Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Christoph Hellwig]

On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> physical sector size of loop devices. On ppc64, this causes loop devices
> to have 64k as the physical sector size.

Eek.  We'll need to revert the loop change ASAP!

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Omar Sandoval]

On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> > Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> > physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> > physical sector size of loop devices. On ppc64, this causes loop devices
> > to have 64k as the physical sector size.
> 
> Eek.  We'll need to revert the loop change ASAP!

Most annoying patch series ever. It hasn't made it to Linus' tree yet,
right? We can revert (although the later change depends on that), fold
in a fix, or apply a fix on top of it, whatever Jens prefers.

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Chandan Rajendra]

On Tuesday, September 5, 2017 12:12:08 PM IST Omar Sandoval wrote:
> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> > On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> > > Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> > > physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> > > physical sector size of loop devices. On ppc64, this causes loop devices
> > > to have 64k as the physical sector size.
> > 
> > Eek.  We'll need to revert the loop change ASAP!
> 
> Most annoying patch series ever. It hasn't made it to Linus' tree yet,
> right? We can revert (although the later change depends on that), fold
> in a fix, or apply a fix on top of it, whatever Jens prefers.
> 
> 

My bad. 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 is the commit id from 
Linux-next. I don't see this commit in Linus's git tree.

-- 
chandan

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Jens Axboe]

On 09/05/2017 01:37 AM, Chandan Rajendra wrote:
> On Tuesday, September 5, 2017 12:12:08 PM IST Omar Sandoval wrote:
>> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
>>> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
>>>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
>>>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
>>>> physical sector size of loop devices. On ppc64, this causes loop devices
>>>> to have 64k as the physical sector size.
>>>
>>> Eek.  We'll need to revert the loop change ASAP!
>>
>> Most annoying patch series ever. It hasn't made it to Linus' tree yet,
>> right? We can revert (although the later change depends on that), fold
>> in a fix, or apply a fix on top of it, whatever Jens prefers.
>>
>>
> 
> My bad. 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 is the commit id from 
> Linux-next. I don't see this commit in Linus's git tree.

Right, it's only queued up, and scheduled for the 2nd part of the
block changes for 4.14. It should have been PAGE_CACHE_SIZE, but
we don't have that anymore...

Omar, are you sending a patch to fix this up?

-- 
Jens Axboe

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Dave Chinner]

On Tue, Sep 05, 2017 at 09:00:26AM -0600, Jens Axboe wrote:
> On 09/05/2017 01:37 AM, Chandan Rajendra wrote:
> > On Tuesday, September 5, 2017 12:12:08 PM IST Omar Sandoval wrote:
> >> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> >>> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> >>>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> >>>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> >>>> physical sector size of loop devices. On ppc64, this causes loop devices
> >>>> to have 64k as the physical sector size.
> >>>
> >>> Eek.  We'll need to revert the loop change ASAP!
> >>
> >> Most annoying patch series ever. It hasn't made it to Linus' tree yet,
> >> right? We can revert (although the later change depends on that), fold
> >> in a fix, or apply a fix on top of it, whatever Jens prefers.
> >>
> >>
> > 
> > My bad. 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 is the commit id from 
> > Linux-next. I don't see this commit in Linus's git tree.
> 
> Right, it's only queued up, and scheduled for the 2nd part of the
> block changes for 4.14. It should have been PAGE_CACHE_SIZE, but
> we don't have that anymore...

But PAGE_CACHE_SIZE was equal to PAGE_SIZE, so that would have been
wrong, too.

I just don't see why this is necessary, given that buffered IO
through the upper filesystem will already be doing page
sized/aligned IO where possible (because that's what the page cache
does!). And for direct IO the loop device should just export the
underlying host filesystem logical/physical sector sizes, which
should be optimal for the backing storage to begin with.

Someone want to enlighten me as to what problem is being solved
here?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Omar Sandoval]

