| 1 |
/* SPDX-License-Identifier: LGPL-2.1-or-later */ |
| 2 |
|
| 3 |
#include <sched.h> |
| 4 |
#include <sys/prctl.h> |
| 5 |
#include <sys/wait.h> |
| 6 |
#include <unistd.h> |
| 7 |
|
| 8 |
#include "async.h" |
| 9 |
#include "errno-util.h" |
| 10 |
#include "fd-util.h" |
| 11 |
#include "log.h" |
| 12 |
#include "process-util.h" |
| 13 |
#include "rm-rf.h" |
| 14 |
#include "signal-util.h" |
| 15 |
|
| 16 |
int asynchronous_sync(PidRef *ret_pid) { |
| 17 |
int r; |
| 18 |
|
| 19 |
/* This forks off an invocation of fork() as a child process, in order to initiate synchronization to |
| 20 |
* disk. Note that we implement this as helper process rather than thread as we don't want the sync() to hang our |
| 21 |
* original process ever, and a thread would do that as the process can't exit with threads hanging in blocking |
| 22 |
* syscalls. */ |
| 23 |
|
| 24 |
r = pidref_safe_fork("(sd-sync)", FORK_RESET_SIGNALS|FORK_CLOSE_ALL_FDS|(ret_pid ? 0 : FORK_DETACH), ret_pid); |
| 25 |
if (r < 0) |
| 26 |
return r; |
| 27 |
if (r == 0) { |
| 28 |
/* Child process */ |
| 29 |
sync(); |
| 30 |
_exit(EXIT_SUCCESS); |
| 31 |
} |
| 32 |
|
| 33 |
return 0; |
| 34 |
} |
| 35 |
|
| 36 |
int asynchronous_fsync(int fd, PidRef *ret_pid) { |
| 37 |
int r; |
| 38 |
|
| 39 |
assert(fd >= 0); |
| 40 |
/* Same as asynchronous_sync() above, but calls fsync() on a specific fd */ |
| 41 |
|
| 42 |
r = pidref_safe_fork_full( |
| 43 |
"(sd-fsync)", |
| 44 |
/* stdio_fds= */ NULL, |
| 45 |
/* except_fds= */ &fd, |
| 46 |
/* n_except_fds= */ 1, |
| 47 |
FORK_RESET_SIGNALS|FORK_CLOSE_ALL_FDS|(ret_pid ? 0 : FORK_DETACH), |
| 48 |
ret_pid); |
| 49 |
if (r < 0) |
| 50 |
return r; |
| 51 |
if (r == 0) { |
| 52 |
/* Child process */ |
| 53 |
(void) fsync(fd); |
| 54 |
_exit(EXIT_SUCCESS); |
| 55 |
} |
| 56 |
|
| 57 |
return 0; |
| 58 |
} |
| 59 |
|
| 60 |
/* We encode the fd to close in the userdata pointer as an unsigned value. The highest bit indicates whether |
| 61 |
* we need to fork again */ |
| 62 |
#define NEED_DOUBLE_FORK (1U << (sizeof(unsigned) * 8 - 1)) |
| 63 |
|
| 64 |
static int close_func(void *p) { |
| 65 |
unsigned v = PTR_TO_UINT(p); |
| 66 |
|
| 67 |
(void) prctl(PR_SET_NAME, (unsigned long*) "(sd-close)"); |
| 68 |
|
| 69 |
/* Note: 💣 This function is invoked in a child process created via glibc's clone() wrapper. In such |
| 70 |
* children memory allocation is not allowed, since glibc does not release malloc mutexes in |
| 71 |
* clone() 💣 */ |
| 72 |
|
| 73 |
if (v & NEED_DOUBLE_FORK) { |
| 74 |
pid_t pid; |
| 75 |
|
| 76 |
v &= ~NEED_DOUBLE_FORK; |
| 77 |
|
| 78 |
/* This inner child will be reparented to the subreaper/PID 1. Here we turn on SIGCHLD, so |
| 79 |
* that the reaper knows when it's time to reap. */ |
| 80 |
pid = clone_with_nested_stack(close_func, SIGCHLD|CLONE_FILES, UINT_TO_PTR(v)); |
| 81 |
if (pid >= 0) |
| 82 |
return 0; |
| 83 |
} |
| 84 |
|
| 85 |
close((int) v); /* no assert() here, we are in the child and the result would be eaten up anyway */ |
| 86 |
return 0; |
| 87 |
} |
| 88 |
|
| 89 |
int asynchronous_close(int fd) { |
| 90 |
unsigned v; |
| 91 |
pid_t pid; |
| 92 |
int r; |
| 93 |
|
| 94 |
/* This is supposed to behave similar to safe_close(), but actually invoke close() asynchronously, so |
| 95 |
* that it will never block. Ideally the kernel would have an API for this, but it doesn't, so we |
