Line data Source code
1 : #include "fd_vm_syscall.h"
2 :
3 : #include "../../../ballet/base64/fd_base64.h"
4 : #include "../../../ballet/utf8/fd_utf8.h"
5 : #include "../../runtime/sysvar/fd_sysvar.h"
6 : #include "../../runtime/sysvar/fd_sysvar_clock.h"
7 : #include "../../runtime/sysvar/fd_sysvar_epoch_schedule.h"
8 : #include "../../runtime/context/fd_exec_txn_ctx.h"
9 : #include "../../runtime/context/fd_exec_instr_ctx.h"
10 :
11 : int
12 : fd_vm_syscall_abort( /**/ void * _vm,
13 : FD_PARAM_UNUSED ulong r1,
14 : FD_PARAM_UNUSED ulong r2,
15 : FD_PARAM_UNUSED ulong r3,
16 : FD_PARAM_UNUSED ulong r4,
17 : FD_PARAM_UNUSED ulong r5,
18 0 : FD_PARAM_UNUSED ulong * _ret ) {
19 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/mod.rs#L630 */
20 0 : fd_vm_t * vm = (fd_vm_t *)_vm;
21 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_ABORT );
22 0 : return FD_VM_SYSCALL_ERR_ABORT;
23 0 : }
24 :
25 : /* FD_TRANSLATE_STRING returns a read only pointer to the host address of
26 : a valid utf8 string, or it errors.
27 :
28 : Analogous of Agave's translate_string_and_do().
29 : https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/mod.rs#L601
30 :
31 : As of v0.2.6, the only two usages are in syscall panic and syscall log. */
32 9 : #define FD_TRANSLATE_STRING( vm, vaddr, msg_sz ) (__extension__({ \
33 9 : char const * msg = FD_VM_MEM_SLICE_HADDR_LD( vm, vaddr, FD_VM_ALIGN_RUST_U8, msg_sz ); \
34 9 : if( FD_UNLIKELY( !fd_utf8_verify( msg, msg_sz ) ) ) { \
35 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_STRING ); \
36 0 : return FD_VM_SYSCALL_ERR_INVALID_STRING; \
37 0 : } \
38 9 : msg; \
39 9 : }))
40 :
41 : int
42 : fd_vm_syscall_sol_panic( /**/ void * _vm,
43 : /**/ ulong file_vaddr,
44 : /**/ ulong file_sz,
45 : /**/ ulong line,
46 : /**/ ulong column,
47 : FD_PARAM_UNUSED ulong r5,
48 0 : FD_PARAM_UNUSED ulong * _ret ) {
49 0 : fd_vm_t * vm = (fd_vm_t *)_vm;
50 :
51 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/mod.rs#L637
52 :
53 : Note: this syscall is not used by the Rust SDK, only by the C SDK.
54 : Rust transforms `panic!()` into a log, followed by an abort.
55 : It's unclear if this syscall actually makes any sense... */
56 0 : FD_VM_CU_UPDATE( vm, file_sz );
57 :
58 : /* Validate string */
59 0 : FD_TRANSLATE_STRING( vm, file_vaddr, file_sz );
60 :
61 : /* Note: we truncate the log, ignoring file, line, column.
62 : As mentioned above, it's unclear if anyone is even using this syscall,
63 : so dealing with the complexity of Agave's log is a waste of time. */
64 0 : (void)line;
65 0 : (void)column;
66 :
67 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_PANIC );
68 0 : return FD_VM_SYSCALL_ERR_PANIC;
69 0 : }
70 :
71 : int
72 : fd_vm_syscall_sol_log( /**/ void * _vm,
73 : /**/ ulong msg_vaddr,
74 : /**/ ulong msg_sz,
75 : FD_PARAM_UNUSED ulong r3,
76 : FD_PARAM_UNUSED ulong r4,
77 : FD_PARAM_UNUSED ulong r5,
78 9 : /**/ ulong * _ret ) {
79 9 : fd_vm_t * vm = (fd_vm_t *)_vm;
80 :
81 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L5 */
82 :
83 9 : FD_VM_CU_UPDATE( vm, fd_ulong_max( msg_sz, FD_VM_SYSCALL_BASE_COST ) );
84 :
85 : /* Note: when msg_sz==0, msg can be undefined.
86 : fd_log_collector_program_log() handles it. */
87 9 : const char * buf = FD_TRANSLATE_STRING( vm, msg_vaddr, msg_sz );
88 0 : fd_log_collector_program_log( vm->instr_ctx, buf, msg_sz );
89 :
90 9 : *_ret = 0UL;
91 9 : return FD_VM_SUCCESS;
92 18 : }
93 :
94 : int
95 : fd_vm_syscall_sol_log_64( void * _vm,
96 : ulong r1,
97 : ulong r2,
98 : ulong r3,
99 : ulong r4,
100 : ulong r5,
101 3 : ulong * _ret ) {
102 3 : fd_vm_t * vm = (fd_vm_t *)_vm;
103 :
104 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L37 */
105 :
106 3 : FD_VM_CU_UPDATE( vm, FD_VM_LOG_64_UNITS );
107 :
108 : /* Max msg_sz: 46 - 15 + 16*5 = 111 < 127 => we can use printf */
109 0 : fd_log_collector_printf_dangerous_max_127( vm->instr_ctx,
110 3 : "Program log: 0x%lx, 0x%lx, 0x%lx, 0x%lx, 0x%lx", r1, r2, r3, r4, r5 );
111 :
112 3 : *_ret = 0UL;
113 3 : return FD_VM_SUCCESS;
114 3 : }
115 :
116 : int
117 : fd_vm_syscall_sol_log_compute_units( /**/ void * _vm,
118 : FD_PARAM_UNUSED ulong r1,
119 : FD_PARAM_UNUSED ulong r2,
120 : FD_PARAM_UNUSED ulong r3,
121 : FD_PARAM_UNUSED ulong r4,
122 : FD_PARAM_UNUSED ulong r5,
123 0 : /**/ ulong * _ret ) {
124 0 : fd_vm_t * vm = (fd_vm_t *)_vm;
125 :
126 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L60 */
127 :
128 0 : FD_VM_CU_UPDATE( vm, FD_VM_SYSCALL_BASE_COST );
129 :
130 : /* Max msg_sz: 40 - 3 + 20 = 57 < 127 => we can use printf */
