*
* Copyright (C) 2009-2012 Cray, Inc.
*
+ * Copyright (c) 2014, 2016, Intel Corporation.
+ *
* Derived from work by: Eric Barton <eric@bartonsoftware.com>
* Author: Nic Henke <nic@cray.com>
* Author: James Shimek <jshimek@cray.com>
#ifndef _GNILND_GNILND_H_
#define _GNILND_GNILND_H_
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/string.h>
+#include <linux/stat.h>
+#include <linux/errno.h>
+#ifdef HAVE_LINUX_KERNEL_LOCK
+#include <linux/smp_lock.h>
+#endif
+#include <linux/unistd.h>
+#include <linux/uio.h>
+#include <linux/time.h>
+#include <asm/timex.h>
+
+#include <asm/uaccess.h>
+#include <asm/io.h>
+
+#include <linux/init.h>
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/stat.h>
+#include <linux/list.h>
+#include <linux/kmod.h>
+#include <linux/sysctl.h>
+#include <linux/kthread.h>
+#include <linux/nmi.h>
+
+#include <net/sock.h>
+#include <linux/in.h>
+#include <linux/nmi.h>
+
#define DEBUG_SUBSYSTEM S_LND
#include <libcfs/libcfs.h>
#include <lnet/lnet.h>
#include <lnet/lib-lnet.h>
-#include <lnet/lnet-sysctl.h>
#include <gni_pub.h>
#include "gnilnd_version.h"
-#include "gnilnd_hss_ops.h"
+
/* tunables determined at compile time */
#define GNILND_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
-#define GNILND_BASE_TIMEOUT 60 /* default sane timeout */
#define GNILND_TO2KA(t) (((t)-1)/2) /* timeout -> keepalive interval */
#define GNILND_MIN_RECONNECT_TO (GNILND_BASE_TIMEOUT/4)
#define GNILND_MAX_RECONNECT_TO GNILND_BASE_TIMEOUT
#define GNILND_HARDWARE_TIMEOUT 15 /* maximum time for data to travel between nodes */
#define GNILND_MDD_TIMEOUT 15 /* MDD hold timeout in minutes */
+#define GNILND_SCHED_TIMEOUT 1
+#define GNILND_DGRAM_TIMEOUT 2
+#define GNILND_FAST_MAPPING_TRY \
+ *kgnilnd_tunables.kgn_max_retransmits /* maximum number to attempt mapping of a tx */
+#define GNILND_MAP_RETRY_RATE 1 /* interval between mapping attempts in jiffies */
+
+/* map failure timeout */
+#define GNILND_MAP_TIMEOUT \
+ (cfs_time_seconds(*kgnilnd_tunables.kgn_timeout * \
+ *kgnilnd_tunables.kgn_timeout))
/* reaper thread wakup interval */
#define GNILND_REAPER_THREAD_WAKE 1
/* fixed constants */
#define GNILND_MAXDEVS 1 /* max # of GNI devices currently supported */
#define GNILND_MBOX_CREDITS 256 /* number of credits per mailbox */
-#define GNILND_COOKIE 0xa3579 /* cookie used by along with ptag by GNI */
-
+#define GNILND_CONN_MAGIC 0xa100f /* magic value for verifying connection validity */
/* checksum values */
#define GNILND_CHECKSUM_OFF 0 /* checksum turned off */
#define GNILND_CHECKSUM_SMSG_HEADER 1 /* Only checksum SMSG header */
/* tune down some COMPUTE options as they won't see the same number of connections and
* don't need the throughput of multiple threads by default */
#if defined(CONFIG_CRAY_COMPUTE)
+#ifdef CONFIG_MK1OM
+#define GNILND_SCHED_THREADS 2 /* default # of kgnilnd_scheduler threads */
+#else
#define GNILND_SCHED_THREADS 1 /* default # of kgnilnd_scheduler threads */
+#endif
#define GNILND_FMABLK 64 /* default number of mboxes per fmablk */
+#define GNILND_SCHED_NICE 0 /* default nice value for scheduler threads */
+#define GNILND_COMPUTE 1 /* compute image */
+#define GNILND_FAST_RECONNECT 1 /* Fast Reconnect option */
#else
-#define GNILND_SCHED_THREADS 3 /* default # of kgnilnd_scheduler threads */
#define GNILND_FMABLK 1024 /* default number of mboxes per fmablk */
+#define GNILND_SCHED_NICE -20 /* default nice value for scheduler threads */
+#define GNILND_COMPUTE 0 /* service image */
+#define GNILND_FAST_RECONNECT 0 /* Fast Reconnect option */
#endif
/* EXTRA_BITS are there to allow us to hide NOOP/CLOSE and anything else out of band */
/* need sane upper bound to limit copy overhead */
#define GNILND_MAX_IMMEDIATE (64<<10)
+/* Max number of connections to keep in purgatory per peer */
+#define GNILND_PURGATORY_MAX 5
+/* Closing, don't put in purgatory */
+#define GNILND_NOPURG 222
+
/* payload size to add to the base mailbox size
* This is subtracting 2 from the concurrent_sends as 4 messages are included in the size
