Re: [PATCH V3] fuse: Remove __GFP_FS flag to avoid allocator recursing

[ppvk@codeaurora.org]

On 2020-09-23 18:19, Matthew Wilcox wrote:
> On Wed, Sep 23, 2020 at 11:27:55AM +0530, Pradeep P V K wrote:
>> Changes since V2:
>> - updated memalloc_nofs_save() to allocation paths that potentially
>>   can cause deadlock.
> 
> That's the exact opposite of what I said to do.  Again, the *THREAD*
> is the thing which must not block, not the *ALLOCATION*.  So you
> set this flag *ON THE THREAD*, not *WHEN IT DOES AN ALLOCATION*.
> If that's not clear, please ask again.

The fuse threads are created and started in external libfuse userspace
library functions but not in Kernel. The lowest entry point for these 
threads
to enter in kernel is fuse_dev_read()/fuse_dev_splice_read().

So, can we suppose to use memalloc_nofs_save() from
external userspace library functions ?

Even if we used, can you confirm, if the context of memalloc_nofs_save()
can be persist in kernel ?  (when the thread enters into kernel space).

Also, i didn't see memalloc_nofs_save() been used/called from any
external userspace library functions.


Thanks and Regards,
Pradeep