#define NAL_CMD_DEL_AUTOCONN 105
#define NAL_CMD_ADD_AUTOCONN 106
#define NAL_CMD_GET_AUTOCONN 107
+#define NAL_CMD_GET_TXDESC 108
enum {
DEBUG_DAEMON_START = 1,
int jt_ptl_connect(int argc, char **argv);
int jt_ptl_disconnect(int argc, char **argv);
int jt_ptl_push_connection(int argc, char **argv);
+int jt_ptl_print_active_txs(int argc, char **argv);
int jt_ptl_ping(int argc, char **argv);
int jt_ptl_shownid(int argc, char **argv);
int jt_ptl_mynid(int argc, char **argv);
int jt_ptl_connect(int argc, char **argv);
int jt_ptl_disconnect(int argc, char **argv);
int jt_ptl_push_connection(int argc, char **argv);
+int jt_ptl_print_active_txs(int argc, char **argv);
int jt_ptl_ping(int argc, char **argv);
int jt_ptl_shownid(int argc, char **argv);
int jt_ptl_mynid(int argc, char **argv);
}
int
+kqswnal_get_tx_desc (struct portal_ioctl_data *data)
+{
+ unsigned long flags;
+ struct list_head *tmp;
+ kqswnal_tx_t *ktx;
+ int index = data->ioc_count;
+ int rc = -ENOENT;
+
+ spin_lock_irqsave (&kqswnal_data.kqn_idletxd_lock, flags);
+
+ list_for_each (tmp, &kqswnal_data.kqn_activetxds) {
+ if (index-- != 0)
+ continue;
+
+ ktx = list_entry (tmp, kqswnal_tx_t, ktx_list);
+
+ data->ioc_pbuf1 = (char *)ktx;
+ data->ioc_count = NTOH__u32(ktx->ktx_wire_hdr->type);
+ data->ioc_size = NTOH__u32(PTL_HDR_LENGTH(ktx->ktx_wire_hdr));
+ data->ioc_nid = NTOH__u64(ktx->ktx_wire_hdr->dest_nid);
+ data->ioc_nid2 = ktx->ktx_nid;
+ data->ioc_misc = ktx->ktx_launcher;
+ data->ioc_flags = (list_empty (&ktx->ktx_delayed_list) ? 0 : 1) |
+ ((!ktx->ktx_forwarding) ? 0 : 2) |
+ ((!ktx->ktx_isnblk) ? 0 : 4);
+
+ rc = 0;
+ break;
+ }
+
+ spin_unlock_irqrestore (&kqswnal_data.kqn_idletxd_lock, flags);
+ return (rc);
+}
+
+int
kqswnal_cmd (struct portal_ioctl_data *data, void *private)
{
LASSERT (data != NULL);
switch (data->ioc_nal_cmd) {
+ case NAL_CMD_GET_TXDESC:
+ return (kqswnal_get_tx_desc (data));
+
case NAL_CMD_REGISTER_MYNID:
CDEBUG (D_IOCTL, "setting NID offset to "LPX64" (was "LPX64")\n",
data->ioc_nid - kqswnal_data.kqn_elanid,
INIT_LIST_HEAD (&kqswnal_data.kqn_idletxds);
INIT_LIST_HEAD (&kqswnal_data.kqn_nblk_idletxds);
+ INIT_LIST_HEAD (&kqswnal_data.kqn_activetxds);
spin_lock_init (&kqswnal_data.kqn_idletxd_lock);
init_waitqueue_head (&kqswnal_data.kqn_idletxd_waitq);
INIT_LIST_HEAD (&kqswnal_data.kqn_idletxd_fwdq);
ktx->ktx_basepage = basepage + premapped_pages; /* message mapping starts here */
ktx->ktx_npages = KQSW_NTXMSGPAGES - premapped_pages; /* for this many pages */
- if (i < KQSW_NTXMSGS)
- ktx->ktx_idle = &kqswnal_data.kqn_idletxds;
- else
- ktx->ktx_idle = &kqswnal_data.kqn_nblk_idletxds;
+ INIT_LIST_HEAD (&ktx->ktx_delayed_list);
- list_add_tail (&ktx->ktx_list, ktx->ktx_idle);
+ ktx->ktx_isnblk = (i >= KQSW_NTXMSGS);
+ list_add_tail (&ktx->ktx_list,
+ ktx->ktx_isnblk ? &kqswnal_data.kqn_nblk_idletxds :
+ &kqswnal_data.kqn_idletxds);
}
/**********************************************************************/
#include <portals/p30.h>
#include <portals/lib-p30.h>
-#define KQSW_CHECKSUM 0
+#define KQSW_CHECKSUM 0
#if KQSW_CHECKSUM
typedef unsigned long kqsw_csum_t;
-#define KQSW_CSUM_SIZE (2 * sizeof (kqsw_csum_t))
+#define KQSW_CSUM_SIZE (2 * sizeof (kqsw_csum_t))
#else
-#define KQSW_CSUM_SIZE 0
+#define KQSW_CSUM_SIZE 0
#endif
-#define KQSW_HDR_SIZE (sizeof (ptl_hdr_t) + KQSW_CSUM_SIZE)
+#define KQSW_HDR_SIZE (sizeof (ptl_hdr_t) + KQSW_CSUM_SIZE)
/*
* Elan NAL
*/
-#define EP_SVC_LARGE_PORTALS_SMALL (0x10) /* Portals over elan port number (large payloads) */
-#define EP_SVC_LARGE_PORTALS_LARGE (0x11) /* Portals over elan port number (small payloads) */
+#define EP_SVC_LARGE_PORTALS_SMALL (0x10) /* Portals over elan port number (large payloads) */
+#define EP_SVC_LARGE_PORTALS_LARGE (0x11) /* Portals over elan port number (small payloads) */
/* NB small/large message sizes are GLOBAL constants */
/*
* NB no mention of PAGE_SIZE for interoperability
*/
#if PTL_LARGE_MTU
-# define KQSW_MAXPAYLOAD (256<<10) /* biggest message this NAL will cope with */
+# define KQSW_MAXPAYLOAD (256<<10) /* biggest message this NAL will cope with */
#else
-# define KQSW_MAXPAYLOAD (64<<10) /* biggest message this NAL will cope with */
+# define KQSW_MAXPAYLOAD (64<<10) /* biggest message this NAL will cope with */
#endif
-#define KQSW_SMALLPAYLOAD ((4<<10) - KQSW_HDR_SIZE) /* small/large ep receiver breakpoint */