On Wed, Sep 06, 2017 at 08:06:31AM +1000, Dave Chinner wrote:
> On Tue, Sep 05, 2017 at 09:00:26AM -0600, Jens Axboe wrote:
> > On 09/05/2017 01:37 AM, Chandan Rajendra wrote:
> > > On Tuesday, September 5, 2017 12:12:08 PM IST Omar Sandoval wrote:
> > >> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> > >>> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> > >>>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> > >>>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> > >>>> physical sector size of loop devices. On ppc64, this causes loop devices
> > >>>> to have 64k as the physical sector size.
> > >>>
> > >>> Eek.  We'll need to revert the loop change ASAP!
> > >>
> > >> Most annoying patch series ever. It hasn't made it to Linus' tree yet,
> > >> right? We can revert (although the later change depends on that), fold
> > >> in a fix, or apply a fix on top of it, whatever Jens prefers.
> > >>
> > >>
> > > 
> > > My bad. 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 is the commit id from 
> > > Linux-next. I don't see this commit in Linus's git tree.
> > 
> > Right, it's only queued up, and scheduled for the 2nd part of the
> > block changes for 4.14. It should have been PAGE_CACHE_SIZE, but
> > we don't have that anymore...
> 
> But PAGE_CACHE_SIZE was equal to PAGE_SIZE, so that would have been
> wrong, too.
> 
> I just don't see why this is necessary, given that buffered IO
> through the upper filesystem will already be doing page
> sized/aligned IO where possible (because that's what the page cache
> does!). And for direct IO the loop device should just export the
> underlying host filesystem logical/physical sector sizes, which
> should be optimal for the backing storage to begin with.
> 
> Someone want to enlighten me as to what problem is being solved
> here?

I already sent a patch to fix this:
https://marc.info/?l=linux-block&m=150464670510301&w=2

Loop was using the default physical block size of 512, which is a lie
according to the definition of the physical block size:

/**
 * blk_queue_physical_block_size - set physical block size for the queue
 * @q:  the request queue for the device
 * @size:  the physical block size, in bytes
 *
 * Description:
 *   This should be set to the lowest possible sector size that the
 *   hardware can operate on without reverting to read-modify-write
 *   operations.
 */

Clearly for buffered loop devices, this unit is a page. The change was
pedantic, so whatever, my patch above makes things backwards compatible.

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Dave Chinner]

On Tue, Sep 05, 2017 at 03:18:51PM -0700, Omar Sandoval wrote:
> On Wed, Sep 06, 2017 at 08:06:31AM +1000, Dave Chinner wrote:
> > On Tue, Sep 05, 2017 at 09:00:26AM -0600, Jens Axboe wrote:
> > > On 09/05/2017 01:37 AM, Chandan Rajendra wrote:
> > > > On Tuesday, September 5, 2017 12:12:08 PM IST Omar Sandoval wrote:
> > > >> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> > > >>> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> > > >>>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> > > >>>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> > > >>>> physical sector size of loop devices. On ppc64, this causes loop devices
> > > >>>> to have 64k as the physical sector size.
> > > >>>
> > > >>> Eek.  We'll need to revert the loop change ASAP!
> > > >>
> > > >> Most annoying patch series ever. It hasn't made it to Linus' tree yet,
> > > >> right? We can revert (although the later change depends on that), fold
> > > >> in a fix, or apply a fix on top of it, whatever Jens prefers.
> > > >>
> > > >>
> > > > 
> > > > My bad. 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 is the commit id from 
> > > > Linux-next. I don't see this commit in Linus's git tree.
> > > 
> > > Right, it's only queued up, and scheduled for the 2nd part of the
> > > block changes for 4.14. It should have been PAGE_CACHE_SIZE, but
> > > we don't have that anymore...
> > 
> > But PAGE_CACHE_SIZE was equal to PAGE_SIZE, so that would have been
> > wrong, too.
> > 
> > I just don't see why this is necessary, given that buffered IO
> > through the upper filesystem will already be doing page
> > sized/aligned IO where possible (because that's what the page cache
> > does!). And for direct IO the loop device should just export the
> > underlying host filesystem logical/physical sector sizes, which
> > should be optimal for the backing storage to begin with.
> > 
> > Someone want to enlighten me as to what problem is being solved
> > here?
> 
> I already sent a patch to fix this:
> https://marc.info/?l=linux-block&m=150464670510301&w=2
> 
> Loop was using the default physical block size of 512, which is a lie
> according to the definition of the physical block size:

Loop devices are special. They /have to lie/ because they are the
only mechanism we have for mounting image files with sector sizes
smaller than the host storage sizes.  i.e. loop devices require
compatibility and flexibility first and so need to default to 
"most compatible" behaviour, not "most performant".

If you're really "sick of these stupid changes" then perhaps you
should consider cc'ing linux-fsdevel for loop device changes so
filesystem developers have a chance to catch these filesystem side
problems in new loopdev code before it's merged anywhere....

FWIW, if you're going to set some kind of optimal default for the
loop devices then - like all other stacked block devices - the
defaults need to be pulled from the underlying storage (i.e. the
filesystem that hosts the file) rather than making them up out of
thin air.

Cheers,

Dave.