131 0 : fd_log_collector_printf_dangerous_max_127( vm->instr_ctx,
132 0 : "Program consumption: %lu units remaining", vm->cu );
133 :
134 0 : *_ret = 0UL;
135 0 : return FD_VM_SUCCESS;
136 0 : }
137 :
138 : int
139 : fd_vm_syscall_sol_log_pubkey( /**/ void * _vm,
140 : /**/ ulong pubkey_vaddr,
141 : FD_PARAM_UNUSED ulong r2,
142 : FD_PARAM_UNUSED ulong r3,
143 : FD_PARAM_UNUSED ulong r4,
144 : FD_PARAM_UNUSED ulong r5,
145 0 : /**/ ulong * _ret ) {
146 0 : fd_vm_t * vm = (fd_vm_t *)_vm;
147 :
148 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L84 */
149 :
150 0 : FD_VM_CU_UPDATE( vm, FD_VM_LOG_PUBKEY_UNITS );
151 :
152 0 : void const * pubkey = FD_VM_MEM_HADDR_LD( vm, pubkey_vaddr, FD_VM_ALIGN_RUST_PUBKEY, sizeof(fd_pubkey_t) );
153 :
154 0 : char msg[ FD_BASE58_ENCODED_32_SZ ]; ulong msg_sz;
155 0 : if( FD_UNLIKELY( fd_base58_encode_32( pubkey, &msg_sz, msg )==NULL ) ) {
156 0 : return FD_VM_SYSCALL_ERR_INVALID_STRING;
157 0 : }
158 :
159 0 : fd_log_collector_program_log( vm->instr_ctx, msg, msg_sz );
160 :
161 0 : *_ret = 0UL;
162 0 : return FD_VM_SUCCESS;
163 0 : }
164 :
165 : int
166 : fd_vm_syscall_sol_log_data( /**/ void * _vm,
167 : /**/ ulong slice_vaddr,
168 : /**/ ulong slice_cnt,
169 : FD_PARAM_UNUSED ulong r3,
170 : FD_PARAM_UNUSED ulong r4,
171 : FD_PARAM_UNUSED ulong r5,
172 3 : /**/ ulong * _ret ) {
173 3 : fd_vm_t * vm = (fd_vm_t *)_vm;
174 :
175 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L109
176 :
177 : Note: this is implemented following Agave's perverse behavior.
178 : We need to loop the slice multiple times to match the exact error,
179 : first compute budget, then memory mapping.
180 : And finally, we can loop to log. */
181 :
182 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L121 */
183 :
184 3 : FD_VM_CU_UPDATE( vm, FD_VM_SYSCALL_BASE_COST );
185 :
186 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L123-L128 */
187 :
188 6 : fd_vm_vec_t const * slice = (fd_vm_vec_t const *)FD_VM_MEM_SLICE_HADDR_LD( vm, slice_vaddr, FD_VM_ALIGN_RUST_SLICE_U8_REF,
189 6 : fd_ulong_sat_mul( slice_cnt, sizeof(fd_vm_vec_t) ) );
190 :
191 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L130-L135 */
192 :
193 3 : FD_VM_CU_UPDATE( vm, fd_ulong_sat_mul( FD_VM_SYSCALL_BASE_COST, slice_cnt ) );
194 :
195 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L136-L141 */
196 :
197 18 : for( ulong i=0UL; i<slice_cnt; i++ ) {
198 15 : FD_VM_CU_UPDATE( vm, slice[i].len );
199 15 : }
200 :
201 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L145-L152 */
202 :
203 3 : ulong msg_sz = 14UL; /* "Program data: ", with space */
204 18 : for( ulong i=0UL; i<slice_cnt; i++ ) {
205 15 : ulong cur_len = slice[i].len;
206 : /* This fails the syscall in case of memory mapping issues */
207 30 : FD_VM_MEM_SLICE_HADDR_LD( vm, slice[i].addr, FD_VM_ALIGN_RUST_U8, cur_len );
208 : /* Every buffer will be base64 encoded + space separated */
209 0 : msg_sz += (slice[i].len + 2)/3*4 + (i > 0);
210 30 : }
211 :
212 : /* https://github.com/anza-xyz/agave/blob/v2.0.6/programs/bpf_loader/src/syscalls/logging.rs#L156 */
213 :
214 3 : char msg[ FD_LOG_COLLECTOR_MAX ];
215 3 : ulong bytes_written = fd_log_collector_check_and_truncate( &vm->instr_ctx->txn_ctx->log_collector, msg_sz );
216 3 : if( FD_LIKELY( bytes_written < ULONG_MAX ) ) {
217 3 : fd_memcpy( msg, "Program data: ", 14 );
218 3 : char * buf = msg + 14;
219 :
220 18 : for( ulong i=0UL; i<slice_cnt; i++ ) {
221 15 : ulong cur_len = slice[i].len;
222 30 : void const * bytes = FD_VM_MEM_SLICE_HADDR_LD( vm, slice[i].addr, FD_VM_ALIGN_RUST_U8, cur_len );
223 :
224 15 : if( i ) { *buf = ' '; ++buf; } /* skip first */
225 30 : buf += fd_base64_encode( buf, bytes, cur_len );
226 30 : }
227 3 : FD_TEST( (ulong)(buf-msg)==msg_sz );
228 :
229 3 : fd_log_collector_msg( vm->instr_ctx, msg, msg_sz );
230 3 : }
231 :
232 3 : *_ret = 0;
233 3 : return FD_VM_SUCCESS;
234 3 : }
235 :
236 : int
237 : fd_vm_syscall_sol_alloc_free( /**/ void * _vm,
238 : /**/ ulong sz,
239 : /**/ ulong free_vaddr,
240 : FD_PARAM_UNUSED ulong r3,
241 : FD_PARAM_UNUSED ulong r4,
242 : FD_PARAM_UNUSED ulong r5,
243 0 : /**/ ulong * _ret ) {
244 0 : fd_vm_t * vm = (fd_vm_t *)_vm;
245 :
246 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mod.rs#L666 */
247 :
248 : /* This syscall is ... uh ... problematic. But the community has
249 : already recognized this and deprecated it:
250 :
251 : https://github.com/solana-labs/solana/blob/v1.17.23/sdk/src/feature_set.rs#L846
252 :
253 : Unfortunately, old code never dies so, practically, this will need
254 : to be supported until the heat death of the universe.