* gni_smsg_buff_size_needed calculates, the MAX_PAYLOAD is added to
#define GNILND_LASTRX(conn) (time_after(conn->gnc_last_rx, conn->gnc_last_rx_cq) \
? conn->gnc_last_rx : conn->gnc_last_rx_cq)
+/* fmablk registration failures timeout before failing node */
+#define GNILND_REGFAILTO_DISABLE -1
+
/************************************************************************
* Enum, flag and tag data
*/
#define GNILND_MSG_GET_NAK 0x08 /* gnm_u.completion (no GET match: src->sink) */
#define GNILND_MSG_GET_DONE 0x09 /* gnm_u.completion (src->sink) */
#define GNILND_MSG_CLOSE 0x0a /* empty gnm_u */
+#define GNILND_MSG_PUT_REQ_REV 0x0b /* gnm_u.get (src->sink) */
+#define GNILND_MSG_PUT_DONE_REV 0x0c /* gnm_u.completion (sink->src) */
+#define GNILND_MSG_PUT_NAK_REV 0x0d /* gnm_u.completion (no PUT match: sink->src) */
+#define GNILND_MSG_GET_REQ_REV 0x0e /* gnm_u.get (sink->src ) */
+#define GNILND_MSG_GET_ACK_REV 0x0f /* gnm_u.getack (GET matched: src->sink) */
+#define GNILND_MSG_GET_DONE_REV 0x10 /* gnm_u.completion (sink -> src) */
+#define GNILND_MSG_GET_NAK_REV 0x11 /* gnm_u.completeion (no GET match: sink -> src) */
/* defines for gnc_*scheduled states */
#define GNILND_CONN_IDLE 0
#define GNILND_DEL_PEER 1
#define GNILND_CLEAR_PURGATORY 2
+#define GNILND_PEER_UP 0
+#define GNILND_PEER_DOWN 1
+#define GNILND_PEER_TIMED_OUT 2
+#define GNILND_PEER_UNKNOWN 3
+
+/* defines for reverse RDMA states */
+#define GNILND_REVERSE_NONE 0
+#define GNILND_REVERSE_GET 1
+#define GNILND_REVERSE_PUT 2
+#define GNILND_REVERSE_BOTH (GNILND_REVERSE_GET | GNILND_REVERSE_PUT)
+
typedef enum kgn_fmablk_state {
GNILND_FMABLK_IDLE = 0, /* is allocated or ready to be freed */
GNILND_FMABLK_PHYS, /* allocated out of slab of physical memory */
} WIRE_ATTR kgn_rdma_desc_t;
typedef struct {
- lnet_hdr_t gnim_hdr; /* LNet header */
+ struct lnet_hdr gnim_hdr; /* LNet header */
/* LNet payload is in FMA "Message Data" */
} WIRE_ATTR kgn_immediate_msg_t;
typedef struct {
- lnet_hdr_t gnprm_hdr; /* LNet header */
+ struct lnet_hdr gnprm_hdr; /* LNet header */
__u64 gnprm_cookie; /* opaque completion cookie */
} WIRE_ATTR kgn_putreq_msg_t;
typedef struct {
__u64 gnpam_src_cookie; /* reflected completion cookie */
__u64 gnpam_dst_cookie; /* opaque completion cookie */
+ __u16 gnpam_payload_cksum; /* checksum for get msg */
kgn_rdma_desc_t gnpam_desc; /* sender's sink buffer */
} WIRE_ATTR kgn_putack_msg_t;
typedef struct {
- lnet_hdr_t gngm_hdr; /* LNet header */
+ struct lnet_hdr gngm_hdr; /* LNet header */
__u64 gngm_cookie; /* opaque completion cookie */
+ __u16 gngm_payload_cksum; /* checksum for put msg */
kgn_rdma_desc_t gngm_desc; /* sender's sink buffer */
} WIRE_ATTR kgn_get_msg_t;
int *kgn_max_immediate; /* immediate payload breakpoint */
int *kgn_checksum; /* checksum data */
int *kgn_checksum_dump; /* dump raw data to D_INFO log when checksumming */
- int *kgn_bte_hash; /* hashing on BTE transfers */
- int *kgn_bte_adapt; /* adaptive routing on BTE transfers */
+ int *kgn_bte_put_dlvr_mode; /* BTE Put delivery mode */
+ int *kgn_bte_get_dlvr_mode; /* BTE Get delivery mode */
int *kgn_bte_relaxed_ordering; /* relaxed ordering (PASSPW) on BTE transfers */
int *kgn_ptag; /* PTAG for cdm_create */
- int *kgn_max_retransmits; /* max number of FMA retransmits */
+ int *kgn_pkey; /* PKEY for cdm_create */
+ int *kgn_max_retransmits; /* max number of FMA retransmits before entering delay list */
int *kgn_nwildcard; /* # wildcard per net to post */
int *kgn_nice; /* nice value for kgnilnd threads */
int *kgn_rdmaq_intervals; /* # intervals per second for rdmaq throttle */
int *kgn_loops; /* # of loops sched does before flush/heartbeat tickle */
int *kgn_peer_hash_size; /* size of kgn_peers */
int *kgn_peer_health; /* enable/disable peer health */
+ int *kgn_peer_timeout; /* Override of the default peer_timeout used by peer_health */
int *kgn_vmap_cksum; /* enable/disable vmap of kiov checksums */
int *kgn_mbox_per_block; /* mailboxes per fmablk */
