On Sat, Mar 12, 2011 at 5:58 AM, Arne Jansen <lists@xxxxxxxxxxxxxx> wrote: > Jens Axboe wrote: >> >> On 2011-03-12 14:33, Arne Jansen wrote: >>> >>> Jens Axboe wrote: >>>> >>>> On 2011-03-12 10:29, Arne Jansen wrote: >>>>> >>>>> Hi, >>>>> >>>>> I'm trying to re-use struct bio after completion, including the >>>>> allocated pages. Normally I re-initialize the fields bi_sector, >>>>> bi_size, bi_next, bi_flags, bi_comp_cpu and bi_bdev. >>>>> This works perfectly well, as long as no media errors are encountered. >>>>> After a media error, all subsequent reads with this bio fail. Are >>>>> there any more fields that need to get re-initialized? Or, better, >>>>> is there a function to reset the bio? >>>> >>>> bio_init()? Sounds like you are not setting BIO_UPTODATE when resetting >>>> it. >>>> >>> bio_init does way too much, as it discards the io_vec, refcnt, >>> destructor, >>> private etc. >>> I set the flags to 1 << BIO_UPTODATE, but that proved to be fatal, too, >>> as >>> it clobbers the POOL_FLAGS, which leads to an oops when the last >>> reference >>> drops. These bio-beasts are just not made to be reused. One needs way too >>> much internal knowledge to reinitialize them properly. >>> Therefore, I'd really like to have a call like bio_reinit, which keeps >>> the >>> io_vec and all the owners private information, but resets the parts that >>> get used by the stack like bi_next. Would that make sense? >> >> The thing is, you can't really reuse without violating the allocator >> principle of finite life times. Otherwise you risk hanging the system. >> If you want to reuse, the bio should come out of your own pool or >> through the bio_kmalloc() interface, not from bio_alloc() with the >> default pool. >> >> So there is no reinit, as there hasn't been a good use case for that. >> The fact that flags has pool information in the top bits is a sure sign >> you are violating that. >> > > ok, assuming I allocate them with bio_kmalloc, the problem of reinit would > remain. Now that I own them, I can't reuse them. > This particular use case here is a scrub for btrfs. I have a limited set of > bios and pages, that I reuse as long as the scrub runs. Of course I could > just free up and reallocate everything after each read, but as I know how > much work there is ahead, I can just as well keep them. Have a look at drivers/md/raid5.c::ops_run_io() the stripe_head bios are allocated once at initialization and reused for the life of the array. -- Dan -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