+#define KQSW_SMALLPAYLOAD ((4<<10) - KQSW_HDR_SIZE) /* small/large ep receiver breakpoint */
-#define KQSW_TX_MAXCONTIG (1<<10) /* largest payload that gets made contiguous on transmit */
+#define KQSW_TX_MAXCONTIG (1<<10) /* largest payload that gets made contiguous on transmit */
-#define KQSW_NTXMSGS 8 /* # normal transmit messages */
-#define KQSW_NNBLK_TXMSGS 256 /* # reserved transmit messages if can't block */
+#define KQSW_NTXMSGS 8 /* # normal transmit messages */
+#define KQSW_NNBLK_TXMSGS 256 /* # reserved transmit messages if can't block */
-#define KQSW_NRXMSGS_LARGE 64 /* # large receive buffers */
-#define KQSW_EP_ENVELOPES_LARGE 128 /* # large ep envelopes */
+#define KQSW_NRXMSGS_LARGE 64 /* # large receive buffers */
+#define KQSW_EP_ENVELOPES_LARGE 128 /* # large ep envelopes */
-#define KQSW_NRXMSGS_SMALL 256 /* # small receive buffers */
-#define KQSW_EP_ENVELOPES_SMALL 2048 /* # small ep envelopes */
+#define KQSW_NRXMSGS_SMALL 256 /* # small receive buffers */
+#define KQSW_EP_ENVELOPES_SMALL 2048 /* # small ep envelopes */
-#define KQSW_RESCHED 100 /* # busy loops that forces scheduler to yield */
+#define KQSW_RESCHED 100 /* # busy loops that forces scheduler to yield */
/*
* derived constants
*/
-#define KQSW_TX_BUFFER_SIZE (KQSW_HDR_SIZE + KQSW_TX_MAXCONTIG)
+#define KQSW_TX_BUFFER_SIZE (KQSW_HDR_SIZE + KQSW_TX_MAXCONTIG)
/* The pre-allocated tx buffer (hdr + small payload) */
-#define KQSW_NTXMSGPAGES (btopr(KQSW_TX_BUFFER_SIZE) + 1 + btopr(KQSW_MAXPAYLOAD) + 1)
+#define KQSW_NTXMSGPAGES (btopr(KQSW_TX_BUFFER_SIZE) + 1 + btopr(KQSW_MAXPAYLOAD) + 1)
/* Reserve elan address space for pre-allocated and pre-mapped transmit
* buffer and a full payload too. Extra pages allow for page alignment */
-#define KQSW_NRXMSGPAGES_SMALL (btopr(KQSW_HDR_SIZE + KQSW_SMALLPAYLOAD))
+#define KQSW_NRXMSGPAGES_SMALL (btopr(KQSW_HDR_SIZE + KQSW_SMALLPAYLOAD))
/* receive hdr/payload always contiguous and page aligned */
-#define KQSW_NRXMSGBYTES_SMALL (KQSW_NRXMSGPAGES_SMALL * PAGE_SIZE)
+#define KQSW_NRXMSGBYTES_SMALL (KQSW_NRXMSGPAGES_SMALL * PAGE_SIZE)
-#define KQSW_NRXMSGPAGES_LARGE (btopr(KQSW_HDR_SIZE + KQSW_MAXPAYLOAD))
+#define KQSW_NRXMSGPAGES_LARGE (btopr(KQSW_HDR_SIZE + KQSW_MAXPAYLOAD))
/* receive hdr/payload always contiguous and page aligned */
-#define KQSW_NRXMSGBYTES_LARGE (KQSW_NRXMSGPAGES_LARGE * PAGE_SIZE)
+#define KQSW_NRXMSGBYTES_LARGE (KQSW_NRXMSGPAGES_LARGE * PAGE_SIZE)
/* biggest complete packet we can receive (or transmit) */
typedef struct
{
struct list_head krx_list; /* enqueue -> thread */
- EP_RCVR *krx_eprx; /* port to post receives to */
+ EP_RCVR *krx_eprx; /* port to post receives to */
EP_RXD *krx_rxd; /* receive descriptor (for repost) */
E3_Addr krx_elanaddr; /* Elan address of buffer (contiguous in elan vm) */
int krx_npages; /* # pages in receive buffer */
typedef struct
{
- struct list_head ktx_list; /* enqueue idle/delayed */
- struct list_head *ktx_idle; /* where to put when idle */
- char ktx_state; /* What I'm doing */
+ struct list_head ktx_list; /* enqueue idle/active */
+ struct list_head ktx_delayed_list; /* enqueue delayedtxds */
+ int ktx_isnblk:1; /* reserved descriptor? */
+ int ktx_forwarding:1; /* forwarding (rather than local send) */
uint32_t ktx_basepage; /* page offset in reserved elan tx vaddrs for mapping pages */
int ktx_npages; /* pages reserved for mapping messages */
int ktx_nmappedpages; /* # pages mapped for current message */
- EP_IOVEC ktx_iov[EP_MAXFRAG]; /* msg frags (elan vaddrs) */
+ EP_IOVEC ktx_iov[EP_MAXFRAG]; /* msg frags (elan vaddrs) */
int ktx_niov; /* # message frags */
int ktx_port; /* destination ep port */
ptl_nid_t ktx_nid; /* destination node */
void *ktx_args[2]; /* completion passthru */
- E3_Addr ktx_ebuffer; /* elan address of ktx_buffer */
+ E3_Addr ktx_ebuffer; /* elan address of ktx_buffer */
char *ktx_buffer; /* pre-allocated contiguous buffer for hdr + small payloads */
-} kqswnal_tx_t;
-#define KTX_IDLE 0 /* MUST BE ZERO (so zeroed ktx is idle) */
-#define KTX_SENDING 1 /* local send */
-#define KTX_FORWARDING 2 /* routing a packet */
+ /* debug/info fields */
+ pid_t ktx_launcher; /* pid of launching process */
+ ptl_hdr_t *ktx_wire_hdr; /* portals header (wire endian) */
+} kqswnal_tx_t;
typedef struct
{
kqswnal_tx_t *kqn_txds; /* all the transmit descriptors */
struct list_head kqn_idletxds; /* transmit descriptors free to use */
