#include <stdlib.h>
#include <rpc/pmap_clnt.h>
#include <string.h>
+#include <signal.h>
#include <memory.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include "bluesky.h"
extern BlueSkyFS *fs;
+static int outstanding_rpcs = 0;
+static struct bluesky_stats *rpc_recv_stats, *rpc_send_stats;
+
/* TCP port number to use for NFS protocol. (Should be 2049.) */
#define NFS_SERVICE_PORT 2051
static void async_rpc_write(RPCConnection *rpc,
const char *buf, gsize len);
+static void async_rpc_flush(RPCConnection *rpc);
struct rpc_fail_reply {
uint32_t xid;
uint32_t accept_stat;
};
-/* Routines for XDR-encoding to a growable string. */
-static bool_t xdr_string_putlong(XDR *xdrs, const long *lp)
-{
- GString *str = (GString *)xdrs->x_private;
- uint32_t data = htonl(*lp);
- g_string_set_size(str, str->len + 4);
- memcpy(str->str + str->len - 4, &data, 4);
- return TRUE;
-}
-
-static bool_t xdr_string_putbytes(XDR *xdrs, const char *addr, u_int len)
-{
- GString *str = (GString *)xdrs->x_private;
- g_string_set_size(str, str->len + len);
- memcpy(str->str + str->len - len, addr, len);
- return TRUE;
-}
-
-static u_int xdr_string_getpos(const XDR *xdrs)
-{
- GString *str = (GString *)xdrs->x_private;
- return str->len;
-}
-
-static bool_t xdr_string_putint32(XDR *xdrs, const int32_t *ip)
-{
- GString *str = (GString *)xdrs->x_private;
- uint32_t data = htonl(*ip);
- g_string_set_size(str, str->len + 4);
- memcpy(str->str + str->len - 4, &data, 4);
- return TRUE;
-}
-
-static int32_t *xdr_string_inline(XDR *xdrs, u_int len)
-{
- GString *str = (GString *)xdrs->x_private;
- g_string_set_size(str, str->len + len);
- return (int32_t *)(str->str + str->len - len);
-}
-
-static void xdr_string_destroy(XDR *xdrs)
-{
-}
-
-static struct xdr_ops xdr_string_ops = {
- .x_putlong = xdr_string_putlong,
- .x_putbytes = xdr_string_putbytes,
- .x_getpostn = xdr_string_getpos,
- .x_putint32 = xdr_string_putint32,
- .x_inline = xdr_string_inline,
- .x_destroy = xdr_string_destroy,
-};
-
-static void xdr_string_create(XDR *xdrs, GString *string, enum xdr_op op)
-{
- xdrs->x_op = op;
- xdrs->x_ops = &xdr_string_ops;
- xdrs->x_private = (char *)string;
- xdrs->x_base = NULL;
- xdrs->x_handy = 0;
-}
-
static void
async_rpc_send_failure(RPCRequest *req, enum accept_stat stat)
{
struct rpc_fail_reply header;
+ g_atomic_int_add(&outstanding_rpcs, -1);
+
header.xid = htonl(req->xid);
header.type = htonl(1); /* REPLY */
header.stat = htonl(MSG_ACCEPTED);
header.verf_len = 0;
header.accept_stat = htonl(stat);
+ g_mutex_lock(req->connection->send_lock);
uint32_t fragment = htonl(sizeof(header) | 0x80000000);
- async_rpc_write(req->connection, (const char *)&fragment, sizeof(fragment));
+ if (!req->connection->udp_transport)
+ async_rpc_write(req->connection, (const char *)&fragment,
+ sizeof(fragment));
async_rpc_write(req->connection, (const char *)&header, sizeof(header));
- g_io_channel_flush(req->connection->channel, NULL);
+ async_rpc_flush(req->connection);
+ g_mutex_unlock(req->connection->send_lock);
if (req->args != NULL) {
char buf[4];
if (!req->xdr_args_free(&xdr, req->args)) {
fprintf(stderr, "unable to free arguments");
}
+ g_free(req->args);
}
if (req->raw_args != NULL)
g_string_free(req->raw_args, TRUE);
+ while (req->cleanup != NULL) {
+ struct cleanup_list *c = req->cleanup;
+ req->cleanup = c->next;
+ c->func(c->arg);
+ g_free(c);
+ }
+
+ if (req->connection->udp_transport) {
