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 #define DEBUG_SUBSYSTEM S_LND
31 #include <libcfs/libcfs.h>
32 #include <lnet/lnet.h>
33 #include <lnet/lib-lnet.h>
34 #include <lnet/lnet-sysctl.h>
37 #include "gnilnd_version.h"
38 #include "gnilnd_hss_ops.h"
40 /* tunables determined at compile time */
41 #define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
42 #define GNILND_BASE_TIMEOUT 60 /* default sane timeout */
43 #define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
44 #define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
45 #define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
46 #define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
47 #define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
49 /* reaper thread wakup interval */
50 #define GNILND_REAPER_THREAD_WAKE 1
51 /* reaper thread checks each conn NCHECKS time every kgnilnd_data.kgn_new_min_timeout */
52 #define GNILND_REAPER_NCHECKS 4
55 #define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
56 #define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
57 #define GNILND_COOKIE 0xa3579 /* cookie used by along with ptag by GNI */
60 #define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
61 #define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
62 #define GNILND_CHECKSUM_SMSG 2 /* checksum entire SMSG packet */
63 #define GNILND_CHECKSUM_SMSG_BTE 3 /* Full checksum support */
65 /* tune down some COMPUTE options as they won't see the same number of connections and
66 * don't need the throughput of multiple threads by default */
67 #if defined(CONFIG_CRAY_COMPUTE)
68 #define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
69 #define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
71 #define GNILND_SCHED_THREADS 3 /* default # of kgnilnd_scheduler threads */
72 #define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
75 /* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
76 #define GNILND_EXTRA_BITS 1
77 /* maximum number of conns & bits for cqid in the SMSG event data */
78 #define GNILND_CQID_NBITS (21 - GNILND_EXTRA_BITS)
79 #define GNILND_MSGID_TX_NBITS (32 - GNILND_CQID_NBITS)
80 #define GNILND_MAX_CQID (1 << GNILND_CQID_NBITS)
81 #define GNILND_MAX_MSG_ID (1 << GNILND_MSGID_TX_NBITS)
82 #define GNILND_MAX_MSG_SIZE (*kgnilnd_tunables.kgn_max_immediate + sizeof(kgn_msg_t))
84 /* need sane upper bound to limit copy overhead */
85 #define GNILND_MAX_IMMEDIATE (64<<10)
87 /* payload size to add to the base mailbox size
88 * This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
89 * gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
90 * the calculation return from that function.*/
91 #define GNILND_MBOX_PAYLOAD \
92 (GNILND_MAX_MSG_SIZE * \
93 ((*kgnilnd_tunables.kgn_concurrent_sends - 2) * 2));
95 /* timeout -> deadman timer for kgni mdd holds */
96 #define GNILND_TIMEOUT2DEADMAN ((*kgnilnd_tunables.kgn_mdd_timeout) * 1000 * 60)
98 /* timeout for failing sends in t is in jiffies*/
99 #define GNILND_TIMEOUTRX(t) (t + cfs_time_seconds(*kgnilnd_tunables.kgn_hardware_timeout))
101 /* time when to release from purgatory in the reaper thread in jiffies */
102 #define GNILND_PURG_RELEASE(t) (GNILND_TIMEOUTRX(t) * 3)
104 /* Macro for finding last_rx 2 datapoints are compared
105 * and the most recent one in jiffies is returned.
107 #define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
108 ? conn->gnc_last_rx : conn->gnc_last_rx_cq)
110 /************************************************************************
111 * Enum, flag and tag data
113 #define GNILND_INIT_NOTHING 0
114 #define GNILND_INIT_DATA 1
115 #define GNILND_INIT_ALL 2
117 /* If you change the ordering away from MAPPED = UNMAPPED + 1, things break */
118 #define GNILND_BUF_NONE 0 /* buffer type not set */
119 #define GNILND_BUF_IMMEDIATE 1 /* immediate data */
120 #define GNILND_BUF_IMMEDIATE_KIOV 2 /* immediate data */
121 #define GNILND_BUF_PHYS_UNMAPPED 3 /* physical: not mapped yet */
122 #define GNILND_BUF_PHYS_MAPPED 4 /* physical: mapped already */
123 #define GNILND_BUF_VIRT_UNMAPPED 5 /* virtual: not mapped yet */
124 #define GNILND_BUF_VIRT_MAPPED 6 /* virtual: mapped already */
126 #define GNILND_TX_WAITING_REPLY (1<<1) /* expecting to receive reply */
127 #define GNILND_TX_WAITING_COMPLETION (1<<2) /* waiting for smsg_send to complete */
128 #define GNILND_TX_PENDING_RDMA (1<<3) /* RDMA transaction pending until we get prev. completion */
129 #define GNILND_TX_QUIET_ERROR (1<<4) /* don't print error on tx_done */
130 #define GNILND_TX_FAIL_SMSG (1<<5) /* pass down error injection for SMSG fail */
132 /* stash above max CQID to avoid any collision */
133 #define GNILND_MSGID_NOOP (GNILND_MAX_CQID + 128)
134 #define GNILND_MSGID_CLOSE (GNILND_MSGID_NOOP + 1)
136 /* kgn_msg_t::gnm_type */
137 #define GNILND_MSG_NONE 0x00 /* illegal message */
138 #define GNILND_MSG_NOOP 0x01 /* empty gnm_u (keepalive) */
139 #define GNILND_MSG_IMMEDIATE 0x02 /* gnm_u.immediate */
140 #define GNILND_MSG_PUT_REQ 0x03 /* gnm_u.putreq (src->sink) */
141 #define GNILND_MSG_PUT_NAK 0x04 /* gnm_u.completion (no PUT match: sink->src) */
142 #define GNILND_MSG_PUT_ACK 0x05 /* gnm_u.putack (PUT matched: sink->src) */
143 #define GNILND_MSG_PUT_DONE 0x06 /* gnm_u.completion (src->sink) */
144 #define GNILND_MSG_GET_REQ 0x07 /* gnm_u.get (sink->src) */
145 #define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
146 #define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
147 #define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
149 /* defines for gnc_*scheduled states */
150 #define GNILND_CONN_IDLE 0
151 #define GNILND_CONN_SCHED 1
152 #define GNILND_CONN_WANTS_SCHED 2
153 #define GNILND_CONN_PROCESS 3
155 #define GNILND_DEV_IDLE 0
156 #define GNILND_DEV_IRQ 1
157 #define GNILND_DEV_LOOP 2
159 #define GNILND_DGRAM_IDLE 0
160 #define GNILND_DGRAM_SCHED 1
161 #define GNILND_DGRAM_PROCESS 2
163 #define GNILND_PEER_IDLE 0
164 #define GNILND_PEER_CONNECT 1
165 #define GNILND_PEER_POSTING 2
166 #define GNILND_PEER_POSTED 3
167 #define GNILND_PEER_NEEDS_DEATH 4
168 #define GNILND_PEER_KILL 5
170 /* for gnc_close_recvd */
171 #define GNILND_CLOSE_RX 1
172 #define GNILND_CLOSE_INJECT1 2
173 #define GNILND_CLOSE_INJECT2 3
174 #define GNILND_CLOSE_EARLY 4
176 /* defines for why quiesce trigger set */
177 #define GNILND_QUIESCE_IDLE 0
178 #define GNILND_QUIESCE_ADMIN 1
179 #define GNILND_QUIESCE_RESET 2
180 #define GNILND_QUIESCE_HW_QUIESCE 3
182 #define GNILND_PEER_CLEAN 0
183 #define GNILND_PEER_PERSISTING 1
185 #define GNILND_DEL_CONN 0
186 #define GNILND_DEL_PEER 1
187 #define GNILND_CLEAR_PURGATORY 2
189 typedef enum kgn_fmablk_state {
190 GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
191 GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
192 GNILND_FMABLK_VIRT, /* 'standard' vmalloc hunk */
193 GNILND_FMABLK_FREED, /* after free */
194 } kgn_fmablk_state_t;
196 typedef enum kgn_tx_list_state {
197 GNILND_TX_IDLE = 0, /* TX is on the idle list, kgn_idle_txs */
198 GNILND_TX_ALLOCD, /* TX has been alloced (off of idle), could be in any state transition */
199 GNILND_TX_PEERQ, /* TX on peer->gnp_tx_queue (no live conn) */
200 GNILND_TX_MAPQ, /* TX on dev:gnd_map_tx for buffer mapping */
201 GNILND_TX_FMAQ, /* TX waiting to be send on conn FMA */
202 GNILND_TX_LIVE_FMAQ, /* TX live on the FMA wire, waiting for completion or reply */
203 GNILND_TX_RDMAQ, /* TX waiting to send FMA confirmation to auth RDMA PUT */
204 GNILND_TX_LIVE_RDMAQ, /* TX live on the RDMA wire, waiting for completion */
205 GNILND_TX_DYING, /* TX got caught on MAPQ or RDMAQ while conn was closing, needs someone to call tx_done */
206 GNILND_TX_FREED /* TX is free! */
207 } kgn_tx_list_state_t;
209 typedef enum kgn_conn_state {
210 /* don't start @ 0 - prevent memset(0) badness */
211 GNILND_CONN_DUMMY = 0,
213 GNILND_CONN_CONNECTING,
214 GNILND_CONN_ESTABLISHED,
218 GNILND_CONN_DESTROY_EP
221 /* changing these requires a change to GNILND_CONNREQ_VERSION and
222 * will result in dropped packets instead of NAKs. Adding to this is
223 * acceptable without changing the CONNREQ_VERSION, but code should
224 * be ready to handle NAKs on version mismatch */
225 typedef enum kgn_connreq_type {
226 GNILND_CONNREQ_REQ = 1, /* how YOU doin' ? */
227 GNILND_CONNREQ_NAK, /* NO soup for you! */
228 GNILND_CONNREQ_CLOSE, /* we should see other people */
229 } kgn_connreq_type_t;
231 typedef enum kgn_dgram_state {
232 /* don't use 0 to avoid thinking a memset of zero is valid data */
233 GNILND_DGRAM_USED = 1,
234 GNILND_DGRAM_POSTING,
236 GNILND_DGRAM_PROCESSING,
237 GNILND_DGRAM_CANCELED,
241 typedef enum kgn_dgram_type {
242 GNILND_DGRAM_REQ = 1, /* how YOU doin' ? */
243 GNILND_DGRAM_WC_REQ, /* you talkin' to ME? */
244 GNILND_DGRAM_NAK, /* NO soup for you! */
245 GNILND_DGRAM_CLOSE, /* we should see other people */
248 /************************************************************************
249 * Wire message structs. These are sent in sender's byte order
250 * (i.e. receiver checks magic and flips if required).
