Line data Source code
1 : #define _GNU_SOURCE
2 : #include "run.h"
3 :
4 : #include <sys/wait.h>
5 : #include "generated/main_seccomp.h"
6 : #if defined(__aarch64__)
7 : #include "generated/pidns_arm64_seccomp.h"
8 : #else
9 : #include "generated/pidns_seccomp.h"
10 : #endif
11 :
12 : #include "../../../platform/fd_sys_util.h"
13 : #include "../../../platform/fd_file_util.h"
14 : #include "../../../platform/fd_net_util.h"
15 : #include "../../../../disco/net/fd_net_tile.h"
16 :
17 : #include "../configure/configure.h"
18 :
19 : #include <dirent.h>
20 : #include <sched.h>
21 : #include <stdio.h>
22 : #include <stdlib.h> /* getenv */
23 : #include <poll.h>
24 : #include <unistd.h>
25 : #include <errno.h>
26 : #include <fcntl.h>
27 : #include <sys/prctl.h>
28 : #include <sys/resource.h>
29 : #include <sys/mman.h>
30 : #include <sys/stat.h>
31 : #include <linux/capability.h>
32 :
33 : #include "../../../../util/tile/fd_tile_private.h"
34 :
35 : extern fd_topo_obj_callbacks_t * CALLBACKS[];
36 :
37 0 : #define NAME "run"
38 :
39 : void
40 : run_cmd_perm( args_t * args,
41 : fd_cap_chk_t * chk,
42 0 : config_t const * config ) {
43 0 : (void)args;
44 :
45 0 : ulong mlock_limit = fd_topo_mlock_max_tile( &config->topo );
46 :
47 0 : fd_cap_chk_raise_rlimit( chk, NAME, RLIMIT_MEMLOCK, mlock_limit, "call `rlimit(2)` to increase `RLIMIT_MEMLOCK` so all memory can be locked with `mlock(2)`" );
48 0 : fd_cap_chk_raise_rlimit( chk, NAME, RLIMIT_NICE, 40, "call `setpriority(2)` to increase thread priorities" );
49 0 : fd_cap_chk_raise_rlimit( chk, NAME, RLIMIT_NOFILE, CONFIGURE_NR_OPEN_FILES,
50 0 : "call `rlimit(2) to increase `RLIMIT_NOFILE` to allow more open files for Agave" );
51 0 : fd_cap_chk_cap( chk, NAME, CAP_NET_RAW, "call `socket(2)` to bind to a raw socket for use by XDP" );
52 0 : fd_cap_chk_cap( chk, NAME, CAP_SYS_ADMIN, "call `bpf(2)` with the `BPF_OBJ_GET` command to initialize XDP" );
53 0 : if( fd_sandbox_requires_cap_sys_admin( config->uid, config->gid ) )
54 0 : fd_cap_chk_cap( chk, NAME, CAP_SYS_ADMIN, "call `unshare(2)` with `CLONE_NEWUSER` to sandbox the process in a user namespace" );
55 0 : if( FD_LIKELY( getuid() != config->uid ) )
56 0 : fd_cap_chk_cap( chk, NAME, CAP_SETUID, "call `setresuid(2)` to switch uid to the sandbox user" );
57 0 : if( FD_LIKELY( getgid()!=config->gid ) )
58 0 : fd_cap_chk_cap( chk, NAME, CAP_SETGID, "call `setresgid(2)` to switch gid to the sandbox user" );
59 0 : if( FD_UNLIKELY( config->tiles.metric.prometheus_listen_port<1024 ) )
60 0 : fd_cap_chk_cap( chk, NAME, CAP_NET_BIND_SERVICE, "call `bind(2)` to bind to a privileged port for serving metrics" );
61 0 : if( FD_UNLIKELY( config->tiles.gui.gui_listen_port<1024 ) )
62 0 : fd_cap_chk_cap( chk, NAME, CAP_NET_BIND_SERVICE, "call `bind(2)` to bind to a privileged port for serving the GUI" );
63 0 : }
64 :
65 : struct pidns_clone_args {
66 : config_t const * config;
67 : int * pipefd;
68 : int closefd;
69 : };
70 :
71 : extern char fd_log_private_path[ 1024 ]; /* empty string on start */
72 :
73 : static pid_t pid_namespace;
74 :
75 0 : #define FD_LOG_ERR_NOEXIT(a) do { long _fd_log_msg_now = fd_log_wallclock(); fd_log_private_1( 4, _fd_log_msg_now, __FILE__, __LINE__, __func__, fd_log_private_0 a ); } while(0)
76 :
77 : extern int * fd_log_private_shared_lock;
78 :
79 : static void
80 0 : parent_signal( int sig ) {
81 0 : if( FD_LIKELY( pid_namespace ) ) kill( pid_namespace, SIGKILL );
82 :
83 : /* A pretty gross hack. For the local process, clear the lock so that
84 : we can always print the messages without waiting on another process,
85 : particularly if one of those processes might have just died. The
86 : signal handler is re-entrant so this also avoids a deadlock since
87 : the log lock is not re-entrant. */
88 0 : int lock = 0;
89 0 : fd_log_private_shared_lock = &lock;
90 :
91 0 : if( -1!=fd_log_private_logfile_fd() ) FD_LOG_ERR_NOEXIT(( "Received signal %s\nLog at \"%s\"", fd_io_strsignal( sig ), fd_log_private_path ));
92 0 : else FD_LOG_ERR_NOEXIT(( "Received signal %s", fd_io_strsignal( sig ) ));
93 :
94 0 : if( FD_LIKELY( sig==SIGINT ) ) fd_sys_util_exit_group( 128+SIGINT );
95 0 : else fd_sys_util_exit_group( 0 );
96 0 : }
97 :
98 : static void
99 0 : install_parent_signals( void ) {
100 0 : struct sigaction sa = {
101 0 : .sa_handler = parent_signal,
102 0 : .sa_flags = 0,
103 0 : };
104 0 : if( FD_UNLIKELY( sigaction( SIGTERM, &sa, NULL ) ) )
105 0 : FD_LOG_ERR(( "sigaction(SIGTERM) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
106 0 : if( FD_UNLIKELY( sigaction( SIGINT, &sa, NULL ) ) )
107 0 : FD_LOG_ERR(( "sigaction(SIGINT) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
108 :
109 0 : sa.sa_handler = SIG_IGN;
110 0 : if( FD_UNLIKELY( sigaction( SIGUSR1, &sa, NULL ) ) )
111 0 : FD_LOG_ERR(( "sigaction(SIGUSR1) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
112 0 : if( FD_UNLIKELY( sigaction( SIGUSR2, &sa, NULL ) ) )
113 0 : FD_LOG_ERR(( "sigaction(SIGUSR2) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
114 0 : }
115 :
116 : void *
117 0 : create_clone_stack( void ) {
118 0 : ulong mmap_sz = FD_TILE_PRIVATE_STACK_SZ + 2UL*FD_SHMEM_NORMAL_PAGE_SZ;
119 0 : uchar * stack = (uchar *)mmap( NULL, mmap_sz, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, (off_t)0 );
120 0 : if( FD_UNLIKELY( stack==MAP_FAILED ) )
121 0 : FD_LOG_ERR(( "mmap() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
122 :
123 : /* Make space for guard lo and guard hi */
124 0 : if( FD_UNLIKELY( munmap( stack, FD_SHMEM_NORMAL_PAGE_SZ ) ) )
125 0 : FD_LOG_ERR(( "munmap failed (%i-%s)", errno, fd_io_strerror( errno ) ));
126 0 : stack += FD_SHMEM_NORMAL_PAGE_SZ;
127 0 : if( FD_UNLIKELY( munmap( stack + FD_TILE_PRIVATE_STACK_SZ, FD_SHMEM_NORMAL_PAGE_SZ ) ) )
128 0 : FD_LOG_ERR(( "munmap failed (%i-%s)", errno, fd_io_strerror( errno ) ));
