Line data Source code
1 : #include "fd_wksp_private.h"
2 :
3 : int
4 178 : fd_wksp_private_lock( fd_wksp_t * wksp ) {
5 : # if FD_WKSP_LOCK_RECLAIM
6 : int warning = 0;
7 : #endif
8 178 : ulong me = fd_log_group_id();
9 :
10 178 : ulong * _owner = &wksp->owner;
11 178 : for(;;) {
12 :
13 : /* Note that we emulate CAS on platforms without FD_HAS_ATOMIC
14 : to minimize the amount of code differences we have to test. On
15 : platforms without FD_HAS_ATOMIC, a workspace should not be used
16 : concurrently though. */
17 :
18 178 : FD_COMPILER_MFENCE();
19 178 : # if FD_HAS_ATOMIC
20 178 : ulong pid = FD_ATOMIC_CAS( _owner, ULONG_MAX, me );
21 : # else
22 : ulong pid = FD_VOLATILE_CONST( *_owner );
23 : if( pid==ULONG_MAX ) FD_VOLATILE( *_owner ) = me;
24 : # endif
25 178 : FD_COMPILER_MFENCE();
26 :
27 178 : if( FD_LIKELY( pid==ULONG_MAX ) ) return FD_WKSP_SUCCESS;
28 :
29 : # if FD_WKSP_LOCK_RECLAIM
30 : int status = fd_log_group_id_query( pid );
31 : if( FD_UNLIKELY( status==FD_LOG_GROUP_ID_QUERY_DEAD ) ) { /* A process died while holding the lock, try to recover the lock */
32 :
33 : FD_COMPILER_MFENCE();
34 : # if FD_HAS_ATOMIC
35 : ulong cur = FD_ATOMIC_CAS( _owner, pid, me );
36 : # else
37 : ulong cur = FD_VOLATILE_CONST( *_owner );
38 : if( cur==pid ) FD_VOLATILE( *_owner ) = me;
39 : # endif
40 : FD_COMPILER_MFENCE();
41 :
42 : if( FD_LIKELY( cur==pid ) ) { /* We recovered the lock from the dead pid, try to fix up incomplete ops */
43 :
44 : FD_LOG_WARNING(( "Process %lu died in an operation on wksp %s; verifying", pid, wksp->name ));
45 : if( FD_LIKELY( !fd_wksp_verify( wksp ) ) ) { /* logs details of issues detected */
46 : FD_LOG_NOTICE(( "wksp verified" ));
47 : return FD_WKSP_SUCCESS;
48 : }
49 :
50 : FD_LOG_WARNING(( "Issues detected; rebuilding" ));
51 : if( FD_UNLIKELY( fd_wksp_rebuild( wksp, wksp->seed ) ) ) { /* Rebuild failed (logs details of issues detected) */
52 : /* Return control of the lock to the previous owner */
53 : FD_COMPILER_MFENCE();
54 : FD_VOLATILE( *_owner ) = pid;
55 : FD_COMPILER_MFENCE();
56 : FD_LOG_WARNING(( "corrupt wksp detected" ));
57 : return FD_WKSP_ERR_CORRUPT;
58 : }
59 :
60 : FD_LOG_NOTICE(( "wksp rebuilt" ));
61 : return FD_WKSP_SUCCESS;
62 :
63 : }
64 :
65 : /* Somebody beat us to recovering the lock ... try again */
66 :
67 : } else if( FD_UNLIKELY( status!=FD_LOG_GROUP_ID_QUERY_LIVE ) ) { /* Unclear pid status ... issue a warning and try again */
68 :
69 : if( FD_UNLIKELY( !warning ) ) {
70 : FD_LOG_WARNING(( "wksp %s is owned by unknown pid %li; attempting to recover", wksp->name, pid ));
71 : warning = 1;
72 : }
73 :
74 : }
75 :
76 : /* At this point, either another thread in this process has the
77 : lock, another active thread in another process has the lock,
78 : another unknown status thread in other process has the lock or
79 : another thread beat us to reclaim the lock from a dead process.
