SoW-2020-0002: Trace Hit Counters: trigger error reporting integration
[lttng-tools.git] / src / common / shm.c
diff --git a/src/common/shm.c b/src/common/shm.c
new file mode 100644 (file)
index 0000000..74b240c
--- /dev/null
@@ -0,0 +1,233 @@
+/*
+ * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
+ * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ *
+ * SPDX-License-Identifier: GPL-2.0-only
+ *
+ */
+
+#define _LGPL_SOURCE
+#include <fcntl.h>
+#include <limits.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <urcu.h>
+
+#include <common/error.h>
+
+#include "shm.h"
+
+/*
+ * Using fork to set umask in the child process (not multi-thread safe). We
+ * deal with the shm_open vs ftruncate race (happening when the sessiond owns
+ * the shm and does not let everybody modify it, to ensure safety against
+ * shm_unlink) by simply letting the mmap fail and retrying after a few
+ * seconds. For global shm, everybody has rw access to it until the sessiond
+ * starts.
+ */
+static int get_wait_shm(char *shm_path, size_t mmap_size, int global)
+{
+       int wait_shm_fd, ret;
+       mode_t mode;
+
+       assert(shm_path);
+
+       /* Default permissions */
+       mode = S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP;
+
+       /*
+        * Change owner of the shm path.
+        */
+       if (global) {
+               /*
+                * If global session daemon, any application can
+                * register. Make it initially writeable so applications
+                * registering concurrently can do ftruncate() by
+                * themselves.
+                */
+               mode |= S_IROTH | S_IWOTH;
+       }
+
+       /*
+        * We're alone in a child process, so we can modify the process-wide
+        * umask.
+        */
+       umask(~mode);
+
+       /*
+        * Try creating shm (or get rw access). We don't do an exclusive open,
+        * because we allow other processes to create+ftruncate it concurrently.
+        *
+        * A sysctl, fs.protected_regular may prevent the session daemon from
+        * opening a previously created shm when the O_CREAT flag is provided.
+        * Systemd enables this ABI-breaking change by default since v241.
+        *
+        * First, attempt to use the create-or-open semantic that is
+        * desired here. If this fails with EACCES, work around this broken
+        * behaviour and attempt to open the shm without the O_CREAT flag.
+        *
+        * The two attempts are made in this order since applications are
+        * expected to race with the session daemon to create this shm.
+        * Attempting an shm_open() without the O_CREAT flag first could fail
+        * because the file doesn't exist. It could then be created by an
+        * application, which would cause a second try with the O_CREAT flag to
+        * fail with EACCES.
+        *
+        * Note that this introduces a new failure mode where a user could
+        * launch an application (creating the shm) and unlink the shm while
+        * the session daemon is launching, causing the second attempt
+        * to fail. This is not recovered-from as unlinking the shm will
+        * prevent userspace tracing from succeeding anyhow: the sessiond would
+        * use a now-unlinked shm, while the next application would create
+        * a new named shm.
+        */
+       wait_shm_fd = shm_open(shm_path, O_RDWR | O_CREAT, mode);
+       if (wait_shm_fd < 0) {
+               if (errno == EACCES) {
+                       /* Work around sysctl fs.protected_regular. */
+                       DBG("shm_open of %s returned EACCES, this may be caused "
+                                       "by the fs.protected_regular sysctl. "
+                                       "Attempting to open the shm without "
+                                       "creating it.", shm_path);
+                       wait_shm_fd = shm_open(shm_path, O_RDWR, mode);
+               }
+               if (wait_shm_fd < 0) {
+                       PERROR("Failed to open wait shm at %s", shm_path);
+                       goto error;
+               }
+       }
+
+       ret = ftruncate(wait_shm_fd, mmap_size);
+       if (ret < 0) {
+               PERROR("ftruncate wait shm");
+               exit(EXIT_FAILURE);
+       }
+
+#ifndef __FreeBSD__
+       if (global) {
+               ret = fchown(wait_shm_fd, 0, 0);
+               if (ret < 0) {
+                       PERROR("fchown");
+                       exit(EXIT_FAILURE);
+               }
+               /*
+                * If global session daemon, any application can
+                * register so the shm needs to be set in read-only mode
+                * for others.
