背景
我们执行一个大SQL时(长度大于512M),会返回如下错误:
ERROR: invalid memory alloc request size 1073741824
复现
我们首先复现出来这个问题
- 创建表
create table byteatable(id int, obj bytea);
- 插入512M大对象
#!/bin/bash data='a' for ((i=1;i<=29;i++)); do data=$data$data done echo 'build ok' psql -U postgres -d postgres << EOF insert into byteatable(id,obj) values (1,"$data"); EOF echo 'OK'
执行脚本后就能复现出来了
ERROR: invalid memory alloc request size 1073741824
BUG分析
我们先找到出现这个错误的位置。
源码位置:
void *
MemoryContextAlloc(MemoryContext context, Size size)
{
void *ret;
AssertArg(MemoryContextIsValid(context));
if (!AllocSizeIsValid(size))
elog(ERROR, "invalid memory alloc request size %zu", size);
context->isReset = false;
ret = (*context->methods->alloc) (context, size);
VALGRIND_MEMPOOL_ALLOC(context, ret, size);
return ret;
}
#define AllocSizeIsValid(size) ((Size) (size) <= MaxAllocSize)
#define MaxAllocSize ((Size) 0x3fffffff) /* 1 gigabyte - 1 */
这里限制的内存是1G - 1,而我们插入的大SQL需要的内存没有1G,那么为什么还提示非法的申请内存大小呢?通过调试跟踪,我们发现是在词法分析的时刻出错的。主要问题是在词法分析的内存申请机制上,申请的内存肯定不能少于sql的长度,如果当前申请的内存不够用,那么将重新申请当前内存乘以2的内存大小。
int literallen; /* actual current string length */
int literalalloc; /* current allocated buffer size */
//当前申请内存初始化
yyext->literalalloc = 1024;
static void
addlit(char *ytext, int yleng, core_yyscan_t yyscanner)
{
/* enlarge buffer if needed */
if ((yyextra->literallen + yleng) >= yyextra->literalalloc)
{
do {
yyextra->literalalloc *= 2;
} while ((yyextra->literallen + yleng) >= yyextra->literalalloc);
yyextra->literalbuf = (char *) repalloc(yyextra->literalbuf,
yyextra->literalalloc);
}
/* append new data */
memcpy(yyextra->literalbuf + yyextra->literallen, ytext, yleng);
yyextra->literallen += yleng;
}
static void
addlitchar(unsigned char ychar, core_yyscan_t yyscanner)
{
/* enlarge buffer if needed */
if ((yyextra->literallen + 1) >= yyextra->literalalloc)
{
yyextra->literalalloc *= 2;
yyextra->literalbuf = (char *) repalloc(yyextra->literalbuf,
yyextra->literalalloc);
}
}
从源码中可以看出,每次申请原申请内存的2倍,即yyextra->literalalloc *= 2; 而最大申请内存限制是
#define MaxAllocSize ((Size) 0x3fffffff) /* 1 gigabyte - 1 */
所以我们在词法分析能申请的最大内存是2^29 = 536870912,如果词法分析SQL语句需要的内存大于536870912,那么申请的内存需要再乘2,就会得到2^30 = 1073741824,超过MaxAllocSize=0x3fffffff= 1073741823。所以会提示错误:
ERROR: invalid memory alloc request size 1073741824
当然不仅仅是插入一个大对象才会引起这个问题,只要是SQL语句长度大于512M都是出现这个错误,我们可以使用select复现:
do language plpgsql $$
declare
v_text text := 'a';
begin
for i in 1..29 loop
v_text:=v_text||v_text;
end loop;
execute $_$select '$_$||v_text||$_$'$_$;
raise notice 'execute a sql large than 512MB success.';
exception when others then
raise notice 'execute a sql large than 512MB failed.';
end;
$$;
BUG修复
其实申请MaxAllocSize是可行的,通过修改源码实现,每当申请的内存大于MaxAllocSize并且SQL需要的长度小于MaxAllocSize时,我们就申请MaxAllocSize大小的内存。
static void
addlit(char *ytext, int yleng, core_yyscan_t yyscanner)
{
/* enlarge buffer if needed */
if ((yyextra->literallen + yleng) >= yyextra->literalalloc)
{
do {
yyextra->literalalloc *= 2;
} while ((yyextra->literallen + yleng) >= yyextra->literalalloc);
/* we can not alloc more than MaxAllocSize */
if (yyextra->literalalloc > MaxAllocSize && (yyextra->literallen + yleng) < MaxAllocSize)
yyextra->literalalloc = MaxAllocSize;
yyextra->literalbuf = (char *) repalloc(yyextra->literalbuf,
yyextra->literalalloc);
}
/* append new data */
memcpy(yyextra->literalbuf + yyextra->literallen, ytext, yleng);
yyextra->literallen += yleng;
}
static void
addlitchar(unsigned char ychar, core_yyscan_t yyscanner)
{
/* enlarge buffer if needed */
if ((yyextra->literallen + 1) >= yyextra->literalalloc)
{
yyextra->literalalloc *= 2;
/* we can not alloc more than MaxAllocSize */
if (yyextra->literalalloc > MaxAllocSize && (yyextra->literallen + 1) < MaxAllocSize)
yyextra->literalalloc = MaxAllocSize;
yyextra->literalbuf = (char *) repalloc(yyextra->literalbuf,
yyextra->literalalloc);
}
/* append new data */
yyextra->literalbuf[yyextra->literallen] = ychar;
yyextra->literallen += 1;
}
修复之后可以正常插入了
$ sh pgtest.sh
build ok
INSERT 0 1
OK
时间: 2024-09-29 22:04:39