Re: [PATCH V2] fuse: Remove __GFP_FS flag to avoid allocator recursing

From: Matthew Wilcox <willy@infradead.org>
To: Pradeep P V K <ppvk@codeaurora.org>
Cc: miklos@szeredi.hu, linux-fsdevel@vger.kernel.org,
	stummala@codeaurora.org, sayalil@codeaurora.org
Subject: Re: [PATCH V2] fuse: Remove __GFP_FS flag to avoid allocator recursing
Date: Wed, 16 Sep 2020 15:56:34 +0100	[thread overview]
Message-ID: <20200916145634.GN5449@casper.infradead.org> (raw)
In-Reply-To: <1600238380-33350-1-git-send-email-ppvk@codeaurora.org>

On Wed, Sep 16, 2020 at 12:09:40PM +0530, Pradeep P V K wrote:
> Changes since V1:
> - Used memalloc_nofs_save() in all allocation paths of fuse daemons
>   to avoid use __GFP_FS flag as per Matthew comments.

That's not how to use memalloc_nofs_save().  You call it when entering a
context in which any memory allocation would cause a deadlock.  You don't
look for every place which allocates memory and wrap the memory allocation
calls in memalloc_nofs_save() because you're likely to miss one.

>  static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
>  {
> +	ssize_t size;
> +	unsigned nofs_flag;

This is almost certainly too low in the call stack.