Filesystems like ext4 can submit writes in multiples of blocksizes.
But we still can't allow the writes to be split. Hence let's check if
the iomap_length() is same as iter->len or not.

This shouldn't affect XFS since it anyways checks for this in
xfs_file_write_iter() to not support atomic write size request of more
than FS blocksize.

Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
---
 fs/iomap/direct-io.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index ed4764e3b8f0..1d33b4239b3e 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c
@@ -306,7 +306,7 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
 	size_t copied = 0;
 	size_t orig_count;
 
-	if (atomic && length != fs_block_size)
+	if (atomic && length != iter->len)
 		return -EINVAL;
 
 	if ((pos | length) & (bdev_logical_block_size(iomap->bdev) - 1) ||
-- 
2.46.0

On 25/10/2024 04:45, Ritesh Harjani (IBM) wrote:
> Filesystems like ext4 can submit writes in multiples of blocksizes.
> But we still can't allow the writes to be split. Hence let's check if
> the iomap_length() is same as iter->len or not.
> 
> This shouldn't affect XFS since it anyways checks for this in
> xfs_file_write_iter() to not support atomic write size request of more
> than FS blocksize.
> 
> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
> ---
>   fs/iomap/direct-io.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
> index ed4764e3b8f0..1d33b4239b3e 100644
> --- a/fs/iomap/direct-io.c
> +++ b/fs/iomap/direct-io.c
> @@ -306,7 +306,7 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
>   	size_t copied = 0;
>   	size_t orig_count;
>   
> -	if (atomic && length != fs_block_size)
> +	if (atomic && length != iter->len)
>   		return -EINVAL;

Here you expect just one iter for an atomic write always.

In 6/6, you are saying that iomap does not allow an atomic write which 
covers unwritten and written extents, right?

But for writing a single fs block atomically, we don't mandate it to be 
in unwritten state. So there is a difference in behavior in writing a 
single FS block vs multiple FS blocks atomically.

So we have 3x choices, as I see:
a. add a check now in iomap that the extent is in written state (for an 
atomic write)
b. add extent zeroing code, as I was trying for originally
c. document this peculiarity

Thanks,
John

Hi John, 

John Garry <john.g.garry@oracle.com> writes:

> On 25/10/2024 04:45, Ritesh Harjani (IBM) wrote:
>> Filesystems like ext4 can submit writes in multiples of blocksizes.
>> But we still can't allow the writes to be split. Hence let's check if
>> the iomap_length() is same as iter->len or not.
>> 
>> This shouldn't affect XFS since it anyways checks for this in
>> xfs_file_write_iter() to not support atomic write size request of more
>> than FS blocksize.
>> 
>> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
>> ---
>>   fs/iomap/direct-io.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>> 
>> diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
>> index ed4764e3b8f0..1d33b4239b3e 100644
>> --- a/fs/iomap/direct-io.c
>> +++ b/fs/iomap/direct-io.c
>> @@ -306,7 +306,7 @@ static loff_t iomap_dio_bio_iter(const struct iomap_iter *iter,
>>   	size_t copied = 0;
>>   	size_t orig_count;
>>   
>> -	if (atomic && length != fs_block_size)
>> +	if (atomic && length != iter->len)
>>   		return -EINVAL;
>
> Here you expect just one iter for an atomic write always.

Here we are lifting the limitation of iomap to support entire iter->len
rather than just 1 fsblock. 

>
> In 6/6, you are saying that iomap does not allow an atomic write which 
> covers unwritten and written extents, right?

No, it's not that. If FS does not provide a full mapping to iomap in
->iomap_begin then the writes will get split. For atomic writes this
should not happen, hence the check in iomap above to return -EINVAL if
mapped length does not match iter->len.

>
> But for writing a single fs block atomically, we don't mandate it to be 
> in unwritten state. So there is a difference in behavior in writing a 
> single FS block vs multiple FS blocks atomically.

Same as mentioned above. We can't have atomic writes to get split.
This patch is just lifting the restriction of iomap to allow more than
blocksize but the mapped length should still meet iter->len, as
otherwise the writes can get split.

>
> So we have 3x choices, as I see:
> a. add a check now in iomap that the extent is in written state (for an 
> atomic write)
> b. add extent zeroing code, as I was trying for originally
> c. document this peculiarity
>
> Thanks,
> John

On 25/10/2024 10:31, Ritesh Harjani (IBM) wrote:
>>>    
>>> -	if (atomic && length != fs_block_size)
>>> +	if (atomic && length != iter->len)
>>>    		return -EINVAL;
>> Here you expect just one iter for an atomic write always.
> Here we are lifting the limitation of iomap to support entire iter->len
> rather than just 1 fsblock.

Sure

> 
>> In 6/6, you are saying that iomap does not allow an atomic write which
>> covers unwritten and written extents, right?
> No, it's not that. If FS does not provide a full mapping to iomap in
> ->iomap_begin then the writes will get split. 

but why would it provide multiple mapping?

> For atomic writes this
> should not happen, hence the check in iomap above to return -EINVAL if
> mapped length does not match iter->len.
> 
>> But for writing a single fs block atomically, we don't mandate it to be
>> in unwritten state. So there is a difference in behavior in writing a
>> single FS block vs multiple FS blocks atomically.
> Same as mentioned above. We can't have atomic writes to get split.
> This patch is just lifting the restriction of iomap to allow more than
> blocksize but the mapped length should still meet iter->len, as
> otherwise the writes can get split.

Sure, I get this. But I wonder why would we be getting multiple 
mappings? Why cannot the FS always provide a single mapping?

Thanks,
John

John Garry <john.g.garry@oracle.com> writes:

> On 25/10/2024 10:31, Ritesh Harjani (IBM) wrote:
>>>>    
>>>> -	if (atomic && length != fs_block_size)
>>>> +	if (atomic && length != iter->len)
>>>>    		return -EINVAL;
>>> Here you expect just one iter for an atomic write always.
>> Here we are lifting the limitation of iomap to support entire iter->len
>> rather than just 1 fsblock.
>
> Sure
>
>> 
>>> In 6/6, you are saying that iomap does not allow an atomic write which
>>> covers unwritten and written extents, right?
>> No, it's not that. If FS does not provide a full mapping to iomap in
>> ->iomap_begin then the writes will get split. 
>
> but why would it provide multiple mapping?
>
>> For atomic writes this
>> should not happen, hence the check in iomap above to return -EINVAL if
>> mapped length does not match iter->len.
>> 
>>> But for writing a single fs block atomically, we don't mandate it to be
>>> in unwritten state. So there is a difference in behavior in writing a
>>> single FS block vs multiple FS blocks atomically.
>> Same as mentioned above. We can't have atomic writes to get split.
>> This patch is just lifting the restriction of iomap to allow more than
>> blocksize but the mapped length should still meet iter->len, as
>> otherwise the writes can get split.
>
> Sure, I get this. But I wonder why would we be getting multiple 
> mappings? Why cannot the FS always provide a single mapping?

FS can decide to split the mappings when it couldn't allocate a single
large mapping of the requested length. Could be due to - 
- already allocated extent followed by EOF, 
- already allocated extent followed by a hole
- already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)
- delalloc (not delalloc since we invalidate respective page cache pages before doing DIO).
- fragmentation or ENOSPC - For ext4 bigalloc this will not happen since
we reserve the entire cluster. So we know there should be space. But I
am not sure how other filesystems might end up implementing this functionality.

Thanks!

