Linux Audio

Check our new training course

Loading...
v3.1
  1/* bpf_jit_comp.c : BPF JIT compiler
  2 *
  3 * Copyright (C) 2011 Eric Dumazet (eric.dumazet@gmail.com)
  4 *
  5 * This program is free software; you can redistribute it and/or
  6 * modify it under the terms of the GNU General Public License
  7 * as published by the Free Software Foundation; version 2
  8 * of the License.
  9 */
 10#include <linux/moduleloader.h>
 11#include <asm/cacheflush.h>
 12#include <linux/netdevice.h>
 13#include <linux/filter.h>
 
 
 
 
 
 
 
 
 
 14
 15/*
 16 * Conventions :
 17 *  EAX : BPF A accumulator
 18 *  EBX : BPF X accumulator
 19 *  RDI : pointer to skb   (first argument given to JIT function)
 20 *  RBP : frame pointer (even if CONFIG_FRAME_POINTER=n)
 21 *  ECX,EDX,ESI : scratch registers
 22 *  r9d : skb->len - skb->data_len (headlen)
 23 *  r8  : skb->data
 24 * -8(RBP) : saved RBX value
 25 * -16(RBP)..-80(RBP) : BPF_MEMWORDS values
 26 */
 27int bpf_jit_enable __read_mostly;
 28
 29/*
 30 * assembly code in arch/x86/net/bpf_jit.S
 31 */
 32extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[];
 33extern u8 sk_load_word_ind[], sk_load_half_ind[], sk_load_byte_ind[];
 34
 35static inline u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
 36{
 37	if (len == 1)
 38		*ptr = bytes;
 39	else if (len == 2)
 40		*(u16 *)ptr = bytes;
 41	else {
 42		*(u32 *)ptr = bytes;
 43		barrier();
 44	}
 45	return ptr + len;
 46}
 47
 48#define EMIT(bytes, len)	do { prog = emit_code(prog, bytes, len); } while (0)
 
 49
 50#define EMIT1(b1)		EMIT(b1, 1)
 51#define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
 52#define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
 53#define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
 54#define EMIT1_off32(b1, off)	do { EMIT1(b1); EMIT(off, 4);} while (0)
 55
 56#define CLEAR_A() EMIT2(0x31, 0xc0) /* xor %eax,%eax */
 57#define CLEAR_X() EMIT2(0x31, 0xdb) /* xor %ebx,%ebx */
 
 
 
 
 
 
 58
 59static inline bool is_imm8(int value)
 60{
 61	return value <= 127 && value >= -128;
 62}
 63
 64static inline bool is_near(int offset)
 65{
 66	return offset <= 127 && offset >= -128;
 67}
 68
 69#define EMIT_JMP(offset)						\
 70do {									\
 71	if (offset) {							\
 72		if (is_near(offset))					\
 73			EMIT2(0xeb, offset); /* jmp .+off8 */		\
 74		else							\
 75			EMIT1_off32(0xe9, offset); /* jmp .+off32 */	\
 76	}								\
 77} while (0)
 78
 79/* list of x86 cond jumps opcodes (. + s8)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 80 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
 81 */
 82#define X86_JB  0x72
 83#define X86_JAE 0x73
 84#define X86_JE  0x74
 85#define X86_JNE 0x75
 86#define X86_JBE 0x76
 87#define X86_JA  0x77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 88
 89#define EMIT_COND_JMP(op, offset)				\
 90do {								\
 91	if (is_near(offset))					\
 92		EMIT2(op, offset); /* jxx .+off8 */		\
 93	else {							\
 94		EMIT2(0x0f, op + 0x10);				\
 95		EMIT(offset, 4); /* jxx .+off32 */		\
 96	}							\
 97} while (0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 98
 99#define COND_SEL(CODE, TOP, FOP)	\
100	case CODE:			\
101		t_op = TOP;		\
102		f_op = FOP;		\
103		goto cond_branch
104
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
105
106#define SEEN_DATAREF 1 /* might call external helpers */
107#define SEEN_XREG    2 /* ebx is used */
108#define SEEN_MEM     4 /* use mem[] for temporary storage */
 
 
 
 
 
 
 
109
110static inline void bpf_flush_icache(void *start, void *end)
111{
112	mm_segment_t old_fs = get_fs();
113
114	set_fs(KERNEL_DS);
115	smp_wmb();
116	flush_icache_range((unsigned long)start, (unsigned long)end);
117	set_fs(old_fs);
 
 
 
 
 
118}
119
 
 
 
 
 
 
 
 
 
120
121void bpf_jit_compile(struct sk_filter *fp)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
122{
123	u8 temp[64];
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124	u8 *prog;
125	unsigned int proglen, oldproglen = 0;
126	int ilen, i;
127	int t_offset, f_offset;
128	u8 t_op, f_op, seen = 0, pass;
129	u8 *image = NULL;
130	u8 *func;
131	int pc_ret0 = -1; /* bpf index of first RET #0 instruction (if any) */
132	unsigned int cleanup_addr; /* epilogue code offset */
133	unsigned int *addrs;
134	const struct sock_filter *filter = fp->insns;
135	int flen = fp->len;
136
137	if (!bpf_jit_enable)
138		return;
139
140	addrs = kmalloc(flen * sizeof(*addrs), GFP_KERNEL);
141	if (addrs == NULL)
142		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
143
144	/* Before first pass, make a rough estimation of addrs[]
145	 * each bpf instruction is translated to less than 64 bytes
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
146	 */
147	for (proglen = 0, i = 0; i < flen; i++) {
148		proglen += 64;
149		addrs[i] = proglen;
 
 
 
 
 
 
150	}
151	cleanup_addr = proglen; /* epilogue address */
152
153	for (pass = 0; pass < 10; pass++) {
154		/* no prologue/epilogue for trivial filters (RET something) */
155		proglen = 0;
156		prog = temp;
 
157
158		if (seen) {
159			EMIT4(0x55, 0x48, 0x89, 0xe5); /* push %rbp; mov %rsp,%rbp */
160			EMIT4(0x48, 0x83, 0xec, 96);	/* subq  $96,%rsp	*/
161			/* note : must save %rbx in case bpf_error is hit */
162			if (seen & (SEEN_XREG | SEEN_DATAREF))
163				EMIT4(0x48, 0x89, 0x5d, 0xf8); /* mov %rbx, -8(%rbp) */
164			if (seen & SEEN_XREG)
165				CLEAR_X(); /* make sure we dont leek kernel memory */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
166
167			/*
168			 * If this filter needs to access skb data,
169			 * loads r9 and r8 with :
170			 *  r9 = skb->len - skb->data_len
171			 *  r8 = skb->data
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
172			 */
173			if (seen & SEEN_DATAREF) {
174				if (offsetof(struct sk_buff, len) <= 127)
175					/* mov    off8(%rdi),%r9d */
176					EMIT4(0x44, 0x8b, 0x4f, offsetof(struct sk_buff, len));
177				else {
178					/* mov    off32(%rdi),%r9d */
179					EMIT3(0x44, 0x8b, 0x8f);
180					EMIT(offsetof(struct sk_buff, len), 4);
181				}
182				if (is_imm8(offsetof(struct sk_buff, data_len)))
183					/* sub    off8(%rdi),%r9d */
184					EMIT4(0x44, 0x2b, 0x4f, offsetof(struct sk_buff, data_len));
185				else {
186					EMIT3(0x44, 0x2b, 0x8f);
187					EMIT(offsetof(struct sk_buff, data_len), 4);
188				}
189
190				if (is_imm8(offsetof(struct sk_buff, data)))
191					/* mov off8(%rdi),%r8 */
192					EMIT4(0x4c, 0x8b, 0x47, offsetof(struct sk_buff, data));
193				else {
194					/* mov off32(%rdi),%r8 */
195					EMIT3(0x4c, 0x8b, 0x87);
196					EMIT(offsetof(struct sk_buff, data), 4);
197				}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
198			}
199		}
200
201		switch (filter[0].code) {
202		case BPF_S_RET_K:
203		case BPF_S_LD_W_LEN:
204		case BPF_S_ANC_PROTOCOL:
205		case BPF_S_ANC_IFINDEX:
206		case BPF_S_ANC_MARK:
207		case BPF_S_ANC_RXHASH:
208		case BPF_S_ANC_CPU:
209		case BPF_S_ANC_QUEUE:
210		case BPF_S_LD_W_ABS:
211		case BPF_S_LD_H_ABS:
212		case BPF_S_LD_B_ABS:
213			/* first instruction sets A register (or is RET 'constant') */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
214			break;
215		default:
216			/* make sure we dont leak kernel information to user */
217			CLEAR_A(); /* A = 0 */
218		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
219
220		for (i = 0; i < flen; i++) {
221			unsigned int K = filter[i].k;
 
 
 
 
 
 
 
 
 
 
 
222
223			switch (filter[i].code) {
224			case BPF_S_ALU_ADD_X: /* A += X; */
225				seen |= SEEN_XREG;
226				EMIT2(0x01, 0xd8);		/* add %ebx,%eax */
227				break;
228			case BPF_S_ALU_ADD_K: /* A += K; */
229				if (!K)
230					break;
231				if (is_imm8(K))
232					EMIT3(0x83, 0xc0, K);	/* add imm8,%eax */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
233				else
234					EMIT1_off32(0x05, K);	/* add imm32,%eax */
235				break;
236			case BPF_S_ALU_SUB_X: /* A -= X; */
237				seen |= SEEN_XREG;
238				EMIT2(0x29, 0xd8);		/* sub    %ebx,%eax */
239				break;
240			case BPF_S_ALU_SUB_K: /* A -= K */
241				if (!K)
242					break;
243				if (is_imm8(K))
244					EMIT3(0x83, 0xe8, K); /* sub imm8,%eax */
245				else
246					EMIT1_off32(0x2d, K); /* sub imm32,%eax */
 
