2 * Copyright (C) 2016-2018 Netronome Systems, Inc.
4 * This software is dual licensed under the GNU General License Version 2,
5 * June 1991 as shown in the file COPYING in the top-level directory of this
6 * source tree or the BSD 2-Clause License provided below. You have the
7 * option to license this software under the complete terms of either license.
9 * The BSD 2-Clause License:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * 1. Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * 2. Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #define pr_fmt(fmt) "NFP net bpf: " fmt
36 #include <linux/bug.h>
37 #include <linux/bpf.h>
38 #include <linux/filter.h>
39 #include <linux/kernel.h>
40 #include <linux/pkt_cls.h>
41 #include <linux/reciprocal_div.h>
42 #include <linux/unistd.h>
45 #include "../nfp_asm.h"
46 #include "../nfp_net_ctrl.h"
48 /* --- NFP prog --- */
49 /* Foreach "multiple" entries macros provide pos and next<n> pointers.
50 * It's safe to modify the next pointers (but not pos).
52 #define nfp_for_each_insn_walk2(nfp_prog, pos, next) \
53 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \
54 next = list_next_entry(pos, l); \
55 &(nfp_prog)->insns != &pos->l && \
56 &(nfp_prog)->insns != &next->l; \
57 pos = nfp_meta_next(pos), \
58 next = nfp_meta_next(pos))
60 #define nfp_for_each_insn_walk3(nfp_prog, pos, next, next2) \
61 for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \
62 next = list_next_entry(pos, l), \
63 next2 = list_next_entry(next, l); \
64 &(nfp_prog)->insns != &pos->l && \
65 &(nfp_prog)->insns != &next->l && \
66 &(nfp_prog)->insns != &next2->l; \
67 pos = nfp_meta_next(pos), \
68 next = nfp_meta_next(pos), \
69 next2 = nfp_meta_next(next))
72 nfp_meta_has_prev(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
74 return meta->l.prev != &nfp_prog->insns;
77 static void nfp_prog_push(struct nfp_prog *nfp_prog, u64 insn)
79 if (nfp_prog->__prog_alloc_len / sizeof(u64) == nfp_prog->prog_len) {
80 pr_warn("instruction limit reached (%u NFP instructions)\n",
82 nfp_prog->error = -ENOSPC;
86 nfp_prog->prog[nfp_prog->prog_len] = insn;
90 static unsigned int nfp_prog_current_offset(struct nfp_prog *nfp_prog)
92 return nfp_prog->prog_len;
96 nfp_prog_confirm_current_offset(struct nfp_prog *nfp_prog, unsigned int off)
98 /* If there is a recorded error we may have dropped instructions;
99 * that doesn't have to be due to translator bug, and the translation
100 * will fail anyway, so just return OK.
104 return !WARN_ON_ONCE(nfp_prog_current_offset(nfp_prog) != off);
107 /* --- Emitters --- */
109 __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op,
110 u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, enum cmd_ctx_swap ctx,
115 insn = FIELD_PREP(OP_CMD_A_SRC, areg) |
116 FIELD_PREP(OP_CMD_CTX, ctx) |
117 FIELD_PREP(OP_CMD_B_SRC, breg) |
118 FIELD_PREP(OP_CMD_TOKEN, cmd_tgt_act[op].token) |
119 FIELD_PREP(OP_CMD_XFER, xfer) |
120 FIELD_PREP(OP_CMD_CNT, size) |
121 FIELD_PREP(OP_CMD_SIG, ctx != CMD_CTX_NO_SWAP) |
122 FIELD_PREP(OP_CMD_TGT_CMD, cmd_tgt_act[op].tgt_cmd) |
123 FIELD_PREP(OP_CMD_INDIR, indir) |
124 FIELD_PREP(OP_CMD_MODE, mode);
126 nfp_prog_push(nfp_prog, insn);
130 emit_cmd_any(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer,
131 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx, bool indir)
133 struct nfp_insn_re_regs reg;
136 err = swreg_to_restricted(reg_none(), lreg, rreg, ®, false);
138 nfp_prog->error = err;
142 pr_err("cmd can't swap arguments\n");
143 nfp_prog->error = -EFAULT;
146 if (reg.dst_lmextn || reg.src_lmextn) {
147 pr_err("cmd can't use LMextn\n");
148 nfp_prog->error = -EFAULT;
152 __emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, ctx,
157 emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer,
158 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx)
160 emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, ctx, false);
164 emit_cmd_indir(struct nfp_prog *nfp_prog, enum cmd_tgt_map op, u8 mode, u8 xfer,
165 swreg lreg, swreg rreg, u8 size, enum cmd_ctx_swap ctx)
167 emit_cmd_any(nfp_prog, op, mode, xfer, lreg, rreg, size, ctx, true);
171 __emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, enum br_ev_pip ev_pip,
172 enum br_ctx_signal_state css, u16 addr, u8 defer)
174 u16 addr_lo, addr_hi;
177 addr_lo = addr & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO));
178 addr_hi = addr != addr_lo;
181 FIELD_PREP(OP_BR_MASK, mask) |
182 FIELD_PREP(OP_BR_EV_PIP, ev_pip) |
183 FIELD_PREP(OP_BR_CSS, css) |
184 FIELD_PREP(OP_BR_DEFBR, defer) |
185 FIELD_PREP(OP_BR_ADDR_LO, addr_lo) |
186 FIELD_PREP(OP_BR_ADDR_HI, addr_hi);
188 nfp_prog_push(nfp_prog, insn);
192 emit_br_relo(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer,
193 enum nfp_relo_type relo)
195 if (mask == BR_UNC && defer > 2) {
196 pr_err("BUG: branch defer out of bounds %d\n", defer);
197 nfp_prog->error = -EFAULT;
201 __emit_br(nfp_prog, mask,
202 mask != BR_UNC ? BR_EV_PIP_COND : BR_EV_PIP_UNCOND,
203 BR_CSS_NONE, addr, defer);
205 nfp_prog->prog[nfp_prog->prog_len - 1] |=
206 FIELD_PREP(OP_RELO_TYPE, relo);
210 emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer)
212 emit_br_relo(nfp_prog, mask, addr, defer, RELO_BR_REL);
216 __emit_br_bit(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 addr, u8 defer,
217 bool set, bool src_lmextn)
219 u16 addr_lo, addr_hi;
222 addr_lo = addr & (OP_BR_BIT_ADDR_LO >> __bf_shf(OP_BR_BIT_ADDR_LO));
223 addr_hi = addr != addr_lo;
225 insn = OP_BR_BIT_BASE |
226 FIELD_PREP(OP_BR_BIT_A_SRC, areg) |
227 FIELD_PREP(OP_BR_BIT_B_SRC, breg) |
228 FIELD_PREP(OP_BR_BIT_BV, set) |
229 FIELD_PREP(OP_BR_BIT_DEFBR, defer) |
230 FIELD_PREP(OP_BR_BIT_ADDR_LO, addr_lo) |
231 FIELD_PREP(OP_BR_BIT_ADDR_HI, addr_hi) |
232 FIELD_PREP(OP_BR_BIT_SRC_LMEXTN, src_lmextn);
234 nfp_prog_push(nfp_prog, insn);
238 emit_br_bit_relo(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr,
239 u8 defer, bool set, enum nfp_relo_type relo)
241 struct nfp_insn_re_regs reg;
244 /* NOTE: The bit to test is specified as an rotation amount, such that
245 * the bit to test will be placed on the MSB of the result when
246 * doing a rotate right. For bit X, we need right rotate X + 1.
250 err = swreg_to_restricted(reg_none(), src, reg_imm(bit), ®, false);
252 nfp_prog->error = err;
256 __emit_br_bit(nfp_prog, reg.areg, reg.breg, addr, defer, set,
259 nfp_prog->prog[nfp_prog->prog_len - 1] |=
260 FIELD_PREP(OP_RELO_TYPE, relo);
264 emit_br_bset(struct nfp_prog *nfp_prog, swreg src, u8 bit, u16 addr, u8 defer)
266 emit_br_bit_relo(nfp_prog, src, bit, addr, defer, true, RELO_BR_REL);
270 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi,
271 enum immed_width width, bool invert,
272 enum immed_shift shift, bool wr_both,
273 bool dst_lmextn, bool src_lmextn)
277 insn = OP_IMMED_BASE |
278 FIELD_PREP(OP_IMMED_A_SRC, areg) |
279 FIELD_PREP(OP_IMMED_B_SRC, breg) |
280 FIELD_PREP(OP_IMMED_IMM, imm_hi) |
281 FIELD_PREP(OP_IMMED_WIDTH, width) |
282 FIELD_PREP(OP_IMMED_INV, invert) |
283 FIELD_PREP(OP_IMMED_SHIFT, shift) |
284 FIELD_PREP(OP_IMMED_WR_AB, wr_both) |
285 FIELD_PREP(OP_IMMED_SRC_LMEXTN, src_lmextn) |
286 FIELD_PREP(OP_IMMED_DST_LMEXTN, dst_lmextn);
288 nfp_prog_push(nfp_prog, insn);
292 emit_immed(struct nfp_prog *nfp_prog, swreg dst, u16 imm,
293 enum immed_width width, bool invert, enum immed_shift shift)
295 struct nfp_insn_ur_regs reg;
298 if (swreg_type(dst) == NN_REG_IMM) {
299 nfp_prog->error = -EFAULT;
303 err = swreg_to_unrestricted(dst, dst, reg_imm(imm & 0xff), ®);
305 nfp_prog->error = err;
309 /* Use reg.dst when destination is No-Dest. */
310 __emit_immed(nfp_prog,
311 swreg_type(dst) == NN_REG_NONE ? reg.dst : reg.areg,
312 reg.breg, imm >> 8, width, invert, shift,
313 reg.wr_both, reg.dst_lmextn, reg.src_lmextn);
317 __emit_shf(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab,
318 enum shf_sc sc, u8 shift,
319 u16 areg, enum shf_op op, u16 breg, bool i8, bool sw, bool wr_both,
320 bool dst_lmextn, bool src_lmextn)
324 if (!FIELD_FIT(OP_SHF_SHIFT, shift)) {
325 nfp_prog->error = -EFAULT;
329 if (sc == SHF_SC_L_SHF)
333 FIELD_PREP(OP_SHF_A_SRC, areg) |
334 FIELD_PREP(OP_SHF_SC, sc) |
335 FIELD_PREP(OP_SHF_B_SRC, breg) |
336 FIELD_PREP(OP_SHF_I8, i8) |
337 FIELD_PREP(OP_SHF_SW, sw) |
338 FIELD_PREP(OP_SHF_DST, dst) |
339 FIELD_PREP(OP_SHF_SHIFT, shift) |
340 FIELD_PREP(OP_SHF_OP, op) |
341 FIELD_PREP(OP_SHF_DST_AB, dst_ab) |
342 FIELD_PREP(OP_SHF_WR_AB, wr_both) |
343 FIELD_PREP(OP_SHF_SRC_LMEXTN, src_lmextn) |
344 FIELD_PREP(OP_SHF_DST_LMEXTN, dst_lmextn);
346 nfp_prog_push(nfp_prog, insn);
350 emit_shf(struct nfp_prog *nfp_prog, swreg dst,
351 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc, u8 shift)
353 struct nfp_insn_re_regs reg;
356 err = swreg_to_restricted(dst, lreg, rreg, ®, true);
358 nfp_prog->error = err;
362 __emit_shf(nfp_prog, reg.dst, reg.dst_ab, sc, shift,
363 reg.areg, op, reg.breg, reg.i8, reg.swap, reg.wr_both,
364 reg.dst_lmextn, reg.src_lmextn);
368 emit_shf_indir(struct nfp_prog *nfp_prog, swreg dst,
369 swreg lreg, enum shf_op op, swreg rreg, enum shf_sc sc)
371 if (sc == SHF_SC_R_ROT) {
372 pr_err("indirect shift is not allowed on rotation\n");
373 nfp_prog->error = -EFAULT;
377 emit_shf(nfp_prog, dst, lreg, op, rreg, sc, 0);
381 __emit_alu(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab,
382 u16 areg, enum alu_op op, u16 breg, bool swap, bool wr_both,
383 bool dst_lmextn, bool src_lmextn)
388 FIELD_PREP(OP_ALU_A_SRC, areg) |
389 FIELD_PREP(OP_ALU_B_SRC, breg) |
390 FIELD_PREP(OP_ALU_DST, dst) |
391 FIELD_PREP(OP_ALU_SW, swap) |
392 FIELD_PREP(OP_ALU_OP, op) |
393 FIELD_PREP(OP_ALU_DST_AB, dst_ab) |
394 FIELD_PREP(OP_ALU_WR_AB, wr_both) |
395 FIELD_PREP(OP_ALU_SRC_LMEXTN, src_lmextn) |
396 FIELD_PREP(OP_ALU_DST_LMEXTN, dst_lmextn);
398 nfp_prog_push(nfp_prog, insn);
402 emit_alu(struct nfp_prog *nfp_prog, swreg dst,
403 swreg lreg, enum alu_op op, swreg rreg)
405 struct nfp_insn_ur_regs reg;
408 err = swreg_to_unrestricted(dst, lreg, rreg, ®);
410 nfp_prog->error = err;
414 __emit_alu(nfp_prog, reg.dst, reg.dst_ab,
415 reg.areg, op, reg.breg, reg.swap, reg.wr_both,
416 reg.dst_lmextn, reg.src_lmextn);
420 __emit_mul(struct nfp_prog *nfp_prog, enum alu_dst_ab dst_ab, u16 areg,
421 enum mul_type type, enum mul_step step, u16 breg, bool swap,
422 bool wr_both, bool dst_lmextn, bool src_lmextn)
427 FIELD_PREP(OP_MUL_A_SRC, areg) |
428 FIELD_PREP(OP_MUL_B_SRC, breg) |
429 FIELD_PREP(OP_MUL_STEP, step) |
430 FIELD_PREP(OP_MUL_DST_AB, dst_ab) |
431 FIELD_PREP(OP_MUL_SW, swap) |
432 FIELD_PREP(OP_MUL_TYPE, type) |
433 FIELD_PREP(OP_MUL_WR_AB, wr_both) |
434 FIELD_PREP(OP_MUL_SRC_LMEXTN, src_lmextn) |
435 FIELD_PREP(OP_MUL_DST_LMEXTN, dst_lmextn);
437 nfp_prog_push(nfp_prog, insn);
441 emit_mul(struct nfp_prog *nfp_prog, swreg lreg, enum mul_type type,
442 enum mul_step step, swreg rreg)
444 struct nfp_insn_ur_regs reg;
448 if (type == MUL_TYPE_START && step != MUL_STEP_NONE) {
449 nfp_prog->error = -EINVAL;
453 if (step == MUL_LAST || step == MUL_LAST_2) {
454 /* When type is step and step Number is LAST or LAST2, left
455 * source is used as destination.
