1*4882a593Smuzhiyun /* SPDX-License-Identifier: LGPL-2.1 OR MIT */
2*4882a593Smuzhiyun /*
3*4882a593Smuzhiyun * rseq-x86.h
4*4882a593Smuzhiyun *
5*4882a593Smuzhiyun * (C) Copyright 2016-2018 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6*4882a593Smuzhiyun */
7*4882a593Smuzhiyun
8*4882a593Smuzhiyun #include <stdint.h>
9*4882a593Smuzhiyun
10*4882a593Smuzhiyun /*
11*4882a593Smuzhiyun * RSEQ_SIG is used with the following reserved undefined instructions, which
12*4882a593Smuzhiyun * trap in user-space:
13*4882a593Smuzhiyun *
14*4882a593Smuzhiyun * x86-32: 0f b9 3d 53 30 05 53 ud1 0x53053053,%edi
15*4882a593Smuzhiyun * x86-64: 0f b9 3d 53 30 05 53 ud1 0x53053053(%rip),%edi
16*4882a593Smuzhiyun */
17*4882a593Smuzhiyun #define RSEQ_SIG 0x53053053
18*4882a593Smuzhiyun
19*4882a593Smuzhiyun /*
20*4882a593Smuzhiyun * Due to a compiler optimization bug in gcc-8 with asm goto and TLS asm input
21*4882a593Smuzhiyun * operands, we cannot use "m" input operands, and rather pass the __rseq_abi
22*4882a593Smuzhiyun * address through a "r" input operand.
23*4882a593Smuzhiyun */
24*4882a593Smuzhiyun
25*4882a593Smuzhiyun /* Offset of cpu_id and rseq_cs fields in struct rseq. */
26*4882a593Smuzhiyun #define RSEQ_CPU_ID_OFFSET 4
27*4882a593Smuzhiyun #define RSEQ_CS_OFFSET 8
28*4882a593Smuzhiyun
29*4882a593Smuzhiyun #ifdef __x86_64__
30*4882a593Smuzhiyun
31*4882a593Smuzhiyun #define RSEQ_ASM_TP_SEGMENT %%fs
32*4882a593Smuzhiyun
33*4882a593Smuzhiyun #define rseq_smp_mb() \
34*4882a593Smuzhiyun __asm__ __volatile__ ("lock; addl $0,-128(%%rsp)" ::: "memory", "cc")
35*4882a593Smuzhiyun #define rseq_smp_rmb() rseq_barrier()
36*4882a593Smuzhiyun #define rseq_smp_wmb() rseq_barrier()
37*4882a593Smuzhiyun
38*4882a593Smuzhiyun #define rseq_smp_load_acquire(p) \
39*4882a593Smuzhiyun __extension__ ({ \
40*4882a593Smuzhiyun __typeof(*p) ____p1 = RSEQ_READ_ONCE(*p); \
41*4882a593Smuzhiyun rseq_barrier(); \
42*4882a593Smuzhiyun ____p1; \
43*4882a593Smuzhiyun })
44*4882a593Smuzhiyun
45*4882a593Smuzhiyun #define rseq_smp_acquire__after_ctrl_dep() rseq_smp_rmb()
46*4882a593Smuzhiyun
47*4882a593Smuzhiyun #define rseq_smp_store_release(p, v) \
48*4882a593Smuzhiyun do { \
49*4882a593Smuzhiyun rseq_barrier(); \
50*4882a593Smuzhiyun RSEQ_WRITE_ONCE(*p, v); \
51*4882a593Smuzhiyun } while (0)
52*4882a593Smuzhiyun
53*4882a593Smuzhiyun #ifdef RSEQ_SKIP_FASTPATH
54*4882a593Smuzhiyun #include "rseq-skip.h"
55*4882a593Smuzhiyun #else /* !RSEQ_SKIP_FASTPATH */
56*4882a593Smuzhiyun
57*4882a593Smuzhiyun #define __RSEQ_ASM_DEFINE_TABLE(label, version, flags, \
58*4882a593Smuzhiyun start_ip, post_commit_offset, abort_ip) \
59*4882a593Smuzhiyun ".pushsection __rseq_cs, \"aw\"\n\t" \
60*4882a593Smuzhiyun ".balign 32\n\t" \
61*4882a593Smuzhiyun __rseq_str(label) ":\n\t" \
62*4882a593Smuzhiyun ".long " __rseq_str(version) ", " __rseq_str(flags) "\n\t" \
63*4882a593Smuzhiyun ".quad " __rseq_str(start_ip) ", " __rseq_str(post_commit_offset) ", " __rseq_str(abort_ip) "\n\t" \
64*4882a593Smuzhiyun ".popsection\n\t" \
65*4882a593Smuzhiyun ".pushsection __rseq_cs_ptr_array, \"aw\"\n\t" \
66*4882a593Smuzhiyun ".quad " __rseq_str(label) "b\n\t" \
67*4882a593Smuzhiyun ".popsection\n\t"
68*4882a593Smuzhiyun
69*4882a593Smuzhiyun
70*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_TABLE(label, start_ip, post_commit_ip, abort_ip) \
71*4882a593Smuzhiyun __RSEQ_ASM_DEFINE_TABLE(label, 0x0, 0x0, start_ip, \
72*4882a593Smuzhiyun (post_commit_ip - start_ip), abort_ip)
73*4882a593Smuzhiyun
74*4882a593Smuzhiyun /*
75*4882a593Smuzhiyun * Exit points of a rseq critical section consist of all instructions outside
76*4882a593Smuzhiyun * of the critical section where a critical section can either branch to or
77*4882a593Smuzhiyun * reach through the normal course of its execution. The abort IP and the
78*4882a593Smuzhiyun * post-commit IP are already part of the __rseq_cs section and should not be
79*4882a593Smuzhiyun * explicitly defined as additional exit points. Knowing all exit points is
80*4882a593Smuzhiyun * useful to assist debuggers stepping over the critical section.
81*4882a593Smuzhiyun */
82*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_EXIT_POINT(start_ip, exit_ip) \
83*4882a593Smuzhiyun ".pushsection __rseq_exit_point_array, \"aw\"\n\t" \
84*4882a593Smuzhiyun ".quad " __rseq_str(start_ip) ", " __rseq_str(exit_ip) "\n\t" \
85*4882a593Smuzhiyun ".popsection\n\t"
86*4882a593Smuzhiyun
87*4882a593Smuzhiyun #define RSEQ_ASM_STORE_RSEQ_CS(label, cs_label, rseq_cs) \
88*4882a593Smuzhiyun RSEQ_INJECT_ASM(1) \
89*4882a593Smuzhiyun "leaq " __rseq_str(cs_label) "(%%rip), %%rax\n\t" \
90*4882a593Smuzhiyun "movq %%rax, " __rseq_str(rseq_cs) "\n\t" \
91*4882a593Smuzhiyun __rseq_str(label) ":\n\t"
92*4882a593Smuzhiyun
93*4882a593Smuzhiyun #define RSEQ_ASM_CMP_CPU_ID(cpu_id, current_cpu_id, label) \
94*4882a593Smuzhiyun RSEQ_INJECT_ASM(2) \
95*4882a593Smuzhiyun "cmpl %[" __rseq_str(cpu_id) "], " __rseq_str(current_cpu_id) "\n\t" \
96*4882a593Smuzhiyun "jnz " __rseq_str(label) "\n\t"
97*4882a593Smuzhiyun
98*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_ABORT(label, teardown, abort_label) \
99*4882a593Smuzhiyun ".pushsection __rseq_failure, \"ax\"\n\t" \
100*4882a593Smuzhiyun /* Disassembler-friendly signature: ud1 <sig>(%rip),%edi. */ \
101*4882a593Smuzhiyun ".byte 0x0f, 0xb9, 0x3d\n\t" \
102*4882a593Smuzhiyun ".long " __rseq_str(RSEQ_SIG) "\n\t" \
103*4882a593Smuzhiyun __rseq_str(label) ":\n\t" \
104*4882a593Smuzhiyun teardown \
105*4882a593Smuzhiyun "jmp %l[" __rseq_str(abort_label) "]\n\t" \
106*4882a593Smuzhiyun ".popsection\n\t"
107*4882a593Smuzhiyun
108*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_CMPFAIL(label, teardown, cmpfail_label) \
109*4882a593Smuzhiyun ".pushsection __rseq_failure, \"ax\"\n\t" \
110*4882a593Smuzhiyun __rseq_str(label) ":\n\t" \
111*4882a593Smuzhiyun teardown \
112*4882a593Smuzhiyun "jmp %l[" __rseq_str(cmpfail_label) "]\n\t" \
113*4882a593Smuzhiyun ".popsection\n\t"
114*4882a593Smuzhiyun
115*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_storev(intptr_t * v,intptr_t expect,intptr_t newv,int cpu)116*4882a593Smuzhiyun int rseq_cmpeqv_storev(intptr_t *v, intptr_t expect, intptr_t newv, int cpu)
117*4882a593Smuzhiyun {
118*4882a593Smuzhiyun RSEQ_INJECT_C(9)
119*4882a593Smuzhiyun
120*4882a593Smuzhiyun __asm__ __volatile__ goto (
121*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
122*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
123*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
124*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
125*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
126*4882a593Smuzhiyun #endif
127*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
128*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
129*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
130*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
131*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
132*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
133*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
134*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
135*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
136*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
137*4882a593Smuzhiyun "jnz %l[error2]\n\t"
138*4882a593Smuzhiyun #endif
139*4882a593Smuzhiyun /* final store */
140*4882a593Smuzhiyun "movq %[newv], %[v]\n\t"
141*4882a593Smuzhiyun "2:\n\t"
142*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
143*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
144*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
145*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
146*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
147*4882a593Smuzhiyun [v] "m" (*v),
148*4882a593Smuzhiyun [expect] "r" (expect),
149*4882a593Smuzhiyun [newv] "r" (newv)
150*4882a593Smuzhiyun : "memory", "cc", "rax"
151*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
152*4882a593Smuzhiyun : abort, cmpfail
153*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
154*4882a593Smuzhiyun , error1, error2
155*4882a593Smuzhiyun #endif
156*4882a593Smuzhiyun );
157*4882a593Smuzhiyun rseq_after_asm_goto();
158*4882a593Smuzhiyun return 0;
159*4882a593Smuzhiyun abort:
160*4882a593Smuzhiyun rseq_after_asm_goto();
161*4882a593Smuzhiyun RSEQ_INJECT_FAILED
162*4882a593Smuzhiyun return -1;
163*4882a593Smuzhiyun cmpfail:
164*4882a593Smuzhiyun rseq_after_asm_goto();
165*4882a593Smuzhiyun return 1;
166*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
167*4882a593Smuzhiyun error1:
168*4882a593Smuzhiyun rseq_after_asm_goto();
169*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
170*4882a593Smuzhiyun error2:
171*4882a593Smuzhiyun rseq_after_asm_goto();
172*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
173*4882a593Smuzhiyun #endif
174*4882a593Smuzhiyun }
175*4882a593Smuzhiyun
176*4882a593Smuzhiyun /*
177*4882a593Smuzhiyun * Compare @v against @expectnot. When it does _not_ match, load @v
178*4882a593Smuzhiyun * into @load, and store the content of *@v + voffp into @v.
