/*
* emit REX byte
*/
static void
emit_rex(struct bpf_jit_state *st, uint32_t op, uint32_t reg, uint32_t rm)
{
uint8_t rex;
/* mark operand registers as used*/
USED(st->reguse, reg);
USED(st->reguse, rm);
rex = 0;
if (BPF_CLASS(op) == EBPF_ALU64 ||
op == (BPF_ST | BPF_MEM | EBPF_DW) ||
op == (BPF_STX | BPF_MEM | EBPF_DW) ||
op == (BPF_STX | EBPF_XADD | EBPF_DW) ||
op == (BPF_LD | BPF_IMM | EBPF_DW) ||
(BPF_CLASS(op) == BPF_LDX &&
BPF_MODE(op) == BPF_MEM &&
BPF_SIZE(op) != BPF_W))
rex |= REX_W;
if (IS_EXT_REG(reg))
rex |= REX_R;
if (IS_EXT_REG(rm))
rex |= REX_B;
/* store using SIL, DIL */
if (op == (BPF_STX | BPF_MEM | BPF_B) && (reg == RDI || reg == RSI))
rex |= REX_PREFIX;
if (rex != 0) {
rex |= REX_PREFIX;
emit_bytes(st, &rex, sizeof(rex));
}
}
void flattenGraphTraversal(struct flatten_item * graph, struct bpf_program * filter, int indice, int level)
{
int j;
struct bpf_insn *insn = &filter->bf_insns[ graph[indice].bf_instruct_link ];
if(level > 50)
return;
for(j = 0;j< level;j++)
{
printf(" ");
}
printf("(%d : %d) fnode (code:%u - k:%u)\n",level, indice, insn->code, insn->k);
/*printf("(%d) %s\n",level, bpf_image(insn,indice));*/
if(BPF_CLASS(insn->code) == BPF_JMP)
{
if(BPF_OP(insn->code) == BPF_JA)
{
flattenGraphTraversal(graph, filter, graph[indice].true_instruct, level+1);
}
else
{
flattenGraphTraversal(graph, filter, graph[indice].true_instruct, level+1);
flattenGraphTraversal(graph, filter, graph[indice].false_instruct, level+1);
}
}
else if(BPF_CLASS(insn->code) != BPF_RET)
{
flattenGraphTraversal(graph, filter, graph[indice].true_instruct, level+1);
}
}
/* Because we really don't have an IR, this stuff is a little messy. */
static int
F(int code, int v0, int v1)
{
u_int hash;
int val;
struct valnode *p;
hash = (u_int)code ^ (v0 << 4) ^ (v1 << 8);
hash %= MODULUS;
for (p = hashtbl[hash]; p; p = p->next)
if (p->code == code && p->v0 == v0 && p->v1 == v1)
return p->val;
val = ++curval;
if (BPF_MODE(code) == BPF_IMM &&
(BPF_CLASS(code) == BPF_LD || BPF_CLASS(code) == BPF_LDX)) {
vmap[val].const_val = v0;
vmap[val].is_const = 1;
}
p = next_vnode++;
p->val = val;
p->code = code;
p->v0 = v0;
p->v1 = v1;
p->next = hashtbl[hash];
hashtbl[hash] = p;
return val;
}
/*
* cette fonction compare le graphe existant avec le nouveau graphe applati
*
* @param graph : le graphe applati
* @param filter : le graphe bpf correspondant
* @param indice : l'indice du noeud a comparer
* @param node : la liste de noeud pouvant contenir un noeud similaire
* @return : le nombre de noeud equivalent dans l'arbre
*
*/
unsigned int compareGraphExploreChild(struct flatten_item * graph, struct bpf_program * filter, int indice, struct filter_node * node,unsigned int * collide)
{
int i = graph[indice].bf_instruct_link;
unsigned int local_collide = 0, child_collide_true = 0, child_collide_false = 0;
unsigned int ret = 0;
struct filter_node * current_node = node;
/*recherche d'un noeud racine correspondant*/
for( ; current_node != NULL ; current_node = current_node->next)
{
if(current_node->item->code == filter->bf_insns[i].code && current_node->item->k == filter->bf_insns[i].k)
{
break;
}
}
/*pas de noeud correspondant, on devra insérer ici*/
if(current_node == NULL)
{
if(indice > 0 && node!= NULL)
{
/*on a un noeud pouvant causer un breakpoint*/
local_collide = 1;
}
*collide = local_collide;
return 0;
}
else
{
if(indice > 0 && node != NULL && node->next!=NULL)
{
local_collide = 1;
}
}
if(BPF_CLASS(filter->bf_insns[i].code) == BPF_JMP && BPF_OP(filter->bf_insns[i].code) != BPF_JA) /*noeud conditionnel*/
{
ret = 1 + compareGraphExploreChild(graph, filter, graph[indice].true_instruct, current_node->item->next_child_t,&child_collide_true)
+ compareGraphExploreChild(graph, filter, graph[indice].false_instruct, current_node->item->next_child_f, &child_collide_false);
}
else if(BPF_CLASS(filter->bf_insns[i].code) != BPF_RET) /*noeud normal*/
{
ret = 1 + compareGraphExploreChild(graph, filter, graph[indice].true_instruct, current_node->item->next_child_t,&child_collide_true);
}
/*else
{
on ne se soucie pas des noeuds terminaux, ils seront ajoutés quoi qu'il arrive, c'est necessaire pour la suppression
}*/
*collide = local_collide + (child_collide_true > child_collide_false)?child_collide_true:child_collide_false;
return ret;
}
/*
* XXX Copied from sys/net/bpf_filter.c and modified.
*
* Return true if the 'fcode' is a valid filter program.
* The constraints are that each jump be forward and to a valid
* code. The code must terminate with either an accept or reject.
*
* The kernel needs to be able to verify an application's filter code.
* Otherwise, a bogus program could easily crash the system.
*/
static int
bpf_validate(const struct bpf_insn *f, int len)
{
register int i;
register const struct bpf_insn *p;
/* Do not accept negative length filter. */
if (len < 0)
return (0);
/* An empty filter means accept all. */
if (len == 0)
return (1);
for (i = 0; i < len; ++i) {
p = &f[i];
/*
* Check that the code is valid.
*/
if (!BPF_VALIDATE_CODE(p->code))
return (0);
/*
* Check that that jumps are forward, and within
* the code block.
*/
if (BPF_CLASS(p->code) == BPF_JMP) {
register u_int offset;
if (p->code == (BPF_JMP|BPF_JA))
offset = p->k;
else
offset = p->jt > p->jf ? p->jt : p->jf;
if (offset >= (u_int)(len - i) - 1)
return (0);
continue;
}
/*
* Check that memory operations use valid addresses.
