/*
* Forward write_c() on a cheri_fd to the underlying file descriptor.
*/
struct cheri_fd_ret
cheri_fd_write(__capability const void *buf_c, size_t nbytes)
{
struct cheri_fd_ret ret;
__capability struct cheri_fd *cfp;
void *buf;
/* XXXRW: Object-capability user permission check on idc. */
/* XXXRW: Change to check permissions directly and throw exception. */
if (!(cheri_getperm(buf_c) & CHERI_PERM_LOAD)) {
ret.cfr_retval0 = -1;
ret.cfr_retval1 = EPROT;
return (ret);
}
buf = (void *)buf_c;
/* Check that cheri_fd hasn't been revoked. */
cfp = cheri_getidc();
if (cfp->cf_fd == -1) {
ret.cfr_retval0 = -1;
ret.cfr_retval1 = EBADF;
return (ret);
}
/* Forward to operating system. */
ret.cfr_retval0 = write(cfp->cf_fd, buf,
min(nbytes, cheri_getlen(buf_c) - cheri_getoffset(buf_c)));
ret.cfr_retval1 = (ret.cfr_retval0 < 0 ? errno : 0);
return (ret);
}
/*
* Forward fstat() on a cheri_fd to the underlying file descriptor.
*/
struct cheri_fd_ret
cheri_fd_fstat(__capability struct stat *sb_c)
{
struct cheri_fd_ret ret;
__capability struct cheri_fd *cfp;
struct stat *sb;
/* XXXRW: Object-capability user permission check on idc. */
/* XXXRW: Change to check permissions directly and throw exception. */
if (!(cheri_getperm(sb_c) & CHERI_PERM_STORE) ||
!(cheri_getlen(sb_c) >= sizeof(*sb))) {
ret.cfr_retval0 = -1;
ret.cfr_retval1 = EPROT;
return (ret);
}
sb = (void *)sb_c;
/* Check that the cheri_fd hasn't been revoked. */
cfp = cheri_getidc();
if (cfp->cf_fd == -1) {
ret.cfr_retval0 = -1;
ret.cfr_retval1 = EBADF;
return (ret);
}
/* Forward to operating system. */
ret.cfr_retval0 = fstat(cfp->cf_fd, sb);
ret.cfr_retval1 = (ret.cfr_retval0 < 0 ? errno : 0);
return (ret);
}
void *
kernel_realloc(void *cp, size_t nbytes)
{
size_t cur_space; /* Space in the current bucket */
size_t smaller_space; /* Space in the next smaller bucket */
union overhead *op;
char *res;
if (cp == NULL)
return (kernel_malloc(nbytes));
op = find_overhead(cp);
if (op == NULL)
return (NULL);
cur_space = (1 << (op->ov_index + 3)) - sizeof(*op);
/* avoid the copy if same size block */
/*
* XXX-BD: Arguably we should be tracking the actual allocation
* not just the bucket size so that we can do a full malloc+memcpy
* when the caller has restricted the length of the pointer passed
* realloc() but is growing the buffer within the current bucket.
*
* As it is, this code contains a leak where realloc recovers access
* to the contents in foo:
* char *foo = malloc(10);
* strcpy(foo, "abcdefghi");
* cheri_setbouds(foo, 5);
* foo = realloc(foo, 10);
*/
smaller_space = (1 << (op->ov_index + 2)) - sizeof(*op);
if (nbytes <= cur_space && nbytes > smaller_space)
return (cheri_andperm(cheri_setbounds(op + 1, nbytes),
cheri_getperm(cp)));
if ((res = kernel_malloc(nbytes)) == NULL)
return (NULL);
/*
* Only copy data the caller had access to even if this is less
* than the size of the original allocation. This risks surprise
* for some programmers, but to do otherwise risks information leaks.
