X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fportals%2Fknals%2Fsocknal%2Fsocknal.h;h=9cfe8589672efce3ba77566109fa1f207644355b;hb=3fed8a248a88129a08d2579a988249e648931582;hp=227a24ff6fb781341820faac491bb4e557d4f59c;hpb=c5050e412572b00cbe93d8517d2d1f767bebfa92;p=fs%2Flustre-release.git diff --git a/lustre/portals/knals/socknal/socknal.h b/lustre/portals/knals/socknal/socknal.h index 227a24f..9cfe858 100644 --- a/lustre/portals/knals/socknal/socknal.h +++ b/lustre/portals/knals/socknal/socknal.h @@ -44,7 +44,9 @@ #include #include +#include +#include #include #include #include @@ -55,18 +57,16 @@ #include #include -#define DEBUG_SUBSYSTEM S_SOCKNAL +#define DEBUG_SUBSYSTEM S_NAL #include +#include +#include #include #include +#include #include -#if CONFIG_SMP -# define SOCKNAL_N_SCHED num_online_cpus() /* # socknal schedulers */ -#else -# define SOCKNAL_N_SCHED 1 /* # socknal schedulers */ -#endif #define SOCKNAL_N_AUTOCONNECTD 4 /* # socknal autoconnect daemons */ #define SOCKNAL_MIN_RECONNECT_INTERVAL HZ /* first failed connection retry... */ @@ -78,26 +78,39 @@ #define SOCKNAL_TYPED_CONNS 1 /* unidirectional large, bidirectional small? */ #define SOCKNAL_ZC_MIN_FRAG (2<<10) /* default smallest zerocopy fragment */ #define SOCKNAL_MIN_BULK (1<<10) /* smallest "large" message */ -#define SOCKNAL_USE_KEEPALIVES 0 /* use tcp/ip keepalive? */ +#define SOCKNAL_BUFFER_SIZE (8<<20) /* default socket buffer size */ +#define SOCKNAL_NAGLE 0 /* enable/disable NAGLE? */ +#define SOCKNAL_IRQ_AFFINITY 1 /* enable/disable IRQ affinity? */ +#define SOCKNAL_KEEPALIVE_IDLE 0 /* # seconds idle before 1st probe */ +#define SOCKNAL_KEEPALIVE_COUNT 10 /* # unanswered probes to determine peer death */ +#define SOCKNAL_KEEPALIVE_INTVL 1 /* seconds between probes */ #define SOCKNAL_PEER_HASH_SIZE 101 /* # peer lists */ -#define SOCKNAL_NLTXS 128 /* # normal transmit messages */ -#define SOCKNAL_NNBLK_LTXS 128 /* # transmit messages reserved if can't block */ - #define SOCKNAL_SMALL_FWD_NMSGS 128 /* # small messages I can be forwarding at any time */ #define SOCKNAL_LARGE_FWD_NMSGS 64 /* # large messages I can be forwarding at any time */ #define SOCKNAL_SMALL_FWD_PAGES 1 /* # pages in a small message fwd buffer */ -#define SOCKNAL_LARGE_FWD_PAGES (PAGE_ALIGN (sizeof (ptl_hdr_t) + PTL_MTU) >> PAGE_SHIFT) +#define SOCKNAL_LARGE_FWD_PAGES (PAGE_ALIGN(PTL_MTU) >> PAGE_SHIFT) /* # pages in a large message fwd buffer */ #define SOCKNAL_RESCHED 100 /* # scheduler loops before reschedule */ +#define SOCKNAL_ENOMEM_RETRY 1 /* jiffies between retries */ + +#define SOCKNAL_MAX_INTERFACES 16 /* Largest number of interfaces we bind */ + +#define SOCKNAL_ROUND_ROBIN 0 /* round robin / load balance */ #define SOCKNAL_TX_LOW_WATER(sk) (((sk)->sk_sndbuf*8)/10) +#define SOCKNAL_SINGLE_FRAG_TX 0 /* disable multi-fragment sends */ +#define SOCKNAL_SINGLE_FRAG_RX 0 /* disable multi-fragment receives */ +#define SOCKNAL_RISK_KMAP_DEADLOCK 0 /* risk kmap deadlock on multi-frag I/O + * (backs off to single-frag if disabled) */ + #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,72)) +# define sk_allocation allocation # define sk_data_ready data_ready # define sk_write_space write_space # define sk_user_data user_data @@ -108,13 +121,16 @@ #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0)) # define sk_wmem_queued