int *kgn_nphys_mbox; /* # mailboxes to preallocate with physical memory */
int *kgn_net_hash_size; /* size of kgn_net_ht */
int *kgn_hardware_timeout; /* max time for a message to get across the network */
int *kgn_mdd_timeout; /* max time for ghal to hold an mdd in minutes */
-#if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
- cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
-#endif
+ int *kgn_sched_timeout; /* max time for scheduler to run before yielding */
+ int *kgn_dgram_timeout; /* max time for dgram mover to run before scheduling */
+ int *kgn_sched_nice; /* nice value for kgnilnd scheduler threads */
+ int *kgn_reverse_rdma; /* Reverse RDMA setting */
+ int *kgn_eager_credits; /* allocated eager buffers */
+ int *kgn_fast_reconn; /* fast reconnection on conn timeout */
+ int *kgn_efault_lbug; /* LBUG on receiving an EFAULT */
+ int *kgn_max_purgatory; /* # conns/peer to keep in purgatory */
+ int *kgn_reg_fail_timeout; /* registration failure timeout */
+ int *kgn_thread_affinity; /* bind scheduler threads to cpus */
+ int *kgn_to_reconn_disable;/* disable reconnect after timeout */
+ int *kgn_thread_safe; /* use thread safe kgni API */
} kgn_tunables_t;
typedef struct kgn_mbox_info {
lnet_nid_t mbx_prev_nid;
+ lnet_nid_t mbx_prev_purg_nid;
unsigned long mbx_create_conn_memset;
unsigned long mbx_add_purgatory;
unsigned long mbx_detach_of_purgatory;
unsigned long mbx_release_from_purgatory;
unsigned long mbx_release_purg_active_dgram;
+ int mbx_nallocs;
+ int mbx_nallocs_total;
} kgn_mbox_info_t;
typedef struct kgn_fma_memblock {
int gnd_id; /* device id, also index in kgn_devices */
__u32 gnd_nid; /* ph host ID translated to NID */
struct list_head gnd_fma_buffs; /* list of FMA memory blocks */
- struct semaphore gnd_fmablk_sem; /* semaphore for FMA block memory alloc/free */
+ struct mutex gnd_fmablk_mutex; /* mutex for FMA block memory alloc/free */
spinlock_t gnd_fmablk_lock; /* lock for mbox alloc/release */
atomic_t gnd_nfmablk; /* # of fmablk live */
atomic_t gnd_fmablk_vers; /* gnd_fma_bufs stamp */
atomic_t gnd_neps; /* # EP allocated to conns */
short gnd_ready; /* stuff to do in scheduler thread */
struct list_head gnd_ready_conns; /* connections ready to tx/rx */
+ struct list_head gnd_delay_conns; /* connections in need of dla/or smsg credits */
struct list_head gnd_map_tx; /* TX: needing buffer mapping */
wait_queue_head_t gnd_waitq; /* scheduler wakeup */
spinlock_t gnd_lock; /* serialise gnd_ready_conns */
int gnd_dgram_ready; /* dgrams need movin' */
struct list_head *gnd_dgrams; /* nid hash to dgrams */
atomic_t gnd_ndgrams; /* # dgrams extant */
+ atomic_t gnd_nwcdgrams; /* # wildcard dgrams to post*/
spinlock_t gnd_dgram_lock; /* serialize gnd_dgrams */
struct list_head gnd_map_list; /* list of all mapped regions */
int gnd_map_version; /* version flag for map list */
+ struct timer_list gnd_map_timer; /* wakey-wakey */
atomic_t gnd_n_mdd; /* number of total MDD - fma, tx, etc */
atomic_t gnd_n_mdd_held; /* number of total MDD held - fma, tx, etc */
atomic_t gnd_nq_map; /* # queued waiting for mapping (MDD/GART) */
__u32 gnd_map_nvirt; /* # TX virt mappings */
__u64 gnd_map_virtnob; /* # TX virt bytes mapped */
spinlock_t gnd_map_lock; /* serialize gnd_map_XXX */
+ unsigned long gnd_next_map; /* next mapping attempt in jiffies */
+ int gnd_map_attempt; /* last map attempt # */
+ unsigned long gnd_last_map; /* map timeout base */
struct list_head gnd_rdmaq; /* RDMA to be sent */
spinlock_t gnd_rdmaq_lock; /* play nice with others */
atomic64_t gnd_rdmaq_bytes_out; /* # bytes authorized */
atomic_t gnd_n_yield;
atomic_t gnd_n_schedule;
atomic_t gnd_canceled_dgrams; /* # of outstanding cancels */
+ struct rw_semaphore gnd_conn_sem; /* serialize connection changes/data movement */
+ void *gnd_smdd_hold_buf; /* buffer to keep smdd */
+ gni_mem_handle_t gnd_smdd_hold_hndl; /* buffer mem handle */
} kgn_device_t;
typedef struct kgn_net {
int tx_buftype; /* payload buffer type */