255 :
256 : The most serious issue is that there is nothing to stop VM code
257 : making a decision based on the _location_ of the returned
258 : allocation. If different validator implementations use different
259 : allocator algorithms, though each implementation would behave
260 : functionally correct in isolation, the VM code that uses it would
261 : actually break consensus.
262 :
263 : As a result, every validator needs to use a bit-for-bit identical
264 : allocation algorithm. Fortunately, Solana is just using a basic
265 : bump allocator:
266 :
267 : https://github.com/solana-labs/solana/blob/v1.17.23/program-runtime/src/invoke_context.rs#L122-L148
268 :
269 : vm->heap_{sz,max} and the below replicate this exactly.
270 :
271 : Another major issue is that this alloc doesn't always conform
272 : typical malloc/free semantics (e.g. C/C++ requires malloc to have
273 : an alignment safe for primitive types ... 8 for the Solana machine
274 : model). This is clearly to support backward compat with older VM
275 : code (though ideally a malloc syscall should have behaved like ...
276 : well ... malloc from day 1). So the alignment behavior below is a
277 : bug-for-bug replication of that:
278 :
279 : https://github.com/solana-labs/solana/blob/v1.17.23/programs/bpf_loader/src/syscalls/mod.rs#L645-L681
280 : https://github.com/solana-labs/solana/blob/v1.17.23/sdk/program/src/entrypoint.rs#L265-L266
281 :
282 : More generally and already ranted about elsewhere, any code that
283 : uses malloc/free style dynamic allocation is inherently broken. So
284 : this syscall should have never existed in the first place ... it
285 : just feeds the trolls. The above is just additional implementation
286 : horror because people consistent think malloc/free is much simpler
287 : than it actually is. This is also an example of how quickly
288 : mistakes fossilize and become a thorn-in-the-side forever.
289 :
290 : IMPORTANT SAFETY TIP! heap_start must be non zero and both
291 : heap_start and heap_end should have an alignment of at least 8.
292 : This existing runtime policies around heap implicitly satisfy this.
293 :
294 : IMPORTANT SAFETY TIP! The specification for Rust's align_offset
295 : doesn't seem to provide a strong guarantee that it will return the
296 : minimal positive offset necessary to align pointers. It is
297 : possible for a "conforming" Rust compiler to break consensus by
298 : using a different align_offset implementation that aligned pointer
299 : between different compilations of the Solana validator and the
300 : below. */
301 :
302 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mod.rs#L676-L680 */
303 :
304 0 : ulong align = fd_vm_is_check_align_enabled( vm ) ? 8UL : FD_VM_ALIGN_RUST_U8;
305 :
306 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mod.rs#L681-L683
307 : Nothing to do. This section can't error, see:
308 : https://doc.rust-lang.org/1.81.0/src/core/alloc/layout.rs.html#70
309 : https://doc.rust-lang.org/1.81.0/src/core/alloc/layout.rs.html#100 */
310 :
311 :
312 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mod.rs#L684
313 : Nothing to do.
314 : TODO: unclear if it throw InstructionError::CallDepth
315 : https://github.com/anza-xyz/agave/blob/v2.0.8/program-runtime/src/invoke_context.rs#L662 */
316 :
317 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mod.rs#L685-L693 */
318 :
319 : /* Non-zero free address implies that this is a free() call. Since
320 : this is a bump allocator, free is a no-op. */
321 0 : if( FD_UNLIKELY( free_vaddr ) ) {
322 0 : *_ret = 0UL;
323 0 : return FD_VM_SUCCESS;
324 0 : }
325 :
326 :
327 0 : ulong heap_sz = fd_ulong_align_up( vm->heap_sz, align );
328 0 : ulong heap_vaddr = fd_ulong_sat_add ( heap_sz, FD_VM_MEM_MAP_HEAP_REGION_START );
329 0 : /**/ heap_sz = fd_ulong_sat_add ( heap_sz, sz );
330 :
331 0 : if( FD_UNLIKELY( heap_sz > vm->heap_max ) ) { /* Not enough free memory */
332 0 : *_ret = 0UL;
333 0 : return FD_VM_SUCCESS;
334 0 : }
335 :
336 0 : vm->heap_sz = heap_sz;
337 :
338 0 : *_ret = heap_vaddr;
339 0 : return FD_VM_SUCCESS;
340 0 : }
341 :
342 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mem_ops.rs#L145 */
343 : int
344 : fd_vm_memmove( fd_vm_t * vm,
345 : ulong dst_vaddr,
346 : ulong src_vaddr,
347 66 : ulong sz ) {
348 66 : if( FD_UNLIKELY( !sz ) ) {
349 0 : return FD_VM_SUCCESS;
350 0 : }
351 :
352 66 : if( !vm->direct_mapping ) {
353 : /* https://github.com/anza-xyz/agave/blob/v2.3.1/programs/bpf_loader/src/syscalls/mem_ops.rs#L188-L192 */
354 24 : fd_vm_haddr_query_t dst_ref_mut_query = {
355 24 : .vaddr = dst_vaddr,
356 24 : .align = FD_VM_ALIGN_RUST_U8,
357 24 : .sz = sz,
358 24 : .is_slice = 1,
359 24 : };
360 :
361 24 : fd_vm_haddr_query_t * queries[] = { &dst_ref_mut_query };
362 24 : FD_VM_TRANSLATE_MUT( vm, queries );
363 :
364 48 : void const * src = FD_VM_MEM_HADDR_LD( vm, src_vaddr, FD_VM_ALIGN_RUST_U8, sz );
365 0 : memmove( dst_ref_mut_query.haddr, src, sz );
366 48 : } else {
367 : /* If the src and dst vaddrs overlap and src_vaddr < dst_vaddr, Agave iterates through input regions backwards
368 : to maintain correct memmove behavior for overlapping cases. Although this logic should only apply to the src and dst
369 : vaddrs being in the input data region (since that is the only possible case you could have overlapping, chunked-up memmoves),
370 : Agave will iterate backwards in ANY region. If it eventually reaches the end of a region after iterating backwards and
371 : hits an access violation, the bytes from [region_begin, start_vaddr] will still be written to, causing fuzzing mismatches.
