Re: [PATCH] dm-crypt: limit the number of allocated pages

From: Tom Yan <tom.ty89@gmail.com>
To: Mikulas Patocka <mpatocka@redhat.com>
Cc: Ondrej Kozina <okozina@redhat.com>,
	dm-devel@redhat.com, hch@lst.de,
	Mike Snitzer <msnitzer@redhat.com>, Milan Broz <mbroz@redhat.com>
Subject: Re: [PATCH] dm-crypt: limit the number of allocated pages
Date: Fri, 25 Aug 2017 12:58:59 +0800	[thread overview]
Message-ID: <CAGnHSEnDFc+XJDj9LtHjxJy+vSYS7HNK=F=nUH7tzsWPQozCeA@mail.gmail.com> (raw)
In-Reply-To: <alpine.LRH.2.02.1708191318110.2380@file01.intranet.prod.int.rdu2.redhat.com>

On 20 August 2017 at 01:34, Mikulas Patocka <mpatocka@redhat.com> wrote:
>
> If you issue a single ioctl that takes extreme amount of time - the kernel
> warns about being blocked for extreme amount of time - what else should it
> do?
>

But as I said, it does NOT warn about being blocked if the ioctl is
issued for a SCSI device (i.e. blkdiscard -z /dev/sdX), so why the
warning occurs / cannot be avoided in the case of a dm(-crypt)
container?

One thing I can see is, when blkdiscard -z is running on a dm-crypt
container, iostat list it with a big avgqu-sz (not to be mixed up with
avgrq-sz), starting from ~40000 (the underlying device is a 37G
partition), gradually dropping to 0 (when the job is finished) with a
step of ~40; all other items are 0 for the container (except %util,
also 100% for most of the time). While for the underlying device, no
matter if it is written to through a dm-crypt container, its lines
stays at something like these:

Device:         rrqm/s   wrqm/s     r/s     w/s    rkB/s    wkB/s
avgrq-sz avgqu-sz   await r_await w_await  svctm  %util
sdb               0.00     0.00    0.00   85.00     0.00 43520.00
1024.00   144.71 1664.89    0.00 1664.89  11.76 100.00
sdb               0.00     0.00    0.00   86.00     0.00 44032.00
1024.00   142.34 1694.70    0.00 1694.70  11.63 100.00