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 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 : uint is_writable; /* If the region can be written to or is read-only */
29 : };
30 : typedef struct fd_vm_input_region fd_vm_input_region_t;
31 :
32 : /* fd_vm_acc_region_meta_t holds metadata about a given account. An array of these
33 : structs will map an instruction account index to its respective input memory
34 : region location. */
35 :
36 : struct __attribute((aligned(8UL))) fd_vm_acc_region_meta {
37 : uint region_idx;
38 : uchar has_data_region;
39 : uchar has_resizing_region;
40 : };
41 : typedef struct fd_vm_acc_region_meta fd_vm_acc_region_meta_t;
42 :
43 : /* In Agave, all the regions are 16-byte aligned in host address space. There is then an alignment check
44 : which is done inside each syscall memory translation, checking if the data is aligned in host address
45 : space. This is a layering violation, as it leaks the host address layout into the consensus model.
46 :
47 : In the future we will change this alignment check in the vm to purely operate on the virtual address space,
48 : taking advantage of the fact that Agave regions are known to be aligned. For now, we align our regions to
49 : either 8 or 16 bytes, as there are no 16-byte alignment translations in the syscalls currently:
50 : stack: 16 byte aligned
51 : heap: 16 byte aligned
52 : input: 8 byte aligned
53 : rodata: 8 byte aligned
54 :
55 : https://github.com/solana-labs/rbpf/blob/cd19a25c17ec474e6fa01a3cc3efa325f44cd111/src/ebpf.rs#L39-L40 */
56 12075 : #define FD_VM_HOST_REGION_ALIGN (16UL)
57 :
58 : struct __attribute__((aligned(FD_VM_HOST_REGION_ALIGN))) fd_vm {
59 :
60 : /* VM configuration */
61 :
62 : /* FIXME: suspect these three should be replaced by some kind of VM
63 : enabled feature struct (though syscalls do seem to make additional
64 : non-trivial use of instr_ctx). */
65 :
66 : fd_exec_instr_ctx_t * instr_ctx; /* FIXME: DOCUMENT */
67 :
68 : /* FIXME: frame_max should be run time configurable by compute budget.
69 : If there is no reasonable upper bound on this, shadow and stack
70 : will need to be provided by users. */
71 :
72 : //ulong frame_max; /* Maximum number of stack frames, in [0,FD_VM_STACK_FRAME_MAX] */
73 : ulong heap_max; /* Maximum amount of heap in bytes, in [0,FD_VM_HEAP_MAX] */
74 : ulong entry_cu; /* Initial number of compute units for this program, in [0,FD_VM_COMPUTE_UNIT_LIMIT] */
75 :
76 : /* FIXME: The below are practically an exact match to the
77 : fields of an fd_sbpf_program_t (sans ELF info) */
78 :
79 : uchar const * rodata; /* Program read only data, indexed [0,rodata_sz), aligned 8 */
80 : ulong rodata_sz; /* Program read only data size in bytes, FIXME: BOUNDS? */
81 : ulong const * text; /* Program sBPF words, indexed [0,text_cnt), aligned 8 */
82 : ulong text_cnt; /* Program sBPF word count, all text words are inside the rodata */
83 : ulong text_off; /* ==(ulong)text - (ulong)rodata, relocation offset in bytes we must apply to indirect calls
84 : (callx/CALL_REGs), IMPORTANT SAFETY TIP! THIS IS IN BYTES, NOT WORDS! */
85 : ulong text_sz; /* Program sBPF size in bytes, == text_cnt*8 */
86 :
87 : ulong entry_pc; /* Initial program counter, in [0,text_cnt)
88 : FIXME: MAKE SURE NOT INTO MW INSTRUCTION, MAKE SURE VALID CALLDEST? */
89 : ulong const * calldests; /* Bit vector of local functions that can be called into, bit indexed in [0,text_cnt) */
90 : /* FIXME: ADD BIT VECTOR OF FORBIDDEN BRANCH TARGETS (E.G.