wmem_queued +# define sk_err err #endif typedef struct /* pool of forwarding buffers */ { spinlock_t fmp_lock; /* serialise */ - struct list_head fmp_idle_fmbs; /* buffers waiting for a connection */ + struct list_head fmp_idle_fmbs; /* free buffers */ struct list_head fmp_blocked_conns; /* connections waiting for a buffer */ + int fmp_nactive_fmbs; /* # buffers in use */ + int fmp_buff_pages; /* # pages per buffer */ } ksock_fmb_pool_t; @@ -130,56 +146,71 @@ typedef struct /* per scheduler state */ int kss_nconns; /* # connections assigned to this scheduler */ } ksock_sched_t; -typedef struct { +typedef struct +{ int ksni_valid:1; /* been set yet? */ int ksni_bound:1; /* bound to a cpu yet? */ int ksni_sched:6; /* which scheduler (assumes < 64) */ } ksock_irqinfo_t; -typedef struct { - int ksnd_init; /* initialisation state */ +typedef struct +{ + __u32 ksni_ipaddr; /* interface's IP address */ + __u32 ksni_netmask; /* interface's network mask */ + int ksni_nroutes; /* # routes using (active) */ + int ksni_npeers; /* # peers using (passive) */ +} ksock_interface_t; + +typedef struct +{ int ksnd_io_timeout; /* "stuck" socket timeout (seconds) */ int ksnd_eager_ack; /* make TCP ack eagerly? */ int ksnd_typed_conns; /* drive sockets by type? */ int ksnd_min_bulk; /* smallest "large" message */ + int ksnd_buffer_size; /* socket buffer size */ + int ksnd_nagle; /* enable NAGLE? */ + int ksnd_irq_affinity; /* enable IRQ affinity? */ + int ksnd_keepalive_idle; /* # idle secs before 1st probe */ + int ksnd_keepalive_count; /* # probes */ + int ksnd_keepalive_intvl; /* time between probes */ #if SOCKNAL_ZC unsigned int ksnd_zc_min_frag; /* minimum zero copy frag size */ #endif struct ctl_table_header *ksnd_sysctl; /* sysctl interface */ +} ksock_tunables_t; + +typedef struct +{ + int ksnd_init; /* initialisation state */ __u64 ksnd_incarnation; /* my epoch */ rwlock_t ksnd_global_lock; /* stabilize peer/conn ops */ struct list_head *ksnd_peers; /* hash table of all my known peers */ int ksnd_peer_hash_size; /* size of ksnd_peers */ - nal_cb_t *ksnd_nal_cb; - spinlock_t ksnd_nal_cb_lock; /* lib cli/sti lock */ - - atomic_t ksnd_nthreads; /* # live threads */ + int ksnd_nthreads; /* # live threads */ int ksnd_shuttingdown; /* tell threads to exit */ - ksock_sched_t *ksnd_schedulers; /* scheduler state */ + int ksnd_nschedulers; /* # schedulers */ + ksock_sched_t *ksnd_schedulers; /* their state */ atomic_t ksnd_npeers; /* total # peers extant */ atomic_t ksnd_nclosing_conns; /* # closed conns extant */ kpr_router_t ksnd_router; /* THE router */ - void *ksnd_fmbs; /* all the pre-allocated FMBs */ ksock_fmb_pool_t ksnd_small_fmp; /* small message forwarding buffers */ ksock_fmb_pool_t ksnd_large_fmp; /* large message forwarding buffers */ - void *ksnd_ltxs; /* all the pre-allocated LTXs */ - spinlock_t ksnd_idle_ltx_lock; /* serialise ltx alloc/free */ - struct list_head ksnd_idle_ltx_list; /* where to get an idle LTX */ - struct list_head ksnd_idle_nblk_ltx_list; /* where to get an idle LTX if you can't block */ - wait_queue_head_t ksnd_idle_ltx_waitq; /* where to block for an idle LTX */ - int ksnd_active_ltxs; /* #active ltxs */ + atomic_t ksnd_nactive_ltxs; /* #active ltxs */ struct list_head ksnd_deathrow_conns; /* conns to be closed */ struct list_head ksnd_zombie_conns; /* conns to be