void bpf_jit_compile(struct bpf_prog *fp)
{
unsigned int proglen;
unsigned int alloclen;
u32 *image = NULL;
u32 *code_base;
unsigned int *addrs;
struct codegen_context cgctx;
int pass;
int flen = fp->len;
if (!bpf_jit_enable)
return;
addrs = kzalloc((flen+1) * sizeof(*addrs), GFP_KERNEL);
if (addrs == NULL)
return;
/*
* There are multiple assembly passes as the generated code will change
* size as it settles down, figuring out the max branch offsets/exit
* paths required.
*
* The range of standard conditional branches is +/- 32Kbytes. Since
* BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
* finish with 8 bytes/instruction. Not feasible, so long jumps are
* used, distinct from short branches.
*
* Current:
*
* For now, both branch types assemble to 2 words (short branches padded
* with a NOP); this is less efficient, but assembly will always complete
* after exactly 3 passes:
*
* First pass: No code buffer; Program is "faux-generated" -- no code
* emitted but maximum size of output determined (and addrs[] filled
* in). Also, we note whether we use M[], whether we use skb data, etc.
* All generation choices assumed to be 'worst-case', e.g. branches all
* far (2 instructions), return path code reduction not available, etc.
*
* Second pass: Code buffer allocated with size determined previously.
* Prologue generated to support features we have seen used. Exit paths
* determined and addrs[] is filled in again, as code may be slightly
* smaller as a result.
*
* Third pass: Code generated 'for real', and branch destinations
* determined from now-accurate addrs[] map.
*
* Ideal:
*
* If we optimise this, near branches will be shorter. On the
* first assembly pass, we should err on the side of caution and
* generate the biggest code. On subsequent passes, branches will be
* generated short or long and code size will reduce. With smaller
* code, more branches may fall into the short category, and code will
* reduce more.
*
* Finally, if we see one pass generate code the same size as the
* previous pass we have converged and should now generate code for
* real. Allocating at the end will also save the memory that would
* otherwise be wasted by the (small) current code shrinkage.
* Preferably, we should do a small number of passes (e.g. 5) and if we
* haven't converged by then, get impatient and force code to generate
* as-is, even if the odd branch would be left long. The chances of a
* long jump are tiny with all but the most enormous of BPF filter
* inputs, so we should usually converge on the third pass.
*/
cgctx.idx = 0;
cgctx.seen = 0;
cgctx.pc_ret0 = -1;
/* Scouting faux-generate pass 0 */
if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
/* We hit something illegal or unsupported. */
goto out;
/*
* Pretend to build prologue, given the features we've seen. This will
* update ctgtx.idx as it pretends to output instructions, then we can
* calculate total size from idx.
*/
bpf_jit_build_prologue(fp, 0, &cgctx);
bpf_jit_build_epilogue(0, &cgctx);
proglen = cgctx.idx * 4;
alloclen = proglen + FUNCTION_DESCR_SIZE;
image = module_alloc(alloclen);
if (!image)
goto out;
code_base = image + (FUNCTION_DESCR_SIZE/4);
/* Code generation passes 1-2 */
for (pass = 1; pass < 3; pass++) {
/* Now build the prologue, body code & epilogue for real. */
cgctx.idx = 0;
bpf_jit_build_prologue(fp, code_base, &cgctx);
bpf_jit_build_body(fp, code_base, &cgctx, addrs);
bpf_jit_build_epilogue(code_base, &cgctx);
if (bpf_jit_enable > 1)
pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
proglen - (cgctx.idx * 4), cgctx.seen);
}
if (bpf_jit_enable > 1)
/* Note that we output the base address of the code_base
* rather than image, since opcodes are in code_base.