91 : INTO THE MIDDLE OF A MULTIWORD INSTRUCTION) */
92 :
93 : fd_sbpf_syscalls_t const * syscalls; /* The map of syscalls (sharable over multiple concurrently running vm) */
94 :
95 : fd_vm_trace_t * trace; /* Location to stream traces (no tracing if NULL) */
96 :
97 : /* VM execution and syscall state */
98 :
99 : /* These are used to communicate the execution and syscall state to
100 : users and syscalls. These are initialized based on the above when
101 : a program starts executing. When program halts or faults, these
102 : provide precise execution diagnostics to the user (and potential
103 : breakpoint/continue functionality in the future). When the vm
104 : makes a syscall, the vm will set these precisely and, when a
105 : syscall returns, the vm will update its internal execution state
106 : appropriately. */
107 :
108 : /* IMPORTANT SAFETY TIP! THE BEHAVIOR OF THE SYSCALL ALLOCATOR FOR
109 : HEAP_SZ MUST EXACTLY MATCH THE SOLANA VALIDATOR ALLOCATOR:
110 :
111 : https://github.com/solana-labs/solana/blob/v1.17.23/program-runtime/src/invoke_context.rs#L122-L148
112 :
113 : BIT-FOR-BIT AND BUG-FOR-BUG. SEE THE SYSCALL_ALLOC_FREE FOR MORE
114 : DETAILS. */
115 :
116 : ulong pc; /* The current instruction, in [0,text_cnt) in normal execution, may be out of bounds in a fault */
117 : ulong ic; /* The number of instructions which have been executed */
118 : ulong cu; /* The remaining CUs left for the transaction, positive in normal execution, may be zero in a fault */
119 : ulong frame_cnt; /* The current number of stack frames pushed, in [0,frame_max] */
120 :
121 : ulong heap_sz; /* Heap size in bytes, in [0,heap_max] */
122 :
123 : /* VM memory */
124 :
125 : /* The vm classifies the 64-bit vm address space into 6 regions:
126 :
127 : 0 - unmapped lo
128 : 1 - program -> [FD_VM_MEM_MAP_PROGRAM_REGION_START,FD_VM_MEM_MAP_PROGRAM_REGION_START+4GiB)
129 : 2 - stack -> [FD_VM_MEM_MAP_STACK_REGION_START, FD_VM_MEM_MAP_STACK_REGION_START +4GiB)
130 : 3 - heap -> [FD_VM_MEM_MAP_HEAP_REGION_START, FD_VM_MEM_MAP_HEAP_REGION_START +4GiB)
131 : 4 - input -> [FD_VM_MEM_MAP_INPUT_REGION_START, FD_VM_MEM_MAP_INPUT_REGION_START +4GiB)
132 : 5 - unmapped hi
133 :
134 : These mappings are encoded in a software TLB consisting of three
135 : 6-element arrays: region_haddr, region_ld_sz and region_st_sz.
136 :
137 : region_haddr[i] gives the location in host address space of the
138 : first byte in region i. region_{ld,st}_sz[i] gives the number of
139 : mappable bytes in this region for {loads,stores}. Note that
140 : region_{ld,st}_sz[i]<2^32. Further note that
141 : [region_haddr[i],region_haddr[i]+region_{ld,st}_sz[i]) does not
142 : wrap around in host address space and does not overlap with any
143 : other usages.
144 :
145 : region_{ld,st}_sz[0] and region_{ld,st}_sz[5] are zero such that
146 : requests to access data from a positive sz range in these regions
147 : will fail, making regions 0 and 5 unreadable and unwritable. As
148 : such, region_haddr[0] and region_haddr[5] are arbitrary; NULL is
149 : used as the obvious default.
150 :
151 : region_st_sz[1] is also zero such that requests to store data to
152 : any positive sz range in this region will fail, making region 1
153 : unwritable.
154 :
155 : When the direct mapping feature is enabled, the input region will
156 : no longer be a contigious buffer of host memory. Instead
157 : it will compose of several fragmented regions of memory each with
158 : its own read/write privleges and size. Address translation to the
159 : input region will now have to rely on a binary search lookup of the
160 : start of the appropriate area of physical memory. It also involves
161 : doing a check against if the region can be written to. */
162 :
163 : /* FIXME: If accessing memory beyond the end of the current heap
164 : region is not allowed, sol_alloc_free will need to update the tlb
165 : arrays during program execution (this is trivial). At the same
166 : time, given sol_alloc_free is deprecated, this is unlikely to be
167 : the case. */
168 :
169 : ulong region_haddr[6];
170 : uint region_ld_sz[6];
171 : uint region_st_sz[6];
172 :
173 : /* fd_vm_input_region_t and fd_vm_acc_to_mem arrays are passed in by the bpf
174 : loaders into fd_vm_init.
175 : TODO: It might make more sense to allocate space for these in the VM. */
176 : fd_vm_input_region_t * input_mem_regions; /* An array of input mem regions represent the input region.