*/
memcpy(res, cp, (nbytes <= cheri_getlen(cp)) ? nbytes : cheri_getlen(cp));
res = cheri_andperm(res, cheri_getperm(cp));
kernel_free(cp);
return (res);
}
register_t
cheritest_libcheri_userfn_getstack(void)
{
struct cheri_stack cs;
struct cheri_stack_frame *csfp;
u_int stack_depth;
int retval;
retval = sysarch(CHERI_GET_STACK, &cs);
if (retval != 0)
cheritest_failure_err("sysarch(CHERI_GET_STACK) failed");
/* Does stack layout look sensible enough to continue? */
if ((cs.cs_tsize % CHERI_FRAME_SIZE) != 0)
cheritest_failure_errx(
"stack size (%ld) not a multiple of frame size",
cs.cs_tsize);
stack_depth = cs.cs_tsize / CHERI_FRAME_SIZE;
if ((cs.cs_tsp % CHERI_FRAME_SIZE) != 0)
cheritest_failure_errx(
"stack pointer (%ld) not a multiple of frame size",
cs.cs_tsp);
/* Validate that two stack frames are found. */
if (cs.cs_tsp != cs.cs_tsize - (register_t)(2 * CHERI_FRAME_SIZE))
cheritest_failure_errx("stack contains %d frames; expected "
"2", (cs.cs_tsize - (2 * CHERI_FRAME_SIZE)) /
CHERI_FRAME_SIZE);
/* Validate that the first is a saved ambient context. */
csfp = &cs.cs_frames[stack_depth - 1];
if (cheri_getbase(csfp->csf_pcc) != cheri_getbase(cheri_getpcc()) ||
cheri_getlen(csfp->csf_pcc) != cheri_getlen(cheri_getpcc()))
cheritest_failure_errx("frame 0: not global code cap");
/* Validate that the second is cheritest_objectp. */
csfp = &cs.cs_frames[stack_depth - 2];
if ((cheri_getbase(csfp->csf_pcc) != cheri_getbase(
sandbox_object_getobject(cheritest_objectp).co_codecap)) ||
cheri_getlen(csfp->csf_pcc) != cheri_getlen(
sandbox_object_getobject(cheritest_objectp).co_codecap))
cheritest_failure_errx("frame 1: not sandbox code cap");
return (0);
}
static void
sb_read_fn(png_structp png_ptr, png_bytep data, png_size_t length)
{
void *io_ptr = png_get_io_ptr(png_ptr);
#if 0
printf("in sb_read_fn, data base 0x%jx offset 0x%jx length 0x%zx (min len 0x%zx)\n",
cheri_getbase(data), cheri_getoffset(data),
cheri_getlen(data), length);
#endif
libpng_sb_read_callback(io_ptr, cheri_setlen(data, length), length);
}
/*
* Check for high-precision bounds for a variety of small object sizes,
* allocated from the stack. These should be precise regardless of capability
* compression, as the allocator promises to align things suitably. Test both
* static and dynamic allocation.
*/
static void
test_bounds_precise(__capability void *c, size_t expected_len)
{
size_t len, offset;
/* Confirm precise lower bound: offset of zero. */
offset = cheri_getoffset(c);
if (offset != 0)
cheritest_failure_errx("offset (%jd) not zero", offset);
/* Confirm precise upper bound: length of expected size for type. */
len = cheri_getlen(c);
if (len != expected_len)
cheritest_failure_errx("length (%jd) not expected %jd", len,
expected_len);
cheritest_success();
}
context_t act_init(context_t own_context, init_info_t* info, size_t init_base, size_t init_entry) {
KERNEL_TRACE("init", "activation init");
internel_if.message_send = kernel_seal(act_send_message_get_trampoline(), act_ref_type);
internel_if.message_reply = kernel_seal(act_send_return_get_trampoline(), act_sync_ref_type);
setup_syscall_interface(&internel_if);
kernel_next_act = 0;
// This is a dummy. Our first context has already been created
reg_frame_t frame;
bzero(&frame, sizeof(struct reg_frame));
// Register the kernel (exception) activation
act_t * kernel_act = &kernel_acts[0];
act_register(&frame, &kernel_queue.queue, "kernel", status_terminated, NULL, cheri_getbase(cheri_getpcc()));
/* The kernel context already exists and we set it here */
kernel_act->context = own_context;
// Create and register the init activation
KERNEL_TRACE("act", "Retroactively creating init activation");
/* Not a dummy here. We will subset our own c0/pcc for init. init is loaded directly after the kernel */
bzero(&frame, sizeof(struct reg_frame));
size_t length = cheri_getlen(cheri_getdefault()) - init_base;
frame.cf_c0 = cheri_setbounds(cheri_setoffset(cheri_getdefault(), init_base), length);
capability pcc = cheri_setbounds(cheri_setoffset(cheri_getpcc(), init_base), length);
KERNEL_TRACE("act", "assuming init has virtual entry point %lx", init_entry);
frame.cf_c12 = frame.cf_pcc = cheri_setoffset(pcc, init_entry);
/* provide config info to init. c3 is the conventional register */
frame.cf_c3 = info;
act_t * init_act = &kernel_acts[namespace_num_boot];
act_register_create(&frame, &init_queue.queue, "init", status_alive, NULL);
/* The boot activation should be the current activation */
sched_schedule(init_act);
return init_act->context;
}
/*
* Unwind the trusted stack by the specified number of frames (or all).
*/
int
cheri_stack_unwind(ucontext_t *uap, register_t ret, u_int op,
u_int num_frames)
{
struct cheri_frame *cfp;
struct cheri_stack cs;
struct cheri_stack_frame *csfp;
u_int stack_size, stack_frames;
register_t saved_mcreg0;
if (op != CHERI_STACK_UNWIND_OP_N &&
op != CHERI_STACK_UNWIND_OP_ALL) {
errno = EINVAL;
return (-1);
}
/*
* Request to unwind zero frames is a no-op: no state transformation
* is needed.
*/
if ((op == CHERI_STACK_UNWIND_OP_N) && (num_frames == 0))
return (0);
/*
* Retrieve trusted stack and validate before attempting to unwind.