253 #define GNILND_MSG_MAGIC LNET_PROTO_GNI_MAGIC /* unique magic */
254 #define GNILND_DGRAM_MAGIC 0x0DDBA11
256 /* kgn_msg_t - FMA/SMSG wire struct
258 * - added checksum to FMA
259 * moved seq before paylod
260 * WIRE_ATTR added for alignment
262 * added gnm_payload_len for FMA payload size
264 * added gncm_retval to completion, allowing return code transmission
267 * changed how CQID and TX ids are assigned
269 * added retval on CLOSE
271 * added payload checksumming
273 * reworked checksumming a bit, changed payload checksums
275 #define GNILND_MSG_VERSION 8
276 /* kgn_connreq_t connection request datagram wire struct
281 #define GNILND_CONNREQ_VERSION 2
283 typedef struct kgn_gniparams {
284 __u32 gnpr_host_id; /* ph. host ID of the NIC */
285 __u32 gnpr_cqid; /* cqid I want peer to use when sending events to me */
286 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
287 } WIRE_ATTR kgn_gniparams_t;
289 typedef struct kgn_nak_data {
290 __s32 gnnd_errno; /* errno reason for NAK */
292 } WIRE_ATTR kgn_nak_data_t;
294 /* the first bits of the connreq struct CANNOT CHANGE FORM EVER
295 * without breaking the ability for us to properly NAK someone */
296 typedef struct kgn_connreq { /* connection request/response */
297 __u32 gncr_magic; /* I'm an gnilnd connreq */
298 __u32 gncr_cksum; /* checksum (0 == disabled) */
299 __u16 gncr_type; /* REQ, NAK, etc */
300 __u16 gncr_version; /* this is my version number */
301 __u32 gncr_timeout; /* sender's timeout */
302 __u64 gncr_srcnid; /* sender's NID */
303 __u64 gncr_dstnid; /* who sender expects to listen */
304 __u64 gncr_peerstamp; /* sender's instance stamp */
305 __u64 gncr_connstamp; /* sender's connection stamp */
307 /* everything before this needs to stay static, adding after should
308 * result in a change to GNILND_CONNREQ_VERSION */
311 kgn_gniparams_t gncr_gnparams; /* sender's endpoint info */
312 kgn_nak_data_t gncr_nakdata; /* data (rc, etc) for NAK */
314 } WIRE_ATTR kgn_connreq_t;
317 gni_mem_handle_t gnrd_key;
320 } WIRE_ATTR kgn_rdma_desc_t;
323 lnet_hdr_t gnim_hdr; /* LNet header */
324 /* LNet payload is in FMA "Message Data" */
325 } WIRE_ATTR kgn_immediate_msg_t;
328 lnet_hdr_t gnprm_hdr; /* LNet header */
329 __u64 gnprm_cookie; /* opaque completion cookie */
330 } WIRE_ATTR kgn_putreq_msg_t;
333 __u64 gnpam_src_cookie; /* reflected completion cookie */
334 __u64 gnpam_dst_cookie; /* opaque completion cookie */
335 kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
336 } WIRE_ATTR kgn_putack_msg_t;
339 lnet_hdr_t gngm_hdr; /* LNet header */
340 __u64 gngm_cookie; /* opaque completion cookie */
341 kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
342 } WIRE_ATTR kgn_get_msg_t;
345 int gncm_retval; /* error on NAK, size on REQ */
346 __u64 gncm_cookie; /* reflected completion cookie */
347 } WIRE_ATTR kgn_completion_msg_t;
349 typedef struct { /* NB must fit in FMA "Prefix" */
350 __u32 gnm_magic; /* I'm an gni message */
351 __u16 gnm_version; /* this is my version number */
352 __u16 gnm_type; /* msg type */
353 __u64 gnm_srcnid; /* sender's NID */
354 __u64 gnm_connstamp; /* sender's connection stamp */
355 __u32 gnm_seq; /* incrementing sequence number */
356 __u16 gnm_cksum; /* checksum (0 == no checksum ) */
357 __u16 gnm_payload_cksum; /* payload checksum (0 == no checksum ) */
358 __u32 gnm_payload_len; /* size of the FMA payload sent */
360 kgn_immediate_msg_t immediate;
361 kgn_putreq_msg_t putreq;
362 kgn_putack_msg_t putack;
364 kgn_completion_msg_t completion;
366 } WIRE_ATTR kgn_msg_t;
368 /************************************************************************
369 * runtime tunable data
372 typedef struct kgn_tunables {
373 int *kgn_min_reconnect_interval; /* connreq starting timeout & retransmit interval */
374 int *kgn_max_reconnect_interval; /* ...exponentially increasing to this */
375 int *kgn_credits; /* # concurrent sends */
376 int *kgn_fma_cq_size; /* # entries in receive CQ */
377 int *kgn_peer_credits; /* # LNet peer credits */
378 int *kgn_concurrent_sends; /* max # of max_immediate in mbox */
379 int *kgn_timeout; /* comms timeout (seconds) */
380 int *kgn_max_immediate; /* immediate payload breakpoint */
381 int *kgn_checksum; /* checksum data */
382 int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
383 int *kgn_bte_hash; /* hashing on BTE transfers */
384 int *kgn_bte_adapt; /* adaptive routing on BTE transfers */
385 int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
386 int *kgn_ptag; /* PTAG for cdm_create */
387 int *kgn_max_retransmits; /* max number of FMA retransmits */
388 int *kgn_nwildcard; /* # wildcard per net to post */
389 int *kgn_nice; /* nice value for kgnilnd threads */
390 int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
391 int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
392 int *kgn_peer_hash_size; /* size of kgn_peers */
393 int *kgn_peer_health; /* enable/disable peer health */
394 int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
395 int *kgn_mbox_per_block; /* mailboxes per fmablk */
396 int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
397 int *kgn_mbox_credits; /* max credits per fma */
398 int *kgn_sched_threads; /* number of kgnilnd_scheduler threads */
399 int *kgn_net_hash_size; /* size of kgn_net_ht */
400 int *kgn_hardware_timeout; /* max time for a message to get across the network */
401 int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
402 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
403 cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
407 typedef struct kgn_mbox_info {
408 lnet_nid_t mbx_prev_nid;
409 unsigned long mbx_create_conn_memset;
410 unsigned long mbx_add_purgatory;
411 unsigned long mbx_detach_of_purgatory;
412 unsigned long mbx_release_from_purgatory;
413 unsigned long mbx_release_purg_active_dgram;
416 typedef struct kgn_fma_memblock {
417 struct list_head gnm_bufflist; /* memblock is part of device's gnd_fma_buffs */
418 kgn_fmablk_state_t gnm_state; /* how this memory allocated & state of it */
419 int gnm_hold_timeout; /* hold_timeout if used at unmap time */
420 int gnm_num_mboxs; /* total mboxes allocated */
421 int gnm_avail_mboxs; /* number of available mailboxes in the block */
422 int gnm_held_mboxs; /* number of purgatory held mailboxes */
423 int gnm_mbox_size; /* size of the single mailbox */
424 int gnm_next_avail_mbox; /* next available mailbox */
425 long gnm_max_timeout; /* max timeout for possible purgatory hold */
426 unsigned int gnm_blk_size; /* how big is our hunk o memory ? */
427 void *gnm_block; /* pointer to mem. block */
428 gni_mem_handle_t gnm_hndl; /* mem. handle of the block */
429 unsigned long *gnm_bit_array; /* bit array tracking allocation of mailboxes */
430 kgn_mbox_info_t *gnm_mbox_info; /* array of mbox_information about each mbox */
431 } kgn_fma_memblock_t;
433 typedef struct kgn_device {
434 gni_nic_handle_t gnd_handle; /* device handle */
435 gni_cdm_handle_t gnd_domain; /* GNI communication domain */
436 gni_err_handle_t gnd_err_handle; /* device error handle */
437 unsigned long gnd_sched_alive; /* scheduler thread alive stamp */
438 gni_cq_handle_t gnd_rcv_fma_cqh; /* FMA rcv. completion queue handle */
439 gni_cq_handle_t gnd_snd_rdma_cqh; /* rdma send completion queue handle */
440 gni_cq_handle_t gnd_snd_fma_cqh; /* rdma send completion queue handle */
441 struct mutex gnd_cq_mutex; /* CQ access serialization */
442 __u32 gnd_host_id; /* ph. host ID of the NIC */
443 int gnd_id; /* device id, also index in kgn_devices */
444 __u32 gnd_nid; /* ph host ID translated to NID */
445 struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
446 struct semaphore gnd_fmablk_sem; /* semaphore for FMA block memory alloc/free */
447 spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
448 atomic_t gnd_nfmablk; /* # of fmablk live */