247				break;
248			case BPF_S_ALU_MUL_X: /* A *= X; */
249				seen |= SEEN_XREG;
250				EMIT3(0x0f, 0xaf, 0xc3);	/* imul %ebx,%eax */
251				break;
252			case BPF_S_ALU_MUL_K: /* A *= K */
253				if (is_imm8(K))
254					EMIT3(0x6b, 0xc0, K); /* imul imm8,%eax,%eax */
255				else {
256					EMIT2(0x69, 0xc0);		/* imul imm32,%eax */
257					EMIT(K, 4);
258				}
259				break;
260			case BPF_S_ALU_DIV_X: /* A /= X; */
261				seen |= SEEN_XREG;
262				EMIT2(0x85, 0xdb);	/* test %ebx,%ebx */
263				if (pc_ret0 != -1)
264					EMIT_COND_JMP(X86_JE, addrs[pc_ret0] - (addrs[i] - 4));
265				else {
266					EMIT_COND_JMP(X86_JNE, 2 + 5);
267					CLEAR_A();
268					EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 4)); /* jmp .+off32 */
269				}
270				EMIT4(0x31, 0xd2, 0xf7, 0xf3); /* xor %edx,%edx; div %ebx */
271				break;
272			case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */
273				EMIT3(0x48, 0x69, 0xc0); /* imul imm32,%rax,%rax */
274				EMIT(K, 4);
275				EMIT4(0x48, 0xc1, 0xe8, 0x20); /* shr $0x20,%rax */
276				break;
277			case BPF_S_ALU_AND_X:
278				seen |= SEEN_XREG;
279				EMIT2(0x21, 0xd8);		/* and %ebx,%eax */
280				break;
281			case BPF_S_ALU_AND_K:
282				if (K >= 0xFFFFFF00) {
283					EMIT2(0x24, K & 0xFF); /* and imm8,%al */
284				} else if (K >= 0xFFFF0000) {
285					EMIT2(0x66, 0x25);	/* and imm16,%ax */
286					EMIT2(K, 2);
287				} else {
288					EMIT1_off32(0x25, K);	/* and imm32,%eax */
289				}
290				break;
291			case BPF_S_ALU_OR_X:
292				seen |= SEEN_XREG;
293				EMIT2(0x09, 0xd8);		/* or %ebx,%eax */
294				break;
295			case BPF_S_ALU_OR_K:
296				if (is_imm8(K))
297					EMIT3(0x83, 0xc8, K); /* or imm8,%eax */
298				else
299					EMIT1_off32(0x0d, K);	/* or imm32,%eax */
300				break;
301			case BPF_S_ALU_LSH_X: /* A <<= X; */
302				seen |= SEEN_XREG;
303				EMIT4(0x89, 0xd9, 0xd3, 0xe0);	/* mov %ebx,%ecx; shl %cl,%eax */
304				break;
305			case BPF_S_ALU_LSH_K:
306				if (K == 0)
307					break;
308				else if (K == 1)
309					EMIT2(0xd1, 0xe0); /* shl %eax */
310				else
311					EMIT3(0xc1, 0xe0, K);
312				break;
313			case BPF_S_ALU_RSH_X: /* A >>= X; */
314				seen |= SEEN_XREG;
315				EMIT4(0x89, 0xd9, 0xd3, 0xe8);	/* mov %ebx,%ecx; shr %cl,%eax */
316				break;
317			case BPF_S_ALU_RSH_K: /* A >>= K; */
318				if (K == 0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
319					break;
320				else if (K == 1)
321					EMIT2(0xd1, 0xe8); /* shr %eax */
322				else
323					EMIT3(0xc1, 0xe8, K);
324				break;
325			case BPF_S_ALU_NEG:
326				EMIT2(0xf7, 0xd8);		/* neg %eax */
327				break;
328			case BPF_S_RET_K:
329				if (!K) {
330					if (pc_ret0 == -1)
331						pc_ret0 = i;
332					CLEAR_A();
333				} else {
334					EMIT1_off32(0xb8, K);	/* mov $imm32,%eax */
335				}
336				/* fallinto */
337			case BPF_S_RET_A:
338				if (seen) {
339					if (i != flen - 1) {
340						EMIT_JMP(cleanup_addr - addrs[i]);
341						break;
342					}
343					if (seen & SEEN_XREG)
344						EMIT4(0x48, 0x8b, 0x5d, 0xf8);  /* mov  -8(%rbp),%rbx */
345					EMIT1(0xc9);		/* leaveq */
346				}
347				EMIT1(0xc3);		/* ret */
348				break;
349			case BPF_S_MISC_TAX: /* X = A */
350				seen |= SEEN_XREG;
351				EMIT2(0x89, 0xc3);	/* mov    %eax,%ebx */
352				break;
353			case BPF_S_MISC_TXA: /* A = X */
354				seen |= SEEN_XREG;
355				EMIT2(0x89, 0xd8);	/* mov    %ebx,%eax */
356				break;
357			case BPF_S_LD_IMM: /* A = K */
358				if (!K)
359					CLEAR_A();
360				else
361					EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
362				break;
363			case BPF_S_LDX_IMM: /* X = K */
364				seen |= SEEN_XREG;
365				if (!K)
366					CLEAR_X();
367				else
368					EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */
369				break;
370			case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */
371				seen |= SEEN_MEM;
372				EMIT3(0x8b, 0x45, 0xf0 - K*4);
373				break;
374			case BPF_S_LDX_MEM: /* X = mem[K] : mov off8(%rbp),%ebx */
375				seen |= SEEN_XREG | SEEN_MEM;
376				EMIT3(0x8b, 0x5d, 0xf0 - K*4);
377				break;
378			case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */
379				seen |= SEEN_MEM;
380				EMIT3(0x89, 0x45, 0xf0 - K*4);
381				break;
382			case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */
383				seen |= SEEN_XREG | SEEN_MEM;
384				EMIT3(0x89, 0x5d, 0xf0 - K*4);
385				break;
386			case BPF_S_LD_W_LEN: /*	A = skb->len; */
387				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
388				if (is_imm8(offsetof(struct sk_buff, len)))
389					/* mov    off8(%rdi),%eax */
390					EMIT3(0x8b, 0x47, offsetof(struct sk_buff, len));
391				else {
392					EMIT2(0x8b, 0x87);
393					EMIT(offsetof(struct sk_buff, len), 4);
394				}
395				break;
396			case BPF_S_LDX_W_LEN: /* X = skb->len; */
397				seen |= SEEN_XREG;
398				if (is_imm8(offsetof(struct sk_buff, len)))
399					/* mov off8(%rdi),%ebx */
400					EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len));
401				else {
402					EMIT2(0x8b, 0x9f);
403					EMIT(offsetof(struct sk_buff, len), 4);
404				}
405				break;
406			case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
407				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
408				if (is_imm8(offsetof(struct sk_buff, protocol))) {
409					/* movzwl off8(%rdi),%eax */
410					EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, protocol));
411				} else {
412					EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
413					EMIT(offsetof(struct sk_buff, protocol), 4);
414				}
415				EMIT2(0x86, 0xc4); /* ntohs() : xchg   %al,%ah */
416				break;
417			case BPF_S_ANC_IFINDEX:
418				if (is_imm8(offsetof(struct sk_buff, dev))) {
419					/* movq off8(%rdi),%rax */
420					EMIT4(0x48, 0x8b, 0x47, offsetof(struct sk_buff, dev));
421				} else {
422					EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */
423					EMIT(offsetof(struct sk_buff, dev), 4);
424				}
425				EMIT3(0x48, 0x85, 0xc0);	/* test %rax,%rax */
426				EMIT_COND_JMP(X86_JE, cleanup_addr - (addrs[i] - 6));
427				BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
428				EMIT2(0x8b, 0x80);	/* mov off32(%rax),%eax */
429				EMIT(offsetof(struct net_device, ifindex), 4);
430				break;
431			case BPF_S_ANC_MARK:
432				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
433				if (is_imm8(offsetof(struct sk_buff, mark))) {
434					/* mov off8(%rdi),%eax */
435					EMIT3(0x8b, 0x47, offsetof(struct sk_buff, mark));
436				} else {
437					EMIT2(0x8b, 0x87);
438					EMIT(offsetof(struct sk_buff, mark), 4);
439				}
440				break;
441			case BPF_S_ANC_RXHASH:
442				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
443				if (is_imm8(offsetof(struct sk_buff, rxhash))) {
444					/* mov off8(%rdi),%eax */
445					EMIT3(0x8b, 0x47, offsetof(struct sk_buff, rxhash));
446				} else {
447					EMIT2(0x8b, 0x87);
448					EMIT(offsetof(struct sk_buff, rxhash), 4);
449				}
450				break;
451			case BPF_S_ANC_QUEUE:
452				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
453				if (is_imm8(offsetof(struct sk_buff, queue_mapping))) {
454					/* movzwl off8(%rdi),%eax */
455					EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, queue_mapping));
456				} else {
457					EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
458					EMIT(offsetof(struct sk_buff, queue_mapping), 4);
459				}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
460				break;
461			case BPF_S_ANC_CPU:
462#ifdef CONFIG_SMP
463				EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */
464				EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */
465#else
466				CLEAR_A();
467#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
468				break;
469			case BPF_S_LD_W_ABS:
470				func = sk_load_word;
471common_load:			seen |= SEEN_DATAREF;
472				if ((int)K < 0)
473					goto out;
474				t_offset = func - (image + addrs[i]);
475				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
476				EMIT1_off32(0xe8, t_offset); /* call */
477				break;
478			case BPF_S_LD_H_ABS:
479				func = sk_load_half;
480				goto common_load;
481			case BPF_S_LD_B_ABS:
482				func = sk_load_byte;
483				goto common_load;
484			case BPF_S_LDX_B_MSH:
485				if ((int)K < 0) {
486					if (pc_ret0 != -1) {
487						EMIT_JMP(addrs[pc_ret0] - addrs[i]);
488						break;
 
 
 
 
 
 
489					}
490					CLEAR_A();
491					EMIT_JMP(cleanup_addr - addrs[i]);
492					break;
493				}
494				seen |= SEEN_DATAREF | SEEN_XREG;
495				t_offset = sk_load_byte_msh - (image + addrs[i]);
496				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
497				EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */
498				break;
499			case BPF_S_LD_W_IND:
500				func = sk_load_word_ind;
501common_load_ind:		seen |= SEEN_DATAREF | SEEN_XREG;
502				t_offset = func - (image + addrs[i]);
503				EMIT1_off32(0xbe, K);	/* mov imm32,%esi   */
504				EMIT1_off32(0xe8, t_offset);	/* call sk_load_xxx_ind */
505				break;
506			case BPF_S_LD_H_IND:
507				func = sk_load_half_ind;
508				goto common_load_ind;
509			case BPF_S_LD_B_IND:
510				func = sk_load_byte_ind;
511				goto common_load_ind;
512			case BPF_S_JMP_JA:
513				t_offset = addrs[i + K] - addrs[i];
514				EMIT_JMP(t_offset);
515				break;
516			COND_SEL(BPF_S_JMP_JGT_K, X86_JA, X86_JBE);
517			COND_SEL(BPF_S_JMP_JGE_K, X86_JAE, X86_JB);
518			COND_SEL(BPF_S_JMP_JEQ_K, X86_JE, X86_JNE);
519			COND_SEL(BPF_S_JMP_JSET_K,X86_JNE, X86_JE);
520			COND_SEL(BPF_S_JMP_JGT_X, X86_JA, X86_JBE);
521			COND_SEL(BPF_S_JMP_JGE_X, X86_JAE, X86_JB);
522			COND_SEL(BPF_S_JMP_JEQ_X, X86_JE, X86_JNE);
523			COND_SEL(BPF_S_JMP_JSET_X,X86_JNE, X86_JE);
524
525cond_branch:			f_offset = addrs[i + filter[i].jf] - addrs[i];
526				t_offset = addrs[i + filter[i].jt] - addrs[i];
527
528				/* same targets, can avoid doing the test :) */
529				if (filter[i].jt == filter[i].jf) {
530					EMIT_JMP(t_offset);
531					break;
532				}
 
 
 
 
 
 
 
533
534				switch (filter[i].code) {
535				case BPF_S_JMP_JGT_X:
536				case BPF_S_JMP_JGE_X:
537				case BPF_S_JMP_JEQ_X:
538					seen |= SEEN_XREG;
539					EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */
540					break;
541				case BPF_S_JMP_JSET_X:
542					seen |= SEEN_XREG;
543					EMIT2(0x85, 0xd8); /* test %ebx,%eax */
544					break;
545				case BPF_S_JMP_JEQ_K:
546					if (K == 0) {
547						EMIT2(0x85, 0xc0); /* test   %eax,%eax */
548						break;
549					}
550				case BPF_S_JMP_JGT_K:
551				case BPF_S_JMP_JGE_K:
552					if (K <= 127)
553						EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */
554					else
555						EMIT1_off32(0x3d, K); /* cmp imm32,%eax */
556					break;
557				case BPF_S_JMP_JSET_K:
558					if (K <= 0xFF)
559						EMIT2(0xa8, K); /* test imm8,%al */
560					else if (!(K & 0xFFFF00FF))
561						EMIT3(0xf6, 0xc4, K >> 8); /* test imm8,%ah */
562					else if (K <= 0xFFFF) {
563						EMIT2(0x66, 0xa9); /* test imm16,%ax */
564						EMIT(K, 2);
565					} else {
566						EMIT1_off32(0xa9, K); /* test imm32,%eax */
 
 
567					}
568					break;
569				}
570				if (filter[i].jt != 0) {
571					if (filter[i].jf)
572						t_offset += is_near(f_offset) ? 2 : 6;
573					EMIT_COND_JMP(t_op, t_offset);
574					if (filter[i].jf)
575						EMIT_JMP(f_offset);
576					break;
577				}
578				EMIT_COND_JMP(f_op, f_offset);
579				break;
580			default:
581				/* hmm, too complex filter, give up with jit compiler */
582				goto out;
583			}
584			ilen = prog - temp;
585			if (image) {
586				if (unlikely(proglen + ilen > oldproglen)) {
587					pr_err("bpb_jit_compile fatal error\n");
588					kfree(addrs);
589					module_free(NULL, image);
590					return;
 
 
 
 
 
 
 
 
 
 
 
 
591				}
592				memcpy(image + proglen, temp, ilen);
 
 
 
 
 
 
 
 
 
 
 
 
593			}
594			proglen += ilen;
595			addrs[i] = proglen;
596			prog = temp;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
597		}
598		/* last bpf instruction is always a RET :
599		 * use it to give the cleanup instruction(s) addr
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
600		 */
601		cleanup_addr = proglen - 1; /* ret */
602		if (seen)
603			cleanup_addr -= 1; /* leaveq */
604		if (seen & SEEN_XREG)
605			cleanup_addr -= 4; /* mov  -8(%rbp),%rbx */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
606
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
607		if (image) {
608			WARN_ON(proglen != oldproglen);
 
 
 
 
609			break;
610		}
611		if (proglen == oldproglen) {
612			image = module_alloc(max_t(unsigned int,
613						   proglen,
614						   sizeof(struct work_struct)));
615			if (!image)
616				goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
617		}
618		oldproglen = proglen;
 
619	}
 
620	if (bpf_jit_enable > 1)
621		pr_err("flen=%d proglen=%u pass=%d image=%p\n",
622		       flen, proglen, pass, image);
623
624	if (image) {
625		if (bpf_jit_enable > 1)
626			print_hex_dump(KERN_ERR, "JIT code: ", DUMP_PREFIX_ADDRESS,
627				       16, 1, image, proglen, false);
628
629		bpf_flush_icache(image, image + proglen);
 
 
 
 
 
 
 
 
 
 
 
630
631		fp->bpf_func = (void *)image;
 
 
 
 
 
 
632	}
633out:
634	kfree(addrs);
635	return;
 
 
636}
637
638static void jit_free_defer(struct work_struct *arg)
639{
640	module_free(NULL, arg);
641}
642
643/* run from softirq, we must use a work_struct to call
644 * module_free() from process context
645 */
646void bpf_jit_free(struct sk_filter *fp)
647{
648	if (fp->bpf_func != sk_run_filter) {
649		struct work_struct *work = (struct work_struct *)fp->bpf_func;
650
651		INIT_WORK(work, jit_free_defer);
652		schedule_work(work);
653	}
654}
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * bpf_jit_comp.c: BPF JIT compiler
   4 *
   5 * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
   6 * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
 