| 96 |
* work around it, and hide this as a far away as we can. |
| 97 |
* |
| 98 |
* It is important to us that we don't use threads (via glibc pthread) in PID 1, hence we'll do a |
| 99 |
* minimal subprocess instead which shares our fd table via CLONE_FILES. */ |
| 100 |
|
| 101 |
if (fd < 0) |
| 102 |
return -EBADF; /* already invalid */ |
| 103 |
|
| 104 |
PROTECT_ERRNO; |
| 105 |
|
| 106 |
v = (unsigned) fd; |
| 107 |
|
| 108 |
/* We want to fork off a process that is automatically reaped. For that we'd usually double-fork. But |
| 109 |
* we can optimize this a bit: if we are PID 1 or a subreaper anyway (the systemd service manager |
| 110 |
* process qualifies as this), we can avoid the double forking, since the double forked process would |
| 111 |
* be reparented back to us anyway. */ |
| 112 |
r = is_reaper_process(); |
| 113 |
if (r < 0) |
| 114 |
log_debug_errno(r, "Cannot determine if we are a reaper process, assuming we are not: %m"); |
| 115 |
if (r <= 0) |
| 116 |
v |= NEED_DOUBLE_FORK; |
| 117 |
|
| 118 |
pid = clone_with_nested_stack(close_func, CLONE_FILES | ((v & NEED_DOUBLE_FORK) ? 0 : SIGCHLD), UINT_TO_PTR(v)); |
| 119 |
if (pid < 0) |
| 120 |
safe_close(fd); /* local fallback */ |
| 121 |
else if (v & NEED_DOUBLE_FORK) { |
| 122 |
|
| 123 |
/* Reap the intermediate child. Key here is that we specify __WCLONE, since we didn't ask for |
| 124 |
* any signal to be sent to us on process exit, and otherwise waitid() would refuse waiting |
| 125 |
* then. |
| 126 |
* |
| 127 |
* We usually prefer calling waitid(), but before kernel 4.7 it didn't support __WCLONE while |
| 128 |
* waitpid() did. Hence let's use waitpid() here, it's good enough for our purposes here. */ |
| 129 |
for (;;) |
| 130 |
if (waitpid(pid, NULL, __WCLONE) >= 0 || errno != EINTR) |
| 131 |
break; |
| 132 |
} |
| 133 |
|
| 134 |
return -EBADF; /* return an invalidated fd */ |
| 135 |
} |
| 136 |
|
| 137 |
void asynchronous_close_many(const int fds[], size_t n_fds) { |
| 138 |
assert(fds || n_fds == 0); |
| 139 |
|
| 140 |
FOREACH_ARRAY(i, fds, n_fds) |
| 141 |
asynchronous_close(*i); |
| 142 |
} |
| 143 |
|
| 144 |
int asynchronous_rm_rf(const char *p, RemoveFlags flags) { |
| 145 |
int r; |
| 146 |
|
| 147 |
assert(p); |
| 148 |
|
| 149 |
/* Forks off a child that destroys the specified path. This will be best effort only, i.e. the child |
| 150 |
* will attempt to do its thing, but we won't wait for it or check its success. */ |
| 151 |
|
| 152 |
r = pidref_safe_fork("(sd-rmrf)", FORK_RESET_SIGNALS|FORK_CLOSE_ALL_FDS|FORK_DETACH, /* ret= */ NULL); |
| 153 |
if (r != 0) |
| 154 |
return r; |
| 155 |
|
| 156 |
/* Child */ |
| 157 |
|
| 158 |
/* Let's block SIGTERM here, to grant the operation more time on e.g. final killing spree |
| 159 |
* during shutdown. If this gets stalled pid1 would eventually send SIGKILL to us. */ |
| 160 |
BLOCK_SIGNALS(SIGTERM); |
| 161 |
|
| 162 |
r = rm_rf(p, flags); |
| 163 |
if (r < 0) { |
| 164 |
log_debug_errno(r, "Failed to rm -rf '%s', ignoring: %m", p); |
| 165 |
_exit(EXIT_FAILURE); /* This is a detached process, hence no one really cares, but who knows |
| 166 |
* maybe it's good for debugging/tracing to return an exit code |
| 167 |
* indicative of our failure here. */ |
| 168 |
} |
| 169 |
|
| 170 |
_exit(EXIT_SUCCESS); |
| 171 |
} |