179*4882a593Smuzhiyun */
180*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpnev_storeoffp_load(intptr_t * v,intptr_t expectnot,long voffp,intptr_t * load,int cpu)181*4882a593Smuzhiyun int rseq_cmpnev_storeoffp_load(intptr_t *v, intptr_t expectnot,
182*4882a593Smuzhiyun long voffp, intptr_t *load, int cpu)
183*4882a593Smuzhiyun {
184*4882a593Smuzhiyun RSEQ_INJECT_C(9)
185*4882a593Smuzhiyun
186*4882a593Smuzhiyun __asm__ __volatile__ goto (
187*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
188*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
189*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
190*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
191*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
192*4882a593Smuzhiyun #endif
193*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
194*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
195*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
196*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
197*4882a593Smuzhiyun "movq %[v], %%rbx\n\t"
198*4882a593Smuzhiyun "cmpq %%rbx, %[expectnot]\n\t"
199*4882a593Smuzhiyun "je %l[cmpfail]\n\t"
200*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
201*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
202*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
203*4882a593Smuzhiyun "movq %[v], %%rbx\n\t"
204*4882a593Smuzhiyun "cmpq %%rbx, %[expectnot]\n\t"
205*4882a593Smuzhiyun "je %l[error2]\n\t"
206*4882a593Smuzhiyun #endif
207*4882a593Smuzhiyun "movq %%rbx, %[load]\n\t"
208*4882a593Smuzhiyun "addq %[voffp], %%rbx\n\t"
209*4882a593Smuzhiyun "movq (%%rbx), %%rbx\n\t"
210*4882a593Smuzhiyun /* final store */
211*4882a593Smuzhiyun "movq %%rbx, %[v]\n\t"
212*4882a593Smuzhiyun "2:\n\t"
213*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
214*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
215*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
216*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
217*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
218*4882a593Smuzhiyun /* final store input */
219*4882a593Smuzhiyun [v] "m" (*v),
220*4882a593Smuzhiyun [expectnot] "r" (expectnot),
221*4882a593Smuzhiyun [voffp] "er" (voffp),
222*4882a593Smuzhiyun [load] "m" (*load)
223*4882a593Smuzhiyun : "memory", "cc", "rax", "rbx"
224*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
225*4882a593Smuzhiyun : abort, cmpfail
226*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
227*4882a593Smuzhiyun , error1, error2
228*4882a593Smuzhiyun #endif
229*4882a593Smuzhiyun );
230*4882a593Smuzhiyun rseq_after_asm_goto();
231*4882a593Smuzhiyun return 0;
232*4882a593Smuzhiyun abort:
233*4882a593Smuzhiyun rseq_after_asm_goto();
234*4882a593Smuzhiyun RSEQ_INJECT_FAILED
235*4882a593Smuzhiyun return -1;
236*4882a593Smuzhiyun cmpfail:
237*4882a593Smuzhiyun rseq_after_asm_goto();
238*4882a593Smuzhiyun return 1;
239*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
240*4882a593Smuzhiyun error1:
241*4882a593Smuzhiyun rseq_after_asm_goto();
242*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
243*4882a593Smuzhiyun error2:
244*4882a593Smuzhiyun rseq_after_asm_goto();
245*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
246*4882a593Smuzhiyun #endif
247*4882a593Smuzhiyun }
248*4882a593Smuzhiyun
249*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_addv(intptr_t * v,intptr_t count,int cpu)250*4882a593Smuzhiyun int rseq_addv(intptr_t *v, intptr_t count, int cpu)
251*4882a593Smuzhiyun {
252*4882a593Smuzhiyun RSEQ_INJECT_C(9)
253*4882a593Smuzhiyun
254*4882a593Smuzhiyun __asm__ __volatile__ goto (
255*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
256*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
257*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
258*4882a593Smuzhiyun #endif
259*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
260*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
261*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
262*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
263*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
264*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
265*4882a593Smuzhiyun #endif
266*4882a593Smuzhiyun /* final store */
267*4882a593Smuzhiyun "addq %[count], %[v]\n\t"
268*4882a593Smuzhiyun "2:\n\t"
269*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
270*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
271*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
272*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
273*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
274*4882a593Smuzhiyun /* final store input */
275*4882a593Smuzhiyun [v] "m" (*v),
276*4882a593Smuzhiyun [count] "er" (count)
277*4882a593Smuzhiyun : "memory", "cc", "rax"
278*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
279*4882a593Smuzhiyun : abort
280*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
281*4882a593Smuzhiyun , error1
282*4882a593Smuzhiyun #endif
283*4882a593Smuzhiyun );
284*4882a593Smuzhiyun rseq_after_asm_goto();
285*4882a593Smuzhiyun return 0;
286*4882a593Smuzhiyun abort:
287*4882a593Smuzhiyun rseq_after_asm_goto();
288*4882a593Smuzhiyun RSEQ_INJECT_FAILED
289*4882a593Smuzhiyun return -1;
290*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
291*4882a593Smuzhiyun error1:
292*4882a593Smuzhiyun rseq_after_asm_goto();
293*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
294*4882a593Smuzhiyun #endif
295*4882a593Smuzhiyun }
296*4882a593Smuzhiyun
297*4882a593Smuzhiyun #define RSEQ_ARCH_HAS_OFFSET_DEREF_ADDV
298*4882a593Smuzhiyun
299*4882a593Smuzhiyun /*
300*4882a593Smuzhiyun * pval = *(ptr+off)
301*4882a593Smuzhiyun * *pval += inc;
302*4882a593Smuzhiyun */
303*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_offset_deref_addv(intptr_t * ptr,long off,intptr_t inc,int cpu)304*4882a593Smuzhiyun int rseq_offset_deref_addv(intptr_t *ptr, long off, intptr_t inc, int cpu)
305*4882a593Smuzhiyun {
306*4882a593Smuzhiyun RSEQ_INJECT_C(9)
307*4882a593Smuzhiyun
308*4882a593Smuzhiyun __asm__ __volatile__ goto (
309*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
310*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
311*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
312*4882a593Smuzhiyun #endif
313*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
314*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
315*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
316*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
317*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
318*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
319*4882a593Smuzhiyun #endif
320*4882a593Smuzhiyun /* get p+v */
321*4882a593Smuzhiyun "movq %[ptr], %%rbx\n\t"
322*4882a593Smuzhiyun "addq %[off], %%rbx\n\t"
323*4882a593Smuzhiyun /* get pv */
324*4882a593Smuzhiyun "movq (%%rbx), %%rcx\n\t"
325*4882a593Smuzhiyun /* *pv += inc */
326*4882a593Smuzhiyun "addq %[inc], (%%rcx)\n\t"
327*4882a593Smuzhiyun "2:\n\t"
328*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
329*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
330*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
331*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
332*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
333*4882a593Smuzhiyun /* final store input */
334*4882a593Smuzhiyun [ptr] "m" (*ptr),
335*4882a593Smuzhiyun [off] "er" (off),
336*4882a593Smuzhiyun [inc] "er" (inc)
337*4882a593Smuzhiyun : "memory", "cc", "rax", "rbx", "rcx"
338*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
339*4882a593Smuzhiyun : abort
340*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
341*4882a593Smuzhiyun , error1
342*4882a593Smuzhiyun #endif
343*4882a593Smuzhiyun );
344*4882a593Smuzhiyun return 0;
345*4882a593Smuzhiyun abort:
