Line data Source code
1 : #ifndef HEADER_fd_src_flamenco_vm_fd_vm_h
2 : #define HEADER_fd_src_flamenco_vm_fd_vm_h
3 :
4 : #include "fd_vm_base.h"
5 :
6 : /* A fd_vm_t is an opaque handle of a virtual machine that can execute
7 : sBPF programs. */
8 :
9 : struct fd_vm;
10 : typedef struct fd_vm fd_vm_t;
11 :
12 : /**********************************************************************/
13 : /* FIXME: MOVE TO FD_VM_PRIVATE WHEN CONSTRUCTORS READY */
14 :
15 : /* A fd_vm_shadow_t holds stack frame information not accessible from
16 : within a program. */
17 :
18 : struct fd_vm_shadow { ulong r6; ulong r7; ulong r8; ulong r9; ulong r10; ulong pc; };
19 : typedef struct fd_vm_shadow fd_vm_shadow_t;
20 :
21 : /* fd_vm_input_region_t holds information about fragmented memory regions
22 : within the larger input region. */
23 :
24 : struct __attribute__((aligned(8UL))) fd_vm_input_region {
25 : ulong vaddr_offset; /* Represents offset from the start of the input region. */
26 : ulong haddr; /* Host address corresponding to the start of the mem region. */
27 : uint region_sz; /* Size of the memory region. */
28 : uchar is_writable; /* If the region can be written to or is read-only */
29 : uchar is_acct_data; /* Set if this is an account data region (either orig data or resize buffer). */
30 : };
31 : typedef struct fd_vm_input_region fd_vm_input_region_t;
32 :
33 : /* fd_vm_acc_region_meta_t holds metadata about a given account. An array of these
34 : structs will map an instruction account index to its respective input memory
35 : region location. */
36 :
37 : struct __attribute((aligned(8UL))) fd_vm_acc_region_meta {
38 : uint region_idx;
39 : uchar has_data_region;
40 : uchar has_resizing_region;
41 : /* offset of the accounts metadata region, relative to the start of the input region.
42 : importantly, this excludes any duplicate account markers at the beginning of the "full" metadata region. */
43 : ulong metadata_region_offset;
44 : };
45 : typedef struct fd_vm_acc_region_meta fd_vm_acc_region_meta_t;
46 :
47 : /* In Agave, all the regions are 16-byte aligned in host address space. There is then an alignment check
48 : which is done inside each syscall memory translation, checking if the data is aligned in host address
49 : space. This is a layering violation, as it leaks the host address layout into the consensus model.
50 :
51 : In the future we will change this alignment check in the vm to purely operate on the virtual address space,
52 : taking advantage of the fact that Agave regions are known to be aligned. For now, we align our regions to
53 : either 8 or 16 bytes, as there are no 16-byte alignment translations in the syscalls currently:
54 : stack: 16 byte aligned
55 : heap: 16 byte aligned
56 : input: 8 byte aligned
57 : rodata: 8 byte aligned
58 :
59 : https://github.com/solana-labs/rbpf/blob/cd19a25c17ec474e6fa01a3cc3efa325f44cd111/src/ebpf.rs#L39-L40 */
60 0 : #define FD_VM_HOST_REGION_ALIGN (16UL)
61 :
62 : struct __attribute__((aligned(FD_VM_HOST_REGION_ALIGN))) fd_vm {
63 :
64 : /* VM configuration */
65 :
66 : /* FIXME: suspect these three should be replaced by some kind of VM
67 : enabled feature struct (though syscalls do seem to make additional
68 : non-trivial use of instr_ctx). */
69 :
70 : fd_exec_instr_ctx_t * instr_ctx; /* FIXME: DOCUMENT */
71 :
72 : /* FIXME: frame_max should be run time configurable by compute budget.
