| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* pg_numa.c |
| 4 |
* Basic NUMA portability routines |
| 5 |
* |
| 6 |
* |
| 7 |
* Copyright (c) 2025, PostgreSQL Global Development Group |
| 8 |
* |
| 9 |
* |
| 10 |
* IDENTIFICATION |
| 11 |
* src/port/pg_numa.c |
| 12 |
* |
| 13 |
*------------------------------------------------------------------------- |
| 14 |
*/ |
| 15 |
|
| 16 |
#include "c.h" |
| 17 |
#include <unistd.h> |
| 18 |
|
| 19 |
#include "miscadmin.h" |
| 20 |
#include "port/pg_numa.h" |
| 21 |
|
| 22 |
/* |
| 23 |
* At this point we provide support only for Linux thanks to libnuma, but in |
| 24 |
* future support for other platforms e.g. Win32 or FreeBSD might be possible |
| 25 |
* too. For Win32 NUMA APIs see |
| 26 |
* https://learn.microsoft.com/en-us/windows/win32/procthread/numa-support |
| 27 |
*/ |
| 28 |
#ifdef USE_LIBNUMA |
| 29 |
|
| 30 |
#include <numa.h> |
| 31 |
#include <numaif.h> |
| 32 |
|
| 33 |
/* |
| 34 |
* numa_move_pages() chunk size, has to be <= 16 to work around a kernel bug |
| 35 |
* in do_pages_stat() (chunked by DO_PAGES_STAT_CHUNK_NR). By using the same |
| 36 |
* chunk size, we make it work even on unfixed kernels. |
| 37 |
* |
| 38 |
* 64-bit system are not affected by the bug, and so use much larger chunks. |
| 39 |
*/ |
| 40 |
#if SIZEOF_SIZE_T == 4 |
| 41 |
#define NUMA_QUERY_CHUNK_SIZE 16 |
| 42 |
#else |
| 43 |
#define NUMA_QUERY_CHUNK_SIZE 1024 |
| 44 |
#endif |
| 45 |
|
| 46 |
/* libnuma requires initialization as per numa(3) on Linux */ |
| 47 |
int |
| 48 |
pg_numa_init(void) |
| 49 |
{ |
| 50 |
int r; |
| 51 |
|
| 52 |
/* |
| 53 |
* XXX libnuma versions before 2.0.19 don't handle EPERM by disabling |
| 54 |
* NUMA, which then leads to unexpected failures later. This affects |
| 55 |
* containers that disable get_mempolicy by a seccomp profile. |
| 56 |
*/ |
| 57 |
if (get_mempolicy(NULL, NULL, 0, 0, 0) < 0 && (errno == EPERM)) |
| 58 |
r = -1; |
| 59 |
else |
| 60 |
r = numa_available(); |
| 61 |
|
| 62 |
return r; |
| 63 |
} |
| 64 |
|
| 65 |
/* |
| 66 |
* We use move_pages(2) syscall here - instead of get_mempolicy(2) - as the |
| 67 |
* first one allows us to batch and query about many memory pages in one single |
| 68 |
* giant system call that is way faster. |
| 69 |
* |
| 70 |
* We call numa_move_pages() for smaller chunks of the whole array. The first |
| 71 |
* reason is to work around a kernel bug, but also to allow interrupting the |
| 72 |
* query between the calls (for many pointers processing the whole array can |
| 73 |
* take a lot of time). |
| 74 |
*/ |
| 75 |
int |
| 76 |
pg_numa_query_pages(int pid, unsigned long count, void **pages, int *status) |
| 77 |
{ |
| 78 |
unsigned long next = 0; |
| 79 |
int ret = 0; |
| 80 |
|
| 81 |
/* |
| 82 |
* Chunk pointers passed to numa_move_pages to NUMA_QUERY_CHUNK_SIZE |
| 83 |
* items, to work around a kernel bug in do_pages_stat(). |
| 84 |
*/ |
| 85 |
while (next < count) |
| 86 |
{ |
| 87 |
unsigned long count_chunk = Min(count - next, |
| 88 |
NUMA_QUERY_CHUNK_SIZE); |
| 89 |
|
| 90 |
#ifndef FRONTEND |
| 91 |
CHECK_FOR_INTERRUPTS(); |
| 92 |
#endif |
| 93 |
|
| 94 |
/* |
| 95 |
* Bail out if any of the chunks errors out (ret<0). We ignore (ret>0) |
| 96 |
* which is used to return number of nonmigrated pages, but we're not |
| 97 |
* migrating any pages here. |
| 98 |
*/ |
| 99 |
ret = numa_move_pages(pid, count_chunk, &pages[next], NULL, &status[next], 0); |
| 100 |
if (ret < 0) |
| 101 |
{ |
| 102 |
/* plain error, return as is */ |
| 103 |
return ret; |
| 104 |
} |
| 105 |
|
| 106 |
next += count_chunk; |
| 107 |
} |
| 108 |
|
| 109 |
/* should have consumed the input array exactly */ |
| 110 |
Assert(next == count); |
| 111 |
|
| 112 |
return 0; |
| 113 |
} |
| 114 |
|
| 115 |
int |
| 116 |
pg_numa_get_max_node(void) |
| 117 |
{ |
| 118 |
return numa_max_node(); |
| 119 |
} |
| 120 |
|
| 121 |
#else |
| 122 |
|
| 123 |
/* Empty wrappers */ |
| 124 |
int |
| 125 |
pg_numa_init(void) |
| 126 |
{ |
| 127 |
/* We state that NUMA is not available */ |
| 128 |
return -1; |
| 129 |
} |
| 130 |
|
| 131 |
int |
| 132 |
pg_numa_query_pages(int pid, unsigned long count, void **pages, int *status) |
| 133 |
{ |
| 134 |
return 0; |
| 135 |
} |
| 136 |
|
| 137 |
int |
| 138 |
pg_numa_get_max_node(void) |
| 139 |
{ |
| 140 |
return 0; |
| 141 |
} |
| 142 |
|
| 143 |
#endif |