On 01/27/2017 05:10 AM, Hannes Reinecke wrote:

On 01/26/2017 08:48 PM, Jens Axboe wrote:

quoted

I've been diving into the problems that Hannes reported, both the
stalls related to shared tag maps and multiple hardware queues,
but also the cases where we get suboptimal merging.

This series attempts to fix that.

Hannes, this should be identical to what I sent you as a bundled up
patch earlier today. Would be great if you could run this through
your testing.

The patches are against for-4.11/block

Yep, full success:

4k  seq read : io=59850MB, bw=997.36MB/s, iops=255241, runt= 60028msec
4k rand read : io=399676KB, bw=6127.9KB/s, iops=1531, runt= 65231msec
4k seq write: io=12194MB, bw=207576KB/s, iops=51893, runt= 60155msec
4k rand write: io=111100KB, bw=1315.7KB/s, iops=328, runt= 84445msec

Perfect, do you want to add a tested-by or something?

And how does performance compare to !mq with deadline or similar?

Although I'm getting a lockdep splat:
------------[ cut here ]------------
WARNING: CPU: 4 PID: 3211 at kernel/locking/lockdep.c:3514 lock_release+
DEBUG_LOCKS_WARN_ON(depth <= 0)
Modules linked in: nfsv3 nfs_acl rpcsec_gss_krb5 auth_rpcgss nfsv4 nfs l
uhci_hcd ahci ehci_hcd libahci ttm crc32c_intel serio_raw hpsa usbcore
CPU: 4 PID: 3211 Comm: fio Not tainted 4.10.0-rc3+ #598
Hardware name: HP ProLiant DL380p Gen8, BIOS P70 09/18/2013

That's a known issue in the base that I have, you're supposed to run
it with master pulled in. Annoying, sorry, but it's not my patches.

-- 
Jens Axboe