457 err = swreg_to_unrestricted(lreg, reg_none(), rreg, ®);
460 err = swreg_to_unrestricted(reg_none(), lreg, rreg, ®);
465 nfp_prog->error = err;
469 __emit_mul(nfp_prog, reg.dst_ab, areg, type, step, reg.breg, reg.swap,
470 reg.wr_both, reg.dst_lmextn, reg.src_lmextn);
474 __emit_ld_field(struct nfp_prog *nfp_prog, enum shf_sc sc,
475 u8 areg, u8 bmask, u8 breg, u8 shift, bool imm8,
476 bool zero, bool swap, bool wr_both,
477 bool dst_lmextn, bool src_lmextn)
482 FIELD_PREP(OP_LDF_A_SRC, areg) |
483 FIELD_PREP(OP_LDF_SC, sc) |
484 FIELD_PREP(OP_LDF_B_SRC, breg) |
485 FIELD_PREP(OP_LDF_I8, imm8) |
486 FIELD_PREP(OP_LDF_SW, swap) |
487 FIELD_PREP(OP_LDF_ZF, zero) |
488 FIELD_PREP(OP_LDF_BMASK, bmask) |
489 FIELD_PREP(OP_LDF_SHF, shift) |
490 FIELD_PREP(OP_LDF_WR_AB, wr_both) |
491 FIELD_PREP(OP_LDF_SRC_LMEXTN, src_lmextn) |
492 FIELD_PREP(OP_LDF_DST_LMEXTN, dst_lmextn);
494 nfp_prog_push(nfp_prog, insn);
498 emit_ld_field_any(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src,
499 enum shf_sc sc, u8 shift, bool zero)
501 struct nfp_insn_re_regs reg;
504 /* Note: ld_field is special as it uses one of the src regs as dst */
505 err = swreg_to_restricted(dst, dst, src, ®, true);
507 nfp_prog->error = err;
511 __emit_ld_field(nfp_prog, sc, reg.areg, bmask, reg.breg, shift,
512 reg.i8, zero, reg.swap, reg.wr_both,
513 reg.dst_lmextn, reg.src_lmextn);
517 emit_ld_field(struct nfp_prog *nfp_prog, swreg dst, u8 bmask, swreg src,
518 enum shf_sc sc, u8 shift)
520 emit_ld_field_any(nfp_prog, dst, bmask, src, sc, shift, false);
524 __emit_lcsr(struct nfp_prog *nfp_prog, u16 areg, u16 breg, bool wr, u16 addr,
525 bool dst_lmextn, bool src_lmextn)
529 insn = OP_LCSR_BASE |
530 FIELD_PREP(OP_LCSR_A_SRC, areg) |
531 FIELD_PREP(OP_LCSR_B_SRC, breg) |
532 FIELD_PREP(OP_LCSR_WRITE, wr) |
533 FIELD_PREP(OP_LCSR_ADDR, addr / 4) |
534 FIELD_PREP(OP_LCSR_SRC_LMEXTN, src_lmextn) |
535 FIELD_PREP(OP_LCSR_DST_LMEXTN, dst_lmextn);
537 nfp_prog_push(nfp_prog, insn);
540 static void emit_csr_wr(struct nfp_prog *nfp_prog, swreg src, u16 addr)
542 struct nfp_insn_ur_regs reg;
545 /* This instruction takes immeds instead of reg_none() for the ignored
546 * operand, but we can't encode 2 immeds in one instr with our normal
547 * swreg infra so if param is an immed, we encode as reg_none() and
548 * copy the immed to both operands.
550 if (swreg_type(src) == NN_REG_IMM) {
551 err = swreg_to_unrestricted(reg_none(), src, reg_none(), ®);
554 err = swreg_to_unrestricted(reg_none(), src, reg_imm(0), ®);
557 nfp_prog->error = err;
561 __emit_lcsr(nfp_prog, reg.areg, reg.breg, true, addr,
562 false, reg.src_lmextn);
565 /* CSR value is read in following immed[gpr, 0] */
566 static void __emit_csr_rd(struct nfp_prog *nfp_prog, u16 addr)
568 __emit_lcsr(nfp_prog, 0, 0, false, addr, false, false);
571 static void emit_nop(struct nfp_prog *nfp_prog)
573 __emit_immed(nfp_prog, UR_REG_IMM, UR_REG_IMM, 0, 0, 0, 0, 0, 0, 0);
576 /* --- Wrappers --- */
577 static bool pack_immed(u32 imm, u16 *val, enum immed_shift *shift)
579 if (!(imm & 0xffff0000)) {
581 *shift = IMMED_SHIFT_0B;
582 } else if (!(imm & 0xff0000ff)) {
584 *shift = IMMED_SHIFT_1B;
585 } else if (!(imm & 0x0000ffff)) {
587 *shift = IMMED_SHIFT_2B;
595 static void wrp_immed(struct nfp_prog *nfp_prog, swreg dst, u32 imm)
597 enum immed_shift shift;
600 if (pack_immed(imm, &val, &shift)) {
601 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, false, shift);
602 } else if (pack_immed(~imm, &val, &shift)) {
603 emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, true, shift);
605 emit_immed(nfp_prog, dst, imm & 0xffff, IMMED_WIDTH_ALL,
606 false, IMMED_SHIFT_0B);
607 emit_immed(nfp_prog, dst, imm >> 16, IMMED_WIDTH_WORD,
608 false, IMMED_SHIFT_2B);
613 wrp_immed_relo(struct nfp_prog *nfp_prog, swreg dst, u32 imm,
614 enum nfp_relo_type relo)
617 pr_err("relocation of a large immediate!\n");
618 nfp_prog->error = -EFAULT;
621 emit_immed(nfp_prog, dst, imm, IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B);
623 nfp_prog->prog[nfp_prog->prog_len - 1] |=
624 FIELD_PREP(OP_RELO_TYPE, relo);
627 /* ur_load_imm_any() - encode immediate or use tmp register (unrestricted)
628 * If the @imm is small enough encode it directly in operand and return
629 * otherwise load @imm to a spare register and return its encoding.
631 static swreg ur_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg)
633 if (FIELD_FIT(UR_REG_IMM_MAX, imm))
636 wrp_immed(nfp_prog, tmp_reg, imm);
640 /* re_load_imm_any() - encode immediate or use tmp register (restricted)
641 * If the @imm is small enough encode it directly in operand and return
642 * otherwise load @imm to a spare register and return its encoding.
644 static swreg re_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, swreg tmp_reg)
646 if (FIELD_FIT(RE_REG_IMM_MAX, imm))
649 wrp_immed(nfp_prog, tmp_reg, imm);
653 static void wrp_nops(struct nfp_prog *nfp_prog, unsigned int count)
659 static void wrp_mov(struct nfp_prog *nfp_prog, swreg dst, swreg src)
661 emit_alu(nfp_prog, dst, reg_none(), ALU_OP_NONE, src);
664 static void wrp_reg_mov(struct nfp_prog *nfp_prog, u16 dst, u16 src)
666 wrp_mov(nfp_prog, reg_both(dst), reg_b(src));
669 /* wrp_reg_subpart() - load @field_len bytes from @offset of @src, write the
670 * result to @dst from low end.
673 wrp_reg_subpart(struct nfp_prog *nfp_prog, swreg dst, swreg src, u8 field_len,
676 enum shf_sc sc = offset ? SHF_SC_R_SHF : SHF_SC_NONE;
677 u8 mask = (1 << field_len) - 1;
679 emit_ld_field_any(nfp_prog, dst, mask, src, sc, offset * 8, true);
682 /* wrp_reg_or_subpart() - load @field_len bytes from low end of @src, or the
683 * result to @dst from offset, there is no change on the other bits of @dst.
686 wrp_reg_or_subpart(struct nfp_prog *nfp_prog, swreg dst, swreg src,
687 u8 field_len, u8 offset)
689 enum shf_sc sc = offset ? SHF_SC_L_SHF : SHF_SC_NONE;
690 u8 mask = ((1 << field_len) - 1) << offset;
692 emit_ld_field(nfp_prog, dst, mask, src, sc, 32 - offset * 8);
696 addr40_offset(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset,
697 swreg *rega, swreg *regb)
699 if (offset == reg_imm(0)) {
700 *rega = reg_a(src_gpr);
701 *regb = reg_b(src_gpr + 1);
705 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(src_gpr), ALU_OP_ADD, offset);
706 emit_alu(nfp_prog, imm_b(nfp_prog), reg_b(src_gpr + 1), ALU_OP_ADD_C,
708 *rega = imm_a(nfp_prog);
709 *regb = imm_b(nfp_prog);
712 /* NFP has Command Push Pull bus which supports bluk memory operations. */
713 static int nfp_cpp_memcpy(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
715 bool descending_seq = meta->ldst_gather_len < 0;
716 s16 len = abs(meta->ldst_gather_len);
722 off = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog));
723 src_40bit_addr = meta->ptr.type == PTR_TO_MAP_VALUE;
724 src_base = reg_a(meta->insn.src_reg * 2);
725 xfer_num = round_up(len, 4) / 4;
728 addr40_offset(nfp_prog, meta->insn.src_reg * 2, off, &src_base,
731 /* Setup PREV_ALU fields to override memory read length. */
733 wrp_immed(nfp_prog, reg_none(),
734 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1));
736 /* Memory read from source addr into transfer-in registers. */
737 emit_cmd_any(nfp_prog, CMD_TGT_READ32_SWAP,
738 src_40bit_addr ? CMD_MODE_40b_BA : CMD_MODE_32b, 0,
739 src_base, off, xfer_num - 1, CMD_CTX_SWAP, len > 32);
741 /* Move from transfer-in to transfer-out. */
742 for (i = 0; i < xfer_num; i++)
743 wrp_mov(nfp_prog, reg_xfer(i), reg_xfer(i));
745 off = re_load_imm_any(nfp_prog, meta->paired_st->off, imm_b(nfp_prog));
748 /* Use single direct_ref write8. */
749 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0,
750 reg_a(meta->paired_st->dst_reg * 2), off, len - 1,
752 } else if (len <= 32 && IS_ALIGNED(len, 4)) {
753 /* Use single direct_ref write32. */
754 emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
755 reg_a(meta->paired_st->dst_reg * 2), off, xfer_num - 1,
757 } else if (len <= 32) {
758 /* Use single indirect_ref write8. */
759 wrp_immed(nfp_prog, reg_none(),
760 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, len - 1));
761 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0,
762 reg_a(meta->paired_st->dst_reg * 2), off,
763 len - 1, CMD_CTX_SWAP);
764 } else if (IS_ALIGNED(len, 4)) {
765 /* Use single indirect_ref write32. */
766 wrp_immed(nfp_prog, reg_none(),
767 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1));
768 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
769 reg_a(meta->paired_st->dst_reg * 2), off,
770 xfer_num - 1, CMD_CTX_SWAP);
771 } else if (len <= 40) {
772 /* Use one direct_ref write32 to write the first 32-bytes, then
773 * another direct_ref write8 to write the remaining bytes.
775 emit_cmd(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
776 reg_a(meta->paired_st->dst_reg * 2), off, 7,
779 off = re_load_imm_any(nfp_prog, meta->paired_st->off + 32,
781 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 8,
782 reg_a(meta->paired_st->dst_reg * 2), off, len - 33,
785 /* Use one indirect_ref write32 to write 4-bytes aligned length,
786 * then another direct_ref write8 to write the remaining bytes.
790 wrp_immed(nfp_prog, reg_none(),
791 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 2));
792 emit_cmd_indir(nfp_prog, CMD_TGT_WRITE32_SWAP, CMD_MODE_32b, 0,
793 reg_a(meta->paired_st->dst_reg * 2), off,
794 xfer_num - 2, CMD_CTX_SWAP);
795 new_off = meta->paired_st->off + (xfer_num - 1) * 4;
796 off = re_load_imm_any(nfp_prog, new_off, imm_b(nfp_prog));
797 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b,
798 xfer_num - 1, reg_a(meta->paired_st->dst_reg * 2), off,
799 (len & 0x3) - 1, CMD_CTX_SWAP);
802 /* TODO: The following extra load is to make sure data flow be identical
803 * before and after we do memory copy optimization.
805 * The load destination register is not guaranteed to be dead, so we
806 * need to make sure it is loaded with the value the same as before
807 * this transformation.
809 * These extra loads could be removed once we have accurate register
814 else if (BPF_SIZE(meta->insn.code) != BPF_DW)
815 xfer_num = xfer_num - 1;
817 xfer_num = xfer_num - 2;
819 switch (BPF_SIZE(meta->insn.code)) {
821 wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2),
822 reg_xfer(xfer_num), 1,
823 IS_ALIGNED(len, 4) ? 3 : (len & 3) - 1);
826 wrp_reg_subpart(nfp_prog, reg_both(meta->insn.dst_reg * 2),
827 reg_xfer(xfer_num), 2, (len & 3) ^ 2);
830 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2),
834 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2),
836 wrp_mov(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1),
837 reg_xfer(xfer_num + 1));
841 if (BPF_SIZE(meta->insn.code) != BPF_DW)
842 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
848 data_ld(struct nfp_prog *nfp_prog, swreg offset, u8 dst_gpr, int size)
853 /* We load the value from the address indicated in @offset and then
854 * shift out the data we don't need. Note: this is big endian!
857 shift = size < 4 ? 4 - size : 0;
859 emit_cmd(nfp_prog, CMD_TGT_READ8, CMD_MODE_32b, 0,
860 pptr_reg(nfp_prog), offset, sz - 1, CMD_CTX_SWAP);
864 emit_shf(nfp_prog, reg_both(dst_gpr), reg_none(), SHF_OP_NONE,
865 reg_xfer(0), SHF_SC_R_SHF, shift * 8);
867 for (; i * 4 < size; i++)
868 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i));
871 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0);
877 data_ld_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr,
878 swreg lreg, swreg rreg, int size, enum cmd_mode mode)
883 /* We load the value from the address indicated in rreg + lreg and then
884 * mask out the data we don't need. Note: this is little endian!
887 mask = size < 4 ? GENMASK(size - 1, 0) : 0;
889 emit_cmd(nfp_prog, CMD_TGT_READ32_SWAP, mode, 0,
890 lreg, rreg, sz / 4 - 1, CMD_CTX_SWAP);
894 emit_ld_field_any(nfp_prog, reg_both(dst_gpr), mask,
895 reg_xfer(0), SHF_SC_NONE, 0, true);
897 for (; i * 4 < size; i++)
898 wrp_mov(nfp_prog, reg_both(dst_gpr + i), reg_xfer(i));
901 wrp_immed(nfp_prog, reg_both(dst_gpr + 1), 0);
907 data_ld_host_order_addr32(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset,
910 return data_ld_host_order(nfp_prog, dst_gpr, reg_a(src_gpr), offset,
915 data_ld_host_order_addr40(struct nfp_prog *nfp_prog, u8 src_gpr, swreg offset,
920 addr40_offset(nfp_prog, src_gpr, offset, ®a, ®b);
922 return data_ld_host_order(nfp_prog, dst_gpr, rega, regb,
923 size, CMD_MODE_40b_BA);
927 construct_data_ind_ld(struct nfp_prog *nfp_prog, u16 offset, u16 src, u8 size)
931 /* Calculate the true offset (src_reg + imm) */
932 tmp_reg = ur_load_imm_any(nfp_prog, offset, imm_b(nfp_prog));
933 emit_alu(nfp_prog, imm_both(nfp_prog), reg_a(src), ALU_OP_ADD, tmp_reg);
935 /* Check packet length (size guaranteed to fit b/c it's u8) */
936 emit_alu(nfp_prog, imm_a(nfp_prog),
937 imm_a(nfp_prog), ALU_OP_ADD, reg_imm(size));
938 emit_alu(nfp_prog, reg_none(),
939 plen_reg(nfp_prog), ALU_OP_SUB, imm_a(nfp_prog));
940 emit_br_relo(nfp_prog, BR_BLO, BR_OFF_RELO, 0, RELO_BR_GO_ABORT);
943 return data_ld(nfp_prog, imm_b(nfp_prog), 0, size);
946 static int construct_data_ld(struct nfp_prog *nfp_prog, u16 offset, u8 size)
950 /* Check packet length */
951 tmp_reg = ur_load_imm_any(nfp_prog, offset + size, imm_a(nfp_prog));
952 emit_alu(nfp_prog, reg_none(), plen_reg(nfp_prog), ALU_OP_SUB, tmp_reg);
953 emit_br_relo(nfp_prog, BR_BLO, BR_OFF_RELO, 0, RELO_BR_GO_ABORT);
956 tmp_reg = re_load_imm_any(nfp_prog, offset, imm_b(nfp_prog));
957 return data_ld(nfp_prog, tmp_reg, 0, size);
961 data_stx_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset,
966 for (i = 0; i * 4 < size; i++)
967 wrp_mov(nfp_prog, reg_xfer(i), reg_a(src_gpr + i));
969 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0,
970 reg_a(dst_gpr), offset, size - 1, CMD_CTX_SWAP);
976 data_st_host_order(struct nfp_prog *nfp_prog, u8 dst_gpr, swreg offset,
979 wrp_immed(nfp_prog, reg_xfer(0), imm);
981 wrp_immed(nfp_prog, reg_xfer(1), imm >> 32);
983 emit_cmd(nfp_prog, CMD_TGT_WRITE8_SWAP, CMD_MODE_32b, 0,
984 reg_a(dst_gpr), offset, size - 1, CMD_CTX_SWAP);
990 (*lmem_step)(struct nfp_prog *nfp_prog, u8 gpr, u8 gpr_byte, s32 off,
991 unsigned int size, bool first, bool new_gpr, bool last, bool lm3,
995 wrp_lmem_load(struct nfp_prog *nfp_prog, u8 dst, u8 dst_byte, s32 off,
996 unsigned int size, bool first, bool new_gpr, bool last, bool lm3,
999 bool should_inc = needs_inc && new_gpr && !last;
1006 if (WARN_ON_ONCE(dst_byte + size > 4 || off % 4 + size > 4))
1011 /* Move the entire word */
1013 wrp_mov(nfp_prog, reg_both(dst),
1014 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx));
1018 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX))
1023 mask = (1 << size) - 1;
1026 if (WARN_ON_ONCE(mask > 0xf))
1029 shf = abs(src_byte - dst_byte) * 8;
1030 if (src_byte == dst_byte) {
1032 } else if (src_byte < dst_byte) {
1039 /* ld_field can address fewer indexes, if offset too large do RMW.