*/
if (sysarch(CHERI_GET_STACK, &cs) != 0)
return (-1);
if ((cs.cs_tsize % CHERI_FRAME_SIZE) != 0 ||
(cs.cs_tsp > cs.cs_tsize) ||
(cs.cs_tsp % CHERI_FRAME_SIZE) != 0) {
errno = ERANGE;
return (-1);
}
/*
* See if there is room on the stack for that much unwinding.
*/
stack_size = cs.cs_tsize / CHERI_FRAME_SIZE;
stack_frames = (cs.cs_tsize - cs.cs_tsp) / CHERI_FRAME_SIZE;
if (op == CHERI_STACK_UNWIND_OP_ALL)
num_frames = stack_frames;
if ((num_frames < 0) || (stack_frames < num_frames)) {
errno = ERANGE;
return (-1);
}
/*
* Restore state from the last frame being unwound.
*/
csfp = &cs.cs_frames[stack_size - (stack_frames - num_frames) - 1];
#if 0
/* Make sure we will be returning to ambient authority. */
if (cheri_getbase(csfp->csf_pcc) != cheri_getbase(cheri_getpcc()) ||
cheri_getlen(csfp->csf_pcc) != cheri_getlen(cheri_getpcc()))
return (-1);
#endif
/*
* Pop stack desired number of frames.
*/
cs.cs_tsp += num_frames * CHERI_FRAME_SIZE;
assert(cs.cs_tsp <= cs.cs_tsize);
#ifdef __CHERI_PURE_CAPABILITY__
cfp = &uap->uc_mcontext.mc_cheriframe;
#else
cfp = (struct cheri_frame *)uap->uc_mcontext.mc_cp2state;
if (cfp == NULL || uap->uc_mcontext.mc_cp2state_len != sizeof(*cfp)) {
errno = ERANGE;
return (-1);
}
#endif
/*
* Zero the capability register file, explicitly restoring $pcc and
* $idc from the last trusted-stack frame.
*/
memset(cfp, 0, sizeof(*cfp));
cfp->cf_idc = csfp->csf_idc;
cfp->cf_pcc = csfp->csf_pcc;
/*
* Zero the general-purpose register file. restore not only $pc, but
* also the slot for $zero, which will hold a magic number across
* sigcode and sigreturn(). Also set a return value.
*
* XXXRW: The kernel unwinder sets V1 to the signal number?
*/
saved_mcreg0 = uap->uc_mcontext.mc_regs[0];
memset(uap->uc_mcontext.mc_regs, 0, sizeof(uap->uc_mcontext.mc_regs));
uap->uc_mcontext.mc_regs[0] = saved_mcreg0;
uap->uc_mcontext.mc_pc = cheri_getoffset(cfp->cf_pcc);
uap->uc_mcontext.mc_regs[V0] = ret;
/*
* Update kernel view of trusted stack.
*/
if (sysarch(CHERI_SET_STACK, &cs) != 0)
return (-1);
return (0);
}
act_t * act_register(reg_frame_t *frame, queue_t *queue, const char *name,
status_e create_in_status, act_control_t *parent, size_t base) {
(void)parent;
KERNEL_TRACE("act", "Registering activation %s", name);
if(kernel_next_act >= MAX_ACTIVATIONS) {
kernel_panic("no act slot");
}
act_t * act = kernel_acts + kernel_next_act;
act->image_base = base;
//TODO bit of a hack. the kernel needs to know what namespace service to use
if(kernel_next_act == namespace_num_namespace) {
KERNEL_TRACE("act", "found namespace");
ns_ref = act_create_sealed_ref(act);
}
#ifndef __LITE__
/* set name */
kernel_assert(ACT_NAME_MAX_LEN > 0);
int name_len = 0;
if(VCAP(name, 1, VCAP_R)) {
name_len = imin(cheri_getlen(name), ACT_NAME_MAX_LEN-1);
}
for(int i = 0; i < name_len; i++) {
char c = name[i];
act->name[i] = c; /* todo: sanitize the name if we do not trust it */
}
act->name[name_len] = '\0';
#endif
/* set status */
act->status = create_in_status;
/*Some "documentation" for the interface between the kernel and activation start *
* These fields are setup by the caller of act_register *
* *
* a0 : user GP argument (goes to main) *
* c3 : user Cap argument (goes to main) *
* *
* These fields are setup by act_register itself. Although the queue is an argument to the function *
* *
* c21 : self control reference *
* c23 : namespace reference (may be null for init and namespace) *
* c24 : kernel interface table *
* c25 : queue */
/* set namespace */
frame->cf_c21 = (capability)act_create_sealed_ctrl_ref(act);
frame->cf_c23 = (capability)ns_ref;
frame->cf_c24 = (capability)get_if();
frame->cf_c25 = (capability)queue;
/* set queue */
msg_queue_init(act, queue);
/* set expected sequence to not expecting */
act->sync_state.sync_token = 0;
act->sync_state.sync_condition = 0;
/* set scheduling status */
sched_create(act);
/*update next_act */
kernel_next_act++;
KERNEL_TRACE("register", "image base of %s is %lx", act->name, act->image_base);
KERNEL_TRACE("act", "%s OK! ", __func__);
return act;
}