-ritesh

On 25/10/2024 11:35, Ritesh Harjani (IBM) wrote:
>>> Same as mentioned above. We can't have atomic writes to get split.
>>> This patch is just lifting the restriction of iomap to allow more than
>>> blocksize but the mapped length should still meet iter->len, as
>>> otherwise the writes can get split.
>> Sure, I get this. But I wonder why would we be getting multiple
>> mappings? Why cannot the FS always provide a single mapping?
> FS can decide to split the mappings when it couldn't allocate a single
> large mapping of the requested length. Could be due to -
> - already allocated extent followed by EOF,
> - already allocated extent followed by a hole
> - already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)

This is the sort of scenario which I am concerned with. This issue has 
been discussed at length for XFS forcealign support for atomic writes.

So far, the user can atomic write a single FS block regardless of 
whether the extent in which it would be part of is in written or 
unwritten state.

Now the rule will be to write multiple FS blocks atomically, all blocks 
need to be in same written or unwritten state.

This oddity at least needs to be documented.

Better yet would be to not have this restriction.

> - delalloc (not delalloc since we invalidate respective page cache pages before doing DIO).
> - fragmentation or ENOSPC - For ext4 bigalloc this will not happen since
> we reserve the entire cluster. So we know there should be space. But I
> am not sure how other filesystems might end up implementing this functionality.

Thanks,
John

John Garry <john.g.garry@oracle.com> writes:

> On 25/10/2024 11:35, Ritesh Harjani (IBM) wrote:
>>>> Same as mentioned above. We can't have atomic writes to get split.
>>>> This patch is just lifting the restriction of iomap to allow more than
>>>> blocksize but the mapped length should still meet iter->len, as
>>>> otherwise the writes can get split.
>>> Sure, I get this. But I wonder why would we be getting multiple
>>> mappings? Why cannot the FS always provide a single mapping?
>> FS can decide to split the mappings when it couldn't allocate a single
>> large mapping of the requested length. Could be due to -
>> - already allocated extent followed by EOF,
>> - already allocated extent followed by a hole
>> - already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)
>
> This is the sort of scenario which I am concerned with. This issue has 
> been discussed at length for XFS forcealign support for atomic writes.

extsize and forcealign is being worked for ext4 as well where we can
add such support, sure.

>
> So far, the user can atomic write a single FS block regardless of 
> whether the extent in which it would be part of is in written or 
> unwritten state.
>
> Now the rule will be to write multiple FS blocks atomically, all blocks 
> need to be in same written or unwritten state.

FS needs to ensure that the writes does not get torned. So for whatever reason
FS splits the mapping then we need to return an -EINVAL error to not
allow such writes to get torned. This patch just does that.

But I get your point. More below.

>
> This oddity at least needs to be documented.

Got it. Yes, we can do that.

>
> Better yet would be to not have this restriction.
>

I haven't thought of a clever way where we don't have to zero out the
rest of the unwritten mapping. With ext4 bigalloc since the entire
cluster is anyway reserved - I was thinking if we can come up with a
clever way for doing atomic writes to the entire user requested size w/o
zeroing out.

Zeroing out the other unwritten extent is also a cost penalty to the
user anyways. So user will anyway will have to be made aware of not to
attempt writes of fashion which can cause them such penalties. 

As patch-6 mentions this is a base support for bs = ps systems for
enabling atomic writes using bigalloc. For now we return -EINVAL when we
can't allocate a continuous user requested mapping which means it won't
support operations of types 8k followed by 16k.

We can document this behavior as other things are documented for atomic
writes on ext4 with bigalloc e.g. pow_of_2 length writes, natural
alignment w.r.t pos and length etc.

Does that sound ok?

>> - delalloc (not delalloc since we invalidate respective page cache pages before doing DIO).
>> - fragmentation or ENOSPC - For ext4 bigalloc this will not happen since
>> we reserve the entire cluster. So we know there should be space. But I
>> am not sure how other filesystems might end up implementing this functionality.
>
> Thanks,
> John

-ritesh

On 25/10/2024 12:19, Ritesh Harjani (IBM) wrote:
> John Garry <john.g.garry@oracle.com> writes:
> 
>> On 25/10/2024 11:35, Ritesh Harjani (IBM) wrote:
>>>>> Same as mentioned above. We can't have atomic writes to get split.
>>>>> This patch is just lifting the restriction of iomap to allow more than
>>>>> blocksize but the mapped length should still meet iter->len, as
>>>>> otherwise the writes can get split.
>>>> Sure, I get this. But I wonder why would we be getting multiple
>>>> mappings? Why cannot the FS always provide a single mapping?
>>> FS can decide to split the mappings when it couldn't allocate a single
>>> large mapping of the requested length. Could be due to -
>>> - already allocated extent followed by EOF,
>>> - already allocated extent followed by a hole
>>> - already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)
>>
>> This is the sort of scenario which I am concerned with. This issue has
>> been discussed at length for XFS forcealign support for atomic writes.
> 
> extsize and forcealign is being worked for ext4 as well where we can
> add such support, sure.
> 
>>
>> So far, the user can atomic write a single FS block regardless of
>> whether the extent in which it would be part of is in written or
>> unwritten state.
>>
>> Now the rule will be to write multiple FS blocks atomically, all blocks
>> need to be in same written or unwritten state.
> 
> FS needs to ensure that the writes does not get torned. So for whatever reason
> FS splits the mapping then we need to return an -EINVAL error to not
> allow such writes to get torned. This patch just does that.
> 
> But I get your point. More below.
> 
>>
>> This oddity at least needs to be documented.
> 
> Got it. Yes, we can do that.
> 
>>
>> Better yet would be to not have this restriction.
>>
> 
> I haven't thought of a clever way where we don't have to zero out the
> rest of the unwritten mapping. With ext4 bigalloc since the entire
> cluster is anyway reserved - I was thinking if we can come up with a
> clever way for doing atomic writes to the entire user requested size w/o
> zeroing out.

This following was main method which was being attempted:

https://lore.kernel.org/linux-fsdevel/20240429174746.2132161-15-john.g.garry@oracle.com/

There were other ideas in different versions of the forcelign/xfs block 
atomic writes series.

> 
> Zeroing out the other unwritten extent is also a cost penalty to the
> user anyways.

Sure, unless we have a special inode flag to say "pre-zero the extent".

> So user will anyway will have to be made aware of not to
> attempt writes of fashion which can cause them such penalties.
> 
> As patch-6 mentions this is a base support for bs = ps systems for
> enabling atomic writes using bigalloc. For now we return -EINVAL when we
> can't allocate a continuous user requested mapping which means it won't
> support operations of types 8k followed by 16k.
> 

That's my least-preferred option.

I think better would be reject atomic writes that cover unwritten 
extents always - but that boat is about to sail...

