2 * Copyright (C) 2004 Cluster File Systems, Inc.
4 * Copyright (C) 2009-2012 Cray, Inc.
6 * Copyright (c) 2013, 2014, Intel Corporation.
8 * Derived from work by: Eric Barton <eric@bartonsoftware.com>
9 * Author: Nic Henke <nic@cray.com>
10 * Author: James Shimek <jshimek@cray.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 #ifndef _GNILND_GNILND_H_
29 #define _GNILND_GNILND_H_
31 #ifdef HAVE_COMPAT_RDMA
32 #include <linux/compat-2.6.h>
34 #include <linux/module.h>
35 #include <linux/kernel.h>
37 #include <linux/string.h>
38 #include <linux/stat.h>
39 #include <linux/errno.h>
40 #ifdef HAVE_LINUX_KERNEL_LOCK
41 #include <linux/smp_lock.h>
43 #include <linux/unistd.h>
44 #include <linux/uio.h>
45 #include <linux/time.h>
46 #include <asm/timex.h>
48 #include <asm/uaccess.h>
51 #include <linux/init.h>
53 #include <linux/file.h>
54 #include <linux/stat.h>
55 #include <linux/list.h>
56 #include <linux/kmod.h>
57 #include <linux/sysctl.h>
58 #include <linux/kthread.h>
59 #include <linux/nmi.h>
63 #include <linux/nmi.h>
65 #define DEBUG_SUBSYSTEM S_LND
67 #include <libcfs/linux/kp30.h>
68 #include <libcfs/libcfs.h>
69 #include <lnet/lnet.h>
70 #include <lnet/lib-lnet.h>
73 #include "gnilnd_version.h"
76 /* tunables determined at compile time */
77 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
78 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
79 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
80 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
81 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
82 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
83 #define GNILND_SCHED_TIMEOUT 1
84 #define GNILND_DGRAM_TIMEOUT 2
85 #define GNILND_FAST_MAPPING_TRY \
86 *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
87 #define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
89 /* map failure timeout */
90 #define GNILND_MAP_TIMEOUT \
91 (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
92 *kgnilnd_tunables.kgn_timeout))
94 /* reaper thread wakup interval */
95 #define GNILND_REAPER_THREAD_WAKE 1
96 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
97 #define GNILND_REAPER_NCHECKS 4
100 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
101 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
102 #define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
103 /* checksum values */
104 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
105 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
106 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
107 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
109 /* tune down some COMPUTE options as they won't see the same number of connections and
110 * don't need the throughput of multiple threads by default */
111 #if defined(CONFIG_CRAY_COMPUTE)
113 #define GNILND_SCHED_THREADS 2 /* default # of kgnilnd_scheduler threads */
115 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
117 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
118 #define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
119 #define GNILND_COMPUTE 1 /* compute image */
120 #define GNILND_FAST_RECONNECT 1 /* Fast Reconnect option */
122 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
123 #define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
124 #define GNILND_COMPUTE 0 /* service image */
125 #define GNILND_FAST_RECONNECT 0 /* Fast Reconnect option */
128 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
129 #define GNILND_EXTRA_BITS 1
130 /* maximum number of conns & bits for cqid in the SMSG event data */
131 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
132 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
133 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
134 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
135 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
137 /* need sane upper bound to limit copy overhead */
138 #define GNILND_MAX_IMMEDIATE (64<<10)
140 /* Max number of connections to keep in purgatory per peer */
141 #define GNILND_PURGATORY_MAX 5
143 /* payload size to add to the base mailbox size
144 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
145 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
146 * the calculation return from that function.*/
147 #define GNILND_MBOX_PAYLOAD \
148 (GNILND_MAX_MSG_SIZE * \
149 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
151 /* timeout -> deadman timer for kgni mdd holds */
152 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
154 /* timeout for failing sends in t is in jiffies*/
155 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
157 /* time when to release from purgatory in the reaper thread in jiffies */
158 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
160 /* Macro for finding last_rx 2 datapoints are compared
161 * and the most recent one in jiffies is returned.
163 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
164 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
166 /************************************************************************
167 * Enum, flag and tag data
169 #define GNILND_INIT_NOTHING 0
170 #define GNILND_INIT_DATA 1
171 #define GNILND_INIT_ALL 2
173 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
174 #define GNILND_BUF_NONE 0 /* buffer type not set */
175 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
176 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
177 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
178 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
179 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
180 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
182 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
183 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
184 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
185 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
186 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
188 /* stash above max CQID to avoid any collision */
189 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
190 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
192 /* kgn_msg_t::gnm_type */
193 #define GNILND_MSG_NONE 0x00 /* illegal message */
194 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
195 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
196 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
197 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
198 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
199 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
200 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
201 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
202 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
203 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
204 #define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
205 #define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
206 #define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
207 #define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
208 #define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
209 #define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
210 #define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
212 /* defines for gnc_*scheduled states */
213 #define GNILND_CONN_IDLE 0
214 #define GNILND_CONN_SCHED 1
215 #define GNILND_CONN_WANTS_SCHED 2
216 #define GNILND_CONN_PROCESS 3
218 #define GNILND_DEV_IDLE 0
219 #define GNILND_DEV_IRQ 1
220 #define GNILND_DEV_LOOP 2
222 #define GNILND_DGRAM_IDLE 0
223 #define GNILND_DGRAM_SCHED 1
224 #define GNILND_DGRAM_PROCESS 2
226 #define GNILND_PEER_IDLE 0
227 #define GNILND_PEER_CONNECT 1
228 #define GNILND_PEER_POSTING 2
229 #define GNILND_PEER_POSTED 3
230 #define GNILND_PEER_NEEDS_DEATH 4
231 #define GNILND_PEER_KILL 5
233 /* for gnc_close_recvd */
234 #define GNILND_CLOSE_RX 1
235 #define GNILND_CLOSE_INJECT1 2
236 #define GNILND_CLOSE_INJECT2 3
237 #define GNILND_CLOSE_EARLY 4
239 /* defines for why quiesce trigger set */
240 #define GNILND_QUIESCE_IDLE 0
241 #define GNILND_QUIESCE_ADMIN 1
242 #define GNILND_QUIESCE_RESET 2
243 #define GNILND_QUIESCE_HW_QUIESCE 3
245 #define GNILND_PEER_CLEAN 0
246 #define GNILND_PEER_PERSISTING 1
248 #define GNILND_DEL_CONN 0
249 #define GNILND_DEL_PEER 1
250 #define GNILND_CLEAR_PURGATORY 2
252 #define GNILND_RCA_NODE_UP 0
253 #define GNILND_RCA_NODE_DOWN 1
254 #define GNILND_RCA_NODE_UNKNOWN 2
256 /* defines for reverse RDMA states */
257 #define GNILND_REVERSE_NONE 0
258 #define GNILND_REVERSE_GET 1
259 #define GNILND_REVERSE_PUT 2
260 #define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
262 typedef enum kgn_fmablk_state {
263 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
264 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
265 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
266 GNILND_FMABLK_FREED, /* after free */
267 } kgn_fmablk_state_t;
269 typedef enum kgn_tx_list_state {
270 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
271 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
272 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
273 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
274 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
275 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
276 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
277 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
278 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
279 GNILND_TX_FREED /* TX is free! */
280 } kgn_tx_list_state_t;
282 typedef enum kgn_conn_state {
283 /* don't start @ 0 - prevent memset(0) badness */
284 GNILND_CONN_DUMMY = 0,
286 GNILND_CONN_CONNECTING,
287 GNILND_CONN_ESTABLISHED,
291 GNILND_CONN_DESTROY_EP
294 /* changing these requires a change to GNILND_CONNREQ_VERSION and
295 * will result in dropped packets instead of NAKs. Adding to this is
296 * acceptable without changing the CONNREQ_VERSION, but code should
297 * be ready to handle NAKs on version mismatch */
298 typedef enum kgn_connreq_type {
299 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
300 GNILND_CONNREQ_NAK, /* NO soup for you! */
301 GNILND_CONNREQ_CLOSE, /* we should see other people */
302 } kgn_connreq_type_t;
304 typedef enum kgn_dgram_state {
305 /* don't use 0 to avoid thinking a memset of zero is valid data */
306 GNILND_DGRAM_USED = 1,
307 GNILND_DGRAM_POSTING,
309 GNILND_DGRAM_PROCESSING,
310 GNILND_DGRAM_CANCELED,
314 typedef enum kgn_dgram_type {
315 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
316 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
317 GNILND_DGRAM_NAK, /* NO soup for you! */
318 GNILND_DGRAM_CLOSE, /* we should see other people */
321 /************************************************************************
322 * Wire message structs. These are sent in sender's byte order
323 * (i.e. receiver checks magic and flips if required).