- struct list_head kqn_nblk_idletxds; /* reserve of */
+ struct list_head kqn_nblk_idletxds; /* reserved free transmit descriptors */
+ struct list_head kqn_activetxds; /* transmit descriptors being used */
spinlock_t kqn_idletxd_lock; /* serialise idle txd access */
wait_queue_head_t kqn_idletxd_waitq; /* sender blocks here waiting for idle txd */
struct list_head kqn_idletxd_fwdq; /* forwarded packets block here waiting for idle txd */
spinlock_t kqn_statelock; /* cb_cli/cb_sti */
nal_cb_t *kqn_cb; /* -> kqswnal_lib */
- EP_DEV *kqn_epdev; /* elan device */
- EP_XMTR *kqn_eptx; /* elan transmitter */
- EP_RCVR *kqn_eprx_small; /* elan receiver (small messages) */
- EP_RCVR *kqn_eprx_large; /* elan receiver (large messages) */
- ELAN3_DMA_HANDLE *kqn_eptxdmahandle; /* elan reserved tx vaddrs */
- ELAN3_DMA_HANDLE *kqn_eprxdmahandle; /* elan reserved rx vaddrs */
+ EP_DEV *kqn_epdev; /* elan device */
+ EP_XMTR *kqn_eptx; /* elan transmitter */
+ EP_RCVR *kqn_eprx_small; /* elan receiver (small messages) */
+ EP_RCVR *kqn_eprx_large; /* elan receiver (large messages) */
+ ELAN3_DMA_HANDLE *kqn_eptxdmahandle; /* elan reserved tx vaddrs */
+ ELAN3_DMA_HANDLE *kqn_eprxdmahandle; /* elan reserved rx vaddrs */
kpr_router_t kqn_router; /* connection to Kernel Portals Router module */
ptl_nid_t kqn_nid_offset; /* this cluster's NID offset */
} kqswnal_data_t;
/* kqn_init state */
-#define KQN_INIT_NOTHING 0 /* MUST BE ZERO so zeroed state is initialised OK */
-#define KQN_INIT_DATA 1
-#define KQN_INIT_PTL 2
-#define KQN_INIT_ALL 3
+#define KQN_INIT_NOTHING 0 /* MUST BE ZERO so zeroed state is initialised OK */
+#define KQN_INIT_DATA 1
+#define KQN_INIT_PTL 2
+#define KQN_INIT_ALL 3
extern nal_cb_t kqswnal_lib;
extern nal_t kqswnal_api;
kqswnal_put_idle_tx (kqswnal_tx_t *ktx)
{
kpr_fwd_desc_t *fwd = NULL;
- struct list_head *idle = ktx->ktx_idle;
unsigned long flags;
- kqswnal_unmap_tx (ktx); /* release temporary mappings */
- ktx->ktx_state = KTX_IDLE;
+ kqswnal_unmap_tx (ktx); /* release temporary mappings */
spin_lock_irqsave (&kqswnal_data.kqn_idletxd_lock, flags);
- list_add (&ktx->ktx_list, idle);
+ list_del (&ktx->ktx_list); /* take off active list */
- /* reserved for non-blocking tx */
- if (idle == &kqswnal_data.kqn_nblk_idletxds) {
+ if (ktx->ktx_isnblk) {
+ /* reserved for non-blocking tx */
+ list_add (&ktx->ktx_list, &kqswnal_data.kqn_nblk_idletxds);
spin_unlock_irqrestore (&kqswnal_data.kqn_idletxd_lock, flags);
return;
}
+ list_add (&ktx->ktx_list, &kqswnal_data.kqn_idletxds);
+
/* anything blocking for a tx descriptor? */
if (!list_empty(&kqswnal_data.kqn_idletxd_fwdq)) /* forwarded packet? */
{
if (!list_empty (&kqswnal_data.kqn_idletxds)) {
ktx = list_entry (kqswnal_data.kqn_idletxds.next,
kqswnal_tx_t, ktx_list);
- list_del (&ktx->ktx_list);
break;
}
ktx = list_entry (kqswnal_data.kqn_nblk_idletxds.next,
kqswnal_tx_t, ktx_list);
- list_del (&ktx->ktx_list);
break;
}
!list_empty (&kqswnal_data.kqn_idletxds));
}
+ if (ktx != NULL) {
+ list_del (&ktx->ktx_list);
+ list_add (&ktx->ktx_list, &kqswnal_data.kqn_activetxds);
+ ktx->ktx_launcher = current->pid;
+ }
+
spin_unlock_irqrestore (&kqswnal_data.kqn_idletxd_lock, flags);
/* Idle descs can't have any mapped (as opposed to pre-mapped) pages */
void
kqswnal_tx_done (kqswnal_tx_t *ktx, int error)
{
- switch (ktx->ktx_state) {
- case KTX_FORWARDING: /* router asked me to forward this packet */
+ if (ktx->ktx_forwarding) /* router asked me to forward this packet */
kpr_fwd_done (&kqswnal_data.kqn_router,
(kpr_fwd_desc_t *)ktx->ktx_args[0], error);
- break;
-
- case KTX_SENDING: /* packet sourced locally */
+ else /* packet sourced locally */
lib_finalize (&kqswnal_lib, ktx->ktx_args[0],
(lib_msg_t *)ktx->ktx_args[1]);
- break;
-
- default:
- LASSERT (0);
- }
kqswnal_put_idle_tx (ktx);
}
spin_lock_irqsave (&kqswnal_data.kqn_sched_lock, flags);
- list_add_tail (&ktx->ktx_list, &kqswnal_data.kqn_delayedtxds);
+ list_add_tail (&ktx->ktx_delayed_list, &kqswnal_data.kqn_delayedtxds);
if (waitqueue_active (&kqswnal_data.kqn_sched_waitq))
wake_up (&kqswnal_data.kqn_sched_waitq);
}
memcpy (ktx->ktx_buffer, hdr, sizeof (*hdr)); /* copy hdr from caller's stack */
+ ktx->ktx_wire_hdr = (ptl_hdr_t *)ktx->ktx_buffer;
#if KQSW_CHECKSUM
csum = kqsw_csum (0, (char *)hdr, sizeof (*hdr));
}
}
- ktx->ktx_port = (payload_nob <= KQSW_SMALLPAYLOAD) ?
- EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
- ktx->ktx_nid = nid;
- ktx->ktx_state = KTX_SENDING; /* => lib_finalize() on completion */
- ktx->ktx_args[0] = private;
- ktx->ktx_args[1] = cookie;
+ ktx->ktx_port = (payload_nob <= KQSW_SMALLPAYLOAD) ?
+ EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
+ ktx->ktx_nid = nid;
+ ktx->ktx_forwarding = 0; /* => lib_finalize() on completion */
+ ktx->ktx_args[0] = private;
+ ktx->ktx_args[1] = cookie;
rc = kqswnal_launch (ktx);
if (rc != 0) { /* failed? */
ktx->ktx_iov[0].Base = ktx->ktx_ebuffer; /* already mapped */
ktx->ktx_iov[0].Len = nob;
ktx->ktx_niov = 1;
+
+ ktx->ktx_wire_hdr = (ptl_hdr_t *)ktx->ktx_buffer;
}
else
{
rc = kqswnal_map_tx_iov (ktx, nob, niov, iov);
if (rc != 0)
goto failed;
+
+ ktx->ktx_wire_hdr = (ptl_hdr_t *)iov[0].iov_base;
}
- ktx->ktx_port = (nob <= (sizeof (ptl_hdr_t) + KQSW_SMALLPAYLOAD)) ?
- EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
- ktx->ktx_nid = nid;
- ktx->ktx_state = KTX_FORWARDING; /* kpr_put_packet() on completion */
- ktx->ktx_args[0] = fwd;
+ ktx->ktx_port = (nob <= (sizeof (ptl_hdr_t) + KQSW_SMALLPAYLOAD)) ?
+ EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
+ ktx->ktx_nid = nid;
+ ktx->ktx_forwarding = 1;
+ ktx->ktx_args[0] = fwd;
rc = kqswnal_launch (ktx);
if (rc == 0)
{
ktx = list_entry(kqswnal_data.kqn_delayedtxds.next,
kqswnal_tx_t, ktx_list);
- list_del (&ktx->ktx_list);
+ list_del_init (&ktx->ktx_delayed_list);
spin_unlock_irqrestore(&kqswnal_data.kqn_sched_lock,
flags);
return 0;
}
+int
+jt_ptl_print_active_txs (int argc, char **argv)
+{
+ struct portal_ioctl_data data;
+ int index;
+ int rc;
+
+ if (!g_nal_is_compatible (argv[0], QSWNAL, 0))
+ return -1;
+
+ for (index = 0;;index++) {
+ PORTAL_IOC_INIT (data);
+ data.ioc_nal = g_nal;
+ data.ioc_nal_cmd = NAL_CMD_GET_TXDESC;
+ data.ioc_count = index;
+
+ rc = l_ioctl (PORTALS_DEV_ID, IOC_PORTAL_NAL_CMD, &data);
+ if (rc != 0)
+ break;
+
+ printf ("%p: %5s payload %6d bytes to "LPX64" via "LPX64" by pid %6d: %s,%s,%s\n",
+ data.ioc_pbuf1,
+ data.ioc_count == PTL_MSG_ACK ? "ACK" :
+ data.ioc_count == PTL_MSG_PUT ? "PUT" :
+ data.ioc_count == PTL_MSG_GET ? "GET" :
+ data.ioc_count == PTL_MSG_REPLY ? "REPLY" : "<wierd message>",
+ data.ioc_size,
+ data.ioc_nid,
+ data.ioc_nid2,
+ data.ioc_misc,
+ (data.ioc_flags & 1) ? "delayed" : "active",
+ (data.ioc_flags & 2) ? "forwarding" : "sending",
+ (data.ioc_flags & 4) ? "nblk" : "normal");
+ }
+
+ if (index == 0)
+ printf ("<no active descs>\n");
+ return 0;
+}
+
int jt_ptl_ping(int argc, char **argv)
{
int rc;
{"connect", jt_ptl_connect, 0, "connect to a remote nid (args: host port [xi])"},
{"disconnect", jt_ptl_disconnect, 0, "disconnect from a remote nid (args: [nid] [host]"},
{"push", jt_ptl_push_connection, 0, "flush connection to a remote nid (args: [nid]"},
+ {"active_tx", jt_ptl_print_active_txs, 0, "print active transmits (no args)"},
{"ping", jt_ptl_ping, 0, "do a ping test (args: nid [count] [size] [timeout])"},
{"shownid", jt_ptl_shownid, 0, "print the local NID"},
{"mynid", jt_ptl_mynid, 0, "inform the socknal of the local NID (args: [hostname])"},
#define NAL_CMD_DEL_AUTOCONN 105
#define NAL_CMD_ADD_AUTOCONN 106
#define NAL_CMD_GET_AUTOCONN 107
+#define NAL_CMD_GET_TXDESC 108
enum {
DEBUG_DAEMON_START = 1,
int jt_ptl_connect(int argc, char **argv);
int jt_ptl_disconnect(int argc, char **argv);
int jt_ptl_push_connection(int argc, char **argv);
+int jt_ptl_print_active_txs(int argc, char **argv);
int jt_ptl_ping(int argc, char **argv);
int jt_ptl_shownid(int argc, char **argv);
int jt_ptl_mynid(int argc, char **argv);
}
int
+kqswnal_get_tx_desc (struct portal_ioctl_data *data)
+{
+ unsigned long flags;
+ struct list_head *tmp;
+ kqswnal_tx_t *ktx;
+ int index = data->ioc_count;
+ int rc = -ENOENT;
+
+ spin_lock_irqsave (&kqswnal_data.kqn_idletxd_lock, flags);
+
+ list_for_each (tmp, &kqswnal_data.kqn_activetxds) {
+ if (index-- != 0)
+ continue;
+
+ ktx = list_entry (tmp, kqswnal_tx_t, ktx_list);
