Line data Source code
1 : #include "../metrics/fd_metrics.h"
2 : #include "../stem/fd_stem.h"
3 : #include "../topo/fd_topo.h"
4 :
5 : #include <fcntl.h>
6 : #include <errno.h>
7 : #include <stdlib.h>
8 : #include <sys/types.h> /* SEEK_SET */
9 : #include <time.h>
10 : #include <unistd.h>
11 :
12 : #include "generated/fd_cswtch_tile_seccomp.h"
13 :
14 0 : #define REPORT_INTERVAL_MILLIS (100L)
15 :
16 : typedef struct {
17 : long next_report_nanos;
18 :
19 : ulong tile_cnt;
20 : long first_seen_died[ FD_TILE_MAX ];
21 : int status_fds[ FD_TILE_MAX ];
22 : volatile ulong * metrics[ FD_TILE_MAX ];
23 : } fd_cswtch_ctx_t;
24 :
25 : FD_FN_CONST static inline ulong
26 0 : scratch_align( void ) {
27 0 : return 128UL;
28 0 : }
29 :
30 : FD_FN_PURE static inline ulong
31 0 : scratch_footprint( fd_topo_tile_t const * tile ) {
32 0 : (void)tile;
33 0 : ulong l = FD_LAYOUT_INIT;
34 0 : l = FD_LAYOUT_APPEND( l, alignof( fd_cswtch_ctx_t ), sizeof( fd_cswtch_ctx_t ) );
35 0 : return FD_LAYOUT_FINI( l, scratch_align() );
36 0 : }
37 :
38 : static void
39 : before_credit( fd_cswtch_ctx_t * ctx,
40 : fd_stem_context_t * stem,
41 0 : int * charge_busy ) {
42 0 : (void)stem;
43 :
44 0 : long now = fd_log_wallclock();
45 0 : if( now<ctx->next_report_nanos ) {
46 0 : long diff = ctx->next_report_nanos - now;
47 0 : diff = fd_long_min( diff, 2e6 /* 2ms */ );
48 0 : struct timespec const ts = {
49 0 : .tv_sec = diff / (long)1e9,
50 0 : .tv_nsec = diff % (long)1e9
51 0 : };
52 0 : clock_nanosleep( CLOCK_REALTIME, 0, &ts, NULL );
53 0 : return;
54 0 : }
55 0 : ctx->next_report_nanos += REPORT_INTERVAL_MILLIS*1000L*1000L;
56 :
57 0 : *charge_busy = 1;
58 :
59 0 : for( ulong i=0UL; i<ctx->tile_cnt; i++ ) {
60 0 : if( FD_UNLIKELY( -1==lseek( ctx->status_fds[ i ], 0, SEEK_SET ) ) ) FD_LOG_ERR(( "lseek failed (%i-%s)", errno, strerror( errno ) ));
61 :
62 0 : char contents[ 4096 ] = {0};
63 0 : ulong contents_len = 0UL;
64 :
65 0 : int process_died = 0;
66 0 : while( 1 ) {
67 0 : if( FD_UNLIKELY( contents_len>=sizeof( contents ) ) ) FD_LOG_ERR(( "contents overflow" ));
68 0 : long n = read( ctx->status_fds[ i ], contents + contents_len, sizeof( contents ) - contents_len );
69 0 : if( FD_UNLIKELY( -1==n ) ) {
70 0 : if( FD_UNLIKELY( errno==ESRCH ) ) {
71 0 : process_died = 1;
72 0 : break;
73 0 : }
74 0 : FD_LOG_ERR(( "read failed (%i-%s)", errno, strerror( errno ) ));
75 0 : }
76 0 : if( FD_LIKELY( 0==n ) ) break;
77 0 : contents_len += (ulong)n;
78 0 : }
79 :
80 0 : if( FD_UNLIKELY( process_died ) ) {
81 : /* The tile died, but it's a tile which is allowed to shutdown, so
82 : just stop updating metrics for it. */
83 0 : if( FD_UNLIKELY( ctx->metrics[ i ][ FD_METRICS_GAUGE_TILE_STATUS_OFF ] ) ) continue;
84 0 : }
85 :
86 : /* Supervisor is going to bring the whole process tree down if any
87 : of the target PIDs died, so we can ignore this and wait. */
88 0 : if( FD_UNLIKELY( process_died ) ) {
89 0 : if( FD_UNLIKELY( !ctx->first_seen_died[ i ] ) ) {
90 0 : ctx->first_seen_died[ i ] = now;
91 0 : } else if( FD_LIKELY( ctx->first_seen_died[ i ]==LONG_MAX ) ) {
92 : /* We already reported this, so we can ignore it. */