346*4882a593Smuzhiyun RSEQ_INJECT_FAILED
347*4882a593Smuzhiyun return -1;
348*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
349*4882a593Smuzhiyun error1:
350*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
351*4882a593Smuzhiyun #endif
352*4882a593Smuzhiyun }
353*4882a593Smuzhiyun
354*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trystorev_storev(intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t newv2,intptr_t newv,int cpu)355*4882a593Smuzhiyun int rseq_cmpeqv_trystorev_storev(intptr_t *v, intptr_t expect,
356*4882a593Smuzhiyun intptr_t *v2, intptr_t newv2,
357*4882a593Smuzhiyun intptr_t newv, int cpu)
358*4882a593Smuzhiyun {
359*4882a593Smuzhiyun RSEQ_INJECT_C(9)
360*4882a593Smuzhiyun
361*4882a593Smuzhiyun __asm__ __volatile__ goto (
362*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
363*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
364*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
365*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
366*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
367*4882a593Smuzhiyun #endif
368*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
369*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
370*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
371*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
372*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
373*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
374*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
375*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
376*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
377*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
378*4882a593Smuzhiyun "jnz %l[error2]\n\t"
379*4882a593Smuzhiyun #endif
380*4882a593Smuzhiyun /* try store */
381*4882a593Smuzhiyun "movq %[newv2], %[v2]\n\t"
382*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
383*4882a593Smuzhiyun /* final store */
384*4882a593Smuzhiyun "movq %[newv], %[v]\n\t"
385*4882a593Smuzhiyun "2:\n\t"
386*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
387*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
388*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
389*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
390*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
391*4882a593Smuzhiyun /* try store input */
392*4882a593Smuzhiyun [v2] "m" (*v2),
393*4882a593Smuzhiyun [newv2] "r" (newv2),
394*4882a593Smuzhiyun /* final store input */
395*4882a593Smuzhiyun [v] "m" (*v),
396*4882a593Smuzhiyun [expect] "r" (expect),
397*4882a593Smuzhiyun [newv] "r" (newv)
398*4882a593Smuzhiyun : "memory", "cc", "rax"
399*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
400*4882a593Smuzhiyun : abort, cmpfail
401*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
402*4882a593Smuzhiyun , error1, error2
403*4882a593Smuzhiyun #endif
404*4882a593Smuzhiyun );
405*4882a593Smuzhiyun rseq_after_asm_goto();
406*4882a593Smuzhiyun return 0;
407*4882a593Smuzhiyun abort:
408*4882a593Smuzhiyun rseq_after_asm_goto();
409*4882a593Smuzhiyun RSEQ_INJECT_FAILED
410*4882a593Smuzhiyun return -1;
411*4882a593Smuzhiyun cmpfail:
412*4882a593Smuzhiyun rseq_after_asm_goto();
413*4882a593Smuzhiyun return 1;
414*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
415*4882a593Smuzhiyun error1:
416*4882a593Smuzhiyun rseq_after_asm_goto();
417*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
418*4882a593Smuzhiyun error2:
419*4882a593Smuzhiyun rseq_after_asm_goto();
420*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
421*4882a593Smuzhiyun #endif
422*4882a593Smuzhiyun }
423*4882a593Smuzhiyun
424*4882a593Smuzhiyun /* x86-64 is TSO. */
425*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trystorev_storev_release(intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t newv2,intptr_t newv,int cpu)426*4882a593Smuzhiyun int rseq_cmpeqv_trystorev_storev_release(intptr_t *v, intptr_t expect,
427*4882a593Smuzhiyun intptr_t *v2, intptr_t newv2,
428*4882a593Smuzhiyun intptr_t newv, int cpu)
429*4882a593Smuzhiyun {
430*4882a593Smuzhiyun return rseq_cmpeqv_trystorev_storev(v, expect, v2, newv2, newv, cpu);
431*4882a593Smuzhiyun }
432*4882a593Smuzhiyun
433*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_cmpeqv_storev(intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t expect2,intptr_t newv,int cpu)434*4882a593Smuzhiyun int rseq_cmpeqv_cmpeqv_storev(intptr_t *v, intptr_t expect,
435*4882a593Smuzhiyun intptr_t *v2, intptr_t expect2,
436*4882a593Smuzhiyun intptr_t newv, int cpu)
437*4882a593Smuzhiyun {
438*4882a593Smuzhiyun RSEQ_INJECT_C(9)
439*4882a593Smuzhiyun
440*4882a593Smuzhiyun __asm__ __volatile__ goto (
441*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
442*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
443*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
444*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
445*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
446*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error3])
447*4882a593Smuzhiyun #endif
448*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
449*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
450*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
451*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
452*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
453*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
454*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
455*4882a593Smuzhiyun "cmpq %[v2], %[expect2]\n\t"
456*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
457*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
458*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
459*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
460*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
461*4882a593Smuzhiyun "jnz %l[error2]\n\t"
462*4882a593Smuzhiyun "cmpq %[v2], %[expect2]\n\t"
463*4882a593Smuzhiyun "jnz %l[error3]\n\t"
464*4882a593Smuzhiyun #endif
465*4882a593Smuzhiyun /* final store */
466*4882a593Smuzhiyun "movq %[newv], %[v]\n\t"
467*4882a593Smuzhiyun "2:\n\t"
468*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
469*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
470*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
471*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
472*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
473*4882a593Smuzhiyun /* cmp2 input */
474*4882a593Smuzhiyun [v2] "m" (*v2),
475*4882a593Smuzhiyun [expect2] "r" (expect2),
476*4882a593Smuzhiyun /* final store input */
477*4882a593Smuzhiyun [v] "m" (*v),
478*4882a593Smuzhiyun [expect] "r" (expect),
479*4882a593Smuzhiyun [newv] "r" (newv)
480*4882a593Smuzhiyun : "memory", "cc", "rax"
481*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
482*4882a593Smuzhiyun : abort, cmpfail
483*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
484*4882a593Smuzhiyun , error1, error2, error3
485*4882a593Smuzhiyun #endif
486*4882a593Smuzhiyun );
487*4882a593Smuzhiyun rseq_after_asm_goto();
488*4882a593Smuzhiyun return 0;
489*4882a593Smuzhiyun abort:
490*4882a593Smuzhiyun rseq_after_asm_goto();
491*4882a593Smuzhiyun RSEQ_INJECT_FAILED
492*4882a593Smuzhiyun return -1;
493*4882a593Smuzhiyun cmpfail:
494*4882a593Smuzhiyun rseq_after_asm_goto();
495*4882a593Smuzhiyun return 1;
496*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
497*4882a593Smuzhiyun error1:
498*4882a593Smuzhiyun rseq_after_asm_goto();
499*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
500*4882a593Smuzhiyun error2:
501*4882a593Smuzhiyun rseq_after_asm_goto();
502*4882a593Smuzhiyun rseq_bug("1st expected value comparison failed");
503*4882a593Smuzhiyun error3:
504*4882a593Smuzhiyun rseq_after_asm_goto();
505*4882a593Smuzhiyun rseq_bug("2nd expected value comparison failed");
