Version: 1
-Previous: fb361b017597e7093e057cf60662253925d238ee
-Head: 8a093195e258cfb4cf392119c8eb9da58c15f587
+Previous: 499387c6788c5f6a48f50c17c2d742ef2017eb73
+Head: a70ee35a8dfd07cca88c8b3b6fa6308bd0d3e3d0
Applied:
transpose: 8081f96f8481d898d8de86cbadac1ae459c93306
- fork: 3600ccfec6ae26e1b44d492eac50fec4f4169d0d
- refresh-temp: 8a093195e258cfb4cf392119c8eb9da58c15f587
+ fork: a70ee35a8dfd07cca88c8b3b6fa6308bd0d3e3d0
Unapplied:
waitall-buggy: c49c6b56c55385774065f5aa2704078e6ae0ceb8
rs-1sge: 8ec392829399dec7ed5c608b8697a482e5faa2de
Bottom: 5fd229f7e20bde0c0ed3050847f79e5d4c8af2b1
-Top: 5fd229f7e20bde0c0ed3050847f79e5d4c8af2b1
+Top: 704e50345b9783ecab92b6e8a634edd31c1342f6
Author: Sean Hefty <sean.hefty@intel.com>
Date: 2012-07-13 15:25:53 -0700
---
-
+diff --git a/src/preload.c b/src/preload.c
+index 8f281d3..fee191b 100644
+--- a/src/preload.c
++++ b/src/preload.c
+@@ -46,6 +46,8 @@
+ #include <string.h>
+ #include <netinet/in.h>
+ #include <netinet/tcp.h>
++#include <unistd.h>
++#include <semaphore.h>
+
+ #include <rdma/rdma_cma.h>
+ #include <rdma/rdma_verbs.h>
+@@ -81,6 +83,7 @@ struct socket_calls {
+ int (*getsockopt)(int socket, int level, int optname,
+ void *optval, socklen_t *optlen);
+ int (*fcntl)(int socket, int cmd, ... /* arg */);
++ pid_t (*fork)(void);
+ };
+
+ static struct socket_calls real;
+@@ -92,10 +95,13 @@ static pthread_mutex_t mut = PTHREAD_MUTEX_INITIALIZER;
+ static int sq_size;
+ static int rq_size;
+ static int sq_inline;
++static int fork_support;
++static int last_accept = -1;
+
+ enum fd_type {
+ fd_normal,
+- fd_rsocket
++ fd_rsocket,
++ fd_fork
+ };
+
+ struct fd_info {
+@@ -207,6 +213,10 @@ void getenv_options(void)
+ var = getenv("RS_INLINE");
+ if (var)
+ sq_inline = atoi(var);
++
++ var = getenv("RDMAV_FORK_SAFE");
++ if (var)
++ fork_support = atoi(var);
+ }
+
+ static void init_preload(void)
+@@ -244,6 +254,7 @@ static void init_preload(void)
+ real.setsockopt = dlsym(RTLD_NEXT, "setsockopt");
+ real.getsockopt = dlsym(RTLD_NEXT, "getsockopt");
+ real.fcntl = dlsym(RTLD_NEXT, "fcntl");
++ real.fork = dlsym(RTLD_NEXT, "fork");
+
+ rs.socket = dlsym(RTLD_DEFAULT, "rsocket");
+ rs.bind = dlsym(RTLD_DEFAULT, "rbind");
+@@ -378,8 +389,16 @@ int socket(int domain, int type, int protocol)
+ ret = rsocket(domain, type, protocol);
+ recursive = 0;
+ if (ret >= 0) {
+- fd_store(index, ret, fd_rsocket);
+- set_rsocket_options(ret);
++ if (fork_support) {
++ rclose(ret);
++ ret = real_socket(domain, type, protocol);
++ if (ret < 0)
++ return ret;
++ fd_store(index, ret, fd_fork);
++ } else {
++ fd_store(index, ret, fd_rsocket);
++ set_rsocket_options(ret);
++ }
+ return index;
+ }
+ fd_close(index, &ret);
+@@ -431,18 +450,48 @@ int accept(int socket, struct sockaddr *addr, socklen_t *addrlen)
+ }
+
+ fd_store(index, ret, fd_rsocket);
++ last_accept = index;
+ return index;
+ } else {
++ last_accept = -1;
+ return real.accept(fd, addr, addrlen);
+ }
+ }
+
++static int connect_fork(int socket, const struct sockaddr *addr, socklen_t addrlen)
++{
++ struct sockaddr_in *sin;
++ int fd, ret;
++ uint32_t msg;
++
++ fd = fd_getd(socket);
++ ret = real.connect(fd, addr, addrlen);
++ if (ret)
++ return ret;
++
++ ret = real.read(fd, &msg, sizeof msg);
++ if ((ret != sizeof msg) || msg) {
++ fd_store(socket, fd, fd_normal);
++ return 0;
++ }
++
++ ret = transpose_socket(socket, fd_rsocket);
++ if (ret < 0)
++ return ret;
++
++ real.close(fd);
++ return rconnect(ret, addr, addrlen);
++}
++
+ int connect(int socket, const struct sockaddr *addr, socklen_t addrlen)
+ {
+ struct sockaddr_in *sin;
+ int fd, ret;
+
+- if (fd_get(socket, &fd) == fd_rsocket) {
++ switch (fd_get(socket, &fd)) {
++ case fd_fork:
++ return connect_fork(socket, addr, addrlen);
++ case fd_rsocket:
+ sin = (struct sockaddr_in *) addr;
+ if (ntohs(sin->sin_port) > 1024) {
+ ret = rconnect(fd, addr, addrlen);
+@@ -754,3 +803,57 @@ int fcntl(int socket, int cmd, ... /* arg */)
+ va_end(args);
+ return ret;
+ }
++
++/*
++ * We can't fork RDMA connections and pass them from the parent to the child
++ * process. Intercept the fork call, and if we're the child establish the
++ * RDMA connection after calling fork. The assumption is that the last
++ * connection accepted by the server will be processed by the child after the
++ * fork call.