-- 
Dave Chinner
david@fromorbit.com

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Omar Sandoval]

On Wed, Sep 06, 2017 at 09:24:49AM +1000, Dave Chinner wrote:
> On Tue, Sep 05, 2017 at 03:18:51PM -0700, Omar Sandoval wrote:
> > On Wed, Sep 06, 2017 at 08:06:31AM +1000, Dave Chinner wrote:
> > > On Tue, Sep 05, 2017 at 09:00:26AM -0600, Jens Axboe wrote:
> > > > On 09/05/2017 01:37 AM, Chandan Rajendra wrote:
> > > > > On Tuesday, September 5, 2017 12:12:08 PM IST Omar Sandoval wrote:
> > > > >> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> > > > >>> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> > > > >>>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> > > > >>>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> > > > >>>> physical sector size of loop devices. On ppc64, this causes loop devices
> > > > >>>> to have 64k as the physical sector size.
> > > > >>>
> > > > >>> Eek.  We'll need to revert the loop change ASAP!
> > > > >>
> > > > >> Most annoying patch series ever. It hasn't made it to Linus' tree yet,
> > > > >> right? We can revert (although the later change depends on that), fold
> > > > >> in a fix, or apply a fix on top of it, whatever Jens prefers.
> > > > >>
> > > > >>
> > > > > 
> > > > > My bad. 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 is the commit id from 
> > > > > Linux-next. I don't see this commit in Linus's git tree.
> > > > 
> > > > Right, it's only queued up, and scheduled for the 2nd part of the
> > > > block changes for 4.14. It should have been PAGE_CACHE_SIZE, but
> > > > we don't have that anymore...
> > > 
> > > But PAGE_CACHE_SIZE was equal to PAGE_SIZE, so that would have been
> > > wrong, too.
> > > 
> > > I just don't see why this is necessary, given that buffered IO
> > > through the upper filesystem will already be doing page
> > > sized/aligned IO where possible (because that's what the page cache
> > > does!). And for direct IO the loop device should just export the
> > > underlying host filesystem logical/physical sector sizes, which
> > > should be optimal for the backing storage to begin with.
> > > 
> > > Someone want to enlighten me as to what problem is being solved
> > > here?
> > 
> > I already sent a patch to fix this:
> > https://marc.info/?l=linux-block&m=150464670510301&w=2
> > 
> > Loop was using the default physical block size of 512, which is a lie
> > according to the definition of the physical block size:
> 
> Loop devices are special. They /have to lie/ because they are the
> only mechanism we have for mounting image files with sector sizes
> smaller than the host storage sizes.  i.e. loop devices require
> compatibility and flexibility first and so need to default to 
> "most compatible" behaviour, not "most performant".

And that's why the logical block size is 512 by default. The physical
block size doesn't have any bearing on what you can do with the device,
it's a _hint_ for the application. But again, I'm just being overly
pedantic, I should've checked how tools were actually using the physical
block size.

> If you're really "sick of these stupid changes" then perhaps you
> should consider cc'ing linux-fsdevel for loop device changes so
> filesystem developers have a chance to catch these filesystem side
> problems in new loopdev code before it's merged anywhere....

Yup, loop changes haven't been going outside of linux-block but at least
this one should have gone to fsdevel, sorry.

> FWIW, if you're going to set some kind of optimal default for the
> loop devices then - like all other stacked block devices - the
> defaults need to be pulled from the underlying storage (i.e. the
> filesystem that hosts the file) rather than making them up out of
> thin air.

Not worth the trouble for now, I'll just leave this alone now.

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Dave Chinner]

On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> > Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> > physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> > physical sector size of loop devices. On ppc64, this causes loop devices
> > to have 64k as the physical sector size.
> 
> Eek.  We'll need to revert the loop change ASAP!

And, FWIW, making the warning go away if probably a bad idea,
because XFS only supports devices with sector sizes up
to 32k:

#define XFS_MIN_SECTORSIZE_LOG  9       /* i.e. 512 bytes */
#define XFS_MAX_SECTORSIZE_LOG  15      /* i.e. 32768 bytes */

And so it should be warning about devices trying to tell it to use
something larger....

Cheers,

Dave.

-- 
Dave Chinner
david@fromorbit.com

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Eric Sandeen]

On 9/5/17 1:44 AM, Dave Chinner wrote:
> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
>> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
>>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
>>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
>>> physical sector size of loop devices. On ppc64, this causes loop devices
>>> to have 64k as the physical sector size.
>>
>> Eek.  We'll need to revert the loop change ASAP!
> 
> And, FWIW, making the warning go away if probably a bad idea,
> because XFS only supports devices with sector sizes up
> to 32k:

Well, TBH removing this warning was my suggestion, because it's
automatically fixing values that weren't specified by the user in
the first place.  First preference is physical sector size, then
fallback to logical but it doesn't need to be noisy about it.