73 : If there is no reasonable upper bound on this, shadow and stack
74 : will need to be provided by users. */
75 :
76 : //ulong frame_max; /* Maximum number of stack frames, in [0,FD_VM_STACK_FRAME_MAX] */
77 : ulong heap_max; /* Maximum amount of heap in bytes, in [0,FD_VM_HEAP_MAX] */
78 : ulong entry_cu; /* Initial number of compute units for this program, in [0,FD_VM_COMPUTE_UNIT_LIMIT] */
79 :
80 : /* FIXME: The below are practically an exact match to the
81 : fields of an fd_sbpf_program_t (sans ELF info) */
82 :
83 : uchar const * rodata; /* Program read only data, indexed [0,rodata_sz), aligned 8 */
84 : ulong rodata_sz; /* Program read only data size in bytes, FIXME: BOUNDS? */
85 : ulong const * text; /* Program sBPF words, indexed [0,text_cnt), aligned 8 */
86 : ulong text_cnt; /* Program sBPF word count, all text words are inside the rodata */
87 : ulong text_off; /* ==(ulong)text - (ulong)rodata, relocation offset in bytes we must apply to indirect calls
88 : (callx/CALL_REGs), IMPORTANT SAFETY TIP! THIS IS IN BYTES, NOT WORDS! */
89 : ulong text_sz; /* Program sBPF size in bytes, == text_cnt*8 */
90 :
91 : ulong entry_pc; /* Initial program counter, in [0,text_cnt)
92 : FIXME: MAKE SURE NOT INTO MW INSTRUCTION, MAKE SURE VALID CALLDEST? */
93 : ulong const * calldests; /* Bit vector of local functions that can be called into, bit indexed in [0,text_cnt) */
94 : /* FIXME: ADD BIT VECTOR OF FORBIDDEN BRANCH TARGETS (E.G.
95 : INTO THE MIDDLE OF A MULTIWORD INSTRUCTION) */
96 :
97 : fd_sbpf_syscalls_t const * syscalls; /* The map of syscalls (sharable over multiple concurrently running vm) */
98 :
99 : fd_vm_trace_t * trace; /* Location to stream traces (no tracing if NULL) */
100 :
101 : /* VM execution and syscall state */
102 :
103 : /* These are used to communicate the execution and syscall state to
104 : users and syscalls. These are initialized based on the above when
105 : a program starts executing. When program halts or faults, these
106 : provide precise execution diagnostics to the user (and potential
107 : breakpoint/continue functionality in the future). When the vm
108 : makes a syscall, the vm will set these precisely and, when a
109 : syscall returns, the vm will update its internal execution state
110 : appropriately. */
111 :
112 : /* IMPORTANT SAFETY TIP! THE BEHAVIOR OF THE SYSCALL ALLOCATOR FOR
113 : HEAP_SZ MUST EXACTLY MATCH THE SOLANA VALIDATOR ALLOCATOR:
114 :
115 : https://github.com/solana-labs/solana/blob/v1.17.23/program-runtime/src/invoke_context.rs#L122-L148
116 :
117 : BIT-FOR-BIT AND BUG-FOR-BUG. SEE THE SYSCALL_ALLOC_FREE FOR MORE
118 : DETAILS. */
119 :
120 : ulong pc; /* The current instruction, in [0,text_cnt) in normal execution, may be out of bounds in a fault */
121 : ulong ic; /* The number of instructions which have been executed */
122 : ulong cu; /* The remaining CUs left for the transaction, positive in normal execution, may be zero in a fault */
123 : ulong frame_cnt; /* The current number of stack frames pushed, in [0,frame_max] */
124 :
125 : ulong heap_sz; /* Heap size in bytes, in [0,heap_max] */
126 :
127 : /* VM memory */
128 :
129 : /* The vm classifies the 64-bit vm address space into 6 regions:
130 :
131 : 0 - unmapped lo
132 : 1 - program -> [FD_VM_MEM_MAP_PROGRAM_REGION_START,FD_VM_MEM_MAP_PROGRAM_REGION_START+4GiB)
133 : 2 - stack -> [FD_VM_MEM_MAP_STACK_REGION_START, FD_VM_MEM_MAP_STACK_REGION_START +4GiB)
134 : 3 - heap -> [FD_VM_MEM_MAP_HEAP_REGION_START, FD_VM_MEM_MAP_HEAP_REGION_START +4GiB)
135 : 4 - input -> [FD_VM_MEM_MAP_INPUT_REGION_START, FD_VM_MEM_MAP_INPUT_REGION_START +4GiB)
136 : 5 - unmapped hi
137 :
138 : These mappings are encoded in a software TLB consisting of three
139 : 6-element arrays: region_haddr, region_ld_sz and region_st_sz.