int tx_phys_npages; /* # physical pages */
gni_mem_handle_t tx_map_key; /* mapping key */
+ gni_mem_handle_t tx_buffer_copy_map_key; /* mapping key for page aligned copy */
gni_mem_segment_t *tx_phys; /* page descriptors */
kgn_msg_t tx_msg; /* FMA message buffer */
kgn_tx_ev_id_t tx_id; /* who are you, who ? who ? */
int tx_retrans; /* retrans count of RDMA */
int tx_rc; /* if we need to stash the ret code until we see completion */
void *tx_buffer; /* source/sink buffer */
+ void *tx_buffer_copy; /* pointer to page aligned buffer */
+ unsigned int tx_nob_rdma; /* nob actually rdma */
+ unsigned int tx_offset; /* offset of data into copied buffer */
union {
gni_post_descriptor_t tx_rdma_desc; /* rdma descriptor */
struct page *tx_imm_pages[GNILND_MAX_IMMEDIATE/PAGE_SIZE]; /* page array to map kiov for immediate send */
typedef struct kgn_conn {
kgn_device_t *gnc_device; /* which device */
struct kgn_peer *gnc_peer; /* owning peer */
+ int gnc_magic; /* magic value cleared before free */
struct list_head gnc_list; /* stash on peer's conn list - or pending purgatory lists as we clear them */
struct list_head gnc_hashlist; /* stash in connection hash table */
struct list_head gnc_schedlist; /* schedule (on gnd_?_conns) for attention */
struct list_head gnc_fmaq; /* txs queued for FMA */
struct list_head gnc_mdd_list; /* hold list for MDD on hard conn reset */
+ struct list_head gnc_delaylist; /* If on this list schedule anytime we get interrupted */
__u64 gnc_peerstamp; /* peer's unique stamp */
__u64 gnc_peer_connstamp; /* peer's unique connection stamp */
__u64 gnc_my_connstamp; /* my unique connection stamp */
atomic_t gnc_sched_noop; /* # sched triggered NOOP */
unsigned int gnc_timeout; /* infer peer death if no rx for this many seconds */
__u32 gnc_cqid; /* my completion callback id (non-unique) */
- __u32 gnc_tx_seq; /* tx msg sequence number */
- __u32 gnc_rx_seq; /* rx msg sequence number */
+ atomic_t gnc_tx_seq; /* tx msg sequence number */
+ atomic_t gnc_rx_seq; /* rx msg sequence number */
+ struct mutex gnc_smsg_mutex; /* tx smsg sequence serialization */
+ struct mutex gnc_rdma_mutex; /* tx rdma sequence serialization */
__u64 gnc_tx_retrans; /* # retrans on SMSG */
atomic_t gnc_nlive_fma; /* # live FMA */
atomic_t gnc_nq_rdma; /* # queued (on device) RDMA */
int gnc_peer_error; /* errno peer sent us on CLOSE */
kgn_conn_state_t gnc_state; /* connection state */
int gnc_scheduled; /* being attented to */
+ char gnc_sched_caller[30]; /* what function last called schedule */
+ int gnc_sched_line; /* what line # last called schedule */
atomic_t gnc_refcount; /* # users */
spinlock_t gnc_list_lock; /* serialise tx lists, max_rx_age */
gni_ep_handle_t gnc_ephandle; /* GNI endpoint */
int gnc_mbox_id; /* id of mbox in fma_blk */
short gnc_needs_detach; /* flag set in detach_purgatory_all_locked so reaper will clear out purgatory */
short gnc_needs_closing; /* flag set in del_conns when called from kgnilnd_del_peer_or_conn */
+ atomic_t gnc_tx_in_use; /* # of tx's currently in use by another thread use kgnilnd_peer_conn_lock */
+ kgn_dgram_type_t gnc_dgram_type; /* save dgram type used to establish this conn */
+ void *remote_mbox_addr; /* save remote mbox address */
} kgn_conn_t;
typedef struct kgn_mdd_purgatory {
unsigned long gnp_last_alive; /* last time I had valid comms */
int gnp_last_dgram_errno; /* last error dgrams saw */
unsigned long gnp_last_dgram_time; /* last time I tried to connect */
- unsigned long gnp_reconnect_time; /* CURRENT_SECONDS when reconnect OK */
+ unsigned long gnp_reconnect_time; /* get_seconds() when reconnect OK */
unsigned long gnp_reconnect_interval; /* exponential backoff */
atomic_t gnp_dirty_eps; /* # of old but yet to be destroyed EPs from conns */
+ int gnp_state; /* up/down/timedout */
+ unsigned long gnp_down_event_time; /* time peer down */
+ unsigned long gnp_up_event_time; /* time peer back up */
} kgn_peer_t;
/* the kgn_rx_t is a struct for handing to LNET as the private pointer for things