326 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
327 #define GNILND_DGRAM_MAGIC 0x0DDBA11
329 /* kgn_msg_t - FMA/SMSG wire struct
331 * - added checksum to FMA
332 * moved seq before paylod
333 * WIRE_ATTR added for alignment
335 * added gnm_payload_len for FMA payload size
337 * added gncm_retval to completion, allowing return code transmission
340 * changed how CQID and TX ids are assigned
342 * added retval on CLOSE
344 * added payload checksumming
346 * reworked checksumming a bit, changed payload checksums
348 #define GNILND_MSG_VERSION 8
349 /* kgn_connreq_t connection request datagram wire struct
354 #define GNILND_CONNREQ_VERSION 2
356 typedef struct kgn_gniparams {
357 __u32 gnpr_host_id; /* ph. host ID of the NIC */
358 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
359 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
360 } WIRE_ATTR kgn_gniparams_t;
362 typedef struct kgn_nak_data {
363 __s32 gnnd_errno; /* errno reason for NAK */
365 } WIRE_ATTR kgn_nak_data_t;
367 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
368 * without breaking the ability for us to properly NAK someone */
369 typedef struct kgn_connreq { /* connection request/response */
370 __u32 gncr_magic; /* I'm an gnilnd connreq */
371 __u32 gncr_cksum; /* checksum (0 == disabled) */
372 __u16 gncr_type; /* REQ, NAK, etc */
373 __u16 gncr_version; /* this is my version number */
374 __u32 gncr_timeout; /* sender's timeout */
375 __u64 gncr_srcnid; /* sender's NID */
376 __u64 gncr_dstnid; /* who sender expects to listen */
377 __u64 gncr_peerstamp; /* sender's instance stamp */
378 __u64 gncr_connstamp; /* sender's connection stamp */
380 /* everything before this needs to stay static, adding after should
381 * result in a change to GNILND_CONNREQ_VERSION */
384 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
385 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
387 } WIRE_ATTR kgn_connreq_t;
390 gni_mem_handle_t gnrd_key;
393 } WIRE_ATTR kgn_rdma_desc_t;
396 lnet_hdr_t gnim_hdr; /* LNet header */
397 /* LNet payload is in FMA "Message Data" */
398 } WIRE_ATTR kgn_immediate_msg_t;
401 lnet_hdr_t gnprm_hdr; /* LNet header */
402 __u64 gnprm_cookie; /* opaque completion cookie */
403 } WIRE_ATTR kgn_putreq_msg_t;
406 __u64 gnpam_src_cookie; /* reflected completion cookie */
407 __u64 gnpam_dst_cookie; /* opaque completion cookie */
408 __u16 gnpam_payload_cksum; /* checksum for get msg */
409 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
410 } WIRE_ATTR kgn_putack_msg_t;
413 lnet_hdr_t gngm_hdr; /* LNet header */
414 __u64 gngm_cookie; /* opaque completion cookie */
415 __u16 gngm_payload_cksum; /* checksum for put msg */
416 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
417 } WIRE_ATTR kgn_get_msg_t;
420 int gncm_retval; /* error on NAK, size on REQ */
421 __u64 gncm_cookie; /* reflected completion cookie */
422 } WIRE_ATTR kgn_completion_msg_t;
424 typedef struct { /* NB must fit in FMA "Prefix" */
425 __u32 gnm_magic; /* I'm an gni message */
426 __u16 gnm_version; /* this is my version number */
427 __u16 gnm_type; /* msg type */
428 __u64 gnm_srcnid; /* sender's NID */
429 __u64 gnm_connstamp; /* sender's connection stamp */
430 __u32 gnm_seq; /* incrementing sequence number */
431 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
432 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
433 __u32 gnm_payload_len; /* size of the FMA payload sent */
435 kgn_immediate_msg_t immediate;
436 kgn_putreq_msg_t putreq;
437 kgn_putack_msg_t putack;
439 kgn_completion_msg_t completion;
441 } WIRE_ATTR kgn_msg_t;
443 /************************************************************************
444 * runtime tunable data
447 typedef struct kgn_tunables {
448 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
449 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
450 int *kgn_credits; /* # concurrent sends */
451 int *kgn_fma_cq_size; /* # entries in receive CQ */
452 int *kgn_peer_credits; /* # LNet peer credits */
453 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
454 int *kgn_timeout; /* comms timeout (seconds) */
455 int *kgn_max_immediate; /* immediate payload breakpoint */
456 int *kgn_checksum; /* checksum data */
457 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
458 int *kgn_bte_dlvr_mode; /* BTE delivery mode mask */
459 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
460 int *kgn_ptag; /* PTAG for cdm_create */
461 int *kgn_pkey; /* PKEY for cdm_create */
462 int *kgn_max_retransmits; /* max number of FMA retransmits */
463 int *kgn_nwildcard; /* # wildcard per net to post */
464 int *kgn_nice; /* nice value for kgnilnd threads */
465 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
466 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
467 int *kgn_peer_hash_size; /* size of kgn_peers */
468 int *kgn_peer_health; /* enable/disable peer health */
469 int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
470 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
471 int *kgn_mbox_per_block; /* mailboxes per fmablk */
472 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
473 int *kgn_mbox_credits; /* max credits per fma */
474 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
475 int *kgn_net_hash_size; /* size of kgn_net_ht */
476 int *kgn_hardware_timeout; /* max time for a message to get across the network */
477 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
478 int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
479 int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
480 int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
481 int *kgn_reverse_rdma; /* Reverse RDMA setting */
482 int *kgn_eager_credits; /* allocated eager buffers */
483 int *kgn_fast_reconn; /* fast reconnection on conn timeout */
484 int *kgn_efault_lbug; /* LBUG on receiving an EFAULT */
485 int *kgn_max_purgatory; /* # conns/peer to keep in purgatory */
486 int *kgn_thread_affinity; /* bind scheduler threads to cpus */
487 int *kgn_thread_safe; /* use thread safe kgni API */
488 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
489 cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
493 typedef struct kgn_mbox_info {
494 lnet_nid_t mbx_prev_nid;
495 lnet_nid_t mbx_prev_purg_nid;
496 unsigned long mbx_create_conn_memset;
497 unsigned long mbx_add_purgatory;