140 :
141 : region_haddr[i] gives the location in host address space of the
142 : first byte in region i. region_{ld,st}_sz[i] gives the number of
143 : mappable bytes in this region for {loads,stores}. Note that
144 : region_{ld,st}_sz[i]<2^32. Further note that
145 : [region_haddr[i],region_haddr[i]+region_{ld,st}_sz[i]) does not
146 : wrap around in host address space and does not overlap with any
147 : other usages.
148 :
149 : region_{ld,st}_sz[0] and region_{ld,st}_sz[5] are zero such that
150 : requests to access data from a positive sz range in these regions
151 : will fail, making regions 0 and 5 unreadable and unwritable. As
152 : such, region_haddr[0] and region_haddr[5] are arbitrary; NULL is
153 : used as the obvious default.
154 :
155 : region_st_sz[1] is also zero such that requests to store data to
156 : any positive sz range in this region will fail, making region 1
157 : unwritable.
158 :
159 : When the direct mapping feature is enabled, the input region will
160 : no longer be a contigious buffer of host memory. Instead
161 : it will compose of several fragmented regions of memory each with
162 : its own read/write privleges and size. Address translation to the
163 : input region will now have to rely on a binary search lookup of the
164 : start of the appropriate area of physical memory. It also involves
165 : doing a check against if the region can be written to. */
166 :
167 : /* FIXME: If accessing memory beyond the end of the current heap
168 : region is not allowed, sol_alloc_free will need to update the tlb
169 : arrays during program execution (this is trivial). At the same
170 : time, given sol_alloc_free is deprecated, this is unlikely to be
171 : the case. */
172 :
173 : ulong region_haddr[6];
174 : uint region_ld_sz[6];
175 : uint region_st_sz[6];
176 :
177 : /* fd_vm_input_region_t and fd_vm_acc_to_mem arrays are passed in by the bpf
178 : loaders into fd_vm_init.
179 : TODO: It might make more sense to allocate space for these in the VM. */
180 : fd_vm_input_region_t * input_mem_regions; /* An array of input mem regions represent the input region.
181 : The virtual addresses of each region are contigiuous and
182 : strictly increasing. */
183 : uint input_mem_regions_cnt;
184 : fd_vm_acc_region_meta_t * acc_region_metas; /* Represents a mapping from the instruction account indicies
185 : from the instruction context to the input memory region index
186 : of the account's data region in the input space. */
187 : uchar is_deprecated; /* The vm requires additional checks in certain CPIs if the
188 : vm's current instance was initialized by a deprecated program. */
189 :
190 : ulong reg [ FD_VM_REG_MAX ]; /* registers, indexed [0,FD_VM_REG_CNT). Note that FD_VM_REG_MAX>FD_VM_REG_CNT.
