Mike Kravetz [off-list ref] 于2021年11月29日周一 下午12:31写道：

On 11/28/21 03:18, Maxim Levitsky wrote:

quoted

dmesg prints this:

HugeTLB: allocating 64 of page size 1.00 GiB failed.  Only allocated 0 hugepages

Huge pages were allocated on kernel command line (1/2 of 128GB system):

'default_hugepagesz=1G hugepagesz=1G hugepages=64'

This is 3970X and no real support/need for NUMA, thus only fake NUMA node 0 is present.

Reverting the commit helps.

New syntax also works ( hugepages=0:64 )

I can test any patches for this bug.

Argh!  I think preallocation of gigantic pages on all systems with only
a single node is broken.  The issue is at the beginning of
__alloc_bootmem_huge_page:

int __alloc_bootmem_huge_page(struct hstate *h, int nid)
{
        struct huge_bootmem_page *m = NULL; /* initialize for clang */
        int nr_nodes, node;

        if (nid >= nr_online_nodes)
                return 0;

Without using the node specific syntax, nid == NUMA_NO_NODE == -1.  For the
comparison, nid will be converted to an unsigned into to match nr_online_nodes
so we will immediately return 0 instead of doing the allocations.

Zhenguo Yao,
Can you verify and perhaps put together a patch?does

Preallocation of gigantic pages cant‘ work in all the environment, not
only in single node.
I think the issue is because of the replacement
nodes_weight(node_states[N_MEMORY] with
nr_online_nodes in my patch of last version. Sorry for my careless. I
didn't notice that  parameter
nid is int ，but nr_online_nodes is unsigned int. so, this if (nid >=
nr_online_nodes) is always
true when nid is NUMA_NO_NODE(-1).  I will send a fix  as soon as passible.
This is really a low-level mistake ^^

quoted

Also unrelated, is there any progress on allocating 1GB pages on demand so that I could
allocate them only when I run a VM?

That should be possible.  Such support was added back in 2014 with commit
944d9fec8d7a "hugetlb: add support for gigantic page allocation at runtime".

quoted

i don't mind having these pages to be marked as to be used for userspace only,
since as far as I remember its the kernel usage that makes some page unmoveable.

Of course, finding 1GB of contiguous space for a gigantic page is often
difficult at runtime.  So, allocations are likely to fail the longer the
system is up and running and fragmentation increases.

quoted

Last time (many years ago) I tried to create a zone with only userspace pages
(I don't remember what options I used) but it didn't work.

Not too long ago, support was added to use CMA for gigantic page allocation.
See commit cf11e85fc08c "mm: hugetlb: optionally allocate gigantic hugepages
using cma".  This sounds like something you might want to try.
--
Mike Kravetz

quoted

Is there a way to debug what is causing unmoveable pages and doesn't let
/proc/sys/vm/nr_hugepages work (I tried it today and as usual the number
it can allocate steadly decreases over time).