449 atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
450 atomic_t gnd_neps; /* # EP allocated to conns */
451 short gnd_ready; /* stuff to do in scheduler thread */
452 struct list_head gnd_ready_conns; /* connections ready to tx/rx */
453 struct list_head gnd_map_tx; /* TX: needing buffer mapping */
454 wait_queue_head_t gnd_waitq; /* scheduler wakeup */
455 spinlock_t gnd_lock; /* serialise gnd_ready_conns */
456 struct list_head gnd_connd_peers; /* peers waiting for a connection */
457 spinlock_t gnd_connd_lock; /* serialise connd_peers */
458 wait_queue_head_t gnd_dgram_waitq; /* dgram_mover thread wakeup */
459 wait_queue_head_t gnd_dgping_waitq; /* dgram thread ping-pong */
460 int gnd_dgram_ready; /* dgrams need movin' */
461 struct list_head *gnd_dgrams; /* nid hash to dgrams */
462 atomic_t gnd_ndgrams; /* # dgrams extant */
463 spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
464 struct list_head gnd_map_list; /* list of all mapped regions */
465 int gnd_map_version; /* version flag for map list */
466 atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
467 atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
468 atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
469 atomic64_t gnd_nbytes_map; /* bytes of total GART maps - fma, tx, etc */
470 __u32 gnd_map_nphys; /* # TX phys mappings */
471 __u32 gnd_map_physnop; /* # TX phys pages mapped */
472 __u32 gnd_map_nvirt; /* # TX virt mappings */
473 __u64 gnd_map_virtnob; /* # TX virt bytes mapped */
474 spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
475 struct list_head gnd_rdmaq; /* RDMA to be sent */
476 spinlock_t gnd_rdmaq_lock; /* play nice with others */
477 atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
478 atomic64_t gnd_rdmaq_bytes_ok; /* # bytes allowed until deadline */
479 atomic_t gnd_rdmaq_nstalls; /* # stalls due to throttle */
480 unsigned long gnd_rdmaq_deadline; /* when does bucket roll over ? */
481 struct timer_list gnd_rdmaq_timer; /* wakey-wakey */
482 atomic_t gnd_short_ntx; /* TX stats: short messages */
483 atomic64_t gnd_short_txbytes; /* TX stats: short message payload*/
484 atomic_t gnd_rdma_ntx; /* TX stats: rdma messages */
485 atomic64_t gnd_rdma_txbytes; /* TX stats: rdma message payload*/
486 atomic_t gnd_short_nrx; /* RX stats: short messages */
487 atomic64_t gnd_short_rxbytes; /* RX stats: short message payload*/
488 atomic_t gnd_rdma_nrx; /* RX stats: rdma messages */
489 atomic64_t gnd_rdma_rxbytes; /* RX stats: rdma message payload*/
490 atomic_t gnd_fast_try; /* # of times fast send tried */
491 atomic_t gnd_fast_ok; /* # of times fast send ok */
492 atomic_t gnd_fast_block; /* # of times fast send blocked */
493 unsigned long gnd_mutex_delay;
494 atomic_t gnd_n_yield;
495 atomic_t gnd_n_schedule;
496 atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
499 typedef struct kgn_net {
500 struct list_head gnn_list; /* chain on kgni_data::kgn_nets */
501 kgn_device_t *gnn_dev; /* device for this net */
502 lnet_ni_t *gnn_ni; /* network interface instance */
503 atomic_t gnn_refcount; /* # current references */
504 int gnn_shutdown; /* lnd_shutdown set */
505 __u16 gnn_netnum; /* stash netnum for quicker lookup */
508 static inline lnet_nid_t
509 kgnilnd_lnd2lnetnid(lnet_nid_t ni_nid, lnet_nid_t kgnilnd_nid)
511 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(kgnilnd_nid));
514 static inline lnet_nid_t
515 kgnilnd_lnet2lndnid(lnet_nid_t lnet_nid, lnet_nid_t kgnilnd_nid)
517 return LNET_MKNID(LNET_NIDNET(kgnilnd_nid), LNET_NIDADDR(lnet_nid));
520 /* The code for this is a bit ugly - but really this just boils down to a __u64
521 * that can have various parts accessed separately.
523 * The lower 32 bits is the ID
524 * we give to SMSG for our completion event - it needs to be globally unique across
525 * all TX currently in flight. We separate that out into the CQID so that we can
526 * reference the connection (kgnilnd_cqid2conn_locked) and then the msg_id to pull
527 * the actual TX out of the per-connection gnc_tx_ref_table.
529 * The upper 32 bits are just extra stuff we put into the cookie to ensure this TX
530 * has a unique value we can send with RDMA setup messages to ensure the completion for
531 * those is unique across the wire. The extra 32 bits are there to ensure that TX id
532 * reuse is separated.
535 typedef struct kgn_tx_ev_id {
537 __u64 txe_cookie; /* are you my mommy ? */
539 __u32 txe_chips; /* extra bits to ensure ID unique across reuse */
541 __u32 txe_smsg_id; /* ID for SMSG CQ event */
542 /* N.B: Never ever ever ever use the bit shifts directly,
543 * you are just asking for a world of pain and are at the
544 * mercy of the compiler layouts */
546 __u32 txe_cqid :GNILND_CQID_NBITS;
547 __u32 txe_idx :GNILND_MSGID_TX_NBITS;
554 typedef struct kgn_dgram {
555 struct list_head gndg_list; /* on hash dev::gnd_dgrams */
556 kgn_dgram_state_t gndg_state; /* state of this dgram */
557 kgn_dgram_type_t gndg_type; /* REQ, NAK, etc */
558 __u32 gndg_magic; /* saftey word */
559 unsigned long gndg_post_time; /* time when we posted */
560 struct kgn_conn *gndg_conn; /* unbound conn with ep & smsg */
561 kgn_connreq_t gndg_conn_out; /* connreq from local node */
562 kgn_connreq_t gndg_conn_in; /* connreq from remote node */
565 typedef struct kgn_tx { /* message descriptor */
566 struct list_head tx_list; /* TX queues - peer, conn, rdma */
567 kgn_tx_list_state_t tx_list_state;/* where in state machine is this TX ? */
568 struct list_head *tx_list_p; /* pointer to current list */
569 struct kgn_conn *tx_conn; /* owning conn */
570 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
571 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
572 unsigned long tx_cred_wait; /* time spend waiting for smsg creds */
573 struct list_head tx_map_list; /* list entry on device map list */
574 unsigned int tx_nob; /* # bytes of payload */
575 int tx_buftype; /* payload buffer type */
576 int tx_phys_npages; /* # physical pages */
577 gni_mem_handle_t tx_map_key; /* mapping key */
578 gni_mem_segment_t *tx_phys; /* page descriptors */
579 kgn_msg_t tx_msg; /* FMA message buffer */
580 kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
581 __u8 tx_state; /* state of the descriptor */
582 int tx_retrans; /* retrans count of RDMA */
583 int tx_rc; /* if we need to stash the ret code until we see completion */
584 void *tx_buffer; /* source/sink buffer */
586 gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
587 struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
590 /* we only use one or the other */
592 kgn_putack_msg_t tx_putinfo; /* data for differed rdma & re-try */
593 kgn_get_msg_t tx_getinfo; /* data for rdma re-try*/
597 typedef struct kgn_conn {
598 kgn_device_t *gnc_device; /* which device */
599 struct kgn_peer *gnc_peer; /* owning peer */
600 struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
601 struct list_head gnc_hashlist; /* stash in connection hash table */
602 struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
603 struct list_head gnc_fmaq; /* txs queued for FMA */
604 struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
605 __u64 gnc_peerstamp; /* peer's unique stamp */
606 __u64 gnc_peer_connstamp; /* peer's unique connection stamp */
607 __u64 gnc_my_connstamp; /* my unique connection stamp */
608 unsigned long gnc_first_rx; /* when I first received an FMA message (jiffies) */
609 unsigned long gnc_last_tx; /* when I last sent an FMA message (jiffies) */
610 unsigned long gnc_last_rx; /* when I last sent an FMA message (jiffies) */
611 unsigned long gnc_last_tx_cq; /* when I last received an FMA CQ (jiffies) */
612 unsigned long gnc_last_rx_cq; /* when I last received an FMA CQ (jiffies) */
613 unsigned long gnc_last_noop_want; /* time I wanted to send NOOP */
614 unsigned long gnc_last_noop_sent; /* time I did gni_smsg_send on NOOP */