506*4882a593Smuzhiyun #endif
507*4882a593Smuzhiyun }
508*4882a593Smuzhiyun
509*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trymemcpy_storev(intptr_t * v,intptr_t expect,void * dst,void * src,size_t len,intptr_t newv,int cpu)510*4882a593Smuzhiyun int rseq_cmpeqv_trymemcpy_storev(intptr_t *v, intptr_t expect,
511*4882a593Smuzhiyun void *dst, void *src, size_t len,
512*4882a593Smuzhiyun intptr_t newv, int cpu)
513*4882a593Smuzhiyun {
514*4882a593Smuzhiyun uint64_t rseq_scratch[3];
515*4882a593Smuzhiyun
516*4882a593Smuzhiyun RSEQ_INJECT_C(9)
517*4882a593Smuzhiyun
518*4882a593Smuzhiyun __asm__ __volatile__ goto (
519*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
520*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
521*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
522*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
523*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
524*4882a593Smuzhiyun #endif
525*4882a593Smuzhiyun "movq %[src], %[rseq_scratch0]\n\t"
526*4882a593Smuzhiyun "movq %[dst], %[rseq_scratch1]\n\t"
527*4882a593Smuzhiyun "movq %[len], %[rseq_scratch2]\n\t"
528*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
529*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
530*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
531*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
532*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
533*4882a593Smuzhiyun "jnz 5f\n\t"
534*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
535*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
536*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 6f)
537*4882a593Smuzhiyun "cmpq %[v], %[expect]\n\t"
538*4882a593Smuzhiyun "jnz 7f\n\t"
539*4882a593Smuzhiyun #endif
540*4882a593Smuzhiyun /* try memcpy */
541*4882a593Smuzhiyun "test %[len], %[len]\n\t" \
542*4882a593Smuzhiyun "jz 333f\n\t" \
543*4882a593Smuzhiyun "222:\n\t" \
544*4882a593Smuzhiyun "movb (%[src]), %%al\n\t" \
545*4882a593Smuzhiyun "movb %%al, (%[dst])\n\t" \
546*4882a593Smuzhiyun "inc %[src]\n\t" \
547*4882a593Smuzhiyun "inc %[dst]\n\t" \
548*4882a593Smuzhiyun "dec %[len]\n\t" \
549*4882a593Smuzhiyun "jnz 222b\n\t" \
550*4882a593Smuzhiyun "333:\n\t" \
551*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
552*4882a593Smuzhiyun /* final store */
553*4882a593Smuzhiyun "movq %[newv], %[v]\n\t"
554*4882a593Smuzhiyun "2:\n\t"
555*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
556*4882a593Smuzhiyun /* teardown */
557*4882a593Smuzhiyun "movq %[rseq_scratch2], %[len]\n\t"
558*4882a593Smuzhiyun "movq %[rseq_scratch1], %[dst]\n\t"
559*4882a593Smuzhiyun "movq %[rseq_scratch0], %[src]\n\t"
560*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4,
561*4882a593Smuzhiyun "movq %[rseq_scratch2], %[len]\n\t"
562*4882a593Smuzhiyun "movq %[rseq_scratch1], %[dst]\n\t"
563*4882a593Smuzhiyun "movq %[rseq_scratch0], %[src]\n\t",
564*4882a593Smuzhiyun abort)
565*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(5,
566*4882a593Smuzhiyun "movq %[rseq_scratch2], %[len]\n\t"
567*4882a593Smuzhiyun "movq %[rseq_scratch1], %[dst]\n\t"
568*4882a593Smuzhiyun "movq %[rseq_scratch0], %[src]\n\t",
569*4882a593Smuzhiyun cmpfail)
570*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
571*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(6,
572*4882a593Smuzhiyun "movq %[rseq_scratch2], %[len]\n\t"
573*4882a593Smuzhiyun "movq %[rseq_scratch1], %[dst]\n\t"
574*4882a593Smuzhiyun "movq %[rseq_scratch0], %[src]\n\t",
575*4882a593Smuzhiyun error1)
576*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(7,
577*4882a593Smuzhiyun "movq %[rseq_scratch2], %[len]\n\t"
578*4882a593Smuzhiyun "movq %[rseq_scratch1], %[dst]\n\t"
579*4882a593Smuzhiyun "movq %[rseq_scratch0], %[src]\n\t",
580*4882a593Smuzhiyun error2)
581*4882a593Smuzhiyun #endif
582*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
583*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
584*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
585*4882a593Smuzhiyun /* final store input */
586*4882a593Smuzhiyun [v] "m" (*v),
587*4882a593Smuzhiyun [expect] "r" (expect),
588*4882a593Smuzhiyun [newv] "r" (newv),
589*4882a593Smuzhiyun /* try memcpy input */
590*4882a593Smuzhiyun [dst] "r" (dst),
591*4882a593Smuzhiyun [src] "r" (src),
592*4882a593Smuzhiyun [len] "r" (len),
593*4882a593Smuzhiyun [rseq_scratch0] "m" (rseq_scratch[0]),
594*4882a593Smuzhiyun [rseq_scratch1] "m" (rseq_scratch[1]),
595*4882a593Smuzhiyun [rseq_scratch2] "m" (rseq_scratch[2])
596*4882a593Smuzhiyun : "memory", "cc", "rax"
597*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
598*4882a593Smuzhiyun : abort, cmpfail
599*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
600*4882a593Smuzhiyun , error1, error2
601*4882a593Smuzhiyun #endif
602*4882a593Smuzhiyun );
603*4882a593Smuzhiyun rseq_after_asm_goto();
604*4882a593Smuzhiyun return 0;
605*4882a593Smuzhiyun abort:
606*4882a593Smuzhiyun rseq_after_asm_goto();
607*4882a593Smuzhiyun RSEQ_INJECT_FAILED
608*4882a593Smuzhiyun return -1;
609*4882a593Smuzhiyun cmpfail:
610*4882a593Smuzhiyun rseq_after_asm_goto();
611*4882a593Smuzhiyun return 1;
612*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
613*4882a593Smuzhiyun error1:
614*4882a593Smuzhiyun rseq_after_asm_goto();
615*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
616*4882a593Smuzhiyun error2:
617*4882a593Smuzhiyun rseq_after_asm_goto();
618*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
619*4882a593Smuzhiyun #endif
620*4882a593Smuzhiyun }
621*4882a593Smuzhiyun
622*4882a593Smuzhiyun /* x86-64 is TSO. */
623*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trymemcpy_storev_release(intptr_t * v,intptr_t expect,void * dst,void * src,size_t len,intptr_t newv,int cpu)624*4882a593Smuzhiyun int rseq_cmpeqv_trymemcpy_storev_release(intptr_t *v, intptr_t expect,
625*4882a593Smuzhiyun void *dst, void *src, size_t len,
626*4882a593Smuzhiyun intptr_t newv, int cpu)
627*4882a593Smuzhiyun {
628*4882a593Smuzhiyun return rseq_cmpeqv_trymemcpy_storev(v, expect, dst, src, len,
629*4882a593Smuzhiyun newv, cpu);
630*4882a593Smuzhiyun }
631*4882a593Smuzhiyun
632*4882a593Smuzhiyun #endif /* !RSEQ_SKIP_FASTPATH */
633*4882a593Smuzhiyun
634*4882a593Smuzhiyun #elif defined(__i386__)
635*4882a593Smuzhiyun
636*4882a593Smuzhiyun #define RSEQ_ASM_TP_SEGMENT %%gs
637*4882a593Smuzhiyun
638*4882a593Smuzhiyun #define rseq_smp_mb() \
639*4882a593Smuzhiyun __asm__ __volatile__ ("lock; addl $0,-128(%%esp)" ::: "memory", "cc")
640*4882a593Smuzhiyun #define rseq_smp_rmb() \
641*4882a593Smuzhiyun __asm__ __volatile__ ("lock; addl $0,-128(%%esp)" ::: "memory", "cc")
642*4882a593Smuzhiyun #define rseq_smp_wmb() \
643*4882a593Smuzhiyun __asm__ __volatile__ ("lock; addl $0,-128(%%esp)" ::: "memory", "cc")
644*4882a593Smuzhiyun
645*4882a593Smuzhiyun #define rseq_smp_load_acquire(p) \
646*4882a593Smuzhiyun __extension__ ({ \
647*4882a593Smuzhiyun __typeof(*p) ____p1 = RSEQ_READ_ONCE(*p); \
648*4882a593Smuzhiyun rseq_smp_mb(); \
649*4882a593Smuzhiyun ____p1; \
650*4882a593Smuzhiyun })
651*4882a593Smuzhiyun
652*4882a593Smuzhiyun #define rseq_smp_acquire__after_ctrl_dep() rseq_smp_rmb()
653*4882a593Smuzhiyun
654*4882a593Smuzhiyun #define rseq_smp_store_release(p, v) \
655*4882a593Smuzhiyun do { \
656*4882a593Smuzhiyun rseq_smp_mb(); \
657*4882a593Smuzhiyun RSEQ_WRITE_ONCE(*p, v); \
658*4882a593Smuzhiyun } while (0)
659*4882a593Smuzhiyun
660*4882a593Smuzhiyun #ifdef RSEQ_SKIP_FASTPATH
661*4882a593Smuzhiyun #include "rseq-skip.h"
662*4882a593Smuzhiyun #else /* !RSEQ_SKIP_FASTPATH */
663*4882a593Smuzhiyun
664*4882a593Smuzhiyun /*
665*4882a593Smuzhiyun * Use eax as scratch register and take memory operands as input to
666*4882a593Smuzhiyun * lessen register pressure. Especially needed when compiling in O0.