int kgn_nresets; /* number of stack resets */
int kgn_in_reset; /* are we in stack reset ? */
+ __u64 kgn_nid_trans_private;/* private data for each of the HW nid2nic arenas */
+
kgn_device_t kgn_devices[GNILND_MAXDEVS]; /* device/ptag/cq etc */
int kgn_ndevs; /* # devices */
wait_queue_head_t kgn_ruhroh_waitq; /* ruhroh thread wakeup */
int kgn_quiesce_trigger; /* should we quiesce ? */
atomic_t kgn_nquiesce; /* how many quiesced ? */
- struct semaphore kgn_quiesce_sem; /* serialize ruhroh task, startup and shutdown */
+ struct mutex kgn_quiesce_mutex; /* serialize ruhroh task, startup and shutdown */
int kgn_needs_reset; /* we need stack reset */
/* These next three members implement communication from gnilnd into
struct list_head *kgn_conns; /* conns hashed by cqid */
atomic_t kgn_nconns; /* # connections extant */
+ atomic_t kgn_neager_allocs; /* # of eager allocations */
__u64 kgn_peerstamp; /* when I started up */
__u64 kgn_connstamp; /* conn stamp generator */
int kgn_conn_version; /* version flag for conn tables */
wait_queue_head_t kgn_reaper_waitq; /* reaper sleeps here */
spinlock_t kgn_reaper_lock; /* serialise */
- cfs_mem_cache_t *kgn_rx_cache; /* rx descriptor space */
- cfs_mem_cache_t *kgn_tx_cache; /* tx descriptor memory */
- cfs_mem_cache_t *kgn_tx_phys_cache; /* tx phys descriptor memory */
+ struct kmem_cache *kgn_rx_cache; /* rx descriptor space */
+ struct kmem_cache *kgn_tx_cache; /* tx descriptor memory */
+ struct kmem_cache *kgn_tx_phys_cache; /* tx phys descriptor memory */
atomic_t kgn_ntx; /* # tx in use */
- cfs_mem_cache_t *kgn_dgram_cache; /* outgoing datagrams */
+ struct kmem_cache *kgn_dgram_cache; /* outgoing datagrams */
struct page ***kgn_cksum_map_pages; /* page arrays for mapping pages on checksum */
- __u64 kgn_cksum_npages; /* Number of pages allocated for checksumming */
+ __u64 kgn_cksum_npages; /* # pages alloc'd for checksumming */
atomic_t kgn_nvmap_cksum; /* # times we vmapped for checksums */
atomic_t kgn_nvmap_short; /* # times we vmapped for short kiov */
atomic_t kgn_npending_unlink; /* # of peers pending unlink */
atomic_t kgn_npending_conns; /* # of conns with pending closes */
atomic_t kgn_npending_detach; /* # of conns with a pending detach */
-
+ unsigned long kgn_last_scheduled; /* last time schedule was called */
+ unsigned long kgn_last_condresched; /* last time cond_resched was called */
+ atomic_t kgn_rev_offset; /* # of REV rdma w/misaligned offsets */
+ atomic_t kgn_rev_length; /* # of REV rdma have misaligned len */
+ atomic_t kgn_rev_copy_buff; /* # of REV rdma buffer copies */
+ struct socket *kgn_sock; /* for Apollo */
+ unsigned long free_pages_limit; /* # of free pages reserve from fma block allocations */
+ int kgn_enable_gl_mutex; /* kgni api mtx enable */
} kgn_data_t;
extern kgn_data_t kgnilnd_data;
extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
-extern void kgnilnd_schedule_conn(kgn_conn_t *conn);
+extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held);
+extern int _kgnilnd_schedule_delay_conn(kgn_conn_t *conn);
-static inline int
-kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id)
-{
- struct task_struct *thrd = kthread_run(fn, arg, "%s_%02d", name, id);
- if (IS_ERR(thrd))
- return PTR_ERR(thrd);
+/* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
+ * and the line of the calling function to allow us to debug problematic
+ * schedule calls in the future without the programmer having to mark
+ * the location manually.
+ */
+#define kgnilnd_schedule_conn(conn) \
+ _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 0);
- atomic_inc(&kgnilnd_data.kgn_nthreads);
- return 0;
-}
+#define kgnilnd_schedule_conn_refheld(conn, refheld) \
+ _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld, 0);
+
+#define kgnilnd_schedule_conn_nolock(conn) \
+ _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 1);
+
+
+/* Macro wrapper for _kgnilnd_schedule_delay_conn. This will allow us to store
+ * extra data if we need to.