+ /* For UDP, a connection only exists for the duration of a single
+ * message. */
+ g_mutex_free(req->connection->send_lock);
+ g_string_free(req->connection->msgbuf, TRUE);
+ g_string_free(req->connection->sendbuf, TRUE);
+ g_free(req->connection);
+ }
+
g_free(req);
}
void
async_rpc_send_reply(RPCRequest *req, void *result)
{
+ bluesky_time_hires time_end;
+
GString *str = g_string_new("");
XDR xdr_out;
xdr_string_create(&xdr_out, str, XDR_ENCODE);
return;
}
+ g_atomic_int_add(&outstanding_rpcs, -1);
+ bluesky_stats_add(rpc_send_stats, str->len);
+
struct rpc_reply header;
header.xid = htonl(req->xid);
header.type = htonl(1); /* REPLY */
header.verf_len = 0;
header.accept_stat = 0;
+ g_mutex_lock(req->connection->send_lock);
gsize msg_size = str->len;
uint32_t fragment = htonl((msg_size + sizeof(header)) | 0x80000000);
- async_rpc_write(req->connection, (const char *)&fragment, sizeof(fragment));
+ if (!req->connection->udp_transport)
+ async_rpc_write(req->connection, (const char *)&fragment,
+ sizeof(fragment));
async_rpc_write(req->connection, (const char *)&header, sizeof(header));
async_rpc_write(req->connection, str->str, str->len);
- g_io_channel_flush(req->connection->channel, NULL);
+ async_rpc_flush(req->connection);
+ g_mutex_unlock(req->connection->send_lock);
+
+ time_end = bluesky_now_hires();
+
+#if 0
+ printf("RPC[%"PRIx32"]: time = %"PRId64" ns\n",
+ req->xid, time_end - req->time_start);
+#endif
/* Clean up. */
g_string_free(str, TRUE);
if (!req->xdr_args_free(&xdr, req->args)) {
fprintf(stderr, "unable to free arguments");
}
+ g_free(req->args);
}
if (req->raw_args != NULL)
g_string_free(req->raw_args, TRUE);
+ while (req->cleanup != NULL) {
+ struct cleanup_list *c = req->cleanup;
+ req->cleanup = c->next;
+ c->func(c->arg);
+ g_free(c);
+ }
+
+ if (req->connection->udp_transport) {
+ /* For UDP, a connection only exists for the duration of a single
+ * message. */
+ g_mutex_free(req->connection->send_lock);
+ g_string_free(req->connection->msgbuf, TRUE);
+ g_string_free(req->connection->sendbuf, TRUE);
+ g_free(req->connection);
+ }
+
g_free(req);
}
+static const char *nfs_proc_names[] = {
+ [NFSPROC3_NULL] = "NULL",
+ [NFSPROC3_GETATTR] = "GETATTR",
+ [NFSPROC3_SETATTR] = "SETATTR",
+ [NFSPROC3_LOOKUP] = "LOOKUP",
+ [NFSPROC3_ACCESS] = "ACCESS",
+ [NFSPROC3_READLINK] = "READLINK",
+ [NFSPROC3_READ] = "READ",
+ [NFSPROC3_WRITE] = "WRITE",
+ [NFSPROC3_CREATE] = "CREATE",
+ [NFSPROC3_MKDIR] = "MKDIR",
+ [NFSPROC3_SYMLINK] = "SYMLINK",
+ [NFSPROC3_MKNOD] = "MKNOD",
+ [NFSPROC3_REMOVE] = "REMOVE",
+ [NFSPROC3_RMDIR] = "RMDIR",
+ [NFSPROC3_RENAME] = "RENAME",
+ [NFSPROC3_LINK] = "LINK",
+ [NFSPROC3_READDIR] = "READDIR",
+ [NFSPROC3_READDIRPLUS] = "READDIRPLUS",
+ [NFSPROC3_FSSTAT] = "FSSTAT",
+ [NFSPROC3_FSINFO] = "FSINFO",
+ [NFSPROC3_PATHCONF] = "PATHCONF",
+ [NFSPROC3_COMMIT] = "COMMIT",
+};
+
static void
nfs_program_3(RPCRequest *req)
{
xdrproc_t _xdr_argument, _xdr_result;
char *(*local)(char *, RPCRequest *);
- printf("Dispatched NFS RPC message type %d\n", req->req_proc);
+#if 0
+ if (req->req_proc < sizeof(nfs_proc_names) / sizeof(const char *)) {
+ printf("Dispatched NFS RPC message type %s\n",
+ nfs_proc_names[req->req_proc]);
+ } else {
+ printf("Dispatched unknown NFS RPC message type %d\n", req->req_proc);
+ }
+#endif
switch (req->req_proc) {
case NFSPROC3_NULL:
req->xdr_result = _xdr_result;
result = (*local)((char *)req->args, req);
- bluesky_flushd_invoke(fs);
- bluesky_debug_dump(fs);