80 : In any case, we don't have the lock. Wait a while to limit O/S
81 : contention and try again. */
82 :
83 : FD_YIELD();
84 : # else
85 :
86 : /* If we are running without FD_WKSP_LOCK_RECLAIM then it is assumed
87 : that the contention is caused by a tile pinned to another core,
88 : and that this core is itself pinned so spin locking is best. */
89 0 : FD_SPIN_PAUSE();
90 :
91 0 : #endif
92 0 : }
93 :
94 : /* never get here */
95 178 : }
96 :
97 : /* Public APIs ********************************************************/
98 :
99 : ulong
100 : fd_wksp_part_max_est( ulong footprint,
101 1 : ulong sz_typical ) {
102 1 : footprint = fd_ulong_align_dn( footprint, FD_WKSP_ALIGN );
103 1 : ulong data_end = footprint - 1UL;
104 1 : ulong pinfo_off = fd_wksp_private_pinfo_off();
105 1 : ulong consumed = sizeof(fd_wksp_private_pinfo_t) + sz_typical;
106 1 : ulong part_max = (data_end - pinfo_off) / (consumed + (ulong)!consumed); /* avoid div-by-zero */
107 1 : if( FD_UNLIKELY( (!footprint) | (!sz_typical) | (sz_typical>consumed) | (pinfo_off>data_end) ) ) return 0UL;
108 1 : return fd_ulong_min( part_max, FD_WKSP_PRIVATE_PINFO_IDX_NULL );
109 1 : }
110 :
111 : ulong
112 : fd_wksp_data_max_est( ulong footprint,
113 0 : ulong part_max ) {
114 0 : footprint = fd_ulong_align_dn( footprint, FD_WKSP_ALIGN );
115 0 : ulong data_end = footprint - 1UL;
116 0 : ulong data_off = fd_wksp_private_data_off( part_max );
117 0 : if( FD_UNLIKELY( (!part_max) | (part_max>FD_WKSP_PRIVATE_PINFO_IDX_NULL) |
118 0 : (part_max > ((ULONG_MAX - fd_wksp_private_pinfo_off())/sizeof(fd_wksp_private_pinfo_t))) | /* covered above */
119 0 : (!footprint) | (data_off>=data_end) ) ) return 0UL;
120 0 : return data_end - data_off;
121 0 : }
122 :
123 : ulong
124 0 : fd_wksp_align( void ) {
125 0 : return FD_WKSP_ALIGN;
126 0 : }
127 :
128 : ulong
129 : fd_wksp_footprint( ulong part_max,
130 4 : ulong data_max ) {
131 4 : ulong data_off = fd_wksp_private_data_off( part_max );
132 4 : if( FD_UNLIKELY( (!part_max) | (part_max>FD_WKSP_PRIVATE_PINFO_IDX_NULL) | (!data_max) |
133 4 : (part_max > ((ULONG_MAX - fd_wksp_private_pinfo_off())/sizeof(fd_wksp_private_pinfo_t))) | /* Covered above */
134 4 : (data_max > (ULONG_MAX - FD_WKSP_ALIGN + 1UL - data_off - 1UL) ) ) ) return 0UL;
135 4 : return fd_ulong_align_up( data_off + data_max + 1UL, FD_WKSP_ALIGN );
136 4 : }
137 :
138 : void *
139 : fd_wksp_new( void * shmem,
140 : char const * name,
141 : uint seed,
142 : ulong part_max,
143 1 : ulong data_max ) {
144 1 : fd_wksp_t * wksp = (fd_wksp_t *)shmem;
145 :
146 1 : if( FD_UNLIKELY( !wksp ) ) {
147 0 : FD_LOG_WARNING(( "NULL shmem" ));
148 0 : return NULL;
149 0 : }
150 :
151 1 : if( FD_UNLIKELY( !fd_ulong_is_aligned( (ulong)wksp, FD_WKSP_ALIGN ) ) ) {
152 0 : FD_LOG_WARNING(( "bad align" ));
153 0 : return NULL;
154 0 : }
155 :
156 1 : ulong name_len = fd_shmem_name_len( name );
157 1 : if( FD_UNLIKELY( !name_len ) ) {
158 0 : FD_LOG_WARNING(( "bad name" ));
159 0 : return NULL;
160 0 : }
161 :
162 1 : ulong footprint = fd_wksp_footprint( part_max, data_max );
163 1 : if( FD_UNLIKELY( !footprint ) ) {
164 0 : FD_LOG_WARNING(( "bad part_max and/or data_max" ));
165 0 : return NULL;
166 0 : }
167 :
168 1 : fd_memset( wksp, 0, fd_wksp_footprint( part_max, 1UL ) );
169 :
170 1 : wksp->part_max = part_max;
171 1 : wksp->data_max = data_max;
172 1 : wksp->gaddr_lo = fd_wksp_private_data_off( part_max );
173 1 : wksp->gaddr_hi = wksp->gaddr_lo + data_max;
174 1 : fd_memcpy( wksp->name, name, name_len+1UL );
175 1 : wksp->seed = seed;
176 1 : wksp->idle_top_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
177 1 : wksp->part_head_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
178 1 : wksp->part_tail_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