 
   7 */
 
 
   8#include <linux/netdevice.h>
   9#include <linux/filter.h>
  10#include <linux/if_vlan.h>
  11#include <linux/bpf.h>
  12#include <linux/memory.h>
  13#include <linux/sort.h>
  14#include <asm/extable.h>
  15#include <asm/set_memory.h>
  16#include <asm/nospec-branch.h>
  17#include <asm/text-patching.h>
  18#include <asm/asm-prototypes.h>
  19
  20static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  21{
  22	if (len == 1)
  23		*ptr = bytes;
  24	else if (len == 2)
  25		*(u16 *)ptr = bytes;
  26	else {
  27		*(u32 *)ptr = bytes;
  28		barrier();
  29	}
  30	return ptr + len;
  31}
  32
  33#define EMIT(bytes, len) \
  34	do { prog = emit_code(prog, bytes, len); } while (0)
  35
  36#define EMIT1(b1)		EMIT(b1, 1)
  37#define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
  38#define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
  39#define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
 
  40
  41#define EMIT1_off32(b1, off) \
  42	do { EMIT1(b1); EMIT(off, 4); } while (0)
  43#define EMIT2_off32(b1, b2, off) \
  44	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
  45#define EMIT3_off32(b1, b2, b3, off) \
  46	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
  47#define EMIT4_off32(b1, b2, b3, b4, off) \
  48	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
  49
  50static bool is_imm8(int value)
  51{
  52	return value <= 127 && value >= -128;
  53}
  54
  55static bool is_simm32(s64 value)
  56{
  57	return value == (s64)(s32)value;
  58}
  59
  60static bool is_uimm32(u64 value)
  61{
  62	return value == (u64)(u32)value;
  63}
 
 
 
 
 
  64
  65/* mov dst, src */
  66#define EMIT_mov(DST, SRC)								 \
  67	do {										 \
  68		if (DST != SRC)								 \
  69			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
  70	} while (0)
  71
  72static int bpf_size_to_x86_bytes(int bpf_size)
  73{
  74	if (bpf_size == BPF_W)
  75		return 4;
  76	else if (bpf_size == BPF_H)
  77		return 2;
  78	else if (bpf_size == BPF_B)
  79		return 1;
  80	else if (bpf_size == BPF_DW)
  81		return 4; /* imm32 */
  82	else
  83		return 0;
  84}
  85
  86/*
  87 * List of x86 cond jumps opcodes (. + s8)
  88 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
  89 */
  90#define X86_JB  0x72
  91#define X86_JAE 0x73
  92#define X86_JE  0x74
  93#define X86_JNE 0x75
  94#define X86_JBE 0x76
  95#define X86_JA  0x77
  96#define X86_JL  0x7C
  97#define X86_JGE 0x7D
  98#define X86_JLE 0x7E
  99#define X86_JG  0x7F
 100
 101/* Pick a register outside of BPF range for JIT internal work */
 102#define AUX_REG (MAX_BPF_JIT_REG + 1)
 103#define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
 104
 105/*
 106 * The following table maps BPF registers to x86-64 registers.
 107 *
 108 * x86-64 register R12 is unused, since if used as base address
 109 * register in load/store instructions, it always needs an
 110 * extra byte of encoding and is callee saved.
 111 *
 112 * x86-64 register R9 is not used by BPF programs, but can be used by BPF
 113 * trampoline. x86-64 register R10 is used for blinding (if enabled).
 114 */
 115static const int reg2hex[] = {
 116	[BPF_REG_0] = 0,  /* RAX */
 117	[BPF_REG_1] = 7,  /* RDI */
 118	[BPF_REG_2] = 6,  /* RSI */
 119	[BPF_REG_3] = 2,  /* RDX */
 120	[BPF_REG_4] = 1,  /* RCX */
 121	[BPF_REG_5] = 0,  /* R8  */
 122	[BPF_REG_6] = 3,  /* RBX callee saved */
 123	[BPF_REG_7] = 5,  /* R13 callee saved */
 124	[BPF_REG_8] = 6,  /* R14 callee saved */
 125	[BPF_REG_9] = 7,  /* R15 callee saved */
 126	[BPF_REG_FP] = 5, /* RBP readonly */
 127	[BPF_REG_AX] = 2, /* R10 temp register */
 128	[AUX_REG] = 3,    /* R11 temp register */
 129	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
 130};
 131
 132static const int reg2pt_regs[] = {
 133	[BPF_REG_0] = offsetof(struct pt_regs, ax),
 134	[BPF_REG_1] = offsetof(struct pt_regs, di),
 135	[BPF_REG_2] = offsetof(struct pt_regs, si),
 136	[BPF_REG_3] = offsetof(struct pt_regs, dx),
 137	[BPF_REG_4] = offsetof(struct pt_regs, cx),
 138	[BPF_REG_5] = offsetof(struct pt_regs, r8),
 139	[BPF_REG_6] = offsetof(struct pt_regs, bx),
 140	[BPF_REG_7] = offsetof(struct pt_regs, r13),
 141	[BPF_REG_8] = offsetof(struct pt_regs, r14),
 142	[BPF_REG_9] = offsetof(struct pt_regs, r15),
 143};
 144
 145/*
 146 * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
 147 * which need extra byte of encoding.
 148 * rax,rcx,...,rbp have simpler encoding
 149 */
 150static bool is_ereg(u32 reg)
 151{
 152	return (1 << reg) & (BIT(BPF_REG_5) |
 153			     BIT(AUX_REG) |
 154			     BIT(BPF_REG_7) |
 155			     BIT(BPF_REG_8) |
 156			     BIT(BPF_REG_9) |
 157			     BIT(X86_REG_R9) |
 158			     BIT(BPF_REG_AX));
 159}
 160
 161/*
 162 * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
 163 * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
 164 * of encoding. al,cl,dl,bl have simpler encoding.
 165 */
 166static bool is_ereg_8l(u32 reg)
 167{
 168	return is_ereg(reg) ||
 169	    (1 << reg) & (BIT(BPF_REG_1) |
 170			  BIT(BPF_REG_2) |
 171			  BIT(BPF_REG_FP));
 172}
 173
 174static bool is_axreg(u32 reg)
 175{
 176	return reg == BPF_REG_0;
 177}
 178
 179/* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
 180static u8 add_1mod(u8 byte, u32 reg)
 181{
 182	if (is_ereg(reg))
 183		byte |= 1;
 184	return byte;
 185}
 186
 187static u8 add_2mod(u8 byte, u32 r1, u32 r2)
 188{
 189	if (is_ereg(r1))
 190		byte |= 1;
 191	if (is_ereg(r2))
 192		byte |= 4;
 193	return byte;
 194}
 195
 196/* Encode 'dst_reg' register into x86-64 opcode 'byte' */
 197static u8 add_1reg(u8 byte, u32 dst_reg)
 198{
 199	return byte + reg2hex[dst_reg];
 200}
 201
 202/* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
 203static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
 204{
 205	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
 206}
 207
 208/* Some 1-byte opcodes for binary ALU operations */
 209static u8 simple_alu_opcodes[] = {
 210	[BPF_ADD] = 0x01,
 211	[BPF_SUB] = 0x29,
 212	[BPF_AND] = 0x21,
 213	[BPF_OR] = 0x09,
 214	[BPF_XOR] = 0x31,
 215	[BPF_LSH] = 0xE0,
 216	[BPF_RSH] = 0xE8,
 217	[BPF_ARSH] = 0xF8,
 218};
 219
 220static void jit_fill_hole(void *area, unsigned int size)
 221{
 222	/* Fill whole space with INT3 instructions */
 223	memset(area, 0xcc, size);
 224}
 225
 226struct jit_context {
 227	int cleanup_addr; /* Epilogue code offset */
 228};
 229
 230/* Maximum number of bytes emitted while JITing one eBPF insn */
 231#define BPF_MAX_INSN_SIZE	128
 232#define BPF_INSN_SAFETY		64
 233
 234/* Number of bytes emit_patch() needs to generate instructions */
 235#define X86_PATCH_SIZE		5
 236/* Number of bytes that will be skipped on tailcall */
 237#define X86_TAIL_CALL_OFFSET	11
 238
 239static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
 240{
 241	u8 *prog = *pprog;
 242
 243	if (callee_regs_used[0])
 244		EMIT1(0x53);         /* push rbx */
 245	if (callee_regs_used[1])
 246		EMIT2(0x41, 0x55);   /* push r13 */
 247	if (callee_regs_used[2])
 248		EMIT2(0x41, 0x56);   /* push r14 */
 249	if (callee_regs_used[3])
 250		EMIT2(0x41, 0x57);   /* push r15 */
 251	*pprog = prog;
 252}
 253
 254static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
 255{
 256	u8 *prog = *pprog;
 257
 258	if (callee_regs_used[3])
 259		EMIT2(0x41, 0x5F);   /* pop r15 */
 260	if (callee_regs_used[2])
 261		EMIT2(0x41, 0x5E);   /* pop r14 */
 262	if (callee_regs_used[1])
 263		EMIT2(0x41, 0x5D);   /* pop r13 */
 264	if (callee_regs_used[0])
 265		EMIT1(0x5B);         /* pop rbx */
 266	*pprog = prog;
 267}
 268
 269/*
 270 * Emit x86-64 prologue code for BPF program.
 271 * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
 272 * while jumping to another program
 273 */
 274static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
 275			  bool tail_call_reachable, bool is_subprog)
 276{
 277	u8 *prog = *pprog;
 278
 279	/* BPF trampoline can be made to work without these nops,
 280	 * but let's waste 5 bytes for now and optimize later
 281	 */
 282	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
 283	prog += X86_PATCH_SIZE;
 284	if (!ebpf_from_cbpf) {
 285		if (tail_call_reachable && !is_subprog)
 286			EMIT2(0x31, 0xC0); /* xor eax, eax */
 287		else
 288			EMIT2(0x66, 0x90); /* nop2 */
 289	}
 290	EMIT1(0x55);             /* push rbp */
 291	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
 292	/* sub rsp, rounded_stack_depth */
 293	if (stack_depth)
 294		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
 295	if (tail_call_reachable)
 296		EMIT1(0x50);         /* push rax */
 297	*pprog = prog;
 298}
 299
 300static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
 301{
 302	u8 *prog = *pprog;
 303	s64 offset;
 304
 305	offset = func - (ip + X86_PATCH_SIZE);
 306	if (!is_simm32(offset)) {
 307		pr_err("Target call %p is out of range\n", func);
 308		return -ERANGE;
 309	}
 310	EMIT1_off32(opcode, offset);
 311	*pprog = prog;
 312	return 0;
 313}
 314
 315static int emit_call(u8 **pprog, void *func, void *ip)
 316{
 317	return emit_patch(pprog, func, ip, 0xE8);
 318}
 319
 320static int emit_jump(u8 **pprog, void *func, void *ip)
 321{
 322	return emit_patch(pprog, func, ip, 0xE9);
 323}
 324
 325static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
 326				void *old_addr, void *new_addr,
 327				const bool text_live)
 328{
 329	const u8 *nop_insn = x86_nops[5];
 330	u8 old_insn[X86_PATCH_SIZE];
 331	u8 new_insn[X86_PATCH_SIZE];
 332	u8 *prog;
 333	int ret;
 334
 335	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
 336	if (old_addr) {
 337		prog = old_insn;
 338		ret = t == BPF_MOD_CALL ?
 339		      emit_call(&prog, old_addr, ip) :
 340		      emit_jump(&prog, old_addr, ip);
 341		if (ret)
 342			return ret;
 343	}
 344
 345	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
 346	if (new_addr) {
 347		prog = new_insn;
 348		ret = t == BPF_MOD_CALL ?
 349		      emit_call(&prog, new_addr, ip) :
 350		      emit_jump(&prog, new_addr, ip);
 351		if (ret)
 352			return ret;
 353	}
 354
 355	ret = -EBUSY;
 356	mutex_lock(&text_mutex);
 357	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
 358		goto out;
 359	ret = 1;
 360	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
 361		if (text_live)
 362			text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
 363		else
 364			memcpy(ip, new_insn, X86_PATCH_SIZE);
 365		ret = 0;
 366	}
 367out:
 368	mutex_unlock(&text_mutex);
 369	return ret;
 370}
 371
 372int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
 373		       void *old_addr, void *new_addr)
 374{
 375	if (!is_kernel_text((long)ip) &&
 376	    !is_bpf_text_address((long)ip))
 377		/* BPF poking in modules is not supported */
 378		return -EINVAL;
 379
 380	return __bpf_arch_text_poke(ip, t, old_addr, new_addr, true);
 381}
 382
 383static int get_pop_bytes(bool *callee_regs_used)
 384{
 385	int bytes = 0;
 386
 387	if (callee_regs_used[3])
 388		bytes += 2;
 389	if (callee_regs_used[2])
 390		bytes += 2;
 391	if (callee_regs_used[1])
 392		bytes += 2;
 393	if (callee_regs_used[0])
 394		bytes += 1;
 395
 396	return bytes;
 397}
 398
 399/*
 400 * Generate the following code:
 401 *
 402 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
 403 *   if (index >= array->map.max_entries)
 404 *     goto out;
 405 *   if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
 406 *     goto out;
 407 *   prog = array->ptrs[index];
 408 *   if (prog == NULL)
 409 *     goto out;
 410 *   goto *(prog->bpf_func + prologue_size);
 411 * out:
 412 */
 413static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
 414					u32 stack_depth)
 415{
 416	int tcc_off = -4 - round_up(stack_depth, 8);
 417	u8 *prog = *pprog;
 418	int pop_bytes = 0;
 419	int off1 = 42;
 420	int off2 = 31;
 421	int off3 = 9;
 422
 423	/* count the additional bytes used for popping callee regs from stack
 424	 * that need to be taken into account for each of the offsets that
 425	 * are used for bailing out of the tail call
 426	 */
 427	pop_bytes = get_pop_bytes(callee_regs_used);
 428	off1 += pop_bytes;
 429	off2 += pop_bytes;
 430	off3 += pop_bytes;
 431
 432	if (stack_depth) {
 433		off1 += 7;
 434		off2 += 7;
 435		off3 += 7;
 436	}
 
