Re: [v5 01/12] struct device: Add function callback durable_name

Re: [v5 01/12] struct device: Add function callback durable_name

[Tony Asleson]

On 10/1/20 6:48 AM, Greg Kroah-Hartman wrote:
> On Wed, Sep 30, 2020 at 09:35:52AM -0500, Tony Asleson wrote:
>> On 9/30/20 2:38 AM, Greg Kroah-Hartman wrote:
>>> On Tue, Sep 29, 2020 at 05:04:32PM -0500, Tony Asleson wrote:
>>>> I'm trying to figure out a way to positively identify which storage
>>>> device an error belongs to over time.
>>>
>>> "over time" is not the kernel's responsibility.
>>>
>>> This comes up every 5 years or so. The kernel provides you, at runtime,
>>> a mapping between a hardware device and a "logical" device.  It can
>>> provide information to userspace about this mapping, but once that
>>> device goes away, the kernel is free to reuse that logical device again.
>>>
>>> If you want to track what logical devices match up to what physical
>>> device, then do it in userspace, by parsing the log files.
>>
>> I don't understand why people think it's acceptable to ask user space to
>> parse text that is subject to change.
> 
> What text is changing? The format of of the prefix of dev_*() is well
> known and has been stable for 15+ years now, right?  What is difficult
> in parsing it?

Many of the storage layer messages are using printk, not dev_printk.

>>>> Thank you for supplying some feedback and asking questions.  I've been
>>>> asking for suggestions and would very much like to have a discussion on
>>>> how this issue is best solved.  I'm not attached to what I've provided.
>>>> I'm just trying to get towards a solution.
>>>
>>> Again, solve this in userspace, you have the information there at
>>> runtime, why not use it?
>>
>> We usually don't have the needed information if you remove the
>> expectation that user space should parse the human readable portion of
>> the error message.
> 
> I don't expect that userspace should have to parse any human readable
> portion, if they don't want to.  But if you do want it to, it is pretty
> trivial to parse what you have today:
> 
> 	scsi 2:0:0:0: Direct-Access     Generic  STORAGE DEVICE   1531 PQ: 0 ANSI: 6
> 
> If you really have a unique identifier, then great, parse it today:
> 
> 	usb 4-1.3.1: Product: USB3.0 Card Reader
> 	usb 4-1.3.1: Manufacturer: Generic
> 	usb 4-1.3.1: SerialNumber: 000000001531
> 
> What's keeping that from working now?

I believe these examples are using dev_printk.  With dev_printk we don't
need to parse the text, we can use the meta data.

> In fact, I would argue that it does seem to work, as there are many
> commercial tools out there that seem to handle it just fine...

I'm trying to get something that's works for journalctl.

