1// SPDX-License-Identifier: GPL-2.0
2/*
3 * The back-end-agnostic part of Just-In-Time compiler for eBPF bytecode.
4 *
5 * Copyright (c) 2024 Synopsys Inc.
6 * Author: Shahab Vahedi <shahab@synopsys.com>
7 */
8#include <linux/bug.h>
9#include "bpf_jit.h"
10
11/*
12 * Check for the return value. A pattern used often in this file.
13 * There must be a "ret" variable of type "int" in the scope.
14 */
15#define CHECK_RET(cmd) \
16 do { \
17 ret = (cmd); \
18 if (ret < 0) \
19 return ret; \
20 } while (0)
21
22#ifdef ARC_BPF_JIT_DEBUG
23/* Dumps bytes in /var/log/messages at KERN_INFO level (4). */
24static void dump_bytes(const u8 *buf, u32 len, const char *header)
25{
26 u8 line[64];
27 size_t i, j;
28
29 pr_info("-----------------[ %s ]-----------------\n", header);
30
31 for (i = 0, j = 0; i < len; i++) {
32 /* Last input byte? */
33 if (i == len - 1) {
34 j += scnprintf(line + j, 64 - j, "0x%02x", buf[i]);
35 pr_info("%s\n", line);
36 break;
37 }
38 /* End of line? */
39 else if (i % 8 == 7) {
40 j += scnprintf(line + j, 64 - j, "0x%02x", buf[i]);
41 pr_info("%s\n", line);
42 j = 0;
43 } else {
44 j += scnprintf(line + j, 64 - j, "0x%02x, ", buf[i]);
45 }
46 }
47}
48#endif /* ARC_BPF_JIT_DEBUG */
49
50/********************* JIT context ***********************/
51
52/*
53 * buf: Translated instructions end up here.
54 * len: The length of whole block in bytes.
55 * index: The offset at which the _next_ instruction may be put.
56 */
57struct jit_buffer {
58 u8 *buf;
59 u32 len;
60 u32 index;
61};
62
63/*
64 * This is a subset of "struct jit_context" that its information is deemed
65 * necessary for the next extra pass to come.
66 *
67 * bpf_header: Needed to finally lock the region.
68 * bpf2insn: Used to find the translation for instructions of interest.
69 *
70 * Things like "jit.buf" and "jit.len" can be retrieved respectively from
71 * "prog->bpf_func" and "prog->jited_len".
72 */
73struct arc_jit_data {
74 struct bpf_binary_header *bpf_header;
75 u32 *bpf2insn;
76};
77
78/*
79 * The JIT pertinent context that is used by different functions.
80 *
81 * prog: The current eBPF program being handled.
82 * orig_prog: The original eBPF program before any possible change.
83 * jit: The JIT buffer and its length.
84 * bpf_header: The JITed program header. "jit.buf" points inside it.
85 * emit: If set, opcodes are written to memory; else, a dry-run.
86 * do_zext: If true, 32-bit sub-regs must be zero extended.
87 * bpf2insn: Maps BPF insn indices to their counterparts in jit.buf.
88 * bpf2insn_valid: Indicates if "bpf2ins" is populated with the mappings.
89 * jit_data: A piece of memory to transfer data to the next pass.
90 * arc_regs_clobbered: Each bit status determines if that arc reg is clobbered.
91 * save_blink: Whether ARC's "blink" register needs to be saved.
92 * frame_size: Derived from "prog->aux->stack_depth".
93 * epilogue_offset: Used by early "return"s in the code to jump here.
94 * need_extra_pass: A forecast if an "extra_pass" will occur.
95 * is_extra_pass: Indicates if the current pass is an extra pass.
96 * user_bpf_prog: True, if VM opcodes come from a real program.
97 * blinded: True if "constant blinding" step returned a new "prog".
98 * success: Indicates if the whole JIT went OK.
99 */
100struct jit_context {
101 struct bpf_prog *prog;
102 struct bpf_prog *orig_prog;
103 struct jit_buffer jit;
104 struct bpf_binary_header *bpf_header;
105 bool emit;
106 bool do_zext;
107 u32 *bpf2insn;
108 bool bpf2insn_valid;
109 struct arc_jit_data *jit_data;
110 u32 arc_regs_clobbered;
111 bool save_blink;
112 u16 frame_size;
113 u32 epilogue_offset;
114 bool need_extra_pass;
115 bool is_extra_pass;
116 bool user_bpf_prog;
117 bool blinded;
118 bool success;
119};
120
121/*
122 * If we're in ARC_BPF_JIT_DEBUG mode and the debug level is right, dump the
123 * input BPF stream. "bpf_jit_dump()" is not fully suited for this purpose.
124 */
125static void vm_dump(const struct bpf_prog *prog)
126{
127#ifdef ARC_BPF_JIT_DEBUG
128 if (bpf_jit_enable > 1)
129 dump_bytes((u8 *)prog->insns, 8 * prog->len, " VM ");
130#endif
131}
132
133/*
134 * If the right level of debug is set, dump the bytes. There are 2 variants
135 * of this function:
136 *
137 * 1. Use the standard bpf_jit_dump() which is meant only for JITed code.
138 * 2. Use the dump_bytes() to match its "vm_dump()" instance.
139 */
140static void jit_dump(const struct jit_context *ctx)
141{
142#ifdef ARC_BPF_JIT_DEBUG
143 u8 header[8];
144#endif
145 const int pass = ctx->is_extra_pass ? 2 : 1;
146
147 if (bpf_jit_enable <= 1 || !ctx->prog->jited)
148 return;
149
150#ifdef ARC_BPF_JIT_DEBUG
151 scnprintf(header, sizeof(header), "JIT:%d", pass);
152 dump_bytes(ctx->jit.buf, ctx->jit.len, header);
153 pr_info("\n");
154#else
155 bpf_jit_dump(flen: ctx->prog->len, proglen: ctx->jit.len, pass, image: ctx->jit.buf);
156#endif
157}
158
159/* Initialise the context so there's no garbage. */
160static int jit_ctx_init(struct jit_context *ctx, struct bpf_prog *prog)
161{
162 memset(ctx, 0, sizeof(*ctx));
163
164 ctx->orig_prog = prog;
165
166 /* If constant blinding was requested but failed, scram. */
167 ctx->prog = bpf_jit_blind_constants(fp: prog);
168 if (IS_ERR(ptr: ctx->prog))
169 return PTR_ERR(ptr: ctx->prog);
170 ctx->blinded = (ctx->prog != ctx->orig_prog);
171
172 /* If the verifier doesn't zero-extend, then we have to do it. */
173 ctx->do_zext = !ctx->prog->aux->verifier_zext;
174
175 ctx->is_extra_pass = ctx->prog->jited;
176 ctx->user_bpf_prog = ctx->prog->is_func;
177
178 return 0;
179}
180
181/*
182 * Only after the first iteration of normal pass (the dry-run),
183 * there are valid offsets in ctx->bpf2insn array.