 437
 438	/*
 439	 * rdi - pointer to ctx
 440	 * rsi - pointer to bpf_array
 441	 * rdx - index in bpf_array
 442	 */
 443
 444	/*
 445	 * if (index >= array->map.max_entries)
 446	 *	goto out;
 447	 */
 448	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
 449	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
 450	      offsetof(struct bpf_array, map.max_entries));
 451#define OFFSET1 (off1 + RETPOLINE_RCX_BPF_JIT_SIZE) /* Number of bytes to jump */
 452	EMIT2(X86_JBE, OFFSET1);                  /* jbe out */
 453
 454	/*
 455	 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
 456	 *	goto out;
 457	 */
 458	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
 459	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
 460#define OFFSET2 (off2 + RETPOLINE_RCX_BPF_JIT_SIZE)
 461	EMIT2(X86_JA, OFFSET2);                   /* ja out */
 462	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
 463	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
 464
 465	/* prog = array->ptrs[index]; */
 466	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
 467		    offsetof(struct bpf_array, ptrs));
 468
 469	/*
 470	 * if (prog == NULL)
 471	 *	goto out;
 472	 */
 473	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
 474#define OFFSET3 (off3 + RETPOLINE_RCX_BPF_JIT_SIZE)
 475	EMIT2(X86_JE, OFFSET3);                   /* je out */
 476
 477	*pprog = prog;
 478	pop_callee_regs(pprog, callee_regs_used);
 479	prog = *pprog;
 480
 481	EMIT1(0x58);                              /* pop rax */
 482	if (stack_depth)
 483		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
 484			    round_up(stack_depth, 8));
 485
 486	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
 487	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
 488	      offsetof(struct bpf_prog, bpf_func));
 489	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
 490	      X86_TAIL_CALL_OFFSET);
 491	/*
 492	 * Now we're ready to jump into next BPF program
 493	 * rdi == ctx (1st arg)
 494	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
 495	 */
 496	RETPOLINE_RCX_BPF_JIT();
 497
 498	/* out: */
 499	*pprog = prog;
 500}
 501
 502static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
 503				      u8 **pprog, int addr, u8 *image,
 504				      bool *callee_regs_used, u32 stack_depth)
 505{
 506	int tcc_off = -4 - round_up(stack_depth, 8);
 507	u8 *prog = *pprog;
 508	int pop_bytes = 0;
 509	int off1 = 20;
 510	int poke_off;
 511
 512	/* count the additional bytes used for popping callee regs to stack
 513	 * that need to be taken into account for jump offset that is used for
 514	 * bailing out from of the tail call when limit is reached
 515	 */
 516	pop_bytes = get_pop_bytes(callee_regs_used);
 517	off1 += pop_bytes;
 518
 519	/*
 520	 * total bytes for:
 521	 * - nop5/ jmpq $off
 522	 * - pop callee regs
 523	 * - sub rsp, $val if depth > 0
 524	 * - pop rax
 525	 */
 526	poke_off = X86_PATCH_SIZE + pop_bytes + 1;
 527	if (stack_depth) {
 528		poke_off += 7;
 529		off1 += 7;
 530	}
 531
 532	/*
 533	 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
 534	 *	goto out;
 535	 */
 536	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
 537	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
 538	EMIT2(X86_JA, off1);                          /* ja out */
 539	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
 540	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
 541
 542	poke->tailcall_bypass = image + (addr - poke_off - X86_PATCH_SIZE);
 543	poke->adj_off = X86_TAIL_CALL_OFFSET;
 544	poke->tailcall_target = image + (addr - X86_PATCH_SIZE);
 545	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
 546
 547	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
 548		  poke->tailcall_bypass);
 549
 550	*pprog = prog;
 551	pop_callee_regs(pprog, callee_regs_used);
 552	prog = *pprog;
 553	EMIT1(0x58);                                  /* pop rax */
 554	if (stack_depth)
 555		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
 556
 557	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
 558	prog += X86_PATCH_SIZE;
 559	/* out: */
 560
 561	*pprog = prog;
 562}
 563
 564static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
 565{
 566	struct bpf_jit_poke_descriptor *poke;
 567	struct bpf_array *array;
 568	struct bpf_prog *target;
 569	int i, ret;
 570
 571	for (i = 0; i < prog->aux->size_poke_tab; i++) {
 572		poke = &prog->aux->poke_tab[i];
 573		if (poke->aux && poke->aux != prog->aux)
 574			continue;
 575
 576		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
 577
 578		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
 579			continue;
 580
 581		array = container_of(poke->tail_call.map, struct bpf_array, map);
 582		mutex_lock(&array->aux->poke_mutex);
 583		target = array->ptrs[poke->tail_call.key];
 584		if (target) {
 585			/* Plain memcpy is used when image is not live yet
 586			 * and still not locked as read-only. Once poke
 587			 * location is active (poke->tailcall_target_stable),
 588			 * any parallel bpf_arch_text_poke() might occur
 589			 * still on the read-write image until we finally
 590			 * locked it as read-only. Both modifications on
 591			 * the given image are under text_mutex to avoid
 592			 * interference.
 593			 */
 594			ret = __bpf_arch_text_poke(poke->tailcall_target,
 595						   BPF_MOD_JUMP, NULL,
 596						   (u8 *)target->bpf_func +
 597						   poke->adj_off, false);
 598			BUG_ON(ret < 0);
 599			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
 600						   BPF_MOD_JUMP,
 601						   (u8 *)poke->tailcall_target +
 602						   X86_PATCH_SIZE, NULL, false);
 603			BUG_ON(ret < 0);
 604		}
 605		WRITE_ONCE(poke->tailcall_target_stable, true);
 606		mutex_unlock(&array->aux->poke_mutex);
 607	}
 608}
 
 609
 610static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
 611			   u32 dst_reg, const u32 imm32)
 612{
 613	u8 *prog = *pprog;
 614	u8 b1, b2, b3;
 615
 616	/*
 617	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
 618	 * (which zero-extends imm32) to save 2 bytes.
 619	 */
 620	if (sign_propagate && (s32)imm32 < 0) {
 621		/* 'mov %rax, imm32' sign extends imm32 */
 622		b1 = add_1mod(0x48, dst_reg);
 623		b2 = 0xC7;
 624		b3 = 0xC0;
 625		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
 626		goto done;
 627	}
 628
 629	/*
 630	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
 631	 * to save 3 bytes.
 632	 */
 633	if (imm32 == 0) {
 634		if (is_ereg(dst_reg))
 635			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
 636		b2 = 0x31; /* xor */
 637		b3 = 0xC0;
 638		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
 639		goto done;
 640	}
 641
 642	/* mov %eax, imm32 */
 643	if (is_ereg(dst_reg))
 644		EMIT1(add_1mod(0x40, dst_reg));
 645	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
 646done:
 647	*pprog = prog;
 648}
 649
 650static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
 651			   const u32 imm32_hi, const u32 imm32_lo)
 652{
 653	u8 *prog = *pprog;
 654
 655	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
 656		/*
 657		 * For emitting plain u32, where sign bit must not be
 658		 * propagated LLVM tends to load imm64 over mov32
 659		 * directly, so save couple of bytes by just doing
 660		 * 'mov %eax, imm32' instead.
 661		 */
 662		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
 663	} else {
 664		/* movabsq %rax, imm64 */
 665		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
 666		EMIT(imm32_lo, 4);
 667		EMIT(imm32_hi, 4);
 668	}
 669
 670	*pprog = prog;
 671}
 672
 673static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
 674{
 675	u8 *prog = *pprog;
 676
 677	if (is64) {
 678		/* mov dst, src */
 679		EMIT_mov(dst_reg, src_reg);
 680	} else {
 681		/* mov32 dst, src */
 682		if (is_ereg(dst_reg) || is_ereg(src_reg))
 683			EMIT1(add_2mod(0x40, dst_reg, src_reg));
 684		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
 685	}
 686
 687	*pprog = prog;
 688}
 689
 690/* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
 691static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
 692{
 693	u8 *prog = *pprog;
 694
 695	if (is_imm8(off)) {
 696		/* 1-byte signed displacement.
 697		 *
 698		 * If off == 0 we could skip this and save one extra byte, but
 699		 * special case of x86 R13 which always needs an offset is not
 700		 * worth the hassle
 701		 */
 702		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
 703	} else {
 704		/* 4-byte signed displacement */
 705		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
 706	}
 707	*pprog = prog;
 708}
 709
 710/*
 711 * Emit a REX byte if it will be necessary to address these registers
 712 */
 713static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
 714{
 715	u8 *prog = *pprog;
 716
 717	if (is64)
 718		EMIT1(add_2mod(0x48, dst_reg, src_reg));
 719	else if (is_ereg(dst_reg) || is_ereg(src_reg))
 720		EMIT1(add_2mod(0x40, dst_reg, src_reg));
 721	*pprog = prog;
 722}
 723
 724/* LDX: dst_reg = *(u8*)(src_reg + off) */
 725static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
 726{
 727	u8 *prog = *pprog;
 728
 729	switch (size) {
 730	case BPF_B:
 731		/* Emit 'movzx rax, byte ptr [rax + off]' */
 732		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
 733		break;
 734	case BPF_H:
 735		/* Emit 'movzx rax, word ptr [rax + off]' */
 736		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
 737		break;
 738	case BPF_W:
 739		/* Emit 'mov eax, dword ptr [rax+0x14]' */
 740		if (is_ereg(dst_reg) || is_ereg(src_reg))
 741			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
 742		else
 743			EMIT1(0x8B);
 744		break;
 745	case BPF_DW:
 746		/* Emit 'mov rax, qword ptr [rax+0x14]' */
 747		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
 748		break;
 749	}
 750	emit_insn_suffix(&prog, src_reg, dst_reg, off);
 751	*pprog = prog;
 752}
 753
 754/* STX: *(u8*)(dst_reg + off) = src_reg */
 755static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
 756{
 757	u8 *prog = *pprog;
 758
 759	switch (size) {
 760	case BPF_B:
 761		/* Emit 'mov byte ptr [rax + off], al' */
 762		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
 763			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
 764			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
 765		else
 766			EMIT1(0x88);
 767		break;
 768	case BPF_H:
 769		if (is_ereg(dst_reg) || is_ereg(src_reg))
 770			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
 771		else
 772			EMIT2(0x66, 0x89);
 773		break;
 774	case BPF_W:
 775		if (is_ereg(dst_reg) || is_ereg(src_reg))
 776			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
 777		else
 778			EMIT1(0x89);
 779		break;
 780	case BPF_DW:
 781		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
 782		break;
 783	}
 784	emit_insn_suffix(&prog, dst_reg, src_reg, off);
 785	*pprog = prog;
 786}
 787
 788static int emit_atomic(u8 **pprog, u8 atomic_op,
 789		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
 790{
 791	u8 *prog = *pprog;
 792
 793	EMIT1(0xF0); /* lock prefix */
 794
 795	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
 796
 797	/* emit opcode */
 798	switch (atomic_op) {
 799	case BPF_ADD:
 800	case BPF_SUB:
 801	case BPF_AND:
 802	case BPF_OR:
 803	case BPF_XOR:
 804		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
 805		EMIT1(simple_alu_opcodes[atomic_op]);
 806		break;
 807	case BPF_ADD | BPF_FETCH:
 808		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
 809		EMIT2(0x0F, 0xC1);
 810		break;
 811	case BPF_XCHG:
 812		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
 813		EMIT1(0x87);
 814		break;
 815	case BPF_CMPXCHG:
 816		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
 817		EMIT2(0x0F, 0xB1);
 818		break;
 819	default:
 820		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
 821		return -EFAULT;
 822	}
 823
 824	emit_insn_suffix(&prog, dst_reg, src_reg, off);
 825
 826	*pprog = prog;
 827	return 0;
 828}
 829
 830static bool ex_handler_bpf(const struct exception_table_entry *x,
 831			   struct pt_regs *regs, int trapnr,
 832			   unsigned long error_code, unsigned long fault_addr)
 833{
 834	u32 reg = x->fixup >> 8;
 835
 836	/* jump over faulting load and clear dest register */
 837	*(unsigned long *)((void *)regs + reg) = 0;
 838	regs->ip += x->fixup & 0xff;
 839	return true;
 840}
 841
 842static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
 843			     bool *regs_used, bool *tail_call_seen)
 844{
 845	int i;
 846
 847	for (i = 1; i <= insn_cnt; i++, insn++) {
 848		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
 849			*tail_call_seen = true;
 850		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
 851			regs_used[0] = true;
 852		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
 853			regs_used[1] = true;
 854		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
 855			regs_used[2] = true;
 856		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
 857			regs_used[3] = true;
 858	}
 859}
 860
 861static void emit_nops(u8 **pprog, int len)
 862{
 863	u8 *prog = *pprog;
 864	int i, noplen;
 865
 866	while (len > 0) {
 867		noplen = len;
 868
 869		if (noplen > ASM_NOP_MAX)
 870			noplen = ASM_NOP_MAX;
 871
 872		for (i = 0; i < noplen; i++)
 873			EMIT1(x86_nops[noplen][i]);
 874		len -= noplen;
 875	}
 876
 877	*pprog = prog;
 878}
 879
 880#define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
 881
 882static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
 883		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
 884{
 885	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
 886	struct bpf_insn *insn = bpf_prog->insnsi;
 887	bool callee_regs_used[4] = {};
 888	int insn_cnt = bpf_prog->len;
 889	bool tail_call_seen = false;
 890	bool seen_exit = false;
 891	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
 892	int i, excnt = 0;
 893	int ilen, proglen = 0;
 894	u8 *prog = temp;
 895	int err;
 896
 897	detect_reg_usage(insn, insn_cnt, callee_regs_used,
 898			 &tail_call_seen);
 899
 900	/* tail call's presence in current prog implies it is reachable */
 901	tail_call_reachable |= tail_call_seen;
 902
 903	emit_prologue(&prog, bpf_prog->aux->stack_depth,
 904		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
 905		      bpf_prog->aux->func_idx != 0);
 906	push_callee_regs(&prog, callee_regs_used);
 907
 908	ilen = prog - temp;
 909	if (image)
 910		memcpy(image + proglen, temp, ilen);
 911	proglen += ilen;
 912	addrs[0] = proglen;
 913	prog = temp;
 914
 915	for (i = 1; i <= insn_cnt; i++, insn++) {
 916		const s32 imm32 = insn->imm;
 917		u32 dst_reg = insn->dst_reg;
 918		u32 src_reg = insn->src_reg;
 919		u8 b2 = 0, b3 = 0;
 920		u8 *start_of_ldx;
 921		s64 jmp_offset;
 922		u8 jmp_cond;
 923		u8 *func;
 924		int nops;
 925
 926		switch (insn->code) {
 927			/* ALU */
 928		case BPF_ALU | BPF_ADD | BPF_X:
 929		case BPF_ALU | BPF_SUB | BPF_X:
 930		case BPF_ALU | BPF_AND | BPF_X:
 931		case BPF_ALU | BPF_OR | BPF_X:
 932		case BPF_ALU | BPF_XOR | BPF_X:
 933		case BPF_ALU64 | BPF_ADD | BPF_X:
 934		case BPF_ALU64 | BPF_SUB | BPF_X:
 935		case BPF_ALU64 | BPF_AND | BPF_X:
 936		case BPF_ALU64 | BPF_OR | BPF_X:
 937		case BPF_ALU64 | BPF_XOR | BPF_X:
 938			maybe_emit_mod(&prog, dst_reg, src_reg,
 939				       BPF_CLASS(insn->code) == BPF_ALU64);
 940			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
 941			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
 942			break;
 943
 944		case BPF_ALU64 | BPF_MOV | BPF_X:
 945		case BPF_ALU | BPF_MOV | BPF_X:
 946			emit_mov_reg(&prog,
 947				     BPF_CLASS(insn->code) == BPF_ALU64,
 948				     dst_reg, src_reg);
 949			break;
 950
 951			/* neg dst */
 952		case BPF_ALU | BPF_NEG:
 953		case BPF_ALU64 | BPF_NEG:
 954			if (BPF_CLASS(insn->code) == BPF_ALU64)
 955				EMIT1(add_1mod(0x48, dst_reg));
 956			else if (is_ereg(dst_reg))
 957				EMIT1(add_1mod(0x40, dst_reg));
 958			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
 959			break;
 960
 961		case BPF_ALU | BPF_ADD | BPF_K:
 962		case BPF_ALU | BPF_SUB | BPF_K:
 963		case BPF_ALU | BPF_AND | BPF_K:
 964		case BPF_ALU | BPF_OR | BPF_K:
 965		case BPF_ALU | BPF_XOR | BPF_K:
 966		case BPF_ALU64 | BPF_ADD | BPF_K:
 967		case BPF_ALU64 | BPF_SUB | BPF_K:
 968		case BPF_ALU64 | BPF_AND | BPF_K:
 969		case BPF_ALU64 | BPF_OR | BPF_K:
 970		case BPF_ALU64 | BPF_XOR | BPF_K:
 971			if (BPF_CLASS(insn->code) == BPF_ALU64)
 972				EMIT1(add_1mod(0x48, dst_reg));
 973			else if (is_ereg(dst_reg))
 974				EMIT1(add_1mod(0x40, dst_reg));
 975
 976			/*
 977			 * b3 holds 'normal' opcode, b2 short form only valid
 978			 * in case dst is eax/rax.
 979			 */
 980			switch (BPF_OP(insn->code)) {
 981			case BPF_ADD:
 982				b3 = 0xC0;
 983				b2 = 0x05;
 984				break;
 985			case BPF_SUB:
 986				b3 = 0xE8;
 987				b2 = 0x2D;
 988				break;
 989			case BPF_AND:
 990				b3 = 0xE0;
 991				b2 = 0x25;
 992				break;
 993			case BPF_OR:
 994				b3 = 0xC8;
 995				b2 = 0x0D;
 996				break;
 997			case BPF_XOR:
 998				b3 = 0xF0;
 999				b2 = 0x35;
1000				break;
1001			}
 