179 1 : wksp->part_used_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
180 1 : wksp->part_free_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
181 1 : wksp->cycle_tag = 4UL; /* Verify uses tags 0-3 */
182 1 : wksp->owner = 0UL; /* Mark as locked and in construction */
183 :
184 : /* Note that wksp->owner was set to zero above, "locking" the wksp by
185 : group_id 0. And the memset above set all the partition tags to
186 : zero such that there are no allocated partitions. So once we set
187 : magic below, we can finish the initialization by rebuilding and
188 : unlocking. Since fd_log_group_id is non-zero, the zero owner
189 : indicates to any remote observer of the shared memory region that
190 : the wksp is being built for the first time. */
191 :
192 1 : FD_COMPILER_MFENCE();
193 1 : FD_VOLATILE( wksp->magic ) = FD_WKSP_MAGIC;
194 1 : FD_COMPILER_MFENCE();
195 :
196 1 : int err = fd_wksp_rebuild( wksp, seed );
197 1 : if( FD_UNLIKELY( err ) ) { /* Should be impossible at this point */
198 :
199 0 : FD_COMPILER_MFENCE();
200 0 : FD_VOLATILE( wksp->magic ) = 0UL;
201 0 : FD_COMPILER_MFENCE();
202 :
203 0 : FD_LOG_WARNING(( "fd_wksp_rebuild failed (%i-%s)", err, fd_wksp_strerror( err ) ));
204 0 : return NULL;
205 0 : }
206 :
207 : #if FD_HAS_DEEPASAN
208 : fd_wksp_private_asan_poison_data( wksp );
209 : #endif
210 :
211 1 : fd_wksp_private_unlock( wksp );
212 :
213 1 : return wksp;
214 1 : }
215 :
216 : fd_wksp_t *
217 1 : fd_wksp_join( void * shwksp ) {
218 1 : fd_wksp_t * wksp = (fd_wksp_t *)shwksp;
219 :
220 1 : if( FD_UNLIKELY( !wksp ) ) {
221 0 : FD_LOG_WARNING(( "NULL shwksp" ));
222 0 : return NULL;
223 0 : }
224 :
225 1 : if( FD_UNLIKELY( !fd_ulong_is_aligned( (ulong)wksp, FD_WKSP_ALIGN ) ) ) {
226 0 : FD_LOG_WARNING(( "bad align" ));
227 0 : return NULL;
228 0 : }
229 :
230 1 : if( FD_UNLIKELY( wksp->magic!=FD_WKSP_MAGIC ) ) {
231 0 : FD_LOG_WARNING(( "bad magic" ));
232 0 : return NULL;
233 0 : }
234 :
235 1 : return wksp;
236 1 : }
237 :
238 : void *
239 1 : fd_wksp_leave( fd_wksp_t * wksp ) {
240 1 : if( FD_UNLIKELY( !wksp ) ) {
241 0 : FD_LOG_WARNING(( "NULL wksp" ));
242 0 : return NULL;
243 0 : }
244 :
245 1 : return (void *)wksp;
246 1 : }
247 :
248 : void *
249 1 : fd_wksp_delete( void * shwksp ) {
250 1 : fd_wksp_t * wksp = (fd_wksp_t *)shwksp;
251 :
252 1 : if( FD_UNLIKELY( !wksp ) ) {
253 0 : FD_LOG_WARNING(( "NULL shwksp" ));
254 0 : return NULL;
255 0 : }
256 :
257 1 : if( FD_UNLIKELY( !fd_ulong_is_aligned( (ulong)wksp, FD_WKSP_ALIGN ) ) ) {
258 0 : FD_LOG_WARNING(( "bad align" ));
259 0 : return NULL;
260 0 : }
261 :
262 1 : if( FD_UNLIKELY( wksp->magic!=FD_WKSP_MAGIC ) ) {
263 0 : FD_LOG_WARNING(( "bad magic" ));
264 0 : return NULL;
265 0 : }
266 :
267 : /* TODO: consider testing owner */
268 :
269 1 : FD_COMPILER_MFENCE();
270 1 : FD_VOLATILE( wksp->magic ) = 0UL;
271 1 : FD_COMPILER_MFENCE();
272 :
273 : # if FD_HAS_DEEPASAN
274 : /* Unpoison entire wksp region. */
275 : ulong footprint = fd_wksp_footprint( wksp->part_max, wksp->data_max );
276 : void * wksp_data = (void*)((ulong)wksp + fd_wksp_private_pinfo_off());
277 : fd_asan_unpoison( wksp_data, footprint - fd_wksp_private_pinfo_off());
278 : # endif
279 :
280 1 : return wksp;
281 1 : }
282 :
283 0 : char const * fd_wksp_name ( fd_wksp_t const * wksp ) { return wksp->name; }
284 0 : uint fd_wksp_seed ( fd_wksp_t const * wksp ) { return wksp->seed; }
285 0 : ulong fd_wksp_part_max( fd_wksp_t const * wksp ) { return wksp->part_max; }
286 0 : ulong fd_wksp_data_max( fd_wksp_t const * wksp ) { return wksp->data_max; }
287 0 : ulong fd_wksp_gaddr_lo( fd_wksp_t const * wksp ) { return wksp->gaddr_lo; }
288 0 : ulong fd_wksp_gaddr_hi( fd_wksp_t const * wksp ) { return wksp->gaddr_hi; }
289 :
290 : ulong
291 0 : fd_wksp_owner( fd_wksp_t const * wksp ) {