*/
if (p->code == BPF_ST || p->code == BPF_STX ||
p->code == (BPF_LD|BPF_MEM) ||
p->code == (BPF_LDX|BPF_MEM)) {
if (p->k >= BPF_MEMWORDS)
return (0);
continue;
}
/*
* Check for constant division by 0.
*/
if (p->code == (BPF_ALU|BPF_DIV|BPF_K) && p->k == 0)
return (0);
}
return (BPF_CLASS(f[len - 1].code) == BPF_RET);
}
/**
* sk_chk_filter - verify socket filter code
* @filter: filter to verify
* @flen: length of filter
*
* Check the user's filter code. If we let some ugly
* filter code slip through kaboom! The filter must contain
* no references or jumps that are out of range, no illegal instructions
* and no backward jumps. It must end with a RET instruction
*
* Returns 0 if the rule set is legal or a negative errno code if not.
*/
int sk_chk_filter(struct sock_filter *filter, int flen)
{
struct sock_filter *ftest;
int pc;
if (((unsigned int)flen >= (~0U / sizeof(struct sock_filter))) || flen == 0)
return -EINVAL;
/* check the filter code now */
for (pc = 0; pc < flen; pc++) {
/* all jumps are forward as they are not signed */
ftest = &filter[pc];
if (BPF_CLASS(ftest->code) == BPF_JMP) {
/* but they mustn't jump off the end */
if (BPF_OP(ftest->code) == BPF_JA) {
/*
* Note, the large ftest->k might cause loops.
* Compare this with conditional jumps below,
* where offsets are limited. --ANK (981016)
*/
if (ftest->k >= (unsigned)(flen-pc-1))
return -EINVAL;
} else {
/* for conditionals both must be safe */
if (pc + ftest->jt +1 >= flen ||
pc + ftest->jf +1 >= flen)
return -EINVAL;
}
}
/* check that memory operations use valid addresses. */
if (ftest->k >= BPF_MEMWORDS) {
/* but it might not be a memory operation... */
switch (ftest->code) {
case BPF_ST:
case BPF_STX:
case BPF_LD|BPF_MEM:
case BPF_LDX|BPF_MEM:
return -EINVAL;
}
}
}
/*
* The program must end with a return. We don't care where they
* jumped within the script (its always forwards) but in the end
* they _will_ hit this.
*/
return (BPF_CLASS(filter[flen - 1].code) == BPF_RET) ? 0 : -EINVAL;
}
static void
decode_bpf_code(uint16_t code)
{
uint16_t i = code & ~BPF_CLASS(code);
printxval(bpf_class, BPF_CLASS(code), "BPF_???");
switch (BPF_CLASS(code)) {
case BPF_LD:
case BPF_LDX:
tprints(" | ");
printxval(bpf_size, BPF_SIZE(code), "BPF_???");
tprints(" | ");
printxval(bpf_mode, BPF_MODE(code), "BPF_???");
break;
case BPF_ST:
case BPF_STX:
if (i)
tprintf(" | %#x /* %s */", i, "BPF_???");
break;
case BPF_ALU:
tprints(" | ");
printxval(bpf_src, BPF_SRC(code), "BPF_???");
tprints(" | ");
printxval(bpf_op_alu, BPF_OP(code), "BPF_???");
break;
case BPF_JMP:
tprints(" | ");
printxval(bpf_src, BPF_SRC(code), "BPF_???");
tprints(" | ");
printxval(bpf_op_jmp, BPF_OP(code), "BPF_???");
break;
case BPF_RET:
tprints(" | ");
printxval(bpf_rval, BPF_RVAL(code), "BPF_???");
i &= ~BPF_RVAL(code);
if (i)
tprintf(" | %#x /* %s */", i, "BPF_???");
break;
case BPF_MISC:
tprints(" | ");
printxval(bpf_miscop, BPF_MISCOP(code), "BPF_???");
i &= ~BPF_MISCOP(code);
if (i)
tprintf(" | %#x /* %s */", i, "BPF_???");
break;
}
}
static void
fixup_jumps(npf_bpf_t *ctx, u_int start, u_int end, bool swap)
{
struct bpf_program *bp = &ctx->prog;
for (u_int i = start; i < end; i++) {
struct bpf_insn *insn = &bp->bf_insns[i];
const u_int fail_off = end - i;
if (fail_off >= JUMP_MAGIC) {
errx(EXIT_FAILURE, "BPF generation error: "
"the number of instructions is over the limit");
}
if (BPF_CLASS(insn->code) != BPF_JMP) {
continue;
}
if (swap) {
uint8_t jt = insn->jt;
insn->jt = insn->jf;
insn->jf = jt;
}
if (insn->jt == JUMP_MAGIC)
insn->jt = fail_off;
if (insn->jf == JUMP_MAGIC)
insn->jf = fail_off;
}
}
/*
* cette fonction met a plat un graphe bpf
*
* @param graph : le graphe applati à remplir
* @param filter : le graphe bpf à convertir
* @param non_terminal_count : le nombre de noeud non terminaux dans le graphe applati
* @param terminal_count : le nombre de noeud terminaux dans le graphe
* @param i : l'indice de l'instruction dans le graphe bpf
* @param id : le prochain identifiant disponible pour ajouter un noeud
* @return : l'id du noeud ajouté dans l'appel de la fonction
*/
int flatteGraph(struct flatten_item * graph, struct bpf_program * filter,int non_terminal_count, int terminal_count, int i, int * id)
{
int current_id, indice;
if(BPF_CLASS(filter->bf_insns[i].code) == BPF_JMP) /*JUMP INSTRUCTION*/
{
current_id = (*id)++;
graph[current_id].bf_instruct_link = i;
if(BPF_OP(filter->bf_insns[i].code) == BPF_JA)
{
graph[current_id].true_instruct = flatteGraph(graph, filter, non_terminal_count, terminal_count, filter->bf_insns[i].k+i+1,id);
}
else
{
graph[current_id].true_instruct = flatteGraph(graph, filter, non_terminal_count, terminal_count, filter->bf_insns[i].jt+i+1, id);
graph[current_id].false_instruct = flatteGraph(graph, filter, non_terminal_count, terminal_count, filter->bf_insns[i].jf+i+1, id);
}
}
else if(BPF_CLASS(filter->bf_insns[i].code) != BPF_RET) /*OTHER INSTRUCTION*/
{
current_id = (*id)++;
graph[current_id].bf_instruct_link = i;
graph[current_id].true_instruct = flatteGraph(graph, filter, non_terminal_count, terminal_count, i+1,id);
}
else /*TERMINAL INSTRUCTION*/
{
for(indice = 0 ; (indice<terminal_count) && ( graph[non_terminal_count + indice].bf_instruct_link != -1) ; indice++)
{
if(graph[non_terminal_count + indice].bf_instruct_link == i)
{
graph[non_terminal_count + indice].true_instruct += 1;
return non_terminal_count + indice;
}
}
/*noeud terminal non encore recontré, on l'ajoute*/
current_id = non_terminal_count + indice;
graph[current_id].bf_instruct_link = i;
graph[current_id].true_instruct = 1;
graph[current_id].false_instruct = indice;
}
return current_id;
}
/*
* Compute the sets of registers used, defined, and killed by 'b'.