372 : In this case, if we didn't have the reverse flag, we would have thrown an access violation before any bytes were copied.
373 : The same logic applies to memmoves that go past the high end of a region - reverse iteration logic would throw an access
374 : violation before any bytes were copied, while the current logic would copy the bytes until the end of the region.
375 : https://github.com/anza-xyz/agave/blob/v2.1.0/programs/bpf_loader/src/syscalls/mem_ops.rs#L184 */
376 42 : uchar reverse = !!( dst_vaddr >= src_vaddr && dst_vaddr - src_vaddr < sz );
377 :
378 : /* In reverse calculations, start from the rightmost vaddr that will be accessed (note the - 1). */
379 42 : ulong dst_vaddr_begin = reverse ? fd_ulong_sat_add( dst_vaddr, sz - 1UL ) : dst_vaddr;
380 42 : ulong src_vaddr_begin = reverse ? fd_ulong_sat_add( src_vaddr, sz - 1UL ) : src_vaddr;
381 :
382 : /* Find the correct src and dst haddrs to start operating from. If the src or dst vaddrs
383 : belong to the input data region (4), keep track of region statistics to memmove in chunks. */
384 42 : ulong dst_region = FD_VADDR_TO_REGION( dst_vaddr_begin );
385 42 : uchar dst_is_input_mem_region = ( dst_region==FD_VM_INPUT_REGION );
386 42 : ulong dst_offset = dst_vaddr_begin & FD_VM_OFFSET_MASK;
387 42 : ulong dst_region_idx = 0UL;
388 42 : ulong dst_bytes_rem_in_cur_region;
389 42 : uchar * dst_haddr;
390 42 : if( dst_is_input_mem_region ) {
391 18 : FD_VM_MEM_HADDR_AND_REGION_IDX_FROM_INPUT_REGION_CHECKED( vm, dst_offset, dst_region_idx, dst_haddr );
392 18 : if( FD_UNLIKELY( !vm->input_mem_regions[ dst_region_idx ].is_writable ) ) {
393 0 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
394 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
395 0 : }
396 18 : if( FD_UNLIKELY( reverse ) ) {
397 : /* Bytes remaining between region begin and current position (+ 1 for inclusive region beginning). */
398 6 : dst_bytes_rem_in_cur_region = fd_ulong_sat_sub( dst_offset + 1UL, vm->input_mem_regions[ dst_region_idx ].vaddr_offset );
399 12 : } else {
400 : /* Bytes remaining between current position and region end. */
401 12 : dst_bytes_rem_in_cur_region = fd_ulong_sat_sub( vm->input_mem_regions[ dst_region_idx ].region_sz, ( dst_offset - vm->input_mem_regions[ dst_region_idx ].vaddr_offset ) );
402 12 : }
403 24 : } else {
404 24 : dst_haddr = (uchar*)FD_VM_MEM_HADDR_ST_NO_SZ_CHECK( vm, dst_vaddr_begin, FD_VM_ALIGN_RUST_U8 );
405 :
406 18 : if( FD_UNLIKELY( reverse ) ) {
407 : /* Bytes remaining is minimum of the offset from the beginning of the current
408 : region (+1 for inclusive region beginning) and the number of storable bytes in the region. */
409 3 : dst_bytes_rem_in_cur_region = fd_ulong_min( vm->region_st_sz[ dst_region ], dst_offset + 1UL );
410 :
411 15 : } else {
412 : /* Bytes remaining is the number of writable bytes left in the region */
413 15 : dst_bytes_rem_in_cur_region = fd_ulong_sat_sub( vm->region_st_sz[ dst_region ], dst_offset );
414 15 : }
415 18 : }
416 :
417 : /* Logic for src vaddr translation is similar to above excluding any writable checks. */
418 36 : ulong src_region = FD_VADDR_TO_REGION( src_vaddr_begin );
419 36 : uchar src_is_input_mem_region = ( src_region==FD_VM_INPUT_REGION );
420 36 : ulong src_offset = src_vaddr_begin & FD_VM_OFFSET_MASK;
421 36 : ulong src_region_idx = 0UL;
422 36 : ulong src_bytes_rem_in_cur_region;
423 36 : uchar * src_haddr;
424 36 : if( src_is_input_mem_region ) {
425 18 : FD_VM_MEM_HADDR_AND_REGION_IDX_FROM_INPUT_REGION_CHECKED( vm, src_offset, src_region_idx, src_haddr );
426 18 : if( FD_UNLIKELY( reverse ) ) {
427 6 : src_bytes_rem_in_cur_region = fd_ulong_sat_sub( src_offset + 1UL, vm->input_mem_regions[ src_region_idx ].vaddr_offset );
428 12 : } else {
429 12 : src_bytes_rem_in_cur_region = fd_ulong_sat_sub( vm->input_mem_regions[ src_region_idx ].region_sz, ( src_offset - vm->input_mem_regions[ src_region_idx ].vaddr_offset ) );
430 12 : }
431 18 : } else {
432 54 : src_haddr = (uchar*)FD_VM_MEM_HADDR_LD_NO_SZ_CHECK( vm, src_vaddr_begin, FD_VM_ALIGN_RUST_U8 );
433 :
434 18 : if( FD_UNLIKELY( reverse ) ) {
435 3 : src_bytes_rem_in_cur_region = fd_ulong_min( vm->region_ld_sz[ src_region ], src_offset + 1UL );
436 :
437 15 : } else {
438 15 : src_bytes_rem_in_cur_region = fd_ulong_sat_sub( vm->region_ld_sz[ src_region ], src_offset );
439 15 : }
440 54 : }
441 :
442 : /* Short circuit: if the number of copyable bytes stays within all memory regions,
443 : just memmove and return. This is a majority case in mainnet, devnet, and testnet.