292 0 : FD_COMPILER_MFENCE();
293 0 : ulong owner = FD_VOLATILE_CONST( wksp->owner );
294 0 : FD_COMPILER_MFENCE();
295 0 : return owner;
296 0 : }
297 :
298 : char const *
299 0 : fd_wksp_strerror( int err ) {
300 0 : switch( err ) {
301 0 : case FD_WKSP_SUCCESS: return "success";
302 0 : case FD_WKSP_ERR_INVAL: return "inval";
303 0 : case FD_WKSP_ERR_FAIL: return "fail";
304 0 : case FD_WKSP_ERR_CORRUPT: return "corrupt";
305 0 : default: break;
306 0 : }
307 0 : return "unknown";
308 0 : }
309 :
310 : int
311 0 : fd_wksp_verify( fd_wksp_t * wksp ) {
312 :
313 0 : # define TEST(c) do { \
314 0 : if( FD_UNLIKELY( !(c) ) ) { FD_LOG_WARNING(( "FAIL: %s", #c )); return FD_WKSP_ERR_CORRUPT; } \
315 0 : } while(0)
316 :
317 : /* Validate metadata */
318 :
319 0 : TEST( wksp );
320 0 : TEST( wksp->magic==FD_WKSP_MAGIC );
321 :
322 0 : ulong part_max = wksp->part_max;
323 0 : ulong data_max = wksp->data_max;
324 0 : TEST( fd_wksp_footprint( part_max, data_max ) );
325 :
326 0 : ulong gaddr_lo = wksp->gaddr_lo; TEST( gaddr_lo==fd_wksp_private_data_off( part_max ) );
327 0 : ulong gaddr_hi = wksp->gaddr_hi; TEST( gaddr_hi==gaddr_lo+data_max );
328 :
329 0 : TEST( fd_shmem_name_len( wksp->name ) );
330 :
331 : /* seed is arbitrary */
332 :
333 0 : TEST( wksp->cycle_tag >= 4UL );
334 :
335 : /* TODO: consider verifying owner */
336 :
337 0 : fd_wksp_private_pinfo_t * pinfo = fd_wksp_private_pinfo( wksp );
338 :
339 : /* Clear out cycle tags */
340 :
341 0 : for( ulong i=0UL; i<part_max; i++ ) pinfo[ i ].cycle_tag = 0UL;
342 :
343 : /* Verify the idle stack */
344 :
345 0 : ulong idle_cnt = 0UL;
346 :
347 0 : do {
348 0 : ulong i = fd_wksp_private_pinfo_idx( wksp->idle_top_cidx );
349 0 : while( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
350 :
351 : /* Visit i. Note that i has not been validated yet. */
352 :
353 0 : TEST( i<part_max ); /* Validate i */
354 0 : TEST( !pinfo[ i ].cycle_tag ); /* Make sure not visited before */
355 0 : pinfo[ i ].cycle_tag = 1UL; /* Mark as visited in idle stack */
356 0 : idle_cnt++; /* Update the idle cnt */
357 :
358 : /* Advance to the next idle */
359 :
360 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx );
361 0 : }
362 0 : } while(0);
363 :
364 : /* Idle stack looks intact, verify partitioning */
365 :
366 0 : ulong free_cnt = 0UL;
367 0 : ulong used_cnt = 0UL;
368 :
369 0 : do {
370 0 : ulong j = FD_WKSP_PRIVATE_PINFO_IDX_NULL;
371 0 : ulong i = fd_wksp_private_pinfo_idx( wksp->part_head_cidx );
372 0 : ulong g = gaddr_lo;
373 :
374 0 : int last_free = 0;
375 :
376 0 : while( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
377 :
378 : /* At this point, we last visited j. Visit i. Note that j has
379 : been validated but i has not. */
380 :
381 0 : TEST( i<part_max ); /* Validate i */
382 0 : TEST( pinfo[ i ].gaddr_lo==g ); /* Make sure partition is tightly adjacent to previous */
383 0 : TEST( pinfo[ i ].gaddr_hi> g ); /* Make sure partition size is non-zero */
384 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ i ].prev_cidx )==j ); /* Make sure correct prev partition */
385 0 : TEST( !pinfo[ i ].cycle_tag ); /* Make sure not visited before */
386 0 : pinfo[ i ].cycle_tag = 2UL; /* Mark as visited in partitioning */
387 :
388 0 : g = pinfo[ i ].gaddr_hi; /* Extract where the next partition should start */
389 0 : int is_free = !pinfo[ i ].tag; /* Determine if this partition is free or used */
390 0 : TEST( !(last_free & is_free) ); /* Make sure no adjacent free partitions */
391 0 : free_cnt += (ulong) is_free; /* Update the free cnt */
392 0 : used_cnt += (ulong)!is_free; /* Update the used cnt */
393 :
394 : /* Advance to the next partition */
395 :
396 0 : last_free = is_free;
397 :
398 0 : j = i;
399 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].next_cidx );