177 : The virtual addresses of each region are contigiuous and
178 : strictly increasing. */
179 : uint input_mem_regions_cnt;
180 : fd_vm_acc_region_meta_t * acc_region_metas; /* Represents a mapping from the instruction account indicies
181 : from the instruction context to the input memory region index
182 : of the account's data region in the input space. */
183 : uchar is_deprecated; /* The vm requires additional checks in certain CPIs if the
184 : vm's current instance was initialized by a deprecated program. */
185 :
186 : ulong reg [ FD_VM_REG_MAX ]; /* registers, indexed [0,FD_VM_REG_CNT). Note that FD_VM_REG_MAX>FD_VM_REG_CNT.
187 : As such, malformed instructions, which can have src/dst reg index in
188 : [0,FD_VM_REG_MAX), cannot access info outside reg. Aligned 8. */
189 : 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
190 : bottom/top. Aligned 16. */
191 : uchar stack [ FD_VM_STACK_MAX ]; /* stack, indexed [0,FD_VM_STACK_MAX). Divided into FD_VM_STACK_FRAME_MAX
192 : frames. Each frame has a FD_VM_STACK_GUARD_SZ region followed by a
193 : FD_VM_STACK_FRAME_SZ region. reg[10] gives the offset of the start of the
194 : current stack frame. Aligned 16. */
195 : uchar heap [ FD_VM_HEAP_MAX ]; /* syscall heap, [0,heap_sz) used, [heap_sz,heap_max) free. Aligned 8. */
196 :
197 : fd_sha256_t * sha; /* Pre-joined SHA instance. This should be re-initialised before every use. */
198 :
199 : ulong magic; /* ==FD_VM_MAGIC */
200 :
201 : int direct_mapping; /* If direct mapping is enabled or not */
202 : ulong stack_frame_size; /* Size of a stack frame (varies depending on direct mapping being enabled or not) */
203 :
204 : /* Agave reports different error codes (for developers to understand the failure cause) if direct mapping is
205 : enabled AND we halt on a segfault caused by a store on an invalid vaddr. */
206 : ulong segv_store_vaddr;
207 : };
208 :
209 : /* FIXME: MOVE ABOVE INTO PRIVATE WHEN CONSTRUCTORS READY */
210 : /**********************************************************************/
211 :
212 : FD_PROTOTYPES_BEGIN
213 :
214 : /* FIXME: FD_VM_T NEEDS PROPER CONSTRUCTORS */
215 :
216 : /* FD_VM_{ALIGN,FOOTPRINT} describe the alignment and footprint needed
217 : for a memory region to hold a fd_vm_t. ALIGN is a positive
218 : integer power of 2. FOOTPRINT is a multiple of align.
219 : These are provided to facilitate compile time declarations. */
220 12075 : #define FD_VM_ALIGN FD_VM_HOST_REGION_ALIGN
221 12075 : #define FD_VM_FOOTPRINT (527296UL)
222 :
223 : /* fd_vm_{align,footprint} give the needed alignment and footprint
224 : of a memory region suitable to hold an fd_vm_t.
225 : Declaration / aligned_alloc / fd_alloca friendly (e.g. a memory
226 : region declared as "fd_vm_t _vm[1];", or created by
227 : "aligned_alloc(alignof(fd_vm_t),sizeof(fd_vm_t))" or created
228 : by "fd_alloca(alignof(fd_vm_t),sizeof(fd_vm_t))" will all
229 : automatically have the needed alignment and footprint).
230 : fd_vm_{align,footprint} return the same value as
231 : FD_VM_{ALIGN,FOOTPRINT}. */
232 : FD_FN_CONST ulong
233 : fd_vm_align( void );
234 :
235 : FD_FN_CONST ulong
236 : fd_vm_footprint( void );
237 :
238 15348 : #define FD_VM_MAGIC (0xF17EDA2CEF0) /* FIREDANCE SBPF V0 */
239 :
240 : /* fd_vm_new formats memory region with suitable alignment and
241 : footprint suitable for holding a fd_vm_t. Assumes
242 : shmem points on the caller to the first byte of the memory region
243 : owned by the caller to use. Returns shmem on success and NULL on
244 : failure (logs details). The memory region will be owned by the state
245 : on successful return. The caller is not joined on return. */
246 :
247 : void *
248 : fd_vm_new( void * shmem );
249 :
250 : /* fd_vm_join joins the caller to a vm.