+                */
+               mode &= ~S_IWOTH;
+               ret = fchmod(wait_shm_fd, mode);
+               if (ret < 0) {
+                       PERROR("fchmod");
+                       exit(EXIT_FAILURE);
+               }
+       } else {
+               ret = fchown(wait_shm_fd, getuid(), getgid());
+               if (ret < 0) {
+                       PERROR("fchown");
+                       exit(EXIT_FAILURE);
+               }
+       }
+#else
+#warning "FreeBSD does not support setting file mode on shm FD."
+#endif
+
+       DBG("Got the wait shm fd %d", wait_shm_fd);
+
+       return wait_shm_fd;
+
+error:
+       DBG("Failing to get the wait shm fd");
+
+       return -1;
+}
+
+/*
+ * Return the wait shm mmap for UST application notification. The global
+ * variable is used to indicate if the the session daemon is global
+ * (root:tracing) or running with an unprivileged user.
+ *
+ * This returned value is used by futex_wait_update() in futex.c to WAKE all
+ * waiters which are UST application waiting for a session daemon.
+ */
+char *shm_ust_get_mmap(char *shm_path, int global)
+{
+       size_t mmap_size;
+       int wait_shm_fd, ret;
+       char *wait_shm_mmap;
+       long sys_page_size;
+
+       assert(shm_path);
+
+       sys_page_size = sysconf(_SC_PAGE_SIZE);
+       if (sys_page_size < 0) {
+               PERROR("sysconf PAGE_SIZE");
+               goto error;
+       }
+       mmap_size = sys_page_size;
+
+       wait_shm_fd = get_wait_shm(shm_path, mmap_size, global);
+       if (wait_shm_fd < 0) {
+               goto error;
+       }
+
+       wait_shm_mmap = mmap(NULL, mmap_size, PROT_WRITE | PROT_READ,
+                       MAP_SHARED, wait_shm_fd, 0);
+
+       /* close shm fd immediately after taking the mmap reference */
+       ret = close(wait_shm_fd);
+       if (ret) {
+               PERROR("Error closing fd");
+       }
+
+       if (wait_shm_mmap == MAP_FAILED) {
+               DBG("mmap error (can be caused by race with ust).");
+               goto error;
+       }
+
+       return wait_shm_mmap;
+
+error:
+       return NULL;
+}
+
+/*
+ * shm_create_anonymous is never called concurrently within a process.
+ */
+int shm_create_anonymous(const char *owner_name)
+{
+       char tmp_name[NAME_MAX];
+       int shmfd, ret;
+
+       ret = snprintf(tmp_name, NAME_MAX, "/shm-%s-%d", owner_name, getpid());
+       if (ret < 0) {
+               PERROR("snprintf");
+               return -1;
+       }
+       /*
+        * Allocate shm, and immediately unlink its shm oject, keeping only the
+        * file descriptor as a reference to the object.
+        */
+       shmfd = shm_open(tmp_name, O_CREAT | O_EXCL | O_RDWR, 0700);
+       if (shmfd < 0) {
+               PERROR("shm_open");
+               goto error_shm_open;
+       }
+       ret = shm_unlink(tmp_name);
+       if (ret < 0 && errno != ENOENT) {
+               PERROR("shm_unlink");
+               goto error_shm_release;
+       }
+       return shmfd;
+
+error_shm_release:
+       ret = close(shmfd);
+       if (ret) {
+               PERROR("close");
+       }
+error_shm_open:
+       return -1;
+}
This page took 0.025755 seconds and 5 git commands to generate.