++ *
++ * It would be better to establishing the RDMA connection once the child
++ * process tries to use the connection after the fork call (i.e. in a read
++ * or write call), rather than making the previous assumption.
++ */
++pid_t fork(void)
++{
++ struct sockaddr_storage sa;
++ pid_t pid;
++ sem_t *sem;
++ int fd, lfd, newfd, ret, len, param;
++ uint32_t msg;
++
++ init_preload();
++ pid = real.fork();
++ if (pid || !fork_support || (last_accept < 0) ||
++ (fd_get(last_accept, &fd) != fd_fork))
++ goto out;
++
++ sem = sem_open("/rsocket_fork", O_CREAT, 0644, 1);
++ if (sem == SEM_FAILED)
++ goto out;
++
++ lfd = transpose_socket(last_accept, fd_rsocket);
++ if (lfd < 0)
++ goto out;
++
++ param = 1;
++ len = sizeof param;
++ rsetsockopt(lfd, SOL_SOCKET, SO_REUSEADDR, ¶m, len);
++
++ len = sizeof sa;
++ ret = real.getsockname(fd, &sa, &len);
++ if (ret)
++ goto out;
++
++ sem_wait(sem);
++ ret = rbind()
++
++ real.close(fd);
++
++ sem_post(sem);
++ sem_close(sem);
++out:
++ last_accept = -1;
++ return pid;
++}
+++ /dev/null
-Bottom: 5fd229f7e20bde0c0ed3050847f79e5d4c8af2b1
-Top: 704e50345b9783ecab92b6e8a634edd31c1342f6
-Author: Sean Hefty <sean.hefty@intel.com>
-Date: 2012-07-17 11:45:12 -0700
-
-Refresh of fork
-
----
-
-diff --git a/src/preload.c b/src/preload.c
-index 8f281d3..fee191b 100644
---- a/src/preload.c
-+++ b/src/preload.c
-@@ -46,6 +46,8 @@
- #include <string.h>
- #include <netinet/in.h>
- #include <netinet/tcp.h>
-+#include <unistd.h>
-+#include <semaphore.h>
-
- #include <rdma/rdma_cma.h>
- #include <rdma/rdma_verbs.h>
-@@ -81,6 +83,7 @@ struct socket_calls {
- int (*getsockopt)(int socket, int level, int optname,
- void *optval, socklen_t *optlen);
- int (*fcntl)(int socket, int cmd, ... /* arg */);
-+ pid_t (*fork)(void);
- };
-
- static struct socket_calls real;
-@@ -92,10 +95,13 @@ static pthread_mutex_t mut = PTHREAD_MUTEX_INITIALIZER;
- static int sq_size;
- static int rq_size;
- static int sq_inline;
-+static int fork_support;
-+static int last_accept = -1;
-
- enum fd_type {
- fd_normal,
-- fd_rsocket
-+ fd_rsocket,
-+ fd_fork
- };
-
- struct fd_info {
-@@ -207,6 +213,10 @@ void getenv_options(void)
- var = getenv("RS_INLINE");
- if (var)
- sq_inline = atoi(var);
-+
-+ var = getenv("RDMAV_FORK_SAFE");
-+ if (var)
-+ fork_support = atoi(var);
- }
-
- static void init_preload(void)
-@@ -244,6 +254,7 @@ static void init_preload(void)
- real.setsockopt = dlsym(RTLD_NEXT, "setsockopt");
- real.getsockopt = dlsym(RTLD_NEXT, "getsockopt");
- real.fcntl = dlsym(RTLD_NEXT, "fcntl");
-+ real.fork = dlsym(RTLD_NEXT, "fork");
-
- rs.socket = dlsym(RTLD_DEFAULT, "rsocket");
- rs.bind = dlsym(RTLD_DEFAULT, "rbind");
-@@ -378,8 +389,16 @@ int socket(int domain, int type, int protocol)
- ret = rsocket(domain, type, protocol);
- recursive = 0;
- if (ret >= 0) {
-- fd_store(index, ret, fd_rsocket);
-- set_rsocket_options(ret);
-+ if (fork_support) {