Thanks,
John

John Garry <john.g.garry@oracle.com> writes:

> On 25/10/2024 12:19, Ritesh Harjani (IBM) wrote:
>> John Garry <john.g.garry@oracle.com> writes:
>> 
>>> On 25/10/2024 11:35, Ritesh Harjani (IBM) wrote:
>>>>>> Same as mentioned above. We can't have atomic writes to get split.
>>>>>> This patch is just lifting the restriction of iomap to allow more than
>>>>>> blocksize but the mapped length should still meet iter->len, as
>>>>>> otherwise the writes can get split.
>>>>> Sure, I get this. But I wonder why would we be getting multiple
>>>>> mappings? Why cannot the FS always provide a single mapping?
>>>> FS can decide to split the mappings when it couldn't allocate a single
>>>> large mapping of the requested length. Could be due to -
>>>> - already allocated extent followed by EOF,
>>>> - already allocated extent followed by a hole
>>>> - already mapped extent followed by an extent of different type (e.g. written followed by unwritten or unwritten followed by written)
>>>
>>> This is the sort of scenario which I am concerned with. This issue has
>>> been discussed at length for XFS forcealign support for atomic writes.
>> 
>> extsize and forcealign is being worked for ext4 as well where we can
>> add such support, sure.
>> 
>>>
>>> So far, the user can atomic write a single FS block regardless of
>>> whether the extent in which it would be part of is in written or
>>> unwritten state.
>>>
>>> Now the rule will be to write multiple FS blocks atomically, all blocks
>>> need to be in same written or unwritten state.
>> 
>> FS needs to ensure that the writes does not get torned. So for whatever reason
>> FS splits the mapping then we need to return an -EINVAL error to not
>> allow such writes to get torned. This patch just does that.
>> 
>> But I get your point. More below.
>> 
>>>
>>> This oddity at least needs to be documented.
>> 
>> Got it. Yes, we can do that.
>> 
>>>
>>> Better yet would be to not have this restriction.
>>>
>> 
>> I haven't thought of a clever way where we don't have to zero out the
>> rest of the unwritten mapping. With ext4 bigalloc since the entire
>> cluster is anyway reserved - I was thinking if we can come up with a
>> clever way for doing atomic writes to the entire user requested size w/o
>> zeroing out.
>
> This following was main method which was being attempted:
>
> https://lore.kernel.org/linux-fsdevel/20240429174746.2132161-15-john.g.garry@oracle.com/
>
> There were other ideas in different versions of the forcelign/xfs block 
> atomic writes series.
>
>> 
>> Zeroing out the other unwritten extent is also a cost penalty to the
>> user anyways.
>
> Sure, unless we have a special inode flag to say "pre-zero the extent".
>
>> So user will anyway will have to be made aware of not to
>> attempt writes of fashion which can cause them such penalties.
>> 
>> As patch-6 mentions this is a base support for bs = ps systems for
>> enabling atomic writes using bigalloc. For now we return -EINVAL when we
>> can't allocate a continuous user requested mapping which means it won't
>> support operations of types 8k followed by 16k.
>> 
>
> That's my least-preferred option.
>
> I think better would be reject atomic writes that cover unwritten 
> extents always - but that boat is about to sail...

That's what this patch does. For whatever reason if we couldn't allocate
a single contiguous region of requested size for atomic write, then we
reject the request always, isn't it. Or maybe I didn't understand your comment.

If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks()) 
for atomic writes in ext4_dio_write_checks(), similar to how we detect
overwrites case to decide whether we need a read v/s write semaphore. 
So this can check if the user has a partially allocated extent for the
user requested region and if yes, we can return -EINVAL from
ext4_dio_write_iter() itself. 

I think this maybe better option than waiting until ->iomap_begin().
This might also bring all atomic write constraints to be checked in one
place i.e. during ext4_file_write_iter() itself.

Thoughts?

-ritesh

On 25/10/2024 13:36, Ritesh Harjani (IBM) wrote:
>>> So user will anyway will have to be made aware of not to
>>> attempt writes of fashion which can cause them such penalties.
>>>
>>> As patch-6 mentions this is a base support for bs = ps systems for
>>> enabling atomic writes using bigalloc. For now we return -EINVAL when we
>>> can't allocate a continuous user requested mapping which means it won't
>>> support operations of types 8k followed by 16k.
>>>
>> That's my least-preferred option.
>>
>> I think better would be reject atomic writes that cover unwritten
>> extents always - but that boat is about to sail...
> That's what this patch does.

Not really.

Currently we have 2x iomap restrictions:
a. mapping length must equal fs block size
b. bio created must equal total write size

This patch just says that the mapping length must equal total write size 
(instead of a.). So quite similar to b.

> For whatever reason if we couldn't allocate
> a single contiguous region of requested size for atomic write, then we
> reject the request always, isn't it. Or maybe I didn't understand your comment.

As the simplest example, for an atomic write to an empty file, there 
should only be a single mapping returned to iomap_dio_bio_iter() and 
that would be of IOMAP_UNWRITTEN type. And we don't reject that.

> 
> If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks())
> for atomic writes in ext4_dio_write_checks(), similar to how we detect
> overwrites case to decide whether we need a read v/s write semaphore.
> So this can check if the user has a partially allocated extent for the
> user requested region and if yes, we can return -EINVAL from
> ext4_dio_write_iter() itself.
 > > I think this maybe better option than waiting until ->iomap_begin().
> This might also bring all atomic write constraints to be checked in one
> place i.e. during ext4_file_write_iter() itself.

Something like this can be done once we decide how atomic writing to 
regions which cover mixed unwritten and written extents is to be handled.

Thanks,
John

John Garry <john.g.garry@oracle.com> writes:

> On 25/10/2024 13:36, Ritesh Harjani (IBM) wrote:
>>>> So user will anyway will have to be made aware of not to
>>>> attempt writes of fashion which can cause them such penalties.
>>>>
>>>> As patch-6 mentions this is a base support for bs = ps systems for
>>>> enabling atomic writes using bigalloc. For now we return -EINVAL when we
>>>> can't allocate a continuous user requested mapping which means it won't
>>>> support operations of types 8k followed by 16k.
>>>>
>>> That's my least-preferred option.
>>>
>>> I think better would be reject atomic writes that cover unwritten
>>> extents always - but that boat is about to sail...
>> That's what this patch does.
>
> Not really.
>
> Currently we have 2x iomap restrictions:
> a. mapping length must equal fs block size
> b. bio created must equal total write size
>
> This patch just says that the mapping length must equal total write size 
> (instead of a.). So quite similar to b.
>
>> For whatever reason if we couldn't allocate
>> a single contiguous region of requested size for atomic write, then we
>> reject the request always, isn't it. Or maybe I didn't understand your comment.
>
> As the simplest example, for an atomic write to an empty file, there 
> should only be a single mapping returned to iomap_dio_bio_iter() and 
> that would be of IOMAP_UNWRITTEN type. And we don't reject that.
>

Ok. Maybe this is what I am missing. Could you please help me understand
why should such writes be rejected? 

For e.g. 
If FS could allocate a single contiguous IOMAP_UNWRITTEN extent of
atomic write request size, that means - 
1. FS will allocate an unwritten extent.
2. will do writes (using submit_bio) to the unwritten extent. 
3. will do unwritten to written conversion. 

It is ok if either of the above operations fail right? If (3) fails
then the region will still be marked unwritten that means it will read
zero (old contents). (2) can anyway fail and will not result into
partial writes. (1) will anyway not result into any write whatsoever.

So we can never have a situation where there is partial writes leading
to mix of old and new write contents right for such cases? Which is what the
requirement of atomic/untorn write also is?

Sorry am I missing something here?

>> 
>> If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks())
>> for atomic writes in ext4_dio_write_checks(), similar to how we detect
>> overwrites case to decide whether we need a read v/s write semaphore.
>> So this can check if the user has a partially allocated extent for the
>> user requested region and if yes, we can return -EINVAL from
>> ext4_dio_write_iter() itself.
>  > > I think this maybe better option than waiting until ->iomap_begin().
>> This might also bring all atomic write constraints to be checked in one
>> place i.e. during ext4_file_write_iter() itself.
>
> Something like this can be done once we decide how atomic writing to 
> regions which cover mixed unwritten and written extents is to be handled.