129 :
130 : /* Create the guard regions in the extra space */
131 0 : void * guard_lo = (void *)(stack - FD_SHMEM_NORMAL_PAGE_SZ );
132 0 : if( FD_UNLIKELY( mmap( guard_lo, FD_SHMEM_NORMAL_PAGE_SZ, PROT_NONE,
133 0 : MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, (off_t)0 )!=guard_lo ) )
134 0 : FD_LOG_ERR(( "mmap failed (%i-%s)", errno, fd_io_strerror( errno ) ));
135 :
136 0 : void * guard_hi = (void *)(stack + FD_TILE_PRIVATE_STACK_SZ);
137 0 : if( FD_UNLIKELY( mmap( guard_hi, FD_SHMEM_NORMAL_PAGE_SZ, PROT_NONE,
138 0 : MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1, (off_t)0 )!=guard_hi ) )
139 0 : FD_LOG_ERR(( "mmap failed (%i-%s)", errno, fd_io_strerror( errno ) ));
140 :
141 0 : return stack;
142 0 : }
143 :
144 :
145 : static int
146 : execve_agave( int config_memfd,
147 0 : int pipefd ) {
148 0 : if( FD_UNLIKELY( -1==fcntl( pipefd, F_SETFD, 0 ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,0) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
149 0 : pid_t child = fork();
150 0 : if( FD_UNLIKELY( -1==child ) ) FD_LOG_ERR(( "fork() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
151 0 : if( FD_LIKELY( !child ) ) {
152 0 : char _current_executable_path[ PATH_MAX ];
153 0 : FD_TEST( -1!=fd_file_util_self_exe( _current_executable_path ) );
154 :
155 0 : char config_fd[ 32 ];
156 0 : FD_TEST( fd_cstr_printf_check( config_fd, sizeof( config_fd ), NULL, "%d", config_memfd ) );
157 0 : char * args[ 5 ] = { _current_executable_path, "run-agave", "--config-fd", config_fd, NULL };
158 :
159 0 : char * envp[] = { NULL, NULL };
160 0 : char * google_creds = getenv( "GOOGLE_APPLICATION_CREDENTIALS" );
161 0 : char provide_creds[ PATH_MAX+30UL ];
162 0 : if( FD_UNLIKELY( google_creds ) ) {
163 0 : FD_TEST( fd_cstr_printf_check( provide_creds, sizeof( provide_creds ), NULL, "GOOGLE_APPLICATION_CREDENTIALS=%s", google_creds ) );
164 0 : envp[ 0 ] = provide_creds;
165 0 : }
166 :
167 0 : if( FD_UNLIKELY( -1==execve( _current_executable_path, args, envp ) ) ) FD_LOG_ERR(( "execve() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
168 0 : } else {
169 0 : if( FD_UNLIKELY( -1==fcntl( pipefd, F_SETFD, FD_CLOEXEC ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,FD_CLOEXEC) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
170 0 : return child;
171 0 : }
172 0 : return 0;
173 0 : }
174 :
175 : static pid_t
176 : execve_tile( fd_topo_tile_t const * tile,
177 : fd_cpuset_t const * floating_cpu_set,
178 : int floating_priority,
179 : int config_memfd,
180 0 : int pipefd ) {
181 0 : FD_CPUSET_DECL( cpu_set );
182 0 : if( FD_LIKELY( tile->cpu_idx!=ULONG_MAX ) ) {
183 : /* set the thread affinity before we clone the new process to ensure
184 : kernel first touch happens on the desired thread. */
185 0 : fd_cpuset_insert( cpu_set, tile->cpu_idx );
186 0 : if( FD_UNLIKELY( -1==setpriority( PRIO_PROCESS, 0, -19 ) ) ) FD_LOG_ERR(( "setpriority() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
187 0 : } else {
188 0 : fd_memcpy( cpu_set, floating_cpu_set, fd_cpuset_footprint() );
189 0 : if( FD_UNLIKELY( -1==setpriority( PRIO_PROCESS, 0, floating_priority ) ) ) FD_LOG_ERR(( "setpriority() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
190 0 : }
191 :
192 0 : if( FD_UNLIKELY( fd_cpuset_setaffinity( 0, cpu_set ) ) ) {
193 0 : if( FD_LIKELY( errno==EINVAL ) ) {
194 0 : FD_LOG_ERR(( "Unable to set the thread affinity for tile %s:%lu on cpu %lu. It is likely that the affinity "
195 0 : "you have specified for this tile in [layout.affinity] of your configuration file contains a "
196 0 : "CPU (%lu) which does not exist on this machine.",
197 0 : tile->name, tile->kind_id, tile->cpu_idx, tile->cpu_idx ));
198 0 : } else {
199 0 : FD_LOG_ERR(( "sched_setaffinity failed (%i-%s)", errno, fd_io_strerror( errno ) ));
200 0 : }
201 0 : }
202 :
203 : /* Clear CLOEXEC on the side of the pipe we want to pass to the tile. */
204 0 : if( FD_UNLIKELY( -1==fcntl( pipefd, F_SETFD, 0 ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,0) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
205 0 : pid_t child = fork();
206 0 : if( FD_UNLIKELY( -1==child ) ) FD_LOG_ERR(( "fork() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
207 0 : if( FD_LIKELY( !child ) ) {
208 0 : char _current_executable_path[ PATH_MAX ];
209 0 : FD_TEST( -1!=fd_file_util_self_exe( _current_executable_path ) );
210 :
211 0 : char kind_id[ 32 ], config_fd[ 32 ], pipe_fd[ 32 ];
212 0 : FD_TEST( fd_cstr_printf_check( kind_id, sizeof( kind_id ), NULL, "%lu", tile->kind_id ) );
213 0 : FD_TEST( fd_cstr_printf_check( config_fd, sizeof( config_fd ), NULL, "%d", config_memfd ) );
214 0 : FD_TEST( fd_cstr_printf_check( pipe_fd, sizeof( pipe_fd ), NULL, "%d", pipefd ) );
215 0 : char const * args[ 9 ] = { _current_executable_path, "run1", tile->name, kind_id, "--pipe-fd", pipe_fd, "--config-fd", config_fd, NULL };
216 0 : if( FD_UNLIKELY( -1==execve( _current_executable_path, (char **)args, NULL ) ) ) FD_LOG_ERR(( "execve() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
217 0 : } else {
218 0 : if( FD_UNLIKELY( -1==fcntl( pipefd, F_SETFD, FD_CLOEXEC ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,FD_CLOEXEC) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
219 0 : return child;
220 0 : }
221 0 : return 0;
222 0 : }
223 :
224 : extern int * fd_log_private_shared_lock;
225 :
226 : int
227 0 : main_pid_namespace( void * _args ) {
228 0 : struct pidns_clone_args * args = _args;
229 0 : if( FD_UNLIKELY( close( args->pipefd[ 0 ] ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
230 0 : if( FD_UNLIKELY( -1!=args->closefd ) ) {
231 0 : if( FD_UNLIKELY( close( args->closefd ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
232 0 : }
233 :
234 0 : config_t const * config = args->config;
