On Thu, Jul 11, 2019 at 10:30:52PM +0000, co...@sdf.org wrote:
> hi folks, while skimming jemalloc, noticed a comment about our old code
> being better.
* Ideally, there would be a way to specify alignment to mmap() (like
* NetBSD has), but in the absence of such a feature, we have to work
* hard to efficiently create aligned mappings. The reliable, but
* slow method is to create a mapping that is over-sized, then trim the
* excess. However, that always results in one or two calls to
* os_pages_unmap(), and it can leave holes in the process's virtual
* memory map if memory grows downward.
*
* Optimistically try mapping precisely the right amount before falling
* back to the slow method, with the expectation that the optimistic
* approach works most of the time.
So if that doesn't work, it'll loop in pages_map_slow, hoping to get aligned
stuff.
