[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <e10b611e-f925-f12d-bcd2-ba60d86dd8d0@marvell.com>
Date: Tue, 12 May 2020 19:23:28 +0300
From: Igor Russkikh <irusskikh@...vell.com>
To: Luis Chamberlain <mcgrof@...nel.org>
CC: <jeyu@...nel.org>, <akpm@...ux-foundation.org>, <arnd@...db.de>,
<rostedt@...dmis.org>, <mingo@...hat.com>, <aquini@...hat.com>,
<cai@....pw>, <dyoung@...hat.com>, <bhe@...hat.com>,
<peterz@...radead.org>, <tglx@...utronix.de>,
<gpiccoli@...onical.com>, <pmladek@...e.com>, <tiwai@...e.de>,
<schlad@...e.de>, <andriy.shevchenko@...ux.intel.com>,
<keescook@...omium.org>, <daniel.vetter@...ll.ch>,
<will@...nel.org>, <mchehab+samsung@...nel.org>,
<kvalo@...eaurora.org>, <davem@...emloft.net>,
<netdev@...r.kernel.org>, <linux-kernel@...r.kernel.org>,
Ariel Elior <aelior@...vell.com>,
GR-everest-linux-l2 <GR-everest-linux-l2@...vell.com>
Subject: Re: [EXT] [PATCH 09/15] qed: use new module_firmware_crashed()
>> So I think its not a good place to insert this call.
>> Its hard to find exact good place to insert it in qed.
>
> Is there a way to check if what happened was indeed a fw crash?
Our driver has two firmwares (slowpath and fastpath).
For slowpath firmware the way to understand it crashed is to observe command
response timeout. This is in qed_mcp.c, around "The MFW failed to respond to
command" traceout.
For fastpath this is tricky, think you may leave the above place as the only
place to invoke module_firmware_crashed()
>
>> One more thing is that AFAIU taint flag gets permanent on kernel, but
> for
>> example our device can recover itself from some FW crashes, thus it'd be
>> transparent for user.
>
> Similar things are *supposed* to recoverable with other device, however
> this can also sometimes lead to a situation where devices are not usable
> anymore, and require a full driver unload / load.
>
>> Whats the logical purpose of module_firmware_crashed? Does it mean fatal
>> unrecoverable error on device?
>
> Its just to annotate on the module and kernel that this has happened.
>
> I take it you may agree that, firmware crashing *often* is not good
> design,
> and these issues should be reported to / fixed by vendors. In cases
> where driver bugs are reported it is good to see if a firmware crash has
> happened before, so that during analysis this is ruled out.
Probably, but still I see some misalignment here, in sense that taint is about
the kernel state, not about a hardware state indication.
devlink health could really be a much better candidate for such things.
Regards
Igor
Powered by blists - more mailing lists