235 :
236 0 : fd_log_thread_set( "pidns" );
237 0 : ulong pid = fd_sandbox_getpid(); /* Need to read /proc again.. we got a new PID from clone */
238 0 : fd_log_private_group_id_set( pid );
239 0 : fd_log_private_thread_id_set( pid );
240 0 : fd_log_private_stack_discover( FD_TILE_PRIVATE_STACK_SZ,
241 0 : &fd_tile_private_stack0, &fd_tile_private_stack1 );
242 :
243 0 : if( FD_UNLIKELY( !config->development.sandbox ) ) {
244 : /* If no sandbox, then there's no actual PID namespace so we can't
245 : wait() grandchildren for the exit code. Do this as a workaround. */
246 0 : if( FD_UNLIKELY( -1==prctl( PR_SET_CHILD_SUBREAPER, 1, 0, 0, 0 ) ) )
247 0 : FD_LOG_ERR(( "prctl(PR_SET_CHILD_SUBREAPER) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
248 0 : }
249 :
250 : /* Save the current affinity, it will be restored after creating any child tiles */
251 0 : FD_CPUSET_DECL( floating_cpu_set );
252 0 : if( FD_UNLIKELY( fd_cpuset_getaffinity( 0, floating_cpu_set ) ) )
253 0 : FD_LOG_ERR(( "fd_cpuset_getaffinity failed (%i-%s)", errno, fd_io_strerror( errno ) ));
254 :
255 0 : pid_t child_pids[ FD_TOPO_MAX_TILES+1 ];
256 0 : ulong actual_pids[ FD_TOPO_MAX_TILES+1 ];
257 0 : for( ulong i=0UL; i<FD_TOPO_MAX_TILES+1; i++ ) actual_pids[ i ] = ULONG_MAX;
258 0 : char child_names[ FD_TOPO_MAX_TILES+1 ][ 32 ];
259 0 : ulong child_idxs[ FD_TOPO_MAX_TILES+1 ];
260 0 : struct pollfd fds[ FD_TOPO_MAX_TILES+2 ];
261 :
262 0 : int config_memfd = fd_config_to_memfd( config );
263 0 : if( FD_UNLIKELY( -1==config_memfd ) ) FD_LOG_ERR(( "fd_config_to_memfd() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
264 :
265 0 : if( FD_UNLIKELY( config->development.debug_tile ) ) {
266 0 : fd_log_private_shared_lock[1] = 1;
267 0 : }
268 :
269 0 : ulong child_cnt = 0UL;
270 0 : if( FD_LIKELY( !config->is_firedancer && !config->development.no_agave ) ) {
271 0 : int pipefd[ 2 ];
272 0 : if( FD_UNLIKELY( pipe2( pipefd, O_CLOEXEC ) ) ) FD_LOG_ERR(( "pipe2() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
273 0 : fds[ child_cnt ] = (struct pollfd){ .fd = pipefd[ 0 ], .events = 0 };
274 0 : child_pids[ child_cnt ] = execve_agave( config_memfd, pipefd[ 1 ] );
275 0 : FD_TEST( child_pids[ child_cnt ]>0 );
276 0 : actual_pids[ child_cnt ] = (ulong)child_pids[ child_cnt ];
277 0 : child_idxs[ child_cnt ] = ULONG_MAX;
278 0 : if( FD_UNLIKELY( close( pipefd[ 1 ] ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
279 0 : strncpy( child_names[ child_cnt ], "agave", 32 );
280 0 : child_cnt++;
281 0 : }
282 :
283 0 : errno = 0;
284 0 : int save_priority = getpriority( PRIO_PROCESS, 0 );
285 0 : if( FD_UNLIKELY( -1==save_priority && errno ) ) FD_LOG_ERR(( "getpriority() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
286 :
287 0 : int need_xdp = 0==strcmp( config->net.provider, "xdp" );
288 0 : fd_xdp_fds_t xdp_fds[ FD_TOPO_XDP_FDS_MAX ];
289 0 : uint xdp_fds_cnt = FD_TOPO_XDP_FDS_MAX;
290 0 : if( need_xdp ) {
291 0 : fd_topo_install_xdp( &config->topo, xdp_fds, &xdp_fds_cnt, config->net.bind_address_parsed, 0 );
292 0 : }
293 :
294 0 : for( ulong i=0UL; i<config->topo.tile_cnt; i++ ) {
295 0 : fd_topo_tile_t const * tile = &config->topo.tiles[ i ];
296 0 : if( FD_UNLIKELY( tile->is_agave ) ) continue;
297 :
298 0 : if( need_xdp ) {
299 0 : if( FD_UNLIKELY( strcmp( tile->name, "net" ) ) ) {
300 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
301 : /* close XDP related file descriptors */
302 0 : if( FD_UNLIKELY( -1==fcntl( xdp_fds[i].xsk_map_fd, F_SETFD, FD_CLOEXEC ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,FD_CLOEXEC) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
303 0 : if( FD_UNLIKELY( -1==fcntl( xdp_fds[i].prog_link_fd, F_SETFD, FD_CLOEXEC ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,FD_CLOEXEC) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
304 0 : }
305 0 : } else {
306 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
307 0 : if( FD_UNLIKELY( -1==fcntl( xdp_fds[i].xsk_map_fd, F_SETFD, 0 ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,0) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
308 0 : if( FD_UNLIKELY( -1==fcntl( xdp_fds[i].prog_link_fd, F_SETFD, 0 ) ) ) FD_LOG_ERR(( "fcntl(F_SETFD,0) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
309 0 : }
310 0 : }
311 0 : }
312 :
313 0 : int pipefd[ 2 ];
314 0 : if( FD_UNLIKELY( pipe2( pipefd, O_CLOEXEC ) ) ) FD_LOG_ERR(( "pipe2() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
315 0 : fds[ child_cnt ] = (struct pollfd){ .fd = pipefd[ 0 ], .events = 0 };
316 0 : child_pids[ child_cnt ] = execve_tile( tile, floating_cpu_set, save_priority, config_memfd, pipefd[ 1 ] );
317 0 : child_idxs[ child_cnt ] = i;
318 0 : if( FD_UNLIKELY( close( pipefd[ 1 ] ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
319 0 : strncpy( child_names[ child_cnt ], tile->name, 32 );
320 0 : child_cnt++;
321 0 : }
322 :
323 : /* Obtain the actual grandchild PID from the pipe */
324 0 : for( ulong i=0UL; i<child_cnt; i++ ) {
325 0 : if( FD_UNLIKELY( actual_pids[ i ]!=ULONG_MAX ) ) continue;
326 0 : FD_TEST( 8UL==read( fds[ i ].fd, &actual_pids[ i ], 8UL ) );
327 0 : }
328 :
329 0 : if( FD_UNLIKELY( -1==setpriority( PRIO_PROCESS, 0, save_priority ) ) ) FD_LOG_ERR(( "setpriority() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
330 0 : if( FD_UNLIKELY( fd_cpuset_setaffinity( 0, floating_cpu_set ) ) )
331 0 : FD_LOG_ERR(( "fd_cpuset_setaffinity failed (%i-%s)", errno, fd_io_strerror( errno ) ));
332 :
333 0 : if( FD_UNLIKELY( close( config_memfd ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
334 0 : if( FD_UNLIKELY( close( config->log.lock_fd ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
335 0 : if( need_xdp ) {
336 0 : for( uint i=0U; i<xdp_fds_cnt; i++ ) {
337 0 : if( FD_UNLIKELY( close( xdp_fds[i].xsk_map_fd ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