498 unsigned long mbx_detach_of_purgatory;
499 unsigned long mbx_release_from_purgatory;
500 unsigned long mbx_release_purg_active_dgram;
502 int mbx_nallocs_total;
505 typedef struct kgn_fma_memblock {
506 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
507 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
508 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
509 int gnm_num_mboxs; /* total mboxes allocated */
510 int gnm_avail_mboxs; /* number of available mailboxes in the block */
511 int gnm_held_mboxs; /* number of purgatory held mailboxes */
512 int gnm_mbox_size; /* size of the single mailbox */
513 int gnm_next_avail_mbox; /* next available mailbox */
514 long gnm_max_timeout; /* max timeout for possible purgatory hold */
515 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
516 void *gnm_block; /* pointer to mem. block */
517 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
518 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
519 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
520 } kgn_fma_memblock_t;
522 typedef struct kgn_device {
523 gni_nic_handle_t gnd_handle; /* device handle */
524 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
525 gni_err_handle_t gnd_err_handle; /* device error handle */
526 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
527 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
528 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
529 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
530 struct mutex gnd_cq_mutex; /* CQ access serialization */
531 __u32 gnd_host_id; /* ph. host ID of the NIC */
532 int gnd_id; /* device id, also index in kgn_devices */
533 __u32 gnd_nid; /* ph host ID translated to NID */
534 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
535 struct mutex gnd_fmablk_mutex; /* mutex for FMA block memory alloc/free */
536 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
537 atomic_t gnd_nfmablk; /* # of fmablk live */
538 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
539 atomic_t gnd_neps; /* # EP allocated to conns */
540 short gnd_ready; /* stuff to do in scheduler thread */
541 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
542 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
543 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
544 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
545 struct list_head gnd_connd_peers; /* peers waiting for a connection */
546 spinlock_t gnd_connd_lock; /* serialise connd_peers */
547 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
548 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
549 int gnd_dgram_ready; /* dgrams need movin' */
550 struct list_head *gnd_dgrams; /* nid hash to dgrams */
551 atomic_t gnd_ndgrams; /* # dgrams extant */
552 atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post*/
553 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
554 struct list_head gnd_map_list; /* list of all mapped regions */
555 int gnd_map_version; /* version flag for map list */
556 struct timer_list gnd_map_timer; /* wakey-wakey */
557 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
558 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
559 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
560 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
561 __u32 gnd_map_nphys; /* # TX phys mappings */
562 __u32 gnd_map_physnop; /* # TX phys pages mapped */
563 __u32 gnd_map_nvirt; /* # TX virt mappings */
564 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
565 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
566 unsigned long gnd_next_map; /* next mapping attempt in jiffies */
567 int gnd_map_attempt; /* last map attempt # */
568 unsigned long gnd_last_map; /* map timeout base */
569 struct list_head gnd_rdmaq; /* RDMA to be sent */
570 spinlock_t gnd_rdmaq_lock; /* play nice with others */
571 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
572 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
573 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
574 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
575 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
576 atomic_t gnd_short_ntx; /* TX stats: short messages */
577 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
578 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
579 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
580 atomic_t gnd_short_nrx; /* RX stats: short messages */
581 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
582 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
583 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
584 atomic_t gnd_fast_try; /* # of times fast send tried */
585 atomic_t gnd_fast_ok; /* # of times fast send ok */
586 atomic_t gnd_fast_block; /* # of times fast send blocked */
587 unsigned long gnd_mutex_delay;
588 atomic_t gnd_n_yield;
589 atomic_t gnd_n_schedule;
590 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
591 struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
594 typedef struct kgn_net {
595 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
596 kgn_device_t *gnn_dev; /* device for this net */
597 lnet_ni_t *gnn_ni; /* network interface instance */
598 atomic_t gnn_refcount; /* # current references */
599 int gnn_shutdown; /* lnd_shutdown set */
600 __u16 gnn_netnum; /* stash netnum for quicker lookup */
603 static inline lnet_nid_t
604 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
606 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
609 static inline lnet_nid_t
610 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
612 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
615 /* The code for this is a bit ugly - but really this just boils down to a __u64
616 * that can have various parts accessed separately.
618 * The lower 32 bits is the ID
619 * we give to SMSG for our completion event - it needs to be globally unique across
620 * all TX currently in flight. We separate that out into the CQID so that we can
621 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
622 * the actual TX out of the per-connection gnc_tx_ref_table.
624 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
625 * has a unique value we can send with RDMA setup messages to ensure the completion for
626 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
627 * reuse is separated.