400 0 : }
401 :
402 0 : TEST( fd_wksp_private_pinfo_idx( wksp->part_tail_cidx )==j ); /* Make sure correct partition tail */
403 0 : TEST( g==gaddr_hi ); /* Make sure complete partitioning */
404 0 : TEST( (idle_cnt + free_cnt + used_cnt)==part_max ); /* Make sure no lost idle partitions */
405 0 : } while(0);
406 :
407 : /* Idle stack and partitioning look intact, validate used treap */
408 :
409 0 : do {
410 0 : ulong visit_cnt = 0UL;
411 :
412 0 : ulong i = fd_wksp_private_pinfo_idx( wksp->part_used_cidx );
413 0 : ulong s = FD_WKSP_PRIVATE_PINFO_IDX_NULL;
414 0 : ulong g = gaddr_lo;
415 :
416 0 : if( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
417 0 : TEST( i<part_max ); /* Validate i */
418 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx )==s ); /* Validate parent */
419 0 : }
420 :
421 0 : for(;;) {
422 :
423 : /* At this point i is and everything on stack is validated */
424 :
425 0 : if( fd_wksp_private_pinfo_idx_is_null( i ) ) {
426 0 : if( fd_wksp_private_pinfo_idx_is_null( s ) ) break; /* Done */
427 :
428 : /* Pop stack */
429 :
430 0 : i = s;
431 0 : s = fd_wksp_private_pinfo_idx( pinfo[ i ].stack_cidx );
432 :
433 : /* Visit i */
434 :
435 0 : ulong p = fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx ); /* Extract the parent */
436 :
437 0 : TEST( pinfo[ i ].gaddr_lo>=g ); /* Make sure this starts after last visited */
438 0 : TEST( pinfo[ i ].tag ); /* Make sure tagged as a used partition */
439 0 : TEST( pinfo[ i ].cycle_tag==2UL ); /* Make sure in partitioning and not visited yet this traversal */
440 0 : if( !fd_wksp_private_pinfo_idx_is_null( p ) ) /* Make sure heap property satisfied */
441 0 : TEST( pinfo[ p ].heap_prio >= pinfo[ i ].heap_prio );
442 :
443 0 : TEST( !pinfo[ i ].in_same ); /* Make sure unique */
444 0 : TEST( fd_wksp_private_pinfo_idx_is_null( fd_wksp_private_pinfo_idx( pinfo[ i ].same_cidx ) ) ); /* " */
445 :
446 0 : pinfo[ i ].cycle_tag = 3UL; /* Mark as visited this traversal */
447 0 : visit_cnt++; /* Update the visit cnt */
448 0 : g = pinfo[ i ].gaddr_hi; /* Get minimum start for next partition */
449 :
450 : /* Traverse the right subtree */
451 :
452 0 : p = i;
453 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].right_cidx );
454 0 : if( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
455 0 : TEST( i<part_max ); /* Validate i */
456 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx )==p ); /* Validate parent */
457 0 : }
458 :
459 0 : } else {
460 :
461 : /* At this point i and everything on the stack is validated.
462 : Push i to the stack and recurse on the left subtree. */
463 :
464 0 : pinfo[ i ].stack_cidx = fd_wksp_private_pinfo_cidx( s );
465 0 : s = i;
466 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].left_cidx );
467 0 : if( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
468 0 : TEST( i<part_max ); /* Validate i */
469 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx )==s ); /* Validate parent */
470 0 : }
471 :
472 0 : }
473 0 : }
474 :
475 0 : TEST( visit_cnt==used_cnt ); /* Make sure all used partitions in used treap */
476 0 : } while(0);
477 :
478 : /* Idle stack, partitioning and used treap look intact, validate the
479 : free treap. */
480 :
481 0 : do {
482 0 : ulong visit_cnt = 0UL;
483 :
484 0 : ulong i = fd_wksp_private_pinfo_idx( wksp->part_free_cidx );
485 0 : ulong s = FD_WKSP_PRIVATE_PINFO_IDX_NULL;
486 0 : ulong sz = 0UL;
487 :
488 0 : if( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
489 0 : TEST( i<part_max ); /* Validate i */
490 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx )==s ); /* Validate parent */
491 0 : }
492 :
493 0 : for(;;) {
494 :
495 : /* At this point i and everything on the stack is validated */
496 :
497 0 : if( fd_wksp_private_pinfo_idx_is_null( i ) ) {