1002
1003			if (is_imm8(imm32))
1004				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1005			else if (is_axreg(dst_reg))
1006				EMIT1_off32(b2, imm32);
1007			else
1008				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1009			break;
1010
1011		case BPF_ALU64 | BPF_MOV | BPF_K:
1012		case BPF_ALU | BPF_MOV | BPF_K:
1013			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1014				       dst_reg, imm32);
1015			break;
1016
1017		case BPF_LD | BPF_IMM | BPF_DW:
1018			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1019			insn++;
1020			i++;
1021			break;
1022
1023			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1024		case BPF_ALU | BPF_MOD | BPF_X:
1025		case BPF_ALU | BPF_DIV | BPF_X:
1026		case BPF_ALU | BPF_MOD | BPF_K:
1027		case BPF_ALU | BPF_DIV | BPF_K:
1028		case BPF_ALU64 | BPF_MOD | BPF_X:
1029		case BPF_ALU64 | BPF_DIV | BPF_X:
1030		case BPF_ALU64 | BPF_MOD | BPF_K:
1031		case BPF_ALU64 | BPF_DIV | BPF_K:
1032			EMIT1(0x50); /* push rax */
1033			EMIT1(0x52); /* push rdx */
1034
1035			if (BPF_SRC(insn->code) == BPF_X)
1036				/* mov r11, src_reg */
1037				EMIT_mov(AUX_REG, src_reg);
1038			else
1039				/* mov r11, imm32 */
1040				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1041
1042			/* mov rax, dst_reg */
1043			EMIT_mov(BPF_REG_0, dst_reg);
1044
1045			/*
1046			 * xor edx, edx
1047			 * equivalent to 'xor rdx, rdx', but one byte less
1048			 */
1049			EMIT2(0x31, 0xd2);
1050
1051			if (BPF_CLASS(insn->code) == BPF_ALU64)
1052				/* div r11 */
1053				EMIT3(0x49, 0xF7, 0xF3);
1054			else
1055				/* div r11d */
1056				EMIT3(0x41, 0xF7, 0xF3);
1057
1058			if (BPF_OP(insn->code) == BPF_MOD)
1059				/* mov r11, rdx */
1060				EMIT3(0x49, 0x89, 0xD3);
1061			else
1062				/* mov r11, rax */
1063				EMIT3(0x49, 0x89, 0xC3);
1064
1065			EMIT1(0x5A); /* pop rdx */
1066			EMIT1(0x58); /* pop rax */
1067
1068			/* mov dst_reg, r11 */
1069			EMIT_mov(dst_reg, AUX_REG);
1070			break;
1071
1072		case BPF_ALU | BPF_MUL | BPF_K:
1073		case BPF_ALU | BPF_MUL | BPF_X:
1074		case BPF_ALU64 | BPF_MUL | BPF_K:
1075		case BPF_ALU64 | BPF_MUL | BPF_X:
1076		{
1077			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1078
1079			if (dst_reg != BPF_REG_0)
1080				EMIT1(0x50); /* push rax */
1081			if (dst_reg != BPF_REG_3)
1082				EMIT1(0x52); /* push rdx */
1083
1084			/* mov r11, dst_reg */
1085			EMIT_mov(AUX_REG, dst_reg);
1086
1087			if (BPF_SRC(insn->code) == BPF_X)
1088				emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
1089			else
1090				emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
1091
1092			if (is64)
1093				EMIT1(add_1mod(0x48, AUX_REG));
1094			else if (is_ereg(AUX_REG))
1095				EMIT1(add_1mod(0x40, AUX_REG));
1096			/* mul(q) r11 */
1097			EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
1098
1099			if (dst_reg != BPF_REG_3)
1100				EMIT1(0x5A); /* pop rdx */
1101			if (dst_reg != BPF_REG_0) {
1102				/* mov dst_reg, rax */
1103				EMIT_mov(dst_reg, BPF_REG_0);
1104				EMIT1(0x58); /* pop rax */
1105			}
1106			break;
 
 
 
1107		}
1108			/* Shifts */
1109		case BPF_ALU | BPF_LSH | BPF_K:
1110		case BPF_ALU | BPF_RSH | BPF_K:
1111		case BPF_ALU | BPF_ARSH | BPF_K:
1112		case BPF_ALU64 | BPF_LSH | BPF_K:
1113		case BPF_ALU64 | BPF_RSH | BPF_K:
1114		case BPF_ALU64 | BPF_ARSH | BPF_K:
1115			if (BPF_CLASS(insn->code) == BPF_ALU64)
1116				EMIT1(add_1mod(0x48, dst_reg));
1117			else if (is_ereg(dst_reg))
1118				EMIT1(add_1mod(0x40, dst_reg));
1119
1120			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1121			if (imm32 == 1)
1122				EMIT2(0xD1, add_1reg(b3, dst_reg));
1123			else
1124				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1125			break;
1126
1127		case BPF_ALU | BPF_LSH | BPF_X:
1128		case BPF_ALU | BPF_RSH | BPF_X:
1129		case BPF_ALU | BPF_ARSH | BPF_X:
1130		case BPF_ALU64 | BPF_LSH | BPF_X:
1131		case BPF_ALU64 | BPF_RSH | BPF_X:
1132		case BPF_ALU64 | BPF_ARSH | BPF_X:
1133
1134			/* Check for bad case when dst_reg == rcx */
1135			if (dst_reg == BPF_REG_4) {
1136				/* mov r11, dst_reg */
1137				EMIT_mov(AUX_REG, dst_reg);
1138				dst_reg = AUX_REG;
1139			}
1140
1141			if (src_reg != BPF_REG_4) { /* common case */
1142				EMIT1(0x51); /* push rcx */
1143
1144				/* mov rcx, src_reg */
1145				EMIT_mov(BPF_REG_4, src_reg);
1146			}
1147
1148			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1149			if (BPF_CLASS(insn->code) == BPF_ALU64)
1150				EMIT1(add_1mod(0x48, dst_reg));
1151			else if (is_ereg(dst_reg))
1152				EMIT1(add_1mod(0x40, dst_reg));
1153
1154			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1155			EMIT2(0xD3, add_1reg(b3, dst_reg));
1156
1157			if (src_reg != BPF_REG_4)
1158				EMIT1(0x59); /* pop rcx */
1159
1160			if (insn->dst_reg == BPF_REG_4)
1161				/* mov dst_reg, r11 */
1162				EMIT_mov(insn->dst_reg, AUX_REG);
1163			break;
1164
1165		case BPF_ALU | BPF_END | BPF_FROM_BE:
1166			switch (imm32) {
1167			case 16:
1168				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1169				EMIT1(0x66);
1170				if (is_ereg(dst_reg))
1171					EMIT1(0x41);
1172				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1173
1174				/* Emit 'movzwl eax, ax' */
1175				if (is_ereg(dst_reg))
1176					EMIT3(0x45, 0x0F, 0xB7);
1177				else
1178					EMIT2(0x0F, 0xB7);
1179				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
 
 
 
1180				break;
1181			case 32:
1182				/* Emit 'bswap eax' to swap lower 4 bytes */
1183				if (is_ereg(dst_reg))
1184					EMIT2(0x41, 0x0F);
 
1185				else
1186					EMIT1(0x0F);
1187				EMIT1(add_1reg(0xC8, dst_reg));
1188				break;
1189			case 64:
1190				/* Emit 'bswap rax' to swap 8 bytes */
1191				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1192				      add_1reg(0xC8, dst_reg));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1193				break;
1194			}
1195			break;
1196
1197		case BPF_ALU | BPF_END | BPF_FROM_LE:
1198			switch (imm32) {
1199			case 16:
1200				/*
1201				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1202				 * into 64 bit
1203				 */
1204				if (is_ereg(dst_reg))
1205					EMIT3(0x45, 0x0F, 0xB7);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1206				else
1207					EMIT2(0x0F, 0xB7);
1208				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
 
