[Question] Editing a bio write request

[Question] Editing a bio write request

[Jasper Surmont]

I'm writing a device-mapper which has to edit incoming writes, for
example: Incoming bio request contains a biovec with a page full of 0
bytes, and I want to change it to all 0xFF bytes.

I tried using bvec_kmap_local() to get the page address, to then read
the data and if needed adjust it using memcpy or similar. Initial
tests seemed to work, but if I execute things like mkfs on the created
dm, I get a lot of segmentation faults. I narrowed down the problem
that this only happens if I actually write something to the mapped
page. However, I see no reason why it should cause this fault as I
(after checking probably 100 times) don't access invalid memory. I'm
wondering whether my method of editing the write is actually correct,
or if I am doing something else very stupid.

Any help is appreciated

Thanks!

Cheers, Jasper

Re: [Question] Editing a bio write request

[Eric Biggers]

On Wed, May 18, 2022 at 01:56:31PM +0300, Jasper Surmont wrote:
> I'm writing a device-mapper which has to edit incoming writes, for
> example: Incoming bio request contains a biovec with a page full of 0
> bytes, and I want to change it to all 0xFF bytes.
> 
> I tried using bvec_kmap_local() to get the page address, to then read
> the data and if needed adjust it using memcpy or similar. Initial
> tests seemed to work, but if I execute things like mkfs on the created
> dm, I get a lot of segmentation faults. I narrowed down the problem
> that this only happens if I actually write something to the mapped
> page. However, I see no reason why it should cause this fault as I
> (after checking probably 100 times) don't access invalid memory. I'm
> wondering whether my method of editing the write is actually correct,
> or if I am doing something else very stupid.
 
Data in bios can't be modified in-place, as submitters of writes (e.g.,
filesystems) expect that their data buffers aren't modified.  To modify the data
you'd need to allocate a new bio with new pages and submit that instead.  See
dm-crypt for an example of a device-mapper target which does this.

What sort of device-mapper target are you trying to implement, anyway?  If
you're trying to do encryption, dm-crypt already solves that problem.

- Eric

Re: [Question] Editing a bio write request

[Jasper Surmont]

> Data in bios can't be modified in-place, as submitters of writes (e.g.,
> filesystems) expect that their data buffers aren't modified.  To modify the data
> you'd need to allocate a new bio with new pages and submit that instead. 

Oh wow thanks! I feel stupid. Just to be sure so that I don't send 
'noob' questions like this in the future; could I have found this 
anywhere in the documentation? The block layer is a lot to take in and I 
don't always find the right stuff online.


> What sort of device-mapper target are you trying to implement, anyway?  If
> you're trying to do encryption, dm-crypt already solves that problem.

It's more of an experiment, using compression!

Thanks!

Cheers, Jasper