498 0 : if( fd_wksp_private_pinfo_idx_is_null( s ) ) break; /* Done */
499 :
500 : /* Pop stack */
501 :
502 0 : i = s;
503 0 : s = fd_wksp_private_pinfo_idx( pinfo[ i ].stack_cidx );
504 :
505 : /* Visit i */
506 :
507 0 : ulong p = fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx ); /* Extract the parent */
508 0 : ulong isz = fd_wksp_private_pinfo_sz( pinfo + i ); /* Extract the size */
509 :
510 0 : TEST( isz>sz ); /* Make sure this partition i larger than previous */
511 0 : TEST( !pinfo[ i ].tag ); /* Make sure tagged as a free partition */
512 0 : TEST( !pinfo[ i ].in_same ); /* Make sure marked as not in same */
513 :
514 0 : if( !fd_wksp_private_pinfo_idx_is_null( p ) ) { /* Make sure heap property satisfied */
515 0 : TEST( pinfo[ p ].heap_prio >= pinfo[ i ].heap_prio );
516 0 : }
517 :
518 0 : sz = isz; /* Update largest size partition seen so far */
519 :
520 : /* Traverse all same sized partitions */
521 :
522 0 : ulong j = i;
523 0 : for(;;) {
524 :
525 : /* At this point, j is validated */
526 :
527 0 : TEST( pinfo[ j ].cycle_tag==2UL ); /* Make sure in partitioning and not visited yet this traversal */
528 0 : pinfo[ j ].cycle_tag = 3UL; /* Mark as visited this traversal */
529 0 : visit_cnt++;
530 :
531 0 : ulong k = fd_wksp_private_pinfo_idx( pinfo[ j ].same_cidx ); /* Get the next same sized */
532 0 : if( fd_wksp_private_pinfo_idx_is_null( k ) ) break; /* If no more, we are done with this node */
533 0 : TEST( k<part_max ); /* Make sure valid index */
534 0 : TEST( fd_wksp_private_pinfo_sz( pinfo + k )==sz ); /* Make sure same size */
535 0 : TEST( pinfo[ k ].in_same ); /* Make sure marked as in same */
536 0 : TEST( fd_wksp_private_pinfo_idx_is_null( fd_wksp_private_pinfo_idx( pinfo[ k ].left_cidx ) ) );
537 0 : TEST( fd_wksp_private_pinfo_idx_is_null( fd_wksp_private_pinfo_idx( pinfo[ k ].right_cidx ) ) );
538 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ k ].parent_cidx )==j ); /* Make sure correct parent */
539 0 : j = k;
540 0 : }
541 :
542 : /* Recurse on the right subtree */
543 :
544 0 : p = i;
545 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].right_cidx );
546 0 : if( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
547 0 : TEST( i<part_max ); /* Validate i */
548 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx )==p ); /* Validate parent */
549 0 : }
550 :
551 0 : } else {
552 :
553 0 : TEST( i<part_max ); /* Validate i */
554 :
555 : /* At this point i and everything on the stack is validated.
556 : Push i to the stack and recurse on the left subtree. */
557 :
558 0 : pinfo[ i ].stack_cidx = fd_wksp_private_pinfo_cidx( s );
559 0 : s = i;
560 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].left_cidx );
561 0 : if( !fd_wksp_private_pinfo_idx_is_null( i ) ) {
562 0 : TEST( i<part_max ); /* Validate i */
563 0 : TEST( fd_wksp_private_pinfo_idx( pinfo[ i ].parent_cidx )==s ); /* Validate parent */
564 0 : }
565 0 : }
566 0 : }
567 :
568 0 : TEST( visit_cnt==free_cnt ); /* Make sure all free partitions in free treap */
569 :
570 0 : } while(0);
571 :
572 0 : # undef TEST
573 :
574 0 : return FD_WKSP_SUCCESS;
575 0 : }
576 :
577 : int
578 : fd_wksp_rebuild( fd_wksp_t * wksp,
579 1 : uint seed ) {
580 :
581 : /* Load the wksp metadata, don't rebuild if any of it looks even
582 : slightly off. */
583 :
584 1 : if( FD_UNLIKELY( !wksp ) ) {
585 0 : FD_LOG_WARNING(( "NULL wksp" ));
586 0 : return FD_WKSP_ERR_CORRUPT;
587 0 : }
588 :
589 1 : ulong magic = wksp->magic;
590 1 : ulong part_max = wksp->part_max;
591 1 : ulong data_max = wksp->data_max;
592 1 : ulong gaddr_lo = wksp->gaddr_lo;
593 1 : ulong gaddr_hi = wksp->gaddr_hi;
594 1 : ulong cycle_tag = wksp->cycle_tag;
595 :
596 : /* TODO: consider verifying owner */
597 :
598 1 : ulong footprint = fd_wksp_footprint( part_max, data_max );