freed */ - wait_queue_head_t ksnd_reaper_waitq; /* reaper sleep here */ + struct list_head ksnd_enomem_conns; /* conns to be retried */ + wait_queue_head_t ksnd_reaper_waitq; /* reaper sleeps here */ + unsigned long ksnd_reaper_waketime; /* when reaper will wake */ spinlock_t ksnd_reaper_lock; /* serialise */ + int ksnd_enomem_tx; /* test ENOMEM sender */ int ksnd_stall_tx; /* test sluggish sender */ int ksnd_stall_rx; /* test sluggish receiver */ @@ -188,26 +219,24 @@ typedef struct { spinlock_t ksnd_autoconnectd_lock; /* serialise */ ksock_irqinfo_t ksnd_irqinfo[NR_IRQS];/* irq->scheduler lookup */ + + int ksnd_ninterfaces; + ksock_interface_t ksnd_interfaces[SOCKNAL_MAX_INTERFACES]; /* published interfaces */ } ksock_nal_data_t; #define SOCKNAL_INIT_NOTHING 0 #define SOCKNAL_INIT_DATA 1 -#define SOCKNAL_INIT_PTL 2 +#define SOCKNAL_INIT_LIB 2 #define SOCKNAL_INIT_ALL 3 /* A packet just assembled for transmission is represented by 1 or more - * struct iovec fragments and 0 or more ptl_kiov_t fragments. Forwarded - * messages, or messages from an MD with PTL_MD_KIOV _not_ set have 0 - * ptl_kiov_t fragments. Messages from an MD with PTL_MD_KIOV set, have 1 - * struct iovec fragment (the header) and up to PTL_MD_MAX_IOV ptl_kiov_t - * fragments. + * struct iovec fragments (the first frag contains the portals header), + * followed by 0 or more ptl_kiov_t fragments. * * On the receive side, initially 1 struct iovec fragment is posted for - * receive (the header). Once the header has been received, if the message - * requires forwarding or will be received into mapped memory, up to - * PTL_MD_MAX_IOV struct iovec fragments describe the target memory. - * Otherwise up to PTL_MD_MAX_IOV ptl_kiov_t fragments are used. - */ + * receive (the header). Once the header has been received, the payload is + * received into either struct iovec or ptl_kiov_t fragments, depending on + * what the header matched or whether the message needs forwarding. */ struct ksock_conn; /* forward ref */ struct ksock_peer; /* forward ref */ @@ -230,28 +259,24 @@ typedef struct /* transmit packet */ #endif } ksock_tx_t; +typedef struct /* forwarded packet */ +{ + ksock_tx_t ftx_tx; /* send info */ + struct iovec ftx_iov; /* hdr iovec */ +} ksock_ftx_t; + #define KSOCK_ZCCD_2_TX(ptr) list_entry (ptr, ksock_tx_t, tx_zccd) /* network zero copy callback descriptor embedded in ksock_tx_t */ -/* space for the tx frag descriptors: hdr is always 1 iovec - * and payload is PTL_MD_MAX of either type. */ -typedef struct -{ - struct iovec hdr; - union { - struct iovec iov[PTL_MD_MAX_IOV]; - ptl_kiov_t kiov[PTL_MD_MAX_IOV]; - } payload; -} ksock_txiovspace_t; - typedef struct /* locally transmitted packet */ { ksock_tx_t ltx_tx; /* send info */ - struct list_head *ltx_idle; /* where to put when idle */ void *ltx_private; /* lib_finalize() callback arg */ void *ltx_cookie; /* lib_finalize() callback arg */ - ksock_txiovspace_t ltx_iov_space; /* where to stash frag descriptors */ ptl_hdr_t ltx_hdr; /* buffer for packet header */ + int ltx_desc_size; /* bytes allocated for this desc */ + struct iovec ltx_iov[1]; /* iov for hdr + payload */ + ptl_kiov_t ltx_kiov[0]; /* kiov for payload */ } ksock_ltx_t; #define KSOCK_TX_2_KPR_FWD_DESC(ptr) list_entry ((kprfd_scratch_t *)ptr, kpr_fwd_desc_t, kprfd_scratch) @@ -267,15 +292,14 @@ typedef