>>>> We've looked at user space quite a bit and there is an inherit race
>>>> condition with trying to fetch the unique hardware id for a message when
>>>> it gets emitted from the kernel as udev rules haven't even run (assuming
>>>> we even have the meta-data to make the association).
>>>
>>> But one moment later you do have the information, so you can properly
>>> correlate it, right?
>>
>> We could have the information if all the storage paths went through
>> dev_printk.  Here is what we get today when we encounter a read error
>> which uses printk in the block layer:
>>
>> {
>>         "_HOSTNAME" : "pn",
>>         "_TRANSPORT" : "kernel",
>>         "__MONOTONIC_TIMESTAMP" : "1806379233",
>>         "SYSLOG_IDENTIFIER" : "kernel",
>>         "_SOURCE_MONOTONIC_TIMESTAMP" : "1805611354",
>>         "SYSLOG_FACILITY" : "0",
>>         "MESSAGE" : "blk_update_request: critical medium error, dev
>> nvme0n1, sector 10000 op 0x0:(READ) flags 0x80700 phys_seg 3 prio class 0",
>>         "PRIORITY" : "3",
>>         "_MACHINE_ID" : "3f31a0847cea4c95b7a9cec13d07deeb",
>>         "__REALTIME_TIMESTAMP" : "1601471260802301",
>>         "_BOOT_ID" : "b03ed610f21d46ab8243a495ba5a0058",
>>         "__CURSOR" :
>> "s=a063a22bbb384da0b0412e8f652deabb;i=23c2;b=b03ed610f21d46ab8243a495ba5a0058;m=6bab28e1;t=5b087959e3cfd;x=20528862f8f765c9"
>> }
> 
> Ok, messy stuff, don't do that :)
> 
>> Unless you parse the message text you cannot make the association.  If
>> the same message was changed to dev_printk we would get:
>>
>>
>> {
>>         "__REALTIME_TIMESTAMP" : "1589401901093443",
>>         "__CURSOR" :
>> "s=caac9703b34a48fd92f7875adae55a2f;i=1c713;b=e2ae14a9def345aa803a13648b95429c;m=7d25b4f;t=5a58d77b85243;x=b034c2d3fb853870",
>>         "SYSLOG_IDENTIFIER" : "kernel",
>>         "_KERNEL_DEVICE" : "b259:917504",
>>         "__MONOTONIC_TIMESTAMP" : "131226447",
>>         "_UDEV_SYSNAME" : "nvme0n1",
>>         "PRIORITY" : "3",
>>         "_KERNEL_SUBSYSTEM" : "block",
>>         "_SOURCE_MONOTONIC_TIMESTAMP" : "130941917",
>>         "_TRANSPORT" : "kernel",
>>         "_MACHINE_ID" : "3f31a0847cea4c95b7a9cec13d07deeb",
>>         "_HOSTNAME" : "pn",
>>         "SYSLOG_FACILITY" : "0",
>>         "_BOOT_ID" : "e2ae14a9def345aa803a13648b95429c",
>>         "_UDEV_DEVLINK" : [
>>                 "/dev/disk/by-uuid/22fc262a-d621-452a-a951-7761d9fcf0dc",
>>                 "/dev/disk/by-path/pci-0000:00:05.0-nvme-1",
>>
>> "/dev/disk/by-id/nvme-nvme.8086-4445414442454546-51454d55204e564d65204374726c-00000001",
>>                 "/dev/disk/by-id/nvme-QEMU_NVMe_Ctrl_DEADBEEF"
>>         ],
>>         "MESSAGE" : "block nvme0n1: blk_update_request: critical medium
>> error, dev nvme0n1, sector 10000 op 0x0:(READ) flags 0x0 phys_seg 1 prio
>> class 0",
>>         "_UDEV_DEVNODE" : "/dev/nvme0n1"
>> }
> 
> Great, you have a udev sysname, a kernel subsystem and a way to
> associate that with a real device, what more are you wanting?

Did you miss in my example where it's currently a printk?  I showed what
it would look like if it was a dev_printk.

Journald is using _KERNEL_DEVICE to add the _UDEV_DEVLINK information to
the journal entry, it's not parsing the prefix of the message.

The above json is outputted from journalctl when you specify "-o
json-pretty".

>> Journald already knows how to utilize the dev_printk meta data.
> 
> And if you talk to the printk developers (which you seem to be keeping
> out of the loop here), they are ripping out the meta data facility as
> fast as possible.  So don't rely on extending that please.

Again, I'm not trying to keep anyone out of the loop.  Last I knew the
meta data capability wasn't being removed, maybe this has changed?

Ref.

https://lore.kernel.org/lkml/20191007120134.ciywr3wale4gxa6v@pathway.suse.cz/


>> One idea that I've suggested along the way is creating a dev_printk
>> function that doesn't change the message text.  We then avoid breaking
>> people that are parsing.  Is this something that would be acceptable to
>> folks?  It doesn't solve early boot where udev rules haven't even run,
>> but it's better.
> 
> I still fail to understand the root problem here.