667*4882a593Smuzhiyun */
668*4882a593Smuzhiyun #define __RSEQ_ASM_DEFINE_TABLE(label, version, flags, \
669*4882a593Smuzhiyun start_ip, post_commit_offset, abort_ip) \
670*4882a593Smuzhiyun ".pushsection __rseq_cs, \"aw\"\n\t" \
671*4882a593Smuzhiyun ".balign 32\n\t" \
672*4882a593Smuzhiyun __rseq_str(label) ":\n\t" \
673*4882a593Smuzhiyun ".long " __rseq_str(version) ", " __rseq_str(flags) "\n\t" \
674*4882a593Smuzhiyun ".long " __rseq_str(start_ip) ", 0x0, " __rseq_str(post_commit_offset) ", 0x0, " __rseq_str(abort_ip) ", 0x0\n\t" \
675*4882a593Smuzhiyun ".popsection\n\t" \
676*4882a593Smuzhiyun ".pushsection __rseq_cs_ptr_array, \"aw\"\n\t" \
677*4882a593Smuzhiyun ".long " __rseq_str(label) "b, 0x0\n\t" \
678*4882a593Smuzhiyun ".popsection\n\t"
679*4882a593Smuzhiyun
680*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_TABLE(label, start_ip, post_commit_ip, abort_ip) \
681*4882a593Smuzhiyun __RSEQ_ASM_DEFINE_TABLE(label, 0x0, 0x0, start_ip, \
682*4882a593Smuzhiyun (post_commit_ip - start_ip), abort_ip)
683*4882a593Smuzhiyun
684*4882a593Smuzhiyun /*
685*4882a593Smuzhiyun * Exit points of a rseq critical section consist of all instructions outside
686*4882a593Smuzhiyun * of the critical section where a critical section can either branch to or
687*4882a593Smuzhiyun * reach through the normal course of its execution. The abort IP and the
688*4882a593Smuzhiyun * post-commit IP are already part of the __rseq_cs section and should not be
689*4882a593Smuzhiyun * explicitly defined as additional exit points. Knowing all exit points is
690*4882a593Smuzhiyun * useful to assist debuggers stepping over the critical section.
691*4882a593Smuzhiyun */
692*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_EXIT_POINT(start_ip, exit_ip) \
693*4882a593Smuzhiyun ".pushsection __rseq_exit_point_array, \"aw\"\n\t" \
694*4882a593Smuzhiyun ".long " __rseq_str(start_ip) ", 0x0, " __rseq_str(exit_ip) ", 0x0\n\t" \
695*4882a593Smuzhiyun ".popsection\n\t"
696*4882a593Smuzhiyun
697*4882a593Smuzhiyun #define RSEQ_ASM_STORE_RSEQ_CS(label, cs_label, rseq_cs) \
698*4882a593Smuzhiyun RSEQ_INJECT_ASM(1) \
699*4882a593Smuzhiyun "movl $" __rseq_str(cs_label) ", " __rseq_str(rseq_cs) "\n\t" \
700*4882a593Smuzhiyun __rseq_str(label) ":\n\t"
701*4882a593Smuzhiyun
702*4882a593Smuzhiyun #define RSEQ_ASM_CMP_CPU_ID(cpu_id, current_cpu_id, label) \
703*4882a593Smuzhiyun RSEQ_INJECT_ASM(2) \
704*4882a593Smuzhiyun "cmpl %[" __rseq_str(cpu_id) "], " __rseq_str(current_cpu_id) "\n\t" \
705*4882a593Smuzhiyun "jnz " __rseq_str(label) "\n\t"
706*4882a593Smuzhiyun
707*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_ABORT(label, teardown, abort_label) \
708*4882a593Smuzhiyun ".pushsection __rseq_failure, \"ax\"\n\t" \
709*4882a593Smuzhiyun /* Disassembler-friendly signature: ud1 <sig>,%edi. */ \
710*4882a593Smuzhiyun ".byte 0x0f, 0xb9, 0x3d\n\t" \
711*4882a593Smuzhiyun ".long " __rseq_str(RSEQ_SIG) "\n\t" \
712*4882a593Smuzhiyun __rseq_str(label) ":\n\t" \
713*4882a593Smuzhiyun teardown \
714*4882a593Smuzhiyun "jmp %l[" __rseq_str(abort_label) "]\n\t" \
715*4882a593Smuzhiyun ".popsection\n\t"
716*4882a593Smuzhiyun
717*4882a593Smuzhiyun #define RSEQ_ASM_DEFINE_CMPFAIL(label, teardown, cmpfail_label) \
718*4882a593Smuzhiyun ".pushsection __rseq_failure, \"ax\"\n\t" \
719*4882a593Smuzhiyun __rseq_str(label) ":\n\t" \
720*4882a593Smuzhiyun teardown \
721*4882a593Smuzhiyun "jmp %l[" __rseq_str(cmpfail_label) "]\n\t" \
722*4882a593Smuzhiyun ".popsection\n\t"
723*4882a593Smuzhiyun
724*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_storev(intptr_t * v,intptr_t expect,intptr_t newv,int cpu)725*4882a593Smuzhiyun int rseq_cmpeqv_storev(intptr_t *v, intptr_t expect, intptr_t newv, int cpu)
726*4882a593Smuzhiyun {
727*4882a593Smuzhiyun RSEQ_INJECT_C(9)
728*4882a593Smuzhiyun
729*4882a593Smuzhiyun __asm__ __volatile__ goto (
730*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
731*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
732*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
733*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
734*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
735*4882a593Smuzhiyun #endif
736*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
737*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
738*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
739*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
740*4882a593Smuzhiyun "cmpl %[v], %[expect]\n\t"
741*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
742*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
743*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
744*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
745*4882a593Smuzhiyun "cmpl %[v], %[expect]\n\t"
746*4882a593Smuzhiyun "jnz %l[error2]\n\t"
747*4882a593Smuzhiyun #endif
748*4882a593Smuzhiyun /* final store */
749*4882a593Smuzhiyun "movl %[newv], %[v]\n\t"
750*4882a593Smuzhiyun "2:\n\t"
751*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
752*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
753*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
754*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
755*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
756*4882a593Smuzhiyun [v] "m" (*v),
757*4882a593Smuzhiyun [expect] "r" (expect),
758*4882a593Smuzhiyun [newv] "r" (newv)
759*4882a593Smuzhiyun : "memory", "cc", "eax"
760*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
761*4882a593Smuzhiyun : abort, cmpfail
762*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
763*4882a593Smuzhiyun , error1, error2
764*4882a593Smuzhiyun #endif
765*4882a593Smuzhiyun );
766*4882a593Smuzhiyun rseq_after_asm_goto();
767*4882a593Smuzhiyun return 0;
768*4882a593Smuzhiyun abort:
769*4882a593Smuzhiyun rseq_after_asm_goto();
770*4882a593Smuzhiyun RSEQ_INJECT_FAILED
771*4882a593Smuzhiyun return -1;
772*4882a593Smuzhiyun cmpfail:
773*4882a593Smuzhiyun rseq_after_asm_goto();
774*4882a593Smuzhiyun return 1;
775*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
776*4882a593Smuzhiyun error1:
777*4882a593Smuzhiyun rseq_after_asm_goto();
778*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
779*4882a593Smuzhiyun error2:
780*4882a593Smuzhiyun rseq_after_asm_goto();
781*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
782*4882a593Smuzhiyun #endif
783*4882a593Smuzhiyun }
784*4882a593Smuzhiyun
785*4882a593Smuzhiyun /*
786*4882a593Smuzhiyun * Compare @v against @expectnot. When it does _not_ match, load @v
787*4882a593Smuzhiyun * into @load, and store the content of *@v + voffp into @v.