struct /* Kernel portals Socket Forward { /* (socknal->router) */ struct list_head fmb_list; /* queue idle */ kpr_fwd_desc_t fmb_fwd; /* router's descriptor */ - int fmb_npages; /* # pages allocated */ ksock_fmb_pool_t *fmb_pool; /* owning pool */ struct ksock_peer *fmb_peer; /* peer received from */ - struct page *fmb_pages[SOCKNAL_LARGE_FWD_PAGES]; - struct iovec fmb_iov[SOCKNAL_LARGE_FWD_PAGES]; + ptl_hdr_t fmb_hdr; /* message header */ + ptl_kiov_t fmb_kiov[0]; /* payload frags */ } ksock_fmb_t; /* space for the rx frag descriptors; we either read a single contiguous - * header, or PTL_MD_MAX_IOV frags of payload of either type. */ + * header, or up to PTL_MD_MAX_IOV frags of payload of either type. */ typedef union { struct iovec iov[PTL_MD_MAX_IOV]; ptl_kiov_t kiov[PTL_MD_MAX_IOV]; @@ -298,6 +322,7 @@ typedef struct ksock_conn void *ksnc_saved_write_space; /* socket's original write_space() callback */ atomic_t ksnc_refcount; /* # users */ ksock_sched_t *ksnc_scheduler; /* who schedules this connection */ + __u32 ksnc_myipaddr; /* my IP */ __u32 ksnc_ipaddr; /* peer's IP */ int ksnc_port; /* peer's port */ int ksnc_closing; /* being shut down */ @@ -325,9 +350,17 @@ typedef struct ksock_conn struct list_head ksnc_tx_list; /* where I enq waiting for output space */ struct list_head ksnc_tx_queue; /* packets waiting to be sent */ unsigned long ksnc_tx_deadline; /* when (in jiffies) tx times out */ + int ksnc_tx_bufnob; /* send buffer marker */ atomic_t ksnc_tx_nob; /* # bytes queued */ int ksnc_tx_ready; /* write space */ int ksnc_tx_scheduled; /* being progressed */ + +#if !SOCKNAL_SINGLE_FRAG_RX + struct iovec ksnc_rx_scratch_iov[PTL_MD_MAX_IOV]; +#endif +#if !SOCKNAL_SINGLE_FRAG_TX + struct iovec ksnc_tx_scratch_iov[PTL_MD_MAX_IOV]; +#endif } ksock_conn_t; #define KSNR_TYPED_ROUTES ((1 << SOCKNAL_CONN_CONTROL) | \ @@ -340,18 +373,15 @@ typedef struct ksock_route struct list_head ksnr_connect_list; /* chain on autoconnect list */ struct ksock_peer *ksnr_peer; /* owning peer */ atomic_t ksnr_refcount; /* # users */ - int ksnr_sharecount; /* lconf usage counter */ unsigned long ksnr_timeout; /* when (in jiffies) reconnection can happen next */ unsigned int ksnr_retry_interval; /* how long between retries */ - __u32 ksnr_ipaddr; /* an IP address for this peer */ + __u32 ksnr_myipaddr; /* my IP */ + __u32 ksnr_ipaddr; /* IP address to connect to */ int ksnr_port; /* port to connect to */ - int ksnr_buffer_size; /* size of socket buffers */ - unsigned int ksnr_irq_affinity:1; /* set affinity? */ - unsigned int ksnr_nonagel:1; /* disable nagle? */ - unsigned int ksnr_eager:1; /* connect eagery? */ unsigned int ksnr_connecting:4; /* autoconnects in progress by type */ unsigned int ksnr_connected:4; /* connections established by type */ unsigned int ksnr_deleted:1; /* been removed from peer? */ + unsigned int ksnr_share_count; /* created explicitly? */ int ksnr_conn_count; /* # conns established by this route */ } ksock_route_t; @@ -360,31 +390,35 @@ typedef struct ksock_peer struct list_head ksnp_list; /* stash on global peer list */ ptl_nid_t ksnp_nid; /* who's on the other end(s) */ atomic_t ksnp_refcount; /* # users */ + int ksnp_sharecount; /* lconf usage counter */ int ksnp_closing; /* being closed */ int ksnp_error; /* errno on closing last conn */ struct list_head