615 unsigned long gnc_last_noop_cq; /* time when NOOP completed */
616 unsigned long gnc_last_sched_ask; /* time when conn added to ready_conns */
617 unsigned long gnc_last_sched_do; /* time when conn processed from ready_conns */
618 atomic_t gnc_reaper_noop; /* # reaper triggered NOOP */
619 atomic_t gnc_sched_noop; /* # sched triggered NOOP */
620 unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
621 __u32 gnc_cqid; /* my completion callback id (non-unique) */
622 __u32 gnc_tx_seq; /* tx msg sequence number */
623 __u32 gnc_rx_seq; /* rx msg sequence number */
624 __u64 gnc_tx_retrans; /* # retrans on SMSG */
625 atomic_t gnc_nlive_fma; /* # live FMA */
626 atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
627 atomic_t gnc_nlive_rdma; /* # live RDMA */
628 short gnc_close_sent; /* I've sent CLOSE */
629 short gnc_close_recvd; /* I've received CLOSE */
630 short gnc_in_purgatory; /* in the sin bin */
631 int gnc_error; /* errno when conn being closed due to error */
632 int gnc_peer_error; /* errno peer sent us on CLOSE */
633 kgn_conn_state_t gnc_state; /* connection state */
634 int gnc_scheduled; /* being attented to */
635 atomic_t gnc_refcount; /* # users */
636 spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
637 gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
638 kgn_fma_memblock_t *gnc_fma_blk; /* pointer to fma block for our mailbox */
639 gni_smsg_attr_t gnpr_smsg_attr; /* my short msg. attributes */
640 spinlock_t gnc_tx_lock; /* protect tx alloc/free */
641 __u8 gnc_tx_bits[GNILND_MAX_MSG_ID/8]; /* bit table for tx id */
642 int gnc_next_tx; /* next tx to use in tx_ref_table */
643 kgn_tx_t **gnc_tx_ref_table; /* table of TX descriptors for this conn */
644 int gnc_mbox_id; /* id of mbox in fma_blk */
645 short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
646 short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
649 typedef struct kgn_mdd_purgatory {
650 gni_mem_handle_t gmp_map_key; /* mapping key */
651 struct list_head gmp_list; /* entry point for purgatory list */
652 } kgn_mdd_purgatory_t;
654 typedef struct kgn_peer {
655 struct list_head gnp_list; /* stash on global peer list */
656 struct list_head gnp_connd_list; /* schedule on kgn_connd_peers */
657 struct list_head gnp_conns; /* all active connections and all conns in purgatory for the peer */
658 struct list_head gnp_tx_queue; /* msgs waiting for a conn */
659 kgn_net_t *gnp_net; /* net instance for this peer */
660 lnet_nid_t gnp_nid; /* who's on the other end(s) */
661 atomic_t gnp_refcount; /* # users */
662 __u32 gnp_host_id; /* ph. host ID of the peer */
663 short gnp_connecting; /* connection forming */
664 short gnp_pending_unlink; /* need last conn close to trigger unlink */
665 int gnp_last_errno; /* last error conn saw */
666 unsigned long gnp_last_alive; /* last time I had valid comms */
667 int gnp_last_dgram_errno; /* last error dgrams saw */
668 unsigned long gnp_last_dgram_time; /* last time I tried to connect */
669 unsigned long gnp_reconnect_time; /* CURRENT_SECONDS when reconnect OK */
670 unsigned long gnp_reconnect_interval; /* exponential backoff */
671 atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
674 /* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
675 * like lnet_parse. It allows a single pointer to let us get enough
676 * information in _recv and friends */
677 typedef struct kgn_rx {
678 kgn_conn_t *grx_conn; /* connection */
679 kgn_msg_t *grx_msg; /* message */
680 lnet_msg_t *grx_lntmsg; /* lnet msg for this rx (eager only) */
681 int grx_eager; /* if eager, we copied msg to somewhere */
682 struct timespec grx_received; /* time this msg received */
685 typedef struct kgn_data {
686 int kgn_init; /* initialisation state */
687 int kgn_shutdown; /* shut down? */
688 int kgn_wc_kill; /* Should I repost the WC */
689 atomic_t kgn_nthreads; /* # live threads */
690 int kgn_nresets; /* number of stack resets */
691 int kgn_in_reset; /* are we in stack reset ? */
693 kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
694 int kgn_ndevs; /* # devices */
696 int kgn_ruhroh_running; /* ruhroh thread is running */
697 int kgn_ruhroh_shutdown; /* ruhroh thread should or is shut down */
698 wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
699 int kgn_quiesce_trigger; /* should we quiesce ? */
700 atomic_t kgn_nquiesce; /* how many quiesced ? */
701 struct semaphore kgn_quiesce_sem; /* serialize ruhroh task, startup and shutdown */
702 int kgn_needs_reset; /* we need stack reset */
704 /* These next three members implement communication from gnilnd into
705 * the ruhroh task. To ensure correct operation of the task, code that
706 * writes into them must use memory barriers to ensure that the changes
707 * are visible to other cores in the order the members appear below. */
708 __u32 kgn_quiesce_secs; /* seconds to bump timeouts */
709 int kgn_bump_info_rdy; /* we have info needed to bump */
710 int kgn_needs_pause; /* we need to pause for network quiesce */
712 struct list_head *kgn_nets; /* hashtable of kgn_net instances */
713 struct rw_semaphore kgn_net_rw_sem; /* serialise gnn_shutdown, kgn_nets */
715 rwlock_t kgn_peer_conn_lock; /* stabilize peer/conn ops */
716 struct list_head *kgn_peers; /* hash table of all my known peers */
717 atomic_t kgn_npeers; /* # peers extant */
718 int kgn_peer_version; /* version flag for peer tables */
720 struct list_head *kgn_conns; /* conns hashed by cqid */
721 atomic_t kgn_nconns; /* # connections extant */
722 __u64 kgn_peerstamp; /* when I started up */
723 __u64 kgn_connstamp; /* conn stamp generator */
724 int kgn_conn_version; /* version flag for conn tables */
725 int kgn_next_cqid; /* cqid generator */
727 long kgn_new_min_timeout; /* minimum timeout on any new conn */
728 wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
729 spinlock_t kgn_reaper_lock; /* serialise */
731 cfs_mem_cache_t *kgn_rx_cache; /* rx descriptor space */
732 cfs_mem_cache_t *kgn_tx_cache; /* tx descriptor memory */
733 cfs_mem_cache_t *kgn_tx_phys_cache; /* tx phys descriptor memory */
734 atomic_t kgn_ntx; /* # tx in use */
735 cfs_mem_cache_t *kgn_dgram_cache; /* outgoing datagrams */
737 struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
738 __u64 kgn_cksum_npages; /* Number of pages allocated for checksumming */
739 atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
740 atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
742 atomic_t kgn_nkmap_short; /* # time we kmapped for a short kiov */
743 long kgn_rdmaq_override; /* bytes per second override */
745 struct kmem_cache *kgn_mbox_cache; /* mailboxes from not-GART */
747 atomic_t kgn_npending_unlink; /* # of peers pending unlink */
748 atomic_t kgn_npending_conns; /* # of conns with pending closes */
749 atomic_t kgn_npending_detach; /* # of conns with a pending detach */
753 extern kgn_data_t kgnilnd_data;
754 extern kgn_tunables_t kgnilnd_tunables;
756 extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
757 extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
758 extern void kgnilnd_schedule_conn(kgn_conn_t *conn);
761 kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
763 struct task_struct *thrd = kthread_run(fn, arg, "%s_%02d", name, id);
765 return PTR_ERR(thrd);
767 atomic_inc(&kgnilnd_data.kgn_nthreads);
772 kgnilnd_thread_fini(void)
774 atomic_dec(&kgnilnd_data.kgn_nthreads);
777 /* like mutex_trylock but with a jiffies spinner. This is to allow certain
778 * parts of the code to avoid a scheduler trip when the mutex is held
780 * Try to acquire the mutex atomically for 1 jiffie. Returns 1 if the mutex
781 * has been acquired successfully, and 0 on contention.
783 * NOTE: this function follows the spin_trylock() convention, so
784 * it is negated to the down_trylock() return values! Be careful
785 * about this when converting semaphore users to mutexes.