444 : Someone would have to be very crafty and clever to construct a transaction that
445 : deploys and invokes a custom program that does not fall into this branch. */
446 36 : if( FD_LIKELY( sz<=dst_bytes_rem_in_cur_region && sz<=src_bytes_rem_in_cur_region ) ) {
447 21 : if( FD_UNLIKELY( reverse ) ) {
448 : /* In the reverse iteration case, the haddrs point to the end of the region here. Since the
449 : above checks guarantee that there are enough bytes left in the src and dst regions to do
450 : a direct memmove, we can just subtract (sz-1) from the haddrs, memmove, and return. */
451 3 : memmove( dst_haddr - sz + 1UL, src_haddr - sz + 1UL, sz );
452 18 : } else {
453 : /* In normal iteration, the haddrs correspond to the correct starting point for the memcpy,
454 : so no further translation has to be done. */
455 18 : memmove( dst_haddr, src_haddr, sz );
456 18 : }
457 21 : return FD_VM_SUCCESS;
458 21 : }
459 :
460 : /* Copy over the bytes from each region in chunks. */
461 57 : while( sz>0UL ) {
462 : /* End of region case */
463 45 : if( FD_UNLIKELY( src_bytes_rem_in_cur_region==0UL ) ) {
464 : /* Same as above, except no writable checks. */
465 30 : if( FD_LIKELY( !reverse &&
466 30 : src_is_input_mem_region &&
467 30 : src_region_idx+1UL<vm->input_mem_regions_cnt ) ) {
468 12 : if( FD_UNLIKELY( vm->input_mem_regions[ src_region_idx+1UL ].is_acct_data != vm->input_mem_regions[ src_region_idx ].is_acct_data ) ) {
469 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_LENGTH );
470 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
471 0 : }
472 12 : src_region_idx++;
473 12 : src_haddr = (uchar*)vm->input_mem_regions[ src_region_idx ].haddr;
474 18 : } else if( FD_LIKELY( reverse && src_region_idx>0UL ) ) {
475 15 : if( FD_UNLIKELY( vm->input_mem_regions[ src_region_idx-1UL ].is_acct_data != vm->input_mem_regions[ src_region_idx ].is_acct_data ) ) {
476 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_LENGTH );
477 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
478 0 : }
479 15 : src_region_idx--;
480 15 : src_haddr = (uchar*)vm->input_mem_regions[ src_region_idx ].haddr + vm->input_mem_regions[ src_region_idx ].region_sz - 1UL;
481 15 : } else {
482 3 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
483 3 : return FD_VM_SYSCALL_ERR_SEGFAULT;
484 3 : }
485 27 : src_bytes_rem_in_cur_region = vm->input_mem_regions[ src_region_idx ].region_sz;
486 27 : }
487 42 : if( FD_UNLIKELY( dst_bytes_rem_in_cur_region==0UL ) ) {
488 : /* Only proceed if:
489 : - We are in the input memory region
490 : - There are remaining input memory regions to copy from (for both regular and reverse iteration orders)
491 : - The next input memory region is writable
492 : Fail otherwise. */
493 9 : if( FD_LIKELY( !reverse &&
494 9 : dst_is_input_mem_region &&
495 9 : dst_region_idx+1UL<vm->input_mem_regions_cnt &&
496 9 : vm->input_mem_regions[ dst_region_idx+1UL ].is_writable ) ) {
497 6 : if( FD_UNLIKELY( vm->input_mem_regions[ dst_region_idx+1UL ].is_acct_data != vm->input_mem_regions[ dst_region_idx ].is_acct_data ) ) {
498 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_LENGTH );
499 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
500 0 : }
501 : /* In normal iteration, we move the haddr to the beginning of the next region. */
502 6 : dst_region_idx++;
503 6 : dst_haddr = (uchar*)vm->input_mem_regions[ dst_region_idx ].haddr;
504 6 : } else if( FD_LIKELY( reverse &&
505 3 : dst_region_idx>0UL &&
506 3 : vm->input_mem_regions[ dst_region_idx-1UL ].is_writable ) ) {
507 3 : if( FD_UNLIKELY( vm->input_mem_regions[ dst_region_idx-1UL ].is_acct_data != vm->input_mem_regions[ dst_region_idx ].is_acct_data ) ) {
508 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_LENGTH );
509 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
510 0 : }
511 : /* Note that when reverse iterating, we set the haddr to the END of the PREVIOUS region. */
512 3 : dst_region_idx--;
513 3 : dst_haddr = (uchar*)vm->input_mem_regions[ dst_region_idx ].haddr + vm->input_mem_regions[ dst_region_idx ].region_sz - 1UL;
514 3 : } else {
515 0 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
516 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
517 0 : }
518 9 : dst_bytes_rem_in_cur_region = vm->input_mem_regions[ dst_region_idx ].region_sz;
519 9 : }
520 :
521 : /* Number of bytes to operate on in this iteration is the min of:
522 : - number of bytes left to copy
523 : - bytes left in the current src region
524 : - bytes left in the current dst region */
525 42 : ulong num_bytes_to_copy = fd_ulong_min( sz, fd_ulong_min( src_bytes_rem_in_cur_region, dst_bytes_rem_in_cur_region ) );
526 42 : if( FD_UNLIKELY( reverse ) ) {
527 21 : memmove( dst_haddr - num_bytes_to_copy + 1UL, src_haddr - num_bytes_to_copy + 1UL, num_bytes_to_copy );
528 21 : dst_haddr -= num_bytes_to_copy;
529 21 : src_haddr -= num_bytes_to_copy;
530 21 : } else {
531 21 : memmove( dst_haddr, src_haddr, num_bytes_to_copy );
532 21 : dst_haddr += num_bytes_to_copy;
533 21 : src_haddr += num_bytes_to_copy;
534 21 : }
535 :
536 : /* Update size trackers */
537 42 : sz -= num_bytes_to_copy;
538 42 : src_bytes_rem_in_cur_region -= num_bytes_to_copy;
539 42 : dst_bytes_rem_in_cur_region -= num_bytes_to_copy;
540 42 : }
541 15 : }
542 :
543 27 : return FD_VM_SUCCESS;
544 66 : }
545 :
546 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mem_ops.rs#L41 */
547 : int
548 : fd_vm_syscall_sol_memmove( /**/ void * _vm,
549 : /**/ ulong dst_vaddr,
550 : /**/ ulong src_vaddr,
551 : /**/ ulong sz,
552 : FD_PARAM_UNUSED ulong r4,
553 : FD_PARAM_UNUSED ulong r5,
554 36 : /**/ ulong * _ret ) {
555 36 : fd_vm_t * vm = (fd_vm_t *)_vm;
556 :
557 36 : FD_VM_CU_MEM_OP_UPDATE( vm, sz );
558 :
559 0 : *_ret = 0;
560 :
561 : /* No overlap check for memmove. */
562 36 : return fd_vm_memmove( vm, dst_vaddr, src_vaddr, sz );
563 36 : }
564 :
565 : /* https://github.com/anza-xyz/agave/blob/v2.0.8/programs/bpf_loader/src/syscalls/mem_ops.rs#L18 */
566 : int
567 : fd_vm_syscall_sol_memcpy( /**/ void * _vm,
568 : /**/ ulong dst_vaddr,
569 : /**/ ulong src_vaddr,
570 : /**/ ulong sz,
571 : FD_PARAM_UNUSED ulong r4,
572 : FD_PARAM_UNUSED ulong r5,
573 48 : /**/ ulong * _ret ) {
574 48 : fd_vm_t * vm = (fd_vm_t *)_vm;
575 :
576 48 : FD_VM_CU_MEM_OP_UPDATE( vm, sz );
577 :
578 0 : *_ret = 0;
579 :
580 : /* Exact same as memmove, except also check overlap.