> #define XFS_MIN_SECTORSIZE_LOG  9       /* i.e. 512 bytes */
> #define XFS_MAX_SECTORSIZE_LOG  15      /* i.e. 32768 bytes */
> 
> And so it should be warning about devices trying to tell it to use
> something larger....

As long as the logical sector size is small enough, it seems like a
silent adjustment is probably ok,  no?

-Eric

> Cheers,
> 
> Dave.
> 

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Dave Chinner]

On Tue, Sep 05, 2017 at 09:17:42AM -0500, Eric Sandeen wrote:
> 
> 
> On 9/5/17 1:44 AM, Dave Chinner wrote:
> > On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
> >> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
> >>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
> >>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
> >>> physical sector size of loop devices. On ppc64, this causes loop devices
> >>> to have 64k as the physical sector size.
> >>
> >> Eek.  We'll need to revert the loop change ASAP!
> > 
> > And, FWIW, making the warning go away if probably a bad idea,
> > because XFS only supports devices with sector sizes up
> > to 32k:
> 
> Well, TBH removing this warning was my suggestion, because it's
> automatically fixing values that weren't specified by the user in
> the first place.  First preference is physical sector size, then
> fallback to logical but it doesn't need to be noisy about it.
> 
> > #define XFS_MIN_SECTORSIZE_LOG  9       /* i.e. 512 bytes */
> > #define XFS_MAX_SECTORSIZE_LOG  15      /* i.e. 32768 bytes */
> > 
> > And so it should be warning about devices trying to tell it to use
> > something larger....
> 
> As long as the logical sector size is small enough, it seems like a
> silent adjustment is probably ok,  no?

Think 512e drives. Doing 512 byte sector IO is possible, but slow.
Someone might actually want to avoid that by having the filesystem
use 4k sector sizes. However, if for some reason, mkfs selects 512
byte sectors (the logical size) rather than 4k sector size, then
shouldn't we be telling the user we're doing something that has a
"for-the-life-of-the-filesystem" performance impact?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: [PATCH] mkfs: Remove messages printed when blocksize < physical sectorsize

[Eric Sandeen]

On 9/5/17 5:10 PM, Dave Chinner wrote:
> On Tue, Sep 05, 2017 at 09:17:42AM -0500, Eric Sandeen wrote:
>>
>>
>> On 9/5/17 1:44 AM, Dave Chinner wrote:
>>> On Mon, Sep 04, 2017 at 11:31:39PM -0700, Christoph Hellwig wrote:
>>>> On Tue, Sep 05, 2017 at 11:14:42AM +0530, Chandan Rajendra wrote:
>>>>> Linux kernel commit 6c6b6f28b3335fd85ec833ee0005d9c9dca6c003 (loop: set
>>>>> physical block size to PAGE_SIZE) now sets PAGE_SIZE as the default
>>>>> physical sector size of loop devices. On ppc64, this causes loop devices
>>>>> to have 64k as the physical sector size.
>>>>
>>>> Eek.  We'll need to revert the loop change ASAP!
>>>
>>> And, FWIW, making the warning go away if probably a bad idea,
>>> because XFS only supports devices with sector sizes up
>>> to 32k:
>>
>> Well, TBH removing this warning was my suggestion, because it's
>> automatically fixing values that weren't specified by the user in
>> the first place.  First preference is physical sector size, then
>> fallback to logical but it doesn't need to be noisy about it.
>>
>>> #define XFS_MIN_SECTORSIZE_LOG  9       /* i.e. 512 bytes */
>>> #define XFS_MAX_SECTORSIZE_LOG  15      /* i.e. 32768 bytes */
>>>
>>> And so it should be warning about devices trying to tell it to use
>>> something larger....
>>
>> As long as the logical sector size is small enough, it seems like a
>> silent adjustment is probably ok,  no?
> 
> Think 512e drives. Doing 512 byte sector IO is possible, but slow.
> Someone might actually want to avoid that by having the filesystem
> use 4k sector sizes. However, if for some reason, mkfs selects 512
> byte sectors (the logical size) rather than 4k sector size, then
> shouldn't we be telling the user we're doing something that has a
> "for-the-life-of-the-filesystem" performance impact?

Well, sure, but it'll only select 512 if it /has/ to, i.e. if the
block size is < 4k.

So for the simple case of a 512e drive, our default block size is
4k, physical size is 4k, and everything is happy and fine.

If the user /specifies/ a 1k block size on such a device, how much
of a nanny do we really want to be about telling them this is suboptimal?

There are a lot of suboptimal things you can specify on the
mkfs commandline, but we don't generally choose to warn about them...

-Eric

> Cheers,
> 
> Dave.
>