[Cluster-devel] [PATCH v2 21/23] xfs: handle merkle tree block size != fs blocksize != PAGE_SIZE

[Eric Biggers <ebiggers@kernel.org>]

On Thu, Apr 06, 2023 at 09:37:53AM +1000, Dave Chinner wrote:
> On Wed, Apr 05, 2023 at 10:54:06PM +0000, Eric Biggers wrote:
> > On Thu, Apr 06, 2023 at 08:26:46AM +1000, Dave Chinner wrote:
> > > > We could certainly think about moving to a design where fs/verity/ asks the
> > > > filesystem to just *read* a Merkle tree block, without adding it to a cache, and
> > > > then fs/verity/ implements the caching itself.  That would require some large
> > > > changes to each filesystem, though, unless we were to double-cache the Merkle
> > > > tree blocks which would be inefficient.
> > > 
> > > No, that's unnecessary.
> > > 
> > > All we need if for fsverity to require filesystems to pass it byte
> > > addressable data buffers that are externally reference counted. The
> > > filesystem can take a page reference before mapping the page and
> > > passing the kaddr to fsverity, then unmap and drop the reference
> > > when the merkle tree walk is done as per Andrey's new drop callout.
> > > 
> > > fsverity doesn't need to care what the buffer is made from, how it
> > > is cached, what it's life cycle is, etc. The caching mechanism and
> > > reference counting is entirely controlled by the filesystem callout
> > > implementations, and fsverity only needs to deal with memory buffers
> > > that are guaranteed to live for the entire walk of the merkle
> > > tree....
> > 
> > Sure.  Just a couple notes:
> > 
> > First, fs/verity/ does still need to be able to tell whether the buffer is newly
> > instantiated or not.
> 
> Boolean flag from the caller.
> 
> > Second, fs/verity/ uses the ahash API to do the hashing.  ahash is a
> > scatterlist-based API.  Virtual addresses can still be used (see sg_set_buf()),
> > but the memory cannot be vmalloc'ed memory, since virt_to_page() needs to work.
> > Does XFS use vmalloc'ed memory for these buffers?
> 
> Not vmalloc'ed, but vmapped. we allocate the pages individually, but
> then call vm_map_page() to present the higher level code with a
> single contiguous memory range if it is a multi-page buffer.
> 
> We do have the backing info held in the buffer, and that's what we
> use for IO. If fsverity needs a page based scatter/gather list
> for hardware offload, it could ask the filesystem to provide it
> for that given buffer...
> 
> > BTW, converting fs/verity/ from ahash to shash is an option; I've really never
> > been a fan of the scatterlist-based crypto APIs!  The disadvantage of doing
> > this, though, would be that it would remove support for all the hardware crypto
> > drivers.
> >
> > That *might* actually be okay, as that approach to crypto acceleration
> > has mostly fallen out of favor, in favor of CPU-based acceleration.  But I do
> > worry about e.g. someone coming out of the woodwork and saying they need to use
> > fsverity on a low-powered ARM board that has a crypto accelerator like CAAM, and
> > they MUST use their crypto accelerator to get acceptable performance.
> 
> True, but we are very unlikely to be using XFS on such small
> systems and I don't think we really care about XFS performance on
> android sized systems, either.
> 

FYI, I've sent an RFC patch that converts fs/verity/ from ahash to shash:
https://lore.kernel.org/r/20230406003714.94580-1-ebiggers at kernel.org

It would be great if we could do that.  But I need to get a better sense for
whether anyone will complain...

- Eric