630 typedef struct kgn_tx_ev_id {
632 __u64 txe_cookie; /* are you my mommy ? */
634 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
636 __u32 txe_smsg_id; /* ID for SMSG CQ event */
637 /* N.B: Never ever ever ever use the bit shifts directly,
638 * you are just asking for a world of pain and are at the
639 * mercy of the compiler layouts */
641 __u32 txe_cqid :GNILND_CQID_NBITS;
642 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
649 typedef struct kgn_dgram {
650 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
651 kgn_dgram_state_t gndg_state; /* state of this dgram */
652 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
653 __u32 gndg_magic; /* saftey word */
654 unsigned long gndg_post_time; /* time when we posted */
655 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
656 kgn_connreq_t gndg_conn_out; /* connreq from local node */
657 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
660 typedef struct kgn_tx { /* message descriptor */
661 struct list_head tx_list; /* TX queues - peer, conn, rdma */
662 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
663 struct list_head *tx_list_p; /* pointer to current list */
664 struct kgn_conn *tx_conn; /* owning conn */
665 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
666 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
667 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
668 struct list_head tx_map_list; /* list entry on device map list */
669 unsigned int tx_nob; /* # bytes of payload */
670 int tx_buftype; /* payload buffer type */
671 int tx_phys_npages; /* # physical pages */
672 gni_mem_handle_t tx_map_key; /* mapping key */
673 gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
674 gni_mem_segment_t *tx_phys; /* page descriptors */
675 kgn_msg_t tx_msg; /* FMA message buffer */
676 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
677 __u8 tx_state; /* state of the descriptor */
678 int tx_retrans; /* retrans count of RDMA */
679 int tx_rc; /* if we need to stash the ret code until we see completion */
680 void *tx_buffer; /* source/sink buffer */
681 void *tx_buffer_copy; /* pointer to page aligned buffer */
682 unsigned int tx_nob_rdma; /* nob actually rdma */
683 unsigned int tx_offset; /* offset of data into copied buffer */
685 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
686 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
689 /* we only use one or the other */
691 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
692 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
696 typedef struct kgn_conn {
697 kgn_device_t *gnc_device; /* which device */
698 struct kgn_peer *gnc_peer; /* owning peer */
699 int gnc_magic; /* magic value cleared before free */
700 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
701 struct list_head gnc_hashlist; /* stash in connection hash table */
702 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
703 struct list_head gnc_fmaq; /* txs queued for FMA */
704 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
705 __u64 gnc_peerstamp; /* peer's unique stamp */
706 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
707 __u64 gnc_my_connstamp; /* my unique connection stamp */
708 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
709 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
710 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
711 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
712 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
713 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
714 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
715 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
716 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
717 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
718 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
719 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
720 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
721 __u32 gnc_cqid; /* my completion callback id (non-unique) */
722 atomic_t gnc_tx_seq; /* tx msg sequence number */
723 atomic_t gnc_rx_seq; /* rx msg sequence number */
724 struct mutex gnc_smsg_mutex; /* tx smsg sequence serialization */
725 struct mutex gnc_rdma_mutex; /* tx rdma sequence serialization */
726 __u64 gnc_tx_retrans; /* # retrans on SMSG */
727 atomic_t gnc_nlive_fma; /* # live FMA */
728 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
729 atomic_t gnc_nlive_rdma; /* # live RDMA */
730 short gnc_close_sent; /* I've sent CLOSE */
731 short gnc_close_recvd; /* I've received CLOSE */
732 short gnc_in_purgatory; /* in the sin bin */
733 int gnc_error; /* errno when conn being closed due to error */
734 int gnc_peer_error; /* errno peer sent us on CLOSE */
735 kgn_conn_state_t gnc_state; /* connection state */
736 int gnc_scheduled; /* being attented to */
737 char gnc_sched_caller[30]; /* what function last called schedule */
738 int gnc_sched_line; /* what line # last called schedule */
739 atomic_t gnc_refcount; /* # users */
740 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
741 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
742 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
743 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
744 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
745 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
746 int gnc_next_tx; /* next tx to use in tx_ref_table */
747 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
748 int gnc_mbox_id; /* id of mbox in fma_blk */
749 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
750 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
751 atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
752 kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
753 void *remote_mbox_addr; /* save remote mbox address */
756 typedef struct kgn_mdd_purgatory {
757 gni_mem_handle_t gmp_map_key; /* mapping key */
758 struct list_head gmp_list; /* entry point for purgatory list */
759 } kgn_mdd_purgatory_t;
761 typedef struct kgn_peer {
762 struct list_head gnp_list; /* stash on global peer list */
763 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
764 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
765 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
766 kgn_net_t *gnp_net; /* net instance for this peer */
767 lnet_nid_t gnp_nid; /* who's on the other end(s) */
768 atomic_t gnp_refcount; /* # users */
769 __u32 gnp_host_id; /* ph. host ID of the peer */
770 short gnp_connecting; /* connection forming */
771 short gnp_pending_unlink; /* need last conn close to trigger unlink */
772 int gnp_last_errno; /* last error conn saw */
773 unsigned long gnp_last_alive; /* last time I had valid comms */
774 int gnp_last_dgram_errno; /* last error dgrams saw */
775 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
776 unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
777 unsigned long gnp_reconnect_interval; /* exponential backoff */
778 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
779 int gnp_down; /* rca says peer down */
780 unsigned long gnp_down_event_time; /* time peer down */
781 unsigned long gnp_up_event_time; /* time peer back up */
784 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
785 * like lnet_parse. It allows a single pointer to let us get enough
786 * information in _recv and friends */
787 typedef struct kgn_rx {
788 kgn_conn_t *grx_conn; /* connection */
789 kgn_msg_t *grx_msg; /* message */
790 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
791 int grx_eager; /* if eager, we copied msg to somewhere */
792 struct timespec grx_received; /* time this msg received */
795 typedef struct kgn_data {
796 int kgn_init; /* initialisation state */
797 int kgn_shutdown; /* shut down? */
798 int kgn_wc_kill; /* Should I repost the WC */
799 atomic_t kgn_nthreads; /* # live threads */
800 int kgn_nresets; /* number of stack resets */
801 int kgn_in_reset; /* are we in stack reset ? */
803 __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
805 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
806 int kgn_ndevs; /* # devices */
808 int kgn_ruhroh_running; /* ruhroh thread is running */
809 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
810 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
811 int kgn_quiesce_trigger; /* should we quiesce ? */
812 atomic_t kgn_nquiesce; /* how many quiesced ? */
813 struct mutex kgn_quiesce_mutex; /* serialize ruhroh task, startup and shutdown */
814 int kgn_needs_reset; /* we need stack reset */
816 /* These next three members implement communication from gnilnd into
817 * the ruhroh task. To ensure correct operation of the task, code that
818 * writes into them must use memory barriers to ensure that the changes
819 * are visible to other cores in the order the members appear below. */
820 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
821 int kgn_bump_info_rdy; /* we have info needed to bump */
822 int kgn_needs_pause; /* we need to pause for network quiesce */
824 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
825 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
827 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
828 struct list_head *kgn_peers; /* hash table of all my known peers */
829 atomic_t kgn_npeers; /* # peers extant */
830 int kgn_peer_version; /* version flag for peer tables */
832 struct list_head *kgn_conns; /* conns hashed by cqid */
833 atomic_t kgn_nconns; /* # connections extant */
834 atomic_t kgn_neager_allocs; /* # of eager allocations */
835 __u64 kgn_peerstamp; /* when I started up */
836 __u64 kgn_connstamp; /* conn stamp generator */
837 int kgn_conn_version; /* version flag for conn tables */
838 int kgn_next_cqid; /* cqid generator */
840 long kgn_new_min_timeout; /* minimum timeout on any new conn */
841 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
842 spinlock_t kgn_reaper_lock; /* serialise */
844 struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
845 struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
846 struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
847 atomic_t kgn_ntx; /* # tx in use */
848 struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
850 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
851 __u64 kgn_cksum_npages; /* # pages alloc'd for checksumming */
852 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
853 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
855 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
856 long kgn_rdmaq_override; /* bytes per second override */
858 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
860 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
861 atomic_t kgn_npending_conns; /* # of conns with pending closes */
862 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
863 unsigned long kgn_last_scheduled; /* last time schedule was called */
864 unsigned long kgn_last_condresched; /* last time cond_resched was called */
865 atomic_t kgn_rev_offset; /* # of REV rdma w/misaligned offsets */
866 atomic_t kgn_rev_length; /* # of REV rdma have misaligned len */
867 atomic_t kgn_rev_copy_buff; /* # of REV rdma buffer copies */
868 struct socket *kgn_sock; /* for Apollo */
869 unsigned long free_pages_limit; /* # of free pages reserve from fma block allocations */
870 int kgn_enable_gl_mutex; /* kgni api mtx enable */
873 extern kgn_data_t kgnilnd_data;
874 extern kgn_tunables_t kgnilnd_tunables;
876 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
877 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
878 extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
880 /* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
881 * and the line of the calling function to allow us to debug problematic
882 * schedule calls in the future without the programmer having to mark
883 * the location manually.