Mixed extent regions (written + unwritten) is a different case all
together (which can lead to mix of old and new contents).


But here what I am suggesting is to add following constraint in case of
ext4 with bigalloc - 

"Writes to a region which already has partially allocated extent is not supported."

That means we will return -EINVAL if we detect above case in
ext4_file_write_iter() and sure we can document this behavior.

In retrospect, I am not sure why we cannot add a constraint for atomic
writes (e.g. for ext4 bigalloc) and reject such writes outright,
instead of silently incurring a performance penalty by zeroing out the
partial regions by allowing such write request.

-ritesh

On Fri, Oct 25, 2024 at 07:43:17PM +0530, Ritesh Harjani wrote:
> John Garry <john.g.garry@oracle.com> writes:
> 
> > On 25/10/2024 13:36, Ritesh Harjani (IBM) wrote:
> >>>> So user will anyway will have to be made aware of not to
> >>>> attempt writes of fashion which can cause them such penalties.
> >>>>
> >>>> As patch-6 mentions this is a base support for bs = ps systems for
> >>>> enabling atomic writes using bigalloc. For now we return -EINVAL when we
> >>>> can't allocate a continuous user requested mapping which means it won't
> >>>> support operations of types 8k followed by 16k.
> >>>>
> >>> That's my least-preferred option.
> >>>
> >>> I think better would be reject atomic writes that cover unwritten
> >>> extents always - but that boat is about to sail...
> >> That's what this patch does.
> >
> > Not really.
> >
> > Currently we have 2x iomap restrictions:
> > a. mapping length must equal fs block size
> > b. bio created must equal total write size
> >
> > This patch just says that the mapping length must equal total write size 
> > (instead of a.). So quite similar to b.
> >
> >> For whatever reason if we couldn't allocate
> >> a single contiguous region of requested size for atomic write, then we
> >> reject the request always, isn't it. Or maybe I didn't understand your comment.
> >
> > As the simplest example, for an atomic write to an empty file, there 
> > should only be a single mapping returned to iomap_dio_bio_iter() and 
> > that would be of IOMAP_UNWRITTEN type. And we don't reject that.
> >
> 
> Ok. Maybe this is what I am missing. Could you please help me understand
> why should such writes be rejected? 
> 
> For e.g. 
> If FS could allocate a single contiguous IOMAP_UNWRITTEN extent of
> atomic write request size, that means - 
> 1. FS will allocate an unwritten extent.
> 2. will do writes (using submit_bio) to the unwritten extent. 
> 3. will do unwritten to written conversion. 
> 
> It is ok if either of the above operations fail right? If (3) fails
> then the region will still be marked unwritten that means it will read
> zero (old contents). (2) can anyway fail and will not result into
> partial writes. (1) will anyway not result into any write whatsoever.
> 
> So we can never have a situation where there is partial writes leading
> to mix of old and new write contents right for such cases? Which is what the
> requirement of atomic/untorn write also is?
> 
> Sorry am I missing something here?

I must be missing something; to perform an untorn write, two things must
happen --

1. The kernel writes the data to the storage device, and the storage
device either persists all of it, or throws back an error having
persisted none of it.

2. If (1) completes successfully, all file mapping updates for the range
written must be persisted, or an error is thrown back and none of them
are persisted.

iomap doesn't have to know how the filesystem satisfies (2); it just has
to create a single bio containing all data pages or it rejects the
write.

Currently, it's an implementation detail that the XFS directio write
ioend code processes the file mapping updates for the range written by
walking every extent mapping for that range and issuing separate
transactions for each mapping update.  There's nothing that can restart
the walk if it is interrupted.  That's why XFS cannot support multi
fsblock untorn writes to blocks with different status.

As I've said before, the most general solution to this would be to add a
new log intent item that would track the "update all mappings in this
file range" operation so that recovery could restart the walk.  This is
the most technically challenging, so we decided not to implement it
until there is demand.

Having set aside the idea of redesigning ioend, the second-most general
solution is pre-zeroing unwritten extents and holes so that
->iomap_begin implementations can present a single mapping to the bio
constructor.  Technically if there's only one unwritten extent or hole
or cow, xfs can actually satisfy (2) because it only creates one
transaction.

This gets me to the third and much less general solution -- only allow
untorn writes if we know that the ioend only ever has to run a single
transaction.  That's why untorn writes are limited to a single fsblock
for now -- it's a simple solution so that we can get our downstream
customers to kick the tires and start on the next iteration instead of
spending years on waterfalling.

Did you notice that in all of these cases, the capabilities of the
filesystem's ioend processing determines the restrictions on the number
and type of mappings that ->iomap_begin can give to iomap?

Now that we have a second system trying to hook up to the iomap support,
it's clear to me that the restrictions on mappings are specific to each
filesystem.  Therefore, the iomap directio code should not impose
restrictions on the mappings it receives unless they would prevent the
creation of the single aligned bio.

Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return
EINVAL or something if they look at the file mappings and discover that
they cannot perform the ioend without risking torn mapping updates.  In
the long run, ->iomap_begin is where this iomap->len <= iter->len check
really belongs, but hold that thought.

For the multi fsblock case, the ->iomap_begin functions would have to
check that only one metadata update would be necessary in the ioend.
That's where things get murky, since ext4/xfs drop their mapping locks
between calls to ->iomap_begin.  So you'd have to check all the mappings
for unsupported mixed state every time.  Yuck.

It might be less gross to retain the restriction that iomap accepts only
one mapping for the entire file range, like Ritesh has here.  Users
might be ok with us saying that you can't do a 16k atomic write to a
region where you previously did an 8k write until you write the other
8k, even if someone has to write zeroes.  Users might be ok with the
kernel allowing multi-fsblock writes but only if the stars align.  But
to learn the answers to those questions, we have to put /something/ in
the hands of our users.

For now (because we're already at -rc5), let's have xfs/ext4's
->write_iter implementations restrict atomic writes to a single fsblock,
and get both merged into the kernel.  Let's defer the multi fsblock work
to 6.14, though I think we could take this patch.

Does that sound cool?

--D

> >> 
> >> If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks())
> >> for atomic writes in ext4_dio_write_checks(), similar to how we detect
> >> overwrites case to decide whether we need a read v/s write semaphore.
> >> So this can check if the user has a partially allocated extent for the
> >> user requested region and if yes, we can return -EINVAL from
> >> ext4_dio_write_iter() itself.
> >  > > I think this maybe better option than waiting until ->iomap_begin().
> >> This might also bring all atomic write constraints to be checked in one
> >> place i.e. during ext4_file_write_iter() itself.
> >
> > Something like this can be done once we decide how atomic writing to 
> > regions which cover mixed unwritten and written extents is to be handled.
> 
> Mixed extent regions (written + unwritten) is a different case all
> together (which can lead to mix of old and new contents).
> 
> 
> But here what I am suggesting is to add following constraint in case of
> ext4 with bigalloc - 
> 
> "Writes to a region which already has partially allocated extent is not supported."
> 
> That means we will return -EINVAL if we detect above case in
> ext4_file_write_iter() and sure we can document this behavior.
> 
> In retrospect, I am not sure why we cannot add a constraint for atomic
> writes (e.g. for ext4 bigalloc) and reject such writes outright,
> instead of silently incurring a performance penalty by zeroing out the
> partial regions by allowing such write request.
> 
> -ritesh
>