184 */
185static inline bool offsets_available(const struct jit_context *ctx)
186{
187 return ctx->bpf2insn_valid;
188}
189
190/*
191 * "*mem" should be freed when there is no "extra pass" to come,
192 * or the compilation terminated abruptly. A few of such memory
193 * allocations are: ctx->jit_data and ctx->bpf2insn.
194 */
195static inline void maybe_free(struct jit_context *ctx, void **mem)
196{
197 if (*mem) {
198 if (!ctx->success || !ctx->need_extra_pass) {
199 kfree(objp: *mem);
200 *mem = NULL;
201 }
202 }
203}
204
205/*
206 * Free memories based on the status of the context.
207 *
208 * A note about "bpf_header": On successful runs, "bpf_header" is
209 * not freed, because "jit.buf", a sub-array of it, is returned as
210 * the "bpf_func". However, "bpf_header" is lost and nothing points
211 * to it. This should not cause a leakage, because apparently
212 * "bpf_header" can be revived by "bpf_jit_binary_hdr()". This is
213 * how "bpf_jit_free()" in "kernel/bpf/core.c" releases the memory.
214 */
215static void jit_ctx_cleanup(struct jit_context *ctx)
216{
217 if (ctx->blinded) {
218 /* if all went well, release the orig_prog. */
219 if (ctx->success)
220 bpf_jit_prog_release_other(fp: ctx->prog, fp_other: ctx->orig_prog);
221 else
222 bpf_jit_prog_release_other(fp: ctx->orig_prog, fp_other: ctx->prog);
223 }
224
225 maybe_free(ctx, mem: (void **)&ctx->bpf2insn);
226 maybe_free(ctx, mem: (void **)&ctx->jit_data);
227
228 if (!ctx->bpf2insn)
229 ctx->bpf2insn_valid = false;
230
231 /* Freeing "bpf_header" is enough. "jit.buf" is a sub-array of it. */
232 if (!ctx->success && ctx->bpf_header) {
233 bpf_jit_binary_free(hdr: ctx->bpf_header);
234 ctx->bpf_header = NULL;
235 ctx->jit.buf = NULL;
236 ctx->jit.index = 0;
237 ctx->jit.len = 0;
238 }
239
240 ctx->emit = false;
241 ctx->do_zext = false;
242}
243
244/*
245 * Analyse the register usage and record the frame size.
246 * The register usage is determined by consulting the back-end.
247 */
248static void analyze_reg_usage(struct jit_context *ctx)
249{
250 size_t i;
251 u32 usage = 0;
252 const struct bpf_insn *insn = ctx->prog->insnsi;
253
254 for (i = 0; i < ctx->prog->len; i++) {
255 u8 bpf_reg;
256 bool call;
257
258 bpf_reg = insn[i].dst_reg;
259 call = (insn[i].code == (BPF_JMP | BPF_CALL)) ? true : false;
260 usage |= mask_for_used_regs(bpf_reg, is_call: call);
261 }
262
263 ctx->arc_regs_clobbered = usage;
264 ctx->frame_size = ctx->prog->aux->stack_depth;
265}
266
267/* Verify that no instruction will be emitted when there is no buffer. */
268static inline int jit_buffer_check(const struct jit_context *ctx)
269{
270 if (ctx->emit) {
271 if (!ctx->jit.buf) {
272 pr_err("bpf-jit: inconsistence state; no "
273 "buffer to emit instructions.\n");
274 return -EINVAL;
275 } else if (ctx->jit.index > ctx->jit.len) {
276 pr_err("bpf-jit: estimated JIT length is less "
277 "than the emitted instructions.\n");
278 return -EFAULT;
279 }
280 }
281 return 0;
282}
283
284/* On a dry-run (emit=false), "jit.len" is growing gradually. */
285static inline void jit_buffer_update(struct jit_context *ctx, u32 n)
286{
287 if (!ctx->emit)
288 ctx->jit.len += n;
289 else
290 ctx->jit.index += n;
291}
292
293/* Based on "emit", determine the address where instructions are emitted. */
294static inline u8 *effective_jit_buf(const struct jit_context *ctx)
295{
296 return ctx->emit ? (ctx->jit.buf + ctx->jit.index) : NULL;
297}
298
299/* Prologue based on context variables set by "analyze_reg_usage()". */
300static int handle_prologue(struct jit_context *ctx)
301{
302 int ret;
303 u8 *buf = effective_jit_buf(ctx);
304 u32 len = 0;
305
306 CHECK_RET(jit_buffer_check(ctx));
307
308 len = arc_prologue(buf, usage: ctx->arc_regs_clobbered, frame_size: ctx->frame_size);
309 jit_buffer_update(ctx, n: len);
310
311 return 0;
312}
313
314/* The counter part for "handle_prologue()". */
315static int handle_epilogue(struct jit_context *ctx)
316{
317 int ret;
318 u8 *buf = effective_jit_buf(ctx);
319 u32 len = 0;
320
321 CHECK_RET(jit_buffer_check(ctx));
322
323 len = arc_epilogue(buf, usage: ctx->arc_regs_clobbered, frame_size: ctx->frame_size);
324 jit_buffer_update(ctx, n: len);
325
326 return 0;
327}
328
329/* Tell which number of the BPF instruction we are dealing with. */
330static inline s32 get_index_for_insn(const struct jit_context *ctx,
331 const struct bpf_insn *insn)
332{
333 return (insn - ctx->prog->insnsi);
334}
335
336/*
337 * In most of the cases, the "offset" is read from "insn->off". However,
338 * if it is an unconditional BPF_JMP32, then it comes from "insn->imm".
339 *
340 * (Courtesy of "cpu=v4" support)
341 */
342static inline s32 get_offset(const struct bpf_insn *insn)
343{
344 if ((BPF_CLASS(insn->code) == BPF_JMP32) &&
345 (BPF_OP(insn->code) == BPF_JA))
346 return insn->imm;
347 else
348 return insn->off;
349}
350
351/*
352 * Determine to which number of the BPF instruction we're jumping to.
353 *
354 * The "offset" is interpreted as the "number" of BPF instructions
355 * from the _next_ BPF instruction. e.g.:
356 *
357 * 4 means 4 instructions after the next insn
358 * 0 means 0 instructions after the next insn -> fallthrough.
359 * -1 means 1 instruction before the next insn -> jmp to current insn.
360 *
361 * Another way to look at this, "offset" is the number of instructions
362 * that exist between the current instruction and the target instruction.
363 *
364 * It is worth noting that a "mov r,i64", which is 16-byte long, is
365 * treated as two instructions long, therefore "offset" needn't be
366 * treated specially for those. Everything is uniform.
367 */
368static inline s32 get_target_index_for_insn(const struct jit_context *ctx,
369 const struct bpf_insn *insn)
370{
371 return (get_index_for_insn(ctx, insn) + 1) + get_offset(insn);
372}
373
374/* Is there an immediate operand encoded in the "insn"? */
375static inline bool has_imm(const struct bpf_insn *insn)
376{
377 return BPF_SRC(insn->code) == BPF_K;
378}
379
380/* Is the last BPF instruction? */
381static inline bool is_last_insn(const struct bpf_prog *prog, u32 idx)
382{
383 return idx == (prog->len - 1);
384}
385
386/*
387 * Invocation of this function, conditionally signals the need for
388 * an extra pass. The conditions that must be met are:
389 *
390 * 1. The current pass itself shouldn't be an extra pass.
