Re: [PATCH v4] rbd: fix the memory leak of bio_chain_clone

[Alex Elder <elder@inktank.com>]

On 08/08/2012 09:46 AM, Alex Elder wrote:
> On 08/07/2012 07:49 PM, Alex Elder wrote:
>> On 08/03/2012 10:33 AM, Guangliang Zhao wrote:
>>> The bio_pair alloced in bio_chain_clone would not be freed,
>>> this will cause a memory leak. It could be freed actually only
>>> after 3 times release, because the reference count of bio_pair
>>> is initialized to 3 when bio_split and bio_pair_release only
>>> drops the reference count.
>>
>> Last Friday I added this patch to the testing branch and
>> ran the xfstests-on-rbd suite on it, and it triggered some
>> new errors.  I did not have time to investigate them though.
>>
>> I am taking this latest version and will apply it and will
>> run it through the same series of tests.
>
>
> Testing again, test 013 in the xfstests suite failed.  (Sorry
> I don't have a record of where failures occurred on Friday;
> maybe Yehuda does.)  The failure was that xfs_repair found an
> inconsistent file system after running fsstress (I believe it
> was running 20 concurrent copies).
>
> I ran through the tests again, and the second time it did not
> encounter a problem on test 013.  It hit a similar problem
> on test 269, which also runs fsstress, as well as a series
> of "dd" commands.  Again the problem was that the XFS
> filesystem running over an rbd device was found to be corrupt
> after the test.
>
> I will review the patch today to see if I can offer any
> insights about why we might be hitting these apparently
> timing-sensitive errors.

I've been reviewing the code today.  I found a problem with your
patch, which I'll describe a little more below.  I do agree that
there is a problem--a memory leak of the bio_pair created via
bio_split() in bio_chain_clone().

However I think it would be preferable to have a solution that
fixes the underlying problem with how we're using the result of
a bio_split() call instead of what you've proposed here.  The
whole purpose of bio_split() is to handle exactly the kind of
situation we're facing here, and I think it's unwise to try to
invent a different way of handling this scenario.

Granted, it may not be easy to fix the rbd code to fit the model
most other callers use when they call bio_split().  However if
the result of bio_chain_clone() produced one bio chain that ended
with bp->bio1 and a second chain that began with bp->bio2, and
then called bio_pair_release() after either the rbd_req_write()
or rbd_req_read() request, then there might no longer be the
leak.

I haven't tested what I point out below but I think it might offer
a clue about why your patch is resulting in test failures.

					-Alex



>                      -Alex
>
>
>
>> I also had hoped to give your patch a careful review, but
>> so far I haven't managed to do that (yet).
>>
>> I just want you to know you haven't been forgotten...
>>
>>                      -Alex
>>>
>>> The function bio_pair_release must be called three times for
>>> releasing bio_pair, and the callback functions of bios on the
>>> requests will be called when the last release time in bio_pair_release,
>>> however, these functions will also be called in rbd_req_cb. In
>>> other words, they will be called twice, and it may cause serious
>>> consequences.
>>>
>>> This patch clones bio chain from the origin directly instead of
>>> bio_split. The new bio chain can be released whenever we don't
>>> need it.
>>>
>>> This patch can just handle the split of *single page* bios, but
>>> it's enough here for the following reasons:
>>>
>>> Only bios span across multiple osds need to be split, and these bios
>>> *must* be single page because of rbd_merge_bvec. With the function,
>>> the new bvec will not permitted to merge, if it make the bio cross
>>> the osd boundary, except it is the first one. In other words, there
>>> are two types of bio:
>>>
>>>     - the bios don't cross the osd boundary
>>>       They have one or more pages. The value of offset will
>>>       always be 0 in this case, so nothing will be changed, and
>>>       the code changes tmp bios doesn't take effact at all.
>>>
>>>     - the bios cross the osd boundary
>>>       Each one have only one page. These bios need to be split,
>>>       and the offset is used to indicate the next bio, it makes
>>>       sense only in this instance.
>>>
>>> Signed-off-by: Guangliang Zhao <gzhao@suse.com>
>>> ---
>>>   drivers/block/rbd.c |   73
>>> +++++++++++++++++++++-----------------------------
>>>   1 files changed, 31 insertions(+), 42 deletions(-)
>>>
>>> diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c
>>> index 013c7a5..356657d 100644
>>> --- a/drivers/block/rbd.c
>>> +++ b/drivers/block/rbd.c
>>> @@ -712,51 +712,46 @@ static void zero_bio_chain(struct bio *chain,
>>> int start_ofs)
>>>       }
>>>   }
>>>
>>> -/*
>>> - * bio_chain_clone - clone a chain of bios up to a certain length.
>>> - * might return a bio_pair that will need to be released.
>>> +/**
>>> + *      bio_chain_clone - clone a chain of bios up to a certain length.
>>> + *      @old: bio to clone
>>> + *      @offset: start point for bio clone
>>> + *      @len: length of bio chain
>>> + *      @gfp_mask: allocation priority
>>> + *
>>> + *      RETURNS:
>>> + *      Pointer to new bio chain on success, NULL on failure.
>>>    */
>>> -static struct bio *bio_chain_clone(struct bio **old, struct bio **next,
>>> -                   struct bio_pair **bp,
>>> +static struct bio *bio_chain_clone(struct bio **old, int *offset,
>>>                      int len, gfp_t gfpmask)
>>>   {
>>>       struct bio *tmp, *old_chain = *old, *new_chain = NULL, *tail =
>>> NULL;
>>>       int total = 0;
>>>
>>> -    if (*bp) {
>>> -        bio_pair_release(*bp);
>>> -        *bp = NULL;
>>> -    }
>>> -
>>>       while (old_chain && (total < len)) {
>>> +        int need = len - total;
>>> +
>>>           tmp = bio_kmalloc(gfpmask, old_chain->bi_max_vecs);
>>>           if (!tmp)
>>>               goto err_out;
>>>
>>> -        if (total + old_chain->bi_size > len) {
>>> -            struct bio_pair *bp;
>>> -
>>> -            /*
>>> -             * this split can only happen with a single paged bio,
>>> -             * split_bio will BUG_ON if this is not the case
>>> -             */
>>> -            dout("bio_chain_clone split! total=%d remaining=%d"
>>> -                 "bi_size=%d\n",
>>> -                 (int)total, (int)len-total,
>>> -                 (int)old_chain->bi_size);
>>> -
>>> -            /* split the bio. We'll release it either in the next
>>> -               call, or it will have to be released outside */
>>> -            bp = bio_split(old_chain, (len - total) / SECTOR_SIZE);
>>> -            if (!bp)
>>> -                goto err_out;
>>> -
>>> -            __bio_clone(tmp, &bp->bio1);
>>> -
>>> -            *next = &bp->bio2;
>>> +        __bio_clone(tmp, old_chain);
>>> +        tmp->bi_sector += *offset >> SECTOR_SHIFT;
>>> +        tmp->bi_io_vec->bv_offset += *offset >> SECTOR_SHIFT;
>>> +        /*
>>> +         * The bios span across multiple osd objects must be
>>> +         * single paged, rbd_merge_bvec would guarantee it.
>>> +         * So we needn't worry about other things.
>>> +         */
>>> +        if (tmp->bi_size - *offset > need) {
>>> +            tmp->bi_size = need;
>>> +            tmp->bi_io_vec->bv_len = need;
>>> +            *offset += need;