885 #define kgnilnd_schedule_conn(conn) \
886 _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
888 #define kgnilnd_schedule_conn_refheld(conn, refheld) \
889 _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
892 kgnilnd_thread_fini(void)
894 atomic_dec(&kgnilnd_data.kgn_nthreads);
897 static inline void kgnilnd_gl_mutex_lock(struct mutex *lock)
899 if (kgnilnd_data.kgn_enable_gl_mutex)
903 static inline void kgnilnd_gl_mutex_unlock(struct mutex *lock)
905 if (kgnilnd_data.kgn_enable_gl_mutex)
909 static inline void kgnilnd_conn_mutex_lock(struct mutex *lock)
911 if (!kgnilnd_data.kgn_enable_gl_mutex)
915 static inline void kgnilnd_conn_mutex_unlock(struct mutex *lock)
917 if (!kgnilnd_data.kgn_enable_gl_mutex)
921 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
922 * parts of the code to avoid a scheduler trip when the mutex is held
924 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
925 * has been acquired successfully, and 0 on contention.
927 * NOTE: this function follows the spin_trylock() convention, so
928 * it is negated to the down_trylock() return values! Be careful
929 * about this when converting semaphore users to mutexes.
931 * This function must not be used in interrupt context. The
932 * mutex must be released by the same task that acquired it.
934 static inline int __kgnilnd_mutex_trylock(struct mutex *lock)
937 unsigned long timeout;
939 LASSERT(!in_interrupt());
941 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
943 ret = mutex_trylock(lock);
950 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
952 if (!kgnilnd_data.kgn_enable_gl_mutex)
955 return __kgnilnd_mutex_trylock(lock);
958 static inline int kgnilnd_trylock(struct mutex *cq_lock,
959 struct mutex *c_lock)
961 if (kgnilnd_data.kgn_enable_gl_mutex)
962 return __kgnilnd_mutex_trylock(cq_lock);
964 return __kgnilnd_mutex_trylock(c_lock);
967 static inline void *kgnilnd_vmalloc(int size)
969 void *ret = __vmalloc(size, __GFP_HIGHMEM | GFP_NOFS, PAGE_KERNEL);
970 LIBCFS_ALLOC_POST(ret, size);
974 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
977 _kgnilnd_debug_msg(kgn_msg_t *msg,
978 struct libcfs_debug_msg_data *data, const char *fmt, ... );
980 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
982 CFS_CHECK_STACK(msgdata, mask, cdls); \
984 if (((mask) & D_CANTMASK) != 0 || \
985 ((libcfs_debug & (mask)) != 0 && \
986 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
987 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
990 /* for most callers (level is a constant) this is resolved at compile time */
991 #define GNIDBG_MSG(level, msg, fmt, args...) \
993 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
994 static cfs_debug_limit_state_t cdls; \
995 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
996 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
997 "$$ "fmt" from %s ", ## args, \
998 libcfs_nid2str((msg)->gnm_srcnid)); \
1000 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1001 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1002 "$$ "fmt" from %s ", ## args, \
1003 libcfs_nid2str((msg)->gnm_srcnid)); \
1007 /* user puts 'to nid' in msg for us */
1008 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
1010 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1011 static cfs_debug_limit_state_t cdls; \
1012 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1013 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
1014 "$$ "fmt" ", ## args); \
1016 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1017 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
1018 "$$ "fmt" ", ## args); \
1023 _kgnilnd_debug_conn(kgn_conn_t *conn,
1024 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1026 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
1028 CFS_CHECK_STACK(msgdata, mask, cdls); \
1030 if (((mask) & D_CANTMASK) != 0 || \
1031 ((libcfs_debug & (mask)) != 0 && \
1032 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1033 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
1036 /* for most callers (level is a constant) this is resolved at compile time */
1037 #define GNIDBG_CONN(level, conn, fmt, args...) \
1039 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1040 static cfs_debug_limit_state_t cdls; \
1041 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1042 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
1043 "$$ "fmt" ", ## args); \
1045 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1046 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
1047 "$$ "fmt" ", ## args); \
1052 _kgnilnd_debug_tx(kgn_tx_t *tx,
1053 struct libcfs_debug_msg_data *data, const char *fmt, ... );
1055 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
1057 CFS_CHECK_STACK(msgdata, mask, cdls); \
1059 if (((mask) & D_CANTMASK) != 0 || \
1060 ((libcfs_debug & (mask)) != 0 && \
1061 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
1062 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
1065 /* for most callers (level is a constant) this is resolved at compile time */
1066 #define GNIDBG_TX(level, tx, fmt, args...) \
1068 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
1069 static cfs_debug_limit_state_t cdls; \
1070 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
1071 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
1072 "$$ "fmt" ", ## args); \
1074 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
1075 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
1076 "$$ "fmt" ", ## args); \
1080 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
1082 if (unlikely(!(cond))) { \
1083 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
1088 #define GNILND_IS_QUIESCED \
1089 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
1090 atomic_read(&kgnilnd_data.kgn_nthreads))
1092 #define KGNILND_SPIN_QUIESCE \
1094 /* E.T phone home */ \
1095 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
1096 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
1097 while (kgnilnd_data.kgn_quiesce_trigger) { \
1098 msleep_interruptible(MSEC_PER_SEC); \
1100 /* Mom, my homework is done */ \
1101 CDEBUG(D_NET, "Waking up from thread pause\n"); \
1102 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
1105 /* use macros for addref/decref to get the calling function name in the CDEBUG */
1106 #ifndef LIBCFS_DEBUG
1107 #error "this code uses actions inside LASSERT for ref counting"
1110 #define kgnilnd_admin_addref(atomic) \
1112 int val = atomic_inc_return(&atomic); \
1113 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
1114 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1117 #define kgnilnd_admin_decref(atomic) \
1119 int val = atomic_dec_return(&atomic); \
1120 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
1121 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
1124 #define kgnilnd_net_addref(net) \
1126 int val = atomic_inc_return(&net->gnn_refcount); \
1127 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
1128 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
1129 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1132 #define kgnilnd_net_decref(net) \
1134 int val = atomic_dec_return(&net->gnn_refcount); \
1135 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
1136 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
1137 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
1140 #define kgnilnd_peer_addref(peer) \
1142 int val = atomic_inc_return(&peer->gnp_refcount); \
1143 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
1144 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
1145 libcfs_nid2str(peer->gnp_nid), val); \
1148 #define kgnilnd_peer_decref(peer) \
1150 int val = atomic_dec_return(&peer->gnp_refcount); \
1151 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
1152 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
1153 libcfs_nid2str(peer->gnp_nid), val); \
1155 kgnilnd_destroy_peer(peer); \
1158 #define kgnilnd_conn_addref(conn) \
1163 val = atomic_inc_return(&conn->gnc_refcount); \
1164 LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
1165 "conn %p refc %d to %s\n", \
1168 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1170 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1172 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1177 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1178 * through the scheduler thread to get the EP destroyed. This avoids some
1179 * messy semaphore business and allows us to reuse the connd_list and existing
1180 * linkage and avoid creating extra lists just for destroying EPs */
1182 /* Safety Disclaimer:
1183 * Q: If we decrement the refcount and then check it again, is it possible that
1184 * another caller could have passed through this macro concurrently? If so,
1185 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1187 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1188 * once we are refcount <= 2. It hinges around gnc_state and membership of
1189 * gnc_hashlist. There are two ways to find a connection - either ask for
1190 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1191 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1194 * - #1 from create (kgnilnd_create_conn)
1195 * - #2 for EP (kgnilnd_create_conn)
1196 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1197 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1199 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1200 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1202 * There could be more from TX descriptors during the lifetime of a live
1205 * If we nuke the conn before finish_connect, we won't have parallel paths
1206 * because nobody besides the dgram handler for the single outstanding
1207 * dgram can find the connection as it isn't in any searchable tables yet.