IMHO one of the root problems is that many storage messages are still
using printk.  Changing messages to dev_printk has been met with resistance.


> Ok, no, I think I understand what you think the problem is, I just don't
> see why it is up to the kernel to change what we have today when there
> are lots of tools out there working just fine without any kernel changes
> needed.
> 
> So try explaining the problem as you see it please, so we all know where
> to work from.

To me the problem today is the kernel logs information to identify how a
storage device is attached, not what is attached.  I think you agree
with this statement.  The log information is not helpful without the
information to correlate to the actual device.  I think you also agree
with this too as you have mentioned it's user space's responsibility to
collect this so that the correlation can be done.

If the following are *both* true, we have a usable message that has the
correlated data with it in the journal.

1. The storage related kernel message goes through dev_printk
2. At the time of the message being emitted the device symlinks are present.

When those two things are both true, journalctl can do the following
(today):

$ journalctl /dev/disk/by-id/wwn-0x5002538844584d30


However, it usually can't because the above two things are rarely both
true at the same time for a given message for journald when it logs to
the journal.

You keep saying this is a user space issue, but I believe we still need
a bit of help from the kernel at the very least by migrating to
dev_printk or something similar that adds the same meta data without
changing the message text.

Yes, my patch series went one step further and added the device ID as
structured data to the log message, but I was also trying to minimize
the race condition between the kernel emitting a message and journald
not having the information to associate it to the hardware device.

If people have other suggestions please let them be known.

> But again, cutting out the developers of the subsystems you were wanting
> to modify might just be making me really grumpy about this whole
> thing...

Again, I'm sorry I didn't reach out to the correct people.  Hopefully
I've CC'd everyone that is appropriate for this discussion.


Thanks,
Tony

Re: [v5 01/12] struct device: Add function callback durable_name

[Greg Kroah-Hartman]

On Wed, Oct 07, 2020 at 03:10:17PM -0500, Tony Asleson wrote:
> On 10/1/20 6:48 AM, Greg Kroah-Hartman wrote:
> > On Wed, Sep 30, 2020 at 09:35:52AM -0500, Tony Asleson wrote:
> >> On 9/30/20 2:38 AM, Greg Kroah-Hartman wrote:
> >>> On Tue, Sep 29, 2020 at 05:04:32PM -0500, Tony Asleson wrote:
> >>>> I'm trying to figure out a way to positively identify which storage
> >>>> device an error belongs to over time.
> >>>
> >>> "over time" is not the kernel's responsibility.
> >>>
> >>> This comes up every 5 years or so. The kernel provides you, at runtime,
> >>> a mapping between a hardware device and a "logical" device.  It can
> >>> provide information to userspace about this mapping, but once that
> >>> device goes away, the kernel is free to reuse that logical device again.
> >>>
> >>> If you want to track what logical devices match up to what physical
> >>> device, then do it in userspace, by parsing the log files.
> >>
> >> I don't understand why people think it's acceptable to ask user space to
> >> parse text that is subject to change.
> > 
> > What text is changing? The format of of the prefix of dev_*() is well
> > known and has been stable for 15+ years now, right?  What is difficult
> > in parsing it?
> 
> Many of the storage layer messages are using printk, not dev_printk.

Ok, then stop right there.  Fix that up.  Don't try to route around the
standard way of displaying log messages by creating a totally different
way of doing things.