787 * This function must not be used in interrupt context. The
788 * mutex must be released by the same task that acquired it.
790 static inline int kgnilnd_mutex_trylock(struct mutex *lock)
793 unsigned long timeout;
795 LASSERT(!in_interrupt());
797 for (timeout = jiffies + 1; time_before(jiffies, timeout);) {
799 ret = mutex_trylock(lock);
806 /* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
809 _kgnilnd_debug_msg(kgn_msg_t *msg,
810 struct libcfs_debug_msg_data *data, const char *fmt, ... );
812 #define kgnilnd_debug_msg(msgdata, mask, cdls, msg, fmt, a...) \
814 CFS_CHECK_STACK(msgdata, mask, cdls); \
816 if (((mask) & D_CANTMASK) != 0 || \
817 ((libcfs_debug & (mask)) != 0 && \
818 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
819 _kgnilnd_debug_msg((msg), msgdata, fmt, ##a); \
822 /* for most callers (level is a constant) this is resolved at compile time */
823 #define GNIDBG_MSG(level, msg, fmt, args...) \
825 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
826 static cfs_debug_limit_state_t cdls; \
827 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
828 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
829 "$$ "fmt" from %s ", ## args, \
830 libcfs_nid2str((msg)->gnm_srcnid)); \
832 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
833 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
834 "$$ "fmt" from %s ", ## args, \
835 libcfs_nid2str((msg)->gnm_srcnid)); \
839 /* user puts 'to nid' in msg for us */
840 #define GNIDBG_TOMSG(level, msg, fmt, args...) \
842 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
843 static cfs_debug_limit_state_t cdls; \
844 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
845 kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
846 "$$ "fmt" ", ## args); \
848 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
849 kgnilnd_debug_msg(&msgdata, level, NULL, msg, \
850 "$$ "fmt" ", ## args); \
855 _kgnilnd_debug_conn(kgn_conn_t *conn,
856 struct libcfs_debug_msg_data *data, const char *fmt, ... );
858 #define kgnilnd_debug_conn(msgdata, mask, cdls, conn, fmt, a...) \
860 CFS_CHECK_STACK(msgdata, mask, cdls); \
862 if (((mask) & D_CANTMASK) != 0 || \
863 ((libcfs_debug & (mask)) != 0 && \
864 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
865 _kgnilnd_debug_conn((conn), msgdata, fmt, ##a); \
868 /* for most callers (level is a constant) this is resolved at compile time */
869 #define GNIDBG_CONN(level, conn, fmt, args...) \
871 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
872 static cfs_debug_limit_state_t cdls; \
873 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
874 kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
875 "$$ "fmt" ", ## args); \
877 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
878 kgnilnd_debug_conn(&msgdata, level, NULL, conn, \
879 "$$ "fmt" ", ## args); \
884 _kgnilnd_debug_tx(kgn_tx_t *tx,
885 struct libcfs_debug_msg_data *data, const char *fmt, ... );
887 #define kgnilnd_debug_tx(msgdata, mask, cdls, tx, fmt, a...) \
889 CFS_CHECK_STACK(msgdata, mask, cdls); \
891 if (((mask) & D_CANTMASK) != 0 || \
892 ((libcfs_debug & (mask)) != 0 && \
893 (libcfs_subsystem_debug & DEBUG_SUBSYSTEM) != 0)) \
894 _kgnilnd_debug_tx((tx), msgdata, fmt, ##a); \
897 /* for most callers (level is a constant) this is resolved at compile time */
898 #define GNIDBG_TX(level, tx, fmt, args...) \
900 if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
901 static cfs_debug_limit_state_t cdls; \
902 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
903 kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
904 "$$ "fmt" ", ## args); \
906 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, NULL); \
907 kgnilnd_debug_tx(&msgdata, level, NULL, tx, \
908 "$$ "fmt" ", ## args); \
912 #define GNITX_ASSERTF(tx, cond, fmt, a...) \
914 if (unlikely(!(cond))) { \
915 GNIDBG_TX(D_EMERG, tx, "ASSERTION(" #cond ") failed:" fmt, a); \
920 #define GNILND_IS_QUIESCED \
921 (atomic_read(&kgnilnd_data.kgn_nquiesce) == \
922 atomic_read(&kgnilnd_data.kgn_nthreads))
924 #define KGNILND_SPIN_QUIESCE \
926 /* E.T phone home */ \
927 atomic_inc(&kgnilnd_data.kgn_nquiesce); \
928 CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
929 while (kgnilnd_data.kgn_quiesce_trigger) { \
930 set_current_state(TASK_INTERRUPTIBLE); \
931 cfs_schedule_timeout_and_set_state(TASK_INTERRUPTIBLE, \
932 cfs_time_seconds(1)); \
934 /* Mom, my homework is done */ \
935 CDEBUG(D_NET, "Waking up from thread pause\n"); \
936 atomic_dec(&kgnilnd_data.kgn_nquiesce); \
939 /* use macros for addref/decref to get the calling function name in the CDEBUG */
941 #error "this code uses actions inside LASSERT for ref counting"
944 #define kgnilnd_admin_addref(atomic) \
946 int val = atomic_inc_return(&atomic); \
947 LASSERTF(val > 0, #atomic " refcount %d\n", val); \
948 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
951 #define kgnilnd_admin_decref(atomic) \
953 int val = atomic_dec_return(&atomic); \
954 LASSERTF(val >=0, #atomic " refcount %d\n", val); \
955 CDEBUG(D_NETTRACE, #atomic " refcount %d\n", val); \
958 #define kgnilnd_net_addref(net) \
960 int val = atomic_inc_return(&net->gnn_refcount); \
961 LASSERTF(val > 1, "net %p refcount %d\n", net, val); \
962 CDEBUG(D_NETTRACE, "net %p->%s++ (%d)\n", net, \
963 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
966 #define kgnilnd_net_decref(net) \
968 int val = atomic_dec_return(&net->gnn_refcount); \
969 LASSERTF(val >= 0, "net %p refcount %d\n", net, val); \
970 CDEBUG(D_NETTRACE, "net %p->%s-- (%d)\n", net, \
971 libcfs_nid2str(net->gnn_ni->ni_nid), val); \
974 #define kgnilnd_peer_addref(peer) \
976 int val = atomic_inc_return(&peer->gnp_refcount); \
977 LASSERTF(val > 1, "peer %p refcount %d\n", peer, val); \
978 CDEBUG(D_NETTRACE, "peer %p->%s++ (%d)\n", peer, \
979 libcfs_nid2str(peer->gnp_nid), val); \
982 #define kgnilnd_peer_decref(peer) \
984 int val = atomic_dec_return(&peer->gnp_refcount); \
985 LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
986 CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
987 libcfs_nid2str(peer->gnp_nid), val); \
988 if (atomic_read(&peer->gnp_refcount) == 0) \
989 kgnilnd_destroy_peer(peer); \
992 #define kgnilnd_conn_addref(conn) \
997 val = atomic_inc_return(&conn->gnc_refcount); \
998 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1001 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1003 CDEBUG(D_NETTRACE, "conn %p->%s++ (%d)\n", conn, \
1005 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1010 /* we hijack conn_decref && gnc_refcount = 1 to allow us to push the conn
1011 * through the scheduler thread to get the EP destroyed. This avoids some
1012 * messy semaphore business and allows us to reuse the connd_list and existing
1013 * linkage and avoid creating extra lists just for destroying EPs */
1015 /* Safety Disclaimer:
1016 * Q: If we decrement the refcount and then check it again, is it possible that
1017 * another caller could have passed through this macro concurrently? If so,
1018 * then it is possible that both will attempt to call kgnilnd_destroy_conn().
1020 * A: Yes, entirely possible in most cases, but we can't get concurrent users
1021 * once we are refcount <= 2. It hinges around gnc_state and membership of
1022 * gnc_hashlist. There are two ways to find a connection - either ask for
1023 * it from the peer, kgnilnd_find_conn_locked(peer) or from the CQ id,
1024 * kgnilnd_cqid2conn_locked(id). While a conn is live, we'll have at least
1027 * - #1 from create (kgnilnd_create_conn)
1028 * - #2 for EP (kgnilnd_create_conn)
1029 * - #3 - living on peer (gnc_list, kgnilnd_finish_connect)
1030 * - #4 living in global hash (gnc_hashlist, kgnilnd_finish_connect).
1032 * Actually, only 3 live, as at the end of kgnilnd_finish_connect, we drop:
1033 * - #1 - the ref the dgram inherited from kgnilnd_create_conn.
1035 * There could be more from TX descriptors during the lifetime of a live
1038 * If we nuke the conn before finish_connect, we won't have parallel paths
1039 * because nobody besides the dgram handler for the single outstanding
1040 * dgram can find the connection as it isn't in any searchable tables yet.