581 : https://github.com/anza-xyz/agave/blob/v2.2.17/programs/bpf_loader/src/syscalls/mem_ops.rs#L45 */
582 48 : FD_VM_MEM_CHECK_NON_OVERLAPPING( vm, src_vaddr, sz, dst_vaddr, sz );
583 :
584 30 : return fd_vm_memmove( vm, dst_vaddr, src_vaddr, sz );
585 48 : }
586 :
587 : int
588 : fd_vm_syscall_sol_memcmp( /**/ void * _vm,
589 : /**/ ulong m0_vaddr,
590 : /**/ ulong m1_vaddr,
591 : /**/ ulong sz,
592 : /**/ ulong out_vaddr,
593 : FD_PARAM_UNUSED ulong r5,
594 15 : /**/ ulong * _ret ) {
595 15 : *_ret = 0;
596 15 : fd_vm_t * vm = (fd_vm_t *)_vm;
597 :
598 : /* https://github.com/anza-xyz/agave/blob/v2.2.17/programs/bpf_loader/src/syscalls/mem_ops.rs#L84 */
599 :
600 15 : FD_VM_CU_MEM_OP_UPDATE( vm, sz );
601 :
602 : /* Note: though this behaves like a normal C-style memcmp, we can't
603 : use the compilers / libc memcmp directly because the specification
604 : doesn't provide strong enough guarantees about the return value (it
605 : only promises the sign). */
606 :
607 15 : if( !vm->direct_mapping ) {
608 0 : uchar const * m0 = (uchar const *)FD_VM_MEM_SLICE_HADDR_LD( vm, m0_vaddr, FD_VM_ALIGN_RUST_U8, sz );
609 0 : uchar const * m1 = (uchar const *)FD_VM_MEM_SLICE_HADDR_LD( vm, m1_vaddr, FD_VM_ALIGN_RUST_U8, sz );
610 :
611 : /* Silly that this doesn't use r0 to return ... slower, more edge
612 : case, different from libc style memcmp, harder to callers to use,
613 : etc ... probably too late to do anything about it now ... sigh */
614 :
615 : /* https://github.com/anza-xyz/agave/blob/v2.3.1/programs/bpf_loader/src/syscalls/mem_ops.rs#L121-L125 */
616 0 : fd_vm_haddr_query_t cmp_result_ref_mut_query = {
617 0 : .vaddr = out_vaddr,
618 0 : .align = FD_VM_ALIGN_RUST_I32,
619 0 : .sz = 4UL,
620 0 : .is_slice = 0,
621 0 : };
622 :
623 0 : fd_vm_haddr_query_t * queries[] = { &cmp_result_ref_mut_query };
624 0 : FD_VM_TRANSLATE_MUT( vm, queries );
625 :
626 0 : int out = 0;
627 0 : for( ulong i=0UL; i<sz; i++ ) {
628 0 : int i0 = (int)m0[i];
629 0 : int i1 = (int)m1[i];
630 0 : if( i0!=i1 ) {
631 0 : out = i0 - i1;
632 0 : break;
633 0 : }
634 0 : }
635 :
636 0 : fd_memcpy( cmp_result_ref_mut_query.haddr, &out, 4UL ); /* Sigh ... see note above (and might be unaligned ... double sigh) */
637 :
638 0 : return FD_VM_SUCCESS;
639 15 : } else {
640 : /* In the case that direct mapping is enabled, the behavior for memcmps
641 : differ significantly from the non-dm case. The key difference is that
642 : invalid loads will instantly lead to errors in the non-dm case. However,
643 : when direct mapping is enabled, we will first try to memcmp the largest
644 : size valid chunk first, and will exit successfully if a difference is
645 : found without aborting from the VM. A chunk is defined as the largest
646 : valid vaddr range in both memory regions that doesn't span multiple
647 : regions.
648 :
649 : Example:
650 : fd_vm_syscall_sol_memcmp( vm, m0_addr : 0x4000, m1_vaddr : 0x2000, 0x200, ... );
651 : m0's region: m0_addr 0x4000 -> 0x4000 + 0x50 (region sz 0x50)
652 : m1's region: m1_addr 0x2000 -> 0x2000 + 0x100 (region sz 0x100)
653 : sz: 0x200
654 :
655 : Case 1: 0x4000 -> 0x4050 does have the same bytes as 0x2000 -> 0x2050
656 : Case 2: 0x4000 -> 0x4050 does NOT have the same bytes as 0x2000 -> 0x2050
657 :
658 : Pre-DM:
659 : This will fail out before any bytes are compared because the memory
660 : translation is done first.
661 :
662 : Post-DM:
663 : For case 1, the memcmp will return an error and the VM will exit because
664 : the memcmp will eventually try to access 0x4051 which is invalid. First
665 : 0x50 bytes are compared, but the next chunk will lead to an invalid
666 : access.
667 :
668 : For case 2, the memcmp will first translate the first 0x50 bytes and will
669 : see that the bytes are not the same. This will lead to the syscall
670 : exiting out successfully without detecting the access violation.