1209 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1210 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1211 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1212 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1213 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1214 * sent. Now the gnc_state == CLOSED, so we hit
1215 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1216 * - we've nuked them from the peer and CQ id tables, so we own them and
1217 * are guaranteed serial access - hence the complete lack of conn list
1218 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1219 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1220 * kgnilnd_process_conns. Then the next scheduler loop would call
1221 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1222 * kgnilnd_process_conns.
1224 * Clearly, we are totally safe. Clearly.
1227 #define kgnilnd_conn_decref(conn) \
1232 val = atomic_dec_return(&conn->gnc_refcount); \
1233 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1236 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1238 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1240 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1245 (conn->gnc_ephandle != NULL) && \
1246 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1247 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1248 kgnilnd_schedule_conn(conn); \
1249 } else if (val == 0) { \
1250 kgnilnd_destroy_conn(conn); \
1254 static inline struct list_head *
1255 kgnilnd_nid2peerlist(lnet_nid_t nid)
1257 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1259 RETURN(&kgnilnd_data.kgn_peers[hash]);
1262 static inline struct list_head *
1263 kgnilnd_netnum2netlist(__u16 netnum)
1265 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1267 RETURN(&kgnilnd_data.kgn_nets[hash]);
1271 kgnilnd_peer_active(kgn_peer_t *peer)
1273 /* Am I in the peer hash table? */
1274 return (!list_empty(&peer->gnp_list));
1277 /* need write_lock on kgn_peer_conn_lock */
1279 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1281 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1282 peer, libcfs_nid2str(peer->gnp_nid),
1283 !list_empty(&peer->gnp_conns),
1284 !list_empty(&peer->gnp_tx_queue));
1286 /* kgn_peer_conn_lock protects us from conflict with
1287 * kgnilnd_peer_notify and gnp_persistent */
1288 RETURN ((list_empty(&peer->gnp_conns)) &&
1289 (list_empty(&peer->gnp_tx_queue)));
1292 /* returns positive if error was for a clean shutdown of conn */
1294 kgnilnd_conn_clean_errno(int errno)
1296 /* - ESHUTDOWN - LND is unloading
1297 * - EUCLEAN - admin requested via "lctl del_peer"
1298 * - ENETRESET - admin requested via "lctl disconnect" or rca event
1299 * - ENOTRECOVERABLE - stack reset
1300 * - EISCONN - cleared via "lctl push"
1301 * not doing ESTALE - that isn't clean */
1302 RETURN ((errno == 0) ||
1303 (errno == -ESHUTDOWN) ||
1304 (errno == -EUCLEAN) ||
1305 (errno == -ENETRESET) ||
1306 (errno == -EISCONN) ||
1307 (errno == -ENOTRECOVERABLE));
1310 /* returns positive if error results in purgatory hold */
1312 kgnilnd_check_purgatory_errno(int errno)
1314 /* We don't want to save the purgatory lists these cases:
1315 * - EUCLEAN - admin requested via "lctl del_peer"
1316 * - ESHUTDOWN - LND is unloading
1318 RETURN ((errno != -ESHUTDOWN) &&
1319 (errno != -EUCLEAN));
1323 /* returns positive if a purgatory hold is needed */
1325 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1329 if (conn->gnc_peer) {
1330 loopback = conn->gnc_peer->gnp_nid ==
1331 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1333 /* short circuit - a conn that didn't complete
1334 * setup never needs a purgatory hold */
1337 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1338 conn, conn->gnc_peer ?
1339 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1341 loopback, conn->gnc_close_recvd);
1343 /* we only use a purgatory hold if we've not received the CLOSE msg
1344 * from our peer - without that message, we can't know the state of
1345 * the other end of this connection and must put it into purgatory
1346 * to prevent reuse and corruption.
1347 * The theory is that a TX error can be communicated in all other cases
1349 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1350 kgnilnd_check_purgatory_errno(conn->gnc_error));
1353 static inline const char *
1354 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1356 static inline struct list_head *
1357 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1358 kgn_tx_list_state_t to_state)
1361 case GNILND_TX_PEERQ:
1362 return &peer->gnp_tx_queue;
1363 case GNILND_TX_FMAQ:
1364 return &conn->gnc_fmaq;
1365 case GNILND_TX_LIVE_FMAQ:
1366 case GNILND_TX_LIVE_RDMAQ:
1367 case GNILND_TX_DYING:
1369 case GNILND_TX_MAPQ:
1370 return &conn->gnc_device->gnd_map_tx;
1371 case GNILND_TX_RDMAQ:
1372 return &conn->gnc_device->gnd_rdmaq;
1374 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1375 CERROR("invalid state requested: %s\n",
1376 kgnilnd_tx_state2str(to_state));
1382 /* should hold tx, conn or peer lock when calling */
1384 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1385 kgn_conn_t *conn, kgn_tx_list_state_t state,
1388 struct list_head *list = NULL;
1390 /* make sure we have a sane TX state to start */
1391 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1392 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1393 list_empty(&tx->tx_list),
1394 "bad state with tx_list %s",
1395 list_empty(&tx->tx_list) ? "empty" : "not empty");
1397 /* WTF - you are already on that state buttmunch */
1398 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1399 "already at %s", kgnilnd_tx_state2str(state));
1401 /* get proper list from the state requested */
1402 list = kgnilnd_tx_state2list(peer, conn, state);
1406 case GNILND_TX_PEERQ:
1407 kgnilnd_peer_addref(peer);
1409 case GNILND_TX_ALLOCD:
1410 /* no refs needed */
1412 case GNILND_TX_FMAQ:
1413 kgnilnd_conn_addref(conn);
1415 case GNILND_TX_MAPQ:
1416 atomic_inc(&conn->gnc_device->gnd_nq_map);
1417 kgnilnd_conn_addref(conn);
1419 case GNILND_TX_LIVE_FMAQ:
1420 atomic_inc(&conn->gnc_nlive_fma);
1421 kgnilnd_conn_addref(conn);
1423 case GNILND_TX_LIVE_RDMAQ:
1424 atomic_inc(&conn->gnc_nlive_rdma);
1425 kgnilnd_conn_addref(conn);
1427 case GNILND_TX_RDMAQ:
1428 atomic_inc(&conn->gnc_nq_rdma);
1429 kgnilnd_conn_addref(conn);
1431 case GNILND_TX_DYING:
1432 kgnilnd_conn_addref(conn);
1435 CERROR("invalid state requested: %s\n",
1436 kgnilnd_tx_state2str(state));
1441 /* if this changes, change kgnilnd_alloc_tx */
1442 tx->tx_list_state = state;
1444 /* some states don't have lists - we track them in the per conn
1445 * TX table instead. Waste not, want not! */
1447 tx->tx_list_p = list;
1449 list_add_tail(&tx->tx_list, list);
1451 list_add(&tx->tx_list, list);
1453 /* set dummy list_p to make book keeping happy and let debugging
1454 * be a hair easier */
1455 tx->tx_list_p = (void *)state;
1458 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1459 kgnilnd_tx_state2str(state), list);
1463 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1464 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1466 /* These is only 1 "off-list" state */
1467 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1468 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1470 /* new_state == ALLOCD means we are deallocating this tx,