"Darrick J. Wong" <djwong@kernel.org> writes:

> On Fri, Oct 25, 2024 at 07:43:17PM +0530, Ritesh Harjani wrote:
>> John Garry <john.g.garry@oracle.com> writes:
>> 
>> > On 25/10/2024 13:36, Ritesh Harjani (IBM) wrote:
>> >>>> So user will anyway will have to be made aware of not to
>> >>>> attempt writes of fashion which can cause them such penalties.
>> >>>>
>> >>>> As patch-6 mentions this is a base support for bs = ps systems for
>> >>>> enabling atomic writes using bigalloc. For now we return -EINVAL when we
>> >>>> can't allocate a continuous user requested mapping which means it won't
>> >>>> support operations of types 8k followed by 16k.
>> >>>>
>> >>> That's my least-preferred option.
>> >>>
>> >>> I think better would be reject atomic writes that cover unwritten
>> >>> extents always - but that boat is about to sail...
>> >> That's what this patch does.
>> >
>> > Not really.
>> >
>> > Currently we have 2x iomap restrictions:
>> > a. mapping length must equal fs block size
>> > b. bio created must equal total write size
>> >
>> > This patch just says that the mapping length must equal total write size 
>> > (instead of a.). So quite similar to b.
>> >
>> >> For whatever reason if we couldn't allocate
>> >> a single contiguous region of requested size for atomic write, then we
>> >> reject the request always, isn't it. Or maybe I didn't understand your comment.
>> >
>> > As the simplest example, for an atomic write to an empty file, there 
>> > should only be a single mapping returned to iomap_dio_bio_iter() and 
>> > that would be of IOMAP_UNWRITTEN type. And we don't reject that.
>> >
>> 
>> Ok. Maybe this is what I am missing. Could you please help me understand
>> why should such writes be rejected? 
>> 
>> For e.g. 
>> If FS could allocate a single contiguous IOMAP_UNWRITTEN extent of
>> atomic write request size, that means - 
>> 1. FS will allocate an unwritten extent.
>> 2. will do writes (using submit_bio) to the unwritten extent. 
>> 3. will do unwritten to written conversion. 
>> 
>> It is ok if either of the above operations fail right? If (3) fails
>> then the region will still be marked unwritten that means it will read
>> zero (old contents). (2) can anyway fail and will not result into
>> partial writes. (1) will anyway not result into any write whatsoever.
>> 
>> So we can never have a situation where there is partial writes leading
>> to mix of old and new write contents right for such cases? Which is what the
>> requirement of atomic/untorn write also is?
>> 
>> Sorry am I missing something here?
>
> I must be missing something; to perform an untorn write, two things must
> happen --
>
> 1. The kernel writes the data to the storage device, and the storage
> device either persists all of it, or throws back an error having
> persisted none of it.
>
> 2. If (1) completes successfully, all file mapping updates for the range
> written must be persisted, or an error is thrown back and none of them
> are persisted.
>
> iomap doesn't have to know how the filesystem satisfies (2); it just has
> to create a single bio containing all data pages or it rejects the
> write.
>
> Currently, it's an implementation detail that the XFS directio write
> ioend code processes the file mapping updates for the range written by
> walking every extent mapping for that range and issuing separate
> transactions for each mapping update.  There's nothing that can restart
> the walk if it is interrupted.  That's why XFS cannot support multi
> fsblock untorn writes to blocks with different status.
>
> As I've said before, the most general solution to this would be to add a
> new log intent item that would track the "update all mappings in this
> file range" operation so that recovery could restart the walk.  This is
> the most technically challenging, so we decided not to implement it
> until there is demand.
>
> Having set aside the idea of redesigning ioend, the second-most general
> solution is pre-zeroing unwritten extents and holes so that
> ->iomap_begin implementations can present a single mapping to the bio
> constructor.  Technically if there's only one unwritten extent or hole
> or cow, xfs can actually satisfy (2) because it only creates one
> transaction.
>
> This gets me to the third and much less general solution -- only allow
> untorn writes if we know that the ioend only ever has to run a single
> transaction.  That's why untorn writes are limited to a single fsblock
> for now -- it's a simple solution so that we can get our downstream
> customers to kick the tires and start on the next iteration instead of
> spending years on waterfalling.
>
> Did you notice that in all of these cases, the capabilities of the
> filesystem's ioend processing determines the restrictions on the number
> and type of mappings that ->iomap_begin can give to iomap?
>
> Now that we have a second system trying to hook up to the iomap support,
> it's clear to me that the restrictions on mappings are specific to each
> filesystem.  Therefore, the iomap directio code should not impose
> restrictions on the mappings it receives unless they would prevent the
> creation of the single aligned bio.
>
> Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return
> EINVAL or something if they look at the file mappings and discover that
> they cannot perform the ioend without risking torn mapping updates.  In
> the long run, ->iomap_begin is where this iomap->len <= iter->len check
> really belongs, but hold that thought.
>
> For the multi fsblock case, the ->iomap_begin functions would have to
> check that only one metadata update would be necessary in the ioend.
> That's where things get murky, since ext4/xfs drop their mapping locks
> between calls to ->iomap_begin.  So you'd have to check all the mappings
> for unsupported mixed state every time.  Yuck.
>

Thanks Darrick for taking time summarizing what all has been done
and your thoughts here.

> It might be less gross to retain the restriction that iomap accepts only
> one mapping for the entire file range, like Ritesh has here.

less gross :) sure. 

I would like to think of this as, being less restrictive (compared to
only allowing a single fsblock) by adding a constraint on the atomic
write I/O request i.e.  

"Atomic write I/O request to a region in a file is only allowed if that
region has no partially allocated extents. Otherwise, the file system
can fail the I/O operation by returning -EINVAL."

Essentially by adding this constraint to the I/O request, we are
helping the user to prevent atomic writes from accidentally getting
torned and also allowing multi-fsblock writes. So I still think that
might be the right thing to do here or at least a better start. FS can
later work on adding such support where we don't even need above
such constraint on a given atomic write I/O request.

> Users
> might be ok with us saying that you can't do a 16k atomic write to a
> region where you previously did an 8k write until you write the other
> 8k, even if someone has to write zeroes.  Users might be ok with the
> kernel allowing multi-fsblock writes but only if the stars align.

> But
> to learn the answers to those questions, we have to put /something/ in
> the hands of our users.

On this point, I think ext4 might already has those users who might be
using atomic write characteristics of devices to do untorn writes. e.g. 

In [1], Ted has talked about using bigalloc with ext4 for torn write
prevention. [2] talks about using ext4 with bigalloc to prevent torn
writes on aws cloud.

[1]: https://www.youtube.com/watch?v=gIeuiGg-_iw
[2]: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/configure-twp.html

My point being - Looks like the class of users who are using untorn
writes to improve their database performances are already doing so even
w/o any such interfaces being exposed to them (with ext4 bigalloc).

The current feature support of allowing atomic writes to only single
fsblock might not be helpful to these users who can provide that
feedback, who are using ext4 on bs = ps systems with bigalloc. But maybe
let's wait and hear from them whether it is ok if -   

"Atomic write I/O request to a region in a file is only allowed if that
region has no partially allocated extents. Otherwise, the file system
can fail the I/O operation by returning -EINVAL."