1040 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes.
1042 if (idx <= RE_REG_LM_IDX_MAX) {
1043 reg = reg_lm(lm3 ? 3 : 0, idx);
1045 reg = imm_a(nfp_prog);
1046 /* If it's not the first part of the load and we start a new GPR
1047 * that means we are loading a second part of the LMEM word into
1048 * a new GPR. IOW we've already looked that LMEM word and
1049 * therefore it has been loaded into imm_a().
1051 if (first || !new_gpr)
1052 wrp_mov(nfp_prog, reg, reg_lm(0, idx));
1055 emit_ld_field_any(nfp_prog, reg_both(dst), mask, reg, sc, shf, new_gpr);
1058 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3));
1064 wrp_lmem_store(struct nfp_prog *nfp_prog, u8 src, u8 src_byte, s32 off,
1065 unsigned int size, bool first, bool new_gpr, bool last, bool lm3,
1068 bool should_inc = needs_inc && new_gpr && !last;
1075 if (WARN_ON_ONCE(src_byte + size > 4 || off % 4 + size > 4))
1080 /* Move the entire word */
1083 should_inc ? reg_lm_inc(3) : reg_lm(lm3 ? 3 : 0, idx),
1088 if (WARN_ON_ONCE(lm3 && idx > RE_REG_LM_IDX_MAX))
1093 mask = (1 << size) - 1;
1096 if (WARN_ON_ONCE(mask > 0xf))
1099 shf = abs(src_byte - dst_byte) * 8;
1100 if (src_byte == dst_byte) {
1102 } else if (src_byte < dst_byte) {
1109 /* ld_field can address fewer indexes, if offset too large do RMW.
1110 * Because we RMV twice we waste 2 cycles on unaligned 8 byte writes.
1112 if (idx <= RE_REG_LM_IDX_MAX) {
1113 reg = reg_lm(lm3 ? 3 : 0, idx);
1115 reg = imm_a(nfp_prog);
1116 /* Only first and last LMEM locations are going to need RMW,
1117 * the middle location will be overwritten fully.
1120 wrp_mov(nfp_prog, reg, reg_lm(0, idx));
1123 emit_ld_field(nfp_prog, reg, mask, reg_b(src), sc, shf);
1125 if (new_gpr || last) {
1126 if (idx > RE_REG_LM_IDX_MAX)
1127 wrp_mov(nfp_prog, reg_lm(0, idx), reg);
1129 wrp_mov(nfp_prog, reg_none(), reg_lm_inc(3));
1136 mem_op_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
1137 unsigned int size, unsigned int ptr_off, u8 gpr, u8 ptr_gpr,
1138 bool clr_gpr, lmem_step step)
1140 s32 off = nfp_prog->stack_depth + meta->insn.off + ptr_off;
1141 bool first = true, last;
1142 bool needs_inc = false;
1143 swreg stack_off_reg;
1149 if (meta->ptr_not_const) {
1150 /* Use of the last encountered ptr_off is OK, they all have
1151 * the same alignment. Depend on low bits of value being
1152 * discarded when written to LMaddr register.
1154 stack_off_reg = ur_load_imm_any(nfp_prog, meta->insn.off,
1155 stack_imm(nfp_prog));
1157 emit_alu(nfp_prog, imm_b(nfp_prog),
1158 reg_a(ptr_gpr), ALU_OP_ADD, stack_off_reg);
1161 } else if (off + size <= 64) {
1162 /* We can reach bottom 64B with LMaddr0 */
1164 } else if (round_down(off, 32) == round_down(off + size - 1, 32)) {
1165 /* We have to set up a new pointer. If we know the offset
1166 * and the entire access falls into a single 32 byte aligned
1167 * window we won't have to increment the LM pointer.
1168 * The 32 byte alignment is imporant because offset is ORed in
1169 * not added when doing *l$indexN[off].
1171 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 32),
1172 stack_imm(nfp_prog));
1173 emit_alu(nfp_prog, imm_b(nfp_prog),
1174 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg);
1178 stack_off_reg = ur_load_imm_any(nfp_prog, round_down(off, 4),
1179 stack_imm(nfp_prog));
1181 emit_alu(nfp_prog, imm_b(nfp_prog),
1182 stack_reg(nfp_prog), ALU_OP_ADD, stack_off_reg);
1187 emit_csr_wr(nfp_prog, imm_b(nfp_prog), NFP_CSR_ACT_LM_ADDR3);
1188 /* For size < 4 one slot will be filled by zeroing of upper. */
1189 wrp_nops(nfp_prog, clr_gpr && size < 8 ? 2 : 3);
1192 if (clr_gpr && size < 8)
1193 wrp_immed(nfp_prog, reg_both(gpr + 1), 0);
1199 slice_size = min(size, 4 - gpr_byte);
1200 slice_end = min(off + slice_size, round_up(off + 1, 4));
1201 slice_size = slice_end - off;
1203 last = slice_size == size;
1208 ret = step(nfp_prog, gpr, gpr_byte, off, slice_size,
1209 first, gpr != prev_gpr, last, lm3, needs_inc);
1216 gpr_byte += slice_size;
1217 if (gpr_byte >= 4) {
1230 wrp_alu_imm(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u32 imm)
1234 if (alu_op == ALU_OP_AND) {
1236 wrp_immed(nfp_prog, reg_both(dst), 0);
1240 if (alu_op == ALU_OP_OR) {
1242 wrp_immed(nfp_prog, reg_both(dst), ~0U);
1246 if (alu_op == ALU_OP_XOR) {
1248 emit_alu(nfp_prog, reg_both(dst), reg_none(),
1249 ALU_OP_NOT, reg_b(dst));
1254 tmp_reg = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog));
1255 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, tmp_reg);
1259 wrp_alu64_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
1260 enum alu_op alu_op, bool skip)
1262 const struct bpf_insn *insn = &meta->insn;
1263 u64 imm = insn->imm; /* sign extend */
1270 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, imm & ~0U);
1271 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, alu_op, imm >> 32);
1277 wrp_alu64_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
1280 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2;
1282 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src));
1283 emit_alu(nfp_prog, reg_both(dst + 1),
1284 reg_a(dst + 1), alu_op, reg_b(src + 1));
1290 wrp_alu32_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
1291 enum alu_op alu_op, bool skip)
1293 const struct bpf_insn *insn = &meta->insn;
1300 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, insn->imm);
1301 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
1307 wrp_alu32_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
1310 u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2;
1312 emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src));
1313 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
1319 wrp_test_reg_one(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u8 src,
1320 enum br_mask br_mask, u16 off)
1322 emit_alu(nfp_prog, reg_none(), reg_a(dst), alu_op, reg_b(src));
1323 emit_br(nfp_prog, br_mask, off, 0);
1327 wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
1328 enum alu_op alu_op, enum br_mask br_mask)
1330 const struct bpf_insn *insn = &meta->insn;
1332 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2, alu_op,
1333 insn->src_reg * 2, br_mask, insn->off);
1334 wrp_test_reg_one(nfp_prog, insn->dst_reg * 2 + 1, alu_op,
1335 insn->src_reg * 2 + 1, br_mask, insn->off);
1340 static const struct jmp_code_map {
1341 enum br_mask br_mask;
1343 } jmp_code_map[] = {
1344 [BPF_JGT >> 4] = { BR_BLO, true },
1345 [BPF_JGE >> 4] = { BR_BHS, false },
1346 [BPF_JLT >> 4] = { BR_BLO, false },
1347 [BPF_JLE >> 4] = { BR_BHS, true },
1348 [BPF_JSGT >> 4] = { BR_BLT, true },
1349 [BPF_JSGE >> 4] = { BR_BGE, false },
1350 [BPF_JSLT >> 4] = { BR_BLT, false },
1351 [BPF_JSLE >> 4] = { BR_BGE, true },
1354 static const struct jmp_code_map *nfp_jmp_code_get(struct nfp_insn_meta *meta)
1358 op = BPF_OP(meta->insn.code) >> 4;
1359 /* br_mask of 0 is BR_BEQ which we don't use in jump code table */
1360 if (WARN_ONCE(op >= ARRAY_SIZE(jmp_code_map) ||
1361 !jmp_code_map[op].br_mask,
1362 "no code found for jump instruction"))
1365 return &jmp_code_map[op];
1368 static int cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1370 const struct bpf_insn *insn = &meta->insn;
1371 u64 imm = insn->imm; /* sign extend */
1372 const struct jmp_code_map *code;
1373 enum alu_op alu_op, carry_op;
1374 u8 reg = insn->dst_reg * 2;
1377 code = nfp_jmp_code_get(meta);
1381 alu_op = meta->jump_neg_op ? ALU_OP_ADD : ALU_OP_SUB;
1382 carry_op = meta->jump_neg_op ? ALU_OP_ADD_C : ALU_OP_SUB_C;
1384 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
1386 emit_alu(nfp_prog, reg_none(), reg_a(reg), alu_op, tmp_reg);
1388 emit_alu(nfp_prog, reg_none(), tmp_reg, alu_op, reg_a(reg));
1390 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
1392 emit_alu(nfp_prog, reg_none(),
1393 reg_a(reg + 1), carry_op, tmp_reg);
1395 emit_alu(nfp_prog, reg_none(),
1396 tmp_reg, carry_op, reg_a(reg + 1));
1398 emit_br(nfp_prog, code->br_mask, insn->off, 0);
1403 static int cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1405 const struct bpf_insn *insn = &meta->insn;
1406 const struct jmp_code_map *code;
1409 code = nfp_jmp_code_get(meta);
1413 areg = insn->dst_reg * 2;
1414 breg = insn->src_reg * 2;
1422 emit_alu(nfp_prog, reg_none(), reg_a(areg), ALU_OP_SUB, reg_b(breg));
1423 emit_alu(nfp_prog, reg_none(),
1424 reg_a(areg + 1), ALU_OP_SUB_C, reg_b(breg + 1));
1425 emit_br(nfp_prog, code->br_mask, insn->off, 0);
1430 static void wrp_end32(struct nfp_prog *nfp_prog, swreg reg_in, u8 gpr_out)
1432 emit_ld_field(nfp_prog, reg_both(gpr_out), 0xf, reg_in,
1434 emit_ld_field(nfp_prog, reg_both(gpr_out), 0x5, reg_a(gpr_out),
1439 wrp_mul_u32(struct nfp_prog *nfp_prog, swreg dst_hi, swreg dst_lo, swreg lreg,
1440 swreg rreg, bool gen_high_half)
1442 emit_mul(nfp_prog, lreg, MUL_TYPE_START, MUL_STEP_NONE, rreg);
1443 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_1, rreg);
1444 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_2, rreg);
1445 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_3, rreg);
1446 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_32x32, MUL_STEP_4, rreg);
1447 emit_mul(nfp_prog, dst_lo, MUL_TYPE_STEP_32x32, MUL_LAST, reg_none());
1449 emit_mul(nfp_prog, dst_hi, MUL_TYPE_STEP_32x32, MUL_LAST_2,
1452 wrp_immed(nfp_prog, dst_hi, 0);
1456 wrp_mul_u16(struct nfp_prog *nfp_prog, swreg dst_hi, swreg dst_lo, swreg lreg,
1459 emit_mul(nfp_prog, lreg, MUL_TYPE_START, MUL_STEP_NONE, rreg);
1460 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_16x16, MUL_STEP_1, rreg);
1461 emit_mul(nfp_prog, lreg, MUL_TYPE_STEP_16x16, MUL_STEP_2, rreg);
1462 emit_mul(nfp_prog, dst_lo, MUL_TYPE_STEP_16x16, MUL_LAST, reg_none());
1466 wrp_mul(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
1467 bool gen_high_half, bool ropnd_from_reg)
1469 swreg multiplier, multiplicand, dst_hi, dst_lo;
1470 const struct bpf_insn *insn = &meta->insn;
1471 u32 lopnd_max, ropnd_max;
1474 dst_reg = insn->dst_reg;
1475 multiplicand = reg_a(dst_reg * 2);
1476 dst_hi = reg_both(dst_reg * 2 + 1);
1477 dst_lo = reg_both(dst_reg * 2);
1478 lopnd_max = meta->umax_dst;
1479 if (ropnd_from_reg) {
1480 multiplier = reg_b(insn->src_reg * 2);
1481 ropnd_max = meta->umax_src;
1483 u32 imm = insn->imm;
1485 multiplier = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog));
1488 if (lopnd_max > U16_MAX || ropnd_max > U16_MAX)
1489 wrp_mul_u32(nfp_prog, dst_hi, dst_lo, multiplicand, multiplier,
1492 wrp_mul_u16(nfp_prog, dst_hi, dst_lo, multiplicand, multiplier);
1497 static int wrp_div_imm(struct nfp_prog *nfp_prog, u8 dst, u64 imm)
1499 swreg dst_both = reg_both(dst), dst_a = reg_a(dst), dst_b = reg_a(dst);
1500 struct reciprocal_value_adv rvalue;
1504 if (imm > U32_MAX) {
1505 wrp_immed(nfp_prog, dst_both, 0);
1509 /* NOTE: because we are using "reciprocal_value_adv" which doesn't
1510 * support "divisor > (1u << 31)", we need to JIT separate NFP sequence
1511 * to handle such case which actually equals to the result of unsigned
1512 * comparison "dst >= imm" which could be calculated using the following
1515 * alu[--, dst, -, imm]
1517 * alu[dst, imm, +carry, 0]
1520 if (imm > 1U << 31) {
1521 swreg tmp_b = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog));
1523 emit_alu(nfp_prog, reg_none(), dst_a, ALU_OP_SUB, tmp_b);
1524 wrp_immed(nfp_prog, imm_a(nfp_prog), 0);
1525 emit_alu(nfp_prog, dst_both, imm_a(nfp_prog), ALU_OP_ADD_C,
1530 rvalue = reciprocal_value_adv(imm, 32);
1532 if (rvalue.is_wide_m && !(imm & 1)) {
1533 pre_shift = fls(imm & -imm) - 1;
1534 rvalue = reciprocal_value_adv(imm >> pre_shift, 32 - pre_shift);
1538 magic = ur_load_imm_any(nfp_prog, rvalue.m, imm_b(nfp_prog));
1539 if (imm == 1U << exp) {
1540 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b,
1542 } else if (rvalue.is_wide_m) {
1543 wrp_mul_u32(nfp_prog, imm_both(nfp_prog), reg_none(), dst_a,
1545 emit_alu(nfp_prog, dst_both, dst_a, ALU_OP_SUB,
1547 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b,
1549 emit_alu(nfp_prog, dst_both, dst_a, ALU_OP_ADD,
1551 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE, dst_b,
1552 SHF_SC_R_SHF, rvalue.sh - 1);
1555 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE,
1556 dst_b, SHF_SC_R_SHF, pre_shift);
1557 wrp_mul_u32(nfp_prog, dst_both, reg_none(), dst_a, magic, true);
1558 emit_shf(nfp_prog, dst_both, reg_none(), SHF_OP_NONE,
1559 dst_b, SHF_SC_R_SHF, rvalue.sh);
1565 static int adjust_head(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1567 swreg tmp = imm_a(nfp_prog), tmp_len = imm_b(nfp_prog);
1568 struct nfp_bpf_cap_adjust_head *adjust_head;
1569 u32 ret_einval, end;
1571 adjust_head = &nfp_prog->bpf->adjust_head;
1573 /* Optimized version - 5 vs 14 cycles */
1574 if (nfp_prog->adjust_head_location != UINT_MAX) {
1575 if (WARN_ON_ONCE(nfp_prog->adjust_head_location != meta->n))
1578 emit_alu(nfp_prog, pptr_reg(nfp_prog),
1579 reg_a(2 * 2), ALU_OP_ADD, pptr_reg(nfp_prog));
1580 emit_alu(nfp_prog, plen_reg(nfp_prog),
1581 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2));
1582 emit_alu(nfp_prog, pv_len(nfp_prog),
1583 pv_len(nfp_prog), ALU_OP_SUB, reg_a(2 * 2));
1585 wrp_immed(nfp_prog, reg_both(0), 0);
1586 wrp_immed(nfp_prog, reg_both(1), 0);
1588 /* TODO: when adjust head is guaranteed to succeed we can
1589 * also eliminate the following if (r0 == 0) branch.