-+ rclose(ret);
-+ ret = real_socket(domain, type, protocol);
-+ if (ret < 0)
-+ return ret;
-+ fd_store(index, ret, fd_fork);
-+ } else {
-+ fd_store(index, ret, fd_rsocket);
-+ set_rsocket_options(ret);
-+ }
- return index;
- }
- fd_close(index, &ret);
-@@ -431,18 +450,48 @@ int accept(int socket, struct sockaddr *addr, socklen_t *addrlen)
- }
-
- fd_store(index, ret, fd_rsocket);
-+ last_accept = index;
- return index;
- } else {
-+ last_accept = -1;
- return real.accept(fd, addr, addrlen);
- }
- }
-
-+static int connect_fork(int socket, const struct sockaddr *addr, socklen_t addrlen)
-+{
-+ struct sockaddr_in *sin;
-+ int fd, ret;
-+ uint32_t msg;
-+
-+ fd = fd_getd(socket);
-+ ret = real.connect(fd, addr, addrlen);
-+ if (ret)
-+ return ret;
-+
-+ ret = real.read(fd, &msg, sizeof msg);
-+ if ((ret != sizeof msg) || msg) {
-+ fd_store(socket, fd, fd_normal);
-+ return 0;
-+ }
-+
-+ ret = transpose_socket(socket, fd_rsocket);
-+ if (ret < 0)
-+ return ret;
-+
-+ real.close(fd);
-+ return rconnect(ret, addr, addrlen);
-+}
-+
- int connect(int socket, const struct sockaddr *addr, socklen_t addrlen)
- {
- struct sockaddr_in *sin;
- int fd, ret;
-
-- if (fd_get(socket, &fd) == fd_rsocket) {
-+ switch (fd_get(socket, &fd)) {
-+ case fd_fork:
-+ return connect_fork(socket, addr, addrlen);
-+ case fd_rsocket:
- sin = (struct sockaddr_in *) addr;
- if (ntohs(sin->sin_port) > 1024) {
- ret = rconnect(fd, addr, addrlen);
-@@ -754,3 +803,57 @@ int fcntl(int socket, int cmd, ... /* arg */)
- va_end(args);
- return ret;
- }
-+
-+/*
-+ * We can't fork RDMA connections and pass them from the parent to the child
-+ * process. Intercept the fork call, and if we're the child establish the
-+ * RDMA connection after calling fork. The assumption is that the last
-+ * connection accepted by the server will be processed by the child after the
-+ * fork call.
-+ *
-+ * It would be better to establishing the RDMA connection once the child
-+ * process tries to use the connection after the fork call (i.e. in a read
-+ * or write call), rather than making the previous assumption.
-+ */
-+pid_t fork(void)
-+{
-+ struct sockaddr_storage sa;
-+ pid_t pid;
-+ sem_t *sem;
-+ int fd, lfd, newfd, ret, len, param;
-+ uint32_t msg;
-+
-+ init_preload();
-+ pid = real.fork();
-+ if (pid || !fork_support || (last_accept < 0) ||
-+ (fd_get(last_accept, &fd) != fd_fork))
-+ goto out;
-+
-+ sem = sem_open("/rsocket_fork", O_CREAT, 0644, 1);
-+ if (sem == SEM_FAILED)
-+ goto out;
-+
-+ lfd = transpose_socket(last_accept, fd_rsocket);
-+ if (lfd < 0)
-+ goto out;
-+
-+ param = 1;
-+ len = sizeof param;
-+ rsetsockopt(lfd, SOL_SOCKET, SO_REUSEADDR, ¶m, len);
-+
-+ len = sizeof sa;
-+ ret = real.getsockname(fd, &sa, &len);
-+ if (ret)
-+ goto out;
-+
-+ sem_wait(sem);
-+ ret = rbind()
-+
-+ real.close(fd);
-+
-+ sem_post(sem);
-+ sem_close(sem);
-+out:
-+ last_accept = -1;
-+ return pid;
-+}