671 :
672 : https://github.com/anza-xyz/agave/blob/v2.0.10/programs/bpf_loader/src/syscalls/mem_ops.rs#L213
673 : */
674 :
675 : /* TODO: Refactor to use `FD_VM_TRANSLATE_MUT` macro when direct mapping is rewritten */
676 15 : void * _out = FD_VM_MEM_HADDR_ST( vm, out_vaddr, FD_VM_ALIGN_RUST_I32, 4UL );
677 0 : int out = 0;
678 :
679 : /* Lookup host address chunks. Try to do a standard memcpy if the regions
680 : do not cross memory regions. The translation logic is different if the
681 : the virtual address region is the input region vs. not. See the comment
682 : in fd_bpf_loader_serialization for more details on how the input
683 : region is different from other regions. The input data region will try
684 : to lookup the number of remaining bytes in the specific data region. If
685 : the memory access is not in the input data region, assume the bytes in
686 : the current region are bound by the size of the remaining bytes in the
687 : region. */
688 :
689 12 : ulong m0_region = FD_VADDR_TO_REGION( m0_vaddr );
690 12 : ulong m0_offset = m0_vaddr & FD_VM_OFFSET_MASK;
691 12 : ulong m0_region_idx = 0UL;
692 12 : ulong m0_bytes_in_cur_region = sz;
693 12 : uchar * m0_haddr = NULL;
694 12 : if( m0_region==FD_VM_INPUT_REGION ) {
695 6 : m0_region_idx = fd_vm_get_input_mem_region_idx( vm, m0_offset );
696 6 : m0_haddr = (uchar*)(vm->input_mem_regions[ m0_region_idx ].haddr + m0_offset - vm->input_mem_regions[ m0_region_idx ].vaddr_offset);
697 6 : m0_bytes_in_cur_region = fd_ulong_min( sz, fd_ulong_sat_sub( vm->input_mem_regions[ m0_region_idx ].region_sz,
698 6 : ((ulong)m0_haddr - vm->input_mem_regions[ m0_region_idx ].haddr) ) );
699 6 : } else {
700 : /* We can safely load a slice of 1 byte here because we know that we will
701 : not ever read more than the number of bytes that are left in the
702 : region. */
703 6 : m0_bytes_in_cur_region = fd_ulong_min( sz, vm->region_ld_sz[ m0_region ] - m0_offset );
704 12 : m0_haddr = (uchar *)FD_VM_MEM_SLICE_HADDR_LD_SZ_UNCHECKED( vm, m0_vaddr, FD_VM_ALIGN_RUST_U8 );
705 12 : }
706 :
707 12 : ulong m1_region = FD_VADDR_TO_REGION( m1_vaddr );
708 12 : ulong m1_offset = m1_vaddr & FD_VM_OFFSET_MASK;
709 12 : ulong m1_region_idx = 0UL;
710 12 : ulong m1_bytes_in_cur_region = sz;
711 12 : uchar * m1_haddr = NULL;
712 12 : if( m1_region==FD_VM_INPUT_REGION ) {
713 6 : m1_region_idx = fd_vm_get_input_mem_region_idx( vm, m1_offset );
714 6 : m1_haddr = (uchar*)(vm->input_mem_regions[ m1_region_idx ].haddr + m1_offset - vm->input_mem_regions[ m1_region_idx ].vaddr_offset);
715 6 : m1_bytes_in_cur_region = fd_ulong_min( sz, fd_ulong_sat_sub( vm->input_mem_regions[ m1_region_idx ].region_sz,
716 6 : ((ulong)m1_haddr - vm->input_mem_regions[ m1_region_idx ].haddr) ) );
717 6 : } else {
718 6 : m1_bytes_in_cur_region = fd_ulong_min( sz, vm->region_ld_sz[ m1_region ] - m1_offset );
719 12 : m1_haddr = (uchar *)FD_VM_MEM_SLICE_HADDR_LD_SZ_UNCHECKED( vm, m1_vaddr, FD_VM_ALIGN_RUST_U8 );
720 12 : }
721 :
722 : /* Case where the operation spans multiple regions. Copy over the bytes
723 : from each region while iterating to the next one. */
724 : /* TODO: An optimization would be to memcmp chunks at once */
725 12 : ulong m0_idx = 0UL;
726 12 : ulong m1_idx = 0UL;
727 612 : for( ulong i=0UL; i<sz; i++ ) {
728 609 : if( FD_UNLIKELY( !m0_bytes_in_cur_region ) ) {
729 : /* If the memory is not in the input region or it is the last input
730 : memory region, that means that if we don't exit now we will have
731 : an access violation. */
732 6 : if( FD_UNLIKELY( m0_region!=FD_VM_INPUT_REGION || ++m0_region_idx>=vm->input_mem_regions_cnt ) ) {
733 0 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
734 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
735 0 : }
736 6 : if( FD_UNLIKELY( vm->input_mem_regions[ m0_region_idx-1UL ].is_acct_data != vm->input_mem_regions[ m0_region_idx ].is_acct_data ) ) {
737 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_LENGTH );
738 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
739 0 : }
740 : /* Otherwise, query the next input region. */
741 6 : m0_haddr = (uchar*)vm->input_mem_regions[ m0_region_idx ].haddr;
742 6 : m0_idx = 0UL;
743 6 : m0_bytes_in_cur_region = vm->input_mem_regions[ m0_region_idx ].region_sz;
744 6 : }
745 609 : if( FD_UNLIKELY( !m1_bytes_in_cur_region ) ) {
746 0 : if( FD_UNLIKELY( m1_region!=FD_VM_INPUT_REGION || ++m1_region_idx>=vm->input_mem_regions_cnt ) ) {
747 0 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
748 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
749 0 : }
750 0 : if( FD_UNLIKELY( vm->input_mem_regions[ m1_region_idx-1UL ].is_acct_data != vm->input_mem_regions[ m1_region_idx ].is_acct_data ) ) {
751 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_LENGTH );
752 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
753 0 : }