391 * 2. The stream of bytes being JITed must come from a user program.
392 */
393static inline void set_need_for_extra_pass(struct jit_context *ctx)
394{
395 if (!ctx->is_extra_pass)
396 ctx->need_extra_pass = ctx->user_bpf_prog;
397}
398
399/*
400 * Check if the "size" is valid and then transfer the control to
401 * the back-end for the swap.
402 */
403static int handle_swap(u8 *buf, u8 rd, u8 size, u8 endian,
404 bool force, bool do_zext, u8 *len)
405{
406 /* Sanity check on the size. */
407 switch (size) {
408 case 16:
409 case 32:
410 case 64:
411 break;
412 default:
413 pr_err("bpf-jit: invalid size for swap.\n");
414 return -EINVAL;
415 }
416
417 *len = gen_swap(buf, rd, size, endian, force, do_zext);
418
419 return 0;
420}
421
422/* Checks if the (instruction) index is in valid range. */
423static inline bool check_insn_idx_valid(const struct jit_context *ctx,
424 const s32 idx)
425{
426 return (idx >= 0 && idx < ctx->prog->len);
427}
428
429/*
430 * Decouple the back-end from BPF by converting BPF conditions
431 * to internal enum. ARC_CC_* start from 0 and are used as index
432 * to an array. BPF_J* usage must end after this conversion.
433 */
434static int bpf_cond_to_arc(const u8 op, u8 *arc_cc)
435{
436 switch (op) {
437 case BPF_JA:
438 *arc_cc = ARC_CC_AL;
439 break;
440 case BPF_JEQ:
441 *arc_cc = ARC_CC_EQ;
442 break;
443 case BPF_JGT:
444 *arc_cc = ARC_CC_UGT;
445 break;
446 case BPF_JGE:
447 *arc_cc = ARC_CC_UGE;
448 break;
449 case BPF_JSET:
450 *arc_cc = ARC_CC_SET;
451 break;
452 case BPF_JNE:
453 *arc_cc = ARC_CC_NE;
454 break;
455 case BPF_JSGT:
456 *arc_cc = ARC_CC_SGT;
457 break;
458 case BPF_JSGE:
459 *arc_cc = ARC_CC_SGE;
460 break;
461 case BPF_JLT:
462 *arc_cc = ARC_CC_ULT;
463 break;
464 case BPF_JLE:
465 *arc_cc = ARC_CC_ULE;
466 break;
467 case BPF_JSLT:
468 *arc_cc = ARC_CC_SLT;
469 break;
470 case BPF_JSLE:
471 *arc_cc = ARC_CC_SLE;
472 break;
473 default:
474 pr_err("bpf-jit: can't handle condition 0x%02X\n", op);
475 return -EINVAL;
476 }
477 return 0;
478}
479
480/*
481 * Check a few things for a supposedly "jump" instruction:
482 *
483 * 0. "insn" is a "jump" instruction, but not the "call/exit" variant.
484 * 1. The current "insn" index is in valid range.
485 * 2. The index of target instruction is in valid range.
486 */
487static int check_bpf_jump(const struct jit_context *ctx,
488 const struct bpf_insn *insn)
489{
490 const u8 class = BPF_CLASS(insn->code);
491 const u8 op = BPF_OP(insn->code);
492
493 /* Must be a jmp(32) instruction that is not a "call/exit". */
494 if ((class != BPF_JMP && class != BPF_JMP32) ||
495 (op == BPF_CALL || op == BPF_EXIT)) {
496 pr_err("bpf-jit: not a jump instruction.\n");
497 return -EINVAL;
498 }
499
500 if (!check_insn_idx_valid(ctx, idx: get_index_for_insn(ctx, insn))) {
501 pr_err("bpf-jit: the bpf jump insn is not in prog.\n");
502 return -EINVAL;
503 }
504
505 if (!check_insn_idx_valid(ctx, idx: get_target_index_for_insn(ctx, insn))) {
506 pr_err("bpf-jit: bpf jump label is out of range.\n");
507 return -EINVAL;
508 }
509
510 return 0;
511}
512
513/*
514 * Based on input "insn", consult "ctx->bpf2insn" to get the
515 * related index (offset) of the translation in JIT stream.
516 */
517static u32 get_curr_jit_off(const struct jit_context *ctx,
518 const struct bpf_insn *insn)
519{
520 const s32 idx = get_index_for_insn(ctx, insn);
521#ifdef ARC_BPF_JIT_DEBUG
522 BUG_ON(!offsets_available(ctx) || !check_insn_idx_valid(ctx, idx));
523#endif
524 return ctx->bpf2insn[idx];
525}
526
527/*
528 * The input "insn" must be a jump instruction.
529 *
530 * Based on input "insn", consult "ctx->bpf2insn" to get the
531 * related JIT index (offset) of "target instruction" that
532 * "insn" would jump to.
533 */
534static u32 get_targ_jit_off(const struct jit_context *ctx,
535 const struct bpf_insn *insn)
536{
537 const s32 tidx = get_target_index_for_insn(ctx, insn);
538#ifdef ARC_BPF_JIT_DEBUG
539 BUG_ON(!offsets_available(ctx) || !check_insn_idx_valid(ctx, tidx));
540#endif
541 return ctx->bpf2insn[tidx];
542}
543
544/*
545 * This function will return 0 for a feasible jump.
546 *
547 * Consult the back-end to check if it finds it feasible to emit
548 * the necessary instructions based on "cond" and the displacement
549 * between the "from_off" and the "to_off".
550 */
551static int feasible_jit_jump(u32 from_off, u32 to_off, u8 cond, bool j32)
552{
553 int ret = 0;
554
555 if (j32) {
556 if (!check_jmp_32(curr_off: from_off, targ_off: to_off, cond))
557 ret = -EFAULT;
558 } else {
559 if (!check_jmp_64(curr_off: from_off, targ_off: to_off, cond))
560 ret = -EFAULT;
561 }
562
563 if (ret != 0)
564 pr_err("bpf-jit: the JIT displacement is not OK.\n");
565
566 return ret;
567}
568
569/*
570 * This jump handler performs the following steps:
571 *
572 * 1. Compute ARC's internal condition code from BPF's
573 * 2. Determine the bitness of the operation (32 vs. 64)
574 * 3. Sanity check on BPF stream
575 * 4. Sanity check on what is supposed to be JIT's displacement
576 * 5. And finally, emit the necessary instructions
577 *
578 * The last two steps are performed through the back-end.
579 * The value of steps 1 and 2 are necessary inputs for the back-end.