338 0 : if( FD_UNLIKELY( close( xdp_fds[i].prog_link_fd ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
339 0 : }
340 0 : }
341 :
342 0 : int allow_fds[ 4+FD_TOPO_MAX_TILES ];
343 0 : ulong allow_fds_cnt = 0;
344 0 : allow_fds[ allow_fds_cnt++ ] = 2; /* stderr */
345 0 : if( FD_LIKELY( fd_log_private_logfile_fd()!=-1 ) )
346 0 : allow_fds[ allow_fds_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
347 0 : allow_fds[ allow_fds_cnt++ ] = args->pipefd[ 1 ]; /* write end of main pipe */
348 0 : for( ulong i=0UL; i<child_cnt; i++ )
349 0 : allow_fds[ allow_fds_cnt++ ] = fds[ i ].fd; /* read end of child pipes */
350 :
351 0 : struct sock_filter seccomp_filter[ 128UL ];
352 0 : unsigned int instr_cnt;
353 : #if defined(__aarch64__)
354 : populate_sock_filter_policy_pidns_arm64( 128UL, seccomp_filter, (uint)fd_log_private_logfile_fd() );
355 : instr_cnt = sock_filter_policy_pidns_arm64_instr_cnt;
356 : #else
357 0 : populate_sock_filter_policy_pidns( 128UL, seccomp_filter, (uint)fd_log_private_logfile_fd() );
358 0 : instr_cnt = sock_filter_policy_pidns_instr_cnt;
359 0 : #endif
360 :
361 0 : if( FD_LIKELY( config->development.sandbox ) ) {
362 0 : fd_sandbox_enter( config->uid,
363 0 : config->gid,
364 0 : 0,
365 0 : 0,
366 0 : 0,
367 0 : 0,
368 0 : 0,
369 0 : 1UL+child_cnt, /* RLIMIT_NOFILE needs to be set to the nfds argument of poll() */
370 0 : 0UL,
371 0 : 0UL,
372 0 : 0UL,
373 0 : allow_fds_cnt,
374 0 : allow_fds,
375 0 : instr_cnt,
376 0 : seccomp_filter );
377 0 : } else {
378 0 : fd_sandbox_switch_uid_gid( config->uid, config->gid );
379 0 : }
380 :
381 : /* The supervsior process should not share the log lock, because a
382 : child process might die while holding it and we still need to
383 : reap and print errors. */
384 0 : int lock = 0;
385 0 : fd_log_private_shared_lock = &lock;
386 :
387 : /* Reap child process PIDs so they don't show up in `ps` etc. All of
388 : these children should have exited immediately after clone(2)'ing
389 : another child with a huge page based stack. */
390 0 : for( ulong i=0UL; i<child_cnt; i++ ) {
391 0 : int wstatus;
392 0 : int exited_pid = wait4( child_pids[ i ], &wstatus, (int)__WALL, NULL );
393 0 : if( FD_UNLIKELY( -1==exited_pid ) ) {
394 0 : FD_LOG_ERR(( "pidns wait4() failed (%i-%s) %lu %hu", errno, fd_io_strerror( errno ), i, fds[i].revents ));
395 0 : } else if( FD_UNLIKELY( child_pids[ i ]!=exited_pid ) ) {
396 0 : FD_LOG_ERR(( "pidns wait4() returned unexpected pid %d %d", child_pids[ i ], exited_pid ));
397 0 : } else if( FD_UNLIKELY( !WIFEXITED( wstatus ) ) ) {
398 0 : FD_LOG_ERR_NOEXIT(( "tile %lu (%s) exited while booting with signal %d (%s)\n", i, child_names[ i ], WTERMSIG( wstatus ), fd_io_strsignal( WTERMSIG( wstatus ) ) ));
399 0 : fd_sys_util_exit_group( WTERMSIG( wstatus ) ? WTERMSIG( wstatus ) : 1 );
400 0 : }
401 0 : if( FD_UNLIKELY( WEXITSTATUS( wstatus ) ) ) {
402 0 : FD_LOG_ERR_NOEXIT(( "tile %lu (%s) exited while booting with code %d\n", i, child_names[ i ], WEXITSTATUS( wstatus ) ));
403 0 : fd_sys_util_exit_group( WEXITSTATUS( wstatus ) ? WEXITSTATUS( wstatus ) : 1 );
404 0 : }
405 0 : }
406 :
407 0 : fds[ child_cnt ] = (struct pollfd){ .fd = args->pipefd[ 1 ], .events = 0 };
408 0 : strncpy( child_names[ child_cnt ], "parent", 32UL );
409 0 : child_idxs[ child_cnt ] = ULONG_MAX;
410 :
411 : /* We are now the init process of the pid namespace. If the init
412 : process dies, all children are terminated. If any child dies, we
413 : terminate the init process, which will cause the kernel to
414 : terminate all other children bringing all of our processes down as
415 : a group. The parent process will also die if this process dies,
416 : due to getting SIGHUP on the pipe. */
417 0 : while( 1 ) {
418 0 : if( FD_UNLIKELY( -1==poll( fds, 1UL+child_cnt, -1 ) ) ) FD_LOG_ERR(( "poll() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
419 :
420 : /* Parent process died, probably SIGINT, exit gracefully. */
421 0 : if( FD_UNLIKELY( fds[ child_cnt ].revents ) ) fd_sys_util_exit_group( 0 );
422 :
423 : /* Child process died, reap it to figure out exit code. */
424 0 : int wstatus;
425 0 : int exited_pid = wait4( -1, &wstatus, (int)__WALL | (int)WNOHANG, NULL );
426 0 : if( FD_UNLIKELY( -1==exited_pid ) ) {
427 0 : FD_LOG_ERR(( "pidns wait4() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
428 0 : } else if( FD_UNLIKELY( !exited_pid ) ) {
429 : /* Spurious wakeup, no child actually dead yet. */
430 0 : continue;
431 0 : }
432 :
433 : /* Now find the tile corresponding to that PID */
434 0 : FD_TEST( exited_pid>0 );
435 0 : int found = 0;
436 0 : for( ulong i=0UL; i<child_cnt; i++ ) {
437 0 : if( FD_LIKELY( actual_pids[ i ]!=(ulong)exited_pid ) ) continue;
438 :
439 0 : found = 1;
440 0 : fds[ i ].fd = -1; /* Don't poll on this tile anymore */
441 :
442 0 : char * tile_name = child_names[ i ];
443 0 : ulong tile_idx = child_idxs[ i ];
444 0 : ulong tile_id = config->topo.tiles[ tile_idx ].kind_id;
445 :
446 0 : if( FD_UNLIKELY( !WIFEXITED( wstatus ) ) ) {
447 0 : FD_LOG_ERR_NOEXIT(( "tile %s:%lu exited with signal %d (%s)", tile_name, tile_id, WTERMSIG( wstatus ), fd_io_strsignal( WTERMSIG( wstatus ) ) ));
448 0 : fd_sys_util_exit_group( WTERMSIG( wstatus ) ? WTERMSIG( wstatus ) : 1 );
449 0 : } else {
450 0 : int exit_code = WEXITSTATUS( wstatus );
451 0 : if( FD_LIKELY( !exit_code && tile_idx!=ULONG_MAX && config->topo.tiles[ tile_idx ].allow_shutdown ) ) {
452 0 : found = 1;
453 0 : FD_LOG_INFO(( "tile %s:%lu exited gracefully with code %d", tile_name, tile_id, exit_code ));
454 0 : } else {
455 0 : FD_LOG_ERR_NOEXIT(( "tile %s:%lu exited with code %d", tile_name, tile_id, exit_code ));
456 0 : fd_sys_util_exit_group( exit_code ? exit_code : 1 );
457 0 : }
458 0 : }
459 0 : }
460 :