+ */
+#define kgnilnd_schedule_delay_conn(conn) \
+ _kgnilnd_schedule_delay_conn(conn);
static inline void
kgnilnd_thread_fini(void)
atomic_dec(&kgnilnd_data.kgn_nthreads);
}
+static inline void kgnilnd_gl_mutex_lock(struct mutex *lock)
+{
+ if (kgnilnd_data.kgn_enable_gl_mutex)
+ mutex_lock(lock);
+}
+
+static inline void kgnilnd_gl_mutex_unlock(struct mutex *lock)
+{
+ if (kgnilnd_data.kgn_enable_gl_mutex)
+ mutex_unlock(lock);
+}
+
+static inline void kgnilnd_conn_mutex_lock(struct mutex *lock)
+{
+ if (!kgnilnd_data.kgn_enable_gl_mutex)
+ mutex_lock(lock);
+}
+
+static inline void kgnilnd_conn_mutex_unlock(struct mutex *lock)
+{
+ if (!kgnilnd_data.kgn_enable_gl_mutex)
+ mutex_unlock(lock);
+}
+
/* like mutex_trylock but with a jiffies spinner. This is to allow certain
* parts of the code to avoid a scheduler trip when the mutex is held
*
* This function must not be used in interrupt context. The
* mutex must be released by the same task that acquired it.
*/
-static inline int kgnilnd_mutex_trylock(struct mutex *lock)
+static inline int __kgnilnd_mutex_trylock(struct mutex *lock)
{
int ret;
unsigned long timeout;
return 0;
}
+static inline int kgnilnd_mutex_trylock(struct mutex *lock)
+{
+ if (!kgnilnd_data.kgn_enable_gl_mutex)
+ return 1;
+
+ return __kgnilnd_mutex_trylock(lock);
+}
+
+static inline int kgnilnd_trylock(struct mutex *cq_lock,
+ struct mutex *c_lock)
+{
+ if (kgnilnd_data.kgn_enable_gl_mutex)
+ return __kgnilnd_mutex_trylock(cq_lock);
+ else
+ return __kgnilnd_mutex_trylock(c_lock);
+}
+
+static inline void *kgnilnd_vzalloc(int size)
+{
+ void *ret = __vmalloc(size, __GFP_HIGHMEM | GFP_NOIO | __GFP_NORETRY | __GFP_ZERO,
+ PAGE_KERNEL);
+ LIBCFS_ALLOC_POST(ret, size);
+ return ret;
+}
+
+static inline void kgnilnd_vfree(void *ptr, int size)
+{
+ libcfs_kmem_dec(ptr, size);
+ vfree(ptr);
+}
+
/* Copied from DEBUG_REQ in Lustre - the dance is needed to save stack space */
extern void
#define GNIDBG_MSG(level, msg, fmt, args...) \
do { \
if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
- static cfs_debug_limit_state_t cdls; \
+ static struct cfs_debug_limit_state cdls; \
LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
"$$ "fmt" from %s ", ## args, \
#define GNIDBG_TOMSG(level, msg, fmt, args...) \
do { \
if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
- static cfs_debug_limit_state_t cdls; \
+ static struct cfs_debug_limit_state cdls; \
LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
kgnilnd_debug_msg(&msgdata, level, &cdls, msg, \
"$$ "fmt" ", ## args); \
#define GNIDBG_CONN(level, conn, fmt, args...) \
do { \
if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
- static cfs_debug_limit_state_t cdls; \
+ static struct cfs_debug_limit_state cdls; \
LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
kgnilnd_debug_conn(&msgdata, level, &cdls, conn, \
"$$ "fmt" ", ## args); \
#define GNIDBG_TX(level, tx, fmt, args...) \
do { \
if ((level) & (D_ERROR | D_WARNING | D_NETERROR)) { \
- static cfs_debug_limit_state_t cdls; \
+ static struct cfs_debug_limit_state cdls; \
LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, level, &cdls); \
kgnilnd_debug_tx(&msgdata, level, &cdls, tx, \
"$$ "fmt" ", ## args); \
(atomic_read(&kgnilnd_data.kgn_nquiesce) == \
atomic_read(&kgnilnd_data.kgn_nthreads))
-#define KGNILND_SPIN_QUIESCE \
-do { \
- /* E.T phone home */ \
- atomic_inc(&kgnilnd_data.kgn_nquiesce); \
- CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
- while (kgnilnd_data.kgn_quiesce_trigger) { \
- set_current_state(TASK_INTERRUPTIBLE); \
- cfs_schedule_timeout_and_set_state(TASK_INTERRUPTIBLE, \
- cfs_time_seconds(1)); \
- } \
- /* Mom, my homework is done */ \
- CDEBUG(D_NET, "Waking up from thread pause\n"); \
- atomic_dec(&kgnilnd_data.kgn_nquiesce); \