 
 
1209				break;
1210			case 32:
1211				/* Emit 'mov eax, eax' to clear upper 32-bits */
1212				if (is_ereg(dst_reg))
1213					EMIT1(0x45);
1214				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
 
 
1215				break;
1216			case 64:
1217				/* nop */
 
1218				break;
1219			}
1220			break;
1221
1222			/* speculation barrier */
1223		case BPF_ST | BPF_NOSPEC:
1224			if (boot_cpu_has(X86_FEATURE_XMM2))
1225				/* Emit 'lfence' */
1226				EMIT3(0x0F, 0xAE, 0xE8);
1227			break;
1228
1229			/* ST: *(u8*)(dst_reg + off) = imm */
1230		case BPF_ST | BPF_MEM | BPF_B:
1231			if (is_ereg(dst_reg))
1232				EMIT2(0x41, 0xC6);
1233			else
1234				EMIT1(0xC6);
1235			goto st;
1236		case BPF_ST | BPF_MEM | BPF_H:
1237			if (is_ereg(dst_reg))
1238				EMIT3(0x66, 0x41, 0xC7);
1239			else
1240				EMIT2(0x66, 0xC7);
1241			goto st;
1242		case BPF_ST | BPF_MEM | BPF_W:
1243			if (is_ereg(dst_reg))
1244				EMIT2(0x41, 0xC7);
1245			else
1246				EMIT1(0xC7);
1247			goto st;
1248		case BPF_ST | BPF_MEM | BPF_DW:
1249			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1250
1251st:			if (is_imm8(insn->off))
1252				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1253			else
1254				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1255
1256			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1257			break;
1258
1259			/* STX: *(u8*)(dst_reg + off) = src_reg */
1260		case BPF_STX | BPF_MEM | BPF_B:
1261		case BPF_STX | BPF_MEM | BPF_H:
1262		case BPF_STX | BPF_MEM | BPF_W:
1263		case BPF_STX | BPF_MEM | BPF_DW:
1264			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1265			break;
1266
1267			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1268		case BPF_LDX | BPF_MEM | BPF_B:
1269		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1270		case BPF_LDX | BPF_MEM | BPF_H:
1271		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1272		case BPF_LDX | BPF_MEM | BPF_W:
1273		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1274		case BPF_LDX | BPF_MEM | BPF_DW:
1275		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1276			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1277				/* test src_reg, src_reg */
1278				maybe_emit_mod(&prog, src_reg, src_reg, true); /* always 1 byte */
1279				EMIT2(0x85, add_2reg(0xC0, src_reg, src_reg));
1280				/* jne start_of_ldx */
1281				EMIT2(X86_JNE, 0);
1282				/* xor dst_reg, dst_reg */
1283				emit_mov_imm32(&prog, false, dst_reg, 0);
1284				/* jmp byte_after_ldx */
1285				EMIT2(0xEB, 0);
1286
1287				/* populate jmp_offset for JNE above */
1288				temp[4] = prog - temp - 5 /* sizeof(test + jne) */;
1289				start_of_ldx = prog;
1290			}
1291			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1292			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1293				struct exception_table_entry *ex;
1294				u8 *_insn = image + proglen + (start_of_ldx - temp);
1295				s64 delta;
1296
1297				/* populate jmp_offset for JMP above */
1298				start_of_ldx[-1] = prog - start_of_ldx;
1299
1300				if (!bpf_prog->aux->extable)
1301					break;
1302
1303				if (excnt >= bpf_prog->aux->num_exentries) {
1304					pr_err("ex gen bug\n");
1305					return -EFAULT;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1306				}
1307				ex = &bpf_prog->aux->extable[excnt++];
1308
1309				delta = _insn - (u8 *)&ex->insn;
1310				if (!is_simm32(delta)) {
1311					pr_err("extable->insn doesn't fit into 32-bit\n");
1312					return -EFAULT;
 
 
 
 
 
 
 
 
1313				}
1314				ex->insn = delta;
1315
1316				delta = (u8 *)ex_handler_bpf - (u8 *)&ex->handler;
1317				if (!is_simm32(delta)) {
1318					pr_err("extable->handler doesn't fit into 32-bit\n");
1319					return -EFAULT;
 
 
 
1320				}
1321				ex->handler = delta;
1322
1323				if (dst_reg > BPF_REG_9) {
1324					pr_err("verifier error\n");
1325					return -EFAULT;
 
 
 
 
1326				}
1327				/*
1328				 * Compute size of x86 insn and its target dest x86 register.
1329				 * ex_handler_bpf() will use lower 8 bits to adjust
1330				 * pt_regs->ip to jump over this x86 instruction
1331				 * and upper bits to figure out which pt_regs to zero out.
1332				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1333				 * of 4 bytes will be ignored and rbx will be zero inited.
1334				 */
1335				ex->fixup = (prog - temp) | (reg2pt_regs[dst_reg] << 8);
1336			}
1337			break;
1338
1339		case BPF_STX | BPF_ATOMIC | BPF_W:
1340		case BPF_STX | BPF_ATOMIC | BPF_DW:
1341			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1342			    insn->imm == (BPF_OR | BPF_FETCH) ||
1343			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1344				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1345				u32 real_src_reg = src_reg;
1346				u32 real_dst_reg = dst_reg;
1347				u8 *branch_target;
1348
1349				/*
1350				 * Can't be implemented with a single x86 insn.
1351				 * Need to do a CMPXCHG loop.
1352				 */
1353
1354				/* Will need RAX as a CMPXCHG operand so save R0 */
1355				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1356				if (src_reg == BPF_REG_0)
1357					real_src_reg = BPF_REG_AX;
1358				if (dst_reg == BPF_REG_0)
1359					real_dst_reg = BPF_REG_AX;
1360
1361				branch_target = prog;
1362				/* Load old value */
1363				emit_ldx(&prog, BPF_SIZE(insn->code),
1364					 BPF_REG_0, real_dst_reg, insn->off);
1365				/*
1366				 * Perform the (commutative) operation locally,
1367				 * put the result in the AUX_REG.
1368				 */
1369				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1370				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1371				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1372				      add_2reg(0xC0, AUX_REG, real_src_reg));
1373				/* Attempt to swap in new value */
1374				err = emit_atomic(&prog, BPF_CMPXCHG,
1375						  real_dst_reg, AUX_REG,
1376						  insn->off,
1377						  BPF_SIZE(insn->code));
1378				if (WARN_ON(err))
1379					return err;
1380				/*
1381				 * ZF tells us whether we won the race. If it's
1382				 * cleared we need to try again.
1383				 */
1384				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1385				/* Return the pre-modification value */
1386				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1387				/* Restore R0 after clobbering RAX */
1388				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1389				break;
1390			}
1391
1392			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1393					  insn->off, BPF_SIZE(insn->code));
1394			if (err)
1395				return err;
1396			break;
1397
1398			/* call */
1399		case BPF_JMP | BPF_CALL:
1400			func = (u8 *) __bpf_call_base + imm32;
1401			if (tail_call_reachable) {
1402				EMIT3_off32(0x48, 0x8B, 0x85,
1403					    -(bpf_prog->aux->stack_depth + 8));
1404				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1405					return -EINVAL;
1406			} else {
1407				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1408					return -EINVAL;
1409			}
1410			break;
1411
1412		case BPF_JMP | BPF_TAIL_CALL:
1413			if (imm32)
1414				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1415							  &prog, addrs[i], image,
1416							  callee_regs_used,
1417							  bpf_prog->aux->stack_depth);
1418			else
1419				emit_bpf_tail_call_indirect(&prog,
1420							    callee_regs_used,
1421							    bpf_prog->aux->stack_depth);
1422			break;
1423
1424			/* cond jump */
1425		case BPF_JMP | BPF_JEQ | BPF_X:
1426		case BPF_JMP | BPF_JNE | BPF_X:
1427		case BPF_JMP | BPF_JGT | BPF_X:
1428		case BPF_JMP | BPF_JLT | BPF_X:
1429		case BPF_JMP | BPF_JGE | BPF_X:
1430		case BPF_JMP | BPF_JLE | BPF_X:
1431		case BPF_JMP | BPF_JSGT | BPF_X:
1432		case BPF_JMP | BPF_JSLT | BPF_X:
1433		case BPF_JMP | BPF_JSGE | BPF_X:
1434		case BPF_JMP | BPF_JSLE | BPF_X:
1435		case BPF_JMP32 | BPF_JEQ | BPF_X:
1436		case BPF_JMP32 | BPF_JNE | BPF_X:
1437		case BPF_JMP32 | BPF_JGT | BPF_X:
1438		case BPF_JMP32 | BPF_JLT | BPF_X:
1439		case BPF_JMP32 | BPF_JGE | BPF_X:
1440		case BPF_JMP32 | BPF_JLE | BPF_X:
1441		case BPF_JMP32 | BPF_JSGT | BPF_X:
1442		case BPF_JMP32 | BPF_JSLT | BPF_X:
1443		case BPF_JMP32 | BPF_JSGE | BPF_X:
1444		case BPF_JMP32 | BPF_JSLE | BPF_X:
1445			/* cmp dst_reg, src_reg */
1446			maybe_emit_mod(&prog, dst_reg, src_reg,
1447				       BPF_CLASS(insn->code) == BPF_JMP);
1448			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1449			goto emit_cond_jmp;
1450
1451		case BPF_JMP | BPF_JSET | BPF_X:
1452		case BPF_JMP32 | BPF_JSET | BPF_X:
1453			/* test dst_reg, src_reg */
1454			maybe_emit_mod(&prog, dst_reg, src_reg,
1455				       BPF_CLASS(insn->code) == BPF_JMP);
1456			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1457			goto emit_cond_jmp;
1458
1459		case BPF_JMP | BPF_JSET | BPF_K:
1460		case BPF_JMP32 | BPF_JSET | BPF_K:
1461			/* test dst_reg, imm32 */
1462			if (BPF_CLASS(insn->code) == BPF_JMP)
1463				EMIT1(add_1mod(0x48, dst_reg));
1464			else if (is_ereg(dst_reg))
1465				EMIT1(add_1mod(0x40, dst_reg));
1466			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1467			goto emit_cond_jmp;
1468
1469		case BPF_JMP | BPF_JEQ | BPF_K:
1470		case BPF_JMP | BPF_JNE | BPF_K:
1471		case BPF_JMP | BPF_JGT | BPF_K:
1472		case BPF_JMP | BPF_JLT | BPF_K:
1473		case BPF_JMP | BPF_JGE | BPF_K:
1474		case BPF_JMP | BPF_JLE | BPF_K:
1475		case BPF_JMP | BPF_JSGT | BPF_K:
1476		case BPF_JMP | BPF_JSLT | BPF_K:
1477		case BPF_JMP | BPF_JSGE | BPF_K:
1478		case BPF_JMP | BPF_JSLE | BPF_K:
1479		case BPF_JMP32 | BPF_JEQ | BPF_K:
1480		case BPF_JMP32 | BPF_JNE | BPF_K:
1481		case BPF_JMP32 | BPF_JGT | BPF_K:
1482		case BPF_JMP32 | BPF_JLT | BPF_K:
1483		case BPF_JMP32 | BPF_JGE | BPF_K:
1484		case BPF_JMP32 | BPF_JLE | BPF_K:
1485		case BPF_JMP32 | BPF_JSGT | BPF_K:
1486		case BPF_JMP32 | BPF_JSLT | BPF_K:
1487		case BPF_JMP32 | BPF_JSGE | BPF_K:
1488		case BPF_JMP32 | BPF_JSLE | BPF_K:
1489			/* test dst_reg, dst_reg to save one extra byte */
1490			if (imm32 == 0) {
1491				maybe_emit_mod(&prog, dst_reg, dst_reg,
1492					       BPF_CLASS(insn->code) == BPF_JMP);
1493				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1494				goto emit_cond_jmp;
1495			}
1496
1497			/* cmp dst_reg, imm8/32 */
1498			if (BPF_CLASS(insn->code) == BPF_JMP)
1499				EMIT1(add_1mod(0x48, dst_reg));
1500			else if (is_ereg(dst_reg))
1501				EMIT1(add_1mod(0x40, dst_reg));
1502
1503			if (is_imm8(imm32))
1504				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1505			else
1506				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1507
1508emit_cond_jmp:		/* Convert BPF opcode to x86 */
1509			switch (BPF_OP(insn->code)) {
1510			case BPF_JEQ:
1511				jmp_cond = X86_JE;
1512				break;
1513			case BPF_JSET:
1514			case BPF_JNE:
1515				jmp_cond = X86_JNE;
1516				break;
1517			case BPF_JGT:
1518				/* GT is unsigned '>', JA in x86 */
1519				jmp_cond = X86_JA;
1520				break;
1521			case BPF_JLT:
1522				/* LT is unsigned '<', JB in x86 */
1523				jmp_cond = X86_JB;
1524				break;
1525			case BPF_JGE:
1526				/* GE is unsigned '>=', JAE in x86 */
1527				jmp_cond = X86_JAE;
1528				break;
1529			case BPF_JLE:
1530				/* LE is unsigned '<=', JBE in x86 */
1531				jmp_cond = X86_JBE;
1532				break;
1533			case BPF_JSGT:
1534				/* Signed '>', GT in x86 */
1535				jmp_cond = X86_JG;
1536				break;
1537			case BPF_JSLT:
1538				/* Signed '<', LT in x86 */
1539				jmp_cond = X86_JL;
1540				break;
1541			case BPF_JSGE:
1542				/* Signed '>=', GE in x86 */
1543				jmp_cond = X86_JGE;
1544				break;
1545			case BPF_JSLE:
1546				/* Signed '<=', LE in x86 */
1547				jmp_cond = X86_JLE;
1548				break;
1549			default: /* to silence GCC warning */
1550				return -EFAULT;
1551			}
1552			jmp_offset = addrs[i + insn->off] - addrs[i];
1553			if (is_imm8(jmp_offset)) {
1554				if (jmp_padding) {
1555					/* To keep the jmp_offset valid, the extra bytes are
1556					 * padded before the jump insn, so we subtract the
1557					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1558					 *
1559					 * If the previous pass already emits an imm8
1560					 * jmp_cond, then this BPF insn won't shrink, so
1561					 * "nops" is 0.
1562					 *
1563					 * On the other hand, if the previous pass emits an
1564					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1565					 * keep the image from shrinking further.
1566					 *
1567					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1568					 *     is 2 bytes, so the size difference is 4 bytes.
1569					 */
1570					nops = INSN_SZ_DIFF - 2;
1571					if (nops != 0 && nops != 4) {
1572						pr_err("unexpected jmp_cond padding: %d bytes\n",
1573						       nops);
1574						return -EFAULT;
1575					}
1576					emit_nops(&prog, nops);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1577				}
1578				EMIT2(jmp_cond, jmp_offset);
1579			} else if (is_simm32(jmp_offset)) {
1580				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1581			} else {
1582				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1583				return -EFAULT;
1584			}
1585
1586			break;
1587
1588		case BPF_JMP | BPF_JA:
1589			if (insn->off == -1)
1590				/* -1 jmp instructions will always jump
1591				 * backwards two bytes. Explicitly handling
1592				 * this case avoids wasting too many passes
1593				 * when there are long sequences of replaced
1594				 * dead code.
1595				 */
1596				jmp_offset = -2;
1597			else
1598				jmp_offset = addrs[i + insn->off] - addrs[i];
1599
1600			if (!jmp_offset) {
1601				/*
1602				 * If jmp_padding is enabled, the extra nops will
1603				 * be inserted. Otherwise, optimize out nop jumps.
1604				 */
1605				if (jmp_padding) {
1606					/* There are 3 possible conditions.
1607					 * (1) This BPF_JA is already optimized out in
1608					 *     the previous run, so there is no need
1609					 *     to pad any extra byte (0 byte).
1610					 * (2) The previous pass emits an imm8 jmp,
1611					 *     so we pad 2 bytes to match the previous
1612					 *     insn size.
1613					 * (3) Similarly, the previous pass emits an
1614					 *     imm32 jmp, and 5 bytes is padded.
1615					 */
1616					nops = INSN_SZ_DIFF;
1617					if (nops != 0 && nops != 2 && nops != 5) {
1618						pr_err("unexpected nop jump padding: %d bytes\n",
1619						       nops);
1620						return -EFAULT;
1621					}
1622					emit_nops(&prog, nops);
 