580 */
581static int handle_jumps(const struct jit_context *ctx,
582 const struct bpf_insn *insn,
583 u8 *len)
584{
585 u8 cond;
586 int ret = 0;
587 u8 *buf = effective_jit_buf(ctx);
588 const bool j32 = (BPF_CLASS(insn->code) == BPF_JMP32) ? true : false;
589 const u8 rd = insn->dst_reg;
590 u8 rs = insn->src_reg;
591 u32 curr_off = 0, targ_off = 0;
592
593 *len = 0;
594
595 /* Map the BPF condition to internal enum. */
596 CHECK_RET(bpf_cond_to_arc(BPF_OP(insn->code), &cond));
597
598 /* Sanity check on the BPF byte stream. */
599 CHECK_RET(check_bpf_jump(ctx, insn));
600
601 /*
602 * Move the immediate into a temporary register _now_ for 2 reasons:
603 *
604 * 1. "gen_jmp_{32,64}()" deal with operands in registers.
605 *
606 * 2. The "len" parameter will grow so that the current jit offset
607 * (curr_off) will have increased to a point where the necessary
608 * instructions can be inserted by "gen_jmp_{32,64}()".
609 */
610 if (has_imm(insn) && cond != ARC_CC_AL) {
611 if (j32) {
612 *len += mov_r32_i32(BUF(buf, *len), JIT_REG_TMP,
613 imm: insn->imm);
614 } else {
615 *len += mov_r64_i32(BUF(buf, *len), JIT_REG_TMP,
616 imm: insn->imm);
617 }
618 rs = JIT_REG_TMP;
619 }
620
621 /* If the offsets are known, check if the branch can occur. */
622 if (offsets_available(ctx)) {
623 curr_off = get_curr_jit_off(ctx, insn) + *len;
624 targ_off = get_targ_jit_off(ctx, insn);
625
626 /* Sanity check on the back-end side. */
627 CHECK_RET(feasible_jit_jump(curr_off, targ_off, cond, j32));
628 }
629
630 if (j32) {
631 *len += gen_jmp_32(BUF(buf, *len), rd, rs, cond,
632 c_off: curr_off, t_off: targ_off);
633 } else {
634 *len += gen_jmp_64(BUF(buf, *len), rd, rs, cond,
635 c_off: curr_off, t_off: targ_off);
636 }
637
638 return ret;
639}
640
641/* Jump to translated epilogue address. */
642static int handle_jmp_epilogue(struct jit_context *ctx,
643 const struct bpf_insn *insn, u8 *len)
644{
645 u8 *buf = effective_jit_buf(ctx);
646 u32 curr_off = 0, epi_off = 0;
647
648 /* Check the offset only if the data is available. */
649 if (offsets_available(ctx)) {
650 curr_off = get_curr_jit_off(ctx, insn);
651 epi_off = ctx->epilogue_offset;
652
653 if (!check_jmp_64(curr_off, targ_off: epi_off, cond: ARC_CC_AL)) {
654 pr_err("bpf-jit: epilogue offset is not valid.\n");
655 return -EINVAL;
656 }
657 }
658
659 /* Jump to "epilogue offset" (rd and rs don't matter). */
660 *len = gen_jmp_64(buf, rd: 0, rs: 0, cond: ARC_CC_AL, c_off: curr_off, t_off: epi_off);
661
662 return 0;
663}
664
665/* Try to get the resolved address and generate the instructions. */
666static int handle_call(struct jit_context *ctx,
667 const struct bpf_insn *insn,
668 u8 *len)
669{
670 int ret;
671 bool in_kernel_func, fixed = false;
672 u64 addr = 0;
673 u8 *buf = effective_jit_buf(ctx);
674
675 ret = bpf_jit_get_func_addr(prog: ctx->prog, insn, extra_pass: ctx->is_extra_pass,
676 func_addr: &addr, func_addr_fixed: &fixed);
677 if (ret < 0) {
678 pr_err("bpf-jit: can't get the address for call.\n");
679 return ret;
680 }
681 in_kernel_func = (fixed ? true : false);
682
683 /* No valuable address retrieved (yet). */
684 if (!fixed && !addr)
685 set_need_for_extra_pass(ctx);
686
687 *len = gen_func_call(buf, (ARC_ADDR)addr, in_kernel_func);
688
689 if (insn->src_reg != BPF_PSEUDO_CALL) {
690 /* Assigning ABI's return reg to JIT's return reg. */
691 *len += arc_to_bpf_return(BUF(buf, *len));
692 }
693
694 return 0;
695}
696
697/*
698 * Try to generate instructions for loading a 64-bit immediate.
699 * These sort of instructions are usually associated with the 64-bit
700 * relocations: R_BPF_64_64. Therefore, signal the need for an extra
701 * pass if the circumstances are right.
702 */
703static int handle_ld_imm64(struct jit_context *ctx,
704 const struct bpf_insn *insn,
705 u8 *len)
706{
707 const s32 idx = get_index_for_insn(ctx, insn);
708 u8 *buf = effective_jit_buf(ctx);
709
710 /* We're about to consume 2 VM instructions. */
711 if (is_last_insn(prog: ctx->prog, idx)) {
712 pr_err("bpf-jit: need more data for 64-bit immediate.\n");
713 return -EINVAL;
714 }
715
716 *len = mov_r64_i64(buf, reg: insn->dst_reg, lo: insn->imm, hi: (insn + 1)->imm);
717
718 if (bpf_pseudo_func(insn))
719 set_need_for_extra_pass(ctx);
720
721 return 0;
722}
723
724/*
725 * Handles one eBPF instruction at a time. To make this function faster,
726 * it does not call "jit_buffer_check()". Else, it would call it for every
727 * instruction. As a result, it should not be invoked directly. Only
728 * "handle_body()", that has already executed the "check", may call this
729 * function.
730 *
731 * If the "ret" value is negative, something has went wrong. Else,
732 * it mostly holds the value 0 and rarely 1. Number 1 signals
733 * the loop in "handle_body()" to skip the next instruction, because
734 * it has been consumed as part of a 64-bit immediate value.