93 0 : } else if( FD_UNLIKELY( now-ctx->first_seen_died[ i ] < 10L*1000L*1000L*1000L ) ) {
94 : /* Wait 10 seconds for supervisor to kill us before reporting WARNING */
95 0 : } else {
96 0 : FD_LOG_WARNING(( "cannot get context switch metrics for dead tile idx %lu", i ));
97 0 : ctx->first_seen_died[ i ] = LONG_MAX;
98 0 : }
99 0 : continue;
100 0 : }
101 :
102 0 : int found_voluntary = 0;
103 0 : int found_involuntary = 0;
104 :
105 0 : char * line = contents;
106 0 : while( 1 ) {
107 0 : char * next_line = strchr( line, '\n' );
108 0 : if( FD_UNLIKELY( NULL==next_line ) ) break;
109 0 : *next_line = '\0';
110 :
111 0 : char * colon = strchr( line, ':' );
112 0 : if( FD_UNLIKELY( NULL==colon ) ) FD_LOG_ERR(( "no colon in line '%s'", line ));
113 :
114 0 : *colon = '\0';
115 0 : char * key = line;
116 0 : char * value = colon + 1;
117 :
118 0 : while( ' '==*value || '\t'==*value ) value++;
119 :
120 0 : if( FD_LIKELY( !strncmp( key, "voluntary_ctxt_switches", 23UL ) ) ) {
121 0 : char * endptr;
122 0 : ulong voluntary_ctxt_switches = strtoul( value, &endptr, 10 );
123 0 : if( FD_UNLIKELY( *endptr!='\0' || voluntary_ctxt_switches==ULONG_MAX ) ) FD_LOG_ERR(( "strtoul failed" ));
124 0 : ctx->metrics[ i ][ FD_METRICS_COUNTER_TILE_CONTEXT_SWITCH_VOLUNTARY_COUNT_OFF ] = voluntary_ctxt_switches;
125 0 : found_voluntary = 1;
126 0 : } else if( FD_LIKELY( !strncmp( key, "nonvoluntary_ctxt_switches", 26UL ) ) ) {
127 0 : char * endptr;
128 0 : ulong involuntary_ctxt_switches = strtoul( value, &endptr, 10 );
129 0 : if( FD_UNLIKELY( *endptr!='\0' || involuntary_ctxt_switches==ULONG_MAX ) ) FD_LOG_ERR(( "strtoul failed" ));
130 0 : ctx->metrics[ i ][ FD_METRICS_COUNTER_TILE_CONTEXT_SWITCH_INVOLUNTARY_COUNT_OFF ] = involuntary_ctxt_switches;
131 0 : found_involuntary = 1;
132 0 : }
133 :
134 0 : line = next_line + 1;
135 0 : }
136 :
137 0 : if( FD_UNLIKELY( !found_voluntary ) ) FD_LOG_ERR(( "voluntary_ctxt_switches not found" ));
138 0 : if( FD_UNLIKELY( !found_involuntary ) ) FD_LOG_ERR(( "nonvoluntary_ctxt_switches not found" ));
139 0 : }
140 0 : }
141 :
142 : static void
143 : privileged_init( fd_topo_t * topo,
144 0 : fd_topo_tile_t * tile ) {
145 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
146 :
147 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
148 0 : fd_cswtch_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_cswtch_ctx_t ), sizeof( fd_cswtch_ctx_t ) );
149 :
150 0 : FD_TEST( topo->tile_cnt<FD_TILE_MAX );
151 :
152 0 : ctx->tile_cnt = topo->tile_cnt;
153 0 : for( ulong i=0UL; i<topo->tile_cnt; i++ ) {
154 0 : ulong * metrics = fd_metrics_join( fd_topo_obj_laddr( topo, topo->tiles[ i ].metrics_obj_id ) );
155 :
156 0 : for(;;) {
157 0 : ulong pid, tid;
158 0 : if( FD_UNLIKELY( tile->id==i ) ) {
159 0 : pid = fd_sandbox_getpid();
160 0 : tid = fd_sandbox_gettid();
161 0 : } else {
162 0 : pid = fd_metrics_tile( metrics )[ FD_METRICS_GAUGE_TILE_PID_OFF ];
163 0 : tid = fd_metrics_tile( metrics )[ FD_METRICS_GAUGE_TILE_TID_OFF ];
164 0 : if( FD_UNLIKELY( !pid || !tid ) ) {
165 0 : FD_SPIN_PAUSE();
166 0 : continue;
167 0 : }
168 0 : }
169 :
170 0 : char path[ 64 ];
171 0 : FD_TEST( fd_cstr_printf_check( path, sizeof( path ), NULL, "/proc/%lu/task/%lu/status", pid, tid ) );