1042 * This leaves connection close, we'll drop 2 refs (#4 and #3) but only
1043 * after calling kgnilnd_schedule_conn, which would add a new ref (#5). At
1044 * this point gnc_refcount=2 (#2, #5). We have a 'maybe' send of the CLOSE
1045 * now on the next scheduler loop, this could be #6 (schedule_conn again)
1046 * and #7 (TX on gnc_fmaq). Both would be cleared quickly as that TX is
1047 * sent. Now the gnc_state == CLOSED, so we hit
1048 * kgnilnd_complete_closed_conn. At this point, nobody can 'find' this conn
1049 * - we've nuked them from the peer and CQ id tables, so we own them and
1050 * are guaranteed serial access - hence the complete lack of conn list
1051 * locking in kgnilnd_complete_closed_conn. We are free then to mark the
1052 * conn DESTROY_EP (add #6 for schedule_conn), then lose #5 in
1053 * kgnilnd_process_conns. Then the next scheduler loop would call
1054 * kgnilnd_destroy_conn_ep (drop #2 for EP) and lose #6 (refcount=0) in
1055 * kgnilnd_process_conns.
1057 * Clearly, we are totally safe. Clearly.
1060 #define kgnilnd_conn_decref(conn) \
1065 val = atomic_dec_return(&conn->gnc_refcount); \
1066 LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
1069 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1071 CDEBUG(D_NETTRACE, "conn %p->%s-- (%d)\n", conn, \
1073 ? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
1077 if ((atomic_read(&conn->gnc_refcount) == 1) && \
1078 (conn->gnc_ephandle != NULL) && \
1079 (conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
1080 set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
1081 kgnilnd_schedule_conn(conn); \
1082 } else if (atomic_read(&conn->gnc_refcount) == 0) { \
1083 kgnilnd_destroy_conn(conn); \
1087 static inline struct list_head *
1088 kgnilnd_nid2peerlist(lnet_nid_t nid)
1090 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) % *kgnilnd_tunables.kgn_peer_hash_size;
1092 RETURN(&kgnilnd_data.kgn_peers[hash]);
1095 static inline struct list_head *
1096 kgnilnd_netnum2netlist(__u16 netnum)
1098 unsigned int hash = ((unsigned int) netnum) % *kgnilnd_tunables.kgn_net_hash_size;
1100 RETURN(&kgnilnd_data.kgn_nets[hash]);
1104 kgnilnd_peer_active(kgn_peer_t *peer)
1106 /* Am I in the peer hash table? */
1107 return (!list_empty(&peer->gnp_list));
1110 /* need write_lock on kgn_peer_conn_lock */
1112 kgnilnd_can_unlink_peer_locked(kgn_peer_t *peer)
1114 CDEBUG(D_NET, "peer 0x%p->%s conns? %d tx? %d\n",
1115 peer, libcfs_nid2str(peer->gnp_nid),
1116 !list_empty(&peer->gnp_conns),
1117 !list_empty(&peer->gnp_tx_queue));
1119 /* kgn_peer_conn_lock protects us from conflict with
1120 * kgnilnd_peer_notify and gnp_persistent */
1121 RETURN ((list_empty(&peer->gnp_conns)) &&
1122 (list_empty(&peer->gnp_tx_queue)));
1125 /* returns positive if error was for a clean shutdown of conn */
1127 kgnilnd_conn_clean_errno(int errno)
1129 /* - ESHUTDOWN - LND is unloading
1130 * - EUCLEAN - admin requested via "lctl del_peer"
1131 * - ENETRESET - admin requested via "lctl disconnect"
1132 * - ENOTRECOVERABLE - stack reset
1133 * - EISCONN - cleared via "lctl push"
1134 * not doing ESTALE - that isn't clean */
1135 RETURN ((errno == 0) ||
1136 (errno == -ESHUTDOWN) ||
1137 (errno == -EUCLEAN) ||
1138 (errno == -ENETRESET) ||
1139 (errno == -EISCONN) ||
1140 (errno == -ENOTRECOVERABLE));
1143 /* returns positive if error results in purgatory hold */
1145 kgnilnd_check_purgatory_errno(int errno)
1147 /* We don't want to save the purgatory lists these cases:
1148 * - EUCLEAN - admin requested via "lctl del_peer"
1149 * - ESHUTDOWN - LND is unloading
1151 RETURN ((errno != -ESHUTDOWN) &&
1152 (errno != -EUCLEAN));
1156 /* returns positive if a purgatory hold is needed */
1158 kgnilnd_check_purgatory_conn(kgn_conn_t *conn)
1162 if (conn->gnc_peer) {
1163 loopback = conn->gnc_peer->gnp_nid ==
1164 conn->gnc_peer->gnp_net->gnn_ni->ni_nid;
1166 /* short circuit - a conn that didn't complete
1167 * setup never needs a purgatory hold */
1170 CDEBUG(D_NETTRACE, "conn 0x%p->%s loopback %d close_recvd %d\n",
1171 conn, conn->gnc_peer ?
1172 libcfs_nid2str(conn->gnc_peer->gnp_nid) :
1174 loopback, conn->gnc_close_recvd);
1176 /* we only use a purgatory hold if we've not received the CLOSE msg
1177 * from our peer - without that message, we can't know the state of
1178 * the other end of this connection and must put it into purgatory
1179 * to prevent reuse and corruption.
1180 * The theory is that a TX error can be communicated in all other cases
1182 RETURN(likely(!loopback) && !conn->gnc_close_recvd &&
1183 kgnilnd_check_purgatory_errno(conn->gnc_error));
1186 static inline const char *
1187 kgnilnd_tx_state2str(kgn_tx_list_state_t state);
1189 static inline struct list_head *
1190 kgnilnd_tx_state2list(kgn_peer_t *peer, kgn_conn_t *conn,
1191 kgn_tx_list_state_t to_state)
1194 case GNILND_TX_PEERQ:
1195 return &peer->gnp_tx_queue;
1196 case GNILND_TX_FMAQ:
1197 return &conn->gnc_fmaq;
1198 case GNILND_TX_LIVE_FMAQ:
1199 case GNILND_TX_LIVE_RDMAQ:
1200 case GNILND_TX_DYING:
1202 case GNILND_TX_MAPQ:
1203 return &conn->gnc_device->gnd_map_tx;
1204 case GNILND_TX_RDMAQ:
1205 return &conn->gnc_device->gnd_rdmaq;
1207 /* IDLE, FREED or ALLOCD is not valid "on list" state */
1208 CERROR("invalid state requested: %s\n",
1209 kgnilnd_tx_state2str(to_state));
1215 /* should hold tx, conn or peer lock when calling */
1217 kgnilnd_tx_add_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1218 kgn_conn_t *conn, kgn_tx_list_state_t state,
1221 struct list_head *list = NULL;
1223 /* make sure we have a sane TX state to start */
1224 GNITX_ASSERTF(tx, (tx->tx_list_p == NULL &&
1225 tx->tx_list_state == GNILND_TX_ALLOCD) &&
1226 list_empty(&tx->tx_list),
1227 "bad state with tx_list %s",
1228 list_empty(&tx->tx_list) ? "empty" : "not empty");
1230 /* WTF - you are already on that state buttmunch */
1231 GNITX_ASSERTF(tx, state != tx->tx_list_state,
1232 "already at %s", kgnilnd_tx_state2str(state));
1234 /* get proper list from the state requested */
1235 list = kgnilnd_tx_state2list(peer, conn, state);
1239 case GNILND_TX_PEERQ:
1240 kgnilnd_peer_addref(peer);
1242 case GNILND_TX_ALLOCD:
1243 /* no refs needed */
1245 case GNILND_TX_FMAQ:
1246 kgnilnd_conn_addref(conn);
1248 case GNILND_TX_MAPQ:
1249 atomic_inc(&conn->gnc_device->gnd_nq_map);
1250 kgnilnd_conn_addref(conn);
1252 case GNILND_TX_LIVE_FMAQ:
1253 atomic_inc(&conn->gnc_nlive_fma);
1254 kgnilnd_conn_addref(conn);
1256 case GNILND_TX_LIVE_RDMAQ:
1257 atomic_inc(&conn->gnc_nlive_rdma);
1258 kgnilnd_conn_addref(conn);
1260 case GNILND_TX_RDMAQ:
1261 atomic_inc(&conn->gnc_nq_rdma);
1262 kgnilnd_conn_addref(conn);
1264 case GNILND_TX_DYING:
1265 kgnilnd_conn_addref(conn);
1268 CERROR("invalid state requested: %s\n",
1269 kgnilnd_tx_state2str(state));
1274 /* if this changes, change kgnilnd_alloc_tx */
1275 tx->tx_list_state = state;
1277 /* some states don't have lists - we track them in the per conn
1278 * TX table instead. Waste not, want not! */
1280 tx->tx_list_p = list;
1282 list_add_tail(&tx->tx_list, list);
1284 list_add(&tx->tx_list, list);
1286 /* set dummy list_p to make book keeping happy and let debugging