*
* "Used" means that a statement in 'b' uses the register before any
* statement in 'b' defines it, i.e. it uses the value left in
* that register by a predecessor block of this block.
* "Defined" means that a statement in 'b' defines it.
* "Killed" means that a statement in 'b' defines it before any
* statement in 'b' uses it, i.e. it kills the value left in that
* register by a predecessor block of this block.
*/
static void
compute_local_ud(struct block *b)
{
struct slist *s;
atomset def = 0, use = 0, kill = 0;
int atom;
for (s = b->stmts; s; s = s->next) {
if (s->s.code == NOP)
continue;
atom = atomuse(&s->s);
if (atom >= 0) {
if (atom == AX_ATOM) {
if (!ATOMELEM(def, X_ATOM))
use |= ATOMMASK(X_ATOM);
if (!ATOMELEM(def, A_ATOM))
use |= ATOMMASK(A_ATOM);
}
else if (atom < N_ATOMS) {
if (!ATOMELEM(def, atom))
use |= ATOMMASK(atom);
}
else
abort();
}
atom = atomdef(&s->s);
if (atom >= 0) {
if (!ATOMELEM(use, atom))
kill |= ATOMMASK(atom);
def |= ATOMMASK(atom);
}
}
if (BPF_CLASS(b->s.code) == BPF_JMP) {
/*
* XXX - what about RET?
*/
atom = atomuse(&b->s);
if (atom >= 0) {
if (atom == AX_ATOM) {
if (!ATOMELEM(def, X_ATOM))
use |= ATOMMASK(X_ATOM);
if (!ATOMELEM(def, A_ATOM))
use |= ATOMMASK(A_ATOM);
}
else if (atom < N_ATOMS) {
if (!ATOMELEM(def, atom))
use |= ATOMMASK(atom);
}
else
abort();
}
}
b->def = def;
b->kill = kill;
b->in_use = use;
}
static void
make_marks(struct block *p)
{
if (!isMarked(p)) {
Mark(p);
if (BPF_CLASS(p->s.code) != BPF_RET) {
make_marks(JT(p));
make_marks(JF(p));
}
}
}
/*
* emit mov %<sreg>, %<dreg>
*/
static void
emit_mov_reg(struct bpf_jit_state *st, uint32_t op, uint32_t sreg,
uint32_t dreg)
{
const uint8_t ops = 0x89;
/* if operands are 32-bit, then it can be used to clear upper 32-bit */
if (sreg != dreg || BPF_CLASS(op) == BPF_ALU) {
emit_rex(st, op, sreg, dreg);
emit_bytes(st, &ops, sizeof(ops));
emit_modregrm(st, MOD_DIRECT, sreg, dreg);
}
}
static void
opt_root(struct block **b)
{
struct slist *tmp, *s;
s = (*b)->stmts;
(*b)->stmts = 0;
while (BPF_CLASS((*b)->s.code) == BPF_JMP && JT(*b) == JF(*b))
*b = JT(*b);
tmp = (*b)->stmts;
if (tmp != 0)
sappend(s, tmp);
(*b)->stmts = s;
/*
* If the root node is a return, then there is no
* point executing any statements (since the bpf machine
* has no side effects).
*/
if (BPF_CLASS((*b)->s.code) == BPF_RET)
(*b)->stmts = 0;
}
/*
* cette methode explore le graphe existant et compte le nombre de noeud NON TERMINAUX necessaire pour la mise a plat du graphe
*/
unsigned int countNoTerminalNodes(struct bpf_program * filter, int i)
{
if(BPF_CLASS(filter->bf_insns[i].code) == BPF_JMP)
{
if(BPF_OP(filter->bf_insns[i].code) == BPF_JA)
{
return 1 + countNoTerminalNodes(filter,filter->bf_insns[i].k+i+1);
}
else
{
return 1 + countNoTerminalNodes(filter,filter->bf_insns[i].jt+i+1)
+ countNoTerminalNodes(filter,filter->bf_insns[i].jf+i+1);
}
}
else if(BPF_CLASS(filter->bf_insns[i].code) != BPF_RET)
{
return 1 + countNoTerminalNodes(filter,i+1);
}
/*on ne compte pas les noeuds terminaux ici*/
return 0;
}
/*
* cette methode explore le graphe existant et compte le nombre de noeud TERMINAUX necessaire pour la mise a plat du graphe
*/
unsigned int countTerminalNodes(struct bpf_program * filter)
{
unsigned int ret = 0, i;
for(i=0;i<filter->bf_len;i++)
{
if( BPF_CLASS(filter->bf_insns[i].code) == BPF_RET )
{
ret += 1;
}
}
return ret;
}
/*
* Scan the filter program and find possible optimization.