+#define KGNILND_SPIN_QUIESCE \
+do { \
+ /* E.T phone home */ \
+ atomic_inc(&kgnilnd_data.kgn_nquiesce); \
+ CDEBUG(D_NET, "Waiting for thread pause to be over...\n"); \
+ while (kgnilnd_data.kgn_quiesce_trigger) { \
+ msleep_interruptible(MSEC_PER_SEC); \
+ } \
+ /* Mom, my homework is done */ \
+ CDEBUG(D_NET, "Waking up from thread pause\n"); \
+ atomic_dec(&kgnilnd_data.kgn_nquiesce); \
} while(0)
/* use macros for addref/decref to get the calling function name in the CDEBUG */
LASSERTF(val >= 0, "peer %p refcount %d\n", peer, val); \
CDEBUG(D_NETTRACE, "peer %p->%s--(%d)\n", peer, \
libcfs_nid2str(peer->gnp_nid), val); \
- if (atomic_read(&peer->gnp_refcount) == 0) \
+ if (val == 0) \
kgnilnd_destroy_peer(peer); \
} while(0)
\
smp_wmb(); \
val = atomic_inc_return(&conn->gnc_refcount); \
- LASSERTF(val >= 0, "conn %p refc %d to %s\n", \
+ LASSERTF(val > 1 && conn->gnc_magic == GNILND_CONN_MAGIC, \
+ "conn %p refc %d to %s\n", \
conn, val, \
conn->gnc_peer \
? libcfs_nid2str(conn->gnc_peer->gnp_nid) \
: "<?>", \
val); \
smp_rmb(); \
- if ((atomic_read(&conn->gnc_refcount) == 1) && \
+ if ((val == 1) && \
(conn->gnc_ephandle != NULL) && \
(conn->gnc_state != GNILND_CONN_DESTROY_EP)) { \
set_mb(conn->gnc_state, GNILND_CONN_DESTROY_EP); \
kgnilnd_schedule_conn(conn); \
- } else if (atomic_read(&conn->gnc_refcount) == 0) { \
+ } else if (val == 0) { \
kgnilnd_destroy_conn(conn); \
} \
} while (0)
{
/* - ESHUTDOWN - LND is unloading
* - EUCLEAN - admin requested via "lctl del_peer"
- * - ENETRESET - admin requested via "lctl disconnect"
+ * - ENETRESET - admin requested via "lctl disconnect" or rca event
* - ENOTRECOVERABLE - stack reset
* - EISCONN - cleared via "lctl push"
* not doing ESTALE - that isn't clean */
}
/* just insurance */
kgnilnd_conn_addref(conn);
+ kgnilnd_admin_addref(conn->gnc_tx_in_use);
read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
/* we know this is safe - as the TX won't be reused until AFTER
* lctl disconnect or del_peer. */
if (tx == NULL) {
CNETERR("txe_idx %d is gone, ignoring event\n", ev_id->txe_idx);
+ kgnilnd_admin_decref(conn->gnc_tx_in_use);
kgnilnd_conn_decref(conn);
return;
}
int kgnilnd_allocate_phys_fmablk(kgn_device_t *device);
int kgnilnd_map_phys_fmablk(kgn_device_t *device);
-void kgnilnd_unmap_phys_fmablk(kgn_device_t *device);
+void kgnilnd_unmap_fma_blocks(kgn_device_t *device);
void kgnilnd_free_phys_fmablk(kgn_device_t *device);
int kgnilnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
lnet_msg_t *lntmsg, void **new_private);
int kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
int delayed, unsigned int niov,
- struct iovec *iov, lnet_kiov_t *kiov,
+ struct kvec *iov, lnet_kiov_t *kiov,
unsigned int offset, unsigned int mlen, unsigned int rlen);
__u16 kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int offset, unsigned int nob, int dump_blob);
void kgnilnd_tx_done(kgn_tx_t *tx, int completion);
void kgnilnd_txlist_done(struct list_head *txlist, int error);
void kgnilnd_unlink_peer_locked(kgn_peer_t *peer);
-void kgnilnd_schedule_conn(kgn_conn_t *conn);
-void kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
+int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld, int lock_held);
+int kgnilnd_schedule_process_conn(kgn_conn_t *conn, int sched_intent);
void kgnilnd_schedule_dgram(kgn_device_t *dev);
-int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net);
+int kgnilnd_create_peer_safe(kgn_peer_t **peerp, lnet_nid_t nid, kgn_net_t *net, int node_state);
void kgnilnd_add_peer_locked(lnet_nid_t nid, kgn_peer_t *new_stub_peer, kgn_peer_t **peerp);
int kgnilnd_add_peer(kgn_net_t *net, lnet_nid_t nid, kgn_peer_t **peerp);
int kgnilnd_reaper(void *arg);
int kgnilnd_scheduler(void *arg);