599 1 : ulong gaddr_lo_exp = fd_wksp_private_data_off( part_max );
600 1 : ulong gaddr_hi_exp = gaddr_lo_exp + data_max;
601 1 : if( FD_UNLIKELY( (magic!=FD_WKSP_MAGIC) | (!footprint) | (!fd_shmem_name_len( wksp->name )) |
602 1 : (gaddr_lo!=gaddr_lo_exp) | (gaddr_hi!=gaddr_hi_exp) | (cycle_tag<4UL) ) ) {
603 0 : FD_LOG_WARNING(( "bad metadata\n\t"
604 0 : "magic %016lx (exp %016lx)\n\t"
605 0 : "part_max %lu data_max %lu (footprint %lu)\n\t"
606 0 : "gaddr_lo %lu (exp %lu)\n\t"
607 0 : "gaddr_hi %lu (exp %lu)\n\t"
608 0 : "cycle_tag %lu (exp>=4)",
609 0 : magic, FD_WKSP_MAGIC, part_max, data_max, footprint,
610 0 : gaddr_lo, gaddr_lo_exp, gaddr_hi, gaddr_hi_exp, cycle_tag ));
611 0 : return FD_WKSP_ERR_CORRUPT;
612 0 : }
613 :
614 : /* Scan the wksp pinfo and insert any used partitions into the used
615 : treap and put the rest on the idle stack. If there is any sign of
616 : corruption (empty, bad range or overlap between used partitions),
617 : we abort the rebuild (this is almost certainly data corruption of
618 : some form and we don't have enough info to resolve a conflict
619 : without potentially making the situation worse). We do the scan in
620 : reverse order to rebuild the idle stack in forward order.
621 :
622 : Note that we don't ever change the gaddr_lo,gaddr_hi of any tagged
623 : partitions such that operation is guaranteed to never change the
624 : single source of truth. As such, this operation can be interrupted
625 : and restarted arbitrarily safely.*/
626 :
627 1 : fd_wksp_private_pinfo_t * pinfo = fd_wksp_private_pinfo( wksp );
628 :
629 1 : do {
630 1 : wksp->seed = seed;
631 1 : wksp->idle_top_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL ); /* Flush idle stack */
632 1 : wksp->part_used_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL ); /* Flush used treap */
633 1 : wksp->part_free_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL ); /* Flush free treap */
634 :
635 1 : ulong i = part_max;
636 4321157 : while( i ) {
637 4321156 : i--;
638 :
639 : /* Ideally, heap priorities should just be a shuffling of the
640 : integers [0,part_max). fd_uint_hash will generate such a
641 : shuffling for part_max = 2^32. Using the lower 30 bits
642 : (reserving bit 31 for bulk operations) will yield something
643 : very close. We use seed to mix it up some more. */
644 :
645 4321156 : pinfo[ i ].in_same = 0U;
646 4321156 : pinfo[ i ].heap_prio = fd_uint_hash( seed ^ (uint)i ) & ((1U<<30)-1U);
647 4321156 : pinfo[ i ].stack_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
648 4321156 : pinfo[ i ].cycle_tag = 0U;
649 :
650 4321156 : ulong tag = pinfo[ i ].tag;
651 4321156 : if( !tag ) { /* Not used ... make it available for reuse below */
652 4321156 : fd_wksp_private_idle_stack_push( i, wksp, pinfo );
653 4321156 : continue;
654 4321156 : }
655 :
656 0 : pinfo[ i ].prev_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
657 0 : pinfo[ i ].next_cidx = fd_wksp_private_pinfo_cidx( FD_WKSP_PRIVATE_PINFO_IDX_NULL );
658 :
659 0 : if( FD_UNLIKELY( fd_wksp_private_used_treap_insert( i, wksp, pinfo ) ) ) return FD_WKSP_ERR_CORRUPT; /* Logs details */
660 0 : }
661 1 : } while(0);
662 :
663 : /* At this point, a partition is either in the idle stack or used
664 : treap. Further, we have:
665 :
666 : | used | idle
667 : ----------+----------------------------+--------
668 : gaddr_* | non-empty range | 0
669 : | no overlap with other used | 0
670 : tag | non-zero | 0
671 : in_same | 0 | 0
672 : heap_prio | randomized | randomized
673 : prev | NULL | NULL
674 : next | NULL | NULL
675 : left | used treap managed | NULL
676 : right | used treap managed | NULL
677 : same | used treap managed (NULL) | NULL
678 : parent | used treap managed | idle stack managed