461 0 : if( FD_UNLIKELY( !found ) ) FD_LOG_ERR(( "wait4() returned unexpected pid %d", exited_pid ));
462 0 : }
463 :
464 0 : return 0;
465 0 : }
466 :
467 : int
468 : clone_firedancer( config_t const * config,
469 : int close_fd,
470 0 : int * out_pipe ) {
471 : /* This pipe is here just so that the child process knows when the
472 : parent has died (it will get a HUP). */
473 0 : int pipefd[2];
474 0 : if( FD_UNLIKELY( pipe2( pipefd, O_CLOEXEC | O_NONBLOCK ) ) ) FD_LOG_ERR(( "pipe2() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
475 :
476 : /* clone into a pid namespace */
477 0 : int flags = config->development.sandbox ? CLONE_NEWPID : 0;
478 0 : struct pidns_clone_args args = { .config = config, .closefd = close_fd, .pipefd = pipefd, };
479 :
480 0 : void * stack = create_clone_stack();
481 :
482 0 : int pid_namespace = clone( main_pid_namespace, (uchar *)stack + FD_TILE_PRIVATE_STACK_SZ, flags, &args );
483 0 : if( FD_UNLIKELY( pid_namespace<0 ) ) FD_LOG_ERR(( "clone() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
484 :
485 0 : if( FD_UNLIKELY( close( pipefd[ 1 ] ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
486 :
487 0 : *out_pipe = pipefd[ 0 ];
488 0 : return pid_namespace;
489 0 : }
490 :
491 : static void
492 : workspace_path( config_t const * config,
493 : fd_topo_wksp_t const * wksp,
494 0 : char out[ PATH_MAX ] ) {
495 0 : char const * mount_path;
496 0 : switch( wksp->page_sz ) {
497 0 : case FD_SHMEM_HUGE_PAGE_SZ:
498 0 : mount_path = config->hugetlbfs.huge_page_mount_path;
499 0 : break;
500 0 : case FD_SHMEM_GIGANTIC_PAGE_SZ:
501 0 : mount_path = config->hugetlbfs.gigantic_page_mount_path;
502 0 : break;
503 0 : case FD_SHMEM_NORMAL_PAGE_SZ:
504 0 : mount_path = config->hugetlbfs.normal_page_mount_path;
505 0 : break;
506 0 : default:
507 0 : FD_LOG_ERR(( "invalid page size %lu", wksp->page_sz ));
508 0 : }
509 :
510 0 : FD_TEST( fd_cstr_printf_check( out, PATH_MAX, NULL, "%s/%s_%s.wksp", mount_path, config->name, wksp->name ) );
511 0 : }
512 :
513 : static void
514 : warn_unknown_files( config_t const * config,
515 0 : ulong mount_type ) {
516 0 : char const * mount_path;
517 0 : switch( mount_type ) {
518 0 : case 0UL:
519 0 : mount_path = config->hugetlbfs.huge_page_mount_path;
520 0 : break;
521 0 : case 1UL:
522 0 : mount_path = config->hugetlbfs.gigantic_page_mount_path;
523 0 : break;
524 0 : default:
525 0 : FD_LOG_ERR(( "invalid mount type %lu", mount_type ));
526 0 : }
527 :
528 : /* Check if there are any files in mount_path */
529 0 : DIR * dir = opendir( mount_path );
530 0 : if( FD_UNLIKELY( !dir ) ) {
531 0 : if( FD_UNLIKELY( errno!=ENOENT ) ) FD_LOG_ERR(( "error opening `%s` (%i-%s)", mount_path, errno, fd_io_strerror( errno ) ));
532 0 : return;
533 0 : }
534 :
535 0 : struct dirent * entry;
536 0 : while(( FD_LIKELY( entry = readdir( dir ) ) )) {
537 0 : if( FD_UNLIKELY( !strcmp( entry->d_name, ".") || !strcmp( entry->d_name, ".." ) ) ) continue;
538 :
539 0 : char entry_path[ PATH_MAX ];
540 0 : FD_TEST( fd_cstr_printf_check( entry_path, PATH_MAX, NULL, "%s/%s", mount_path, entry->d_name ));
541 :
542 0 : int known_file = 0;
543 0 : for( ulong i=0UL; i<config->topo.wksp_cnt; i++ ) {
544 0 : fd_topo_wksp_t const * wksp = &config->topo.workspaces[ i ];
545 :
546 0 : char expected_path[ PATH_MAX ];
547 0 : workspace_path( config, wksp, expected_path );
548 :
549 0 : if( !strcmp( entry_path, expected_path ) ) {
550 0 : known_file = 1;
551 0 : break;
552 0 : }
553 0 : }
554 :
555 0 : if( mount_type==0UL ) {
556 0 : for( ulong i=0UL; i<config->topo.tile_cnt; i++ ) {
557 0 : fd_topo_tile_t const * tile = &config->topo.tiles [ i ];
558 :
559 0 : char expected_path[ PATH_MAX ];
560 0 : FD_TEST( fd_cstr_printf_check( expected_path, PATH_MAX, NULL, "%s/%s_stack_%s%lu", config->hugetlbfs.huge_page_mount_path, config->name, tile->name, tile->kind_id ) );
561 :
562 0 : if( !strcmp( entry_path, expected_path ) ) {
563 0 : known_file = 1;
564 0 : break;
565 0 : }
566 0 : }
567 0 : }
568 :
569 0 : if( FD_UNLIKELY( !known_file ) ) FD_LOG_WARNING(( "unknown file `%s` found in `%s`", entry->d_name, mount_path ));
570 0 : }
571 :
572 0 : if( FD_UNLIKELY( errno && errno!=ENOENT ) ) FD_LOG_ERR(( "error reading dir `%s` (%i-%s)", mount_path, errno, fd_io_strerror( errno ) ));
573 0 : if( FD_UNLIKELY( closedir( dir ) ) ) FD_LOG_ERR(( "error closing `%s` (%i-%s)", mount_path, errno, fd_io_strerror( errno ) ));
574 0 : }
575 :
576 : void
577 0 : initialize_workspaces( config_t * config ) {
578 : /* Switch to non-root uid/gid for workspace creation. Permissions
579 : checks are still done as the current user. */
580 0 : uint gid = getgid();
581 0 : uint uid = getuid();
582 0 : if( FD_LIKELY( gid!=config->gid && -1==setegid( config->gid ) ) )
583 0 : FD_LOG_ERR(( "setegid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
584 0 : if( FD_LIKELY( uid!=config->uid && -1==seteuid( config->uid ) ) )
585 0 : FD_LOG_ERR(( "seteuid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
586 :
587 0 : for( ulong i=0UL; i<config->topo.wksp_cnt; i++ ) {
588 0 : fd_topo_wksp_t * wksp = &config->topo.workspaces[ i ];
589 :
590 0 : char path[ PATH_MAX ];
591 0 : workspace_path( config, wksp, path );
592 :
593 0 : struct stat st;
594 0 : int result = stat( path, &st );
595 :
596 0 : int update_existing;
597 0 : if( FD_UNLIKELY( !result && config->is_live_cluster ) ) {
598 0 : if( FD_UNLIKELY( -1==unlink( path ) && errno!=ENOENT ) ) FD_LOG_ERR(( "unlink() failed when trying to create workspace `%s` (%i-%s)", path, errno, fd_io_strerror( errno ) ));
599 0 : update_existing = 0;
600 0 : } else if( FD_UNLIKELY( !result ) ) {
601 : /* Creating all of the workspaces is very expensive because the
602 : kernel has to zero out all of the pages. There can be tens or
603 : hundreds of gigabytes of zeroing to do.
