On 23.01.20 20:52, Jens Axboe wrote: > On 1/23/20 10:28 AM, Mike Snitzer wrote: >> On Thu, Jan 23 2020 at 5:35am -0500, >> Mike Snitzer <snitzer@xxxxxxxxxx> wrote: >> >>> On Thu, Jan 23 2020 at 4:17am -0500, >>> Stefan Bader <stefan.bader@xxxxxxxxxxxxx> wrote: >>> >>>> When device-mapper adapted for multi-queue functionality, they >>>> also re-organized the way the make-request function was set. >>>> Before, this happened when the device-mapper logical device was >>>> created. Now it is done once the mapping table gets loaded the >>>> first time (this also decides whether the block device is request >>>> or bio based). >>>> >>>> However in generic_make_request(), the request function gets used >>>> without further checks and this happens if one tries to mount such >>>> a partially set up device. >>>> >>>> This can easily be reproduced with the following steps: >>>> - dmsetup create -n test >>>> - mount /dev/dm-<#> /mnt >>>> >>>> This maybe is something which also should be fixed up in device- >>>> mapper. >>> >>> I'll look closer at other options. >>> >>>> But given there is already a check for an unset queue >>>> pointer and potentially there could be other drivers which do or >>>> might do the same, it sounds like a good move to add another check >>>> to generic_make_request_checks() and to bail out if the request >>>> function has not been set, yet. >>>> >>>> BugLink: https://bugs.launchpad.net/bugs/1860231 >>> >>> >From that bug; >>> "The currently proposed fix introduces no chance of stability >>> regressions. There is a chance of a very small performance regression >>> since an additional pointer comparison is performed on each block layer >>> request but this is unlikely to be noticeable." >>> >>> This captures my immediate concern: slowing down everyone for this DM >>> edge-case isn't desirable. >> >> SO I had a look and there isn't anything easier than adding the proposed >> NULL check in generic_make_request_checks(). Given the many >> conditionals in that function.. what's one more? ;) >> >> I looked at marking the queue frozen to prevent IO via >> blk_queue_enter()'s existing cheeck -- but that quickly felt like an >> abuse, especially in that there isn't a queue unfreeze for bio-based. >> >> Jens, I'll defer to you to judge this patch further. If you're OK with >> it: cool. If not, I'm open to suggestions for how to proceed. >> > > It does kinda suck... The generic_make_request_checks() is a mess, and > this doesn't make it any better. Any reason why we can't solve this > two step setup in a clean fashion instead of patching around it like > this? Feels like a pretty bad hack, tbh. > Tyler spent some time thinking about delaying the allocation of the queue structure until later but that seemed rather dangerous. IIRC there are places during registration of the (generic) block device which expect this to be done. Not sure whether it would be feasible to start with one kind of dummy make_request_fn and then switch that over to the proper one once that decision can be made...
Attachment:
signature.asc
Description: OpenPGP digital signature
