Skip to content

LittleFox94/pmalloc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

.github/workflows/test-and-build.yaml codecov

This is an allocator for physically continuous memory in linux userspace.

Implementation details:

Memory is requested from the system with mmap, allocating the smallest hugepage size available on the system matching the request - so on x86_64 this means either a 4KiB, a 2MiB or a 1GiB page. Memory not used in a page will be used for other allocations.

Each block allocated from the system has a header struct pmem, storing the originally allocated size (how big the mmap'd block is), the amount of memory in use/freed and the physical address of the start of the block, which is retrieved by reading /proc/self/pagemap, this requires special privileges! Check proc(5) for details. This header also contains pointer to the previous and the next one, forming a double-linked list of memory blocks.

When the user of this library makes an allocation, first every already allocated block is checked if it has enough space for that allocation. Only space after the last allocation will be used, so if you a = pmalloc(), b=pmalloc(), free(a), the memory used by a is not usable again - only the space needed for it is added to the size_freed of the block. This design decision makes the metadata to store per block very simple and fits the use case I needed this for - allocating some memory on start of program, maybe allocating some more at some point and freeing at the end.

If a memory blocks size_freed matches the size_used, it will be given back to the system, so removed from the linked list and munmap'd.

The available huge page sizes are retrieved by listing the directories in /sys/kernel/mm/hugepages (retrieve_hugepage_sizes does this). Before doing a mmap, pmalloc will check if enough (mostly 1, if more it might already be problematic) pages of the needed size are available by checking /sys/kernel/mm/hugepages/$dir/free_hugepages. If there aren't, it tries to increase nr_hugepages in the same directory; this might work, but probably won't when the system is online for some time already, since memory fragmentation is a thing (check the linux docs for details: admin-guide/mm/hugetlbpage).

The whole thing is quite chatty on anything unexpected. There is no way to make it non-chatty, deal with it.

Tests are failing!

One of the tests is allocating a 3MB chunk. On x86_64, this needs to allocate a 1GiB huge page. Since those aren't allocated by default, you systems memory might be to fragmented to allocate one. There is a linux cmdline flag to allocate some on boot (again, check out linux docs on that admin-guide/mm/hugetlbpage). If only the 3MiB test is failing, you can consider it as success. I might change the test one day to only log that as error when the system would have had the hugepages for that.

What did you need this for?

Testing LF OS drivers on linux in userspace, especially one of the few drivers LF OS has in kernel space: XHCI DbC (debug class). Unbinding the device driver from the PCI node and then running the driver code against it, using this library for physical memory allocations.

It's very cursed, and also kinda funny - since this is one of the few drivers in LF OS kernel space, it makes a lot of sense to test it in linux userspace - with linux having most drivers in the kernel :D

About

Allocating physical memory in linux userspace

Topics

Resources

Stars

Watchers

Forks

Sponsor this project