735 */
736static int handle_insn(struct jit_context *ctx, u32 idx)
737{
738 const struct bpf_insn *insn = &ctx->prog->insnsi[idx];
739 const u8 code = insn->code;
740 const u8 dst = insn->dst_reg;
741 const u8 src = insn->src_reg;
742 const s16 off = insn->off;
743 const s32 imm = insn->imm;
744 u8 *buf = effective_jit_buf(ctx);
745 u8 len = 0;
746 int ret = 0;
747
748 switch (code) {
749 /* dst += src (32-bit) */
750 case BPF_ALU | BPF_ADD | BPF_X:
751 len = add_r32(buf, rd: dst, rs: src);
752 break;
753 /* dst += imm (32-bit) */
754 case BPF_ALU | BPF_ADD | BPF_K:
755 len = add_r32_i32(buf, rd: dst, imm);
756 break;
757 /* dst -= src (32-bit) */
758 case BPF_ALU | BPF_SUB | BPF_X:
759 len = sub_r32(buf, rd: dst, rs: src);
760 break;
761 /* dst -= imm (32-bit) */
762 case BPF_ALU | BPF_SUB | BPF_K:
763 len = sub_r32_i32(buf, rd: dst, imm);
764 break;
765 /* dst = -dst (32-bit) */
766 case BPF_ALU | BPF_NEG:
767 len = neg_r32(buf, r: dst);
768 break;
769 /* dst *= src (32-bit) */
770 case BPF_ALU | BPF_MUL | BPF_X:
771 len = mul_r32(buf, rd: dst, rs: src);
772 break;
773 /* dst *= imm (32-bit) */
774 case BPF_ALU | BPF_MUL | BPF_K:
775 len = mul_r32_i32(buf, rd: dst, imm);
776 break;
777 /* dst /= src (32-bit) */
778 case BPF_ALU | BPF_DIV | BPF_X:
779 len = div_r32(buf, rd: dst, rs: src, sign_ext: off == 1);
780 break;
781 /* dst /= imm (32-bit) */
782 case BPF_ALU | BPF_DIV | BPF_K:
783 len = div_r32_i32(buf, rd: dst, imm, sign_ext: off == 1);
784 break;
785 /* dst %= src (32-bit) */
786 case BPF_ALU | BPF_MOD | BPF_X:
787 len = mod_r32(buf, rd: dst, rs: src, sign_ext: off == 1);
788 break;
789 /* dst %= imm (32-bit) */
790 case BPF_ALU | BPF_MOD | BPF_K:
791 len = mod_r32_i32(buf, rd: dst, imm, sign_ext: off == 1);
792 break;
793 /* dst &= src (32-bit) */
794 case BPF_ALU | BPF_AND | BPF_X:
795 len = and_r32(buf, rd: dst, rs: src);
796 break;
797 /* dst &= imm (32-bit) */
798 case BPF_ALU | BPF_AND | BPF_K:
799 len = and_r32_i32(buf, rd: dst, imm);
800 break;
801 /* dst |= src (32-bit) */
802 case BPF_ALU | BPF_OR | BPF_X:
803 len = or_r32(buf, rd: dst, rs: src);
804 break;
805 /* dst |= imm (32-bit) */
806 case BPF_ALU | BPF_OR | BPF_K:
807 len = or_r32_i32(buf, rd: dst, imm);
808 break;
809 /* dst ^= src (32-bit) */
810 case BPF_ALU | BPF_XOR | BPF_X:
811 len = xor_r32(buf, rd: dst, rs: src);
812 break;
813 /* dst ^= imm (32-bit) */
814 case BPF_ALU | BPF_XOR | BPF_K:
815 len = xor_r32_i32(buf, rd: dst, imm);
816 break;
817 /* dst <<= src (32-bit) */
818 case BPF_ALU | BPF_LSH | BPF_X:
819 len = lsh_r32(buf, rd: dst, rs: src);
820 break;
821 /* dst <<= imm (32-bit) */
822 case BPF_ALU | BPF_LSH | BPF_K:
823 len = lsh_r32_i32(buf, rd: dst, imm);
824 break;
825 /* dst >>= src (32-bit) [unsigned] */
826 case BPF_ALU | BPF_RSH | BPF_X:
827 len = rsh_r32(buf, rd: dst, rs: src);
828 break;
829 /* dst >>= imm (32-bit) [unsigned] */
830 case BPF_ALU | BPF_RSH | BPF_K:
831 len = rsh_r32_i32(buf, rd: dst, imm);
832 break;
833 /* dst >>= src (32-bit) [signed] */
834 case BPF_ALU | BPF_ARSH | BPF_X:
835 len = arsh_r32(buf, rd: dst, rs: src);
836 break;
837 /* dst >>= imm (32-bit) [signed] */
838 case BPF_ALU | BPF_ARSH | BPF_K:
839 len = arsh_r32_i32(buf, rd: dst, imm);
840 break;
841 /* dst = src (32-bit) */
842 case BPF_ALU | BPF_MOV | BPF_X:
843 len = mov_r32(buf, rd: dst, rs: src, sign_ext: (u8)off);
844 break;
845 /* dst = imm32 (32-bit) */
846 case BPF_ALU | BPF_MOV | BPF_K:
847 len = mov_r32_i32(buf, reg: dst, imm);
848 break;
849 /* dst = swap(dst) */
850 case BPF_ALU | BPF_END | BPF_FROM_LE:
851 case BPF_ALU | BPF_END | BPF_FROM_BE:
852 case BPF_ALU64 | BPF_END | BPF_FROM_LE: {
853 CHECK_RET(handle_swap(buf, dst, imm, BPF_SRC(code),
854 BPF_CLASS(code) == BPF_ALU64,
855 ctx->do_zext, &len));
856 break;
857 }
858 /* dst += src (64-bit) */
859 case BPF_ALU64 | BPF_ADD | BPF_X:
860 len = add_r64(buf, rd: dst, rs: src);
861 break;
862 /* dst += imm32 (64-bit) */
863 case BPF_ALU64 | BPF_ADD | BPF_K:
864 len = add_r64_i32(buf, rd: dst, imm);
865 break;
866 /* dst -= src (64-bit) */
867 case BPF_ALU64 | BPF_SUB | BPF_X:
868 len = sub_r64(buf, rd: dst, rs: src);
869 break;
870 /* dst -= imm32 (64-bit) */
871 case BPF_ALU64 | BPF_SUB | BPF_K:
872 len = sub_r64_i32(buf, rd: dst, imm);
873 break;
874 /* dst = -dst (64-bit) */
875 case BPF_ALU64 | BPF_NEG:
876 len = neg_r64(buf, r: dst);
877 break;
878 /* dst *= src (64-bit) */
879 case BPF_ALU64 | BPF_MUL | BPF_X:
880 len = mul_r64(buf, rd: dst, rs: src);
881 break;
882 /* dst *= imm32 (64-bit) */
883 case BPF_ALU64 | BPF_MUL | BPF_K:
884 len = mul_r64_i32(buf, rd: dst, imm);
885 break;
886 /* dst &= src (64-bit) */
887 case BPF_ALU64 | BPF_AND | BPF_X:
888 len = and_r64(buf, rd: dst, rs: src);
889 break;
890 /* dst &= imm32 (64-bit) */
891 case BPF_ALU64 | BPF_AND | BPF_K:
892 len = and_r64_i32(buf, rd: dst, imm);
893 break;
894 /* dst |= src (64-bit) */
895 case BPF_ALU64 | BPF_OR | BPF_X:
896 len = or_r64(buf, rd: dst, rs: src);
897 break;
898 /* dst |= imm32 (64-bit) */