788*4882a593Smuzhiyun */
789*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpnev_storeoffp_load(intptr_t * v,intptr_t expectnot,long voffp,intptr_t * load,int cpu)790*4882a593Smuzhiyun int rseq_cmpnev_storeoffp_load(intptr_t *v, intptr_t expectnot,
791*4882a593Smuzhiyun long voffp, intptr_t *load, int cpu)
792*4882a593Smuzhiyun {
793*4882a593Smuzhiyun RSEQ_INJECT_C(9)
794*4882a593Smuzhiyun
795*4882a593Smuzhiyun __asm__ __volatile__ goto (
796*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
797*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
798*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
799*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
800*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
801*4882a593Smuzhiyun #endif
802*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
803*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
804*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
805*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
806*4882a593Smuzhiyun "movl %[v], %%ebx\n\t"
807*4882a593Smuzhiyun "cmpl %%ebx, %[expectnot]\n\t"
808*4882a593Smuzhiyun "je %l[cmpfail]\n\t"
809*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
810*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
811*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
812*4882a593Smuzhiyun "movl %[v], %%ebx\n\t"
813*4882a593Smuzhiyun "cmpl %%ebx, %[expectnot]\n\t"
814*4882a593Smuzhiyun "je %l[error2]\n\t"
815*4882a593Smuzhiyun #endif
816*4882a593Smuzhiyun "movl %%ebx, %[load]\n\t"
817*4882a593Smuzhiyun "addl %[voffp], %%ebx\n\t"
818*4882a593Smuzhiyun "movl (%%ebx), %%ebx\n\t"
819*4882a593Smuzhiyun /* final store */
820*4882a593Smuzhiyun "movl %%ebx, %[v]\n\t"
821*4882a593Smuzhiyun "2:\n\t"
822*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
823*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
824*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
825*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
826*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
827*4882a593Smuzhiyun /* final store input */
828*4882a593Smuzhiyun [v] "m" (*v),
829*4882a593Smuzhiyun [expectnot] "r" (expectnot),
830*4882a593Smuzhiyun [voffp] "ir" (voffp),
831*4882a593Smuzhiyun [load] "m" (*load)
832*4882a593Smuzhiyun : "memory", "cc", "eax", "ebx"
833*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
834*4882a593Smuzhiyun : abort, cmpfail
835*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
836*4882a593Smuzhiyun , error1, error2
837*4882a593Smuzhiyun #endif
838*4882a593Smuzhiyun );
839*4882a593Smuzhiyun rseq_after_asm_goto();
840*4882a593Smuzhiyun return 0;
841*4882a593Smuzhiyun abort:
842*4882a593Smuzhiyun rseq_after_asm_goto();
843*4882a593Smuzhiyun RSEQ_INJECT_FAILED
844*4882a593Smuzhiyun return -1;
845*4882a593Smuzhiyun cmpfail:
846*4882a593Smuzhiyun rseq_after_asm_goto();
847*4882a593Smuzhiyun return 1;
848*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
849*4882a593Smuzhiyun error1:
850*4882a593Smuzhiyun rseq_after_asm_goto();
851*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
852*4882a593Smuzhiyun error2:
853*4882a593Smuzhiyun rseq_after_asm_goto();
854*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
855*4882a593Smuzhiyun #endif
856*4882a593Smuzhiyun }
857*4882a593Smuzhiyun
858*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_addv(intptr_t * v,intptr_t count,int cpu)859*4882a593Smuzhiyun int rseq_addv(intptr_t *v, intptr_t count, int cpu)
860*4882a593Smuzhiyun {
861*4882a593Smuzhiyun RSEQ_INJECT_C(9)
862*4882a593Smuzhiyun
863*4882a593Smuzhiyun __asm__ __volatile__ goto (
864*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
865*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
866*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
867*4882a593Smuzhiyun #endif
868*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
869*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
870*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
871*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
872*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
873*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
874*4882a593Smuzhiyun #endif
875*4882a593Smuzhiyun /* final store */
876*4882a593Smuzhiyun "addl %[count], %[v]\n\t"
877*4882a593Smuzhiyun "2:\n\t"
878*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
879*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
880*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
881*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
882*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
883*4882a593Smuzhiyun /* final store input */
884*4882a593Smuzhiyun [v] "m" (*v),
885*4882a593Smuzhiyun [count] "ir" (count)
886*4882a593Smuzhiyun : "memory", "cc", "eax"
887*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
888*4882a593Smuzhiyun : abort
889*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
890*4882a593Smuzhiyun , error1
891*4882a593Smuzhiyun #endif
892*4882a593Smuzhiyun );
893*4882a593Smuzhiyun rseq_after_asm_goto();
894*4882a593Smuzhiyun return 0;
895*4882a593Smuzhiyun abort:
896*4882a593Smuzhiyun rseq_after_asm_goto();
897*4882a593Smuzhiyun RSEQ_INJECT_FAILED
898*4882a593Smuzhiyun return -1;
899*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
900*4882a593Smuzhiyun error1:
901*4882a593Smuzhiyun rseq_after_asm_goto();
902*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
903*4882a593Smuzhiyun #endif
904*4882a593Smuzhiyun }
905*4882a593Smuzhiyun
906*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trystorev_storev(intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t newv2,intptr_t newv,int cpu)907*4882a593Smuzhiyun int rseq_cmpeqv_trystorev_storev(intptr_t *v, intptr_t expect,
908*4882a593Smuzhiyun intptr_t *v2, intptr_t newv2,
909*4882a593Smuzhiyun intptr_t newv, int cpu)
910*4882a593Smuzhiyun {
911*4882a593Smuzhiyun RSEQ_INJECT_C(9)
912*4882a593Smuzhiyun
913*4882a593Smuzhiyun __asm__ __volatile__ goto (
914*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
915*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
916*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
917*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
918*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
919*4882a593Smuzhiyun #endif
920*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
921*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
922*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
923*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
924*4882a593Smuzhiyun "cmpl %[v], %[expect]\n\t"
925*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
926*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
927*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
928*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
929*4882a593Smuzhiyun "cmpl %[v], %[expect]\n\t"
930*4882a593Smuzhiyun "jnz %l[error2]\n\t"
931*4882a593Smuzhiyun #endif
932*4882a593Smuzhiyun /* try store */
933*4882a593Smuzhiyun "movl %[newv2], %%eax\n\t"
934*4882a593Smuzhiyun "movl %%eax, %[v2]\n\t"
935*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
936*4882a593Smuzhiyun /* final store */
937*4882a593Smuzhiyun "movl %[newv], %[v]\n\t"
938*4882a593Smuzhiyun "2:\n\t"
939*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
940*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
941*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
942*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
943*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
944*4882a593Smuzhiyun /* try store input */
945*4882a593Smuzhiyun [v2] "m" (*v2),
946*4882a593Smuzhiyun [newv2] "m" (newv2),
947*4882a593Smuzhiyun /* final store input */
948*4882a593Smuzhiyun [v] "m" (*v),
949*4882a593Smuzhiyun [expect] "r" (expect),
950*4882a593Smuzhiyun [newv] "r" (newv)
951*4882a593Smuzhiyun : "memory", "cc", "eax"
952*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
953*4882a593Smuzhiyun : abort, cmpfail
954*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
955*4882a593Smuzhiyun , error1, error2
956*4882a593Smuzhiyun #endif
957*4882a593Smuzhiyun );
958*4882a593Smuzhiyun rseq_after_asm_goto();
959*4882a593Smuzhiyun return 0;
960*4882a593Smuzhiyun abort:
961*4882a593Smuzhiyun rseq_after_asm_goto();
962*4882a593Smuzhiyun RSEQ_INJECT_FAILED
963*4882a593Smuzhiyun return -1;
964*4882a593Smuzhiyun cmpfail:
965*4882a593Smuzhiyun rseq_after_asm_goto();
966*4882a593Smuzhiyun return 1;
967*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
968*4882a593Smuzhiyun error1:
969*4882a593Smuzhiyun rseq_after_asm_goto();
970*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
971*4882a593Smuzhiyun error2:
972*4882a593Smuzhiyun rseq_after_asm_goto();
973*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
974*4882a593Smuzhiyun #endif
975*4882a593Smuzhiyun }
976*4882a593Smuzhiyun
977*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trystorev_storev_release(intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t newv2,intptr_t newv,int cpu)978*4882a593Smuzhiyun int rseq_cmpeqv_trystorev_storev_release(intptr_t *v, intptr_t expect,
979*4882a593Smuzhiyun intptr_t *v2, intptr_t newv2,
980*4882a593Smuzhiyun intptr_t newv, int cpu)
981*4882a593Smuzhiyun {
982*4882a593Smuzhiyun RSEQ_INJECT_C(9)
983*4882a593Smuzhiyun