251 : Assumes shmem points to the first byte of the memory region holding
252 : the vm. Returns a local handle to the join on success (this is
253 : not necessarily a simple cast of the address) and NULL on failure
254 : (logs details). */
255 : fd_vm_t *
256 : fd_vm_join( void * shmem );
257 :
258 : /* fd_vm_init initializes the given fd_vm_t struct, checking that it is
259 : not null and has the correct magic value.
260 :
261 : It modifies the vm object and also returns the object for convenience.
262 :
263 : FIXME: we should split out the memory mapping setup from this function
264 : to handle those errors separately. */
265 : fd_vm_t *
266 : fd_vm_init(
267 : fd_vm_t * vm,
268 : fd_exec_instr_ctx_t *instr_ctx,
269 : ulong heap_max,
270 : ulong entry_cu,
271 : uchar const * rodata,
272 : ulong rodata_sz,
273 : ulong const * text,
274 : ulong text_cnt,
275 : ulong text_off,
276 : ulong text_sz,
277 : ulong entry_pc,
278 : ulong * calldests,
279 : fd_sbpf_syscalls_t * syscalls,
280 : fd_vm_trace_t * trace,
281 : fd_sha256_t * sha,
282 : fd_vm_input_region_t * mem_regions,
283 : uint mem_regions_cnt,
284 : fd_vm_acc_region_meta_t * acc_region_metas,
285 : uchar is_deprecated,
286 : int direct_mapping );
287 :
288 : /* fd_vm_leave leaves the caller's current local join to a vm.
289 : Returns a pointer to the memory region holding the vm on success
290 : (this is not necessarily a simple cast of the
291 : address) and NULL on failure (logs details). The caller is not
292 : joined on successful return. */
293 : void *
294 : fd_vm_leave( fd_vm_t * vm );
295 :
296 : /* fd_vm_delete unformats a memory region that holds a vm.
297 : Assumes shmem points on the caller to the first
298 : byte of the memory region holding the state and that nobody is
299 : joined. Returns a pointer to the memory region on success and NULL
300 : on failure (logs details). The caller has ownership of the memory
301 : region on successful return. */
302 : void *
303 : fd_vm_delete( void * shmem );
304 :
305 : /* fd_vm_validate validates the sBPF program in the given vm. Returns
306 : success or an error code. Called before executing a sBPF program.
307 : FIXME: DOCUMENT BETTER */
308 :
309 : FD_FN_PURE int
310 : fd_vm_validate( fd_vm_t const * vm );
311 :
312 : /* fd_vm_is_check_align_enabled returns 1 if the vm should check alignment
313 : when doing memory translation. */
314 : FD_FN_PURE static inline int
315 67035 : fd_vm_is_check_align_enabled( fd_vm_t const * vm ) {
316 67035 : return !vm->is_deprecated;
317 67035 : }
318 :
319 : /* fd_vm_is_check_size_enabled returns 1 if the vm should check size
320 : when doing memory translation. */
321 : FD_FN_PURE static inline int
322 0 : fd_vm_is_check_size_enabled( fd_vm_t const * vm ) {
323 0 : return !vm->is_deprecated;
324 0 : }
325 :
326 : /* FIXME: make this trace-aware, and move into fd_vm_init
327 : This is a temporary hack to make the fuzz harness work. */
328 : int
329 : fd_vm_setup_state_for_execution( fd_vm_t * vm ) ;
330 :
331 : /* fd_vm_exec runs vm from program start to program halt or program
332 : fault, appending an execution trace if vm is attached to a trace.
333 :
334 : Since this is running from program start, this will init r1 and r10,
335 : pop all stack frames and free all heap allocations.
336 :
337 : IMPORTANT SAFETY TIP! This currently does not zero out any other
338 : registers, the user stack region or the user heap. (FIXME: SHOULD
339 : IT??)
340 :
341 : Returns FD_VM_SUCCESS (0) on success and an FD_VM_ERR code (negative)
342 : on failure. Reasons for failure include:
343 :
344 : INVAL - NULL vm (or, for fd_vm_exec_trace, the vm is not
345 : attached to trace). FIXME: ADD OTHER INPUT ARG CHECKS?
346 :
347 : SIGTEXT - A jump/call set the program counter outside the text
348 : region or the program counter incremented beyond the
349 : text region. pc will be at the out of bounds location.
350 : ic and cu will not include the out of bounds location.
351 : For a call, the call stack frame was allocated.