*/
static int
bpf_jit_optimize(struct bpf_insn *prog, u_int nins)
{
int flags;
u_int i;
/* Do we return immediately? */
if (BPF_CLASS(prog[0].code) == BPF_RET)
return (BPF_JIT_FRET);
for (flags = 0, i = 0; i < nins; i++) {
switch (prog[i].code) {
case BPF_LD|BPF_W|BPF_ABS:
case BPF_LD|BPF_H|BPF_ABS:
case BPF_LD|BPF_B|BPF_ABS:
case BPF_LD|BPF_W|BPF_IND:
case BPF_LD|BPF_H|BPF_IND:
case BPF_LD|BPF_B|BPF_IND:
case BPF_LDX|BPF_MSH|BPF_B:
flags |= BPF_JIT_FPKT;
break;
case BPF_LD|BPF_MEM:
case BPF_LDX|BPF_MEM:
case BPF_ST:
case BPF_STX:
flags |= BPF_JIT_FMEM;
break;
case BPF_LD|BPF_W|BPF_LEN:
case BPF_LDX|BPF_W|BPF_LEN:
flags |= BPF_JIT_FLEN;
break;
case BPF_JMP|BPF_JA:
case BPF_JMP|BPF_JGT|BPF_K:
case BPF_JMP|BPF_JGE|BPF_K:
case BPF_JMP|BPF_JEQ|BPF_K:
case BPF_JMP|BPF_JSET|BPF_K:
case BPF_JMP|BPF_JGT|BPF_X:
case BPF_JMP|BPF_JGE|BPF_X:
case BPF_JMP|BPF_JEQ|BPF_X:
case BPF_JMP|BPF_JSET|BPF_X:
flags |= BPF_JIT_FJMP;
break;
}
if (flags == BPF_JIT_FLAG_ALL)
break;
}
return (flags);
}
/*
* emit mov <imm>, %<dreg>
*/
static void
emit_mov_imm(struct bpf_jit_state *st, uint32_t op, uint32_t dreg, uint32_t imm)
{
const uint8_t ops = 0xC7;
if (imm == 0) {
/* replace 'mov 0, %<dst>' with 'xor %<dst>, %<dst>' */
op = BPF_CLASS(op) | BPF_XOR | BPF_X;
emit_alu_reg(st, op, dreg, dreg);
return;
}
emit_rex(st, op, 0, dreg);
emit_bytes(st, &ops, sizeof(ops));
emit_modregrm(st, MOD_DIRECT, 0, dreg);
emit_imm(st, imm, sizeof(imm));
}
/*
* emit one of:
* mov %<sreg>, <ofs>(%<dreg>)
* mov <imm>, <ofs>(%<dreg>)
*/
static void
emit_st_common(struct bpf_jit_state *st, uint32_t op, uint32_t sreg,
uint32_t dreg, uint32_t imm, int32_t ofs)
{
uint32_t mods, imsz, opsz, opx;
const uint8_t prfx16 = 0x66;
/* 8 bit instruction opcodes */
static const uint8_t op8[] = {0xC6, 0x88};
/* 16/32/64 bit instruction opcodes */
static const uint8_t ops[] = {0xC7, 0x89};
/* is the instruction has immediate value or src reg? */
opx = (BPF_CLASS(op) == BPF_STX);
opsz = BPF_SIZE(op);
if (opsz == BPF_H)
emit_bytes(st, &prfx16, sizeof(prfx16));
emit_rex(st, op, sreg, dreg);
if (opsz == BPF_B)
emit_bytes(st, &op8[opx], sizeof(op8[opx]));
else
emit_bytes(st, &ops[opx], sizeof(ops[opx]));
imsz = imm_size(ofs);
mods = (imsz == 1) ? MOD_IDISP8 : MOD_IDISP32;
emit_modregrm(st, mods, sreg, dreg);
if (dreg == RSP || dreg == R12)
emit_sib(st, SIB_SCALE_1, dreg, dreg);
emit_imm(st, ofs, imsz);
if (opx == 0) {
imsz = RTE_MIN(bpf_size(opsz), sizeof(imm));
emit_imm(st, imm, imsz);
}
}
static void
decode_bpf_stmt(const struct bpf_filter *filter)
{
#ifdef HAVE_LINUX_FILTER_H
tprints("BPF_STMT(");
decode_bpf_code(filter->code);
tprints(", ");
if (BPF_CLASS(filter->code) == BPF_RET) {
unsigned int action = SECCOMP_RET_ACTION & filter->k;
unsigned int data = filter->k & ~action;
printxval(seccomp_ret_action, action, "SECCOMP_RET_???");
if (data)
tprintf(" | %#x)", data);
else
tprints(")");
} else {
tprintf("%#x)", filter->k);
}
#else
tprintf("BPF_STMT(%#x, %#x)", filter->code, filter->k);
#endif /* HAVE_LINUX_FILTER_H */
}
void innerTraversal(const char * pre,struct filter_item * i, int level)
{
struct filter_node * node_tmp = NULL;
int j;
if(i == NULL)
{
return;
}
for(j = 0;j< level;j++)
{
printf(" ");
}
printf("%s",pre);
printf("(%d) fnode (code:%u - k:%u) \n",level, i->code, i->k);
if(BPF_CLASS(i->code) != BPF_RET)
{
node_tmp = i->next_child_t;
while(node_tmp != NULL)
{
if(i->next_child_f == NULL)
innerTraversal("",node_tmp->item, level+1);
else
innerTraversal("(t):",node_tmp->item, level+1);
node_tmp = node_tmp->next;
}
node_tmp = i->next_child_f;
while(node_tmp != NULL)
{
innerTraversal("(f):",node_tmp->item, level+1);
node_tmp = node_tmp->next;
}
}
}
/*
* Return the register number that is used by s. If A and X are both
* used, return AX_ATOM. If no register is used, return -1.
*
* The implementation should probably change to an array access.
*/
static int
atomuse(struct stmt *s)
{
register int c = s->code;
if (c == NOP)
return -1;
switch (BPF_CLASS(c)) {
case BPF_RET:
return (BPF_RVAL(c) == BPF_A) ? A_ATOM :
(BPF_RVAL(c) == BPF_X) ? X_ATOM : -1;
case BPF_LD:
case BPF_LDX:
return (BPF_MODE(c) == BPF_IND) ? X_ATOM :
(BPF_MODE(c) == BPF_MEM) ? s->k : -1;
case BPF_ST:
return A_ATOM;
case BPF_STX:
return X_ATOM;
case BPF_JMP:
case BPF_ALU:
if (BPF_SRC(c) == BPF_X)
return AX_ATOM;
return A_ATOM;
case BPF_MISC:
return BPF_MISCOP(c) == BPF_TXA ? X_ATOM : A_ATOM;
}
abort();
/* NOTREACHED */
}
static int add_monitor_filter(int s)
{
int idx;
/* rewrite all PASS/FAIL jump offsets */
for (idx = 0; idx < msock_filter.len; idx++)
{
struct sock_filter *insn = &msock_filter_insns[idx];
if (BPF_CLASS(insn->code) == BPF_JMP) {
if (insn->code == (BPF_JMP | BPF_JA)) {
if (insn->k == PASS)
insn->k = msock_filter.len - idx - 2;
else if (insn->k == FAIL)
insn->k = msock_filter.len - idx - 3;
}
if (insn->jt == PASS)
insn->jt = msock_filter.len - idx - 2;
else if (insn->jt == FAIL)
insn->jt = msock_filter.len - idx - 3;
if (insn->jf == PASS)
insn->jf = msock_filter.len - idx - 2;
else if (insn->jf == FAIL)
insn->jf = msock_filter.len - idx - 3;
}
}
if (setsockopt(s, SOL_SOCKET, SO_ATTACH_FILTER,
&msock_filter, sizeof(msock_filter))) {
fprintf(stderr, "nl80211: setsockopt(SO_ATTACH_FILTER) failed: %s\n", strerror(errno));
return -1;
}
return 0;
}
/*
* Return the register number that is defined by 's'. We assume that
* a single stmt cannot define more than one register. If no register
* is defined, return -1.