This is the splitting case.

You are updating the new bio (tmp) by making its size and
iovec length be the number of bytes in the first part of
the original bio--the part being split off.

But you are *not* updating the original bio to reflect this.
That is, I think you need something like:

		old_chain->bi_size -= need;
		old_chain->bi_io_vec->bv_offset += need;
		old_chain->bi_io_vec->bv_len -= need;

>>>           } else {
>>> -            __bio_clone(tmp, old_chain);
>>> -            *next = old_chain->bi_next;
>>> +            old_chain = old_chain->bi_next;
>>> +            tmp->bi_size -= *offset;
>>> +            tmp->bi_io_vec->bv_len -= *offset;
>>> +            *offset = 0;
>>>           }
>>>
>>>           tmp->bi_bdev = NULL;
>>> @@ -769,7 +764,6 @@ static struct bio *bio_chain_clone(struct bio
>>> **old, struct bio **next,
>>>               tail->bi_next = tmp;
>>>               tail = tmp;
>>>           }
>>> -        old_chain = old_chain->bi_next;
>>>
>>>           total += tmp->bi_size;
>>>       }
>>> @@ -1447,13 +1441,12 @@ static void rbd_rq_fn(struct request_queue *q)
>>>   {
>>>       struct rbd_device *rbd_dev = q->queuedata;
>>>       struct request *rq;
>>> -    struct bio_pair *bp = NULL;
>>>
>>>       while ((rq = blk_fetch_request(q))) {
>>>           struct bio *bio;
>>> -        struct bio *rq_bio, *next_bio = NULL;
>>> +        struct bio *rq_bio;
>>>           bool do_write;
>>> -        int size, op_size = 0;
>>> +        int size, op_size = 0, offset = 0;
>>>           u64 ofs;
>>>           int num_segs, cur_seg = 0;
>>>           struct rbd_req_coll *coll;
>>> @@ -1503,7 +1496,7 @@ static void rbd_rq_fn(struct request_queue *q)
>>>                             ofs, size,
>>>                             NULL, NULL);
>>>               kref_get(&coll->kref);
>>> -            bio = bio_chain_clone(&rq_bio, &next_bio, &bp,
>>> +            bio = bio_chain_clone(&rq_bio, &offset,
>>>                             op_size, GFP_ATOMIC);
>>>               if (!bio) {
>>>                   rbd_coll_end_req_index(rq, coll, cur_seg,
>>> @@ -1531,12 +1524,8 @@ next_seg:
>>>               ofs += op_size;
>>>
>>>               cur_seg++;
>>> -            rq_bio = next_bio;
>>>           } while (size > 0);
>>>           kref_put(&coll->kref, rbd_coll_release);
>>> -
>>> -        if (bp)
>>> -            bio_pair_release(bp);
>>>           spin_lock_irq(q->queue_lock);
>>>       }
>>>   }
>>>
>>
>