352 :
353 : SIGSPLIT - A jump/call set the program counter into the middle of
354 : a multiword instruction or a multiword instruction went
355 : past the text region end. pc will be at the split. ic
356 : and cu will not include the split. For a call, the
357 : call stack frame was allocated.
358 :
359 : SIGCALL - A call set the program counter to a non-function
360 : location. pc will be at the non-function location. ic
361 : and cu will include the call but not include the
362 : non-function location. The call stack frame was
363 : allocated.
364 :
365 : SIGSTACK - The call depth limit was exceeded. pc will be at the
366 : call. ic and cu will include the call but not the call
367 : target. The call stack frame was not allocated.
368 :
369 : SIGILL - An invalid instruction was encountered (including an
370 : invalid opcode and an endian swap with an invalid bit
371 : width). pc will be at the invalid instruction. ic and
372 : cu will not include the invalid instruction.
373 :
374 : SIGSEGV - An invalid memory access (outside the program memory
375 : map) was encountered. pc will be at the faulting
376 : instruction. ic and cu will not include the faulting
377 : instruction.
378 :
379 : SIGBUS - An unaligned memory access was encountered. pc will be
380 : at the faulting instruction. ic and cu will not
381 : include the faulting instruction. (Note: currently
382 : mapped to SIGSEGV and then only if check_align is
383 : enabled.)
384 :
385 : SIGRDONLY - A write to read-only memory address was encountered.
386 : pc will be at the faulting instruction. ic and cu will
387 : not include the faulting instruction. (Note: currently
388 : mapped to SIGSEGV.)
389 :
390 : SIGCOST - The compute limit was exceeded. pc will be at the
391 : first non-executed instruction (if pc is a syscall, the
392 : syscall might have been partially executed when it ran
393 : out of budget .. see safety tip below). ic will cover
394 : all executed instructions. cu will be zero.
395 :
396 : This will considers any error returned by a syscall as a fault and
397 : returns the syscall error code here. See syscall documentation for
398 : details here. When a syscall faults, pc will be at the syscall, ic
399 : will include the syscall and cu will include the syscall and any
400 : additional costs the syscall might have incurred up to that point of
401 : the fault.
402 :
403 : IMPORTANT SAFETY TIP! Ideally, a syscall should only modify vm's
404 : state when it knows its overall syscall will be successful.
405 : Unfortunately, this is often not practical (e.g. a syscall starts
406 : processing a list of user provided commands and discovers an error
407 : condition late in the command list that did not exist at syscall
408 : start because the error condition was created by successfully
409 : executed commands earlier in the list). As such, vm's state on a
410 : faulting syscall may not be clean.
411 :
412 : FIXME: SINCE MOST SYSCALLS CAN BE IMPLEMENTED TO HAVE CLEAN FAULTING
413 : BEHAVIOR, PROVIDE A MECHANISM SO USERS CAN EASILY DETECT UNCLEAN
414 : SYSCALL FAULTS?
415 :
416 : For SIGCOST, note that the vm can speculate ahead when processing
417 : instructions. This makes it is possible to have a situation where
418 : a vm faults with, for example, SIGSEGV from a speculatively
419 : executed memory access while a non-speculative execution would have
420 : faulted with SIGCOST on an earlier instruction. In these situations,
421 : pc will be at the faulting speculatively executed instruction, ic
422 : will include all the speculatively executed instructions, cu will be
423 : zero and vm's state will include the impact of all the speculation.
424 :
425 : IMPORTANT SAFETY TIP! While different vm implementations can
426 : disagree on why a program faulted (e.g. SIGCOST versus SIGSEGV in the
427 : example above), they cannot disagree on whether or not a program
428 : faulted. As a result, the specific fault reason must never be
429 : allowed to be part of consensus.
430 :
431 : fd_vm_exec_trace runs with tracing and requires vm to be attached to
432 : a trace. fd_vm_exec_notrace runs without without tracing even if vm
433 : is attached to a trace. */
434 :
435 : int
436 : fd_vm_exec_trace( fd_vm_t * vm );
437 :
438 : int
439 : fd_vm_exec_notrace( fd_vm_t * vm );
440 :
441 : static inline int
442 1203 : fd_vm_exec( fd_vm_t * vm ) {
443 1203 : if( FD_UNLIKELY( vm->trace ) ) return fd_vm_exec_trace ( vm );
444 1203 : else return fd_vm_exec_notrace( vm );
445 1203 : }
446 :
447 : FD_PROTOTYPES_END
448 :
449 : #endif /* HEADER_fd_src_flamenco_vm_fd_vm_h */
|
