Re: [PATCH] [RFC] block: Don't let blk_put_request leak BIOs

[Boaz Harrosh <bharrosh@panasas.com>]

Boaz Harrosh wrote:
> If a block ULD had allocated a request and mapped some memory into it,
> using one of blk_rq_map_xxx routines, but then for some reason failed to
> execute the request through one of the blk_execute_request routines.
> Then the associated BIO would leak, unless ULD resorts to low-level loops
> intimate of block internals.
> 
> [RFC]
>   This code will also catch situations where LLD failed to complete
>   the request before aborting it. Such situations are a BUG. Should we
>   use WARN_ON_ONCE() in that case. The situation above is possible and
>   can happen normally in memory pressure situations so maybe we should
>   devise a bit-flag that ULD denotes that the request was aborted and
>   only WARN_ON if flag was not set.
> 
>   I'm sending this before any-tests so people can comment on possible
>   pitfalls.
> 
> Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
> ---

I've booted a Linux PC with lots of sata disks, connected an iscsi
target, ran OSD tests. It looks like it's working which means
request->bio is set to NULL after it is used in the regular path.

This needs to sit in Linux next and be tested for a long while.

Jens I'll be waiting for your comment and will send a proper
patch for the block bits. We will have to time this with James
to see when the OSD bits can be submitted after that, then TOMO's
patch for un-exporting blk_req_append_bio can be merged. Or maybe
it can all go in one patch through scsi?

Thanks
Boaz