679 : stack | wksp managed | wksp managed
680 : cycle_tag | wksp managed | wksp managed
681 :
682 : In-order traverse the used treap to rebuild the partitioning and
683 : the free treap. */
684 :
685 1 : do {
686 1 : uint * j_next_cidx_ptr = &wksp->part_head_cidx; /* Location of most recently added partition next link */
687 :
688 1 : ulong j = FD_WKSP_PRIVATE_PINFO_IDX_NULL; /* Most recently added partition */
689 1 : ulong g0 = gaddr_lo; /* Most recently added partition end */
690 :
691 1 : ulong i = fd_wksp_private_pinfo_idx( wksp->part_used_cidx );
692 1 : ulong s = FD_WKSP_PRIVATE_PINFO_IDX_NULL;
693 1 : for(;;) {
694 1 : if( fd_wksp_private_pinfo_idx_is_null( i ) ) {
695 1 : if( fd_wksp_private_pinfo_idx_is_null( s ) ) break; /* Done */
696 :
697 : /* Pop traversal stack */
698 :
699 0 : i = s;
700 0 : s = fd_wksp_private_pinfo_idx( pinfo[ i ].stack_cidx );
701 :
702 : /* Visit i */
703 :
704 0 : ulong g1 = pinfo[ i ].gaddr_lo;
705 0 : if( g1 > g0 ) { /* There's a gap between i and the most recently added partition */
706 :
707 : /* Acquire an idle partition to hold the gap */
708 :
709 0 : if( FD_UNLIKELY( fd_wksp_private_idle_stack_is_empty( wksp ) ) ) {
710 0 : FD_LOG_WARNING(( "part_max (%lu) too small to fill gap before partition %lu (tag %lu gaddr_lo %lu gaddr_hi %lu)",
711 0 : part_max, i, pinfo[i].tag, pinfo[i].gaddr_lo, pinfo[i].gaddr_hi ));
712 0 : return FD_WKSP_ERR_CORRUPT;
713 0 : }
714 0 : ulong k = fd_wksp_private_idle_stack_pop( wksp, pinfo );
715 :
716 : /* Populate the acquired partition with the gap details,
717 : append it to the wksp partitioning and insert it into the
718 : free treap. Note that stack_push/pop reset gaddr_lo,
719 : gaddr_hi, tag, in_same, {prev, next, left, right, same,
720 : parent}_cidx. It preserved heap_prio from its original
721 : assignment and didn't touch stack_cidx or cycle_tag. */
722 :
723 0 : pinfo[ k ].gaddr_lo = g0;
724 0 : pinfo[ k ].gaddr_hi = g1;
725 0 : pinfo[ k ].prev_cidx = fd_wksp_private_pinfo_cidx( j );
726 0 : *j_next_cidx_ptr = fd_wksp_private_pinfo_cidx( k );
727 0 : j_next_cidx_ptr = &pinfo[ k ].next_cidx;
728 0 : j = k;
729 0 : g0 = g1;
730 :
731 0 : fd_wksp_private_free_treap_insert( j, wksp, pinfo );
732 0 : }
733 :
734 : /* Add i to the partitioning. */
735 :
736 0 : pinfo[ i ].prev_cidx = fd_wksp_private_pinfo_cidx( j );
737 0 : *j_next_cidx_ptr = fd_wksp_private_pinfo_cidx( i );
738 0 : j_next_cidx_ptr = &pinfo[ i ].next_cidx;
739 0 : j = i;
740 0 : g0 = pinfo[ i ].gaddr_hi;
741 :
742 : /* Traverse the right subtree */
743 :
744 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].right_cidx );
745 :
746 0 : } else {
747 :
748 : /* Push i to the stack and recurse on the left subtree. */
749 :
750 0 : pinfo[ i ].stack_cidx = fd_wksp_private_pinfo_cidx( s );
751 0 : s = i;
752 0 : i = fd_wksp_private_pinfo_idx( pinfo[ i ].left_cidx );
753 :
754 0 : }
755 1 : }
756 :
757 1 : if( g0 < gaddr_hi ) { /* Have final gap to fill */
758 :
759 : /* This works the same as the above */
760 :
761 1 : if( FD_UNLIKELY( fd_wksp_private_idle_stack_is_empty( wksp ) ) ) {
762 0 : FD_LOG_WARNING(( "part_max (%lu) too small to complete partitioning", part_max ));
763 0 : return FD_WKSP_ERR_CORRUPT;
764 0 : }
765 1 : ulong k = fd_wksp_private_idle_stack_pop( wksp, pinfo );
766 :
767 1 : pinfo[ k ].gaddr_lo = g0;
768 1 : pinfo[ k ].gaddr_hi = gaddr_hi;
769 1 : pinfo[ k ].prev_cidx = fd_wksp_private_pinfo_cidx( j );
770 1 : *j_next_cidx_ptr = fd_wksp_private_pinfo_cidx( k );
771 1 : j_next_cidx_ptr = &pinfo[ k ].next_cidx;
772 1 : j = k;
773 : //g0 = gaddr_hi;
774 :
775 1 : fd_wksp_private_free_treap_insert( j, wksp, pinfo );
776 1 : }
777 :
778 1 : wksp->part_tail_cidx = fd_wksp_private_pinfo_cidx( j );
779 :
780 1 : } while(0);
781 :
782 1 : return FD_WKSP_SUCCESS;
783 1 : }
|
