Re: [PATCH 05/12] mm: support async buffered reads in generic_file_buffered_read()

From: Jens Axboe <axboe@kernel.dk>
To: Trond Myklebust <trondmy@hammerspace.com>,
	"io-uring@vger.kernel.org" <io-uring@vger.kernel.org>
Cc: "linux-mm@kvack.org" <linux-mm@kvack.org>,
	"linux-fsdevel@vger.kernel.org" <linux-fsdevel@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 05/12] mm: support async buffered reads in generic_file_buffered_read()
Date: Sun, 24 May 2020 10:30:42 -0600	[thread overview]
Message-ID: <2fa7104a-ea85-55f2-692c-514eb3b88a88@kernel.dk> (raw)
In-Reply-To: <264614fc4fa08df2b0899da1cd38bb07150cd7f3.camel@hammerspace.com>

On 5/24/20 8:05 AM, Trond Myklebust wrote:
> On Sat, 2020-05-23 at 12:57 -0600, Jens Axboe wrote:
>> Use the async page locking infrastructure, if IOCB_WAITQ is set in
>> the
>> passed in iocb. The caller must expect an -EIOCBQUEUED return value,
>> which means that IO is started but not done yet. This is similar to
>> how
>> O_DIRECT signals the same operation. Once the callback is received by
>> the caller for IO completion, the caller must retry the operation.
>>
>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>> ---
>>  mm/filemap.c | 33 ++++++++++++++++++++++++++-------
>>  1 file changed, 26 insertions(+), 7 deletions(-)
>>
>> diff --git a/mm/filemap.c b/mm/filemap.c
>> index c746541b1d49..a3b86c9acdc8 100644
>> --- a/mm/filemap.c
>> +++ b/mm/filemap.c
>> @@ -1219,6 +1219,14 @@ static int __wait_on_page_locked_async(struct
>> page *page,
>>  	return ret;
>>  }
>>  
>> +static int wait_on_page_locked_async(struct page *page,
>> +				     struct wait_page_queue *wait)
>> +{
>> +	if (!PageLocked(page))
>> +		return 0;
>> +	return __wait_on_page_locked_async(compound_head(page), wait,
>> false);
>> +}
>> +
>>  /**
>>   * put_and_wait_on_page_locked - Drop a reference and wait for it to
>> be unlocked
>>   * @page: The page to wait for.
>> @@ -2058,17 +2066,25 @@ static ssize_t
>> generic_file_buffered_read(struct kiocb *iocb,
>>  					index, last_index - index);
>>  		}
>>  		if (!PageUptodate(page)) {
>> -			if (iocb->ki_flags & IOCB_NOWAIT) {
>> -				put_page(page);
>> -				goto would_block;
>> -			}
>> -
>>  			/*
>>  			 * See comment in do_read_cache_page on why
>>  			 * wait_on_page_locked is used to avoid
>> unnecessarily
>>  			 * serialisations and why it's safe.
>>  			 */
>> -			error = wait_on_page_locked_killable(page);
>> +			if (iocb->ki_flags & IOCB_WAITQ) {
>> +				if (written) {
>> +					put_page(page);
>> +					goto out;
>> +				}
>> +				error = wait_on_page_locked_async(page,
>> +								iocb-
>>> private);
> 
> If it is being used in 'generic_file_buffered_read()' as storage for a
> wait queue, then it is hard to consider this a 'private' field.

private isn't the prettiest, and in fact this one in particular is a bit
of a mess. It's not clear if it's caller or callee owned. It's generally
not used, outside of the old usb gadget code, iomap O_DIRECT, and ocfs2.
With FMODE_BUF_RASYNC, the fs obviously can't set it if it uses
->private for buffered IO.

> Perhaps either rename and add type checking, or else add a separate
> field altogether to struct kiocb?

I'd hate to add a new field and increase the size of the kiocb... One
alternative is to do:

	union {
		void *private;
		struct wait_page_queue *ki_waitq;
	};

and still use IOCB_WAITQ to say that ->ki_waitq is valid.

There's also 4 bytes of padding in the kiocb struct. And some fields are
only used for O_DIRECT as well, eg ->ki_cookie which is just used for
polled O_DIRECT. So we could also do:

	union {
		unsigned int ki_cookie;
		struct wait_page_queue *ki_waitq;
	};

and still not grow the kiocb. How about we go with this approach, and
also add:

	if (kiocb->ki_flags & IOCB_HIPRI)
		return -EOPNOTSUPP;

to kiocb_wait_page_queue_init() to make sure that this combination isn't
valid?

-- 
Jens Axboe