1595 ret_einval = nfp_prog_current_offset(nfp_prog) + 14;
1596 end = ret_einval + 2;
1598 /* We need to use a temp because offset is just a part of the pkt ptr */
1599 emit_alu(nfp_prog, tmp,
1600 reg_a(2 * 2), ALU_OP_ADD_2B, pptr_reg(nfp_prog));
1602 /* Validate result will fit within FW datapath constraints */
1603 emit_alu(nfp_prog, reg_none(),
1604 tmp, ALU_OP_SUB, reg_imm(adjust_head->off_min));
1605 emit_br(nfp_prog, BR_BLO, ret_einval, 0);
1606 emit_alu(nfp_prog, reg_none(),
1607 reg_imm(adjust_head->off_max), ALU_OP_SUB, tmp);
1608 emit_br(nfp_prog, BR_BLO, ret_einval, 0);
1610 /* Validate the length is at least ETH_HLEN */
1611 emit_alu(nfp_prog, tmp_len,
1612 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2));
1613 emit_alu(nfp_prog, reg_none(),
1614 tmp_len, ALU_OP_SUB, reg_imm(ETH_HLEN));
1615 emit_br(nfp_prog, BR_BMI, ret_einval, 0);
1617 /* Load the ret code */
1618 wrp_immed(nfp_prog, reg_both(0), 0);
1619 wrp_immed(nfp_prog, reg_both(1), 0);
1621 /* Modify the packet metadata */
1622 emit_ld_field(nfp_prog, pptr_reg(nfp_prog), 0x3, tmp, SHF_SC_NONE, 0);
1624 /* Skip over the -EINVAL ret code (defer 2) */
1625 emit_br(nfp_prog, BR_UNC, end, 2);
1627 emit_alu(nfp_prog, plen_reg(nfp_prog),
1628 plen_reg(nfp_prog), ALU_OP_SUB, reg_a(2 * 2));
1629 emit_alu(nfp_prog, pv_len(nfp_prog),
1630 pv_len(nfp_prog), ALU_OP_SUB, reg_a(2 * 2));
1632 /* return -EINVAL target */
1633 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_einval))
1636 wrp_immed(nfp_prog, reg_both(0), -22);
1637 wrp_immed(nfp_prog, reg_both(1), ~0);
1639 if (!nfp_prog_confirm_current_offset(nfp_prog, end))
1645 static int adjust_tail(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1647 u32 ret_einval, end;
1650 BUILD_BUG_ON(plen_reg(nfp_prog) != reg_b(STATIC_REG_PKT_LEN));
1652 plen = imm_a(nfp_prog);
1653 delta = reg_a(2 * 2);
1655 ret_einval = nfp_prog_current_offset(nfp_prog) + 9;
1656 end = nfp_prog_current_offset(nfp_prog) + 11;
1658 /* Calculate resulting length */
1659 emit_alu(nfp_prog, plen, plen_reg(nfp_prog), ALU_OP_ADD, delta);
1660 /* delta == 0 is not allowed by the kernel, add must overflow to make
1663 emit_br(nfp_prog, BR_BCC, ret_einval, 0);
1665 /* if (new_len < 14) then -EINVAL */
1666 emit_alu(nfp_prog, reg_none(), plen, ALU_OP_SUB, reg_imm(ETH_HLEN));
1667 emit_br(nfp_prog, BR_BMI, ret_einval, 0);
1669 emit_alu(nfp_prog, plen_reg(nfp_prog),
1670 plen_reg(nfp_prog), ALU_OP_ADD, delta);
1671 emit_alu(nfp_prog, pv_len(nfp_prog),
1672 pv_len(nfp_prog), ALU_OP_ADD, delta);
1674 emit_br(nfp_prog, BR_UNC, end, 2);
1675 wrp_immed(nfp_prog, reg_both(0), 0);
1676 wrp_immed(nfp_prog, reg_both(1), 0);
1678 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_einval))
1681 wrp_immed(nfp_prog, reg_both(0), -22);
1682 wrp_immed(nfp_prog, reg_both(1), ~0);
1684 if (!nfp_prog_confirm_current_offset(nfp_prog, end))
1691 map_call_stack_common(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1697 /* We only have to reload LM0 if the key is not at start of stack */
1698 lm_off = nfp_prog->stack_depth;
1699 lm_off += meta->arg2.reg.var_off.value + meta->arg2.reg.off;
1700 load_lm_ptr = meta->arg2.var_off || lm_off;
1702 /* Set LM0 to start of key */
1704 emit_csr_wr(nfp_prog, reg_b(2 * 2), NFP_CSR_ACT_LM_ADDR0);
1705 if (meta->func_id == BPF_FUNC_map_update_elem)
1706 emit_csr_wr(nfp_prog, reg_b(3 * 2), NFP_CSR_ACT_LM_ADDR2);
1708 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO + meta->func_id,
1710 ret_tgt = nfp_prog_current_offset(nfp_prog) + 2;
1712 /* Load map ID into A0 */
1713 wrp_mov(nfp_prog, reg_a(0), reg_a(2));
1715 /* Load the return address into B0 */
1716 wrp_immed_relo(nfp_prog, reg_b(0), ret_tgt, RELO_IMMED_REL);
1718 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt))
1721 /* Reset the LM0 pointer */
1725 emit_csr_wr(nfp_prog, stack_reg(nfp_prog), NFP_CSR_ACT_LM_ADDR0);
1726 wrp_nops(nfp_prog, 3);
1732 nfp_get_prandom_u32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1734 __emit_csr_rd(nfp_prog, NFP_CSR_PSEUDO_RND_NUM);
1735 /* CSR value is read in following immed[gpr, 0] */
1736 emit_immed(nfp_prog, reg_both(0), 0,
1737 IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B);
1738 emit_immed(nfp_prog, reg_both(1), 0,
1739 IMMED_WIDTH_ALL, false, IMMED_SHIFT_0B);
1744 nfp_perf_event_output(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1749 ptr_type = ur_load_imm_any(nfp_prog, meta->arg1.type, imm_a(nfp_prog));
1751 ret_tgt = nfp_prog_current_offset(nfp_prog) + 3;
1753 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO + meta->func_id,
1756 /* Load ptr type into A1 */
1757 wrp_mov(nfp_prog, reg_a(1), ptr_type);
1759 /* Load the return address into B0 */
1760 wrp_immed_relo(nfp_prog, reg_b(0), ret_tgt, RELO_IMMED_REL);
1762 if (!nfp_prog_confirm_current_offset(nfp_prog, ret_tgt))
1769 nfp_queue_select(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1773 jmp_tgt = nfp_prog_current_offset(nfp_prog) + 5;
1775 /* Make sure the queue id fits into FW field */
1776 emit_alu(nfp_prog, reg_none(), reg_a(meta->insn.src_reg * 2),
1777 ALU_OP_AND_NOT_B, reg_imm(0xff));
1778 emit_br(nfp_prog, BR_BEQ, jmp_tgt, 2);
1780 /* Set the 'queue selected' bit and the queue value */
1781 emit_shf(nfp_prog, pv_qsel_set(nfp_prog),
1782 pv_qsel_set(nfp_prog), SHF_OP_OR, reg_imm(1),
1783 SHF_SC_L_SHF, PKT_VEL_QSEL_SET_BIT);
1784 emit_ld_field(nfp_prog,
1785 pv_qsel_val(nfp_prog), 0x1, reg_b(meta->insn.src_reg * 2),
1787 /* Delay slots end here, we will jump over next instruction if queue
1788 * value fits into the field.
1790 emit_ld_field(nfp_prog,
1791 pv_qsel_val(nfp_prog), 0x1, reg_imm(NFP_NET_RXR_MAX),
1794 if (!nfp_prog_confirm_current_offset(nfp_prog, jmp_tgt))
1800 /* --- Callbacks --- */
1801 static int mov_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1803 const struct bpf_insn *insn = &meta->insn;
1804 u8 dst = insn->dst_reg * 2;
1805 u8 src = insn->src_reg * 2;
1807 if (insn->src_reg == BPF_REG_10) {
1808 swreg stack_depth_reg;
1810 stack_depth_reg = ur_load_imm_any(nfp_prog,
1811 nfp_prog->stack_depth,
1812 stack_imm(nfp_prog));
1813 emit_alu(nfp_prog, reg_both(dst),
1814 stack_reg(nfp_prog), ALU_OP_ADD, stack_depth_reg);
1815 wrp_immed(nfp_prog, reg_both(dst + 1), 0);
1817 wrp_reg_mov(nfp_prog, dst, src);
1818 wrp_reg_mov(nfp_prog, dst + 1, src + 1);
1824 static int mov_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1826 u64 imm = meta->insn.imm; /* sign extend */
1828 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2), imm & ~0U);
1829 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), imm >> 32);
1834 static int xor_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1836 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_XOR);
1839 static int xor_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1841 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_XOR, !meta->insn.imm);
1844 static int and_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1846 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_AND);
1849 static int and_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1851 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm);
1854 static int or_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1856 return wrp_alu64_reg(nfp_prog, meta, ALU_OP_OR);
1859 static int or_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1861 return wrp_alu64_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm);
1864 static int add_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1866 const struct bpf_insn *insn = &meta->insn;
1868 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2),
1869 reg_a(insn->dst_reg * 2), ALU_OP_ADD,
1870 reg_b(insn->src_reg * 2));
1871 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1),
1872 reg_a(insn->dst_reg * 2 + 1), ALU_OP_ADD_C,
1873 reg_b(insn->src_reg * 2 + 1));
1878 static int add_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1880 const struct bpf_insn *insn = &meta->insn;
1881 u64 imm = insn->imm; /* sign extend */
1883 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_ADD, imm & ~0U);
1884 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_ADD_C, imm >> 32);
1889 static int sub_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1891 const struct bpf_insn *insn = &meta->insn;
1893 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2),
1894 reg_a(insn->dst_reg * 2), ALU_OP_SUB,
1895 reg_b(insn->src_reg * 2));
1896 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1),
1897 reg_a(insn->dst_reg * 2 + 1), ALU_OP_SUB_C,
1898 reg_b(insn->src_reg * 2 + 1));
1903 static int sub_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1905 const struct bpf_insn *insn = &meta->insn;
1906 u64 imm = insn->imm; /* sign extend */
1908 wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_SUB, imm & ~0U);
1909 wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_SUB_C, imm >> 32);
1914 static int mul_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1916 return wrp_mul(nfp_prog, meta, true, true);
1919 static int mul_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1921 return wrp_mul(nfp_prog, meta, true, false);
1924 static int div_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1926 const struct bpf_insn *insn = &meta->insn;
1928 return wrp_div_imm(nfp_prog, insn->dst_reg * 2, insn->imm);
1931 static int div_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1933 /* NOTE: verifier hook has rejected cases for which verifier doesn't
1934 * know whether the source operand is constant or not.
1936 return wrp_div_imm(nfp_prog, meta->insn.dst_reg * 2, meta->umin_src);
1939 static int neg_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1941 const struct bpf_insn *insn = &meta->insn;
1943 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2), reg_imm(0),
1944 ALU_OP_SUB, reg_b(insn->dst_reg * 2));
1945 emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1), reg_imm(0),
1946 ALU_OP_SUB_C, reg_b(insn->dst_reg * 2 + 1));
1952 * if shift_amt >= 32
1953 * dst_high = dst_low << shift_amt[4:0]
1956 * dst_high = (dst_high, dst_low) >> (32 - shift_amt)
1957 * dst_low = dst_low << shift_amt
1959 * The indirect shift will use the same logic at runtime.
1961 static int __shl_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt)
1963 if (shift_amt < 32) {
1964 emit_shf(nfp_prog, reg_both(dst + 1), reg_a(dst + 1),
1965 SHF_OP_NONE, reg_b(dst), SHF_SC_R_DSHF,
1967 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
1968 reg_b(dst), SHF_SC_L_SHF, shift_amt);
1969 } else if (shift_amt == 32) {
1970 wrp_reg_mov(nfp_prog, dst + 1, dst);
1971 wrp_immed(nfp_prog, reg_both(dst), 0);
1972 } else if (shift_amt > 32) {
1973 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
1974 reg_b(dst), SHF_SC_L_SHF, shift_amt - 32);
1975 wrp_immed(nfp_prog, reg_both(dst), 0);
1981 static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1983 const struct bpf_insn *insn = &meta->insn;
1984 u8 dst = insn->dst_reg * 2;
1986 return __shl_imm64(nfp_prog, dst, insn->imm);
1989 static void shl_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src)
1991 emit_alu(nfp_prog, imm_both(nfp_prog), reg_imm(32), ALU_OP_SUB,
1993 emit_alu(nfp_prog, reg_none(), imm_a(nfp_prog), ALU_OP_OR, reg_imm(0));
1994 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_a(dst + 1), SHF_OP_NONE,
1995 reg_b(dst), SHF_SC_R_DSHF);
1998 /* NOTE: for indirect left shift, HIGH part should be calculated first. */
1999 static void shl_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2001 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
2002 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
2003 reg_b(dst), SHF_SC_L_SHF);
2006 static void shl_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2008 shl_reg64_lt32_high(nfp_prog, dst, src);
2009 shl_reg64_lt32_low(nfp_prog, dst, src);
2012 static void shl_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2014 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
2015 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
2016 reg_b(dst), SHF_SC_L_SHF);
2017 wrp_immed(nfp_prog, reg_both(dst), 0);
2020 static int shl_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2022 const struct bpf_insn *insn = &meta->insn;
2026 dst = insn->dst_reg * 2;
2027 umin = meta->umin_src;
2028 umax = meta->umax_src;
2030 return __shl_imm64(nfp_prog, dst, umin);
2032 src = insn->src_reg * 2;
2034 shl_reg64_lt32(nfp_prog, dst, src);
2035 } else if (umin >= 32) {
2036 shl_reg64_ge32(nfp_prog, dst, src);
2038 /* Generate different instruction sequences depending on runtime
2039 * value of shift amount.
2041 u16 label_ge32, label_end;
2043 label_ge32 = nfp_prog_current_offset(nfp_prog) + 7;
2044 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0);
2046 shl_reg64_lt32_high(nfp_prog, dst, src);
2047 label_end = nfp_prog_current_offset(nfp_prog) + 6;
2048 emit_br(nfp_prog, BR_UNC, label_end, 2);
2049 /* shl_reg64_lt32_low packed in delay slot. */
2050 shl_reg64_lt32_low(nfp_prog, dst, src);
2052 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32))
2054 shl_reg64_ge32(nfp_prog, dst, src);
2056 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end))
2064 * if shift_amt >= 32
2066 * dst_low = dst_high >> shift_amt[4:0]
2068 * dst_high = dst_high >> shift_amt
2069 * dst_low = (dst_high, dst_low) >> shift_amt
2071 * The indirect shift will use the same logic at runtime.
