Mercurial > hg > nginx
view src/os/unix/ngx_gcc_atomic_x86.h @ 9300:5be23505292b default tip
SSI: fixed incorrect or duplicate stub output.
Following 3518:eb3aaf8bd2a9 (0.8.37), r->request_output is only set
if there are data in the first buffer sent in the subrequest. As a
result, following the change mentioned this flag cannot be used to
prevent duplicate ngx_http_ssi_stub_output() calls, since it is not
set if there was already some output, but the first buffer was empty.
Still, when there are multiple subrequests, even an empty subrequest
response might be delayed by the postpone filter, leading to a second
call of ngx_http_ssi_stub_output() during finalization from
ngx_http_writer() the subreqest buffers are released by the postpone
filter. Since r->request_output is not set after the first call, this
resulted in duplicate stub output.
Additionally, checking only the first buffer might be wrong in some
unusual cases. For example, the first buffer might be empty if
$r->flush() is called before printing any data in the embedded Perl
module.
Depending on the postpone_output value and corresponding sizes, this
issue can result in either duplicate or unexpected stub output, or
"zero size buf in writer" alerts.
Following 8124:f5515e727656 (1.23.4), it became slightly easier to
reproduce the issue, as empty static files and empty cache items now
result in a response with an empty buffer. Before the change, an empty
proxied response can be used to reproduce the issue.
Fix is check all buffers and set r->request_output if any non-empty
buffers are sent. This ensures that all unusual cases of non-empty
responses are covered, and also that r->request_output will be set
after the first stub output, preventing duplicate output.
Reported by Jan Gassen.
author | Maxim Dounin <mdounin@mdounin.ru> |
---|---|
date | Thu, 04 Jul 2024 17:41:28 +0300 |
parents | d620f497c50f |
children |
line wrap: on
line source
/* * Copyright (C) Igor Sysoev * Copyright (C) Nginx, Inc. */ #if (NGX_SMP) #define NGX_SMP_LOCK "lock;" #else #define NGX_SMP_LOCK #endif /* * "cmpxchgl r, [m]": * * if (eax == [m]) { * zf = 1; * [m] = r; * } else { * zf = 0; * eax = [m]; * } * * * The "r" means the general register. * The "=a" and "a" are the %eax register. * Although we can return result in any register, we use "a" because it is * used in cmpxchgl anyway. The result is actually in %al but not in %eax, * however, as the code is inlined gcc can test %al as well as %eax, * and icc adds "movzbl %al, %eax" by itself. * * The "cc" means that flags were changed. */ static ngx_inline ngx_atomic_uint_t ngx_atomic_cmp_set(ngx_atomic_t *lock, ngx_atomic_uint_t old, ngx_atomic_uint_t set) { u_char res; __asm__ volatile ( NGX_SMP_LOCK " cmpxchgl %3, %1; " " sete %0; " : "=a" (res) : "m" (*lock), "a" (old), "r" (set) : "cc", "memory"); return res; } /* * "xaddl r, [m]": * * temp = [m]; * [m] += r; * r = temp; * * * The "+r" means the general register. * The "cc" means that flags were changed. */ #if !(( __GNUC__ == 2 && __GNUC_MINOR__ <= 7 ) || ( __INTEL_COMPILER >= 800 )) /* * icc 8.1 and 9.0 compile broken code with -march=pentium4 option: * ngx_atomic_fetch_add() always return the input "add" value, * so we use the gcc 2.7 version. * * icc 8.1 and 9.0 with -march=pentiumpro option or icc 7.1 compile * correct code. */ static ngx_inline ngx_atomic_int_t ngx_atomic_fetch_add(ngx_atomic_t *value, ngx_atomic_int_t add) { __asm__ volatile ( NGX_SMP_LOCK " xaddl %0, %1; " : "+r" (add) : "m" (*value) : "cc", "memory"); return add; } #else /* * gcc 2.7 does not support "+r", so we have to use the fixed * %eax ("=a" and "a") and this adds two superfluous instructions in the end * of code, something like this: "mov %eax, %edx / mov %edx, %eax". */ static ngx_inline ngx_atomic_int_t ngx_atomic_fetch_add(ngx_atomic_t *value, ngx_atomic_int_t add) { ngx_atomic_uint_t old; __asm__ volatile ( NGX_SMP_LOCK " xaddl %2, %1; " : "=a" (old) : "m" (*value), "a" (add) : "cc", "memory"); return old; } #endif /* * on x86 the write operations go in a program order, so we need only * to disable the gcc reorder optimizations */ #define ngx_memory_barrier() __asm__ volatile ("" ::: "memory") /* old "as" does not support "pause" opcode */ #define ngx_cpu_pause() __asm__ (".byte 0xf3, 0x90")