604 :
605 : What would be really nice is if the kernel let us create huge
606 : pages without zeroing them, but it's not possible. The
607 : ftruncate and fallocate calls do not support this type of
608 : resize with the hugetlbfs filesystem.
609 :
610 : Instead.. to prevent repeatedly doing this zeroing every time
611 : we start the validator, we have a small hack here to re-use the
612 : workspace files if they exist. */
613 0 : update_existing = 1;
614 0 : } else if( FD_LIKELY( result && errno==ENOENT ) ) {
615 0 : update_existing = 0;
616 0 : } else {
617 0 : FD_LOG_ERR(( "stat failed when trying to create workspace `%s` (%i-%s)", path, errno, fd_io_strerror( errno ) ));
618 0 : }
619 :
620 0 : if( FD_UNLIKELY( -1==fd_topo_create_workspace( &config->topo, wksp, update_existing ) ) ) {
621 0 : FD_TEST( errno==ENOMEM );
622 :
623 0 : warn_unknown_files( config, wksp->page_sz!=FD_SHMEM_HUGE_PAGE_SZ );
624 :
625 0 : char path[ PATH_MAX ];
626 0 : workspace_path( config, wksp, path );
627 0 : FD_LOG_ERR(( "ENOMEM-Out of memory when trying to create workspace `%s` at `%s` "
628 0 : "with %lu %s pages. Firedancer reserves enough memory for all of its workspaces "
629 0 : "during the `hugetlbfs` configure step, so it is likely you have unknown files "
630 0 : "left over in this directory which are consuming memory, or another program on "
631 0 : "the system is using pages from the same mount.",
632 0 : wksp->name, path, wksp->page_cnt, fd_shmem_page_sz_to_cstr( wksp->page_sz ) ));
633 0 : }
634 0 : fd_topo_join_workspace( &config->topo, wksp, FD_SHMEM_JOIN_MODE_READ_WRITE, 0 );
635 0 : fd_topo_wksp_new( &config->topo, wksp, CALLBACKS );
636 0 : fd_topo_leave_workspace( &config->topo, wksp );
637 0 : }
638 :
639 0 : if( FD_UNLIKELY( seteuid( uid ) ) ) FD_LOG_ERR(( "seteuid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
640 0 : if( FD_UNLIKELY( setegid( gid ) ) ) FD_LOG_ERR(( "setegid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
641 0 : }
642 :
643 : void
644 0 : initialize_stacks( config_t const * config ) {
645 : # if FD_HAS_MSAN
646 : /* MSan calls an external symbolizer using fork() on crashes, which is
647 : incompatible with Firedancer's MAP_SHARED stacks. */
648 : (void)config;
649 : return;
650 : # endif
651 :
652 : /* Switch to non-root uid/gid for workspace creation. Permissions
653 : checks are still done as the current user. */
654 0 : uint gid = getgid();
655 0 : uint uid = getuid();
656 0 : if( FD_LIKELY( gid!=config->gid && -1==setegid( config->gid ) ) )
657 0 : FD_LOG_ERR(( "setegid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
658 0 : if( FD_LIKELY( uid!=config->uid && -1==seteuid( config->uid ) ) )
659 0 : FD_LOG_ERR(( "seteuid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
660 :
661 0 : for( ulong i=0UL; i<config->topo.tile_cnt; i++ ) {
662 0 : fd_topo_tile_t const * tile = &config->topo.tiles[ i ];
663 :
664 0 : char path[ PATH_MAX ];
665 0 : FD_TEST( fd_cstr_printf_check( path, PATH_MAX, NULL, "%s/%s_stack_%s%lu", config->hugetlbfs.huge_page_mount_path, config->name, tile->name, tile->kind_id ) );
666 :
667 0 : struct stat st;
668 0 : int result = stat( path, &st );
669 :
670 0 : int update_existing;
671 0 : if( FD_UNLIKELY( !result && config->is_live_cluster ) ) {
672 0 : if( FD_UNLIKELY( -1==unlink( path ) && errno!=ENOENT ) ) FD_LOG_ERR(( "unlink() failed when trying to create stack workspace `%s` (%i-%s)", path, errno, fd_io_strerror( errno ) ));
673 0 : update_existing = 0;
674 0 : } else if( FD_UNLIKELY( !result ) ) {
675 : /* See above note about zeroing out pages. */
676 0 : update_existing = 1;
677 0 : } else if( FD_LIKELY( result && errno==ENOENT ) ) {
678 0 : update_existing = 0;
679 0 : } else {
680 0 : FD_LOG_ERR(( "stat failed when trying to create workspace `%s` (%i-%s)", path, errno, fd_io_strerror( errno ) ));
681 0 : }
682 :
683 : /* TODO: Use a better CPU idx for the stack if tile is floating */
684 0 : ulong stack_cpu_idx = 0UL;
685 0 : if( FD_LIKELY( tile->cpu_idx<65535UL ) ) stack_cpu_idx = tile->cpu_idx;
686 :
687 0 : char name[ PATH_MAX ];
688 0 : FD_TEST( fd_cstr_printf_check( name, PATH_MAX, NULL, "%s_stack_%s%lu", config->name, tile->name, tile->kind_id ) );
689 :
690 0 : ulong sub_page_cnt[ 1 ] = { 6 };
691 0 : ulong sub_cpu_idx [ 1 ] = { stack_cpu_idx };
692 0 : int err;
693 0 : if( FD_UNLIKELY( update_existing ) ) {
694 0 : err = fd_shmem_update_multi( name, FD_SHMEM_HUGE_PAGE_SZ, 1, sub_page_cnt, sub_cpu_idx, S_IRUSR | S_IWUSR ); /* logs details */
695 0 : } else {
696 0 : err = fd_shmem_create_multi( name, FD_SHMEM_HUGE_PAGE_SZ, 1, sub_page_cnt, sub_cpu_idx, S_IRUSR | S_IWUSR ); /* logs details */
697 0 : }
698 0 : if( FD_UNLIKELY( err && errno==ENOMEM ) ) {
699 0 : warn_unknown_files( config, 0UL );
700 :
701 0 : char path[ PATH_MAX ];
702 0 : FD_TEST( fd_cstr_printf_check( path, PATH_MAX, NULL, "%s/%s_stack_%s%lu", config->hugetlbfs.huge_page_mount_path, config->name, tile->name, tile->kind_id ) );
703 0 : FD_LOG_ERR(( "ENOMEM-Out of memory when trying to create huge page stack for tile `%s` at `%s`. "
704 0 : "Firedancer reserves enough memory for all of its stacks during the `hugetlbfs` configure "
705 0 : "step, so it is likely you have unknown files left over in this directory which are "
706 0 : "consuming memory, or another program on the system is using pages from the same mount.",
707 0 : tile->name, path ));
708 0 : } else if( FD_UNLIKELY( err ) ) FD_LOG_ERR(( "fd_shmem_create_multi failed" ));
709 0 : }
710 :
711 0 : if( FD_UNLIKELY( seteuid( uid ) ) ) FD_LOG_ERR(( "seteuid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
