Re: Question about vmalloc(GFP_NOFS)

[Pavel Tikhomirov <ptikhomirov@virtuozzo.com>]

On 11/19/24 13:48, Pavel Tikhomirov wrote:
> Hello,
> 
> I see that in kernel code we have couple of places where kvmalloc is 
> used with GFP_NOFS flag:
> 
> git grep kvmalloc.*NOFS
> fs/bcachefs/journal_io.c:       new_buf = kvmalloc(new_size, GFP_NOFS| 
> __GFP_NOWARN);
> fs/ext4/xattr.c:                buffer = kvmalloc(value_size, GFP_NOFS);
> fs/f2fs/compress.c:     cc->private = f2fs_kvmalloc(F2FS_I_SB(cc- 
>  >inode), size, GFP_NOFS);
> net/ceph/osdmap.c:      state = kvmalloc(array_size(max, 
> sizeof(*state)), GFP_NOFS);
> net/ceph/osdmap.c:      weight = kvmalloc(array_size(max, 
> sizeof(*weight)), GFP_NOFS);
> net/ceph/osdmap.c:      addr = kvmalloc(array_size(max, sizeof(*addr)), 
> GFP_NOFS);
> 
> and with GFP_NOIO flag too:
> 
> git grep kvmalloc.*NOIO
> drivers/md/dm-integrity.c:      recalc_tags = kvmalloc(recalc_tags_size, 
> GFP_NOIO);
> drivers/md/dm-ioctl.c:  dmi = kvmalloc(param_kernel->data_size, GFP_NOIO 
> | __GFP_HIGH);
> net/ceph/messenger_v2.c:        buf = kvmalloc(len, GFP_NOIO);
> net/ceph/osdmap.c:      work = kvmalloc(work_size, GFP_NOIO);
> 
> And AFAIU documentation https://docs.kernel.org/core-api/gfp_mask-from- 
> fs-io.html#what-about-vmalloc-gfp-nofs vmalloc allocation with GFP_NOFS 
> may end up doing "GFP_KERNEL allocations deep inside the allocator", 
> which can potentially lead to deadlock in IO/FS code paths.
> 
> Does it mean that we should rework all those paths to 
> memalloc_noio_save / memalloc_noio_restore variant? Or is it already 
> safe to use kvmalloc(GFP_NOIO) in modern kernel?
> 
> Or maybe I misunderstand something, sorry in advance if that's the case.
> 

Now when I've already sent a question I seemingly found the answer:

In commit 451769ebb7e79 ("mm/vmalloc: alloc GFP_NO{FS,IO} for vmalloc") 
we add implicit memalloc_noXX_save/memalloc_noXX_restore at this code path:

   +->kvmalloc
     +-> ...
       +-> __kvmalloc_node_noprof
         +-> __vmalloc_node_range_noprof
           +-> __vmalloc_area_node

So kvmalloc should be safe now with GFP_NOIO. Should we correct the 
documentation?

-- 
Best regards, Tikhomirov Pavel
Senior Software Developer, Virtuozzo.