*/
bpf_jit_dump(flen, proglen, pass, code_base);
if (image) {
bpf_flush_icache(code_base, code_base + (proglen/4));
/* Function descriptor nastiness: Address + TOC */
((u64 *)image)[0] = (u64)code_base;
((u64 *)image)[1] = local_paca->kernel_toc;
fp->bpf_func = (void *)image;
fp->jited = true;
}
out:
kfree(addrs);
return;
}
struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
{
u32 proglen;
u32 alloclen;
u8 *image = NULL;
u32 *code_base;
u32 *addrs;
struct powerpc64_jit_data *jit_data;
struct codegen_context cgctx;
int pass;
int flen;
struct bpf_binary_header *bpf_hdr;
struct bpf_prog *org_fp = fp;
struct bpf_prog *tmp_fp;
bool bpf_blinded = false;
bool extra_pass = false;
if (!fp->jit_requested)
return org_fp;
tmp_fp = bpf_jit_blind_constants(org_fp);
if (IS_ERR(tmp_fp))
return org_fp;
if (tmp_fp != org_fp) {
bpf_blinded = true;
fp = tmp_fp;
}
jit_data = fp->aux->jit_data;
if (!jit_data) {
jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
if (!jit_data) {
fp = org_fp;
goto out;
}
fp->aux->jit_data = jit_data;
}
flen = fp->len;
addrs = jit_data->addrs;
if (addrs) {
cgctx = jit_data->ctx;
image = jit_data->image;
bpf_hdr = jit_data->header;
proglen = jit_data->proglen;
alloclen = proglen + FUNCTION_DESCR_SIZE;
extra_pass = true;
goto skip_init_ctx;
}
addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
if (addrs == NULL) {
fp = org_fp;
goto out_addrs;
}
memset(&cgctx, 0, sizeof(struct codegen_context));
/* Make sure that the stack is quadword aligned. */
cgctx.stack_size = round_up(fp->aux->stack_depth, 16);
/* Scouting faux-generate pass 0 */
if (bpf_jit_build_body(fp, 0, &cgctx, addrs, false)) {
/* We hit something illegal or unsupported. */
fp = org_fp;
goto out_addrs;
}
/*
* Pretend to build prologue, given the features we've seen. This will
* update ctgtx.idx as it pretends to output instructions, then we can
* calculate total size from idx.
*/
bpf_jit_build_prologue(0, &cgctx);
bpf_jit_build_epilogue(0, &cgctx);
proglen = cgctx.idx * 4;
alloclen = proglen + FUNCTION_DESCR_SIZE;
bpf_hdr = bpf_jit_binary_alloc(alloclen, &image, 4,
bpf_jit_fill_ill_insns);
if (!bpf_hdr) {
fp = org_fp;
goto out_addrs;
}
skip_init_ctx:
code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
if (extra_pass) {
/*
* Do not touch the prologue and epilogue as they will remain
* unchanged. Only fix the branch target address for subprog
* calls in the body.
*
* This does not change the offsets and lengths of the subprog
* call instruction sequences and hence, the size of the JITed
* image as well.
*/
bpf_jit_fixup_subprog_calls(fp, code_base, &cgctx, addrs);
/* There is no need to perform the usual passes. */
goto skip_codegen_passes;
}
/* Code generation passes 1-2 */
for (pass = 1; pass < 3; pass++) {
/* Now build the prologue, body code & epilogue for real. */
cgctx.idx = 0;
bpf_jit_build_prologue(code_base, &cgctx);
bpf_jit_build_body(fp, code_base, &cgctx, addrs, extra_pass);
bpf_jit_build_epilogue(code_base, &cgctx);
if (bpf_jit_enable > 1)
pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
proglen - (cgctx.idx * 4), cgctx.seen);
}
skip_codegen_passes:
if (bpf_jit_enable > 1)
/*
* Note that we output the base address of the code_base
* rather than image, since opcodes are in code_base.
*/
bpf_jit_dump(flen, proglen, pass, code_base);
#ifdef PPC64_ELF_ABI_v1
/* Function descriptor nastiness: Address + TOC */
((u64 *)image)[0] = (u64)code_base;
((u64 *)image)[1] = local_paca->kernel_toc;
#endif
fp->bpf_func = (void *)image;
fp->jited = 1;
fp->jited_len = alloclen;
bpf_flush_icache(bpf_hdr, (u8 *)bpf_hdr + (bpf_hdr->pages * PAGE_SIZE));
if (!fp->is_func || extra_pass) {
out_addrs:
kfree(addrs);
kfree(jit_data);
fp->aux->jit_data = NULL;
} else {
jit_data->addrs = addrs;
jit_data->ctx = cgctx;
jit_data->proglen = proglen;
jit_data->image = image;
jit_data->header = bpf_hdr;
}
out:
if (bpf_blinded)
bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
return fp;
}