712 0 : if( FD_UNLIKELY( setegid( gid ) ) ) FD_LOG_ERR(( "setegid() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
713 0 : }
714 :
715 : void
716 0 : fdctl_check_configure( config_t const * config ) {
717 0 : configure_result_t check = fd_cfg_stage_hugetlbfs.check( config, FD_CONFIGURE_CHECK_TYPE_RUN );
718 0 : if( FD_UNLIKELY( check.result!=CONFIGURE_OK ) )
719 0 : FD_LOG_ERR(( "Huge pages are not configured correctly: %s. You can run `fdctl configure init hugetlbfs` "
720 0 : "to create the mounts correctly. This must be done after every system restart before running "
721 0 : "Firedancer.", check.message ));
722 :
723 0 : if( FD_LIKELY( 0==strcmp( config->net.provider, "xdp" ) ) ) {
724 0 : if( fd_cfg_stage_bonding.enabled( config ) ) {
725 0 : check = fd_cfg_stage_bonding.check( config, FD_CONFIGURE_CHECK_TYPE_RUN );
726 0 : if( FD_UNLIKELY( check.result!=CONFIGURE_OK ) )
727 0 : FD_LOG_ERR(( "Bonded network device is not configured correctly: %s. You can run `fdctl configure init bonding` "
728 0 : "to configure the bonding driver.", check.message ));
729 0 : }
730 :
731 0 : check = fd_cfg_stage_ethtool_channels.check( config, FD_CONFIGURE_CHECK_TYPE_RUN );
732 0 : if( FD_UNLIKELY( check.result!=CONFIGURE_OK ) )
733 0 : FD_LOG_ERR(( "Network %s. You can run `fdctl configure init ethtool-channels` to set the number of channels on the "
734 0 : "network device correctly.", check.message ));
735 :
736 0 : check = fd_cfg_stage_ethtool_offloads.check( config, FD_CONFIGURE_CHECK_TYPE_RUN );
737 0 : if( FD_UNLIKELY( check.result!=CONFIGURE_OK ) )
738 0 : FD_LOG_ERR(( "Network %s. You can run `fdctl configure init ethtool-offloads` to disable features "
739 0 : "as required.", check.message ));
740 :
741 0 : check = fd_cfg_stage_ethtool_loopback.check( config, FD_CONFIGURE_CHECK_TYPE_RUN );
742 0 : if( FD_UNLIKELY( check.result!=CONFIGURE_OK ) )
743 0 : FD_LOG_ERR(( "Network %s. You can run `fdctl configure init ethtool-loopback` to disable tx-udp-segmentation "
744 0 : "on the loopback device.", check.message ));
745 0 : }
746 :
747 0 : check = fd_cfg_stage_sysctl.check( config, FD_CONFIGURE_CHECK_TYPE_RUN );
748 0 : if( FD_UNLIKELY( check.result!=CONFIGURE_OK ) )
749 0 : FD_LOG_ERR(( "Kernel parameters are not configured correctly: %s. You can run `fdctl configure init sysctl` "
750 0 : "to set kernel parameters correctly.", check.message ));
751 :
752 0 : check = fd_cfg_stage_hyperthreads.check( config, FD_CONFIGURE_CHECK_TYPE_RUN );
753 0 : if( FD_UNLIKELY( check.result!=CONFIGURE_OK ) )
754 0 : FD_LOG_ERR(( "Hyperthreading is not configured correctly: %s. You can run `fdctl configure init hyperthreads` "
755 0 : "to configure hyperthreading correctly.", check.message ));
756 0 : }
757 :
758 : void
759 : run_firedancer_init( config_t * config,
760 : int init_workspaces,
761 0 : int check_configure ) {
762 0 : struct stat st;
763 0 : int err = stat( config->paths.identity_key, &st );
764 0 : if( FD_UNLIKELY( -1==err && errno==ENOENT ) ) FD_LOG_ERR(( "[consensus.identity_path] key does not exist `%s`. You can generate an identity key at this path by running `fdctl keys new %s --config <toml>`", config->paths.identity_key, config->paths.identity_key ));
765 0 : else if( FD_UNLIKELY( -1==err ) ) FD_LOG_ERR(( "could not stat [consensus.identity_path] `%s` (%i-%s)", config->paths.identity_key, errno, fd_io_strerror( errno ) ));
766 :
767 0 : if( FD_UNLIKELY( !config->is_firedancer ) ) {
768 0 : for( ulong i=0UL; i<config->frankendancer.paths.authorized_voter_paths_cnt; i++ ) {
769 0 : err = stat( config->frankendancer.paths.authorized_voter_paths[ i ], &st );
770 0 : if( FD_UNLIKELY( -1==err && errno==ENOENT ) ) FD_LOG_ERR(( "[consensus.authorized_voter_paths] key does not exist `%s`", config->frankendancer.paths.authorized_voter_paths[ i ] ));
771 0 : else if( FD_UNLIKELY( -1==err ) ) FD_LOG_ERR(( "could not stat [consensus.authorized_voter_paths] `%s` (%i-%s)", config->frankendancer.paths.authorized_voter_paths[ i ], errno, fd_io_strerror( errno ) ));
772 0 : }
773 0 : }
774 :
775 : /* FIXME: fdctl_check_configure unconditionally checks for network
776 : stack prerequisites even if the command being run does not
777 : require networking. Hack around that here for now. */
778 0 : if( check_configure ) fdctl_check_configure( config );
779 0 : if( FD_LIKELY( init_workspaces ) ) initialize_workspaces( config );
780 0 : initialize_stacks( config );
781 0 : }
782 :
783 : /* The boot sequence is a little bit involved...
784 :
785 : A process tree is created that looks like,
786 :
787 : + main
788 : +-- pidns
789 : +-- agave
790 : +-- tile 0
791 : +-- tile 1
792 : ...
793 :
794 : What we want is that if any process in the tree dies, all other
795 : processes will also die. This is done as follows,
796 :
797 : (a) pidns is the init process of a PID namespace, so if it dies the
798 : kernel will terminate the child processes.
799 :
800 : (b) main is the parent of pidns, so it can issue a waitpid() on the
801 : child PID, and when it completes terminate itself.
802 :
803 : (c) pidns is the parent of agave and the tiles, so it could
804 : issue a waitpid() of -1 to wait for any of them to terminate,
805 : but how would it know if main has died?
806 :
807 : (d) main creates a pipe, and passes the write end to pidns. If main
808 : dies, the pipe will be closed, and pidns will get a HUP on the
809 : read end. Then pidns creates a pipe per child and passes the
810 : write end to the child. If any of the children die, the pipe
811 : will be closed, and pidns will get a HUP on the read end.