899 case BPF_ALU64 | BPF_OR | BPF_K:
900 len = or_r64_i32(buf, rd: dst, imm);
901 break;
902 /* dst ^= src (64-bit) */
903 case BPF_ALU64 | BPF_XOR | BPF_X:
904 len = xor_r64(buf, rd: dst, rs: src);
905 break;
906 /* dst ^= imm32 (64-bit) */
907 case BPF_ALU64 | BPF_XOR | BPF_K:
908 len = xor_r64_i32(buf, rd: dst, imm);
909 break;
910 /* dst <<= src (64-bit) */
911 case BPF_ALU64 | BPF_LSH | BPF_X:
912 len = lsh_r64(buf, rd: dst, rs: src);
913 break;
914 /* dst <<= imm32 (64-bit) */
915 case BPF_ALU64 | BPF_LSH | BPF_K:
916 len = lsh_r64_i32(buf, rd: dst, imm);
917 break;
918 /* dst >>= src (64-bit) [unsigned] */
919 case BPF_ALU64 | BPF_RSH | BPF_X:
920 len = rsh_r64(buf, rd: dst, rs: src);
921 break;
922 /* dst >>= imm32 (64-bit) [unsigned] */
923 case BPF_ALU64 | BPF_RSH | BPF_K:
924 len = rsh_r64_i32(buf, rd: dst, imm);
925 break;
926 /* dst >>= src (64-bit) [signed] */
927 case BPF_ALU64 | BPF_ARSH | BPF_X:
928 len = arsh_r64(buf, rd: dst, rs: src);
929 break;
930 /* dst >>= imm32 (64-bit) [signed] */
931 case BPF_ALU64 | BPF_ARSH | BPF_K:
932 len = arsh_r64_i32(buf, rd: dst, imm);
933 break;
934 /* dst = src (64-bit) */
935 case BPF_ALU64 | BPF_MOV | BPF_X:
936 len = mov_r64(buf, rd: dst, rs: src, sign_ext: (u8)off);
937 break;
938 /* dst = imm32 (sign extend to 64-bit) */
939 case BPF_ALU64 | BPF_MOV | BPF_K:
940 len = mov_r64_i32(buf, reg: dst, imm);
941 break;
942 /* dst = imm64 */
943 case BPF_LD | BPF_DW | BPF_IMM:
944 CHECK_RET(handle_ld_imm64(ctx, insn, &len));
945 /* Tell the loop to skip the next instruction. */
946 ret = 1;
947 break;
948 /* dst = *(size *)(src + off) */
949 case BPF_LDX | BPF_MEM | BPF_W:
950 case BPF_LDX | BPF_MEM | BPF_H:
951 case BPF_LDX | BPF_MEM | BPF_B:
952 case BPF_LDX | BPF_MEM | BPF_DW:
953 len = load_r(buf, rd: dst, rs: src, off, BPF_SIZE(code), sign_ext: false);
954 break;
955 case BPF_LDX | BPF_MEMSX | BPF_W:
956 case BPF_LDX | BPF_MEMSX | BPF_H:
957 case BPF_LDX | BPF_MEMSX | BPF_B:
958 len = load_r(buf, rd: dst, rs: src, off, BPF_SIZE(code), sign_ext: true);
959 break;
960 /* *(size *)(dst + off) = src */
961 case BPF_STX | BPF_MEM | BPF_W:
962 case BPF_STX | BPF_MEM | BPF_H:
963 case BPF_STX | BPF_MEM | BPF_B:
964 case BPF_STX | BPF_MEM | BPF_DW:
965 len = store_r(buf, rd: src, rs: dst, off, BPF_SIZE(code));
966 break;
967 case BPF_ST | BPF_MEM | BPF_W:
968 case BPF_ST | BPF_MEM | BPF_H:
969 case BPF_ST | BPF_MEM | BPF_B:
970 case BPF_ST | BPF_MEM | BPF_DW:
971 len = store_i(buf, imm, rd: dst, off, BPF_SIZE(code));
972 break;
973 case BPF_JMP | BPF_JA:
974 case BPF_JMP | BPF_JEQ | BPF_X:
975 case BPF_JMP | BPF_JEQ | BPF_K:
976 case BPF_JMP | BPF_JNE | BPF_X:
977 case BPF_JMP | BPF_JNE | BPF_K:
978 case BPF_JMP | BPF_JSET | BPF_X:
979 case BPF_JMP | BPF_JSET | BPF_K:
980 case BPF_JMP | BPF_JGT | BPF_X:
981 case BPF_JMP | BPF_JGT | BPF_K:
982 case BPF_JMP | BPF_JGE | BPF_X:
983 case BPF_JMP | BPF_JGE | BPF_K:
984 case BPF_JMP | BPF_JSGT | BPF_X:
985 case BPF_JMP | BPF_JSGT | BPF_K:
986 case BPF_JMP | BPF_JSGE | BPF_X:
987 case BPF_JMP | BPF_JSGE | BPF_K:
988 case BPF_JMP | BPF_JLT | BPF_X:
989 case BPF_JMP | BPF_JLT | BPF_K:
990 case BPF_JMP | BPF_JLE | BPF_X:
991 case BPF_JMP | BPF_JLE | BPF_K:
992 case BPF_JMP | BPF_JSLT | BPF_X:
993 case BPF_JMP | BPF_JSLT | BPF_K:
994 case BPF_JMP | BPF_JSLE | BPF_X:
995 case BPF_JMP | BPF_JSLE | BPF_K:
996 case BPF_JMP32 | BPF_JA:
997 case BPF_JMP32 | BPF_JEQ | BPF_X:
998 case BPF_JMP32 | BPF_JEQ | BPF_K:
999 case BPF_JMP32 | BPF_JNE | BPF_X:
1000 case BPF_JMP32 | BPF_JNE | BPF_K:
1001 case BPF_JMP32 | BPF_JSET | BPF_X:
1002 case BPF_JMP32 | BPF_JSET | BPF_K:
1003 case BPF_JMP32 | BPF_JGT | BPF_X:
1004 case BPF_JMP32 | BPF_JGT | BPF_K:
1005 case BPF_JMP32 | BPF_JGE | BPF_X:
1006 case BPF_JMP32 | BPF_JGE | BPF_K:
1007 case BPF_JMP32 | BPF_JSGT | BPF_X:
1008 case BPF_JMP32 | BPF_JSGT | BPF_K:
1009 case BPF_JMP32 | BPF_JSGE | BPF_X:
1010 case BPF_JMP32 | BPF_JSGE | BPF_K:
1011 case BPF_JMP32 | BPF_JLT | BPF_X:
1012 case BPF_JMP32 | BPF_JLT | BPF_K:
1013 case BPF_JMP32 | BPF_JLE | BPF_X:
1014 case BPF_JMP32 | BPF_JLE | BPF_K:
1015 case BPF_JMP32 | BPF_JSLT | BPF_X:
1016 case BPF_JMP32 | BPF_JSLT | BPF_K:
1017 case BPF_JMP32 | BPF_JSLE | BPF_X:
1018 case BPF_JMP32 | BPF_JSLE | BPF_K:
1019 CHECK_RET(handle_jumps(ctx, insn, &len));
1020 break;
1021 case BPF_JMP | BPF_CALL:
1022 CHECK_RET(handle_call(ctx, insn, &len));
1023 break;
1024
1025 case BPF_JMP | BPF_EXIT:
1026 /* If this is the last instruction, epilogue will follow. */
1027 if (is_last_insn(prog: ctx->prog, idx))
1028 break;
1029 CHECK_RET(handle_jmp_epilogue(ctx, insn, &len));
1030 break;
1031 default:
1032 pr_err("bpf-jit: can't handle instruction code 0x%02X\n", code);
1033 return -EOPNOTSUPP;
1034 }
1035
1036 if (BPF_CLASS(code) == BPF_ALU) {
1037 /*
1038 * Skip the "swap" instructions. Even 64-bit swaps are of type
1039 * BPF_ALU (and not BPF_ALU64). Therefore, for the swaps, one
1040 * has to look at the "size" of the operations rather than the
1041 * ALU type. "gen_swap()" specifically takes care of that.