-
return;
}
static GMainContext *main_context;
static GMainLoop *main_loop;
+static GThreadPool *rpc_thread_pool;
+
+static volatile int fs_dump_requested = 0;
+
+static void sig_handler(int sig)
+{
+ if (sig == SIGUSR1) {
+ fs_dump_requested = 1;
+ }
+}
+
+static gboolean async_flushd(gpointer data)
+{
+#if 0
+ int rpc_count = g_atomic_int_get(&outstanding_rpcs);
+ if (rpc_count != 0) {
+ g_print("Currently outstanding RPC requests: %d\n", rpc_count);
+ }
+#endif
+
+ if (fs_dump_requested) {
+ bluesky_debug_dump(fs);
+ bluesky_stats_dump_all();
+ fs_dump_requested = 0;
+ }
+
+ bluesky_flushd_invoke(fs);
+ return TRUE;
+}
+
+static void async_rpc_task(gpointer data, gpointer user_data)
+{
+ nfs_program_3((RPCRequest *)data);
+}
+
static async_rpc_init()
{
main_context = g_main_context_new();
main_loop = g_main_loop_new(main_context, FALSE);
+
+ rpc_thread_pool = g_thread_pool_new(async_rpc_task, NULL,
+ bluesky_max_threads, FALSE, NULL);
+
+ /* Arrange to have the cache writeback code run every five seconds. */
+ GSource *source = g_timeout_source_new_seconds(5);
+ g_source_set_callback(source, async_flushd, NULL, NULL);
+ g_source_attach(source, main_context);
+ g_source_unref(source);
+
+ /* Signal USR1 is used to request a debugging dump of filesyste info */
+ struct sigaction sa;
+ sa.sa_handler = sig_handler;
+ sigemptyset(&sa.sa_mask);
+ sa.sa_flags = SA_RESTART;
+ if (sigaction(SIGUSR1, &sa, NULL) < 0) {
+ perror("sigaction");
+ }
}
struct rpc_call_header {
* and the transport should be closed. */
static gboolean async_rpc_dispatch(RPCConnection *rpc)
{
+ bluesky_time_hires time_start = bluesky_now_hires();
int i;
GString *msg = rpc->msgbuf;
const char *buf = msg->str;
+ bluesky_stats_add(rpc_recv_stats, msg->len);
+
if (msg->len < sizeof(struct rpc_call_header)) {
fprintf(stderr, "Short RPC message: only %zd bytes!\n", msg->len);
return FALSE;
return FALSE;
}
+ g_atomic_int_add(&outstanding_rpcs, 1);
+
RPCRequest *req = g_new0(RPCRequest, 1);
req->connection = rpc;
+ req->time_start = time_start;
req->xid = xid;
if (ntohl(header->prog) != NFS_PROGRAM) {
buf += sizeof(struct rpc_call_header);
for (i = 0; i < 2; i++) {
struct rpc_auth *auth = (struct rpc_auth *)buf;
- if (buf - msg->str + sizeof(struct rpc_auth) > msg->len)
+ if (buf - msg->str + sizeof(struct rpc_auth) > msg->len) {
+ g_atomic_int_add(&outstanding_rpcs, -1);
return FALSE;
+ }
gsize authsize = ntohl(auth->len) + sizeof(struct rpc_auth);
- if (authsize > MAX_RPC_MSGSIZE)
+ if (authsize > MAX_RPC_MSGSIZE) {
+ g_atomic_int_add(&outstanding_rpcs, -1);
return FALSE;
+ }
buf += authsize;
}
- if (buf - msg->str > msg->len)
+ if (buf - msg->str > msg->len) {
+ g_atomic_int_add(&outstanding_rpcs, -1);
return FALSE;
+ }
req->raw_args = msg;
req->raw_args_header_bytes = buf - msg->str;
req->req_proc = ntohl(header->proc);
rpc->msgbuf = g_string_new("");
- nfs_program_3(req);
+ if (bluesky_options.sync_frontends) {
+ nfs_program_3(req);
+ } else {
+ g_thread_pool_push(rpc_thread_pool, req, NULL);
+ }
return TRUE;
}
static void async_rpc_write(RPCConnection *rpc,
const char *buf, gsize len)
{
+ if (rpc->udp_transport) {
+ g_string_append_len(rpc->sendbuf, buf, len);
+ return;
+ }
+
+ /* Normal TCP path */
while (len > 0) {
gsize written = 0;
switch (g_io_channel_write_chars(rpc->channel, buf, len,
// g_io_channel_flush(rpc->channel, NULL);
}
+/* Flush a completed message out to the RPC socket */
+static void async_rpc_flush(RPCConnection *rpc)
