 void kgnilnd_proc_init(void);
1759 void kgnilnd_proc_fini(void);
1762 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1764 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1765 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1766 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram);
1768 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1769 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1770 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1771 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1773 int kgnilnd_dgram_waitq(void *arg);
1775 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1777 /* struct2str functions - we don't use a default: case to cause the compile
1778 * to fail if there is a missing case. This allows us to hide these down here
1779 * out of the way but ensure we'll catch any updates to the enum/types
1782 #define DO_TYPE(x) case x: return #x;
1783 static inline const char *
1784 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1786 /* Only want single char string for this */
1788 case GNILND_FMABLK_IDLE:
1790 case GNILND_FMABLK_PHYS:
1792 case GNILND_FMABLK_VIRT:
1794 case GNILND_FMABLK_FREED:
1797 return "<unknown state>";
1800 static inline const char *
1801 kgnilnd_msgtype2str(int type)
1804 DO_TYPE(GNILND_MSG_NONE);
1805 DO_TYPE(GNILND_MSG_NOOP);
1806 DO_TYPE(GNILND_MSG_IMMEDIATE);
1807 DO_TYPE(GNILND_MSG_PUT_REQ);
1808 DO_TYPE(GNILND_MSG_PUT_NAK);
1809 DO_TYPE(GNILND_MSG_PUT_ACK);
1810 DO_TYPE(GNILND_MSG_PUT_DONE);
1811 DO_TYPE(GNILND_MSG_GET_REQ);
1812 DO_TYPE(GNILND_MSG_GET_NAK);
1813 DO_TYPE(GNILND_MSG_GET_DONE);
1814 DO_TYPE(GNILND_MSG_CLOSE);
1815 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1816 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1817 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1818 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1819 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1820 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1821 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1823 return "<unknown msg type>";
1826 static inline const char *
1827 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1830 DO_TYPE(GNILND_TX_IDLE);
1831 DO_TYPE(GNILND_TX_ALLOCD);
1832 DO_TYPE(GNILND_TX_PEERQ);
1833 DO_TYPE(GNILND_TX_MAPQ);
1834 DO_TYPE(GNILND_TX_FMAQ);
1835 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1836 DO_TYPE(GNILND_TX_RDMAQ);
1837 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1838 DO_TYPE(GNILND_TX_DYING);
1839 DO_TYPE(GNILND_TX_FREED);
1841 return "<unknown state>";
1844 static inline const char *
1845 kgnilnd_conn_state2str(kgn_conn_t *conn)
1847 kgn_conn_state_t state = conn->gnc_state;
1849 DO_TYPE(GNILND_CONN_DUMMY);
1850 DO_TYPE(GNILND_CONN_LISTEN);
1851 DO_TYPE(GNILND_CONN_CONNECTING);
1852 DO_TYPE(GNILND_CONN_ESTABLISHED);
1853 DO_TYPE(GNILND_CONN_CLOSING);
1854 DO_TYPE(GNILND_CONN_CLOSED);
1855 DO_TYPE(GNILND_CONN_DONE);
1856 DO_TYPE(GNILND_CONN_DESTROY_EP);
1861 static inline const char *
1862 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1864 kgn_connreq_type_t type = connreq->gncr_type;
1867 DO_TYPE(GNILND_CONNREQ_REQ);
1868 DO_TYPE(GNILND_CONNREQ_NAK);
1869 DO_TYPE(GNILND_CONNREQ_CLOSE);
1874 static inline const char *
1875 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1877 kgn_dgram_state_t state = dgram->gndg_state;
1880 DO_TYPE(GNILND_DGRAM_USED);
1881 DO_TYPE(GNILND_DGRAM_POSTING);
1882 DO_TYPE(GNILND_DGRAM_POSTED);
1883 DO_TYPE(GNILND_DGRAM_PROCESSING);
1884 DO_TYPE(GNILND_DGRAM_DONE);
1885 DO_TYPE(GNILND_DGRAM_CANCELED);
1890 static inline const char *
1891 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1893 kgn_dgram_type_t type = dgram->gndg_type;
1896 DO_TYPE(GNILND_DGRAM_REQ);
1897 DO_TYPE(GNILND_DGRAM_WC_REQ);
1898 DO_TYPE(GNILND_DGRAM_NAK);
1899 DO_TYPE(GNILND_DGRAM_CLOSE);
1904 static inline const char *
1905 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
1908 DO_TYPE(GNILND_DGRAM_REQ);
1909 DO_TYPE(GNILND_DGRAM_WC_REQ);
1910 DO_TYPE(GNILND_DGRAM_NAK);
1911 DO_TYPE(GNILND_DGRAM_CLOSE);
1918 /* API wrapper functions - include late to pick up all of the other defines */
1919 #include "gnilnd_api_wrap.h"
1921 /* pulls in tunables per platform and adds in nid/nic conversion
1922 * if RCA wasn't available at build time */
1923 #include "gnilnd_hss_ops.h"
1925 #if defined(CONFIG_CRAY_GEMINI)
1926 #include "gnilnd_gemini.h"
1927 #elif defined(CONFIG_CRAY_ARIES)
1928 #include "gnilnd_aries.h"
1930 #error "Undefined Network Hardware Type"
1933 #endif /* _GNILND_GNILND_H_ */