812 :
813 : Then pidns can call poll() on both the write end of the main
814 : pipe and the read end of all the child pipes. If any of them
815 : raises SIGHUP, then pidns knows that the parent or a child has
816 : died, and it can terminate itself, which due to (a) and (b)
817 : will kill all other processes. */
818 : void
819 : run_firedancer( config_t * config,
820 : int parent_pipefd,
821 0 : int init_workspaces ) {
822 : /* dump the topology we are using to the output log */
823 0 : fd_topo_print_log( 0, &config->topo );
824 :
825 0 : run_firedancer_init( config, init_workspaces, 1 );
826 :
827 0 : #if defined(__x86_64__) || defined(__aarch64__)
828 :
829 : #ifndef SYS_landlock_create_ruleset
830 : #define SYS_landlock_create_ruleset 444
831 : #endif
832 :
833 0 : #ifndef LANDLOCK_CREATE_RULESET_VERSION
834 0 : #define LANDLOCK_CREATE_RULESET_VERSION (1U << 0)
835 0 : #endif
836 :
837 0 : #endif
838 0 : long abi = syscall( SYS_landlock_create_ruleset, NULL, 0, LANDLOCK_CREATE_RULESET_VERSION );
839 0 : if( -1L==abi && (errno==ENOSYS || errno==EOPNOTSUPP ) ) {
840 0 : FD_LOG_WARNING(( "The Landlock access control system is not supported by your Linux kernel. Firedancer uses landlock to "
841 0 : "provide an additional layer of security to the sandbox, but it is not required." ));
842 0 : }
843 :
844 0 : if( FD_UNLIKELY( close( 0 ) ) ) FD_LOG_ERR(( "close(0) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
845 0 : if( FD_UNLIKELY( fd_log_private_logfile_fd()!=1 && close( 1 ) ) ) FD_LOG_ERR(( "close(1) failed (%i-%s)", errno, fd_io_strerror( errno ) ));
846 :
847 0 : int pipefd;
848 0 : pid_namespace = clone_firedancer( config, parent_pipefd, &pipefd );
849 :
850 : /* Print the location of the logfile on SIGINT or SIGTERM, and also
851 : kill the child. They are connected by a pipe which the child is
852 : polling so we don't strictly need to kill the child, but its helpful
853 : to do that before printing the log location line, else it might
854 : get interleaved due to timing windows in the shutdown. */
855 0 : install_parent_signals();
856 :
857 0 : if( FD_UNLIKELY( close( config->log.lock_fd ) ) ) FD_LOG_ERR(( "close() failed (%i-%s)", errno, fd_io_strerror( errno ) ));
858 :
859 0 : struct sock_filter seccomp_filter[ 128UL ];
860 0 : populate_sock_filter_policy_main( 128UL, seccomp_filter, (uint)fd_log_private_logfile_fd(), (uint)pid_namespace );
861 :
862 0 : int allow_fds[ 4 ];
863 0 : ulong allow_fds_cnt = 0;
864 0 : allow_fds[ allow_fds_cnt++ ] = 2; /* stderr */
865 0 : if( FD_LIKELY( fd_log_private_logfile_fd()!=-1 ) )
866 0 : allow_fds[ allow_fds_cnt++ ] = fd_log_private_logfile_fd(); /* logfile, or maybe stdout */
867 0 : allow_fds[ allow_fds_cnt++ ] = pipefd; /* read end of main pipe */
868 0 : if( FD_UNLIKELY( parent_pipefd!=-1 ) )
869 0 : allow_fds[ allow_fds_cnt++ ] = parent_pipefd; /* write end of parent pipe */
870 :
871 0 : if( FD_LIKELY( config->development.sandbox ) ) {
872 0 : fd_sandbox_enter( config->uid,
873 0 : config->gid,
874 0 : 0,
875 0 : 0,
876 0 : 0,
877 0 : 1, /* Keep controlling terminal for main so it can receive Ctrl+C */
878 0 : 0,
879 0 : 0UL,
880 0 : 0UL,
881 0 : 0UL,
882 0 : 0UL,
883 0 : allow_fds_cnt,
884 0 : allow_fds,
885 0 : sock_filter_policy_main_instr_cnt,
886 0 : seccomp_filter );
887 0 : } else {
888 0 : fd_sandbox_switch_uid_gid( config->uid, config->gid );
889 0 : }
890 :
891 : /* The supervsior process should not share the log lock, because a
892 : child process might die while holding it and we still need to
893 : reap and print errors. */
894 0 : int lock = 0;
895 0 : fd_log_private_shared_lock = &lock;
896 :
897 : /* the only clean way to exit is SIGINT or SIGTERM on this parent process,
898 : so if wait4() completes, it must be an error */
899 0 : int wstatus;
900 0 : if( FD_UNLIKELY( -1==wait4( pid_namespace, &wstatus, (int)__WALL, NULL ) ) )
901 0 : FD_LOG_ERR(( "main wait4() failed (%i-%s)\nLog at \"%s\"", errno, fd_io_strerror( errno ), fd_log_private_path ));
902 :
903 0 : if( FD_UNLIKELY( WIFSIGNALED( wstatus ) ) ) fd_sys_util_exit_group( WTERMSIG( wstatus ) ? WTERMSIG( wstatus ) : 1 );
904 0 : else fd_sys_util_exit_group( WEXITSTATUS( wstatus ) ? WEXITSTATUS( wstatus ) : 1 );
905 0 : }
906 :
907 : void
908 : run_cmd_fn( args_t * args FD_PARAM_UNUSED,
909 0 : config_t * config ) {
910 0 : #define CHECK_PORT_NON_ZERO( field ) \
911 0 : if( FD_UNLIKELY( config->field==0 ) ) { \
912 0 : FD_LOG_ERR(( #field " is not set in your configuration file. Please set it to a non-zero value." )); \
913 0 : }
914 :
915 0 : if( FD_UNLIKELY( !config->gossip.entrypoints_cnt && !config->development.bootstrap ) )
916 0 : FD_LOG_ERR(( "No entrypoints specified in configuration file under [gossip.entrypoints], but "
917 0 : "at least one is needed to determine how to connect to the Solana cluster. If "
918 0 : "you want to start a new cluster in a development environment, use `fddev` instead "
919 0 : "of `fdctl`. If you want to use an existing genesis, set [development.bootstrap] "
920 0 : "to \"true\" in the configuration file." ));
921 :
922 0 : for( ulong i=0; i<config->gossip.entrypoints_cnt; i++ ) {
923 0 : if( FD_UNLIKELY( !strcmp( config->gossip.entrypoints[ i ], "" ) ) )
924 0 : FD_LOG_ERR(( "One of the entrypoints in your configuration file under [gossip.entrypoints] is "
925 0 : "empty. Please remove the empty entrypoint or set it correctly. "));
926 0 : }
927 :
928 0 : CHECK_PORT_NON_ZERO( gossip.port );
929 0 : CHECK_PORT_NON_ZERO( tiles.quic.quic_transaction_listen_port );
930 0 : CHECK_PORT_NON_ZERO( tiles.quic.regular_transaction_listen_port );
931 0 : CHECK_PORT_NON_ZERO( tiles.shred.shred_listen_port );
932 0 : CHECK_PORT_NON_ZERO( tiles.metric.prometheus_listen_port );
933 0 : CHECK_PORT_NON_ZERO( tiles.gui.gui_listen_port );
934 :
935 0 : #undef CHECK_PORT_NON_ZERO
936 :
937 0 : run_firedancer( config, -1, 1 );
938 0 : }
939 :
940 : action_t fd_action_run1 = {
941 : .name = "run1",
942 : .args = run1_cmd_args,
943 : .fn = run1_cmd_fn,
944 : .perm = NULL,
945 : .description = "Start up a single Firedancer tile"
946 : };
947 :
948 : action_t fd_action_run = {
949 : .name = "run",
950 : .args = NULL,
951 : .fn = run_cmd_fn,
952 : .require_config = 1,
953 : .perm = run_cmd_perm,
954 : .description = "Start up a Firedancer validator",
955 : .permission_err = "insufficient permissions to execute command `%s`. It is recommended "
956 : "to start Firedancer as the root user, but you can also start it "
957 : "with the missing capabilities listed above. The program only needs "
958 : "to start with elevated permissions to do privileged operations at "
959 : "boot, and will immediately drop permissions and switch to the user "
960 : "specified in your configuration file once they are complete. Firedancer "
961 : "will not execute outside of the boot process as root, and will refuse "
962 : "to start if it cannot drop privileges. Firedancer needs to be started "
963 : "privileged to configure high performance networking with XDP.",
964 : };
|
