On Fri, Oct 15, 2021 at 10:27:17AM +0200, Vlastimil Babka wrote:
On 10/14/21 12:16, Mike Rapoport wrote:
On Thu, Oct 14, 2021 at 11:33:03AM +0200, Vlastimil Babka wrote:
On 10/14/21 10:54, kernel test robot wrote:
In my local testing of the patch, when stackdepot was initialized through page owner init, it was using kvmalloc() so slab_is_available() was true. Looks like the exact order of slab vs page_owner alloc is non-deterministic, could be arch-dependent or just random ordering of init calls. A wrong order will exploit the apparent fact that slab_is_available() is not a good indicator of using memblock vs page allocator, and we would need a better one. Thoughts?
The order of slab vs page_owner is deterministic, but it is different for FLATMEM and SPARSEMEM. And page_ext_init_flatmem_late() that initializes page_ext for FLATMEM is called exactly between buddy and slab setup:
Oh, so it was due to FLATMEM, thanks for figuring that out!
static void __init mm_init(void) { ...
mem_init(); mem_init_print_info(); /* page_owner must be initialized after buddy is ready */ page_ext_init_flatmem_late(); kmem_cache_init();
... }
I've stared for a while at page_ext init and it seems that the page_ext_init_flatmem_late() can be simply dropped because there is anyway a call to invoke_init_callbacks() in page_ext_init() that is called much later in the boot process.
Yeah, but page_ext_init() only does something for SPARSEMEM, and is empty on FLATMEM. Otherwise it would be duplicating all the work. So I'll just move page_ext_init_flatmem_late() below kmem_cache_init() in mm_init().
I hope at some point we'll have cleaner mm_init(), but for now simply moving page_ext_init_flatmem_late() should work.
Thanks again!
Welcome :)