submit bio directly from userspace
Hi,
i'm developing block layer workload simulator/player and need to generate bios as close as possible to the ones in original workload. To play workload i employ aio + O_DIRECT, but all IOs end up as bios with SYNC flag set. So i wonder is there any way to control bios flags from userspace or, maybe, any other options to generate workload from userspace except aio + O_DIRECT with more precise control over bio flags?
Quoted 10 lines · Mike Krinkin
Hi,
i'm developing block layer workload simulator/player and need to generate bios as close as possible to the ones in original workload. To play workload i employ aio +
O_DIRECT, but all IOs end up as bios with SYNC flag set. So i wonder is there any way to control bios flags from userspace or, maybe, any other options to generate workload from userspace except aio +O_DIRECTwith more precise control over bio flags?
100% agree, I've investigated existing io-load replay tools and seems that they all use O_DIRECT(like btreplay) or buffered-aio/{sync_file_range,fadvise} ( like fio). So it looks like they all very inaccurate.
Jens, what do to do to replay real load in order to test fs/ioscheduler harness? Are you interested in having such fine-tuning interface (probably guarded by CAP_SYS_ADMIN)?
There are several places where we can pass such flag
1) iocb->aio_flag, add IOCB_FLAG_DIRECT_ASYNC
2) add new flag O_DIRECT_ASYNC to file->f_flags
Quoted 12 lines · Dmitry Monakhov
Quoted 10 lines · Mike Krinkin
Hi,
i'm developing block layer workload simulator/player and need to generate bios as close as possible to the ones in original workload. To play workload i employ aio +
O_DIRECT, but all IOs end up as bios with SYNC flag set. So i wonder is there any way to control bios flags from userspace or, maybe, any other options to generate workload from userspace except aio +O_DIRECTwith more precise control over bio flags?
Quoted 8 lines
100% agree, I've investigated existing io-load replay tools and seems that they all use
O_DIRECT(like btreplay)or buffered-aio/{sync_file_range,fadvise} ( like fio). So it looks like they all very inaccurate.Jens, what do to do to replay real load in order to test fs/ioscheduler harness? Are you interested in having such fine-tuning interface (probably guarded by
CAP_SYS_ADMIN)?
A long time ago I wrote binject, which was meant to be a way to directly inject bio's on a queue:
http://git.kernel.dk/cgit/linux-block/log/?h=binject
It hasn't been moved forward, however. I haven't looked, is is SYNC the only flag addition we are worried about here? There are a slew of other flags that may impact how IO is treated. For instance, file system has tagged the request META. We can't pass that in either. Same is true for FLUSH or FUA writes.
Signature · 1 line
Jens Axboe