Control: forwarded -1 https://github.com/jemalloc/jemalloc/issues/467 Control: tags -1 fixed-upstream
On 10/11/16 18:37, Thadeu Lima de Souza Cascardo wrote: > clone -1 -2 > reassign -2 libjemalloc1 > retitle -2 jemalloc uses a hard coded page size detected during build > bye > > > So, I traced this to jemalloc using the incorrect page size during > runtime. In fact, it detects the page size during build and set it up. > This has a large potential for a mess. And what a mess! > > So, jemalloc in jessie has been built on a 4KiB-page system, and, this, > has a hard-coded page size of 4KiB. While jemalloc in sid has a > 64KiB-page size. When we try to build mariadb on jessie on a 4KiB-page > size system, everything goes well. When we build it on partch, with a > 64-bit 64KiB-page size kernel, things break, if it's a jessie jemalloc. > > Later upstream versions seem to support multiple page sizes during > build. It looks like jemalloc 5.0.0 will support multiple page sizes automatically as long as you pass the biggest possible page size when you configure it: https://github.com/jemalloc/jemalloc/pull/769 That patch is pretty big though, so it won't help with stretch. > For MariaDB specifically, one option would be to build it without > jemalloc. Other users would still be likely affected, however. Limiting it to specific architectures is probably ok though. Thanks, James
signature.asc
Description: OpenPGP digital signature