2073 static int __shr_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt)
2075 if (shift_amt < 32) {
2076 emit_shf(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE,
2077 reg_b(dst), SHF_SC_R_DSHF, shift_amt);
2078 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
2079 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt);
2080 } else if (shift_amt == 32) {
2081 wrp_reg_mov(nfp_prog, dst, dst + 1);
2082 wrp_immed(nfp_prog, reg_both(dst + 1), 0);
2083 } else if (shift_amt > 32) {
2084 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
2085 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt - 32);
2086 wrp_immed(nfp_prog, reg_both(dst + 1), 0);
2092 static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2094 const struct bpf_insn *insn = &meta->insn;
2095 u8 dst = insn->dst_reg * 2;
2097 return __shr_imm64(nfp_prog, dst, insn->imm);
2100 /* NOTE: for indirect right shift, LOW part should be calculated first. */
2101 static void shr_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2103 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
2104 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_NONE,
2105 reg_b(dst + 1), SHF_SC_R_SHF);
2108 static void shr_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2110 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
2111 emit_shf_indir(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE,
2112 reg_b(dst), SHF_SC_R_DSHF);
2115 static void shr_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2117 shr_reg64_lt32_low(nfp_prog, dst, src);
2118 shr_reg64_lt32_high(nfp_prog, dst, src);
2121 static void shr_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2123 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_imm(0));
2124 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_NONE,
2125 reg_b(dst + 1), SHF_SC_R_SHF);
2126 wrp_immed(nfp_prog, reg_both(dst + 1), 0);
2129 static int shr_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2131 const struct bpf_insn *insn = &meta->insn;
2135 dst = insn->dst_reg * 2;
2136 umin = meta->umin_src;
2137 umax = meta->umax_src;
2139 return __shr_imm64(nfp_prog, dst, umin);
2141 src = insn->src_reg * 2;
2143 shr_reg64_lt32(nfp_prog, dst, src);
2144 } else if (umin >= 32) {
2145 shr_reg64_ge32(nfp_prog, dst, src);
2147 /* Generate different instruction sequences depending on runtime
2148 * value of shift amount.
2150 u16 label_ge32, label_end;
2152 label_ge32 = nfp_prog_current_offset(nfp_prog) + 6;
2153 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0);
2154 shr_reg64_lt32_low(nfp_prog, dst, src);
2155 label_end = nfp_prog_current_offset(nfp_prog) + 6;
2156 emit_br(nfp_prog, BR_UNC, label_end, 2);
2157 /* shr_reg64_lt32_high packed in delay slot. */
2158 shr_reg64_lt32_high(nfp_prog, dst, src);
2160 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32))
2162 shr_reg64_ge32(nfp_prog, dst, src);
2164 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end))
2171 /* Code logic is the same as __shr_imm64 except ashr requires signedness bit
2172 * told through PREV_ALU result.
2174 static int __ashr_imm64(struct nfp_prog *nfp_prog, u8 dst, u8 shift_amt)
2176 if (shift_amt < 32) {
2177 emit_shf(nfp_prog, reg_both(dst), reg_a(dst + 1), SHF_OP_NONE,
2178 reg_b(dst), SHF_SC_R_DSHF, shift_amt);
2179 /* Set signedness bit. */
2180 emit_alu(nfp_prog, reg_none(), reg_a(dst + 1), ALU_OP_OR,
2182 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR,
2183 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt);
2184 } else if (shift_amt == 32) {
2185 /* NOTE: this also helps setting signedness bit. */
2186 wrp_reg_mov(nfp_prog, dst, dst + 1);
2187 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR,
2188 reg_b(dst + 1), SHF_SC_R_SHF, 31);
2189 } else if (shift_amt > 32) {
2190 emit_alu(nfp_prog, reg_none(), reg_a(dst + 1), ALU_OP_OR,
2192 emit_shf(nfp_prog, reg_both(dst), reg_none(), SHF_OP_ASHR,
2193 reg_b(dst + 1), SHF_SC_R_SHF, shift_amt - 32);
2194 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR,
2195 reg_b(dst + 1), SHF_SC_R_SHF, 31);
2201 static int ashr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2203 const struct bpf_insn *insn = &meta->insn;
2204 u8 dst = insn->dst_reg * 2;
2206 return __ashr_imm64(nfp_prog, dst, insn->imm);
2209 static void ashr_reg64_lt32_high(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2211 /* NOTE: the first insn will set both indirect shift amount (source A)
2212 * and signedness bit (MSB of result).
2214 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_b(dst + 1));
2215 emit_shf_indir(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR,
2216 reg_b(dst + 1), SHF_SC_R_SHF);
2219 static void ashr_reg64_lt32_low(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2221 /* NOTE: it is the same as logic shift because we don't need to shift in
2222 * signedness bit when the shift amount is less than 32.
2224 return shr_reg64_lt32_low(nfp_prog, dst, src);
2227 static void ashr_reg64_lt32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2229 ashr_reg64_lt32_low(nfp_prog, dst, src);
2230 ashr_reg64_lt32_high(nfp_prog, dst, src);
2233 static void ashr_reg64_ge32(struct nfp_prog *nfp_prog, u8 dst, u8 src)
2235 emit_alu(nfp_prog, reg_none(), reg_a(src), ALU_OP_OR, reg_b(dst + 1));
2236 emit_shf_indir(nfp_prog, reg_both(dst), reg_none(), SHF_OP_ASHR,
2237 reg_b(dst + 1), SHF_SC_R_SHF);
2238 emit_shf(nfp_prog, reg_both(dst + 1), reg_none(), SHF_OP_ASHR,
2239 reg_b(dst + 1), SHF_SC_R_SHF, 31);
2242 /* Like ashr_imm64, but need to use indirect shift. */
2243 static int ashr_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2245 const struct bpf_insn *insn = &meta->insn;
2249 dst = insn->dst_reg * 2;
2250 umin = meta->umin_src;
2251 umax = meta->umax_src;
2253 return __ashr_imm64(nfp_prog, dst, umin);
2255 src = insn->src_reg * 2;
2257 ashr_reg64_lt32(nfp_prog, dst, src);
2258 } else if (umin >= 32) {
2259 ashr_reg64_ge32(nfp_prog, dst, src);
2261 u16 label_ge32, label_end;
2263 label_ge32 = nfp_prog_current_offset(nfp_prog) + 6;
2264 emit_br_bset(nfp_prog, reg_a(src), 5, label_ge32, 0);
2265 ashr_reg64_lt32_low(nfp_prog, dst, src);
2266 label_end = nfp_prog_current_offset(nfp_prog) + 6;
2267 emit_br(nfp_prog, BR_UNC, label_end, 2);
2268 /* ashr_reg64_lt32_high packed in delay slot. */
2269 ashr_reg64_lt32_high(nfp_prog, dst, src);
2271 if (!nfp_prog_confirm_current_offset(nfp_prog, label_ge32))
2273 ashr_reg64_ge32(nfp_prog, dst, src);
2275 if (!nfp_prog_confirm_current_offset(nfp_prog, label_end))
2282 static int mov_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2284 const struct bpf_insn *insn = &meta->insn;
2286 wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->src_reg * 2);
2287 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
2292 static int mov_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2294 const struct bpf_insn *insn = &meta->insn;
2296 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm);
2297 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
2302 static int xor_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2304 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_XOR);
2307 static int xor_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2309 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_XOR, !~meta->insn.imm);
2312 static int and_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2314 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_AND);
2317 static int and_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2319 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm);
2322 static int or_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2324 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_OR);
2327 static int or_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2329 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm);
2332 static int add_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2334 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_ADD);
2337 static int add_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2339 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_ADD, !meta->insn.imm);
2342 static int sub_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2344 return wrp_alu32_reg(nfp_prog, meta, ALU_OP_SUB);
2347 static int sub_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2349 return wrp_alu32_imm(nfp_prog, meta, ALU_OP_SUB, !meta->insn.imm);
2352 static int mul_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2354 return wrp_mul(nfp_prog, meta, false, true);
2357 static int mul_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2359 return wrp_mul(nfp_prog, meta, false, false);
2362 static int div_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2364 return div_reg64(nfp_prog, meta);
2367 static int div_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2369 return div_imm64(nfp_prog, meta);
2372 static int neg_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2374 u8 dst = meta->insn.dst_reg * 2;
2376 emit_alu(nfp_prog, reg_both(dst), reg_imm(0), ALU_OP_SUB, reg_b(dst));
2377 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
2382 static int shl_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2384 const struct bpf_insn *insn = &meta->insn;
2387 return 1; /* TODO: zero shift means indirect */
2389 emit_shf(nfp_prog, reg_both(insn->dst_reg * 2),
2390 reg_none(), SHF_OP_NONE, reg_b(insn->dst_reg * 2),
2391 SHF_SC_L_SHF, insn->imm);
2392 wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
2397 static int end_reg32(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2399 const struct bpf_insn *insn = &meta->insn;
2400 u8 gpr = insn->dst_reg * 2;
2402 switch (insn->imm) {
2404 emit_ld_field(nfp_prog, reg_both(gpr), 0x9, reg_b(gpr),
2406 emit_ld_field(nfp_prog, reg_both(gpr), 0xe, reg_a(gpr),
2409 wrp_immed(nfp_prog, reg_both(gpr + 1), 0);
2412 wrp_end32(nfp_prog, reg_a(gpr), gpr);
2413 wrp_immed(nfp_prog, reg_both(gpr + 1), 0);
2416 wrp_mov(nfp_prog, imm_a(nfp_prog), reg_b(gpr + 1));
2418 wrp_end32(nfp_prog, reg_a(gpr), gpr + 1);
2419 wrp_end32(nfp_prog, imm_a(nfp_prog), gpr);
2426 static int imm_ld8_part2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2428 struct nfp_insn_meta *prev = nfp_meta_prev(meta);
2432 dst = prev->insn.dst_reg * 2;
2433 imm_lo = prev->insn.imm;
2434 imm_hi = meta->insn.imm;
2436 wrp_immed(nfp_prog, reg_both(dst), imm_lo);
2438 /* mov is always 1 insn, load imm may be two, so try to use mov */
2439 if (imm_hi == imm_lo)
2440 wrp_mov(nfp_prog, reg_both(dst + 1), reg_a(dst));
2442 wrp_immed(nfp_prog, reg_both(dst + 1), imm_hi);
2447 static int imm_ld8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2449 meta->double_cb = imm_ld8_part2;
2453 static int data_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2455 return construct_data_ld(nfp_prog, meta->insn.imm, 1);
2458 static int data_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2460 return construct_data_ld(nfp_prog, meta->insn.imm, 2);
2463 static int data_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2465 return construct_data_ld(nfp_prog, meta->insn.imm, 4);
2468 static int data_ind_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2470 return construct_data_ind_ld(nfp_prog, meta->insn.imm,
2471 meta->insn.src_reg * 2, 1);
2474 static int data_ind_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2476 return construct_data_ind_ld(nfp_prog, meta->insn.imm,
2477 meta->insn.src_reg * 2, 2);
2480 static int data_ind_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2482 return construct_data_ind_ld(nfp_prog, meta->insn.imm,
2483 meta->insn.src_reg * 2, 4);
2487 mem_ldx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2488 unsigned int size, unsigned int ptr_off)
2490 return mem_op_stack(nfp_prog, meta, size, ptr_off,
2491 meta->insn.dst_reg * 2, meta->insn.src_reg * 2,
2492 true, wrp_lmem_load);
2495 static int mem_ldx_skb(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2498 swreg dst = reg_both(meta->insn.dst_reg * 2);
2500 switch (meta->insn.off) {
2501 case offsetof(struct __sk_buff, len):
2502 if (size != FIELD_SIZEOF(struct __sk_buff, len))
2504 wrp_mov(nfp_prog, dst, plen_reg(nfp_prog));
2506 case offsetof(struct __sk_buff, data):
2507 if (size != FIELD_SIZEOF(struct __sk_buff, data))
2509 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog));
2511 case offsetof(struct __sk_buff, data_end):
2512 if (size != FIELD_SIZEOF(struct __sk_buff, data_end))
2514 emit_alu(nfp_prog, dst,
2515 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog));
2521 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
2526 static int mem_ldx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2529 swreg dst = reg_both(meta->insn.dst_reg * 2);
2531 switch (meta->insn.off) {
2532 case offsetof(struct xdp_md, data):
2533 if (size != FIELD_SIZEOF(struct xdp_md, data))
2535 wrp_mov(nfp_prog, dst, pptr_reg(nfp_prog));
2537 case offsetof(struct xdp_md, data_end):
2538 if (size != FIELD_SIZEOF(struct xdp_md, data_end))
2540 emit_alu(nfp_prog, dst,
2541 plen_reg(nfp_prog), ALU_OP_ADD, pptr_reg(nfp_prog));
2547 wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
2553 mem_ldx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2558 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog));
2560 return data_ld_host_order_addr32(nfp_prog, meta->insn.src_reg * 2,
2561 tmp_reg, meta->insn.dst_reg * 2, size);