984*4882a593Smuzhiyun __asm__ __volatile__ goto (
985*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
986*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
987*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
988*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
989*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
990*4882a593Smuzhiyun #endif
991*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
992*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
993*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
994*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
995*4882a593Smuzhiyun "movl %[expect], %%eax\n\t"
996*4882a593Smuzhiyun "cmpl %[v], %%eax\n\t"
997*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
998*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
999*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1000*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
1001*4882a593Smuzhiyun "movl %[expect], %%eax\n\t"
1002*4882a593Smuzhiyun "cmpl %[v], %%eax\n\t"
1003*4882a593Smuzhiyun "jnz %l[error2]\n\t"
1004*4882a593Smuzhiyun #endif
1005*4882a593Smuzhiyun /* try store */
1006*4882a593Smuzhiyun "movl %[newv2], %[v2]\n\t"
1007*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
1008*4882a593Smuzhiyun "lock; addl $0,-128(%%esp)\n\t"
1009*4882a593Smuzhiyun /* final store */
1010*4882a593Smuzhiyun "movl %[newv], %[v]\n\t"
1011*4882a593Smuzhiyun "2:\n\t"
1012*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
1013*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
1014*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
1015*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
1016*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
1017*4882a593Smuzhiyun /* try store input */
1018*4882a593Smuzhiyun [v2] "m" (*v2),
1019*4882a593Smuzhiyun [newv2] "r" (newv2),
1020*4882a593Smuzhiyun /* final store input */
1021*4882a593Smuzhiyun [v] "m" (*v),
1022*4882a593Smuzhiyun [expect] "m" (expect),
1023*4882a593Smuzhiyun [newv] "r" (newv)
1024*4882a593Smuzhiyun : "memory", "cc", "eax"
1025*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
1026*4882a593Smuzhiyun : abort, cmpfail
1027*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1028*4882a593Smuzhiyun , error1, error2
1029*4882a593Smuzhiyun #endif
1030*4882a593Smuzhiyun );
1031*4882a593Smuzhiyun rseq_after_asm_goto();
1032*4882a593Smuzhiyun return 0;
1033*4882a593Smuzhiyun abort:
1034*4882a593Smuzhiyun rseq_after_asm_goto();
1035*4882a593Smuzhiyun RSEQ_INJECT_FAILED
1036*4882a593Smuzhiyun return -1;
1037*4882a593Smuzhiyun cmpfail:
1038*4882a593Smuzhiyun rseq_after_asm_goto();
1039*4882a593Smuzhiyun return 1;
1040*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1041*4882a593Smuzhiyun error1:
1042*4882a593Smuzhiyun rseq_after_asm_goto();
1043*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
1044*4882a593Smuzhiyun error2:
1045*4882a593Smuzhiyun rseq_after_asm_goto();
1046*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
1047*4882a593Smuzhiyun #endif
1048*4882a593Smuzhiyun
1049*4882a593Smuzhiyun }
1050*4882a593Smuzhiyun
1051*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_cmpeqv_storev(intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t expect2,intptr_t newv,int cpu)1052*4882a593Smuzhiyun int rseq_cmpeqv_cmpeqv_storev(intptr_t *v, intptr_t expect,
1053*4882a593Smuzhiyun intptr_t *v2, intptr_t expect2,
1054*4882a593Smuzhiyun intptr_t newv, int cpu)
1055*4882a593Smuzhiyun {
1056*4882a593Smuzhiyun RSEQ_INJECT_C(9)
1057*4882a593Smuzhiyun
1058*4882a593Smuzhiyun __asm__ __volatile__ goto (
1059*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
1060*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
1061*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1062*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
1063*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
1064*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error3])
1065*4882a593Smuzhiyun #endif
1066*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
1067*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
1068*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
1069*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
1070*4882a593Smuzhiyun "cmpl %[v], %[expect]\n\t"
1071*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
1072*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
1073*4882a593Smuzhiyun "cmpl %[expect2], %[v2]\n\t"
1074*4882a593Smuzhiyun "jnz %l[cmpfail]\n\t"
1075*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
1076*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1077*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), %l[error1])
1078*4882a593Smuzhiyun "cmpl %[v], %[expect]\n\t"
1079*4882a593Smuzhiyun "jnz %l[error2]\n\t"
1080*4882a593Smuzhiyun "cmpl %[expect2], %[v2]\n\t"
1081*4882a593Smuzhiyun "jnz %l[error3]\n\t"
1082*4882a593Smuzhiyun #endif
1083*4882a593Smuzhiyun "movl %[newv], %%eax\n\t"
1084*4882a593Smuzhiyun /* final store */
1085*4882a593Smuzhiyun "movl %%eax, %[v]\n\t"
1086*4882a593Smuzhiyun "2:\n\t"
1087*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
1088*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4, "", abort)
1089*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
1090*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
1091*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
1092*4882a593Smuzhiyun /* cmp2 input */
1093*4882a593Smuzhiyun [v2] "m" (*v2),
1094*4882a593Smuzhiyun [expect2] "r" (expect2),
1095*4882a593Smuzhiyun /* final store input */
1096*4882a593Smuzhiyun [v] "m" (*v),
1097*4882a593Smuzhiyun [expect] "r" (expect),
1098*4882a593Smuzhiyun [newv] "m" (newv)
1099*4882a593Smuzhiyun : "memory", "cc", "eax"
1100*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
1101*4882a593Smuzhiyun : abort, cmpfail
1102*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1103*4882a593Smuzhiyun , error1, error2, error3
1104*4882a593Smuzhiyun #endif
1105*4882a593Smuzhiyun );
1106*4882a593Smuzhiyun rseq_after_asm_goto();
1107*4882a593Smuzhiyun return 0;
1108*4882a593Smuzhiyun abort:
1109*4882a593Smuzhiyun rseq_after_asm_goto();
1110*4882a593Smuzhiyun RSEQ_INJECT_FAILED
1111*4882a593Smuzhiyun return -1;
1112*4882a593Smuzhiyun cmpfail:
1113*4882a593Smuzhiyun rseq_after_asm_goto();
1114*4882a593Smuzhiyun return 1;
1115*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1116*4882a593Smuzhiyun error1:
1117*4882a593Smuzhiyun rseq_after_asm_goto();
1118*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
1119*4882a593Smuzhiyun error2:
1120*4882a593Smuzhiyun rseq_after_asm_goto();
1121*4882a593Smuzhiyun rseq_bug("1st expected value comparison failed");
1122*4882a593Smuzhiyun error3:
1123*4882a593Smuzhiyun rseq_after_asm_goto();
1124*4882a593Smuzhiyun rseq_bug("2nd expected value comparison failed");
1125*4882a593Smuzhiyun #endif
1126*4882a593Smuzhiyun }
1127*4882a593Smuzhiyun
1128*4882a593Smuzhiyun /* TODO: implement a faster memcpy. */
1129*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trymemcpy_storev(intptr_t * v,intptr_t expect,void * dst,void * src,size_t len,intptr_t newv,int cpu)1130*4882a593Smuzhiyun int rseq_cmpeqv_trymemcpy_storev(intptr_t *v, intptr_t expect,
1131*4882a593Smuzhiyun void *dst, void *src, size_t len,
1132*4882a593Smuzhiyun intptr_t newv, int cpu)
1133*4882a593Smuzhiyun {
1134*4882a593Smuzhiyun uint32_t rseq_scratch[3];
1135*4882a593Smuzhiyun
1136*4882a593Smuzhiyun RSEQ_INJECT_C(9)
1137*4882a593Smuzhiyun
1138*4882a593Smuzhiyun __asm__ __volatile__ goto (
1139*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
1140*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
1141*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1142*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
1143*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
1144*4882a593Smuzhiyun #endif
1145*4882a593Smuzhiyun "movl %[src], %[rseq_scratch0]\n\t"
1146*4882a593Smuzhiyun "movl %[dst], %[rseq_scratch1]\n\t"
1147*4882a593Smuzhiyun "movl %[len], %[rseq_scratch2]\n\t"
1148*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
1149*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
1150*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
1151*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
1152*4882a593Smuzhiyun "movl %[expect], %%eax\n\t"
1153*4882a593Smuzhiyun "cmpl %%eax, %[v]\n\t"
1154*4882a593Smuzhiyun "jnz 5f\n\t"
1155*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
1156*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1157*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 6f)
1158*4882a593Smuzhiyun "movl %[expect], %%eax\n\t"
1159*4882a593Smuzhiyun "cmpl %%eax, %[v]\n\t"
1160*4882a593Smuzhiyun "jnz 7f\n\t"
1161*4882a593Smuzhiyun #endif
1162*4882a593Smuzhiyun /* try memcpy */
1163*4882a593Smuzhiyun "test %[len], %[len]\n\t" \
1164*4882a593Smuzhiyun "jz 333f\n\t" \
1165*4882a593Smuzhiyun "222:\n\t" \
1166*4882a593Smuzhiyun "movb (%[src]), %%al\n\t" \
1167*4882a593Smuzhiyun "movb %%al, (%[dst])\n\t" \
1168*4882a593Smuzhiyun "inc %[src]\n\t" \
1169*4882a593Smuzhiyun "inc %[dst]\n\t" \
1170*4882a593Smuzhiyun "dec %[len]\n\t" \
1171*4882a593Smuzhiyun "jnz 222b\n\t" \
1172*4882a593Smuzhiyun "333:\n\t" \
1173*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
1174*4882a593Smuzhiyun "movl %[newv], %%eax\n\t"