754 0 : m1_haddr = (uchar*)vm->input_mem_regions[ m1_region_idx ].haddr;
755 0 : m1_idx = 0UL;
756 0 : m1_bytes_in_cur_region = vm->input_mem_regions[ m1_region_idx ].region_sz;
757 0 : }
758 :
759 609 : int i0 = (int)m0_haddr[ m0_idx ];
760 609 : int i1 = (int)m1_haddr[ m1_idx ];
761 609 : if( i0!=i1 ) {
762 9 : out = i0 - i1;
763 9 : break;
764 9 : }
765 :
766 600 : m0_bytes_in_cur_region--;
767 600 : m1_bytes_in_cur_region--;
768 600 : m0_idx++;
769 600 : m1_idx++;
770 600 : }
771 12 : fd_memcpy( _out, &out, 4UL ); /* Sigh ... see note above (and might be unaligned ... double sigh) */
772 12 : return FD_VM_SUCCESS;
773 12 : }
774 15 : }
775 :
776 : int
777 : fd_vm_syscall_sol_memset( /**/ void * _vm,
778 : /**/ ulong dst_vaddr,
779 : /**/ ulong c,
780 : /**/ ulong sz,
781 : FD_PARAM_UNUSED ulong r4,
782 : FD_PARAM_UNUSED ulong r5,
783 27 : /**/ ulong * _ret ) {
784 27 : fd_vm_t * vm = (fd_vm_t *)_vm;
785 27 : *_ret = 0;
786 :
787 : /* https://github.com/anza-xyz/agave/blob/v2.2.17/programs/bpf_loader/src/syscalls/mem_ops.rs#L142 */
788 :
789 27 : FD_VM_CU_MEM_OP_UPDATE( vm, sz );
790 :
791 27 : if( FD_UNLIKELY( !sz ) ) {
792 0 : return FD_VM_SUCCESS;
793 0 : }
794 :
795 27 : ulong region = FD_VADDR_TO_REGION( dst_vaddr );
796 27 : ulong offset = dst_vaddr & FD_VM_OFFSET_MASK;
797 27 : uchar * haddr;
798 :
799 27 : int b = (int)(c & 255UL);
800 :
801 27 : if( !vm->direct_mapping ) {
802 : /* https://github.com/anza-xyz/agave/blob/v2.3.1/programs/bpf_loader/src/syscalls/mem_ops.rs#L155-L159 */
803 9 : fd_vm_haddr_query_t haddr_query = {
804 9 : .vaddr = dst_vaddr,
805 9 : .align = FD_VM_ALIGN_RUST_U8,
806 9 : .sz = sz,
807 9 : .is_slice = 1,
808 9 : };
809 :
810 9 : fd_vm_haddr_query_t * queries[] = { &haddr_query };
811 9 : FD_VM_TRANSLATE_MUT( vm, queries );
812 6 : fd_memset( haddr_query.haddr, b, sz );
813 18 : } else if( region!=FD_VM_INPUT_REGION ) {
814 : /* Here we special case non-input region memsets: we try to memset
815 : as many bytes as possible until it reaches an unwritable section.
816 : This is done in order to ensure error-code conformance with
817 : Agave. */
818 9 : haddr = (uchar*)FD_VM_MEM_HADDR_ST_FAST( vm, dst_vaddr );
819 9 : ulong bytes_in_cur_region = fd_ulong_sat_sub( vm->region_st_sz[ region ], offset );
820 9 : ulong bytes_to_set = fd_ulong_min( sz, bytes_in_cur_region );
821 9 : fd_memset( haddr, b, bytes_to_set );
822 9 : if( FD_UNLIKELY( bytes_to_set<sz ) ) {
823 3 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
824 3 : return FD_VM_SYSCALL_ERR_SEGFAULT;
825 3 : }
826 9 : } else {
827 : /* In this case, we are in the input region AND direct mapping is
828 : enabled. Get the haddr and input region and check if it's
829 : writable. This means that we may potentially iterate over
830 : multiple regions. */
831 9 : ulong region_idx;
832 9 : FD_VM_MEM_HADDR_AND_REGION_IDX_FROM_INPUT_REGION_CHECKED( vm, offset, region_idx, haddr );
833 0 : ulong offset_in_cur_region = offset - vm->input_mem_regions[ region_idx ].vaddr_offset;
834 9 : ulong bytes_in_cur_region = fd_ulong_sat_sub( vm->input_mem_regions[ region_idx ].region_sz, offset_in_cur_region );
835 :
836 : /* Check that current region is writable */
837 9 : if( FD_UNLIKELY( !vm->input_mem_regions[ region_idx ].is_writable ) ) {
838 0 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
839 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
840 0 : }
841 :
842 : /* Memset goes into multiple regions. */
843 30 : while( sz>0UL ) {
844 :
845 : /* Memset bytes */
846 30 : ulong num_bytes_to_set = fd_ulong_min( sz, bytes_in_cur_region );
847 30 : fd_memset( haddr, b, num_bytes_to_set );
848 30 : sz -= num_bytes_to_set;
849 :
850 30 : if( !sz ) {
851 6 : break;
852 6 : }
853 :
854 : /* If no more regions left, break. */
855 24 : if( ++region_idx==vm->input_mem_regions_cnt ) {
856 0 : break;
857 0 : }
858 :
859 : /* Check that new region is writable. */
860 24 : if( FD_UNLIKELY( !vm->input_mem_regions[ region_idx ].is_writable ) ) {
861 3 : break;
862 3 : }
863 :
864 : /* If new region crosses into/out of account region, error out. */
865 21 : if( FD_UNLIKELY( vm->input_mem_regions[ region_idx ].is_acct_data !=
866 21 : vm->input_mem_regions[ region_idx-1UL ].is_acct_data ) ) {
867 0 : FD_VM_ERR_FOR_LOG_SYSCALL( vm, FD_VM_SYSCALL_ERR_INVALID_LENGTH );
868 0 : return FD_VM_SYSCALL_ERR_SEGFAULT;
869 0 : }
870 :
871 : /* Move haddr to next region. */
872 21 : haddr = (uchar*)vm->input_mem_regions[ region_idx ].haddr;
873 21 : bytes_in_cur_region = vm->input_mem_regions[ region_idx ].region_sz;
874 21 : }
875 :
876 : /* If we were not able to successfully set all the bytes, throw an error. */
877 9 : if( FD_UNLIKELY( sz>0 ) ) {
878 3 : FD_VM_ERR_FOR_LOG_EBPF( vm, FD_VM_ERR_EBPF_ACCESS_VIOLATION );
879 3 : return FD_VM_SYSCALL_ERR_SEGFAULT;
880 3 : }
881 9 : }
882 18 : return FD_VM_SUCCESS;
883 27 : }
|