1287 * be a hair easier */
1288 tx->tx_list_p = (void *)state;
1291 GNIDBG_TX(D_NET, tx, "onto %s->0x%p",
1292 kgnilnd_tx_state2str(state), list);
1296 kgnilnd_tx_del_state_locked(kgn_tx_t *tx, kgn_peer_t *peer,
1297 kgn_conn_t *conn, kgn_tx_list_state_t new_state)
1299 /* These is only 1 "off-list" state */
1300 GNITX_ASSERTF(tx, new_state == GNILND_TX_ALLOCD,
1301 "invalid new_state %s", kgnilnd_tx_state2str(new_state));
1303 /* new_state == ALLOCD means we are deallocating this tx,
1304 * so make sure it was on a valid list to start with */
1305 GNITX_ASSERTF(tx, (tx->tx_list_p != NULL) &&
1306 (((tx->tx_list_state == GNILND_TX_LIVE_FMAQ) ||
1307 (tx->tx_list_state == GNILND_TX_LIVE_RDMAQ) ||
1308 (tx->tx_list_state == GNILND_TX_DYING)) == list_empty(&tx->tx_list)),
1311 GNIDBG_TX(D_NET, tx, "off %p", tx->tx_list_p);
1314 switch (tx->tx_list_state) {
1315 case GNILND_TX_PEERQ:
1316 kgnilnd_peer_decref(peer);
1318 case GNILND_TX_FREED:
1319 case GNILND_TX_IDLE:
1320 case GNILND_TX_ALLOCD:
1321 /* no refs needed */
1323 case GNILND_TX_DYING:
1324 kgnilnd_conn_decref(conn);
1326 case GNILND_TX_FMAQ:
1327 kgnilnd_conn_decref(conn);
1329 case GNILND_TX_MAPQ:
1330 atomic_dec(&conn->gnc_device->gnd_nq_map);
1331 kgnilnd_conn_decref(conn);
1333 case GNILND_TX_LIVE_FMAQ:
1334 atomic_dec(&conn->gnc_nlive_fma);
1335 kgnilnd_conn_decref(conn);
1337 case GNILND_TX_LIVE_RDMAQ:
1338 atomic_dec(&conn->gnc_nlive_rdma);
1339 kgnilnd_conn_decref(conn);
1341 case GNILND_TX_RDMAQ:
1342 atomic_dec(&conn->gnc_nq_rdma);
1343 kgnilnd_conn_decref(conn);
1344 /* don't need to assert on default, already did in set */
1347 /* for ALLOCD, this might already be true, but no harm doing it again */
1348 list_del_init(&tx->tx_list);
1349 tx->tx_list_p = NULL;
1350 tx->tx_list_state = new_state;
1354 kgnilnd_tx_mapped(kgn_tx_t *tx)
1356 return (tx->tx_buftype == GNILND_BUF_VIRT_MAPPED ||
1357 tx->tx_buftype == GNILND_BUF_PHYS_MAPPED);
1360 static inline struct list_head *
1361 kgnilnd_cqid2connlist(__u32 cqid)
1363 unsigned int hash = cqid % *kgnilnd_tunables.kgn_peer_hash_size;
1365 return (&kgnilnd_data.kgn_conns [hash]);
1368 static inline kgn_conn_t *
1369 kgnilnd_cqid2conn_locked(__u32 cqid)
1371 struct list_head *conns = kgnilnd_cqid2connlist(cqid);
1372 struct list_head *tmp;
1375 list_for_each(tmp, conns) {
1376 conn = list_entry(tmp, kgn_conn_t, gnc_hashlist);
1378 if (conn->gnc_cqid == cqid)
1385 /* returns 1..GNILND_MAX_CQID on success, 0 on failure */
1387 kgnilnd_get_cqid_locked(void)
1393 cqid = kgnilnd_data.kgn_next_cqid++;
1394 if (kgnilnd_data.kgn_next_cqid >= GNILND_MAX_CQID) {
1398 kgnilnd_data.kgn_next_cqid = 1;
1401 } while (kgnilnd_cqid2conn_locked(cqid) != NULL);
1407 kgnilnd_validate_tx_ev_id(kgn_tx_ev_id_t *ev_id, kgn_tx_t **txp, kgn_conn_t **connp)
1409 kgn_tx_t *tx = NULL;
1410 kgn_conn_t *conn = NULL;
1412 /* set to NULL so any early return is an error */
1416 LASSERTF((ev_id->txe_idx > 0) &&
1417 (ev_id->txe_idx < GNILND_MAX_MSG_ID),
1418 "bogus txe_idx %d >= %d\n",
1419 ev_id->txe_idx, GNILND_MAX_MSG_ID);
1421 LASSERTF((ev_id->txe_cqid > 0) &&
1422 (ev_id->txe_cqid < GNILND_MAX_CQID),
1423 "bogus txe_cqid %d >= %d\n",
1424 ev_id->txe_cqid, GNILND_MAX_CQID);
1426 read_lock(&kgnilnd_data.kgn_peer_conn_lock);
1427 conn = kgnilnd_cqid2conn_locked(ev_id->txe_cqid);
1430 /* Conn was destroyed? */
1431 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1432 CDEBUG(D_NET, "CQID %d lookup failed\n", ev_id->txe_cqid);
1435 /* just insurance */
1436 kgnilnd_conn_addref(conn);
1437 read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
1439 /* we know this is safe - as the TX won't be reused until AFTER
1440 * the conn is unlinked from the cqid hash, so we can use the TX
1441 * (serializing to avoid any cache oddness) freely from the conn tx ref table */
1443 spin_lock(&conn->gnc_tx_lock);
1444 tx = conn->gnc_tx_ref_table[ev_id->txe_idx];
1445 spin_unlock(&conn->gnc_tx_lock);
1447 /* We could have a tx that was cleared out by other forces
1448 * lctl disconnect or del_peer. */
1450 CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
1451 kgnilnd_conn_decref(conn);
1455 /* check tx->tx_msg magic to make sure kgni didn't eat it */
1456 GNITX_ASSERTF(tx, tx->tx_msg.gnm_magic == GNILND_MSG_MAGIC,
1457 "came back from kgni with bad magic %x", tx->tx_msg.gnm_magic);
1459 GNITX_ASSERTF(tx, tx->tx_id.txe_idx == ev_id->txe_idx,
1460 "conn 0x%p->%s tx_ref_table hosed: wanted txe_idx %d "
1461 "found tx %p txe_idx %d",
1462 conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
1463 ev_id->txe_idx, tx, tx->tx_id.txe_idx);
1465 GNITX_ASSERTF(tx, tx->tx_conn != NULL, "tx with NULL connection", NULL);
1467 GNITX_ASSERTF(tx, tx->tx_conn == conn, "tx conn does not equal conn", NULL);
1472 GNIDBG_TX(D_NET, tx, "validated to 0x%p", conn);
1475 /* set_normalized_timepsec isn't exported from the kernel, so
1476 * we need to do the same thing inline */
1477 static inline struct timespec
1478 kgnilnd_ts_sub(struct timespec lhs, struct timespec rhs)
1484 sec = lhs.tv_sec - rhs.tv_sec;
1485 nsec = lhs.tv_nsec - rhs.tv_nsec;
1487 while (nsec >= NSEC_PER_SEC) {
1488 nsec -= NSEC_PER_SEC;
1492 nsec += NSEC_PER_SEC;
1501 kgnilnd_count_list(struct list_head *q)
1503 struct list_head *e;
1506 list_for_each(e, q) {
1513 /* kgnilnd_find_net adds a reference to the net it finds
1514 * this is so the net will not be removed before the calling function
1515 * has time to use the data returned. This reference needs to be released
1516 * by the calling function once it has finished using the returned net
1520 kgnilnd_find_net(lnet_nid_t nid, kgn_net_t **netp)
1525 rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
1531 list_for_each_entry(net, kgnilnd_netnum2netlist(LNET_NETNUM(LNET_NIDNET(nid))), gnn_list) {
1532 if (!net->gnn_shutdown && LNET_NIDNET(net->gnn_ni->ni_nid) == LNET_NIDNET(nid)) {
1533 kgnilnd_net_addref(net);
1534 up_read(&kgnilnd_data.kgn_net_rw_sem);
1540 up_read(&kgnilnd_data.kgn_net_rw_sem);
1545 #ifdef CONFIG_DEBUG_SLAB
1546 #define KGNILND_POISON(ptr, c, s) do {} while(0)
1548 #define KGNILND_POISON(ptr, c, s) memset(ptr, c, s)
1551 int kgnilnd_dev_init(kgn_device_t *dev);
1552 void kgnilnd_dev_fini(kgn_device_t *dev);
1553 int kgnilnd_startup(lnet_ni_t *ni);
1554 void kgnilnd_shutdown(lnet_ni_t *ni);
1555 int kgnilnd_base_startup(void);
1556 void kgnilnd_base_shutdown(void);
1558 int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
1559 int kgnilnd_map_phys_fmablk(kgn_device_t *device);
1560 void kgnilnd_unmap_phys_fmablk(kgn_device_t *device);
1561 void kgnilnd_free_phys_fmablk(kgn_device_t *device);
1563 int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
1564 void kgnilnd_query(lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when);
1565 int kgnilnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1566 int kgnilnd_eager_recv(lnet_ni_t *ni, void *private,
1567 lnet_msg_t *lntmsg, void **new_private);
1568 int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
1569 int delayed, unsigned int niov,
1570 struct iovec *iov, lnet_kiov_t *kiov,
1571 unsigned int offset, unsigned int mlen, unsigned int rlen);