+
+ data->ioc_pbuf1 = (char *)ktx;
+ data->ioc_count = NTOH__u32(ktx->ktx_wire_hdr->type);
+ data->ioc_size = NTOH__u32(PTL_HDR_LENGTH(ktx->ktx_wire_hdr));
+ data->ioc_nid = NTOH__u64(ktx->ktx_wire_hdr->dest_nid);
+ data->ioc_nid2 = ktx->ktx_nid;
+ data->ioc_misc = ktx->ktx_launcher;
+ data->ioc_flags = (list_empty (&ktx->ktx_delayed_list) ? 0 : 1) |
+ ((!ktx->ktx_forwarding) ? 0 : 2) |
+ ((!ktx->ktx_isnblk) ? 0 : 4);
+
+ rc = 0;
+ break;
+ }
+
+ spin_unlock_irqrestore (&kqswnal_data.kqn_idletxd_lock, flags);
+ return (rc);
+}
+
+int
kqswnal_cmd (struct portal_ioctl_data *data, void *private)
{
LASSERT (data != NULL);
switch (data->ioc_nal_cmd) {
+ case NAL_CMD_GET_TXDESC:
+ return (kqswnal_get_tx_desc (data));
+
case NAL_CMD_REGISTER_MYNID:
CDEBUG (D_IOCTL, "setting NID offset to "LPX64" (was "LPX64")\n",
data->ioc_nid - kqswnal_data.kqn_elanid,
INIT_LIST_HEAD (&kqswnal_data.kqn_idletxds);
INIT_LIST_HEAD (&kqswnal_data.kqn_nblk_idletxds);
+ INIT_LIST_HEAD (&kqswnal_data.kqn_activetxds);
spin_lock_init (&kqswnal_data.kqn_idletxd_lock);
init_waitqueue_head (&kqswnal_data.kqn_idletxd_waitq);
INIT_LIST_HEAD (&kqswnal_data.kqn_idletxd_fwdq);
ktx->ktx_basepage = basepage + premapped_pages; /* message mapping starts here */
ktx->ktx_npages = KQSW_NTXMSGPAGES - premapped_pages; /* for this many pages */
- if (i < KQSW_NTXMSGS)
- ktx->ktx_idle = &kqswnal_data.kqn_idletxds;
- else
- ktx->ktx_idle = &kqswnal_data.kqn_nblk_idletxds;
+ INIT_LIST_HEAD (&ktx->ktx_delayed_list);
- list_add_tail (&ktx->ktx_list, ktx->ktx_idle);
+ ktx->ktx_isnblk = (i >= KQSW_NTXMSGS);
+ list_add_tail (&ktx->ktx_list,
+ ktx->ktx_isnblk ? &kqswnal_data.kqn_nblk_idletxds :
+ &kqswnal_data.kqn_idletxds);
}
/**********************************************************************/
#include <portals/p30.h>
#include <portals/lib-p30.h>
-#define KQSW_CHECKSUM 0
+#define KQSW_CHECKSUM 0
#if KQSW_CHECKSUM
typedef unsigned long kqsw_csum_t;
-#define KQSW_CSUM_SIZE (2 * sizeof (kqsw_csum_t))
+#define KQSW_CSUM_SIZE (2 * sizeof (kqsw_csum_t))
#else
-#define KQSW_CSUM_SIZE 0
+#define KQSW_CSUM_SIZE 0
#endif
-#define KQSW_HDR_SIZE (sizeof (ptl_hdr_t) + KQSW_CSUM_SIZE)
+#define KQSW_HDR_SIZE (sizeof (ptl_hdr_t) + KQSW_CSUM_SIZE)
/*
* Elan NAL
*/
-#define EP_SVC_LARGE_PORTALS_SMALL (0x10) /* Portals over elan port number (large payloads) */
-#define EP_SVC_LARGE_PORTALS_LARGE (0x11) /* Portals over elan port number (small payloads) */
+#define EP_SVC_LARGE_PORTALS_SMALL (0x10) /* Portals over elan port number (large payloads) */
+#define EP_SVC_LARGE_PORTALS_LARGE (0x11) /* Portals over elan port number (small payloads) */
/* NB small/large message sizes are GLOBAL constants */
/*
* NB no mention of PAGE_SIZE for interoperability
*/
#if PTL_LARGE_MTU
-# define KQSW_MAXPAYLOAD (256<<10) /* biggest message this NAL will cope with */
+# define KQSW_MAXPAYLOAD (256<<10) /* biggest message this NAL will cope with */
#else
-# define KQSW_MAXPAYLOAD (64<<10) /* biggest message this NAL will cope with */
+# define KQSW_MAXPAYLOAD (64<<10) /* biggest message this NAL will cope with */
#endif
-#define KQSW_SMALLPAYLOAD ((4<<10) - KQSW_HDR_SIZE) /* small/large ep receiver breakpoint */
+#define KQSW_SMALLPAYLOAD ((4<<10) - KQSW_HDR_SIZE) /* small/large ep receiver breakpoint */
-#define KQSW_TX_MAXCONTIG (1<<10) /* largest payload that gets made contiguous on transmit */
+#define KQSW_TX_MAXCONTIG (1<<10) /* largest payload that gets made contiguous on transmit */
-#define KQSW_NTXMSGS 8 /* # normal transmit messages */
-#define KQSW_NNBLK_TXMSGS 256 /* # reserved transmit messages if can't block */
+#define KQSW_NTXMSGS 8 /* # normal transmit messages */
+#define KQSW_NNBLK_TXMSGS 256 /* # reserved transmit messages if can't block */
-#define KQSW_NRXMSGS_LARGE 64 /* # large receive buffers */
-#define KQSW_EP_ENVELOPES_LARGE 128 /* # large ep envelopes */
+#define KQSW_NRXMSGS_LARGE 64 /* # large receive buffers */
+#define KQSW_EP_ENVELOPES_LARGE 128 /* # large ep envelopes */