1471 * so make sure it was on a valid list to start with */
1472 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1473 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1474 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1475 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1478 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1481 switch (tx->tx_list_state) {
1482 case GNILND_TX_PEERQ:
1483 kgnilnd_peer_decref(peer);
1485 case GNILND_TX_FREED:
1486 case GNILND_TX_IDLE:
1487 case GNILND_TX_ALLOCD:
1488 /* no refs needed */
1490 case GNILND_TX_DYING:
1491 kgnilnd_conn_decref(conn);
1493 case GNILND_TX_FMAQ:
1494 kgnilnd_conn_decref(conn);
1496 case GNILND_TX_MAPQ:
1497 atomic_dec(&conn->gnc_device->gnd_nq_map);
1498 kgnilnd_conn_decref(conn);
1500 case GNILND_TX_LIVE_FMAQ:
1501 atomic_dec(&conn->gnc_nlive_fma);
1502 kgnilnd_conn_decref(conn);
1504 case GNILND_TX_LIVE_RDMAQ:
1505 atomic_dec(&conn->gnc_nlive_rdma);
1506 kgnilnd_conn_decref(conn);
1508 case GNILND_TX_RDMAQ:
1509 atomic_dec(&conn->gnc_nq_rdma);
1510 kgnilnd_conn_decref(conn);
1511 /* don't need to assert on default, already did in set */
1514 /* for ALLOCD, this might already be true, but no harm doing it again */
1515 list_del_init(&tx->tx_list);
1516 tx->tx_list_p = NULL;
1517 tx->tx_list_state = new_state;
1521 kgnilnd_tx_mapped(kgn_tx_t *tx)
1523 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1524 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1527 static inline struct list_head *
1528 kgnilnd_cqid2connlist(__u32 cqid)
1530 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1532 return (&kgnilnd_data.kgn_conns [hash]);
1535 static inline kgn_conn_t *
1536 kgnilnd_cqid2conn_locked(__u32 cqid)
1538 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1539 struct list_head *tmp;
1542 list_for_each(tmp, conns) {
1543 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1545 if (conn->gnc_cqid == cqid)
1552 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1554 kgnilnd_get_cqid_locked(void)
1560 cqid = kgnilnd_data.kgn_next_cqid++;
1561 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1565 kgnilnd_data.kgn_next_cqid = 1;
1568 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1574 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1576 kgn_tx_t *tx = NULL;
1577 kgn_conn_t *conn = NULL;
1579 /* set to NULL so any early return is an error */
1583 LASSERTF((ev_id->txe_idx > 0) &&
1584 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1585 "bogus txe_idx %d >= %d\n",
1586 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1588 LASSERTF((ev_id->txe_cqid > 0) &&
1589 (ev_id->txe_cqid < GNILND_MAX_CQID),
1590 "bogus txe_cqid %d >= %d\n",
1591 ev_id->txe_cqid, GNILND_MAX_CQID);
1593 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1594 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1597 /* Conn was destroyed? */
1598 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1599 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1602 /* just insurance */
1603 kgnilnd_conn_addref(conn);
1604 kgnilnd_admin_addref(conn->gnc_tx_in_use);
1605 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1607 /* we know this is safe - as the TX won't be reused until AFTER
1608 * the conn is unlinked from the cqid hash, so we can use the TX
1609 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1611 spin_lock(&conn->gnc_tx_lock);
1612 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1613 spin_unlock(&conn->gnc_tx_lock);
1615 /* We could have a tx that was cleared out by other forces
1616 * lctl disconnect or del_peer. */
1618 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1619 kgnilnd_admin_decref(conn->gnc_tx_in_use);
1620 kgnilnd_conn_decref(conn);
1624 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1625 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1626 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1628 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1629 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1630 "found tx %p txe_idx %d",
1631 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1632 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1634 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1636 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1641 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1644 /* set_normalized_timepsec isn't exported from the kernel, so
1645 * we need to do the same thing inline */
1646 static inline struct timespec
1647 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1653 sec = lhs.tv_sec - rhs.tv_sec;
1654 nsec = lhs.tv_nsec - rhs.tv_nsec;
1656 while (nsec >= NSEC_PER_SEC) {
1657 nsec -= NSEC_PER_SEC;
1661 nsec += NSEC_PER_SEC;
1670 kgnilnd_count_list(struct list_head *q)
1672 struct list_head *e;
1675 list_for_each(e, q) {
1682 /* kgnilnd_find_net adds a reference to the net it finds
1683 * this is so the net will not be removed before the calling function
1684 * has time to use the data returned. This reference needs to be released
1685 * by the calling function once it has finished using the returned net
1689 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1694 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1700 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1701 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1702 kgnilnd_net_addref(net);
1703 up_read(&kgnilnd_data.kgn_net_rw_sem);
1709 up_read(&kgnilnd_data.kgn_net_rw_sem);
1714 #ifdef CONFIG_DEBUG_SLAB
1715 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1717 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1720 int kgnilnd_dev_init(kgn_device_t *dev);
1721 void kgnilnd_dev_fini(kgn_device_t *dev);
1722 int kgnilnd_startup(lnet_ni_t *ni);
1723 void kgnilnd_shutdown(lnet_ni_t *ni);
1724 int kgnilnd_base_startup(void);
1725 void kgnilnd_base_shutdown(void);
1727 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1728 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1729 void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
1730 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1732 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1733 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1734 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1735 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1736 lnet_msg_t *lntmsg, void **new_private);
1737 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1738 int delayed, unsigned int niov,
1739 struct iovec *iov, lnet_kiov_t *kiov,
1740 unsigned int offset, unsigned int mlen, unsigned int rlen);
1742 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1744 /* purgatory functions */
1745 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1746 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1747 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1748 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1750 void kgnilnd_update_reaper_timeout(long timeout);
1751 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1752 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1753 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1754 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1755 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1756 int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