191 : As such, malformed instructions, which can have src/dst reg index in
192 : [0,FD_VM_REG_MAX), cannot access info outside reg. Aligned 8. */
193 : fd_vm_shadow_t shadow[ FD_VM_STACK_FRAME_MAX ]; /* shadow stack, indexed [0,frame_cnt), if frame_cnt>0, 0/frame_cnt-1 is
194 : bottom/top. Aligned 16. */
195 : uchar stack [ FD_VM_STACK_MAX ]; /* stack, indexed [0,FD_VM_STACK_MAX). Divided into FD_VM_STACK_FRAME_MAX
196 : frames. Each frame has a FD_VM_STACK_GUARD_SZ region followed by a
197 : FD_VM_STACK_FRAME_SZ region. reg[10] gives the offset of the start of the
198 : current stack frame. Aligned 16. */
199 : uchar heap [ FD_VM_HEAP_MAX ]; /* syscall heap, [0,heap_sz) used, [heap_sz,heap_max) free. Aligned 8. */
200 :
201 : fd_sha256_t * sha; /* Pre-joined SHA instance. This should be re-initialised before every use. */
202 :
203 : ulong magic; /* ==FD_VM_MAGIC */
204 :
205 : int direct_mapping; /* If direct mapping is enabled or not */
206 : ulong stack_frame_size; /* Size of a stack frame (varies depending on direct mapping being enabled or not) */
207 :
208 : /* Agave reports different error codes (for developers to understand the failure cause) if direct mapping is
209 : enabled AND we halt on a segfault caused by a store on an invalid vaddr. */
210 : ulong segv_store_vaddr;
211 :
212 : ulong sbpf_version; /* SBPF version, SIMD-0161 */
213 : };
214 :
215 : /* FIXME: MOVE ABOVE INTO PRIVATE WHEN CONSTRUCTORS READY */
216 : /**********************************************************************/
217 :
218 : FD_PROTOTYPES_BEGIN
219 :
220 : /* FIXME: FD_VM_T NEEDS PROPER CONSTRUCTORS */
221 :
222 : /* FD_VM_{ALIGN,FOOTPRINT} describe the alignment and footprint needed
223 : for a memory region to hold a fd_vm_t. ALIGN is a positive
224 : integer power of 2. FOOTPRINT is a multiple of align.
225 : These are provided to facilitate compile time declarations. */
226 0 : #define FD_VM_ALIGN FD_VM_HOST_REGION_ALIGN
227 0 : #define FD_VM_FOOTPRINT (527808UL)
228 :
229 : /* fd_vm_{align,footprint} give the needed alignment and footprint
230 : of a memory region suitable to hold an fd_vm_t.
231 : Declaration / aligned_alloc / fd_alloca friendly (e.g. a memory
232 : region declared as "fd_vm_t _vm[1];", or created by
233 : "aligned_alloc(alignof(fd_vm_t),sizeof(fd_vm_t))" or created
234 : by "fd_alloca(alignof(fd_vm_t),sizeof(fd_vm_t))" will all
235 : automatically have the needed alignment and footprint).
236 : fd_vm_{align,footprint} return the same value as
237 : FD_VM_{ALIGN,FOOTPRINT}. */
238 : FD_FN_CONST ulong
239 : fd_vm_align( void );
240 :
241 : FD_FN_CONST ulong
242 : fd_vm_footprint( void );
243 :
244 192 : #define FD_VM_MAGIC (0xF17EDA2CEF0) /* FIREDANCE SBPF V0 */
245 :
246 : /* fd_vm_new formats memory region with suitable alignment and
247 : footprint suitable for holding a fd_vm_t. Assumes
248 : shmem points on the caller to the first byte of the memory region
249 : owned by the caller to use. Returns shmem on success and NULL on
250 : failure (logs details). The memory region will be owned by the state
251 : on successful return. The caller is not joined on return. */
252 :
253 : void *
254 : fd_vm_new( void * shmem );
255 :
256 : /* fd_vm_join joins the caller to a vm.
257 : Assumes shmem points to the first byte of the memory region holding
258 : the vm. Returns a local handle to the join on success (this is
259 : not necessarily a simple cast of the address) and NULL on failure
260 : (logs details). */
261 : fd_vm_t *
262 : fd_vm_join( void * shmem );
263 :
264 : /* fd_vm_init initializes the given fd_vm_t struct, checking that it is
265 : not null and has the correct magic value.
266 :
267 : It modifies the vm object and also returns the object for convenience.