-#define KQSW_NRXMSGS_SMALL 256 /* # small receive buffers */
-#define KQSW_EP_ENVELOPES_SMALL 2048 /* # small ep envelopes */
+#define KQSW_NRXMSGS_SMALL 256 /* # small receive buffers */
+#define KQSW_EP_ENVELOPES_SMALL 2048 /* # small ep envelopes */
-#define KQSW_RESCHED 100 /* # busy loops that forces scheduler to yield */
+#define KQSW_RESCHED 100 /* # busy loops that forces scheduler to yield */
/*
* derived constants
*/
-#define KQSW_TX_BUFFER_SIZE (KQSW_HDR_SIZE + KQSW_TX_MAXCONTIG)
+#define KQSW_TX_BUFFER_SIZE (KQSW_HDR_SIZE + KQSW_TX_MAXCONTIG)
/* The pre-allocated tx buffer (hdr + small payload) */
-#define KQSW_NTXMSGPAGES (btopr(KQSW_TX_BUFFER_SIZE) + 1 + btopr(KQSW_MAXPAYLOAD) + 1)
+#define KQSW_NTXMSGPAGES (btopr(KQSW_TX_BUFFER_SIZE) + 1 + btopr(KQSW_MAXPAYLOAD) + 1)
/* Reserve elan address space for pre-allocated and pre-mapped transmit
* buffer and a full payload too. Extra pages allow for page alignment */
-#define KQSW_NRXMSGPAGES_SMALL (btopr(KQSW_HDR_SIZE + KQSW_SMALLPAYLOAD))
+#define KQSW_NRXMSGPAGES_SMALL (btopr(KQSW_HDR_SIZE + KQSW_SMALLPAYLOAD))
/* receive hdr/payload always contiguous and page aligned */
-#define KQSW_NRXMSGBYTES_SMALL (KQSW_NRXMSGPAGES_SMALL * PAGE_SIZE)
+#define KQSW_NRXMSGBYTES_SMALL (KQSW_NRXMSGPAGES_SMALL * PAGE_SIZE)
-#define KQSW_NRXMSGPAGES_LARGE (btopr(KQSW_HDR_SIZE + KQSW_MAXPAYLOAD))
+#define KQSW_NRXMSGPAGES_LARGE (btopr(KQSW_HDR_SIZE + KQSW_MAXPAYLOAD))
/* receive hdr/payload always contiguous and page aligned */
-#define KQSW_NRXMSGBYTES_LARGE (KQSW_NRXMSGPAGES_LARGE * PAGE_SIZE)
+#define KQSW_NRXMSGBYTES_LARGE (KQSW_NRXMSGPAGES_LARGE * PAGE_SIZE)
/* biggest complete packet we can receive (or transmit) */
typedef struct
{
struct list_head krx_list; /* enqueue -> thread */
- EP_RCVR *krx_eprx; /* port to post receives to */
+ EP_RCVR *krx_eprx; /* port to post receives to */
EP_RXD *krx_rxd; /* receive descriptor (for repost) */
E3_Addr krx_elanaddr; /* Elan address of buffer (contiguous in elan vm) */
int krx_npages; /* # pages in receive buffer */
typedef struct
{
- struct list_head ktx_list; /* enqueue idle/delayed */
- struct list_head *ktx_idle; /* where to put when idle */
- char ktx_state; /* What I'm doing */
+ struct list_head ktx_list; /* enqueue idle/active */
+ struct list_head ktx_delayed_list; /* enqueue delayedtxds */
+ int ktx_isnblk:1; /* reserved descriptor? */
+ int ktx_forwarding:1; /* forwarding (rather than local send) */
uint32_t ktx_basepage; /* page offset in reserved elan tx vaddrs for mapping pages */
int ktx_npages; /* pages reserved for mapping messages */
int ktx_nmappedpages; /* # pages mapped for current message */
- EP_IOVEC ktx_iov[EP_MAXFRAG]; /* msg frags (elan vaddrs) */
+ EP_IOVEC ktx_iov[EP_MAXFRAG]; /* msg frags (elan vaddrs) */
int ktx_niov; /* # message frags */
int ktx_port; /* destination ep port */
ptl_nid_t ktx_nid; /* destination node */
void *ktx_args[2]; /* completion passthru */
- E3_Addr ktx_ebuffer; /* elan address of ktx_buffer */
+ E3_Addr ktx_ebuffer; /* elan address of ktx_buffer */
char *ktx_buffer; /* pre-allocated contiguous buffer for hdr + small payloads */
-} kqswnal_tx_t;
-#define KTX_IDLE 0 /* MUST BE ZERO (so zeroed ktx is idle) */
-#define KTX_SENDING 1 /* local send */
-#define KTX_FORWARDING 2 /* routing a packet */
+ /* debug/info fields */
+ pid_t ktx_launcher; /* pid of launching process */
+ ptl_hdr_t *ktx_wire_hdr; /* portals header (wire endian) */
+} kqswnal_tx_t;
typedef struct
{
kqswnal_tx_t *kqn_txds; /* all the transmit descriptors */
struct list_head kqn_idletxds; /* transmit descriptors free to use */
- struct list_head kqn_nblk_idletxds; /* reserve of */
+ struct list_head kqn_nblk_idletxds; /* reserved free transmit descriptors */
+ struct list_head kqn_activetxds; /* transmit descriptors being used */
spinlock_t kqn_idletxd_lock; /* serialise idle txd access */
wait_queue_head_t kqn_idletxd_waitq; /* sender blocks here waiting for idle txd */