int kgnilnd_dgram_mover(void *arg);
+int kgnilnd_rca(void *arg);
+int kgnilnd_thread_start(int(*fn)(void *arg), void *arg, char *name, int id);
int kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev);
int kgnilnd_conn_isdup_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
void kgnilnd_cancel_peer_connect_locked(kgn_peer_t *peer, struct list_head *zombies);
int kgnilnd_close_stale_conns_locked(kgn_peer_t *peer, kgn_conn_t *newconn);
void kgnilnd_peer_alive(kgn_peer_t *peer);
-void kgnilnd_peer_notify(kgn_peer_t *peer, int error);
+void kgnilnd_peer_notify(kgn_peer_t *peer, int error, int alive);
void kgnilnd_close_conn_locked(kgn_conn_t *conn, int error);
void kgnilnd_close_conn(kgn_conn_t *conn, int error);
void kgnilnd_complete_closed_conn(kgn_conn_t *conn);
void kgnilnd_destroy_conn_ep(kgn_conn_t *conn);
int kgnilnd_close_peer_conns_locked(kgn_peer_t *peer, int why);
+int kgnilnd_report_node_state(lnet_nid_t nid, int down);
+void kgnilnd_wakeup_rca_thread(void);
+int kgnilnd_start_rca_thread(void);
+int kgnilnd_get_node_state(__u32 nid);
int kgnilnd_tunables_init(void);
-void kgnilnd_tunables_fini(void);
void kgnilnd_init_msg(kgn_msg_t *msg, int type, lnet_nid_t source);
void kgnilnd_bump_timeouts(__u32 nap_time, char *reason);
int kgnilnd_find_and_cancel_dgram(kgn_device_t *dev, lnet_nid_t dst_nid);
void kgnilnd_cancel_dgram_locked(kgn_dgram_t *dgram);
-void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram);
+void kgnilnd_release_dgram(kgn_device_t *dev, kgn_dgram_t *dgram, int shutdown);
int kgnilnd_setup_wildcard_dgram(kgn_device_t *dev);
int kgnilnd_cancel_net_dgrams(kgn_net_t *net);
int kgnilnd_cancel_wc_dgrams(kgn_device_t *dev);
+int kgnilnd_cancel_dgrams(kgn_device_t *dev);
void kgnilnd_wait_for_canceled_dgrams(kgn_device_t *dev);
int kgnilnd_dgram_waitq(void *arg);
DO_TYPE(GNILND_MSG_GET_NAK);
DO_TYPE(GNILND_MSG_GET_DONE);
DO_TYPE(GNILND_MSG_CLOSE);
+ DO_TYPE(GNILND_MSG_PUT_REQ_REV);
+ DO_TYPE(GNILND_MSG_PUT_DONE_REV);
+ DO_TYPE(GNILND_MSG_PUT_NAK_REV);
+ DO_TYPE(GNILND_MSG_GET_REQ_REV);
+ DO_TYPE(GNILND_MSG_GET_ACK_REV);
+ DO_TYPE(GNILND_MSG_GET_DONE_REV);
+ DO_TYPE(GNILND_MSG_GET_NAK_REV);
}
return "<unknown msg type>";
}
return "<?type?>";
}
+static inline const char *
+kgnilnd_conn_dgram_type2str(kgn_dgram_type_t type)
+{
+ switch (type) {
+ DO_TYPE(GNILND_DGRAM_REQ);
+ DO_TYPE(GNILND_DGRAM_WC_REQ);
+ DO_TYPE(GNILND_DGRAM_NAK);
+ DO_TYPE(GNILND_DGRAM_CLOSE);
+ }
+ return "<?type?>";
+}
#undef DO_TYPE
+/* pulls in tunables per platform and adds in nid/nic conversion
+ * if RCA wasn't available at build time */
+#include "gnilnd_hss_ops.h"
/* API wrapper functions - include late to pick up all of the other defines */
#include "gnilnd_api_wrap.h"
+#if defined(CONFIG_CRAY_GEMINI)
+ #include "gnilnd_gemini.h"
+#elif defined(CONFIG_CRAY_ARIES)
+ #include "gnilnd_aries.h"
+#else
+ #error "Undefined Network Hardware Type"
+#endif
+
+extern uint32_t kgni_driver_version;
+
+static inline void
+kgnilnd_check_kgni_version(void)
+{
+ uint32_t *kdv;
+
+ kgnilnd_data.kgn_enable_gl_mutex = 1;
+ kdv = symbol_get(kgni_driver_version);
+ if (!kdv) {
+ LCONSOLE_INFO("Not using thread safe locking -"
+ " no symbol kgni_driver_version\n");
+ return;
+ }
+
+ /* Thread-safe kgni implemented in minor ver 0x44/45, code rev 0xb9 */
+ if (*kdv < GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9)) {
+ symbol_put(kgni_driver_version);
+ LCONSOLE_INFO("Not using thread safe locking, gni version 0x%x,"
+ " need >= 0x%x\n", *kdv,
+ GNI_VERSION_CHECK(0, GNILND_KGNI_TS_MINOR_VER, 0xb9));
+ return;
+ }
+
+ symbol_put(kgni_driver_version);
+
+ if (!*kgnilnd_tunables.kgn_thread_safe) {
+ return;
+ }
+
+ /* Use thread-safe locking */
+ kgnilnd_data.kgn_enable_gl_mutex = 0;
+}
+
#endif /* _GNILND_GNILND_H_ */