268 :
269 : FIXME: we should split out the memory mapping setup from this function
270 : to handle those errors separately. */
271 : fd_vm_t *
272 : fd_vm_init(
273 : fd_vm_t * vm,
274 : fd_exec_instr_ctx_t *instr_ctx,
275 : ulong heap_max,
276 : ulong entry_cu,
277 : uchar const * rodata,
278 : ulong rodata_sz,
279 : ulong const * text,
280 : ulong text_cnt,
281 : ulong text_off,
282 : ulong text_sz,
283 : ulong entry_pc,
284 : ulong * calldests,
285 : ulong sbpf_version,
286 : fd_sbpf_syscalls_t * syscalls,
287 : fd_vm_trace_t * trace,
288 : fd_sha256_t * sha,
289 : fd_vm_input_region_t * mem_regions,
290 : uint mem_regions_cnt,
291 : fd_vm_acc_region_meta_t * acc_region_metas,
292 : uchar is_deprecated,
293 : int direct_mapping );
294 :
295 : /* fd_vm_leave leaves the caller's current local join to a vm.
296 : Returns a pointer to the memory region holding the vm on success
297 : (this is not necessarily a simple cast of the
298 : address) and NULL on failure (logs details). The caller is not
299 : joined on successful return. */
300 : void *
301 : fd_vm_leave( fd_vm_t * vm );
302 :
303 : /* fd_vm_delete unformats a memory region that holds a vm.
304 : Assumes shmem points on the caller to the first
305 : byte of the memory region holding the state and that nobody is
306 : joined. Returns a pointer to the memory region on success and NULL
307 : on failure (logs details). The caller has ownership of the memory
308 : region on successful return. */
309 : void *
310 : fd_vm_delete( void * shmem );
311 :
312 : /* fd_vm_validate validates the sBPF program in the given vm. Returns
313 : success or an error code. Called before executing a sBPF program.
314 : FIXME: DOCUMENT BETTER */
315 :
316 : FD_FN_PURE int
317 : fd_vm_validate( fd_vm_t const * vm );
318 :
319 : /* fd_vm_is_check_align_enabled returns 1 if the vm should check alignment
320 : when doing memory translation. */
321 : FD_FN_PURE static inline int
322 162 : fd_vm_is_check_align_enabled( fd_vm_t const * vm ) {
323 162 : return !vm->is_deprecated;
324 162 : }
325 :
326 : /* fd_vm_is_check_size_enabled returns 1 if the vm should check size
327 : when doing memory translation. */
328 : FD_FN_PURE static inline int
329 0 : fd_vm_is_check_size_enabled( fd_vm_t const * vm ) {
330 0 : return !vm->is_deprecated;
331 0 : }
332 :
333 : /* FIXME: make this trace-aware, and move into fd_vm_init
334 : This is a temporary hack to make the fuzz harness work. */
335 : int
336 : fd_vm_setup_state_for_execution( fd_vm_t * vm ) ;
337 :
338 : /* fd_vm_exec runs vm from program start to program halt or program
339 : fault, appending an execution trace if vm is attached to a trace.
340 :
341 : Since this is running from program start, this will init r1 and r10,
342 : pop all stack frames and free all heap allocations.
343 :
344 : IMPORTANT SAFETY TIP! This currently does not zero out any other
345 : registers, the user stack region or the user heap. (FIXME: SHOULD
346 : IT??)
347 :
348 : Returns FD_VM_SUCCESS (0) on success and an FD_VM_ERR code (negative)
349 : on failure. Reasons for failure include:
350 :
351 : INVAL - NULL vm (or, for fd_vm_exec_trace, the vm is not
352 : attached to trace). FIXME: ADD OTHER INPUT ARG CHECKS?
353 :
354 : SIGTEXT - A jump/call set the program counter outside the text
355 : region or the program counter incremented beyond the
356 : text region. pc will be at the out of bounds location.
357 : ic and cu will not include the out of bounds location.
358 : For a call, the call stack frame was allocated.
359 :
360 : SIGSPLIT - A jump/call set the program counter into the middle of
361 : a multiword instruction or a multiword instruction went
362 : past the text region end. pc will be at the split. ic
363 : and cu will not include the split. For a call, the
364 : call stack frame was allocated.
365 :
366 : SIGCALL - A call set the program counter to a non-function
367 : location. pc will be at the non-function location. ic
368 : and cu will include the call but not include the
369 : non-function location. The call stack frame was
370 : allocated.