1042 */
1043 if (BPF_OP(code) != BPF_END && ctx->do_zext)
1044 len += zext(BUF(buf, len), rd: dst);
1045 }
1046
1047 jit_buffer_update(ctx, n: len);
1048
1049 return ret;
1050}
1051
1052static int handle_body(struct jit_context *ctx)
1053{
1054 int ret;
1055 bool populate_bpf2insn = false;
1056 const struct bpf_prog *prog = ctx->prog;
1057
1058 CHECK_RET(jit_buffer_check(ctx));
1059
1060 /*
1061 * Record the mapping for the instructions during the dry-run.
1062 * Doing it this way allows us to have the mapping ready for
1063 * the jump instructions during the real compilation phase.
1064 */
1065 if (!ctx->emit)
1066 populate_bpf2insn = true;
1067
1068 for (u32 i = 0; i < prog->len; i++) {
1069 /* During the dry-run, jit.len grows gradually per BPF insn. */
1070 if (populate_bpf2insn)
1071 ctx->bpf2insn[i] = ctx->jit.len;
1072
1073 CHECK_RET(handle_insn(ctx, i));
1074 if (ret > 0) {
1075 /* "ret" is 1 if two (64-bit) chunks were consumed. */
1076 ctx->bpf2insn[i + 1] = ctx->bpf2insn[i];
1077 i++;
1078 }
1079 }
1080
1081 /* If bpf2insn had to be populated, then it is done at this point. */
1082 if (populate_bpf2insn)
1083 ctx->bpf2insn_valid = true;
1084
1085 return 0;
1086}
1087
1088/*
1089 * Initialize the memory with "unimp_s" which is the mnemonic for
1090 * "unimplemented" instruction and always raises an exception.
1091 *
1092 * The instruction is 2 bytes. If "size" is odd, there is not much
1093 * that can be done about the last byte in "area". Because, the
1094 * CPU always fetches instructions in two bytes. Therefore, the
1095 * byte beyond the last one is going to accompany it during a
1096 * possible fetch. In the most likely case of a little endian
1097 * system, that beyond-byte will become the major opcode and
1098 * we have no control over its initialisation.
1099 */
1100static void fill_ill_insn(void *area, unsigned int size)
1101{
1102 const u16 unimp_s = 0x79e0;
1103
1104 if (size & 1) {
1105 *((u8 *)area + (size - 1)) = 0xff;
1106 size -= 1;
1107 }
1108
1109 memset16(s: area, v: unimp_s, n: size >> 1);
1110}
1111
1112/* Piece of memory that can be allocated at the beginning of jit_prepare(). */
1113static int jit_prepare_early_mem_alloc(struct jit_context *ctx)
1114{
1115 ctx->bpf2insn = kcalloc(ctx->prog->len, sizeof(ctx->jit.len),
1116 GFP_KERNEL);
1117
1118 if (!ctx->bpf2insn) {
1119 pr_err("bpf-jit: could not allocate memory for "
1120 "mapping of the instructions.\n");
1121 return -ENOMEM;
1122 }
1123
1124 return 0;
1125}
1126
1127/*
1128 * Memory allocations that rely on parameters known at the end of
1129 * jit_prepare().
1130 */
1131static int jit_prepare_final_mem_alloc(struct jit_context *ctx)
1132{
1133 const size_t alignment = sizeof(u32);
1134
1135 ctx->bpf_header = bpf_jit_binary_alloc(proglen: ctx->jit.len, image_ptr: &ctx->jit.buf,
1136 alignment, bpf_fill_ill_insns: fill_ill_insn);
1137 if (!ctx->bpf_header) {
1138 pr_err("bpf-jit: could not allocate memory for translation.\n");
1139 return -ENOMEM;
1140 }
1141
1142 if (ctx->need_extra_pass) {
1143 ctx->jit_data = kzalloc(sizeof(*ctx->jit_data), GFP_KERNEL);
1144 if (!ctx->jit_data)
1145 return -ENOMEM;
1146 }
1147
1148 return 0;
1149}
1150
1151/*
1152 * The first phase of the translation without actually emitting any
1153 * instruction. It helps in getting a forecast on some aspects, such
1154 * as the length of the whole program or where the epilogue starts.
1155 *
1156 * Whenever the necessary parameters are known, memories are allocated.
1157 */
1158static int jit_prepare(struct jit_context *ctx)
1159{
1160 int ret;
1161
1162 /* Dry run. */
1163 ctx->emit = false;
1164
1165 CHECK_RET(jit_prepare_early_mem_alloc(ctx));
1166
1167 /* Get the length of prologue section after some register analysis. */
1168 analyze_reg_usage(ctx);
1169 CHECK_RET(handle_prologue(ctx));
1170
1171 CHECK_RET(handle_body(ctx));
1172
1173 /* Record at which offset epilogue begins. */
1174 ctx->epilogue_offset = ctx->jit.len;
1175
1176 /* Process the epilogue section now. */
1177 CHECK_RET(handle_epilogue(ctx));
1178
1179 CHECK_RET(jit_prepare_final_mem_alloc(ctx));
1180
1181 return 0;
1182}
1183
1184/*
1185 * jit_compile() is the real compilation phase. jit_prepare() is
1186 * invoked before jit_compile() as a dry-run to make sure everything
1187 * will go OK and allocate the necessary memory.
1188 *
1189 * In the end, jit_compile() checks if it has produced the same number
1190 * of instructions as jit_prepare() would.
1191 */
1192static int jit_compile(struct jit_context *ctx)
1193{
1194 int ret;
1195
1196 /* Let there be code. */
1197 ctx->emit = true;
1198
1199 CHECK_RET(handle_prologue(ctx));
1200
1201 CHECK_RET(handle_body(ctx));
1202
1203 CHECK_RET(handle_epilogue(ctx));
1204
1205 if (ctx->jit.index != ctx->jit.len) {
1206 pr_err("bpf-jit: divergence between the phases; "
1207 "%u vs. %u (bytes).\n",
1208 ctx->jit.len, ctx->jit.index);
1209 return -EFAULT;
1210 }
1211
1212 return 0;
1213}
1214
1215/*
1216 * Calling this function implies a successful JIT. A successful
1217 * translation is signaled by setting the right parameters:
1218 *
1219 * prog->jited=1, prog->jited_len=..., prog->bpf_func=...
1220 */
1221static int jit_finalize(struct jit_context *ctx)
1222{
1223 struct bpf_prog *prog = ctx->prog;
1224
1225 /* We're going to need this information for the "do_extra_pass()". */
1226 if (ctx->need_extra_pass) {
1227 ctx->jit_data->bpf_header = ctx->bpf_header;
1228 ctx->jit_data->bpf2insn = ctx->bpf2insn;
1229 prog->aux->jit_data = (void *)ctx->jit_data;
1230 } else {
1231 /*
1232 * If things seem finalised, then mark the JITed memory
1233 * as R-X and flush it.