 
 
 
 
 
 
 
1623				}
 
1624				break;
1625			}
1626emit_jmp:
1627			if (is_imm8(jmp_offset)) {
1628				if (jmp_padding) {
1629					/* To avoid breaking jmp_offset, the extra bytes
1630					 * are padded before the actual jmp insn, so
1631					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1632					 *
1633					 * If the previous pass already emits an imm8
1634					 * jmp, there is nothing to pad (0 byte).
1635					 *
1636					 * If it emits an imm32 jmp (5 bytes) previously
1637					 * and now an imm8 jmp (2 bytes), then we pad
1638					 * (5 - 2 = 3) bytes to stop the image from
1639					 * shrinking further.
1640					 */
1641					nops = INSN_SZ_DIFF - 2;
1642					if (nops != 0 && nops != 3) {
1643						pr_err("unexpected jump padding: %d bytes\n",
1644						       nops);
1645						return -EFAULT;
1646					}
1647					emit_nops(&prog, INSN_SZ_DIFF - 2);
1648				}
1649				EMIT2(0xEB, jmp_offset);
1650			} else if (is_simm32(jmp_offset)) {
1651				EMIT1_off32(0xE9, jmp_offset);
1652			} else {
1653				pr_err("jmp gen bug %llx\n", jmp_offset);
1654				return -EFAULT;
1655			}
1656			break;
1657
1658		case BPF_JMP | BPF_EXIT:
1659			if (seen_exit) {
1660				jmp_offset = ctx->cleanup_addr - addrs[i];
1661				goto emit_jmp;
1662			}
1663			seen_exit = true;
1664			/* Update cleanup_addr */
1665			ctx->cleanup_addr = proglen;
1666			pop_callee_regs(&prog, callee_regs_used);
1667			EMIT1(0xC9);         /* leave */
1668			EMIT1(0xC3);         /* ret */
1669			break;
1670
1671		default:
1672			/*
1673			 * By design x86-64 JIT should support all BPF instructions.
1674			 * This error will be seen if new instruction was added
1675			 * to the interpreter, but not to the JIT, or if there is
1676			 * junk in bpf_prog.
1677			 */
1678			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1679			return -EINVAL;
1680		}
1681
1682		ilen = prog - temp;
1683		if (ilen > BPF_MAX_INSN_SIZE) {
1684			pr_err("bpf_jit: fatal insn size error\n");
1685			return -EFAULT;
1686		}
1687
1688		if (image) {
1689			/*
1690			 * When populating the image, assert that:
1691			 *
1692			 *  i) We do not write beyond the allocated space, and
1693			 * ii) addrs[i] did not change from the prior run, in order
1694			 *     to validate assumptions made for computing branch
1695			 *     displacements.
1696			 */
1697			if (unlikely(proglen + ilen > oldproglen ||
1698				     proglen + ilen != addrs[i])) {
1699				pr_err("bpf_jit: fatal error\n");
1700				return -EFAULT;
1701			}
1702			memcpy(image + proglen, temp, ilen);
1703		}
1704		proglen += ilen;
1705		addrs[i] = proglen;
1706		prog = temp;
1707	}
1708
1709	if (image && excnt != bpf_prog->aux->num_exentries) {
1710		pr_err("extable is not populated\n");
1711		return -EFAULT;
1712	}
1713	return proglen;
1714}
1715
1716static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1717		      int stack_size)
1718{
1719	int i;
1720	/* Store function arguments to stack.
1721	 * For a function that accepts two pointers the sequence will be:
1722	 * mov QWORD PTR [rbp-0x10],rdi
1723	 * mov QWORD PTR [rbp-0x8],rsi
1724	 */
1725	for (i = 0; i < min(nr_args, 6); i++)
1726		emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1727			 BPF_REG_FP,
1728			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1729			 -(stack_size - i * 8));
1730}
1731
1732static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1733			 int stack_size)
1734{
1735	int i;
1736
1737	/* Restore function arguments from stack.
1738	 * For a function that accepts two pointers the sequence will be:
1739	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1740	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1741	 */
1742	for (i = 0; i < min(nr_args, 6); i++)
1743		emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1744			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1745			 BPF_REG_FP,
1746			 -(stack_size - i * 8));
1747}
1748
1749static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1750			   struct bpf_prog *p, int stack_size, bool save_ret)
1751{
1752	u8 *prog = *pprog;
1753	u8 *jmp_insn;
1754
1755	/* arg1: mov rdi, progs[i] */
1756	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1757	if (emit_call(&prog,
1758		      p->aux->sleepable ? __bpf_prog_enter_sleepable :
1759		      __bpf_prog_enter, prog))
1760			return -EINVAL;
1761	/* remember prog start time returned by __bpf_prog_enter */
1762	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1763
1764	/* if (__bpf_prog_enter*(prog) == 0)
1765	 *	goto skip_exec_of_prog;
1766	 */
1767	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1768	/* emit 2 nops that will be replaced with JE insn */
1769	jmp_insn = prog;
1770	emit_nops(&prog, 2);
1771
1772	/* arg1: lea rdi, [rbp - stack_size] */
1773	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1774	/* arg2: progs[i]->insnsi for interpreter */
1775	if (!p->jited)
1776		emit_mov_imm64(&prog, BPF_REG_2,
1777			       (long) p->insnsi >> 32,
1778			       (u32) (long) p->insnsi);
1779	/* call JITed bpf program or interpreter */
1780	if (emit_call(&prog, p->bpf_func, prog))
1781		return -EINVAL;
1782
1783	/*
1784	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1785	 * of the previous call which is then passed on the stack to
1786	 * the next BPF program.
1787	 *
1788	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1789	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1790	 */
1791	if (save_ret)
1792		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1793
1794	/* replace 2 nops with JE insn, since jmp target is known */
1795	jmp_insn[0] = X86_JE;
1796	jmp_insn[1] = prog - jmp_insn - 2;
1797
1798	/* arg1: mov rdi, progs[i] */
1799	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1800	/* arg2: mov rsi, rbx <- start time in nsec */
1801	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1802	if (emit_call(&prog,
1803		      p->aux->sleepable ? __bpf_prog_exit_sleepable :
1804		      __bpf_prog_exit, prog))
1805			return -EINVAL;
1806
1807	*pprog = prog;
1808	return 0;
1809}
1810
1811static void emit_align(u8 **pprog, u32 align)
1812{
1813	u8 *target, *prog = *pprog;
1814
1815	target = PTR_ALIGN(prog, align);
1816	if (target != prog)
1817		emit_nops(&prog, target - prog);
1818
1819	*pprog = prog;
1820}
1821
1822static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1823{
1824	u8 *prog = *pprog;
1825	s64 offset;
1826
1827	offset = func - (ip + 2 + 4);
1828	if (!is_simm32(offset)) {
1829		pr_err("Target %p is out of range\n", func);
1830		return -EINVAL;
1831	}
1832	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1833	*pprog = prog;
1834	return 0;
1835}
1836
1837static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1838		      struct bpf_tramp_progs *tp, int stack_size,
1839		      bool save_ret)
1840{
1841	int i;
1842	u8 *prog = *pprog;
1843
1844	for (i = 0; i < tp->nr_progs; i++) {
1845		if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size,
1846				    save_ret))
1847			return -EINVAL;
1848	}
1849	*pprog = prog;
1850	return 0;
1851}
1852
1853static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1854			      struct bpf_tramp_progs *tp, int stack_size,
1855			      u8 **branches)
1856{
1857	u8 *prog = *pprog;
1858	int i;
1859
1860	/* The first fmod_ret program will receive a garbage return value.
1861	 * Set this to 0 to avoid confusing the program.
1862	 */
1863	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1864	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1865	for (i = 0; i < tp->nr_progs; i++) {
1866		if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, true))
1867			return -EINVAL;
1868
1869		/* mod_ret prog stored return value into [rbp - 8]. Emit:
1870		 * if (*(u64 *)(rbp - 8) !=  0)
1871		 *	goto do_fexit;
1872		 */
1873		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
1874		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1875
1876		/* Save the location of the branch and Generate 6 nops
1877		 * (4 bytes for an offset and 2 bytes for the jump) These nops
1878		 * are replaced with a conditional jump once do_fexit (i.e. the
1879		 * start of the fexit invocation) is finalized.
1880		 */
1881		branches[i] = prog;
1882		emit_nops(&prog, 4 + 2);
1883	}
1884
1885	*pprog = prog;
1886	return 0;
1887}
1888
1889static bool is_valid_bpf_tramp_flags(unsigned int flags)
1890{
1891	if ((flags & BPF_TRAMP_F_RESTORE_REGS) &&
1892	    (flags & BPF_TRAMP_F_SKIP_FRAME))
1893		return false;
1894
1895	/*
1896	 * BPF_TRAMP_F_RET_FENTRY_RET is only used by bpf_struct_ops,
1897	 * and it must be used alone.
1898	 */
1899	if ((flags & BPF_TRAMP_F_RET_FENTRY_RET) &&
1900	    (flags & ~BPF_TRAMP_F_RET_FENTRY_RET))
1901		return false;
1902
1903	return true;
1904}
1905
1906/* Example:
1907 * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1908 * its 'struct btf_func_model' will be nr_args=2
1909 * The assembly code when eth_type_trans is executing after trampoline:
1910 *
1911 * push rbp
1912 * mov rbp, rsp
1913 * sub rsp, 16                     // space for skb and dev
1914 * push rbx                        // temp regs to pass start time
1915 * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
1916 * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
1917 * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1918 * mov rbx, rax                    // remember start time in bpf stats are enabled
1919 * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
1920 * call addr_of_jited_FENTRY_prog
1921 * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1922 * mov rsi, rbx                    // prog start time
1923 * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1924 * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
1925 * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
1926 * pop rbx
1927 * leave
1928 * ret
1929 *
1930 * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1931 * replaced with 'call generated_bpf_trampoline'. When it returns
1932 * eth_type_trans will continue executing with original skb and dev pointers.
1933 *
1934 * The assembly code when eth_type_trans is called from trampoline:
1935 *
1936 * push rbp
1937 * mov rbp, rsp
1938 * sub rsp, 24                     // space for skb, dev, return value
1939 * push rbx                        // temp regs to pass start time
1940 * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
1941 * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
1942 * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1943 * mov rbx, rax                    // remember start time if bpf stats are enabled
1944 * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1945 * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
1946 * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1947 * mov rsi, rbx                    // prog start time
1948 * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1949 * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
1950 * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
1951 * call eth_type_trans+5           // execute body of eth_type_trans
1952 * mov qword ptr [rbp - 8], rax    // save return value
1953 * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1954 * mov rbx, rax                    // remember start time in bpf stats are enabled
1955 * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1956 * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
1957 * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1958 * mov rsi, rbx                    // prog start time
1959 * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1960 * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
1961 * pop rbx
1962 * leave
1963 * add rsp, 8                      // skip eth_type_trans's frame
1964 * ret                             // return to its caller
1965 */
1966int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
1967				const struct btf_func_model *m, u32 flags,
1968				struct bpf_tramp_progs *tprogs,
1969				void *orig_call)