Just use the dev_*() calls, and all will be fine.  Kernel log messages
are not "ABI" in that they have to be preserved in any specific way, so
adding a prefix to them as dev_*() does, will be fine.

thanks,

greg k-h

Re: [v5 01/12] struct device: Add function callback durable_name

[Hannes Reinecke]

On 10/7/20 10:10 PM, Tony Asleson wrote:
> On 10/1/20 6:48 AM, Greg Kroah-Hartman wrote:
>> On Wed, Sep 30, 2020 at 09:35:52AM -0500, Tony Asleson wrote:
>>> On 9/30/20 2:38 AM, Greg Kroah-Hartman wrote:
>>>> On Tue, Sep 29, 2020 at 05:04:32PM -0500, Tony Asleson wrote:
>>>>> I'm trying to figure out a way to positively identify which storage
>>>>> device an error belongs to over time.
>>>>
>>>> "over time" is not the kernel's responsibility.
>>>>
>>>> This comes up every 5 years or so. The kernel provides you, at runtime,
>>>> a mapping between a hardware device and a "logical" device.  It can
>>>> provide information to userspace about this mapping, but once that
>>>> device goes away, the kernel is free to reuse that logical device again.
>>>>
>>>> If you want to track what logical devices match up to what physical
>>>> device, then do it in userspace, by parsing the log files.
>>>
>>> I don't understand why people think it's acceptable to ask user space to
>>> parse text that is subject to change.
>>
>> What text is changing? The format of of the prefix of dev_*() is well
>> known and has been stable for 15+ years now, right?  What is difficult
>> in parsing it?
> 
> Many of the storage layer messages are using printk, not dev_printk.
> 
So that would be the immediate angle of attack ...

>>>>> Thank you for supplying some feedback and asking questions.  I've been
>>>>> asking for suggestions and would very much like to have a discussion on
>>>>> how this issue is best solved.  I'm not attached to what I've provided.
>>>>> I'm just trying to get towards a solution.
>>>>
>>>> Again, solve this in userspace, you have the information there at
>>>> runtime, why not use it?
>>>
>>> We usually don't have the needed information if you remove the
>>> expectation that user space should parse the human readable portion of
>>> the error message.
>>
>> I don't expect that userspace should have to parse any human readable
>> portion, if they don't want to.  But if you do want it to, it is pretty
>> trivial to parse what you have today:
>>
>> 	scsi 2:0:0:0: Direct-Access     Generic  STORAGE DEVICE   1531 PQ: 0 ANSI: 6
>>
>> If you really have a unique identifier, then great, parse it today:
>>
>> 	usb 4-1.3.1: Product: USB3.0 Card Reader
>> 	usb 4-1.3.1: Manufacturer: Generic
>> 	usb 4-1.3.1: SerialNumber: 000000001531
>>
>> What's keeping that from working now?
> 
> I believe these examples are using dev_printk.  With dev_printk we don't
> need to parse the text, we can use the meta data.
> So it looks as most of your usecase would be solved by moving to 
dev_printk().
Why not work on that instead?
I do presume this will have immediate benefits for everybody, and will 
have approval from everyone.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke                Kernel Storage Architect
hare@suse.de                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer

Re: [v5 01/12] struct device: Add function callback durable_name

[Finn Thain]

On Wed, 7 Oct 2020, Tony Asleson wrote:

> The log information is not helpful without the information to correlate 
> to the actual device.

Log messages that associate one entity with another can be generated 
whenever such an association comes into existence, which is probably when 
devices get probed.

E.g. a host:channel:target:lun identifier gets associated with a block 
device name by the dev_printk() calls in sd_probe():

[    3.600000] sd 0:0:0:0: [sda] Attached SCSI disk

BTW, if you think of {"0:0:0:0","sda"} as a row in some normalized table 
and squint a bit, this problem is not unlike the replication of database 
tables over a message queue.

Re: [v5 01/12] struct device: Add function callback durable_name

[Martin K. Petersen]

Greg,

>> > What text is changing? The format of of the prefix of dev_*() is well
>> > known and has been stable for 15+ years now, right?  What is difficult
>> > in parsing it?
>> 
>> Many of the storage layer messages are using printk, not dev_printk.
>
> Ok, then stop right there.  Fix that up.  Don't try to route around the
> standard way of displaying log messages by creating a totally different
> way of doing things.

Couldn't agree more!

-- 
Martin K. Petersen	Oracle Linux Engineering