*
* The implementation should probably change to an array access.
*/
static int
atomdef(struct stmt *s)
{
if (s->code == NOP)
return -1;
switch (BPF_CLASS(s->code)) {
case BPF_LD:
case BPF_ALU:
return A_ATOM;
case BPF_LDX:
return X_ATOM;
case BPF_ST:
case BPF_STX:
return s->k;
case BPF_MISC:
return BPF_MISCOP(s->code) == BPF_TAX ? X_ATOM : A_ATOM;
}
return -1;
}
/*
* Returns true if successful. Returns false if a branch has
* an offset that is too large. If so, we have marked that
* branch so that on a subsequent iteration, it will be treated
* properly.
*/
static int
convert_code_r(struct block *p)
{
struct bpf_insn *dst;
struct slist *src;
u_int slen;
u_int off;
int extrajmps; /* number of extra jumps inserted */
struct slist **offset = NULL;
if (p == 0 || isMarked(p))
return (1);
Mark(p);
if (convert_code_r(JF(p)) == 0)
return (0);
if (convert_code_r(JT(p)) == 0)
return (0);
slen = slength(p->stmts);
dst = ftail -= (slen + 1 + p->longjt + p->longjf);
/* inflate length by any extra jumps */
p->offset = dst - fstart;
/* generate offset[] for convenience */
if (slen) {
offset = (struct slist **)calloc(slen, sizeof(struct slist *));
if (!offset) {
bpf_error("not enough core");
/*NOTREACHED*/
}
}
src = p->stmts;
for (off = 0; off < slen && src; off++) {
#if 0
printf("off=%d src=%x\n", off, src);
#endif
offset[off] = src;
src = src->next;
}
off = 0;
for (src = p->stmts; src; src = src->next) {
if (src->s.code == NOP)
continue;
dst->code = (u_short)src->s.code;
dst->k = src->s.k;
/* fill block-local relative jump */
if (BPF_CLASS(src->s.code) != BPF_JMP || src->s.code == (BPF_JMP|BPF_JA)) {
#if 0
if (src->s.jt || src->s.jf) {
bpf_error("illegal jmp destination");
/*NOTREACHED*/
}
#endif
goto filled;
}
if (off == slen - 2) /*???*/
goto filled;
{
u_int i;
int jt, jf;
static const char ljerr[] = "%s for block-local relative jump: off=%d";
#if 0
printf("code=%x off=%d %x %x\n", src->s.code,
off, src->s.jt, src->s.jf);
#endif
if (!src->s.jt || !src->s.jf) {
bpf_error(ljerr, "no jmp destination", off);
/*NOTREACHED*/
}
jt = jf = 0;
for (i = 0; i < slen; i++) {
if (offset[i] == src->s.jt) {
if (jt) {
bpf_error(ljerr, "multiple matches", off);
/*NOTREACHED*/
}
dst->jt = i - off - 1;
jt++;
}
if (offset[i] == src->s.jf) {
if (jf) {
bpf_error(ljerr, "multiple matches", off);
/*NOTREACHED*/
}
dst->jf = i - off - 1;
jf++;
}
}
if (!jt || !jf) {
bpf_error(ljerr, "no destination found", off);
/*NOTREACHED*/
}
}
filled:
++dst;
++off;
}
if (offset)
free(offset);
#ifdef BDEBUG
bids[dst - fstart] = p->id + 1;
#endif
dst->code = (u_short)p->s.code;
dst->k = p->s.k;
if (JT(p)) {
extrajmps = 0;
off = JT(p)->offset - (p->offset + slen) - 1;
if (off >= 256) {
/* offset too large for branch, must add a jump */
if (p->longjt == 0) {
/* mark this instruction and retry */
p->longjt++;
return(0);
}
/* branch if T to following jump */
dst->jt = extrajmps;
extrajmps++;
dst[extrajmps].code = BPF_JMP|BPF_JA;
dst[extrajmps].k = off - extrajmps;
}
else
dst->jt = off;
off = JF(p)->offset - (p->offset + slen) - 1;
if (off >= 256) {
/* offset too large for branch, must add a jump */
if (p->longjf == 0) {
/* mark this instruction and retry */
p->longjf++;
return(0);
}
/* branch if F to following jump */
/* if two jumps are inserted, F goes to second one */
dst->jf = extrajmps;
extrajmps++;
dst[extrajmps].code = BPF_JMP|BPF_JA;
dst[extrajmps].k = off - extrajmps;
}
else
dst->jf = off;
}
return (1);
}
static char *bpf_dump(const struct sock_filter bpf, int n)
{
int v;
const char *fmt, *op;
static char image[256];
char operand[64];
v = bpf.k;
switch (bpf.code) {
default:
op = "unimp";
fmt = "0x%x";
v = bpf.code;
break;
case BPF_RET | BPF_K:
op = "ret";
fmt = "#0x%x";
break;
case BPF_RET | BPF_A:
op = "ret";
fmt = "";
break;
case BPF_LD | BPF_W | BPF_ABS:
op = "ld";
fmt = "[%d]";
break;
case BPF_LD | BPF_H | BPF_ABS:
op = "ldh";
fmt = "[%d]";
break;
case BPF_LD | BPF_B | BPF_ABS:
op = "ldb";
fmt = "[%d]";
break;
case BPF_LD | BPF_W | BPF_LEN:
op = "ld";
fmt = "#pktlen";
break;
case BPF_LD | BPF_W | BPF_IND:
op = "ld";
fmt = "[x + %d]";
break;
case BPF_LD | BPF_H | BPF_IND:
op = "ldh";
fmt = "[x + %d]";
break;
case BPF_LD | BPF_B | BPF_IND:
op = "ldb";
fmt = "[x + %d]";
break;
case BPF_LD | BPF_IMM:
op = "ld";
fmt = "#0x%x";
break;
case BPF_LDX | BPF_IMM:
op = "ldx";
fmt = "#0x%x";
break;
case BPF_LDX | BPF_MSH | BPF_B:
op = "ldxb";
fmt = "4*([%d]&0xf)";
break;
case BPF_LD | BPF_MEM:
op = "ld";
fmt = "M[%d]";
break;
case BPF_LDX | BPF_MEM:
op = "ldx";
fmt = "M[%d]";
break;
case BPF_ST:
op = "st";
fmt = "M[%d]";
break;
case BPF_STX:
op = "stx";
fmt = "M[%d]";
break;
case BPF_JMP | BPF_JA:
op = "ja";
fmt = "%d";
v = n + 1 + bpf.k;