>
> For now (because we're already at -rc5), let's have xfs/ext4's
> ->write_iter implementations restrict atomic writes to a single fsblock,
> and get both merged into the kernel.

Yes, I agree with the approach. I agree that we should get a consensus
on this from folks.

Let me split this series up and address the review comments on patch
[1-4]. Patch-5 & 6 can be worked once we have conclusion on this and can
be eyed for 6.14.

> Let's defer the multi fsblock work
> to 6.14, though I think we could take this patch.

It's ok to consider this patch along with multi-fsblock work then i.e.
for 6.14.

>
> Does that sound cool?
>
> --D

Thanks Darrick :)

-ritesh

>> >> 
>> >> If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks())
>> >> for atomic writes in ext4_dio_write_checks(), similar to how we detect
>> >> overwrites case to decide whether we need a read v/s write semaphore.
>> >> So this can check if the user has a partially allocated extent for the
>> >> user requested region and if yes, we can return -EINVAL from
>> >> ext4_dio_write_iter() itself.
>> >  > > I think this maybe better option than waiting until ->iomap_begin().
>> >> This might also bring all atomic write constraints to be checked in one
>> >> place i.e. during ext4_file_write_iter() itself.
>> >
>> > Something like this can be done once we decide how atomic writing to 
>> > regions which cover mixed unwritten and written extents is to be handled.
>> 
>> Mixed extent regions (written + unwritten) is a different case all
>> together (which can lead to mix of old and new contents).
>> 
>> 
>> But here what I am suggesting is to add following constraint in case of
>> ext4 with bigalloc - 
>> 
>> "Writes to a region which already has partially allocated extent is not supported."
>> 
>> That means we will return -EINVAL if we detect above case in
>> ext4_file_write_iter() and sure we can document this behavior.
>> 
>> In retrospect, I am not sure why we cannot add a constraint for atomic
>> writes (e.g. for ext4 bigalloc) and reject such writes outright,
>> instead of silently incurring a performance penalty by zeroing out the
>> partial regions by allowing such write request.
>> 
>> -ritesh
>>

On Sat, Oct 26, 2024 at 10:05:44AM +0530, Ritesh Harjani wrote:
> "Darrick J. Wong" <djwong@kernel.org> writes:
> 
> > On Fri, Oct 25, 2024 at 07:43:17PM +0530, Ritesh Harjani wrote:
> >> John Garry <john.g.garry@oracle.com> writes:
> >> 
> >> > On 25/10/2024 13:36, Ritesh Harjani (IBM) wrote:
> >> >>>> So user will anyway will have to be made aware of not to
> >> >>>> attempt writes of fashion which can cause them such penalties.
> >> >>>>
> >> >>>> As patch-6 mentions this is a base support for bs = ps systems for
> >> >>>> enabling atomic writes using bigalloc. For now we return -EINVAL when we
> >> >>>> can't allocate a continuous user requested mapping which means it won't
> >> >>>> support operations of types 8k followed by 16k.
> >> >>>>
> >> >>> That's my least-preferred option.
> >> >>>
> >> >>> I think better would be reject atomic writes that cover unwritten
> >> >>> extents always - but that boat is about to sail...
> >> >> That's what this patch does.
> >> >
> >> > Not really.
> >> >
> >> > Currently we have 2x iomap restrictions:
> >> > a. mapping length must equal fs block size
> >> > b. bio created must equal total write size
> >> >
> >> > This patch just says that the mapping length must equal total write size 
> >> > (instead of a.). So quite similar to b.
> >> >
> >> >> For whatever reason if we couldn't allocate
> >> >> a single contiguous region of requested size for atomic write, then we
> >> >> reject the request always, isn't it. Or maybe I didn't understand your comment.
> >> >
> >> > As the simplest example, for an atomic write to an empty file, there 
> >> > should only be a single mapping returned to iomap_dio_bio_iter() and 
> >> > that would be of IOMAP_UNWRITTEN type. And we don't reject that.
> >> >
> >> 
> >> Ok. Maybe this is what I am missing. Could you please help me understand
> >> why should such writes be rejected? 
> >> 
> >> For e.g. 
> >> If FS could allocate a single contiguous IOMAP_UNWRITTEN extent of
> >> atomic write request size, that means - 
> >> 1. FS will allocate an unwritten extent.
> >> 2. will do writes (using submit_bio) to the unwritten extent. 
> >> 3. will do unwritten to written conversion. 
> >> 
> >> It is ok if either of the above operations fail right? If (3) fails
> >> then the region will still be marked unwritten that means it will read
> >> zero (old contents). (2) can anyway fail and will not result into
> >> partial writes. (1) will anyway not result into any write whatsoever.
> >> 
> >> So we can never have a situation where there is partial writes leading
> >> to mix of old and new write contents right for such cases? Which is what the
> >> requirement of atomic/untorn write also is?
> >> 
> >> Sorry am I missing something here?
> >
> > I must be missing something; to perform an untorn write, two things must
> > happen --
> >
> > 1. The kernel writes the data to the storage device, and the storage
> > device either persists all of it, or throws back an error having
> > persisted none of it.
> >
> > 2. If (1) completes successfully, all file mapping updates for the range
> > written must be persisted, or an error is thrown back and none of them
> > are persisted.
> >
> > iomap doesn't have to know how the filesystem satisfies (2); it just has
> > to create a single bio containing all data pages or it rejects the
> > write.
> >
> > Currently, it's an implementation detail that the XFS directio write
> > ioend code processes the file mapping updates for the range written by
> > walking every extent mapping for that range and issuing separate
> > transactions for each mapping update.  There's nothing that can restart
> > the walk if it is interrupted.  That's why XFS cannot support multi
> > fsblock untorn writes to blocks with different status.
> >
> > As I've said before, the most general solution to this would be to add a
> > new log intent item that would track the "update all mappings in this
> > file range" operation so that recovery could restart the walk.  This is
> > the most technically challenging, so we decided not to implement it
> > until there is demand.
> >
> > Having set aside the idea of redesigning ioend, the second-most general
> > solution is pre-zeroing unwritten extents and holes so that
> > ->iomap_begin implementations can present a single mapping to the bio
> > constructor.  Technically if there's only one unwritten extent or hole
> > or cow, xfs can actually satisfy (2) because it only creates one
> > transaction.
> >
> > This gets me to the third and much less general solution -- only allow
> > untorn writes if we know that the ioend only ever has to run a single
> > transaction.  That's why untorn writes are limited to a single fsblock
> > for now -- it's a simple solution so that we can get our downstream
> > customers to kick the tires and start on the next iteration instead of
> > spending years on waterfalling.
> >
> > Did you notice that in all of these cases, the capabilities of the
> > filesystem's ioend processing determines the restrictions on the number
> > and type of mappings that ->iomap_begin can give to iomap?
> >
> > Now that we have a second system trying to hook up to the iomap support,
> > it's clear to me that the restrictions on mappings are specific to each
> > filesystem.  Therefore, the iomap directio code should not impose
> > restrictions on the mappings it receives unless they would prevent the
> > creation of the single aligned bio.
> >
> > Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return
> > EINVAL or something if they look at the file mappings and discover that
> > they cannot perform the ioend without risking torn mapping updates.  In
> > the long run, ->iomap_begin is where this iomap->len <= iter->len check
> > really belongs, but hold that thought.
> >
> > For the multi fsblock case, the ->iomap_begin functions would have to
> > check that only one metadata update would be necessary in the ioend.
> > That's where things get murky, since ext4/xfs drop their mapping locks
> > between calls to ->iomap_begin.  So you'd have to check all the mappings
> > for unsupported mixed state every time.  Yuck.
> >
> 
> Thanks Darrick for taking time summarizing what all has been done
> and your thoughts here.
> 
> > It might be less gross to retain the restriction that iomap accepts only
> > one mapping for the entire file range, like Ritesh has here.
> 
> less gross :) sure. 
> 
> I would like to think of this as, being less restrictive (compared to
> only allowing a single fsblock) by adding a constraint on the atomic
> write I/O request i.e.  
> 
> "Atomic write I/O request to a region in a file is only allowed if that
> region has no partially allocated extents. Otherwise, the file system
> can fail the I/O operation by returning -EINVAL."
> 
> Essentially by adding this constraint to the I/O request, we are
> helping the user to prevent atomic writes from accidentally getting
> torned and also allowing multi-fsblock writes. So I still think that
> might be the right thing to do here or at least a better start. FS can
> later work on adding such support where we don't even need above
> such constraint on a given atomic write I/O request.