ksnp_conns; /* all active connections */ struct list_head ksnp_routes; /* routes */ struct list_head ksnp_tx_queue; /* waiting packets */ unsigned long ksnp_last_alive; /* when (in jiffies) I was last alive */ + int ksnp_n_passive_ips; /* # of... */ + __u32 ksnp_passive_ips[SOCKNAL_MAX_INTERFACES]; /* preferred local interfaces */ } ksock_peer_t; -extern nal_cb_t ksocknal_lib; +extern lib_nal_t ksocknal_lib; extern ksock_nal_data_t ksocknal_data; +extern ksock_tunables_t ksocknal_tunables; static inline struct list_head * -ksocknal_nid2peerlist (ptl_nid_t nid) +ksocknal_nid2peerlist (ptl_nid_t nid) { unsigned int hash = ((unsigned int)nid) % ksocknal_data.ksnd_peer_hash_size; - + return (&ksocknal_data.ksnd_peers [hash]); } static inline int -ksocknal_getconnsock (ksock_conn_t *conn) +ksocknal_getconnsock (ksock_conn_t *conn) { int rc = -ESHUTDOWN; - + read_lock (&ksocknal_data.ksnd_global_lock); if (!conn->ksnc_closing) { rc = 0; @@ -401,6 +435,61 @@ ksocknal_putconnsock (ksock_conn_t *conn) fput (conn->ksnc_sock->file); } +#ifndef CONFIG_SMP +static inline +int ksocknal_nsched(void) +{ + return 1; +} +#else +#include +# if !(defined(CONFIG_X86) && (LINUX_VERSION_CODE >= KERNEL_VERSION(2,4,21))) || defined(CONFIG_X86_64) || (LUSTRE_KERNEL_VERSION < 39) || ((LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)) && !defined(CONFIG_X86_HT)) +static inline int +ksocknal_nsched(void) +{ + return num_online_cpus(); +} + +static inline int +ksocknal_sched2cpu(int i) +{ + return i; +} + +static inline int +ksocknal_irqsched2cpu(int i) +{ + return i; +} +# else +static inline int +ksocknal_nsched(void) +{ + if (smp_num_siblings == 1) + return (num_online_cpus()); + + /* We need to know if this assumption is crap */ + LASSERT (smp_num_siblings == 2); + return (num_online_cpus()/2); +} + +static inline int +ksocknal_sched2cpu(int i) +{ + if (smp_num_siblings == 1) + return i; + + return (i * 2); +} + +static inline int +ksocknal_irqsched2cpu(int i) +{ + return (ksocknal_sched2cpu(i) + 1); +} +# endif +#endif + extern void ksocknal_put_route (ksock_route_t *route); extern void ksocknal_put_peer (ksock_peer_t *peer); extern ksock_peer_t *ksocknal_find_peer_locked (ptl_nid_t nid); @@ -408,7 +497,7 @@ extern ksock_peer_t *ksocknal_get_peer (ptl_nid_t nid); extern int ksocknal_del_route (ptl_nid_t nid, __u32 ipaddr, int single, int keep_conn); extern int ksocknal_create_conn (ksock_route_t *route, - struct socket *sock, int bind_irq, int type); + struct socket *sock, int type); extern void ksocknal_close_conn_locked (ksock_conn_t *conn, int why); extern void ksocknal_terminate_conn (ksock_conn_t *conn); extern void ksocknal_destroy_conn (ksock_conn_t *conn); @@ -429,6 +518,9 @@ extern void ksocknal_data_ready(struct sock *sk, int n); extern void ksocknal_write_space(struct sock *sk); extern int ksocknal_autoconnectd (void *arg); extern int ksocknal_reaper (void *arg); +extern int ksocknal_get_conn_tunables (ksock_conn_t *conn, int *txmem, + int *rxmem, int *nagle); extern int ksocknal_setup_sock (struct socket *sock); -extern int ksocknal_hello (struct socket *sock, - ptl_nid_t *nid, int *type, __u64 *incarnation); +extern int ksocknal_send_hello (ksock_conn_t *conn, __u32 *ipaddrs, int nipaddrs); +extern int ksocknal_recv_hello (ksock_conn_t *conn, + ptl_nid_t *nid, __u64 *incarnation, __u32 *ipaddrs);