172 0 : ctx->status_fds[ i ] = open( path, O_RDONLY );
173 0 : ctx->metrics[ i ] = fd_metrics_tile( metrics );
174 0 : if( FD_UNLIKELY( -1==ctx->status_fds[ i ] ) ) FD_LOG_ERR(( "open failed (%i-%s)", errno, strerror( errno ) ));
175 0 : break;
176 0 : }
177 0 : }
178 0 : }
179 :
180 : static void
181 : unprivileged_init( fd_topo_t * topo,
182 0 : fd_topo_tile_t * tile ) {
183 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
184 :
185 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
186 0 : fd_cswtch_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_cswtch_ctx_t ), sizeof( fd_cswtch_ctx_t ) );
187 :
188 0 : memset( ctx->first_seen_died, 0, sizeof( ctx->first_seen_died ) );
189 0 : ctx->next_report_nanos = fd_log_wallclock();
190 :
191 0 : ulong scratch_top = FD_SCRATCH_ALLOC_FINI( l, 1UL );
192 0 : if( FD_UNLIKELY( scratch_top > (ulong)scratch + scratch_footprint( tile ) ) )
193 0 : FD_LOG_ERR(( "scratch overflow %lu %lu %lu", scratch_top - (ulong)scratch - scratch_footprint( tile ), scratch_top, (ulong)scratch + scratch_footprint( tile ) ));
194 0 : }
195 :
196 : static ulong
197 : populate_allowed_seccomp( fd_topo_t const * topo,
198 : fd_topo_tile_t const * tile,
199 : ulong out_cnt,
200 0 : struct sock_filter * out ) {
201 0 : (void)topo;
202 0 : (void)tile;
203 :
204 0 : populate_sock_filter_policy_fd_cswtch_tile( out_cnt, out, (uint)fd_log_private_logfile_fd() );
205 0 : return sock_filter_policy_fd_cswtch_tile_instr_cnt;
206 0 : }
207 :
208 : static ulong
209 : populate_allowed_fds( fd_topo_t const * topo,
210 : fd_topo_tile_t const * tile,
211 : ulong out_fds_cnt,
212 0 : int * out_fds ) {
213 0 : void * scratch = fd_topo_obj_laddr( topo, tile->tile_obj_id );
214 :
215 0 : FD_SCRATCH_ALLOC_INIT( l, scratch );
216 0 : fd_cswtch_ctx_t * ctx = FD_SCRATCH_ALLOC_APPEND( l, alignof( fd_cswtch_ctx_t ), sizeof( fd_cswtch_ctx_t ) );
217 :
218 0 : if( FD_UNLIKELY( out_fds_cnt<2UL+ctx->tile_cnt ) ) FD_LOG_ERR(( "out_fds_cnt %lu", out_fds_cnt ));
219 :
220 0 : ulong out_cnt = 0UL;
221 0 : out_fds[ out_cnt++ ] = 2; /* stderr */
222 0 : if( FD_LIKELY( -1!=fd_log_private_logfile_fd() ) )
223 0 : out_fds[ out_cnt++ ] = fd_log_private_logfile_fd(); /* logfile */
224 0 : for( ulong i=0UL; i<ctx->tile_cnt; i++ )
225 0 : out_fds[ out_cnt++ ] = ctx->status_fds[ i ]; /* /proc/<pid>/task/<tid>/status descriptor */
226 0 : return out_cnt;
227 0 : }
228 :
229 0 : #define STEM_BURST (1UL)
230 0 : #define STEM_LAZY ((long)10e6) /* 10ms */
231 :
232 0 : #define STEM_CALLBACK_CONTEXT_TYPE fd_cswtch_ctx_t
233 0 : #define STEM_CALLBACK_CONTEXT_ALIGN alignof(fd_cswtch_ctx_t)
234 :
235 0 : #define STEM_CALLBACK_BEFORE_CREDIT before_credit
236 :
237 : #include "../../disco/stem/fd_stem.c"
238 :
239 : fd_topo_run_tile_t fd_tile_cswtch = {
240 : .name = "cswtch",
241 : .populate_allowed_seccomp = populate_allowed_seccomp,
242 : .populate_allowed_fds = populate_allowed_fds,
243 : .scratch_align = scratch_align,
244 : .scratch_footprint = scratch_footprint,
245 : .privileged_init = privileged_init,
246 : .unprivileged_init = unprivileged_init,
247 : .run = stem_run,
248 : };
|