On today's ext4 call, Ted and Ritesh and I realized that there's a bit
more to it than this -- it's not possible to support untorn writes to a
mix of written/(cow,unwritten) mappings even if they all point to the
same physical space.  If the system fails after the storage device
commits the write but before any of the ioend processing is scheduled, a
subsequent read of the previously written blocks will produce the new
data, but reads to the other areas will produce the old contents (or
zeroes, or whatever).  That's a torn write.

Therefore, iomap ought to stick to requiring that ->iomap_begin returns
a single iomap to cover the entire file range for the untorn write.  For
an unwritten extent, the post-recovery read will see either zeroes or
the new contents; for a single-mapping COW it'll see old or new contents
but not both.

(Obviously this still requires that the fs can perform the mapping
updates without tearing too.)

--D

> > Users
> > might be ok with us saying that you can't do a 16k atomic write to a
> > region where you previously did an 8k write until you write the other
> > 8k, even if someone has to write zeroes.  Users might be ok with the
> > kernel allowing multi-fsblock writes but only if the stars align.
> 
> > But
> > to learn the answers to those questions, we have to put /something/ in
> > the hands of our users.
> 
> On this point, I think ext4 might already has those users who might be
> using atomic write characteristics of devices to do untorn writes. e.g. 
> 
> In [1], Ted has talked about using bigalloc with ext4 for torn write
> prevention. [2] talks about using ext4 with bigalloc to prevent torn
> writes on aws cloud.
> 
> [1]: https://www.youtube.com/watch?v=gIeuiGg-_iw
> [2]: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/configure-twp.html
> 
> My point being - Looks like the class of users who are using untorn
> writes to improve their database performances are already doing so even
> w/o any such interfaces being exposed to them (with ext4 bigalloc).
> 
> The current feature support of allowing atomic writes to only single
> fsblock might not be helpful to these users who can provide that
> feedback, who are using ext4 on bs = ps systems with bigalloc. But maybe
> let's wait and hear from them whether it is ok if -   
> 
> "Atomic write I/O request to a region in a file is only allowed if that
> region has no partially allocated extents. Otherwise, the file system
> can fail the I/O operation by returning -EINVAL."
> 
> >
> > For now (because we're already at -rc5), let's have xfs/ext4's
> > ->write_iter implementations restrict atomic writes to a single fsblock,
> > and get both merged into the kernel.
> 
> Yes, I agree with the approach. I agree that we should get a consensus
> on this from folks.
> 
> Let me split this series up and address the review comments on patch
> [1-4]. Patch-5 & 6 can be worked once we have conclusion on this and can
> be eyed for 6.14.
> 
> > Let's defer the multi fsblock work
> > to 6.14, though I think we could take this patch.
> 
> It's ok to consider this patch along with multi-fsblock work then i.e.
> for 6.14.
> 
> >
> > Does that sound cool?
> >
> > --D
> 
> Thanks Darrick :)
> 
> -ritesh
> 
> >> >> 
> >> >> If others prefer - we can maybe add such a check (e.g. ext4_dio_atomic_write_checks())
> >> >> for atomic writes in ext4_dio_write_checks(), similar to how we detect
> >> >> overwrites case to decide whether we need a read v/s write semaphore.
> >> >> So this can check if the user has a partially allocated extent for the
> >> >> user requested region and if yes, we can return -EINVAL from
> >> >> ext4_dio_write_iter() itself.
> >> >  > > I think this maybe better option than waiting until ->iomap_begin().
> >> >> This might also bring all atomic write constraints to be checked in one
> >> >> place i.e. during ext4_file_write_iter() itself.
> >> >
> >> > Something like this can be done once we decide how atomic writing to 
> >> > regions which cover mixed unwritten and written extents is to be handled.
> >> 
> >> Mixed extent regions (written + unwritten) is a different case all
> >> together (which can lead to mix of old and new contents).
> >> 
> >> 
> >> But here what I am suggesting is to add following constraint in case of
> >> ext4 with bigalloc - 
> >> 
> >> "Writes to a region which already has partially allocated extent is not supported."
> >> 
> >> That means we will return -EINVAL if we detect above case in
> >> ext4_file_write_iter() and sure we can document this behavior.
> >> 
> >> In retrospect, I am not sure why we cannot add a constraint for atomic
> >> writes (e.g. for ext4 bigalloc) and reject such writes outright,
> >> instead of silently incurring a performance penalty by zeroing out the
> >> partial regions by allowing such write request.
> >> 
> >> -ritesh
> >> 
>

On Thu, Oct 31, 2024 at 02:36:40PM -0700, Darrick J. Wong wrote:
> On Sat, Oct 26, 2024 at 10:05:44AM +0530, Ritesh Harjani wrote:
> > > This gets me to the third and much less general solution -- only allow
> > > untorn writes if we know that the ioend only ever has to run a single
> > > transaction.  That's why untorn writes are limited to a single fsblock
> > > for now -- it's a simple solution so that we can get our downstream
> > > customers to kick the tires and start on the next iteration instead of
> > > spending years on waterfalling.
> > >
> > > Did you notice that in all of these cases, the capabilities of the
> > > filesystem's ioend processing determines the restrictions on the number
> > > and type of mappings that ->iomap_begin can give to iomap?
> > >
> > > Now that we have a second system trying to hook up to the iomap support,
> > > it's clear to me that the restrictions on mappings are specific to each
> > > filesystem.  Therefore, the iomap directio code should not impose
> > > restrictions on the mappings it receives unless they would prevent the
> > > creation of the single aligned bio.
> > >
> > > Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return
> > > EINVAL or something if they look at the file mappings and discover that
> > > they cannot perform the ioend without risking torn mapping updates.  In
> > > the long run, ->iomap_begin is where this iomap->len <= iter->len check
> > > really belongs, but hold that thought.
> > >
> > > For the multi fsblock case, the ->iomap_begin functions would have to
> > > check that only one metadata update would be necessary in the ioend.
> > > That's where things get murky, since ext4/xfs drop their mapping locks
> > > between calls to ->iomap_begin.  So you'd have to check all the mappings
> > > for unsupported mixed state every time.  Yuck.
> > >
> > 
> > Thanks Darrick for taking time summarizing what all has been done
> > and your thoughts here.
> > 
> > > It might be less gross to retain the restriction that iomap accepts only
> > > one mapping for the entire file range, like Ritesh has here.
> > 
> > less gross :) sure. 
> > 
> > I would like to think of this as, being less restrictive (compared to
> > only allowing a single fsblock) by adding a constraint on the atomic
> > write I/O request i.e.  
> > 
> > "Atomic write I/O request to a region in a file is only allowed if that
> > region has no partially allocated extents. Otherwise, the file system
> > can fail the I/O operation by returning -EINVAL."
> > 
> > Essentially by adding this constraint to the I/O request, we are
> > helping the user to prevent atomic writes from accidentally getting
> > torned and also allowing multi-fsblock writes. So I still think that
> > might be the right thing to do here or at least a better start. FS can
> > later work on adding such support where we don't even need above
> > such constraint on a given atomic write I/O request.
> 
> On today's ext4 call, Ted and Ritesh and I realized that there's a bit
> more to it than this -- it's not possible to support untorn writes to a
> mix of written/(cow,unwritten) mappings even if they all point to the
> same physical space.  If the system fails after the storage device
> commits the write but before any of the ioend processing is scheduled, a
> subsequent read of the previously written blocks will produce the new
> data, but reads to the other areas will produce the old contents (or
> zeroes, or whatever).  That's a torn write.