+{
+ if (rpc->udp_transport) {
+ sendto(g_io_channel_unix_get_fd(rpc->channel),
+ rpc->sendbuf->str, rpc->sendbuf->len, 0,
+ (struct sockaddr *)&rpc->peer, sizeof(struct sockaddr_in));
+ return;
+ } else {
+ g_io_channel_flush(rpc->channel, NULL);
+ }
+}
+
static gboolean async_rpc_do_read(GIOChannel *channel,
GIOCondition condition,
gpointer data)
fprintf(stderr, "Unexpected error or end of file on RPC stream %d!\n",
g_io_channel_unix_get_fd(rpc->channel));
g_io_channel_shutdown(rpc->channel, TRUE, NULL);
+ /* TODO: Clean up connection object. */
return FALSE;
}
/* We were reading in the fragment body. */
rpc->frag_len -= bytes_read;
- if (rpc->frag_len = 0x80000000) {
+ if (rpc->frag_len == 0x80000000) {
/* We have a complete message since this was the last fragment and
* there are no more bytes in it. Dispatch the message. */
if (!async_rpc_dispatch(rpc)) {
rpc->channel = g_io_channel_unix_new(nfd);
rpc->msgbuf = g_string_new("");
g_io_channel_set_encoding(rpc->channel, NULL, NULL);
+ rpc->send_lock = g_mutex_new();
GSource *source = g_io_create_watch(rpc->channel, G_IO_IN);
g_source_set_callback(source, (GSourceFunc)async_rpc_do_read,
rpc, NULL);
g_source_unref(source);
}
+static gboolean async_rpc_do_udp(GIOChannel *channel,
+ GIOCondition condition,
+ gpointer data)
+{
+ char buf[65536];
+
+ struct sockaddr_in src;
+ socklen_t addrlen = sizeof(struct sockaddr_in);
+ ssize_t len = recvfrom(g_io_channel_unix_get_fd(channel),
+ buf, sizeof(buf), 0,
+ (struct sockaddr *)&src, &addrlen);
+ if (len < 0) {
+ fprintf(stderr, "UDP read error: %m, shutting down UDP\n");
+ return FALSE;
+ }
+
+ g_assert(len < sizeof(buf));
+
+ RPCConnection *rpc = g_new0(RPCConnection, 1);
+ rpc->channel = channel;
+ rpc->msgbuf = g_string_new_len(buf, len);
+ rpc->send_lock = g_mutex_new();
+ rpc->udp_transport = TRUE;
+ memcpy(&rpc->peer, &src, sizeof(struct sockaddr_in));
+ rpc->sendbuf = g_string_new("");
+
+ /* We have a complete message since this was the last fragment and
+ * there are no more bytes in it. Dispatch the message. */
+ async_rpc_dispatch(rpc);
+
+ return TRUE;
+}
+
+static async_rpc_register_listening_udp(int fd)
+{
+ GIOChannel *channel = g_io_channel_unix_new(fd);
+ g_io_channel_set_encoding(channel, NULL, NULL);
+ GSource *source = g_io_create_watch(channel, G_IO_IN);
+ g_source_set_callback(source, (GSourceFunc)async_rpc_do_udp,
+ NULL, NULL);
+ g_source_attach(source, main_context);
+ g_source_unref(source);
+}
+
static gpointer async_rpc_run(gpointer data)
{
g_print("Starting NFS main loop...\n");
{
SVCXPRT *transp;
+ rpc_recv_stats = bluesky_stats_new("NFS RPC Messages In");
+ rpc_send_stats = bluesky_stats_new("NFS RPC Messages Out");
+
async_rpc_init();
/* MOUNT protocol */
async_rpc_register_listening(fd);
+ /* Minimal UDP NFSv3 support */
+ fd = socket(AF_INET, SOCK_DGRAM, IPPROTO_UDP);
+ if (fd < 0) {
+ fprintf(stderr, "Unable to create NFS UDP socket: %m\n");
+ exit(1);
+ }
+
+ addr.sin_family = AF_INET;
+ addr.sin_port = htons(NFS_SERVICE_PORT);
+ addr.sin_addr.s_addr = INADDR_ANY;
+ if (bind(fd, (struct sockaddr *)&addr, sizeof(addr)) < 0) {
+ fprintf(stderr, "Unable to bind to NFS UDP address: %m\n");
+ exit(1);
+ }
+
+ if (!pmap_set(NFS_PROGRAM, NFS_V3, IPPROTO_UDP, NFS_SERVICE_PORT)) {
+ fprintf(stderr, "Could not register NFS UDP RPC service!\n");
+ exit(1);
+ }
+
+ async_rpc_register_listening_udp(fd);
+
g_thread_create(async_rpc_run, NULL, TRUE, NULL);
}