struct list_head kqn_idletxd_fwdq; /* forwarded packets block here waiting for idle txd */
spinlock_t kqn_statelock; /* cb_cli/cb_sti */
nal_cb_t *kqn_cb; /* -> kqswnal_lib */
- EP_DEV *kqn_epdev; /* elan device */
- EP_XMTR *kqn_eptx; /* elan transmitter */
- EP_RCVR *kqn_eprx_small; /* elan receiver (small messages) */
- EP_RCVR *kqn_eprx_large; /* elan receiver (large messages) */
- ELAN3_DMA_HANDLE *kqn_eptxdmahandle; /* elan reserved tx vaddrs */
- ELAN3_DMA_HANDLE *kqn_eprxdmahandle; /* elan reserved rx vaddrs */
+ EP_DEV *kqn_epdev; /* elan device */
+ EP_XMTR *kqn_eptx; /* elan transmitter */
+ EP_RCVR *kqn_eprx_small; /* elan receiver (small messages) */
+ EP_RCVR *kqn_eprx_large; /* elan receiver (large messages) */
+ ELAN3_DMA_HANDLE *kqn_eptxdmahandle; /* elan reserved tx vaddrs */
+ ELAN3_DMA_HANDLE *kqn_eprxdmahandle; /* elan reserved rx vaddrs */
kpr_router_t kqn_router; /* connection to Kernel Portals Router module */
ptl_nid_t kqn_nid_offset; /* this cluster's NID offset */
} kqswnal_data_t;
/* kqn_init state */
-#define KQN_INIT_NOTHING 0 /* MUST BE ZERO so zeroed state is initialised OK */
-#define KQN_INIT_DATA 1
-#define KQN_INIT_PTL 2
-#define KQN_INIT_ALL 3
+#define KQN_INIT_NOTHING 0 /* MUST BE ZERO so zeroed state is initialised OK */
+#define KQN_INIT_DATA 1
+#define KQN_INIT_PTL 2
+#define KQN_INIT_ALL 3
extern nal_cb_t kqswnal_lib;
extern nal_t kqswnal_api;
kqswnal_put_idle_tx (kqswnal_tx_t *ktx)
{
kpr_fwd_desc_t *fwd = NULL;
- struct list_head *idle = ktx->ktx_idle;
unsigned long flags;
- kqswnal_unmap_tx (ktx); /* release temporary mappings */
- ktx->ktx_state = KTX_IDLE;
+ kqswnal_unmap_tx (ktx); /* release temporary mappings */
spin_lock_irqsave (&kqswnal_data.kqn_idletxd_lock, flags);
- list_add (&ktx->ktx_list, idle);
+ list_del (&ktx->ktx_list); /* take off active list */
- /* reserved for non-blocking tx */
- if (idle == &kqswnal_data.kqn_nblk_idletxds) {
+ if (ktx->ktx_isnblk) {
+ /* reserved for non-blocking tx */
+ list_add (&ktx->ktx_list, &kqswnal_data.kqn_nblk_idletxds);
spin_unlock_irqrestore (&kqswnal_data.kqn_idletxd_lock, flags);
return;
}
+ list_add (&ktx->ktx_list, &kqswnal_data.kqn_idletxds);
+
/* anything blocking for a tx descriptor? */
if (!list_empty(&kqswnal_data.kqn_idletxd_fwdq)) /* forwarded packet? */
{
if (!list_empty (&kqswnal_data.kqn_idletxds)) {
ktx = list_entry (kqswnal_data.kqn_idletxds.next,
kqswnal_tx_t, ktx_list);
- list_del (&ktx->ktx_list);
break;
}
ktx = list_entry (kqswnal_data.kqn_nblk_idletxds.next,
kqswnal_tx_t, ktx_list);
- list_del (&ktx->ktx_list);
break;
}
!list_empty (&kqswnal_data.kqn_idletxds));
}
+ if (ktx != NULL) {
+ list_del (&ktx->ktx_list);
+ list_add (&ktx->ktx_list, &kqswnal_data.kqn_activetxds);
+ ktx->ktx_launcher = current->pid;
+ }
+
spin_unlock_irqrestore (&kqswnal_data.kqn_idletxd_lock, flags);
/* Idle descs can't have any mapped (as opposed to pre-mapped) pages */
void
kqswnal_tx_done (kqswnal_tx_t *ktx, int error)
{
- switch (ktx->ktx_state) {
- case KTX_FORWARDING: /* router asked me to forward this packet */
+ if (ktx->ktx_forwarding) /* router asked me to forward this packet */
kpr_fwd_done (&kqswnal_data.kqn_router,
(kpr_fwd_desc_t *)ktx->ktx_args[0], error);
- break;
-
- case KTX_SENDING: /* packet sourced locally */
+ else /* packet sourced locally */
lib_finalize (&kqswnal_lib, ktx->ktx_args[0],
(lib_msg_t *)ktx->ktx_args[1]);
- break;
-
- default:
- LASSERT (0);
- }
kqswnal_put_idle_tx (ktx);
}
spin_lock_irqsave (&kqswnal_data.kqn_sched_lock, flags);
- list_add_tail (&ktx->ktx_list, &kqswnal_data.kqn_delayedtxds);
+ list_add_tail (&ktx->ktx_delayed_list, &kqswnal_data.kqn_delayedtxds);
if (waitqueue_active (&kqswnal_data.kqn_sched_waitq))
wake_up (&kqswnal_data.kqn_sched_waitq);
}
memcpy (ktx->ktx_buffer, hdr, sizeof (*hdr)); /* copy hdr from caller's stack */
+ ktx->ktx_wire_hdr = (ptl_hdr_t *)ktx->ktx_buffer;
#if KQSW_CHECKSUM
csum = kqsw_csum (0, (char *)hdr, sizeof (*hdr));
}
}
- ktx->ktx_port = (payload_nob <= KQSW_SMALLPAYLOAD) ?
- EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
- ktx->ktx_nid = nid;
- ktx->ktx_state = KTX_SENDING; /* => lib_finalize() on completion */
- ktx->ktx_args[0] = private;
- ktx->ktx_args[1] = cookie;
+ ktx->ktx_port = (payload_nob <= KQSW_SMALLPAYLOAD) ?
+ EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
+ ktx->ktx_nid = nid;
+ ktx->ktx_forwarding = 0; /* => lib_finalize() on completion */
+ ktx->ktx_args[0] = private;
+ ktx->ktx_args[1] = cookie;
rc = kqswnal_launch (ktx);
if (rc != 0) { /* failed? */
ktx->ktx_iov[0].Base = ktx->ktx_ebuffer; /* already mapped */
ktx->ktx_iov[0].Len = nob;
ktx->ktx_niov = 1;
+
+ ktx->ktx_wire_hdr = (ptl_hdr_t *)ktx->ktx_buffer;
}
else
{
rc = kqswnal_map_tx_iov (ktx, nob, niov, iov);
if (rc != 0)
goto failed;
+
+ ktx->ktx_wire_hdr = (ptl_hdr_t *)iov[0].iov_base;
}
- ktx->ktx_port = (nob <= (sizeof (ptl_hdr_t) + KQSW_SMALLPAYLOAD)) ?
- EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
- ktx->ktx_nid = nid;
- ktx->ktx_state = KTX_FORWARDING; /* kpr_put_packet() on completion */
- ktx->ktx_args[0] = fwd;
+ ktx->ktx_port = (nob <= (sizeof (ptl_hdr_t) + KQSW_SMALLPAYLOAD)) ?
+ EP_SVC_LARGE_PORTALS_SMALL : EP_SVC_LARGE_PORTALS_LARGE;
+ ktx->ktx_nid = nid;
+ ktx->ktx_forwarding = 1;
+ ktx->ktx_args[0] = fwd;
rc = kqswnal_launch (ktx);
if (rc == 0)
{
ktx = list_entry(kqswnal_data.kqn_delayedtxds.next,
kqswnal_tx_t, ktx_list);
- list_del (&ktx->ktx_list);
+ list_del_init (&ktx->ktx_delayed_list);
spin_unlock_irqrestore(&kqswnal_data.kqn_sched_lock,
flags);
return 0;
}
+int
+jt_ptl_print_active_txs (int argc, char **argv)
+{
+ struct portal_ioctl_data data;
+ int index;
+ int rc;
+
+ if (!g_nal_is_compatible (argv[0], QSWNAL, 0))
+ return -1;
+
+ for (index = 0;;index++) {
+ PORTAL_IOC_INIT (data);
+ data.ioc_nal = g_nal;
+ data.ioc_nal_cmd = NAL_CMD_GET_TXDESC;
+ data.ioc_count = index;
+
+ rc = l_ioctl (PORTALS_DEV_ID, IOC_PORTAL_NAL_CMD, &data);
+ if (rc != 0)
+ break;
+
+ printf ("%p: %5s payload %6d bytes to "LPX64" via "LPX64" by pid %6d: %s,%s,%s\n",
+ data.ioc_pbuf1,
+ data.ioc_count == PTL_MSG_ACK ? "ACK" :
+ data.ioc_count == PTL_MSG_PUT ? "PUT" :
+ data.ioc_count == PTL_MSG_GET ? "GET" :
+ data.ioc_count == PTL_MSG_REPLY ? "REPLY" : "<wierd message>",
+ data.ioc_size,
+ data.ioc_nid,
+ data.ioc_nid2,
+ data.ioc_misc,
+ (data.ioc_flags & 1) ? "delayed" : "active",
+ (data.ioc_flags & 2) ? "forwarding" : "sending",
+ (data.ioc_flags & 4) ? "nblk" : "normal");
+ }
+
+ if (index == 0)
+ printf ("<no active descs>\n");
+ return 0;
+}
+
int jt_ptl_ping(int argc, char **argv)
{
int rc;
{"connect", jt_ptl_connect, 0, "connect to a remote nid (args: host port [xi])"},
{"disconnect", jt_ptl_disconnect, 0, "disconnect from a remote nid (args: [nid] [host]"},
{"push", jt_ptl_push_connection, 0, "flush connection to a remote nid (args: [nid]"},
+ {"active_tx", jt_ptl_print_active_txs, 0, "print active transmits (no args)"},
{"ping", jt_ptl_ping, 0, "do a ping test (args: nid [count] [size] [timeout])"},
{"shownid", jt_ptl_shownid, 0, "print the local NID"},
{"mynid", jt_ptl_mynid, 0, "inform the socknal of the local NID (args: [hostname])"},
"usage: connect <host> <port> [ix]"},
{"disconnect", jt_ptl_disconnect, 0, "disconnect from a remote nid\n"
"usage: disconnect [<nid>]"},
+ {"active_tx", jt_ptl_print_active_txs, 0, "print active transmits (no args)\n"
+ "usage: active_tx"},
{"mynid", jt_ptl_mynid, 0, "inform the socknal of the local nid. "
"The nid defaults to hostname for tcp networks and is automatically "
"setup for elan/myrinet/scimac networks.\n"
git://git.whamcloud.com / fs / lustre-release.git / commitdiff