I'm *really* surprised that people are only realising that IO
completion processing for atomic writes *must be atomic*.

This was the foundational constraint that the forced alignment
proposal for XFS was intended to address. i.e. to prevent fs
operations from violating atomic write IO constraints (e.g. punching
sub-atomic write size holes in the file) so that the physical IO can
be done without tearing and the IO completion processing that
exposes the new data can be done atomically.

> Therefore, iomap ought to stick to requiring that ->iomap_begin returns
> a single iomap to cover the entire file range for the untorn write.  For
> an unwritten extent, the post-recovery read will see either zeroes or
> the new contents; for a single-mapping COW it'll see old or new contents
> but not both.

I'm pretty sure we enforced that in the XFS mapping implemention for
atomic writes using forced alignment. i.e.  we have to return a
correctly aligned, contiguous mapping to iomap or we have to return
-EINVAL to indicate atomic write mapping failed.

Yes, we can check this in iomap, but it's really the filesystem that
has to implement and enforce it...

-Dave.
-- 
Dave Chinner
david@fromorbit.com

On Mon, Nov 04, 2024 at 12:52:42PM +1100, Dave Chinner wrote:
> On Thu, Oct 31, 2024 at 02:36:40PM -0700, Darrick J. Wong wrote:
> > On Sat, Oct 26, 2024 at 10:05:44AM +0530, Ritesh Harjani wrote:
> > > > This gets me to the third and much less general solution -- only allow
> > > > untorn writes if we know that the ioend only ever has to run a single
> > > > transaction.  That's why untorn writes are limited to a single fsblock
> > > > for now -- it's a simple solution so that we can get our downstream
> > > > customers to kick the tires and start on the next iteration instead of
> > > > spending years on waterfalling.
> > > >
> > > > Did you notice that in all of these cases, the capabilities of the
> > > > filesystem's ioend processing determines the restrictions on the number
> > > > and type of mappings that ->iomap_begin can give to iomap?
> > > >
> > > > Now that we have a second system trying to hook up to the iomap support,
> > > > it's clear to me that the restrictions on mappings are specific to each
> > > > filesystem.  Therefore, the iomap directio code should not impose
> > > > restrictions on the mappings it receives unless they would prevent the
> > > > creation of the single aligned bio.
> > > >
> > > > Instead, xfs_direct_write_iomap_begin and ext4_iomap_begin should return
> > > > EINVAL or something if they look at the file mappings and discover that
> > > > they cannot perform the ioend without risking torn mapping updates.  In
> > > > the long run, ->iomap_begin is where this iomap->len <= iter->len check
> > > > really belongs, but hold that thought.
> > > >
> > > > For the multi fsblock case, the ->iomap_begin functions would have to
> > > > check that only one metadata update would be necessary in the ioend.
> > > > That's where things get murky, since ext4/xfs drop their mapping locks
> > > > between calls to ->iomap_begin.  So you'd have to check all the mappings
> > > > for unsupported mixed state every time.  Yuck.
> > > >
> > > 
> > > Thanks Darrick for taking time summarizing what all has been done
> > > and your thoughts here.
> > > 
> > > > It might be less gross to retain the restriction that iomap accepts only
> > > > one mapping for the entire file range, like Ritesh has here.
> > > 
> > > less gross :) sure. 
> > > 
> > > I would like to think of this as, being less restrictive (compared to
> > > only allowing a single fsblock) by adding a constraint on the atomic
> > > write I/O request i.e.  
> > > 
> > > "Atomic write I/O request to a region in a file is only allowed if that
> > > region has no partially allocated extents. Otherwise, the file system
> > > can fail the I/O operation by returning -EINVAL."
> > > 
> > > Essentially by adding this constraint to the I/O request, we are
> > > helping the user to prevent atomic writes from accidentally getting
> > > torned and also allowing multi-fsblock writes. So I still think that
> > > might be the right thing to do here or at least a better start. FS can
> > > later work on adding such support where we don't even need above
> > > such constraint on a given atomic write I/O request.
> > 
> > On today's ext4 call, Ted and Ritesh and I realized that there's a bit
> > more to it than this -- it's not possible to support untorn writes to a
> > mix of written/(cow,unwritten) mappings even if they all point to the
> > same physical space.  If the system fails after the storage device
> > commits the write but before any of the ioend processing is scheduled, a
> > subsequent read of the previously written blocks will produce the new
> > data, but reads to the other areas will produce the old contents (or
> > zeroes, or whatever).  That's a torn write.
> 
> I'm *really* surprised that people are only realising that IO
> completion processing for atomic writes *must be atomic*.

I've been saying for a while; it's just that I didn't realize until last
week that there were more rules than "can't do an untorn write if you
need to do more than 1 mapping update":

Untorn writes are not possible if:
1. More than 1 mapping update is needed
2. 1 mapping update is needed but there's a written block

(1) can be worked around with a log intent item for ioend processing,
but I don't think (2) can at all.  That's why I went back to saying that
untorn writes require that there can only be one mapping.

> This was the foundational constraint that the forced alignment
> proposal for XFS was intended to address. i.e. to prevent fs
> operations from violating atomic write IO constraints (e.g. punching
> sub-atomic write size holes in the file) so that the physical IO can
> be done without tearing and the IO completion processing that
> exposes the new data can be done atomically.
> 
> > Therefore, iomap ought to stick to requiring that ->iomap_begin returns
> > a single iomap to cover the entire file range for the untorn write.  For
> > an unwritten extent, the post-recovery read will see either zeroes or
> > the new contents; for a single-mapping COW it'll see old or new contents
> > but not both.
> 
> I'm pretty sure we enforced that in the XFS mapping implemention for
> atomic writes using forced alignment. i.e.  we have to return a
> correctly aligned, contiguous mapping to iomap or we have to return
> -EINVAL to indicate atomic write mapping failed.

I think you guys did, it's just the ext4 bigalloc thing that started
this back up again. :/

> Yes, we can check this in iomap, but it's really the filesystem that
> has to implement and enforce it...

I think we ought to keep the "1 mapping per untorn write" check in iomap
until someone decides that it's worth the trouble to make a filesystem
handle mixed states correctly.  Mostly as a guard against the
implementations.

--D

> -Dave.
> -- 
> Dave Chinner
> david@fromorbit.com
>