1175*4882a593Smuzhiyun /* final store */
1176*4882a593Smuzhiyun "movl %%eax, %[v]\n\t"
1177*4882a593Smuzhiyun "2:\n\t"
1178*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
1179*4882a593Smuzhiyun /* teardown */
1180*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1181*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1182*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t"
1183*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4,
1184*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1185*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1186*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1187*4882a593Smuzhiyun abort)
1188*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(5,
1189*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1190*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1191*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1192*4882a593Smuzhiyun cmpfail)
1193*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1194*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(6,
1195*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1196*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1197*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1198*4882a593Smuzhiyun error1)
1199*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(7,
1200*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1201*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1202*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1203*4882a593Smuzhiyun error2)
1204*4882a593Smuzhiyun #endif
1205*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
1206*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
1207*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
1208*4882a593Smuzhiyun /* final store input */
1209*4882a593Smuzhiyun [v] "m" (*v),
1210*4882a593Smuzhiyun [expect] "m" (expect),
1211*4882a593Smuzhiyun [newv] "m" (newv),
1212*4882a593Smuzhiyun /* try memcpy input */
1213*4882a593Smuzhiyun [dst] "r" (dst),
1214*4882a593Smuzhiyun [src] "r" (src),
1215*4882a593Smuzhiyun [len] "r" (len),
1216*4882a593Smuzhiyun [rseq_scratch0] "m" (rseq_scratch[0]),
1217*4882a593Smuzhiyun [rseq_scratch1] "m" (rseq_scratch[1]),
1218*4882a593Smuzhiyun [rseq_scratch2] "m" (rseq_scratch[2])
1219*4882a593Smuzhiyun : "memory", "cc", "eax"
1220*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
1221*4882a593Smuzhiyun : abort, cmpfail
1222*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1223*4882a593Smuzhiyun , error1, error2
1224*4882a593Smuzhiyun #endif
1225*4882a593Smuzhiyun );
1226*4882a593Smuzhiyun rseq_after_asm_goto();
1227*4882a593Smuzhiyun return 0;
1228*4882a593Smuzhiyun abort:
1229*4882a593Smuzhiyun rseq_after_asm_goto();
1230*4882a593Smuzhiyun RSEQ_INJECT_FAILED
1231*4882a593Smuzhiyun return -1;
1232*4882a593Smuzhiyun cmpfail:
1233*4882a593Smuzhiyun rseq_after_asm_goto();
1234*4882a593Smuzhiyun return 1;
1235*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1236*4882a593Smuzhiyun error1:
1237*4882a593Smuzhiyun rseq_after_asm_goto();
1238*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
1239*4882a593Smuzhiyun error2:
1240*4882a593Smuzhiyun rseq_after_asm_goto();
1241*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
1242*4882a593Smuzhiyun #endif
1243*4882a593Smuzhiyun }
1244*4882a593Smuzhiyun
1245*4882a593Smuzhiyun /* TODO: implement a faster memcpy. */
1246*4882a593Smuzhiyun static inline __attribute__((always_inline))
rseq_cmpeqv_trymemcpy_storev_release(intptr_t * v,intptr_t expect,void * dst,void * src,size_t len,intptr_t newv,int cpu)1247*4882a593Smuzhiyun int rseq_cmpeqv_trymemcpy_storev_release(intptr_t *v, intptr_t expect,
1248*4882a593Smuzhiyun void *dst, void *src, size_t len,
1249*4882a593Smuzhiyun intptr_t newv, int cpu)
1250*4882a593Smuzhiyun {
1251*4882a593Smuzhiyun uint32_t rseq_scratch[3];
1252*4882a593Smuzhiyun
1253*4882a593Smuzhiyun RSEQ_INJECT_C(9)
1254*4882a593Smuzhiyun
1255*4882a593Smuzhiyun __asm__ __volatile__ goto (
1256*4882a593Smuzhiyun RSEQ_ASM_DEFINE_TABLE(3, 1f, 2f, 4f) /* start, commit, abort */
1257*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[cmpfail])
1258*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1259*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error1])
1260*4882a593Smuzhiyun RSEQ_ASM_DEFINE_EXIT_POINT(1f, %l[error2])
1261*4882a593Smuzhiyun #endif
1262*4882a593Smuzhiyun "movl %[src], %[rseq_scratch0]\n\t"
1263*4882a593Smuzhiyun "movl %[dst], %[rseq_scratch1]\n\t"
1264*4882a593Smuzhiyun "movl %[len], %[rseq_scratch2]\n\t"
1265*4882a593Smuzhiyun /* Start rseq by storing table entry pointer into rseq_cs. */
1266*4882a593Smuzhiyun RSEQ_ASM_STORE_RSEQ_CS(1, 3b, RSEQ_ASM_TP_SEGMENT:RSEQ_CS_OFFSET(%[rseq_offset]))
1267*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 4f)
1268*4882a593Smuzhiyun RSEQ_INJECT_ASM(3)
1269*4882a593Smuzhiyun "movl %[expect], %%eax\n\t"
1270*4882a593Smuzhiyun "cmpl %%eax, %[v]\n\t"
1271*4882a593Smuzhiyun "jnz 5f\n\t"
1272*4882a593Smuzhiyun RSEQ_INJECT_ASM(4)
1273*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1274*4882a593Smuzhiyun RSEQ_ASM_CMP_CPU_ID(cpu_id, RSEQ_ASM_TP_SEGMENT:RSEQ_CPU_ID_OFFSET(%[rseq_offset]), 6f)
1275*4882a593Smuzhiyun "movl %[expect], %%eax\n\t"
1276*4882a593Smuzhiyun "cmpl %%eax, %[v]\n\t"
1277*4882a593Smuzhiyun "jnz 7f\n\t"
1278*4882a593Smuzhiyun #endif
1279*4882a593Smuzhiyun /* try memcpy */
1280*4882a593Smuzhiyun "test %[len], %[len]\n\t" \
1281*4882a593Smuzhiyun "jz 333f\n\t" \
1282*4882a593Smuzhiyun "222:\n\t" \
1283*4882a593Smuzhiyun "movb (%[src]), %%al\n\t" \
1284*4882a593Smuzhiyun "movb %%al, (%[dst])\n\t" \
1285*4882a593Smuzhiyun "inc %[src]\n\t" \
1286*4882a593Smuzhiyun "inc %[dst]\n\t" \
1287*4882a593Smuzhiyun "dec %[len]\n\t" \
1288*4882a593Smuzhiyun "jnz 222b\n\t" \
1289*4882a593Smuzhiyun "333:\n\t" \
1290*4882a593Smuzhiyun RSEQ_INJECT_ASM(5)
1291*4882a593Smuzhiyun "lock; addl $0,-128(%%esp)\n\t"
1292*4882a593Smuzhiyun "movl %[newv], %%eax\n\t"
1293*4882a593Smuzhiyun /* final store */
1294*4882a593Smuzhiyun "movl %%eax, %[v]\n\t"
1295*4882a593Smuzhiyun "2:\n\t"
1296*4882a593Smuzhiyun RSEQ_INJECT_ASM(6)
1297*4882a593Smuzhiyun /* teardown */
1298*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1299*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1300*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t"
1301*4882a593Smuzhiyun RSEQ_ASM_DEFINE_ABORT(4,
1302*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1303*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1304*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1305*4882a593Smuzhiyun abort)
1306*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(5,
1307*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1308*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1309*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1310*4882a593Smuzhiyun cmpfail)
1311*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1312*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(6,
1313*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1314*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1315*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1316*4882a593Smuzhiyun error1)
1317*4882a593Smuzhiyun RSEQ_ASM_DEFINE_CMPFAIL(7,
1318*4882a593Smuzhiyun "movl %[rseq_scratch2], %[len]\n\t"
1319*4882a593Smuzhiyun "movl %[rseq_scratch1], %[dst]\n\t"
1320*4882a593Smuzhiyun "movl %[rseq_scratch0], %[src]\n\t",
1321*4882a593Smuzhiyun error2)
1322*4882a593Smuzhiyun #endif
1323*4882a593Smuzhiyun : /* gcc asm goto does not allow outputs */
1324*4882a593Smuzhiyun : [cpu_id] "r" (cpu),
1325*4882a593Smuzhiyun [rseq_offset] "r" (rseq_offset),
1326*4882a593Smuzhiyun /* final store input */
1327*4882a593Smuzhiyun [v] "m" (*v),
1328*4882a593Smuzhiyun [expect] "m" (expect),
1329*4882a593Smuzhiyun [newv] "m" (newv),
1330*4882a593Smuzhiyun /* try memcpy input */
1331*4882a593Smuzhiyun [dst] "r" (dst),
1332*4882a593Smuzhiyun [src] "r" (src),
1333*4882a593Smuzhiyun [len] "r" (len),
1334*4882a593Smuzhiyun [rseq_scratch0] "m" (rseq_scratch[0]),
1335*4882a593Smuzhiyun [rseq_scratch1] "m" (rseq_scratch[1]),
1336*4882a593Smuzhiyun [rseq_scratch2] "m" (rseq_scratch[2])
1337*4882a593Smuzhiyun : "memory", "cc", "eax"
1338*4882a593Smuzhiyun RSEQ_INJECT_CLOBBER
1339*4882a593Smuzhiyun : abort, cmpfail
1340*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1341*4882a593Smuzhiyun , error1, error2
1342*4882a593Smuzhiyun #endif
1343*4882a593Smuzhiyun );
1344*4882a593Smuzhiyun rseq_after_asm_goto();
1345*4882a593Smuzhiyun return 0;
1346*4882a593Smuzhiyun abort:
1347*4882a593Smuzhiyun rseq_after_asm_goto();
1348*4882a593Smuzhiyun RSEQ_INJECT_FAILED
1349*4882a593Smuzhiyun return -1;
1350*4882a593Smuzhiyun cmpfail:
1351*4882a593Smuzhiyun rseq_after_asm_goto();
1352*4882a593Smuzhiyun return 1;
1353*4882a593Smuzhiyun #ifdef RSEQ_COMPARE_TWICE
1354*4882a593Smuzhiyun error1:
1355*4882a593Smuzhiyun rseq_after_asm_goto();
1356*4882a593Smuzhiyun rseq_bug("cpu_id comparison failed");
1357*4882a593Smuzhiyun error2:
1358*4882a593Smuzhiyun rseq_after_asm_goto();
1359*4882a593Smuzhiyun rseq_bug("expected value comparison failed");
1360*4882a593Smuzhiyun #endif
1361*4882a593Smuzhiyun }
1362*4882a593Smuzhiyun
1363*4882a593Smuzhiyun #endif /* !RSEQ_SKIP_FASTPATH */
1364*4882a593Smuzhiyun
1365*4882a593Smuzhiyun #endif
1366