2565 mem_ldx_emem(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2570 tmp_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog));
2572 return data_ld_host_order_addr40(nfp_prog, meta->insn.src_reg * 2,
2573 tmp_reg, meta->insn.dst_reg * 2, size);
2577 mem_ldx_data_init_pktcache(struct nfp_prog *nfp_prog,
2578 struct nfp_insn_meta *meta)
2580 s16 range_start = meta->pkt_cache.range_start;
2581 s16 range_end = meta->pkt_cache.range_end;
2582 swreg src_base, off;
2586 off = re_load_imm_any(nfp_prog, range_start, imm_b(nfp_prog));
2587 src_base = reg_a(meta->insn.src_reg * 2);
2588 len = range_end - range_start;
2589 xfer_num = round_up(len, REG_WIDTH) / REG_WIDTH;
2591 indir = len > 8 * REG_WIDTH;
2592 /* Setup PREV_ALU for indirect mode. */
2594 wrp_immed(nfp_prog, reg_none(),
2595 CMD_OVE_LEN | FIELD_PREP(CMD_OV_LEN, xfer_num - 1));
2597 /* Cache memory into transfer-in registers. */
2598 emit_cmd_any(nfp_prog, CMD_TGT_READ32_SWAP, CMD_MODE_32b, 0, src_base,
2599 off, xfer_num - 1, CMD_CTX_SWAP, indir);
2603 mem_ldx_data_from_pktcache_unaligned(struct nfp_prog *nfp_prog,
2604 struct nfp_insn_meta *meta,
2607 s16 range_start = meta->pkt_cache.range_start;
2608 s16 insn_off = meta->insn.off - range_start;
2609 swreg dst_lo, dst_hi, src_lo, src_mid;
2610 u8 dst_gpr = meta->insn.dst_reg * 2;
2611 u8 len_lo = size, len_mid = 0;
2612 u8 idx = insn_off / REG_WIDTH;
2613 u8 off = insn_off % REG_WIDTH;
2615 dst_hi = reg_both(dst_gpr + 1);
2616 dst_lo = reg_both(dst_gpr);
2617 src_lo = reg_xfer(idx);
2619 /* The read length could involve as many as three registers. */
2620 if (size > REG_WIDTH - off) {
2621 /* Calculate the part in the second register. */
2622 len_lo = REG_WIDTH - off;
2623 len_mid = size - len_lo;
2625 /* Calculate the part in the third register. */
2626 if (size > 2 * REG_WIDTH - off)
2627 len_mid = REG_WIDTH;
2630 wrp_reg_subpart(nfp_prog, dst_lo, src_lo, len_lo, off);
2633 wrp_immed(nfp_prog, dst_hi, 0);
2637 src_mid = reg_xfer(idx + 1);
2639 if (size <= REG_WIDTH) {
2640 wrp_reg_or_subpart(nfp_prog, dst_lo, src_mid, len_mid, len_lo);
2641 wrp_immed(nfp_prog, dst_hi, 0);
2643 swreg src_hi = reg_xfer(idx + 2);
2645 wrp_reg_or_subpart(nfp_prog, dst_lo, src_mid,
2646 REG_WIDTH - len_lo, len_lo);
2647 wrp_reg_subpart(nfp_prog, dst_hi, src_mid, len_lo,
2648 REG_WIDTH - len_lo);
2649 wrp_reg_or_subpart(nfp_prog, dst_hi, src_hi, REG_WIDTH - len_lo,
2657 mem_ldx_data_from_pktcache_aligned(struct nfp_prog *nfp_prog,
2658 struct nfp_insn_meta *meta,
2661 swreg dst_lo, dst_hi, src_lo;
2664 idx = (meta->insn.off - meta->pkt_cache.range_start) / REG_WIDTH;
2665 dst_gpr = meta->insn.dst_reg * 2;
2666 dst_hi = reg_both(dst_gpr + 1);
2667 dst_lo = reg_both(dst_gpr);
2668 src_lo = reg_xfer(idx);
2670 if (size < REG_WIDTH) {
2671 wrp_reg_subpart(nfp_prog, dst_lo, src_lo, size, 0);
2672 wrp_immed(nfp_prog, dst_hi, 0);
2673 } else if (size == REG_WIDTH) {
2674 wrp_mov(nfp_prog, dst_lo, src_lo);
2675 wrp_immed(nfp_prog, dst_hi, 0);
2677 swreg src_hi = reg_xfer(idx + 1);
2679 wrp_mov(nfp_prog, dst_lo, src_lo);
2680 wrp_mov(nfp_prog, dst_hi, src_hi);
2687 mem_ldx_data_from_pktcache(struct nfp_prog *nfp_prog,
2688 struct nfp_insn_meta *meta, unsigned int size)
2690 u8 off = meta->insn.off - meta->pkt_cache.range_start;
2692 if (IS_ALIGNED(off, REG_WIDTH))
2693 return mem_ldx_data_from_pktcache_aligned(nfp_prog, meta, size);
2695 return mem_ldx_data_from_pktcache_unaligned(nfp_prog, meta, size);
2699 mem_ldx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2702 if (meta->ldst_gather_len)
2703 return nfp_cpp_memcpy(nfp_prog, meta);
2705 if (meta->ptr.type == PTR_TO_CTX) {
2706 if (nfp_prog->type == BPF_PROG_TYPE_XDP)
2707 return mem_ldx_xdp(nfp_prog, meta, size);
2709 return mem_ldx_skb(nfp_prog, meta, size);
2712 if (meta->ptr.type == PTR_TO_PACKET) {
2713 if (meta->pkt_cache.range_end) {
2714 if (meta->pkt_cache.do_init)
2715 mem_ldx_data_init_pktcache(nfp_prog, meta);
2717 return mem_ldx_data_from_pktcache(nfp_prog, meta, size);
2719 return mem_ldx_data(nfp_prog, meta, size);
2723 if (meta->ptr.type == PTR_TO_STACK)
2724 return mem_ldx_stack(nfp_prog, meta, size,
2725 meta->ptr.off + meta->ptr.var_off.value);
2727 if (meta->ptr.type == PTR_TO_MAP_VALUE)
2728 return mem_ldx_emem(nfp_prog, meta, size);
2733 static int mem_ldx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2735 return mem_ldx(nfp_prog, meta, 1);
2738 static int mem_ldx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2740 return mem_ldx(nfp_prog, meta, 2);
2743 static int mem_ldx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2745 return mem_ldx(nfp_prog, meta, 4);
2748 static int mem_ldx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2750 return mem_ldx(nfp_prog, meta, 8);
2754 mem_st_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2757 u64 imm = meta->insn.imm; /* sign extend */
2760 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog));
2762 return data_st_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg,
2766 static int mem_st(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2769 if (meta->ptr.type == PTR_TO_PACKET)
2770 return mem_st_data(nfp_prog, meta, size);
2775 static int mem_st1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2777 return mem_st(nfp_prog, meta, 1);
2780 static int mem_st2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2782 return mem_st(nfp_prog, meta, 2);
2785 static int mem_st4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2787 return mem_st(nfp_prog, meta, 4);
2790 static int mem_st8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2792 return mem_st(nfp_prog, meta, 8);
2796 mem_stx_data(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2801 off_reg = re_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog));
2803 return data_stx_host_order(nfp_prog, meta->insn.dst_reg * 2, off_reg,
2804 meta->insn.src_reg * 2, size);
2808 mem_stx_stack(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2809 unsigned int size, unsigned int ptr_off)
2811 return mem_op_stack(nfp_prog, meta, size, ptr_off,
2812 meta->insn.src_reg * 2, meta->insn.dst_reg * 2,
2813 false, wrp_lmem_store);
2816 static int mem_stx_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2818 switch (meta->insn.off) {
2819 case offsetof(struct xdp_md, rx_queue_index):
2820 return nfp_queue_select(nfp_prog, meta);
2823 WARN_ON_ONCE(1); /* verifier should have rejected bad accesses */
2828 mem_stx(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
2831 if (meta->ptr.type == PTR_TO_PACKET)
2832 return mem_stx_data(nfp_prog, meta, size);
2834 if (meta->ptr.type == PTR_TO_STACK)
2835 return mem_stx_stack(nfp_prog, meta, size,
2836 meta->ptr.off + meta->ptr.var_off.value);
2841 static int mem_stx1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2843 return mem_stx(nfp_prog, meta, 1);
2846 static int mem_stx2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2848 return mem_stx(nfp_prog, meta, 2);
2851 static int mem_stx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2853 if (meta->ptr.type == PTR_TO_CTX)
2854 if (nfp_prog->type == BPF_PROG_TYPE_XDP)
2855 return mem_stx_xdp(nfp_prog, meta);
2856 return mem_stx(nfp_prog, meta, 4);
2859 static int mem_stx8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2861 return mem_stx(nfp_prog, meta, 8);
2865 mem_xadd(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta, bool is64)
2867 u8 dst_gpr = meta->insn.dst_reg * 2;
2868 u8 src_gpr = meta->insn.src_reg * 2;
2869 unsigned int full_add, out;
2870 swreg addra, addrb, off;
2872 off = ur_load_imm_any(nfp_prog, meta->insn.off, imm_b(nfp_prog));
2874 /* We can fit 16 bits into command immediate, if we know the immediate
2875 * is guaranteed to either always or never fit into 16 bit we only
2876 * generate code to handle that particular case, otherwise generate
2879 out = nfp_prog_current_offset(nfp_prog);
2880 full_add = nfp_prog_current_offset(nfp_prog);
2882 if (meta->insn.off) {
2886 if (meta->xadd_maybe_16bit) {
2890 if (meta->xadd_over_16bit)
2892 if (meta->xadd_maybe_16bit && meta->xadd_over_16bit) {
2897 /* Generate the branch for choosing add_imm vs add */
2898 if (meta->xadd_maybe_16bit && meta->xadd_over_16bit) {
2899 swreg max_imm = imm_a(nfp_prog);
2901 wrp_immed(nfp_prog, max_imm, 0xffff);
2902 emit_alu(nfp_prog, reg_none(),
2903 max_imm, ALU_OP_SUB, reg_b(src_gpr));
2904 emit_alu(nfp_prog, reg_none(),
2905 reg_imm(0), ALU_OP_SUB_C, reg_b(src_gpr + 1));
2906 emit_br(nfp_prog, BR_BLO, full_add, meta->insn.off ? 2 : 0);
2910 /* If insn has an offset add to the address */
2911 if (!meta->insn.off) {
2912 addra = reg_a(dst_gpr);
2913 addrb = reg_b(dst_gpr + 1);
2915 emit_alu(nfp_prog, imma_a(nfp_prog),
2916 reg_a(dst_gpr), ALU_OP_ADD, off);
2917 emit_alu(nfp_prog, imma_b(nfp_prog),
2918 reg_a(dst_gpr + 1), ALU_OP_ADD_C, reg_imm(0));
2919 addra = imma_a(nfp_prog);
2920 addrb = imma_b(nfp_prog);
2923 /* Generate the add_imm if 16 bits are possible */
2924 if (meta->xadd_maybe_16bit) {
2925 swreg prev_alu = imm_a(nfp_prog);
2927 wrp_immed(nfp_prog, prev_alu,
2928 FIELD_PREP(CMD_OVE_DATA, 2) |
2930 FIELD_PREP(CMD_OV_LEN, 0x8 | is64 << 2));
2931 wrp_reg_or_subpart(nfp_prog, prev_alu, reg_b(src_gpr), 2, 2);
2932 emit_cmd_indir(nfp_prog, CMD_TGT_ADD_IMM, CMD_MODE_40b_BA, 0,
2933 addra, addrb, 0, CMD_CTX_NO_SWAP);
2935 if (meta->xadd_over_16bit)
2936 emit_br(nfp_prog, BR_UNC, out, 0);
2939 if (!nfp_prog_confirm_current_offset(nfp_prog, full_add))
2942 /* Generate the add if 16 bits are not guaranteed */
2943 if (meta->xadd_over_16bit) {
2944 emit_cmd(nfp_prog, CMD_TGT_ADD, CMD_MODE_40b_BA, 0,
2945 addra, addrb, is64 << 2,
2946 is64 ? CMD_CTX_SWAP_DEFER2 : CMD_CTX_SWAP_DEFER1);
2948 wrp_mov(nfp_prog, reg_xfer(0), reg_a(src_gpr));
2950 wrp_mov(nfp_prog, reg_xfer(1), reg_a(src_gpr + 1));
2953 if (!nfp_prog_confirm_current_offset(nfp_prog, out))
2959 static int mem_xadd4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2961 return mem_xadd(nfp_prog, meta, false);
2964 static int mem_xadd8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2966 return mem_xadd(nfp_prog, meta, true);
2969 static int jump(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2971 emit_br(nfp_prog, BR_UNC, meta->insn.off, 0);
2976 static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
2978 const struct bpf_insn *insn = &meta->insn;
2979 u64 imm = insn->imm; /* sign extend */
2980 swreg or1, or2, tmp_reg;
2982 or1 = reg_a(insn->dst_reg * 2);
2983 or2 = reg_b(insn->dst_reg * 2 + 1);
2986 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
2987 emit_alu(nfp_prog, imm_a(nfp_prog),
2988 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg);
2989 or1 = imm_a(nfp_prog);
2993 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
2994 emit_alu(nfp_prog, imm_b(nfp_prog),
2995 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg);
2996 or2 = imm_b(nfp_prog);
2999 emit_alu(nfp_prog, reg_none(), or1, ALU_OP_OR, or2);
3000 emit_br(nfp_prog, BR_BEQ, insn->off, 0);
3005 static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
3007 const struct bpf_insn *insn = &meta->insn;
3008 u64 imm = insn->imm; /* sign extend */
3017 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
3018 emit_alu(nfp_prog, reg_none(),
3019 reg_a(insn->dst_reg * 2), ALU_OP_AND, tmp_reg);
3020 emit_br(nfp_prog, BR_BNE, insn->off, 0);
3024 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
3025 emit_alu(nfp_prog, reg_none(),
3026 reg_a(insn->dst_reg * 2 + 1), ALU_OP_AND, tmp_reg);
3027 emit_br(nfp_prog, BR_BNE, insn->off, 0);
3033 static int jne_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
3035 const struct bpf_insn *insn = &meta->insn;
3036 u64 imm = insn->imm; /* sign extend */
3040 emit_alu(nfp_prog, reg_none(), reg_a(insn->dst_reg * 2),
3041 ALU_OP_OR, reg_b(insn->dst_reg * 2 + 1));
3042 emit_br(nfp_prog, BR_BNE, insn->off, 0);
3046 tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
3047 emit_alu(nfp_prog, reg_none(),
3048 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg);
3049 emit_br(nfp_prog, BR_BNE, insn->off, 0);
3051 tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
3052 emit_alu(nfp_prog, reg_none(),
3053 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg);
3054 emit_br(nfp_prog, BR_BNE, insn->off, 0);
3059 static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
3061 const struct bpf_insn *insn = &meta->insn;
3063 emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(insn->dst_reg * 2),
3064 ALU_OP_XOR, reg_b(insn->src_reg * 2));
3065 emit_alu(nfp_prog, imm_b(nfp_prog), reg_a(insn->dst_reg * 2 + 1),
3066 ALU_OP_XOR, reg_b(insn->src_reg * 2 + 1));
3067 emit_alu(nfp_prog, reg_none(),
3068 imm_a(nfp_prog), ALU_OP_OR, imm_b(nfp_prog));
3069 emit_br(nfp_prog, BR_BEQ, insn->off, 0);
3074 static int jset_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
3076 return wrp_test_reg(nfp_prog, meta, ALU_OP_AND, BR_BNE);
3079 static int jne_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
3081 return wrp_test_reg(nfp_prog, meta, ALU_OP_XOR, BR_BNE);
3084 static int call(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
3086 switch (meta->insn.imm) {
3087 case BPF_FUNC_xdp_adjust_head:
3088 return adjust_head(nfp_prog, meta);
3089 case BPF_FUNC_xdp_adjust_tail:
3090 return adjust_tail(nfp_prog, meta);
3091 case BPF_FUNC_map_lookup_elem:
3092 case BPF_FUNC_map_update_elem:
3093 case BPF_FUNC_map_delete_elem:
3094 return map_call_stack_common(nfp_prog, meta);
3095 case BPF_FUNC_get_prandom_u32:
3096 return nfp_get_prandom_u32(nfp_prog, meta);
3097 case BPF_FUNC_perf_event_output:
3098 return nfp_perf_event_output(nfp_prog, meta);
3100 WARN_ONCE(1, "verifier allowed unsupported function\n");
3105 static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
3107 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 0, RELO_BR_GO_OUT);
3112 static const instr_cb_t instr_cb[256] = {
3113 [BPF_ALU64 | BPF_MOV | BPF_X] = mov_reg64,
3114 [BPF_ALU64 | BPF_MOV | BPF_K] = mov_imm64,
3115 [BPF_ALU64 | BPF_XOR | BPF_X] = xor_reg64,
3116 [BPF_ALU64 | BPF_XOR | BPF_K] = xor_imm64,
3117 [BPF_ALU64 | BPF_AND | BPF_X] = and_reg64,
3118 [BPF_ALU64 | BPF_AND | BPF_K] = and_imm64,
3119 [BPF_ALU64 | BPF_OR | BPF_X] = or_reg64,
3120 [BPF_ALU64 | BPF_OR | BPF_K] = or_imm64,
3121 [BPF_ALU64 | BPF_ADD | BPF_X] = add_reg64,
3122 [BPF_ALU64 | BPF_ADD | BPF_K] = add_imm64,
3123 [BPF_ALU64 | BPF_SUB | BPF_X] = sub_reg64,
3124 [BPF_ALU64 | BPF_SUB | BPF_K] = sub_imm64,
3125 [BPF_ALU64 | BPF_MUL | BPF_X] = mul_reg64,
3126 [BPF_ALU64 | BPF_MUL | BPF_K] = mul_imm64,
3127 [BPF_ALU64 | BPF_DIV | BPF_X] = div_reg64,
3128 [BPF_ALU64 | BPF_DIV | BPF_K] = div_imm64,
3129 [BPF_ALU64 | BPF_NEG] = neg_reg64,
3130 [BPF_ALU64 | BPF_LSH | BPF_X] = shl_reg64,
3131 [BPF_ALU64 | BPF_LSH | BPF_K] = shl_imm64,
3132 [BPF_ALU64 | BPF_RSH | BPF_X] = shr_reg64,
3133 [BPF_ALU64 | BPF_RSH | BPF_K] = shr_imm64,
3134 [BPF_ALU64 | BPF_ARSH | BPF_X] = ashr_reg64,
3135 [BPF_ALU64 | BPF_ARSH | BPF_K] = ashr_imm64,
3136 [BPF_ALU | BPF_MOV | BPF_X] = mov_reg,
3137 [BPF_ALU | BPF_MOV | BPF_K] = mov_imm,
3138 [BPF_ALU | BPF_XOR | BPF_X] = xor_reg,
3139 [BPF_ALU | BPF_XOR | BPF_K] = xor_imm,
3140 [BPF_ALU | BPF_AND | BPF_X] = and_reg,
3141 [BPF_ALU | BPF_AND | BPF_K] = and_imm,
3142 [BPF_ALU | BPF_OR | BPF_X] = or_reg,
3143 [BPF_ALU | BPF_OR | BPF_K] = or_imm,
3144 [BPF_ALU | BPF_ADD | BPF_X] = add_reg,
3145 [BPF_ALU | BPF_ADD | BPF_K] = add_imm,
3146 [BPF_ALU | BPF_SUB | BPF_X] = sub_reg,
3147 [BPF_ALU | BPF_SUB | BPF_K] = sub_imm,
3148 [BPF_ALU | BPF_MUL | BPF_X] = mul_reg,
3149 [BPF_ALU | BPF_MUL | BPF_K] = mul_imm,
3150 [BPF_ALU | BPF_DIV | BPF_X] = div_reg,
3151 [BPF_ALU | BPF_DIV | BPF_K] = div_imm,
3152 [BPF_ALU | BPF_NEG] = neg_reg,
3153 [BPF_ALU | BPF_LSH | BPF_K] = shl_imm,
3154 [BPF_ALU | BPF_END | BPF_X] = end_reg32,
3155 [BPF_LD | BPF_IMM | BPF_DW] = imm_ld8,
3156 [BPF_LD | BPF_ABS | BPF_B] = data_ld1,
3157 [BPF_LD | BPF_ABS | BPF_H] = data_ld2,
3158 [BPF_LD | BPF_ABS | BPF_W] = data_ld4,
3159 [BPF_LD | BPF_IND | BPF_B] = data_ind_ld1,
3160 [BPF_LD | BPF_IND | BPF_H] = data_ind_ld2,
3161 [BPF_LD | BPF_IND | BPF_W] = data_ind_ld4,
3162 [BPF_LDX | BPF_MEM | BPF_B] = mem_ldx1,
3163 [BPF_LDX | BPF_MEM | BPF_H] = mem_ldx2,
3164 [BPF_LDX | BPF_MEM | BPF_W] = mem_ldx4,
3165 [BPF_LDX | BPF_MEM | BPF_DW] = mem_ldx8,
3166 [BPF_STX | BPF_MEM | BPF_B] = mem_stx1,
3167 [BPF_STX | BPF_MEM | BPF_H] = mem_stx2,
3168 [BPF_STX | BPF_MEM | BPF_W] = mem_stx4,
3169 [BPF_STX | BPF_MEM | BPF_DW] = mem_stx8,
3170 [BPF_STX | BPF_XADD | BPF_W] = mem_xadd4,
3171 [BPF_STX | BPF_XADD | BPF_DW] = mem_xadd8,
3172 [BPF_ST | BPF_MEM | BPF_B] = mem_st1,
3173 [BPF_ST | BPF_MEM | BPF_H] = mem_st2,
3174 [BPF_ST | BPF_MEM | BPF_W] = mem_st4,
3175 [BPF_ST | BPF_MEM | BPF_DW] = mem_st8,
3176 [BPF_JMP | BPF_JA | BPF_K] = jump,
3177 [BPF_JMP | BPF_JEQ | BPF_K] = jeq_imm,
3178 [BPF_JMP | BPF_JGT | BPF_K] = cmp_imm,
3179 [BPF_JMP | BPF_JGE | BPF_K] = cmp_imm,
3180 [BPF_JMP | BPF_JLT | BPF_K] = cmp_imm,
3181 [BPF_JMP | BPF_JLE | BPF_K] = cmp_imm,
3182 [BPF_JMP | BPF_JSGT | BPF_K] = cmp_imm,
3183 [BPF_JMP | BPF_JSGE | BPF_K] = cmp_imm,
3184 [BPF_JMP | BPF_JSLT | BPF_K] = cmp_imm,
3185 [BPF_JMP | BPF_JSLE | BPF_K] = cmp_imm,
3186 [BPF_JMP | BPF_JSET | BPF_K] = jset_imm,
3187 [BPF_JMP | BPF_JNE | BPF_K] = jne_imm,
3188 [BPF_JMP | BPF_JEQ | BPF_X] = jeq_reg,
3189 [BPF_JMP | BPF_JGT | BPF_X] = cmp_reg,
3190 [BPF_JMP | BPF_JGE | BPF_X] = cmp_reg,
3191 [BPF_JMP | BPF_JLT | BPF_X] = cmp_reg,
3192 [BPF_JMP | BPF_JLE | BPF_X] = cmp_reg,
3193 [BPF_JMP | BPF_JSGT | BPF_X] = cmp_reg,
3194 [BPF_JMP | BPF_JSGE | BPF_X] = cmp_reg,
3195 [BPF_JMP | BPF_JSLT | BPF_X] = cmp_reg,
3196 [BPF_JMP | BPF_JSLE | BPF_X] = cmp_reg,
3197 [BPF_JMP | BPF_JSET | BPF_X] = jset_reg,
3198 [BPF_JMP | BPF_JNE | BPF_X] = jne_reg,
3199 [BPF_JMP | BPF_CALL] = call,
3200 [BPF_JMP | BPF_EXIT] = goto_out,
3203 /* --- Assembler logic --- */
3204 static int nfp_fixup_branches(struct nfp_prog *nfp_prog)
3206 struct nfp_insn_meta *meta, *jmp_dst;
3209 list_for_each_entry(meta, &nfp_prog->insns, l) {
3212 if (meta->insn.code == (BPF_JMP | BPF_CALL))
3214 if (BPF_CLASS(meta->insn.code) != BPF_JMP)
3217 if (list_is_last(&meta->l, &nfp_prog->insns))
3218 br_idx = nfp_prog->last_bpf_off;
3220 br_idx = list_next_entry(meta, l)->off - 1;
3222 if (!nfp_is_br(nfp_prog->prog[br_idx])) {
3223 pr_err("Fixup found block not ending in branch %d %02x %016llx!!\n",
3224 br_idx, meta->insn.code, nfp_prog->prog[br_idx]);
3227 /* Leave special branches for later */
3228 if (FIELD_GET(OP_RELO_TYPE, nfp_prog->prog[br_idx]) !=
3232 if (!meta->jmp_dst) {
3233 pr_err("Non-exit jump doesn't have destination info recorded!!\n");
3237 jmp_dst = meta->jmp_dst;
3239 if (jmp_dst->skip) {
3240 pr_err("Branch landing on removed instruction!!\n");
3244 for (idx = meta->off; idx <= br_idx; idx++) {
3245 if (!nfp_is_br(nfp_prog->prog[idx]))
3247 br_set_offset(&nfp_prog->prog[idx], jmp_dst->off);
3254 static void nfp_intro(struct nfp_prog *nfp_prog)
3256 wrp_immed(nfp_prog, plen_reg(nfp_prog), GENMASK(13, 0));
3257 emit_alu(nfp_prog, plen_reg(nfp_prog),
3258 plen_reg(nfp_prog), ALU_OP_AND, pv_len(nfp_prog));
3261 static void nfp_outro_tc_da(struct nfp_prog *nfp_prog)
3263 /* TC direct-action mode:
3264 * 0,1 ok NOT SUPPORTED[1]
3265 * 2 drop 0x22 -> drop, count as stat1
3266 * 4,5 nuke 0x02 -> drop
3267 * 7 redir 0x44 -> redir, count as stat2
3268 * * unspec 0x11 -> pass, count as stat0
3270 * [1] We can't support OK and RECLASSIFY because we can't tell TC
3271 * the exact decision made. We are forced to support UNSPEC
3272 * to handle aborts so that's the only one we handle for passing
3273 * packets up the stack.
3275 /* Target for aborts */
3276 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog);
3278 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT);
3280 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS);
3281 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16);
3283 /* Target for normal exits */
3284 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog);
3286 /* if R0 > 7 jump to abort */
3287 emit_alu(nfp_prog, reg_none(), reg_imm(7), ALU_OP_SUB, reg_b(0));
3288 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0);
3289 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS);
3291 wrp_immed(nfp_prog, reg_b(2), 0x41221211);
3292 wrp_immed(nfp_prog, reg_b(3), 0x41001211);
3294 emit_shf(nfp_prog, reg_a(1),
3295 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 2);
3297 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0));
3298 emit_shf(nfp_prog, reg_a(2),
3299 reg_imm(0xf), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0);
3301 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0));
3302 emit_shf(nfp_prog, reg_b(2),
3303 reg_imm(0xf), SHF_OP_AND, reg_b(3), SHF_SC_R_SHF, 0);
3305 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT);
3307 emit_shf(nfp_prog, reg_b(2),
3308 reg_a(2), SHF_OP_OR, reg_b(2), SHF_SC_L_SHF, 4);
3309 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16);
3312 static void nfp_outro_xdp(struct nfp_prog *nfp_prog)
3314 /* XDP return codes:
3315 * 0 aborted 0x82 -> drop, count as stat3
3316 * 1 drop 0x22 -> drop, count as stat1
3317 * 2 pass 0x11 -> pass, count as stat0
3318 * 3 tx 0x44 -> redir, count as stat2
3319 * * unknown 0x82 -> drop, count as stat3
3321 /* Target for aborts */
3322 nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog);
3324 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT);
3326 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS);
3327 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x82), SHF_SC_L_SHF, 16);
3329 /* Target for normal exits */
3330 nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog);
3332 /* if R0 > 3 jump to abort */
3333 emit_alu(nfp_prog, reg_none(), reg_imm(3), ALU_OP_SUB, reg_b(0));
3334 emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0);
3336 wrp_immed(nfp_prog, reg_b(2), 0x44112282);
3338 emit_shf(nfp_prog, reg_a(1),
3339 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 3);
3341 emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0));
3342 emit_shf(nfp_prog, reg_b(2),
3343 reg_imm(0xff), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0);
3345 emit_br_relo(nfp_prog, BR_UNC, BR_OFF_RELO, 2, RELO_BR_NEXT_PKT);
3347 wrp_mov(nfp_prog, reg_a(0), NFP_BPF_ABI_FLAGS);
3348 emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16);
3351 static void nfp_outro(struct nfp_prog *nfp_prog)
3353 switch (nfp_prog->type) {
3354 case BPF_PROG_TYPE_SCHED_CLS:
3355 nfp_outro_tc_da(nfp_prog);
3357 case BPF_PROG_TYPE_XDP:
3358 nfp_outro_xdp(nfp_prog);
3365 static int nfp_translate(struct nfp_prog *nfp_prog)
3367 struct nfp_insn_meta *meta;
3370 nfp_intro(nfp_prog);
3371 if (nfp_prog->error)
3372 return nfp_prog->error;
3374 list_for_each_entry(meta, &nfp_prog->insns, l) {
3375 instr_cb_t cb = instr_cb[meta->insn.code];
3377 meta->off = nfp_prog_current_offset(nfp_prog);
3380 nfp_prog->n_translated++;
3384 if (nfp_meta_has_prev(nfp_prog, meta) &&
3385 nfp_meta_prev(meta)->double_cb)
3386 cb = nfp_meta_prev(meta)->double_cb;
3389 err = cb(nfp_prog, meta);
3392 if (nfp_prog->error)
3393 return nfp_prog->error;
3395 nfp_prog->n_translated++;
3398 nfp_prog->last_bpf_off = nfp_prog_current_offset(nfp_prog) - 1;
3400 nfp_outro(nfp_prog);
3401 if (nfp_prog->error)
3402 return nfp_prog->error;
3404 wrp_nops(nfp_prog, NFP_USTORE_PREFETCH_WINDOW);
3405 if (nfp_prog->error)
3406 return nfp_prog->error;
3408 return nfp_fixup_branches(nfp_prog);
3411 /* --- Optimizations --- */
3412 static void nfp_bpf_opt_reg_init(struct nfp_prog *nfp_prog)
3414 struct nfp_insn_meta *meta;
3416 list_for_each_entry(meta, &nfp_prog->insns, l) {
3417 struct bpf_insn insn = meta->insn;
3419 /* Programs converted from cBPF start with register xoring */
3420 if (insn.code == (BPF_ALU64 | BPF_XOR | BPF_X) &&
3421 insn.src_reg == insn.dst_reg)
3424 /* Programs start with R6 = R1 but we ignore the skb pointer */
3425 if (insn.code == (BPF_ALU64 | BPF_MOV | BPF_X) &&
3426 insn.src_reg == 1 && insn.dst_reg == 6)
3429 /* Return as soon as something doesn't match */
3435 /* abs(insn.imm) will fit better into unrestricted reg immediate -
3436 * convert add/sub of a negative number into a sub/add of a positive one.
3438 static void nfp_bpf_opt_neg_add_sub(struct nfp_prog *nfp_prog)
3440 struct nfp_insn_meta *meta;
3442 list_for_each_entry(meta, &nfp_prog->insns, l) {
3443 struct bpf_insn insn = meta->insn;
3448 if (BPF_CLASS(insn.code) != BPF_ALU &&
3449 BPF_CLASS(insn.code) != BPF_ALU64 &&
3450 BPF_CLASS(insn.code) != BPF_JMP)
3452 if (BPF_SRC(insn.code) != BPF_K)
3457 if (BPF_CLASS(insn.code) == BPF_JMP) {
3458 switch (BPF_OP(insn.code)) {
3463 meta->jump_neg_op = true;
3469 if (BPF_OP(insn.code) == BPF_ADD)
3470 insn.code = BPF_CLASS(insn.code) | BPF_SUB;
3471 else if (BPF_OP(insn.code) == BPF_SUB)
3472 insn.code = BPF_CLASS(insn.code) | BPF_ADD;
3476 meta->insn.code = insn.code | BPF_K;
3479 meta->insn.imm = -insn.imm;
3483 /* Remove masking after load since our load guarantees this is not needed */
3484 static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog)
3486 struct nfp_insn_meta *meta1, *meta2;
3487 const s32 exp_mask[] = {
3488 [BPF_B] = 0x000000ffU,
3489 [BPF_H] = 0x0000ffffU,
3490 [BPF_W] = 0xffffffffU,
3493 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) {
3494 struct bpf_insn insn, next;
3499 if (BPF_CLASS(insn.code) != BPF_LD)
3501 if (BPF_MODE(insn.code) != BPF_ABS &&
3502 BPF_MODE(insn.code) != BPF_IND)
3505 if (next.code != (BPF_ALU64 | BPF_AND | BPF_K))
3508 if (!exp_mask[BPF_SIZE(insn.code)])
3510 if (exp_mask[BPF_SIZE(insn.code)] != next.imm)
3513 if (next.src_reg || next.dst_reg)
3516 if (meta2->flags & FLAG_INSN_IS_JUMP_DST)
3523 static void nfp_bpf_opt_ld_shift(struct nfp_prog *nfp_prog)
3525 struct nfp_insn_meta *meta1, *meta2, *meta3;
3527 nfp_for_each_insn_walk3(nfp_prog, meta1, meta2, meta3) {
3528 struct bpf_insn insn, next1, next2;
3531 next1 = meta2->insn;
3532 next2 = meta3->insn;
3534 if (BPF_CLASS(insn.code) != BPF_LD)
3536 if (BPF_MODE(insn.code) != BPF_ABS &&
3537 BPF_MODE(insn.code) != BPF_IND)
3539 if (BPF_SIZE(insn.code) != BPF_W)
3542 if (!(next1.code == (BPF_LSH | BPF_K | BPF_ALU64) &&
3543 next2.code == (BPF_RSH | BPF_K | BPF_ALU64)) &&
3544 !(next1.code == (BPF_RSH | BPF_K | BPF_ALU64) &&
3545 next2.code == (BPF_LSH | BPF_K | BPF_ALU64)))
3548 if (next1.src_reg || next1.dst_reg ||
3549 next2.src_reg || next2.dst_reg)
3552 if (next1.imm != 0x20 || next2.imm != 0x20)
3555 if (meta2->flags & FLAG_INSN_IS_JUMP_DST ||
3556 meta3->flags & FLAG_INSN_IS_JUMP_DST)
3564 /* load/store pair that forms memory copy sould look like the following:
3566 * ld_width R, [addr_src + offset_src]
3567 * st_width [addr_dest + offset_dest], R
3569 * The destination register of load and source register of store should
3570 * be the same, load and store should also perform at the same width.