1573 __u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
1575 /* purgatory functions */
1576 void kgnilnd_add_purgatory_locked(kgn_conn_t *conn, kgn_peer_t *peer);
1577 void kgnilnd_mark_for_detach_purgatory_all_locked(kgn_peer_t *peer);
1578 void kgnilnd_detach_purgatory_locked(kgn_conn_t *conn, struct list_head *conn_list);
1579 void kgnilnd_release_purgatory_list(struct list_head *conn_list);
1581 void kgnilnd_update_reaper_timeout(long timeout);
1582 void kgnilnd_unmap_buffer(kgn_tx_t *tx, int error);
1583 kgn_tx_t *kgnilnd_new_tx_msg(int type, lnet_nid_t source);
1584 void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
1585 void kgnilnd_txlist_done(struct list_head *txlist, int error);
1586 void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
1587 void kgnilnd_schedule_conn(kgn_conn_t *conn);
1588 void kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
1590 void kgnilnd_schedule_dgram(kgn_device_t *dev);
1591 int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net);
1592 void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
1593 int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
1595 kgn_peer_t *kgnilnd_find_peer_locked(lnet_nid_t nid);
1596 int kgnilnd_del_conn_or_peer(kgn_net_t *net, lnet_nid_t nid, int command, int error);
1597 void kgnilnd_peer_increase_reconnect_locked(kgn_peer_t *peer);
1598 void kgnilnd_queue_reply(kgn_conn_t *conn, kgn_tx_t *tx);
1599 void kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx);
1600 void kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target);
1601 int kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full);
1602 void kgnilnd_consume_rx(kgn_rx_t *rx);
1604 void kgnilnd_schedule_device(kgn_device_t *dev);
1605 void kgnilnd_device_callback(__u32 devid, __u64 arg);
1606 void kgnilnd_schedule_device_timer(unsigned long arg);
1608 int kgnilnd_reaper(void *arg);
1609 int kgnilnd_scheduler(void *arg);
1610 int kgnilnd_dgram_mover(void *arg);
1612 int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
1613 int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1614 kgn_conn_t *kgnilnd_find_conn_locked(kgn_peer_t *peer);
1615 int kgnilnd_get_conn(kgn_conn_t **connp, kgn_peer_t);
1616 kgn_conn_t *kgnilnd_find_or_create_conn_locked(kgn_peer_t *peer);
1617 void kgnilnd_peer_cancel_tx_queue(kgn_peer_t *peer);
1618 void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
1619 int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
1620 void kgnilnd_peer_alive(kgn_peer_t *peer);
1621 void kgnilnd_peer_notify(kgn_peer_t *peer, int error);
1622 void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
1623 void kgnilnd_close_conn(kgn_conn_t *conn, int error);
1624 void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
1625 void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
1627 int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
1629 int kgnilnd_tunables_init(void);
1630 void kgnilnd_tunables_fini(void);
1631 void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
1633 void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
1634 void kgnilnd_pause_threads(void);
1635 int kgnilnd_hw_in_quiesce(void);
1636 int kgnilnd_check_hw_quiesce(void);
1637 void kgnilnd_quiesce_wait(char *reason);
1638 void kgnilnd_quiesce_end_callback(gni_nic_handle_t nic_handle, uint64_t msecs);
1639 int kgnilnd_ruhroh_thread(void *arg);
1640 void kgnilnd_reset_stack(void);
1641 void kgnilnd_critical_error(gni_err_handle_t err_handle);
1643 void kgnilnd_insert_sysctl(void);
1644 void kgnilnd_remove_sysctl(void);
1645 void kgnilnd_proc_init(void);
1646 void kgnilnd_proc_fini(void);
1649 void kgnilnd_release_mbox(kgn_conn_t *conn, int purgatory_hold);
1651 int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
1652 void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
1653 void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram);
1655 int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
1656 int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
1657 int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
1658 void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
1660 int kgnilnd_dgram_waitq(void *arg);
1662 int kgnilnd_set_conn_params(kgn_dgram_t *dgram);
1664 /* struct2str functions - we don't use a default: case to cause the compile
1665 * to fail if there is a missing case. This allows us to hide these down here
1666 * out of the way but ensure we'll catch any updates to the enum/types
1669 #define DO_TYPE(x) case x: return #x;
1670 static inline const char *
1671 kgnilnd_fmablk_state2str(kgn_fmablk_state_t state)
1673 /* Only want single char string for this */
1675 case GNILND_FMABLK_IDLE:
1677 case GNILND_FMABLK_PHYS:
1679 case GNILND_FMABLK_VIRT:
1681 case GNILND_FMABLK_FREED:
1684 return "<unknown state>";
1687 static inline const char *
1688 kgnilnd_msgtype2str(int type)
1691 DO_TYPE(GNILND_MSG_NONE);
1692 DO_TYPE(GNILND_MSG_NOOP);
1693 DO_TYPE(GNILND_MSG_IMMEDIATE);
1694 DO_TYPE(GNILND_MSG_PUT_REQ);
1695 DO_TYPE(GNILND_MSG_PUT_NAK);
1696 DO_TYPE(GNILND_MSG_PUT_ACK);
1697 DO_TYPE(GNILND_MSG_PUT_DONE);
1698 DO_TYPE(GNILND_MSG_GET_REQ);
1699 DO_TYPE(GNILND_MSG_GET_NAK);
1700 DO_TYPE(GNILND_MSG_GET_DONE);
1701 DO_TYPE(GNILND_MSG_CLOSE);
1703 return "<unknown msg type>";
1706 static inline const char *
1707 kgnilnd_tx_state2str(kgn_tx_list_state_t state)
1710 DO_TYPE(GNILND_TX_IDLE);
1711 DO_TYPE(GNILND_TX_ALLOCD);
1712 DO_TYPE(GNILND_TX_PEERQ);
1713 DO_TYPE(GNILND_TX_MAPQ);
1714 DO_TYPE(GNILND_TX_FMAQ);
1715 DO_TYPE(GNILND_TX_LIVE_FMAQ);
1716 DO_TYPE(GNILND_TX_RDMAQ);
1717 DO_TYPE(GNILND_TX_LIVE_RDMAQ);
1718 DO_TYPE(GNILND_TX_DYING);
1719 DO_TYPE(GNILND_TX_FREED);
1721 return "<unknown state>";
1724 static inline const char *
1725 kgnilnd_conn_state2str(kgn_conn_t *conn)
1727 kgn_conn_state_t state = conn->gnc_state;
1729 DO_TYPE(GNILND_CONN_DUMMY);
1730 DO_TYPE(GNILND_CONN_LISTEN);
1731 DO_TYPE(GNILND_CONN_CONNECTING);
1732 DO_TYPE(GNILND_CONN_ESTABLISHED);
1733 DO_TYPE(GNILND_CONN_CLOSING);
1734 DO_TYPE(GNILND_CONN_CLOSED);
1735 DO_TYPE(GNILND_CONN_DONE);
1736 DO_TYPE(GNILND_CONN_DESTROY_EP);
1741 static inline const char *
1742 kgnilnd_connreq_type2str(kgn_connreq_t *connreq)
1744 kgn_connreq_type_t type = connreq->gncr_type;
1747 DO_TYPE(GNILND_CONNREQ_REQ);
1748 DO_TYPE(GNILND_CONNREQ_NAK);
1749 DO_TYPE(GNILND_CONNREQ_CLOSE);
1754 static inline const char *
1755 kgnilnd_dgram_state2str(kgn_dgram_t *dgram)
1757 kgn_dgram_state_t state = dgram->gndg_state;
1760 DO_TYPE(GNILND_DGRAM_USED);
1761 DO_TYPE(GNILND_DGRAM_POSTING);
1762 DO_TYPE(GNILND_DGRAM_POSTED);
1763 DO_TYPE(GNILND_DGRAM_PROCESSING);
1764 DO_TYPE(GNILND_DGRAM_DONE);
1765 DO_TYPE(GNILND_DGRAM_CANCELED);
1770 static inline const char *
1771 kgnilnd_dgram_type2str(kgn_dgram_t *dgram)
1773 kgn_dgram_type_t type = dgram->gndg_type;
1776 DO_TYPE(GNILND_DGRAM_REQ);
1777 DO_TYPE(GNILND_DGRAM_WC_REQ);
1778 DO_TYPE(GNILND_DGRAM_NAK);
1779 DO_TYPE(GNILND_DGRAM_CLOSE);
1787 /* API wrapper functions - include late to pick up all of the other defines */
1788 #include "gnilnd_api_wrap.h"
1790 #endif /* _GNILND_GNILND_H_ */