1234 */
1235 if (bpf_jit_binary_lock_ro(hdr: ctx->bpf_header)) {
1236 pr_err("bpf-jit: Could not lock the JIT memory.\n");
1237 return -EFAULT;
1238 }
1239 flush_icache_range(start: (unsigned long)ctx->bpf_header,
1240 end: (unsigned long)
1241 BUF(ctx->jit.buf, ctx->jit.len));
1242 prog->aux->jit_data = NULL;
1243 bpf_prog_fill_jited_linfo(prog, insn_to_jit_off: ctx->bpf2insn);
1244 }
1245
1246 ctx->success = true;
1247 prog->bpf_func = (void *)ctx->jit.buf;
1248 prog->jited_len = ctx->jit.len;
1249 prog->jited = 1;
1250
1251 jit_ctx_cleanup(ctx);
1252 jit_dump(ctx);
1253
1254 return 0;
1255}
1256
1257/*
1258 * A lenient verification for the existence of JIT context in "prog".
1259 * Apparently the JIT internals, namely jit_subprogs() in bpf/verifier.c,
1260 * may request for a second compilation although nothing needs to be done.
1261 */
1262static inline int check_jit_context(const struct bpf_prog *prog)
1263{
1264 if (!prog->aux->jit_data) {
1265 pr_notice("bpf-jit: no jit data for the extra pass.\n");
1266 return 1;
1267 } else {
1268 return 0;
1269 }
1270}
1271
1272/* Reuse the previous pass's data. */
1273static int jit_resume_context(struct jit_context *ctx)
1274{
1275 struct arc_jit_data *jdata =
1276 (struct arc_jit_data *)ctx->prog->aux->jit_data;
1277
1278 if (!jdata) {
1279 pr_err("bpf-jit: no jit data for the extra pass.\n");
1280 return -EINVAL;
1281 }
1282
1283 ctx->jit.buf = (u8 *)ctx->prog->bpf_func;
1284 ctx->jit.len = ctx->prog->jited_len;
1285 ctx->bpf_header = jdata->bpf_header;
1286 ctx->bpf2insn = (u32 *)jdata->bpf2insn;
1287 ctx->bpf2insn_valid = ctx->bpf2insn ? true : false;
1288 ctx->jit_data = jdata;
1289
1290 return 0;
1291}
1292
1293/*
1294 * Patch in the new addresses. The instructions of interest are:
1295 *
1296 * - call
1297 * - ld r64, imm64
1298 *
1299 * For "call"s, it resolves the addresses one more time through the
1300 * handle_call().
1301 *
1302 * For 64-bit immediate loads, it just retranslates them, because the BPF
1303 * core in kernel might have changed the value since the normal pass.
1304 */
1305static int jit_patch_relocations(struct jit_context *ctx)
1306{
1307 const u8 bpf_opc_call = BPF_JMP | BPF_CALL;
1308 const u8 bpf_opc_ldi64 = BPF_LD | BPF_DW | BPF_IMM;
1309 const struct bpf_prog *prog = ctx->prog;
1310 int ret;
1311
1312 ctx->emit = true;
1313 for (u32 i = 0; i < prog->len; i++) {
1314 const struct bpf_insn *insn = &prog->insnsi[i];
1315 u8 dummy;
1316 /*
1317 * Adjust "ctx.jit.index", so "gen_*()" functions below
1318 * can use it for their output addresses.
1319 */
1320 ctx->jit.index = ctx->bpf2insn[i];
1321
1322 if (insn->code == bpf_opc_call) {
1323 CHECK_RET(handle_call(ctx, insn, &dummy));
1324 } else if (insn->code == bpf_opc_ldi64) {
1325 CHECK_RET(handle_ld_imm64(ctx, insn, &dummy));
1326 /* Skip the next instruction. */
1327 ++i;
1328 }
1329 }
1330 return 0;
1331}
1332
1333/*
1334 * A normal pass that involves a "dry-run" phase, jit_prepare(),
1335 * to get the necessary data for the real compilation phase,
1336 * jit_compile().
1337 */
1338static struct bpf_prog *do_normal_pass(struct bpf_prog *prog)
1339{
1340 struct jit_context ctx;
1341
1342 /* Bail out if JIT is disabled. */
1343 if (!prog->jit_requested)
1344 return prog;
1345
1346 if (jit_ctx_init(ctx: &ctx, prog)) {
1347 jit_ctx_cleanup(ctx: &ctx);
1348 return prog;
1349 }
1350
1351 /* Get the lengths and allocate buffer. */
1352 if (jit_prepare(ctx: &ctx)) {
1353 jit_ctx_cleanup(ctx: &ctx);
1354 return prog;
1355 }
1356
1357 if (jit_compile(ctx: &ctx)) {
1358 jit_ctx_cleanup(ctx: &ctx);
1359 return prog;
1360 }
1361
1362 if (jit_finalize(ctx: &ctx)) {
1363 jit_ctx_cleanup(ctx: &ctx);
1364 return prog;
1365 }
1366
1367 return ctx.prog;
1368}
1369
1370/*
1371 * If there are multi-function BPF programs that call each other,
1372 * their translated addresses are not known all at once. Therefore,
1373 * an extra pass is needed to consult the bpf_jit_get_func_addr()
1374 * again to get the newly translated addresses in order to resolve
1375 * the "call"s.
1376 */
1377static struct bpf_prog *do_extra_pass(struct bpf_prog *prog)
1378{
1379 struct jit_context ctx;
1380
1381 /* Skip if there's no context to resume from. */
1382 if (check_jit_context(prog))
1383 return prog;
1384
1385 if (jit_ctx_init(ctx: &ctx, prog)) {
1386 jit_ctx_cleanup(ctx: &ctx);
1387 return prog;
1388 }
1389
1390 if (jit_resume_context(ctx: &ctx)) {
1391 jit_ctx_cleanup(ctx: &ctx);
1392 return prog;
1393 }
1394
1395 if (jit_patch_relocations(ctx: &ctx)) {
1396 jit_ctx_cleanup(ctx: &ctx);
1397 return prog;
1398 }
1399
1400 if (jit_finalize(ctx: &ctx)) {
1401 jit_ctx_cleanup(ctx: &ctx);
1402 return prog;
1403 }
1404
1405 return ctx.prog;
1406}
1407
1408/*
1409 * This function may be invoked twice for the same stream of BPF
1410 * instructions. The "extra pass" happens, when there are
1411 * (re)locations involved that their addresses are not known
1412 * during the first run.
1413 */
1414struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1415{
1416 vm_dump(prog);
1417
1418 /* Was this program already translated? */
1419 if (!prog->jited)
1420 return do_normal_pass(prog);
1421 else
1422 return do_extra_pass(prog);
1423
1424 return prog;
1425}
1426

source code of linux/arch/arc/net/bpf_jit_core.c