1970{
1971	int ret, i, nr_args = m->nr_args;
1972	int stack_size = nr_args * 8;
1973	struct bpf_tramp_progs *fentry = &tprogs[BPF_TRAMP_FENTRY];
1974	struct bpf_tramp_progs *fexit = &tprogs[BPF_TRAMP_FEXIT];
1975	struct bpf_tramp_progs *fmod_ret = &tprogs[BPF_TRAMP_MODIFY_RETURN];
1976	u8 **branches = NULL;
1977	u8 *prog;
1978	bool save_ret;
1979
1980	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
1981	if (nr_args > 6)
1982		return -ENOTSUPP;
1983
1984	if (!is_valid_bpf_tramp_flags(flags))
1985		return -EINVAL;
1986
1987	/* room for return value of orig_call or fentry prog */
1988	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
1989	if (save_ret)
1990		stack_size += 8;
1991
1992	if (flags & BPF_TRAMP_F_SKIP_FRAME)
1993		/* skip patched call instruction and point orig_call to actual
1994		 * body of the kernel function.
1995		 */
1996		orig_call += X86_PATCH_SIZE;
1997
1998	prog = image;
1999
2000	EMIT1(0x55);		 /* push rbp */
2001	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2002	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2003	EMIT1(0x53);		 /* push rbx */
2004
2005	save_regs(m, &prog, nr_args, stack_size);
2006
2007	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2008		/* arg1: mov rdi, im */
2009		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2010		if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2011			ret = -EINVAL;
2012			goto cleanup;
2013		}
2014	}
2015
2016	if (fentry->nr_progs)
2017		if (invoke_bpf(m, &prog, fentry, stack_size,
2018			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2019			return -EINVAL;
2020
2021	if (fmod_ret->nr_progs) {
2022		branches = kcalloc(fmod_ret->nr_progs, sizeof(u8 *),
2023				   GFP_KERNEL);
2024		if (!branches)
2025			return -ENOMEM;
2026
2027		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, stack_size,
2028				       branches)) {
2029			ret = -EINVAL;
2030			goto cleanup;
2031		}
2032	}
2033
2034	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2035		restore_regs(m, &prog, nr_args, stack_size);
2036
2037		/* call original function */
2038		if (emit_call(&prog, orig_call, prog)) {
2039			ret = -EINVAL;
2040			goto cleanup;
2041		}
2042		/* remember return value in a stack for bpf prog to access */
2043		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2044		im->ip_after_call = prog;
2045		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2046		prog += X86_PATCH_SIZE;
2047	}
2048
2049	if (fmod_ret->nr_progs) {
2050		/* From Intel 64 and IA-32 Architectures Optimization
2051		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2052		 * Coding Rule 11: All branch targets should be 16-byte
2053		 * aligned.
2054		 */
2055		emit_align(&prog, 16);
2056		/* Update the branches saved in invoke_bpf_mod_ret with the
2057		 * aligned address of do_fexit.
2058		 */
2059		for (i = 0; i < fmod_ret->nr_progs; i++)
2060			emit_cond_near_jump(&branches[i], prog, branches[i],
2061					    X86_JNE);
2062	}
2063
2064	if (fexit->nr_progs)
2065		if (invoke_bpf(m, &prog, fexit, stack_size, false)) {
2066			ret = -EINVAL;
2067			goto cleanup;
2068		}
2069
2070	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2071		restore_regs(m, &prog, nr_args, stack_size);
2072
2073	/* This needs to be done regardless. If there were fmod_ret programs,
2074	 * the return value is only updated on the stack and still needs to be
2075	 * restored to R0.
2076	 */
2077	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2078		im->ip_epilogue = prog;
2079		/* arg1: mov rdi, im */
2080		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2081		if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2082			ret = -EINVAL;
2083			goto cleanup;
2084		}
2085	}
2086	/* restore return value of orig_call or fentry prog back into RAX */
2087	if (save_ret)
2088		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2089
2090	EMIT1(0x5B); /* pop rbx */
2091	EMIT1(0xC9); /* leave */
2092	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2093		/* skip our return address and return to parent */
2094		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2095	EMIT1(0xC3); /* ret */
2096	/* Make sure the trampoline generation logic doesn't overflow */
2097	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2098		ret = -EFAULT;
2099		goto cleanup;
2100	}
2101	ret = prog - (u8 *)image;
2102
2103cleanup:
2104	kfree(branches);
2105	return ret;
2106}
2107
2108static int emit_fallback_jump(u8 **pprog)
2109{
2110	u8 *prog = *pprog;
2111	int err = 0;
2112
2113#ifdef CONFIG_RETPOLINE
2114	/* Note that this assumes the the compiler uses external
2115	 * thunks for indirect calls. Both clang and GCC use the same
2116	 * naming convention for external thunks.
2117	 */
2118	err = emit_jump(&prog, __x86_indirect_thunk_rdx, prog);
2119#else
2120	EMIT2(0xFF, 0xE2);	/* jmp rdx */
2121#endif
2122	*pprog = prog;
2123	return err;
2124}
2125
2126static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
2127{
2128	u8 *jg_reloc, *prog = *pprog;
2129	int pivot, err, jg_bytes = 1;
2130	s64 jg_offset;
2131
2132	if (a == b) {
2133		/* Leaf node of recursion, i.e. not a range of indices
2134		 * anymore.
2135		 */
2136		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2137		if (!is_simm32(progs[a]))
2138			return -1;
2139		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2140			    progs[a]);
2141		err = emit_cond_near_jump(&prog,	/* je func */
2142					  (void *)progs[a], prog,
2143					  X86_JE);
2144		if (err)
2145			return err;
2146
2147		err = emit_fallback_jump(&prog);	/* jmp thunk/indirect */
2148		if (err)
2149			return err;
2150
2151		*pprog = prog;
2152		return 0;
2153	}
2154
2155	/* Not a leaf node, so we pivot, and recursively descend into
2156	 * the lower and upper ranges.
2157	 */
2158	pivot = (b - a) / 2;
2159	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2160	if (!is_simm32(progs[a + pivot]))
2161		return -1;
2162	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2163
2164	if (pivot > 2) {				/* jg upper_part */
2165		/* Require near jump. */
2166		jg_bytes = 4;
2167		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2168	} else {
2169		EMIT2(X86_JG, 0);
2170	}
2171	jg_reloc = prog;
2172
2173	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2174				  progs);
2175	if (err)
2176		return err;
2177
2178	/* From Intel 64 and IA-32 Architectures Optimization
2179	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2180	 * Coding Rule 11: All branch targets should be 16-byte
2181	 * aligned.
2182	 */
2183	emit_align(&prog, 16);
2184	jg_offset = prog - jg_reloc;
2185	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2186
2187	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2188				  b, progs);
2189	if (err)
2190		return err;
2191
2192	*pprog = prog;
2193	return 0;
2194}
2195
2196static int cmp_ips(const void *a, const void *b)
2197{
2198	const s64 *ipa = a;
2199	const s64 *ipb = b;
2200
2201	if (*ipa > *ipb)
2202		return 1;
2203	if (*ipa < *ipb)
2204		return -1;
2205	return 0;
2206}
2207
2208int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
2209{
2210	u8 *prog = image;
2211
2212	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2213	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
2214}
2215
2216struct x64_jit_data {
2217	struct bpf_binary_header *header;
2218	int *addrs;
2219	u8 *image;
2220	int proglen;
2221	struct jit_context ctx;
2222};
2223
2224#define MAX_PASSES 20
2225#define PADDING_PASSES (MAX_PASSES - 5)
2226
2227struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2228{
2229	struct bpf_binary_header *header = NULL;
2230	struct bpf_prog *tmp, *orig_prog = prog;
2231	struct x64_jit_data *jit_data;
2232	int proglen, oldproglen = 0;
2233	struct jit_context ctx = {};
2234	bool tmp_blinded = false;
2235	bool extra_pass = false;
2236	bool padding = false;
2237	u8 *image = NULL;
2238	int *addrs;
2239	int pass;
2240	int i;
2241
2242	if (!prog->jit_requested)
2243		return orig_prog;
2244
2245	tmp = bpf_jit_blind_constants(prog);
2246	/*
2247	 * If blinding was requested and we failed during blinding,
2248	 * we must fall back to the interpreter.
2249	 */
2250	if (IS_ERR(tmp))
2251		return orig_prog;
2252	if (tmp != prog) {
2253		tmp_blinded = true;
2254		prog = tmp;
2255	}
2256
2257	jit_data = prog->aux->jit_data;
2258	if (!jit_data) {
2259		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2260		if (!jit_data) {
2261			prog = orig_prog;
2262			goto out;
2263		}
2264		prog->aux->jit_data = jit_data;
2265	}
2266	addrs = jit_data->addrs;
2267	if (addrs) {
2268		ctx = jit_data->ctx;
2269		oldproglen = jit_data->proglen;
2270		image = jit_data->image;
2271		header = jit_data->header;
2272		extra_pass = true;
2273		padding = true;
2274		goto skip_init_addrs;
2275	}
2276	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2277	if (!addrs) {
2278		prog = orig_prog;
2279		goto out_addrs;
2280	}
2281
2282	/*
2283	 * Before first pass, make a rough estimation of addrs[]
2284	 * each BPF instruction is translated to less than 64 bytes
2285	 */
2286	for (proglen = 0, i = 0; i <= prog->len; i++) {
2287		proglen += 64;
2288		addrs[i] = proglen;
2289	}
2290	ctx.cleanup_addr = proglen;
2291skip_init_addrs:
2292
2293	/*
2294	 * JITed image shrinks with every pass and the loop iterates
2295	 * until the image stops shrinking. Very large BPF programs
2296	 * may converge on the last pass. In such case do one more
2297	 * pass to emit the final image.
2298	 */
2299	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2300		if (!padding && pass >= PADDING_PASSES)
2301			padding = true;
2302		proglen = do_jit(prog, addrs, image, oldproglen, &ctx, padding);
2303		if (proglen <= 0) {
2304out_image:
2305			image = NULL;
2306			if (header)
2307				bpf_jit_binary_free(header);
2308			prog = orig_prog;
2309			goto out_addrs;
2310		}
2311		if (image) {
2312			if (proglen != oldproglen) {
2313				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2314				       proglen, oldproglen);
2315				goto out_image;
2316			}
2317			break;
2318		}
2319		if (proglen == oldproglen) {
2320			/*
2321			 * The number of entries in extable is the number of BPF_LDX
2322			 * insns that access kernel memory via "pointer to BTF type".
2323			 * The verifier changed their opcode from LDX|MEM|size
2324			 * to LDX|PROBE_MEM|size to make JITing easier.
2325			 */
2326			u32 align = __alignof__(struct exception_table_entry);
2327			u32 extable_size = prog->aux->num_exentries *
2328				sizeof(struct exception_table_entry);
2329
2330			/* allocate module memory for x86 insns and extable */
2331			header = bpf_jit_binary_alloc(roundup(proglen, align) + extable_size,
2332						      &image, align, jit_fill_hole);
2333			if (!header) {
2334				prog = orig_prog;
2335				goto out_addrs;
2336			}
2337			prog->aux->extable = (void *) image + roundup(proglen, align);
2338		}
2339		oldproglen = proglen;
2340		cond_resched();
2341	}
2342
2343	if (bpf_jit_enable > 1)
2344		bpf_jit_dump(prog->len, proglen, pass + 1, image);
 
2345
2346	if (image) {
2347		if (!prog->is_func || extra_pass) {
2348			bpf_tail_call_direct_fixup(prog);
2349			bpf_jit_binary_lock_ro(header);
2350		} else {
2351			jit_data->addrs = addrs;
2352			jit_data->ctx = ctx;
2353			jit_data->proglen = proglen;
2354			jit_data->image = image;
2355			jit_data->header = header;
2356		}
2357		prog->bpf_func = (void *)image;
2358		prog->jited = 1;
2359		prog->jited_len = proglen;
2360	} else {
2361		prog = orig_prog;
2362	}
2363
2364	if (!image || !prog->is_func || extra_pass) {
2365		if (image)
2366			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2367out_addrs:
2368		kvfree(addrs);
2369		kfree(jit_data);
2370		prog->aux->jit_data = NULL;
2371	}
2372out:
2373	if (tmp_blinded)
2374		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2375					   tmp : orig_prog);
2376	return prog;
2377}
2378
2379bool bpf_jit_supports_kfunc_call(void)
2380{
2381	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
2382}