break;
case BPF_JMP | BPF_JGT | BPF_K:
op = "jgt";
fmt = "#0x%x";
break;
case BPF_JMP | BPF_JGE | BPF_K:
op = "jge";
fmt = "#0x%x";
break;
case BPF_JMP | BPF_JEQ | BPF_K:
op = "jeq";
fmt = "#0x%x";
break;
case BPF_JMP | BPF_JSET | BPF_K:
op = "jset";
fmt = "#0x%x";
break;
case BPF_JMP | BPF_JGT | BPF_X:
op = "jgt";
fmt = "x";
break;
case BPF_JMP | BPF_JGE | BPF_X:
op = "jge";
fmt = "x";
break;
case BPF_JMP | BPF_JEQ | BPF_X:
op = "jeq";
fmt = "x";
break;
case BPF_JMP | BPF_JSET | BPF_X:
op = "jset";
fmt = "x";
break;
case BPF_ALU | BPF_ADD | BPF_X:
op = "add";
fmt = "x";
break;
case BPF_ALU | BPF_SUB | BPF_X:
op = "sub";
fmt = "x";
break;
case BPF_ALU | BPF_MUL | BPF_X:
op = "mul";
fmt = "x";
break;
case BPF_ALU | BPF_DIV | BPF_X:
op = "div";
fmt = "x";
break;
case BPF_ALU | BPF_AND | BPF_X:
op = "and";
fmt = "x";
break;
case BPF_ALU | BPF_OR | BPF_X:
op = "or";
fmt = "x";
break;
case BPF_ALU | BPF_LSH | BPF_X:
op = "lsh";
fmt = "x";
break;
case BPF_ALU | BPF_RSH | BPF_X:
op = "rsh";
fmt = "x";
break;
case BPF_ALU | BPF_ADD | BPF_K:
op = "add";
fmt = "#%d";
break;
case BPF_ALU | BPF_SUB | BPF_K:
op = "sub";
fmt = "#%d";
break;
case BPF_ALU | BPF_MUL | BPF_K:
op = "mul";
fmt = "#%d";
break;
case BPF_ALU | BPF_DIV | BPF_K:
op = "div";
fmt = "#%d";
break;
case BPF_ALU | BPF_AND | BPF_K:
op = "and";
fmt = "#0x%x";
break;
case BPF_ALU | BPF_OR | BPF_K:
op = "or";
fmt = "#0x%x";
break;
case BPF_ALU | BPF_LSH | BPF_K:
op = "lsh";
fmt = "#%d";
break;
case BPF_ALU | BPF_RSH | BPF_K:
op = "rsh";
fmt = "#%d";
break;
case BPF_ALU | BPF_NEG:
op = "neg";
fmt = "";
break;
case BPF_MISC | BPF_TAX:
op = "tax";
fmt = "";
break;
case BPF_MISC | BPF_TXA:
op = "txa";
fmt = "";
break;
}
slprintf(operand, sizeof(operand), fmt, v);
slprintf(image, sizeof(image),
(BPF_CLASS(bpf.code) == BPF_JMP &&
BPF_OP(bpf.code) != BPF_JA) ?
" L%d: %s %s, L%d, L%d" : " L%d: %s %s",
n, op, operand, n + 1 + bpf.jt, n + 1 + bpf.jf);
return image;
}
int bpf_validate(const struct sock_fprog *bpf)
{
uint32_t i, from;
const struct sock_filter *p;
if (!bpf)
return 0;
if (bpf->len < 1)
return 0;
for (i = 0; i < bpf->len; ++i) {
p = &bpf->filter[i];
switch (BPF_CLASS(p->code)) {
/*
* Check that memory operations use valid addresses.
*/
case BPF_LD:
case BPF_LDX:
switch (BPF_MODE(p->code)) {
case BPF_IMM:
break;
case BPF_ABS:
case BPF_IND:
case BPF_MSH:
/*
* There's no maximum packet data size
* in userland. The runtime packet length
* check suffices.
*/
break;
case BPF_MEM:
if (p->k >= BPF_MEMWORDS)
return 0;
break;
case BPF_LEN:
break;
default:
return 0;
}
break;
case BPF_ST:
case BPF_STX:
if (p->k >= BPF_MEMWORDS)
return 0;
break;
case BPF_ALU:
switch (BPF_OP(p->code)) {
case BPF_ADD:
case BPF_SUB:
case BPF_MUL:
case BPF_OR:
case BPF_AND:
case BPF_LSH:
case BPF_RSH:
case BPF_NEG:
break;
case BPF_DIV:
/*
* Check for constant division by 0.
*/
if (BPF_RVAL(p->code) == BPF_K && p->k == 0)
return 0;
break;
default:
return 0;
}
break;
case BPF_JMP:
/*
* Check that jumps are within the code block,
* and that unconditional branches don't go
* backwards as a result of an overflow.
* Unconditional branches have a 32-bit offset,
* so they could overflow; we check to make
* sure they don't. Conditional branches have
* an 8-bit offset, and the from address is <=
* BPF_MAXINSNS, and we assume that BPF_MAXINSNS
* is sufficiently small that adding 255 to it
* won't overflow.
*
* We know that len is <= BPF_MAXINSNS, and we
* assume that BPF_MAXINSNS is < the maximum size
* of a u_int, so that i + 1 doesn't overflow.
*
* For userland, we don't know that the from
* or len are <= BPF_MAXINSNS, but we know that
* from <= len, and, except on a 64-bit system,
* it's unlikely that len, if it truly reflects
* the size of the program we've been handed,
* will be anywhere near the maximum size of
* a u_int. We also don't check for backward
* branches, as we currently support them in
* userland for the protochain operation.
*/
from = i + 1;
switch (BPF_OP(p->code)) {
case BPF_JA:
if (from + p->k >= bpf->len)
return 0;
break;
case BPF_JEQ:
case BPF_JGT:
case BPF_JGE:
case BPF_JSET:
if (from + p->jt >= bpf->len ||
from + p->jf >= bpf->len)
return 0;
break;
default:
return 0;
}
break;
case BPF_RET:
break;
case BPF_MISC:
break;
default:
return 0;
}
}
return BPF_CLASS(bpf->filter[bpf->len - 1].code) == BPF_RET;
}
static void
opt_peep(struct block *b)
{
struct slist *s;
struct slist *next, *last;
int val;
s = b->stmts;
if (s == 0)
return;
last = s;
for (/*empty*/; /*empty*/; s = next) {
/*
* Skip over nops.
*/
s = this_op(s);
if (s == 0)
break; /* nothing left in the block */
/*
* Find the next real instruction after that one
* (skipping nops).
