On Fri, Jul 26, 2024 at 8:38 PM Baoquan He <bhe@xxxxxxxxxx> wrote:
>
> On 07/26/24 at 05:29pm, Barry Song wrote:
> > On Fri, Jul 26, 2024 at 5:04 PM Hailong Liu <hailong.liu@xxxxxxxx> wrote:
> > >
> > > On Fri, 26. Jul 12:00, Hailong Liu wrote:
> > > > On Fri, 26. Jul 10:31, Baoquan He wrote:
> > > > [...]
> > > > > > The logic of this patch is somewhat similar to my first one. If high order
> > > > > > allocation fails, it will go normal mapping.
> > > > > >
> > > > > > However I also save the fallback position. The ones before this position are
> > > > > > used for huge mapping, the ones >= position for normal mapping as Barry said.
> > > > > > "support the combination of PMD and PTE mapping". this will take some
> > > > > > times as it needs to address the corner cases and do some tests.
> > > > >
> > > > > Hmm, we may not need to worry about the imperfect mapping. Currently
> > > > > there are two places setting VM_ALLOW_HUGE_VMAP: __kvmalloc_node_noprof()
> > > > > and vmalloc_huge().
> > > > >
> > > > > For vmalloc_huge(), it's called in below three interfaces which are all
> > > > > invoked during boot. Basically they can succeed to get required contiguous
> > > > > physical memory. I guess that's why Tangquan only spot this issue on kvmalloc
> > > > > invocation when the required size exceeds e.g 2M. For kvmalloc_node(),
> > > > > we have told that in the code comment above __kvmalloc_node_noprof(),
> > > > > it's a best effort behaviour.
> > > > >
> > > > Take a __vmalloc_node_range(2.1M, VM_ALLOW_HUGE_VMAP) as a example.
> > > > because the align requirement of huge. the real size is 4M.
> > > > if allocation first order-9 successfully and the next failed. becuase the
> > > > fallback, the layout out pages would be like order9 - 512 * order0
> > > > order9 support huge mapping, but order0 not.
> > > > with the patch above, would call vmap_small_pages_range_noflush() and do normal
> > > > mapping, the huge mapping would not exist.
> > > >
> > > > > mm/mm_init.c <<alloc_large_system_hash>>
> > > > > table = vmalloc_huge(size, gfp_flags);
> > > > > net/ipv4/inet_hashtables.c <<inet_pernet_hashinfo_alloc>>
> > > > > new_hashinfo->ehash = vmalloc_huge(ehash_entries * sizeof(struct inet_ehash_bucket),
> > > > > net/ipv4/udp.c <<udp_pernet_table_alloc>>
> > > > > udptable->hash = vmalloc_huge(hash_entries * 2 * sizeof(struct udp_hslot)
> > > > >
> > > > > Maybe we should add code comment or document to notice people that the
> > > > > contiguous physical pages are not guaranteed for vmalloc_huge() if you
> > > > > use it after boot.
> > > > >
> > > > > >
> > > > > > IMO, the draft can fix the current issue, it also does not have significant side
> > > > > > effects. Barry, what do you think about this patch? If you think it's okay,
> > > > > > I will split this patch into two: one to remove the VM_ALLOW_HUGE_VMAP and the
> > > > > > other to address the current mapping issue.
> > > > > >
> > > > > > --
> > > > > > help you, help me,
> > > > > > Hailong.
> > > > > >
> > > > >
> > > > >
> > > I check the code, the issue only happen in gfp_mask with __GFP_NOFAIL and
> > > fallback to order 0, actuaally without this commit
> > > e9c3cda4d86e ("mm, vmalloc: fix high order __GFP_NOFAIL allocations")
> > > if __vmalloc_area_node allocation failed, it will goto fail and try order-0.
> > >
> > > fail:
> > > if (shift > PAGE_SHIFT) {
> > > shift = PAGE_SHIFT;
> > > align = real_align;
> > > size = real_size;
> > > goto again;
> > > }
> > >
> > > So do we really need fallback to order-0 if nofail?
> >
> > Good catch, this is what I missed. I feel we can revert Michal's fix.
> > And just remove __GFP_NOFAIL bit when we are still allocating
> > by high-order. When "goto again" happens, we will allocate by
> > order-0, in this case, we keep the __GFP_NOFAIL.
>
> With Michal's patch, the fallback will be able to satisfy the allocation
> for nofail case because it fallback to 0-order plus __GFP_NOFAIL. The
My point is that vm_area_alloc_pages() is an internal function and
just an implementation
detail. As long as its caller, __vmalloc_area_node(), can support
NOFAIL, it's fine.
Therefore, we can skip NOFAIL support for high-order allocations in
vm_area_alloc_pages()
and limit GFP_NOFAIL support to order-0.
Good news is that __vmalloc_node_range_noprof() has already a way to fallback
to order-0
fail:
if (shift > PAGE_SHIFT) {
shift = PAGE_SHIFT;
align = real_align;
size = real_size;
goto again;
}
So, we can definitely utilize this fallback instead of implementing it
within vm_area_alloc_pages(),
which would alter the page_order of vm_area and create inconsistency,
crashing the system
due to memory corruption.
With higher-level fallback in __vmalloc_node_range_noprof(), we won't
need an unusual fix-up
for vm_area as you're proposing. The page_order of vm_area will
consistently stay the same.
If there is any way to improvement, we may also add:
diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index caf032f0bd69..03d8148d7a02 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -3806,7 +3806,7 @@ void *__vmalloc_node_range_noprof(unsigned long
size, unsigned long align,
warn_alloc(gfp_mask, NULL,
"vmalloc error: size %lu, vm_struct allocation
failed%s",
real_size, (nofail) ? ". Retrying." : "");
- if (nofail) {
+ if (nofail && shift == PAGE_SHIFT) {
schedule_timeout_uninterruptible(1);
goto again;
}
There's no need to keep retrying for __get_vm_area_node() until success,
as we will succeed when we fall back to order-0.
> 'if (shift > PAGE_SHIFT)' conditional checking and handling may be
> problemtic since it could jump to fail becuase vmap_pages_range()
> invocation failed, or partially allocate huge parges and break down,
> then it will ignore the already allocated pages, and do all the thing again.
>
> The only thing 'if (shift > PAGE_SHIFT)' checking and handling makes
> sense is it fallback to the real_size and real_align. BUT we need handle
> the fail separately, e.g
> 1)__get_vm_area_node() failed;
> 2)vm_area_alloc_pages() failed when shift > PAGE_SHIFT and non-nofail;
> 3)vmap_pages_range() failed;
>
> Honestly, I didn't see where the nofail is mishandled, could you point
> it out specifically? I could miss it.
>
> Thanks
> Baoquan
>
Thanks
Barry
