*
* Copyright (C) 2009-2012 Cray, Inc.
*
- * Copyright (c) 2013, 2014, Intel Corporation.
+ * Copyright (c) 2014, 2016, Intel Corporation.
*
* Derived from work by: Eric Barton <eric@bartonsoftware.com>
* Author: Nic Henke <nic@cray.com>
#ifndef _GNILND_GNILND_H_
#define _GNILND_GNILND_H_
-#ifdef HAVE_COMPAT_RDMA
-#include <linux/compat-2.6.h>
-#endif
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/mm.h>
#define DEBUG_SUBSYSTEM S_LND
-#include <libcfs/linux/kp30.h>
#include <libcfs/libcfs.h>
#include <lnet/lnet.h>
#include <lnet/lib-lnet.h>
/* 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
#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_DEL_PEER 1
#define GNILND_CLEAR_PURGATORY 2
-#define GNILND_RCA_NODE_UP 0
-#define GNILND_RCA_NODE_DOWN 1
-#define GNILND_RCA_NODE_UNKNOWN 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
} 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;
} 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 */
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_dlvr_mode; /* BTE delivery mode mask */
+ 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_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 */
-#if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
- cfs_sysctl_table_header_t *kgn_sysctl; /* sysctl interface */
-#endif
+ 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 {
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 */
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 {
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 */
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_down; /* rca says peer down */
+ 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;
extern void kgnilnd_destroy_peer(kgn_peer_t *peer);
extern void kgnilnd_destroy_conn(kgn_conn_t *conn);
-extern int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
+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);
/* Macro wrapper for _kgnilnd_schedule_conn. This will store the function
* and the line of the calling function to allow us to debug problematic
* the location manually.
*/
#define kgnilnd_schedule_conn(conn) \
- _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0);
+ _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 0);
-#define kgnilnd_schedule_conn_refheld(conn, refheld) \
- _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld);
+#define kgnilnd_schedule_conn_refheld(conn, refheld) \
+ _kgnilnd_schedule_conn(conn, __func__, __LINE__, refheld, 0);
-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);
+#define kgnilnd_schedule_conn_nolock(conn) \
+ _kgnilnd_schedule_conn(conn, __func__, __LINE__, 0, 1);
- atomic_inc(&kgnilnd_data.kgn_nthreads);
- return 0;
-}
+
+/* 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 int kgnilnd_gl_mutex_trylock(struct mutex *lock)
-{
- if (kgnilnd_data.kgn_enable_gl_mutex)
- return mutex_trylock(lock);
- else
- return 1;
-}
-
static inline void kgnilnd_gl_mutex_lock(struct mutex *lock)
{
if (kgnilnd_data.kgn_enable_gl_mutex)
* 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;
- if (!kgnilnd_data.kgn_enable_gl_mutex)
- return 1;
-
LASSERT(!in_interrupt());
for (timeout = jiffies + 1; time_before(jiffies, 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); \
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);
-int _kgnilnd_schedule_conn(kgn_conn_t *conn, const char *caller, int line, int refheld);
+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_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);
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);
/* 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"
#error "Undefined Network Hardware Type"
#endif
-/* API wrapper functions - include late to pick up all of the other defines */
-#include "gnilnd_api_wrap.h"
+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_ */