3571 * If either of addr_src or addr_dest is stack pointer, we don't do the
3572 * CPP optimization as stack is modelled by registers on NFP.
3575 curr_pair_is_memcpy(struct nfp_insn_meta *ld_meta,
3576 struct nfp_insn_meta *st_meta)
3578 struct bpf_insn *ld = &ld_meta->insn;
3579 struct bpf_insn *st = &st_meta->insn;
3581 if (!is_mbpf_load(ld_meta) || !is_mbpf_store(st_meta))
3584 if (ld_meta->ptr.type != PTR_TO_PACKET &&
3585 ld_meta->ptr.type != PTR_TO_MAP_VALUE)
3588 if (st_meta->ptr.type != PTR_TO_PACKET)
3591 if (BPF_SIZE(ld->code) != BPF_SIZE(st->code))
3594 if (ld->dst_reg != st->src_reg)
3597 /* There is jump to the store insn in this pair. */
3598 if (st_meta->flags & FLAG_INSN_IS_JUMP_DST)
3604 /* Currently, we only support chaining load/store pairs if:
3606 * - Their address base registers are the same.
3607 * - Their address offsets are in the same order.
3608 * - They operate at the same memory width.
3609 * - There is no jump into the middle of them.
3612 curr_pair_chain_with_previous(struct nfp_insn_meta *ld_meta,
3613 struct nfp_insn_meta *st_meta,
3614 struct bpf_insn *prev_ld,
3615 struct bpf_insn *prev_st)
3617 u8 prev_size, curr_size, prev_ld_base, prev_st_base, prev_ld_dst;
3618 struct bpf_insn *ld = &ld_meta->insn;
3619 struct bpf_insn *st = &st_meta->insn;
3620 s16 prev_ld_off, prev_st_off;
3622 /* This pair is the start pair. */
3626 prev_size = BPF_LDST_BYTES(prev_ld);
3627 curr_size = BPF_LDST_BYTES(ld);
3628 prev_ld_base = prev_ld->src_reg;
3629 prev_st_base = prev_st->dst_reg;
3630 prev_ld_dst = prev_ld->dst_reg;
3631 prev_ld_off = prev_ld->off;
3632 prev_st_off = prev_st->off;
3634 if (ld->dst_reg != prev_ld_dst)
3637 if (ld->src_reg != prev_ld_base || st->dst_reg != prev_st_base)
3640 if (curr_size != prev_size)
3643 /* There is jump to the head of this pair. */
3644 if (ld_meta->flags & FLAG_INSN_IS_JUMP_DST)
3647 /* Both in ascending order. */
3648 if (prev_ld_off + prev_size == ld->off &&
3649 prev_st_off + prev_size == st->off)
3652 /* Both in descending order. */
3653 if (ld->off + curr_size == prev_ld_off &&
3654 st->off + curr_size == prev_st_off)
3660 /* Return TRUE if cross memory access happens. Cross memory access means
3661 * store area is overlapping with load area that a later load might load
3662 * the value from previous store, for this case we can't treat the sequence
3663 * as an memory copy.
3666 cross_mem_access(struct bpf_insn *ld, struct nfp_insn_meta *head_ld_meta,
3667 struct nfp_insn_meta *head_st_meta)
3669 s16 head_ld_off, head_st_off, ld_off;
3671 /* Different pointer types does not overlap. */
3672 if (head_ld_meta->ptr.type != head_st_meta->ptr.type)
3675 /* load and store are both PTR_TO_PACKET, check ID info. */
3676 if (head_ld_meta->ptr.id != head_st_meta->ptr.id)
3679 /* Canonicalize the offsets. Turn all of them against the original
3682 head_ld_off = head_ld_meta->insn.off + head_ld_meta->ptr.off;
3683 head_st_off = head_st_meta->insn.off + head_st_meta->ptr.off;
3684 ld_off = ld->off + head_ld_meta->ptr.off;
3686 /* Ascending order cross. */
3687 if (ld_off > head_ld_off &&
3688 head_ld_off < head_st_off && ld_off >= head_st_off)
3691 /* Descending order cross. */
3692 if (ld_off < head_ld_off &&
3693 head_ld_off > head_st_off && ld_off <= head_st_off)
3699 /* This pass try to identify the following instructoin sequences.
3701 * load R, [regA + offA]
3702 * store [regB + offB], R
3703 * load R, [regA + offA + const_imm_A]
3704 * store [regB + offB + const_imm_A], R
3705 * load R, [regA + offA + 2 * const_imm_A]
3706 * store [regB + offB + 2 * const_imm_A], R
3709 * Above sequence is typically generated by compiler when lowering
3710 * memcpy. NFP prefer using CPP instructions to accelerate it.
3712 static void nfp_bpf_opt_ldst_gather(struct nfp_prog *nfp_prog)
3714 struct nfp_insn_meta *head_ld_meta = NULL;
3715 struct nfp_insn_meta *head_st_meta = NULL;
3716 struct nfp_insn_meta *meta1, *meta2;
3717 struct bpf_insn *prev_ld = NULL;
3718 struct bpf_insn *prev_st = NULL;
3721 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) {
3722 struct bpf_insn *ld = &meta1->insn;
3723 struct bpf_insn *st = &meta2->insn;
3725 /* Reset record status if any of the following if true:
3726 * - The current insn pair is not load/store.
3727 * - The load/store pair doesn't chain with previous one.
3728 * - The chained load/store pair crossed with previous pair.
3729 * - The chained load/store pair has a total size of memory
3730 * copy beyond 128 bytes which is the maximum length a
3731 * single NFP CPP command can transfer.
3733 if (!curr_pair_is_memcpy(meta1, meta2) ||
3734 !curr_pair_chain_with_previous(meta1, meta2, prev_ld,
3736 (head_ld_meta && (cross_mem_access(ld, head_ld_meta,
3738 head_ld_meta->ldst_gather_len >= 128))) {
3743 s16 prev_ld_off = prev_ld->off;
3744 s16 prev_st_off = prev_st->off;
3745 s16 head_ld_off = head_ld_meta->insn.off;
3747 if (prev_ld_off < head_ld_off) {
3748 head_ld_meta->insn.off = prev_ld_off;
3749 head_st_meta->insn.off = prev_st_off;
3750 head_ld_meta->ldst_gather_len =
3751 -head_ld_meta->ldst_gather_len;
3754 head_ld_meta->paired_st = &head_st_meta->insn;
3755 head_st_meta->skip = true;
3757 head_ld_meta->ldst_gather_len = 0;
3760 /* If the chain is ended by an load/store pair then this
3761 * could serve as the new head of the the next chain.
3763 if (curr_pair_is_memcpy(meta1, meta2)) {
3764 head_ld_meta = meta1;
3765 head_st_meta = meta2;
3766 head_ld_meta->ldst_gather_len =
3768 meta1 = nfp_meta_next(meta1);
3769 meta2 = nfp_meta_next(meta2);
3774 head_ld_meta = NULL;
3775 head_st_meta = NULL;
3784 if (!head_ld_meta) {
3785 head_ld_meta = meta1;
3786 head_st_meta = meta2;
3792 head_ld_meta->ldst_gather_len += BPF_LDST_BYTES(ld);
3793 meta1 = nfp_meta_next(meta1);
3794 meta2 = nfp_meta_next(meta2);
3801 static void nfp_bpf_opt_pkt_cache(struct nfp_prog *nfp_prog)
3803 struct nfp_insn_meta *meta, *range_node = NULL;
3804 s16 range_start = 0, range_end = 0;
3805 bool cache_avail = false;
3806 struct bpf_insn *insn;
3807 s32 range_ptr_off = 0;
3808 u32 range_ptr_id = 0;
3810 list_for_each_entry(meta, &nfp_prog->insns, l) {
3811 if (meta->flags & FLAG_INSN_IS_JUMP_DST)
3812 cache_avail = false;
3819 if (is_mbpf_store_pkt(meta) ||
3820 insn->code == (BPF_JMP | BPF_CALL) ||
3821 is_mbpf_classic_store_pkt(meta) ||
3822 is_mbpf_classic_load(meta)) {
3823 cache_avail = false;
3827 if (!is_mbpf_load(meta))
3830 if (meta->ptr.type != PTR_TO_PACKET || meta->ldst_gather_len) {
3831 cache_avail = false;
3838 goto end_current_then_start_new;
3842 /* Check ID to make sure two reads share the same
3843 * variable offset against PTR_TO_PACKET, and check OFF
3844 * to make sure they also share the same constant
3847 * OFFs don't really need to be the same, because they
3848 * are the constant offsets against PTR_TO_PACKET, so
3849 * for different OFFs, we could canonicalize them to
3850 * offsets against original packet pointer. We don't
3853 if (meta->ptr.id == range_ptr_id &&
3854 meta->ptr.off == range_ptr_off) {
3855 s16 new_start = range_start;
3856 s16 end, off = insn->off;
3857 s16 new_end = range_end;
3858 bool changed = false;
3860 if (off < range_start) {
3865 end = off + BPF_LDST_BYTES(insn);
3866 if (end > range_end) {
3874 if (new_end - new_start <= 64) {
3875 /* Install new range. */
3876 range_start = new_start;
3877 range_end = new_end;
3882 end_current_then_start_new:
3883 range_node->pkt_cache.range_start = range_start;
3884 range_node->pkt_cache.range_end = range_end;
3887 range_node->pkt_cache.do_init = true;
3888 range_ptr_id = range_node->ptr.id;
3889 range_ptr_off = range_node->ptr.off;
3890 range_start = insn->off;
3891 range_end = insn->off + BPF_LDST_BYTES(insn);
3895 range_node->pkt_cache.range_start = range_start;
3896 range_node->pkt_cache.range_end = range_end;
3899 list_for_each_entry(meta, &nfp_prog->insns, l) {
3903 if (is_mbpf_load_pkt(meta) && !meta->ldst_gather_len) {
3904 if (meta->pkt_cache.do_init) {
3905 range_start = meta->pkt_cache.range_start;
3906 range_end = meta->pkt_cache.range_end;
3908 meta->pkt_cache.range_start = range_start;
3909 meta->pkt_cache.range_end = range_end;
3915 static int nfp_bpf_optimize(struct nfp_prog *nfp_prog)
3917 nfp_bpf_opt_reg_init(nfp_prog);
3919 nfp_bpf_opt_neg_add_sub(nfp_prog);
3920 nfp_bpf_opt_ld_mask(nfp_prog);
3921 nfp_bpf_opt_ld_shift(nfp_prog);
3922 nfp_bpf_opt_ldst_gather(nfp_prog);
3923 nfp_bpf_opt_pkt_cache(nfp_prog);
3928 static int nfp_bpf_replace_map_ptrs(struct nfp_prog *nfp_prog)
3930 struct nfp_insn_meta *meta1, *meta2;
3931 struct nfp_bpf_map *nfp_map;
3932 struct bpf_map *map;
3935 nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) {
3936 if (meta1->skip || meta2->skip)
3939 if (meta1->insn.code != (BPF_LD | BPF_IMM | BPF_DW) ||
3940 meta1->insn.src_reg != BPF_PSEUDO_MAP_FD)
3943 map = (void *)(unsigned long)((u32)meta1->insn.imm |
3944 (u64)meta2->insn.imm << 32);
3945 if (bpf_map_offload_neutral(map)) {
3948 nfp_map = map_to_offmap(map)->dev_priv;
3952 meta1->insn.imm = id;
3953 meta2->insn.imm = 0;
3959 static int nfp_bpf_ustore_calc(u64 *prog, unsigned int len)
3961 __le64 *ustore = (__force __le64 *)prog;
3964 for (i = 0; i < len; i++) {
3967 err = nfp_ustore_check_valid_no_ecc(prog[i]);
3971 ustore[i] = cpu_to_le64(nfp_ustore_calc_ecc_insn(prog[i]));
3977 static void nfp_bpf_prog_trim(struct nfp_prog *nfp_prog)
3981 prog = kvmalloc_array(nfp_prog->prog_len, sizeof(u64), GFP_KERNEL);
3985 nfp_prog->__prog_alloc_len = nfp_prog->prog_len * sizeof(u64);
3986 memcpy(prog, nfp_prog->prog, nfp_prog->__prog_alloc_len);
3987 kvfree(nfp_prog->prog);
3988 nfp_prog->prog = prog;
3991 int nfp_bpf_jit(struct nfp_prog *nfp_prog)
3995 ret = nfp_bpf_replace_map_ptrs(nfp_prog);
3999 ret = nfp_bpf_optimize(nfp_prog);
4003 ret = nfp_translate(nfp_prog);
4005 pr_err("Translation failed with error %d (translated: %u)\n",
4006 ret, nfp_prog->n_translated);
4010 nfp_bpf_prog_trim(nfp_prog);
4015 void nfp_bpf_jit_prepare(struct nfp_prog *nfp_prog, unsigned int cnt)
4017 struct nfp_insn_meta *meta;
4019 /* Another pass to record jump information. */
4020 list_for_each_entry(meta, &nfp_prog->insns, l) {
4021 u64 code = meta->insn.code;
4023 if (BPF_CLASS(code) == BPF_JMP && BPF_OP(code) != BPF_EXIT &&
4024 BPF_OP(code) != BPF_CALL) {
4025 struct nfp_insn_meta *dst_meta;
4026 unsigned short dst_indx;
4028 dst_indx = meta->n + 1 + meta->insn.off;
4029 dst_meta = nfp_bpf_goto_meta(nfp_prog, meta, dst_indx,
4032 meta->jmp_dst = dst_meta;
4033 dst_meta->flags |= FLAG_INSN_IS_JUMP_DST;
4038 bool nfp_bpf_supported_opcode(u8 code)
4040 return !!instr_cb[code];
4043 void *nfp_bpf_relo_for_vnic(struct nfp_prog *nfp_prog, struct nfp_bpf_vnic *bv)
4049 prog = kmemdup(nfp_prog->prog, nfp_prog->prog_len * sizeof(u64),
4052 return ERR_PTR(-ENOMEM);
4054 for (i = 0; i < nfp_prog->prog_len; i++) {
4055 enum nfp_relo_type special;
4058 special = FIELD_GET(OP_RELO_TYPE, prog[i]);
4063 br_add_offset(&prog[i], bv->start_off);
4065 case RELO_BR_GO_OUT:
4066 br_set_offset(&prog[i],
4067 nfp_prog->tgt_out + bv->start_off);
4069 case RELO_BR_GO_ABORT:
4070 br_set_offset(&prog[i],
4071 nfp_prog->tgt_abort + bv->start_off);
4073 case RELO_BR_NEXT_PKT:
4074 br_set_offset(&prog[i], bv->tgt_done);
4076 case RELO_BR_HELPER:
4077 val = br_get_offset(prog[i]);
4080 case BPF_FUNC_map_lookup_elem:
4081 val = nfp_prog->bpf->helpers.map_lookup;
4083 case BPF_FUNC_map_update_elem:
4084 val = nfp_prog->bpf->helpers.map_update;
4086 case BPF_FUNC_map_delete_elem:
4087 val = nfp_prog->bpf->helpers.map_delete;
4089 case BPF_FUNC_perf_event_output:
4090 val = nfp_prog->bpf->helpers.perf_event_output;
4093 pr_err("relocation of unknown helper %d\n",
4098 br_set_offset(&prog[i], val);
4100 case RELO_IMMED_REL:
4101 immed_add_value(&prog[i], bv->start_off);
4105 prog[i] &= ~OP_RELO_TYPE;
4108 err = nfp_bpf_ustore_calc(prog, nfp_prog->prog_len);
4116 return ERR_PTR(err);