1757 int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1759 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1760 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
1761 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1762 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1764 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1765 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1766 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1767 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1768 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1769 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1770 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1771 void kgnilnd_consume_rx(kgn_rx_t *rx);
1773 void kgnilnd_schedule_device(kgn_device_t *dev);
1774 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1775 void kgnilnd_schedule_device_timer(unsigned long arg);
1777 int kgnilnd_reaper(void *arg);
1778 int kgnilnd_scheduler(void *arg);
1779 int kgnilnd_dgram_mover(void *arg);
1780 int kgnilnd_rca(void *arg);
1781 int kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id);
1783 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1784 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1785 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1786 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1787 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1788 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1789 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1790 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1791 void kgnilnd_peer_alive(kgn_peer_t *peer);
1792 void kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive);
1793 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1794 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1795 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1796 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1798 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1799 int kgnilnd_report_node_state(lnet_nid_t nid, int down);
1800 void kgnilnd_wakeup_rca_thread(void);
1801 int kgnilnd_start_rca_thread(void);
1802 int kgnilnd_get_node_state(__u32 nid);
1804 int kgnilnd_tunables_init(void);
1805 void kgnilnd_tunables_fini(void);
1806 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1808 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1809 void kgnilnd_pause_threads(void);
1810 int kgnilnd_hw_in_quiesce(void);
1811 int kgnilnd_check_hw_quiesce(void);
1812 void kgnilnd_quiesce_wait(char *reason);
1813 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1814 int kgnilnd_ruhroh_thread(void *arg);
1815 void kgnilnd_reset_stack(void);
1816 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1818 void kgnilnd_insert_sysctl(void);
1819 void kgnilnd_remove_sysctl(void);
1820 void kgnilnd_proc_init(void);
1821 void kgnilnd_proc_fini(void);
1824 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1826 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1827 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1828 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
1830 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1831 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1832 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1833 int kgnilnd_cancel_dgrams(kgn_device_t *dev);
1834 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1836 int kgnilnd_dgram_waitq(void *arg);
1838 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1840 /* struct2str functions - we don't use a default: case to cause the compile
1841 * to fail if there is a missing case. This allows us to hide these down here
1842 * out of the way but ensure we'll catch any updates to the enum/types
1845 #define DO_TYPE(x) case x: return #x;
1846 static inline const char *
1847 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1849 /* Only want single char string for this */
1851 case GNILND_FMABLK_IDLE:
1853 case GNILND_FMABLK_PHYS:
1855 case GNILND_FMABLK_VIRT:
1857 case GNILND_FMABLK_FREED:
1860 return "<unknown state>";
1863 static inline const char *
1864 kgnilnd_msgtype2str(int type)
1867 DO_TYPE(GNILND_MSG_NONE);
1868 DO_TYPE(GNILND_MSG_NOOP);
1869 DO_TYPE(GNILND_MSG_IMMEDIATE);
1870 DO_TYPE(GNILND_MSG_PUT_REQ);
1871 DO_TYPE(GNILND_MSG_PUT_NAK);
1872 DO_TYPE(GNILND_MSG_PUT_ACK);
1873 DO_TYPE(GNILND_MSG_PUT_DONE);
1874 DO_TYPE(GNILND_MSG_GET_REQ);
1875 DO_TYPE(GNILND_MSG_GET_NAK);
1876 DO_TYPE(GNILND_MSG_GET_DONE);
1877 DO_TYPE(GNILND_MSG_CLOSE);
1878 DO_TYPE(GNILND_MSG_PUT_REQ_REV);
1879 DO_TYPE(GNILND_MSG_PUT_DONE_REV);
1880 DO_TYPE(GNILND_MSG_PUT_NAK_REV);
1881 DO_TYPE(GNILND_MSG_GET_REQ_REV);
1882 DO_TYPE(GNILND_MSG_GET_ACK_REV);
1883 DO_TYPE(GNILND_MSG_GET_DONE_REV);
1884 DO_TYPE(GNILND_MSG_GET_NAK_REV);
1886 return "<unknown msg type>";
1889 static inline const char *
1890 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1893 DO_TYPE(GNILND_TX_IDLE);
1894 DO_TYPE(GNILND_TX_ALLOCD);
1895 DO_TYPE(GNILND_TX_PEERQ);
1896 DO_TYPE(GNILND_TX_MAPQ);
1897 DO_TYPE(GNILND_TX_FMAQ);
1898 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1899 DO_TYPE(GNILND_TX_RDMAQ);
1900 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1901 DO_TYPE(GNILND_TX_DYING);
1902 DO_TYPE(GNILND_TX_FREED);
1904 return "<unknown state>";
1907 static inline const char *
1908 kgnilnd_conn_state2str(kgn_conn_t *conn)
1910 kgn_conn_state_t state = conn->gnc_state;
1912 DO_TYPE(GNILND_CONN_DUMMY);
1913 DO_TYPE(GNILND_CONN_LISTEN);
1914 DO_TYPE(GNILND_CONN_CONNECTING);
1915 DO_TYPE(GNILND_CONN_ESTABLISHED);
1916 DO_TYPE(GNILND_CONN_CLOSING);
1917 DO_TYPE(GNILND_CONN_CLOSED);
1918 DO_TYPE(GNILND_CONN_DONE);
1919 DO_TYPE(GNILND_CONN_DESTROY_EP);
1924 static inline const char *
1925 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1927 kgn_connreq_type_t type = connreq->gncr_type;
1930 DO_TYPE(GNILND_CONNREQ_REQ);
1931 DO_TYPE(GNILND_CONNREQ_NAK);
1932 DO_TYPE(GNILND_CONNREQ_CLOSE);
1937 static inline const char *
1938 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1940 kgn_dgram_state_t state = dgram->gndg_state;
1943 DO_TYPE(GNILND_DGRAM_USED);
1944 DO_TYPE(GNILND_DGRAM_POSTING);
1945 DO_TYPE(GNILND_DGRAM_POSTED);
1946 DO_TYPE(GNILND_DGRAM_PROCESSING);
1947 DO_TYPE(GNILND_DGRAM_DONE);
1948 DO_TYPE(GNILND_DGRAM_CANCELED);
1953 static inline const char *
1954 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1956 kgn_dgram_type_t type = dgram->gndg_type;
1959 DO_TYPE(GNILND_DGRAM_REQ);
1960 DO_TYPE(GNILND_DGRAM_WC_REQ);
1961 DO_TYPE(GNILND_DGRAM_NAK);
1962 DO_TYPE(GNILND_DGRAM_CLOSE);
1967 static inline const char *
1968 kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
1971 DO_TYPE(GNILND_DGRAM_REQ);
1972 DO_TYPE(GNILND_DGRAM_WC_REQ);
1973 DO_TYPE(GNILND_DGRAM_NAK);
1974 DO_TYPE(GNILND_DGRAM_CLOSE);
1981 /* pulls in tunables per platform and adds in nid/nic conversion
1982 * if RCA wasn't available at build time */
1983 #include "gnilnd_hss_ops.h"
1985 #if defined(CONFIG_CRAY_GEMINI)
1986 #include "gnilnd_gemini.h"
1987 #elif defined(CONFIG_CRAY_ARIES)
1988 #include "gnilnd_aries.h"
1990 #error "Undefined Network Hardware Type"
1993 /* API wrapper functions - include late to pick up all of the other defines */
1994 #include "gnilnd_api_wrap.h"
1996 #endif /* _GNILND_GNILND_H_ */