*/
next = this_op(s->next);
if (next == 0)
break; /* no next instruction */
last = next;
/*
* st M[k] --> st M[k]
* ldx M[k] tax
*/
if (s->s.code == BPF_ST &&
next->s.code == (BPF_LDX|BPF_MEM) &&
s->s.k == next->s.k) {
done = 0;
next->s.code = BPF_MISC|BPF_TAX;
}
/*
* ld #k --> ldx #k
* tax txa
*/
if (s->s.code == (BPF_LD|BPF_IMM) &&
next->s.code == (BPF_MISC|BPF_TAX)) {
s->s.code = BPF_LDX|BPF_IMM;
next->s.code = BPF_MISC|BPF_TXA;
done = 0;
}
/*
* This is an ugly special case, but it happens
* when you say tcp[k] or udp[k] where k is a constant.
*/
if (s->s.code == (BPF_LD|BPF_IMM)) {
struct slist *add, *tax, *ild;
/*
* Check that X isn't used on exit from this
* block (which the optimizer might cause).
* We know the code generator won't generate
* any local dependencies.
*/
if (ATOMELEM(b->out_use, X_ATOM))
continue;
/*
* Check that the instruction following the ldi
* is an addx, or it's an ldxms with an addx
* following it (with 0 or more nops between the
* ldxms and addx).
*/
if (next->s.code != (BPF_LDX|BPF_MSH|BPF_B))
add = next;
else
add = this_op(next->next);
if (add == 0 || add->s.code != (BPF_ALU|BPF_ADD|BPF_X))
continue;
/*
* Check that a tax follows that (with 0 or more
* nops between them).
*/
tax = this_op(add->next);
if (tax == 0 || tax->s.code != (BPF_MISC|BPF_TAX))
continue;
/*
* Check that an ild follows that (with 0 or more
* nops between them).
*/
ild = this_op(tax->next);
if (ild == 0 || BPF_CLASS(ild->s.code) != BPF_LD ||
BPF_MODE(ild->s.code) != BPF_IND)
continue;
/*
* We want to turn this sequence:
*
* (004) ldi #0x2 {s}
* (005) ldxms [14] {next} -- optional
* (006) addx {add}
* (007) tax {tax}
* (008) ild [x+0] {ild}
*
* into this sequence:
*
* (004) nop
* (005) ldxms [14]
* (006) nop
* (007) nop
* (008) ild [x+2]
*
* XXX We need to check that X is not
* subsequently used, because we want to change
* what'll be in it after this sequence.
*
* We know we can eliminate the accumulator
* modifications earlier in the sequence since
* it is defined by the last stmt of this sequence
* (i.e., the last statement of the sequence loads
* a value into the accumulator, so we can eliminate
* earlier operations on the accumulator).
*/
ild->s.k += s->s.k;
s->s.code = NOP;
add->s.code = NOP;
tax->s.code = NOP;
done = 0;
}
}
/*
* If the comparison at the end of a block is an equality
* comparison against a constant, and nobody uses the value
* we leave in the A register at the end of a block, and
* the operation preceding the comparison is an arithmetic
* operation, we can sometime optimize it away.
*/
if (b->s.code == (BPF_JMP|BPF_JEQ|BPF_K) &&
!ATOMELEM(b->out_use, A_ATOM)) {
/*
* We can optimize away certain subtractions of the
* X register.
*/
if (last->s.code == (BPF_ALU|BPF_SUB|BPF_X)) {
val = b->val[X_ATOM];
if (vmap[val].is_const) {
/*
* If we have a subtract to do a comparison,
* and the X register is a known constant,
* we can merge this value into the
* comparison:
*
* sub x -> nop
* jeq #y jeq #(x+y)
*/
b->s.k += vmap[val].const_val;
last->s.code = NOP;
done = 0;
} else if (b->s.k == 0) {
/*
* If the X register isn't a constant,
* and the comparison in the test is
* against 0, we can compare with the
* X register, instead:
*
* sub x -> nop
* jeq #0 jeq x
*/
last->s.code = NOP;
b->s.code = BPF_JMP|BPF_JEQ|BPF_X;
done = 0;
}
}
/*
* Likewise, a constant subtract can be simplified:
*
* sub #x -> nop
* jeq #y -> jeq #(x+y)
*/
else if (last->s.code == (BPF_ALU|BPF_SUB|BPF_K)) {
last->s.code = NOP;
b->s.k += last->s.k;
done = 0;
}
/*
* And, similarly, a constant AND can be simplified
* if we're testing against 0, i.e.:
*
* and #k nop
* jeq #0 -> jset #k
*/
else if (last->s.code == (BPF_ALU|BPF_AND|BPF_K) &&
b->s.k == 0) {
b->s.k = last->s.k;
b->s.code = BPF_JMP|BPF_K|BPF_JSET;
last->s.code = NOP;
done = 0;
opt_not(b);
}
}
/*
* jset #0 -> never
* jset #ffffffff -> always
*/
if (b->s.code == (BPF_JMP|BPF_K|BPF_JSET)) {
if (b->s.k == 0)
JT(b) = JF(b);
if (b->s.k == (int)0xffffffff)
JF(b) = JT(b);
}
/*
* If we're comparing against the index register, and the index
* register is a known constant, we can just compare against that
* constant.
*/
val = b->val[X_ATOM];
if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_X) {
bpf_int32 v = vmap[val].const_val;
b->s.code &= ~BPF_X;
b->s.k = v;
}
/*
* If the accumulator is a known constant, we can compute the
* comparison result.
