mrb_vm_define_class() may realloc() mrb->c->stack because it calls
mrb_funcall() for inherited hook. If mrb->c->stack is realloc()-ed, regs
refers orphan address.
Think about the following Ruby script:
segv.rb:
begin
lambda do
lambda do
"x" * 1000 # NoMemoryError
end.call
end.call
rescue
raise
end
If memory can't allocate after `"x" * 1000`, mruby crashes.
Because L_RAISE: block in mrb_vm_exec() calls mrb_env_unshare() via
cipop() and mrb_env_unshare() uses allocated memory without NULL check:
L_RAISE: block:
L_RAISE:
// ...
while (ci[0].ridx == ci[-1].ridx) {
cipop(mrb);
// ...
}
cipop():
static void
cipop(mrb_state *mrb)
{
struct mrb_context *c = mrb->c;
if (c->ci->env) {
mrb_env_unshare(mrb, c->ci->env);
}
c->ci--;
}
mrb_env_unshare():
MRB_API void
mrb_env_unshare(mrb_state *mrb, struct REnv *e)
{
size_t len = (size_t)MRB_ENV_STACK_LEN(e);
// p is NULL in this case
mrb_value *p = (mrb_value *)mrb_malloc(mrb, sizeof(mrb_value)*len);
MRB_ENV_UNSHARE_STACK(e);
if (len > 0) {
stack_copy(p, e->stack, len); // p is NULL but used. It causes SEGV.
}
e->stack = p;
mrb_write_barrier(mrb, (struct RBasic *)e);
}
To solve the SEGV, this change always raises NoMemoryError even when
realloc() is failed after the first NoMemoryError in
mrb_realloc(). mrb_unv_unshare() doesn't need to check NULL with this
change.
But it causes infinite loop in the following while:
L_RAISE:
// ...
while (ci[0].ridx == ci[-1].ridx) {
cipop(mrb);
// ...
}
Because cipop() never pops ci.
This change includes cipop() change. The change pops ci even when
mrb_unv_unshare() is failed by NoMemoryError.
This case can be reproduced by the following program:
#include <stdlib.h>
#include <mruby.h>
#include <mruby/compile.h>
static void *
allocf(mrb_state *mrb, void *ptr, size_t size, void *ud)
{
static mrb_bool always_fail = FALSE;
if (size == 1001) {
always_fail = TRUE;
}
if (always_fail) {
return NULL;
}
if (size == 0) {
free(ptr);
return NULL;
} else {
return realloc(ptr, size);
}
}
int
main(int argc, char **argv)
{
mrb_state *mrb;
mrbc_context *c;
FILE *file;
mrb = mrb_open_allocf(allocf, NULL);
c = mrbc_context_new(mrb);
file = fopen(argv[1], "r");
mrb_load_file_cxt(mrb, file, c);
fclose(file);
mrbc_context_free(mrb, c);
mrb_close(mrb);
return EXIT_SUCCESS;
}
Try the following command lines:
% cc -I include -L build/host/lib -O0 -g3 -o no-memory no-memory.c -lmruby -lm
% ./no-memory segv.rb
GitHub: fix#2902, #2917
The current implementation traverses stack to retrieve backtrace. But
stack will be changed when some operations are occurred. It means that
backtrace may be broken after some operations.
This change (1) saves the minimum information to retrieve backtrace when
exception is raised and (2) restores backtrace from the minimum
information when backtrace is needed. It reduces overhead for creating
backtrace Ruby objects.
The space for the minimum information is reused by multiple
exceptions. So memory allocation isn't occurred for each exception.
It isn't needed as it's very similar to mrb_define_method_raw() and also
there's only one place where mrb_proc_ptr() actually has to be called.
Inspired by @furunkel's method cache patch (#2764).
The following program reproduces this problem:
#include <mruby.h>
static mrb_value
recursive(mrb_state *mrb, mrb_value self)
{
mrb_int n;
mrb_get_args(mrb, "i", &n);
if (n == 0) {
mrb_raise(mrb, E_RUNTIME_ERROR, "XXX");
} else {
mrb_funcall(mrb, self, "recursive", 1, mrb_fixnum_value(n - 1));
}
return self;
}
int
main(void)
{
mrb_state *mrb;
mrb = mrb_open();
mrb_define_method(mrb, mrb->kernel_module, "recursive", recursive,
MRB_ARGS_REQ(1));
mrb_funcall(mrb, mrb_top_self(mrb), "recursive", 1, mrb_fixnum_value(30));
mrb_close(mrb);
}
Recursive method call isn't required. It's just for expanding call info
stack.
If mrb_realloc() is called in cipush(), cibase address is changed. So,
we shouldn't compare ci before mrb_realloc() and cibase after
mrb_realloc(). It accesses unknown address and causes crash.
Yukihiro "Matz" Matsumoto