371 :
372 : SIGSTACK - The call depth limit was exceeded. pc will be at the
373 : call. ic and cu will include the call but not the call
374 : target. The call stack frame was not allocated.
375 :
376 : SIGILL - An invalid instruction was encountered (including an
377 : invalid opcode and an endian swap with an invalid bit
378 : width). pc will be at the invalid instruction. ic and
379 : cu will not include the invalid instruction.
380 :
381 : SIGSEGV - An invalid memory access (outside the program memory
382 : map) was encountered. pc will be at the faulting
383 : instruction. ic and cu will not include the faulting
384 : instruction.
385 :
386 : SIGBUS - An unaligned memory access was encountered. pc will be
387 : at the faulting instruction. ic and cu will not
388 : include the faulting instruction. (Note: currently
389 : mapped to SIGSEGV and then only if check_align is
390 : enabled.)
391 :
392 : SIGRDONLY - A write to read-only memory address was encountered.
393 : pc will be at the faulting instruction. ic and cu will
394 : not include the faulting instruction. (Note: currently
395 : mapped to SIGSEGV.)
396 :
397 : SIGCOST - The compute limit was exceeded. pc will be at the
398 : first non-executed instruction (if pc is a syscall, the
399 : syscall might have been partially executed when it ran
400 : out of budget .. see safety tip below). ic will cover
401 : all executed instructions. cu will be zero.
402 :
403 : This will considers any error returned by a syscall as a fault and
404 : returns the syscall error code here. See syscall documentation for
405 : details here. When a syscall faults, pc will be at the syscall, ic
406 : will include the syscall and cu will include the syscall and any
407 : additional costs the syscall might have incurred up to that point of
408 : the fault.
409 :
410 : IMPORTANT SAFETY TIP! Ideally, a syscall should only modify vm's
411 : state when it knows its overall syscall will be successful.
412 : Unfortunately, this is often not practical (e.g. a syscall starts
413 : processing a list of user provided commands and discovers an error
414 : condition late in the command list that did not exist at syscall
415 : start because the error condition was created by successfully
416 : executed commands earlier in the list). As such, vm's state on a
417 : faulting syscall may not be clean.
418 :
419 : FIXME: SINCE MOST SYSCALLS CAN BE IMPLEMENTED TO HAVE CLEAN FAULTING
420 : BEHAVIOR, PROVIDE A MECHANISM SO USERS CAN EASILY DETECT UNCLEAN
421 : SYSCALL FAULTS?
422 :
423 : For SIGCOST, note that the vm can speculate ahead when processing
424 : instructions. This makes it is possible to have a situation where
425 : a vm faults with, for example, SIGSEGV from a speculatively
426 : executed memory access while a non-speculative execution would have
427 : faulted with SIGCOST on an earlier instruction. In these situations,
428 : pc will be at the faulting speculatively executed instruction, ic
429 : will include all the speculatively executed instructions, cu will be
430 : zero and vm's state will include the impact of all the speculation.
431 :
432 : IMPORTANT SAFETY TIP! While different vm implementations can
433 : disagree on why a program faulted (e.g. SIGCOST versus SIGSEGV in the
434 : example above), they cannot disagree on whether or not a program
435 : faulted. As a result, the specific fault reason must never be
436 : allowed to be part of consensus.
437 :
438 : fd_vm_exec_trace runs with tracing and requires vm to be attached to
439 : a trace. fd_vm_exec_notrace runs without without tracing even if vm
440 : is attached to a trace. */
441 :
442 : int
443 : fd_vm_exec_trace( fd_vm_t * vm );
444 :
445 : int
446 : fd_vm_exec_notrace( fd_vm_t * vm );
447 :
448 : static inline int
449 186 : fd_vm_exec( fd_vm_t * vm ) {
450 186 : if( FD_UNLIKELY( vm->trace ) ) return fd_vm_exec_trace ( vm );
451 186 : else return fd_vm_exec_notrace( vm );
452 186 : }
453 :
454 : FD_PROTOTYPES_END
455 :
456 : #endif /* HEADER_fd_src_flamenco_vm_fd_vm_h */
|