*/
val = b->val[A_ATOM];
if (vmap[val].is_const && BPF_SRC(b->s.code) == BPF_K) {
bpf_int32 v = vmap[val].const_val;
switch (BPF_OP(b->s.code)) {
case BPF_JEQ:
v = v == b->s.k;
break;
case BPF_JGT:
v = (unsigned)v > (unsigned)b->s.k;
break;
case BPF_JGE:
v = (unsigned)v >= (unsigned)b->s.k;
break;
case BPF_JSET:
v &= b->s.k;
break;
default:
abort();
}
if (JF(b) != JT(b))
done = 0;
if (v)
JF(b) = JT(b);
else
JT(b) = JF(b);
}
}
static void update_regs_access(cs_struct *ud, cs_detail *detail,
bpf_insn insn_id, unsigned int opcode)
{
if (insn_id == BPF_INS_INVALID)
return;
#define PUSH_READ(r) do { \
detail->regs_read[detail->regs_read_count] = r; \
detail->regs_read_count++; \
} while (0)
#define PUSH_WRITE(r) do { \
detail->regs_write[detail->regs_write_count] = r; \
detail->regs_write_count++; \
} while (0)
/*
* In eBPF mode, only these instructions have implicit registers access:
* - ld{w,h,b,dw} * // w: r0
* - exit // r: r0
*/
if (EBPF_MODE(ud)) {
switch (insn_id) {
default:
break;
case BPF_INS_LDW:
case BPF_INS_LDH:
case BPF_INS_LDB:
case BPF_INS_LDDW:
PUSH_WRITE(BPF_REG_R0);
break;
case BPF_INS_EXIT:
PUSH_READ(BPF_REG_R0);
break;
}
return;
}
/* cBPF mode */
switch (BPF_CLASS(opcode)) {
default:
break;
case BPF_CLASS_LD:
PUSH_WRITE(BPF_REG_A);
break;
case BPF_CLASS_LDX:
PUSH_WRITE(BPF_REG_X);
break;
case BPF_CLASS_ST:
PUSH_READ(BPF_REG_A);
break;
case BPF_CLASS_STX:
PUSH_READ(BPF_REG_X);
break;
case BPF_CLASS_ALU:
PUSH_READ(BPF_REG_A);
PUSH_WRITE(BPF_REG_A);
break;
case BPF_CLASS_JMP:
if (insn_id != BPF_INS_JMP) // except the unconditional jump
PUSH_READ(BPF_REG_A);
break;
/* case BPF_CLASS_RET: */
case BPF_CLASS_MISC:
if (insn_id == BPF_INS_TAX) {
PUSH_READ(BPF_REG_A);
PUSH_WRITE(BPF_REG_X);
}
else {
PUSH_READ(BPF_REG_X);
PUSH_WRITE(BPF_REG_A);
}
break;
}
}
/*
* 1. Convert opcode(id) to BPF_INS_*
* 2. Set regs_read/regs_write/groups
*/
void BPF_get_insn_id(cs_struct *ud, cs_insn *insn, unsigned int opcode)
{
// No need to care the mode (cBPF or eBPF) since all checks has be done in
// BPF_getInstruction, we can simply map opcode to BPF_INS_*.
cs_detail *detail;
bpf_insn id = BPF_INS_INVALID;
bpf_insn_group grp;
detail = insn->detail;
#ifndef CAPSTONE_DIET
#define PUSH_GROUP(grp) do { \
if (detail) { \
detail->groups[detail->groups_count] = grp; \
detail->groups_count++; \
} \
} while(0)
#else
#define PUSH_GROUP
#endif
switch (BPF_CLASS(opcode)) {
default: // will never happen
break;
case BPF_CLASS_LD:
case BPF_CLASS_LDX:
id = op2insn_ld(opcode);
PUSH_GROUP(BPF_GRP_LOAD);
break;
case BPF_CLASS_ST:
case BPF_CLASS_STX:
id = op2insn_st(opcode);
PUSH_GROUP(BPF_GRP_STORE);
break;
case BPF_CLASS_ALU:
id = op2insn_alu(opcode);
PUSH_GROUP(BPF_GRP_ALU);
break;
case BPF_CLASS_JMP:
grp = BPF_GRP_JUMP;
id = op2insn_jmp(opcode);
if (id == BPF_INS_CALL)
grp = BPF_GRP_CALL;
else if (id == BPF_INS_EXIT)
grp = BPF_GRP_RETURN;
PUSH_GROUP(grp);
break;
case BPF_CLASS_RET:
id = BPF_INS_RET;
PUSH_GROUP(BPF_GRP_RETURN);
break;
// BPF_CLASS_MISC and BPF_CLASS_ALU64 have exactly same value
case BPF_CLASS_MISC:
/* case BPF_CLASS_ALU64: */
if (EBPF_MODE(ud)) {
// ALU64 in eBPF
id = op2insn_alu(opcode);
PUSH_GROUP(BPF_GRP_ALU);
}
else {
if (BPF_MISCOP(opcode) == BPF_MISCOP_TXA)
id = BPF_INS_TXA;
else
id = BPF_INS_TAX;
PUSH_GROUP(BPF_GRP_MISC);
}
break;
}
insn->id = id;
#undef PUSH_GROUP
#ifndef CAPSTONE_DIET
if (detail) {
update_regs_access(ud, detail, id, opcode);
}
#endif
}
static void
opt_blk(struct block *b, int do_stmts)
{
struct slist *s;
struct edge *p;
int i;
bpf_int32 aval, xval;
#if 0
for (s = b->stmts; s && s->next; s = s->next)
if (BPF_CLASS(s->s.code) == BPF_JMP) {
do_stmts = 0;
break;
}
#endif
/*
* Initialize the atom values.
*/
p = b->in_edges;
if (p == 0) {
/*
* We have no predecessors, so everything is undefined
* upon entry to this block.
*/
memset((char *)b->val, 0, sizeof(b->val));
} else {
/*
* Inherit values from our predecessors.
*
* First, get the values from the predecessor along the
* first edge leading to this node.
*/
memcpy((char *)b->val, (char *)p->pred->val, sizeof(b->val));
/*
* Now look at all the other nodes leading to this node.
* If, for the predecessor along that edge, a register
* has a different value from the one we have (i.e.,
* control paths are merging, and the merging paths
* assign different values to that register), give the
* register the undefined value of 0.
*/
while ((p = p->next) != NULL) {
for (i = 0; i < N_ATOMS; ++i)
if (b->val[i] != p->pred->val[i])
b->val[i] = 0;
}
}
aval = b->val[A_ATOM];
xval = b->val[X_ATOM];
for (s = b->stmts; s; s = s->next)
opt_stmt(&s->s, b->val, do_stmts);
/*
* This is a special case: if we don't use anything from this
* block, and we load the accumulator or index register with a
* value that is already there, or if this block is a return,
* eliminate all the statements.
*
* XXX - what if it does a store?
*
* XXX - why does it matter whether we use anything from this
* block? If the accumulator or index register doesn't change
* its value, isn't that OK even if we use that value?
*
* XXX - if we load the accumulator with a different value,
* and the block ends with a conditional branch, we obviously
* can't eliminate it, as the branch depends on that value.
* For the index register, the conditional branch only depends
* on the index register value if the test is against the index
* register value rather than a constant; if nothing uses the
* value we put into the index register, and we're not testing
* against the index register's value, and there aren't any
* other problems that would keep us from eliminating this
* block, can we eliminate it?
*/
if (do_stmts &&
((b->out_use == 0 && aval != 0 && b->val[A_ATOM] == aval &&
xval != 0 && b->val[X_ATOM] == xval) ||
BPF_CLASS(b->s.code) == BPF_RET)) {
if (b->stmts != 0) {
b->stmts = 0;
done = 0;
}
} else {
opt_peep(b);
opt_deadstores(b);
}
/*
* Set up values for branch optimizer.
*/
if (BPF_SRC(b->s.code) == BPF_K)
b->oval = K(b->s.k);
else
b->oval = b->val[X_ATOM];
b->et.code = b->s.code;
b->ef.code = -b->s.code;
}
