[LSF/MM TOPIC] Implementing a userland interface for data integrity passthrough

["Darrick J. Wong" <darrick.wong@oracle.com>]

Hello LSF committee,

I'm interested in attending several of the discussions that have
already been brought up on the mailing list:

 * Direct IO rework, specifically the part that has to do with the
   future of DIX/PI interfaces (see below).

 * The IOC_[GS]ETFLAGS mess -- I'd like to drive the new interface to
   a conclusion so that we can deprecate the broken interface and get
   on with the userland part.

 * Copy offloading, mostly with regards to whatever it is ext4 might
   want to do to implement reflink support.  Either this happens
   through magic provided by the underlying storage (XCOPY) or I guess
   we could consider changes the (ext4) extent tree format for anyone
   not posessing such fancy storage.

 * I'd probably also attend an SMR discussion.

 * Fuzzing filesystems, though my interest is in fuzzing images, not
   so much fuzzing the system calls via Trinity.  In my copious spare
   time, I've thought of updating fsfuzzer to know about on-disk
   structures so that we can do better testing than simply randomly
   corrupting bits to see what happens.  In particular I think it'd be
   interesting to see how the code handles almost-too-big numbers and
   the like. (Needs more research.)

 * The ext4 meeting, since it sounds like there will be one.  I'm
   preparing a few patches to try to speed up e2fsck via mmap and
   threaded prefetch, and maybe some other silly parallelization
   tricks. (Needs more research.)  Also I imagine it might be
   useful to touch on whatever the plan is for putting new features
   into a release (32->64bit conversion, userland ext4, metadata
   checksumming...)

I also have my own topic -- implementing a userland interface for
passing integrity metadata through to the storage.  This is the usage
model that I'd set up in my head (at the kernel<->userland boundary):

 1. Program opens a file descriptor.

 2. Program sets up a aio context.

 3. Program queries the fd for supported PI profiles (probably an
    ioctl).

 4. Program uses newly defined IO_CMD_{READV,WRITEV}_PI commands
    to supply PI data and verify the data itself.  A new structure
    will be defined to report the PI profile the app wants to use,
    which fields the app is actually interested in providing or
    verifying, a bitset of which devices should check the PI data (HBA,
    disk, intermediate storage servers), and followed by space for the
    actual PI data; then either we find space in struct iocb to point
    to this buffer, or we do something naughty such as attaching it as
    the first (or last) iovec pointer.

    libaio can take care of all this for a client program.  A separate
    discussion could be had about the interface from libaio to client
    programs, but let's get the kernel<->user piece done first.

 5. Error codes ... perhaps we define a IO_CMD_GET_ERROR command that
    doesn't return an event until it has extended error data to
    supply.  This could be more than just PI failures -- SCSI sense
    data seems like a potential choice.  This is a stretch goal...

The raw kernel interface of course would be passing PI profiles and
data to userspace, for anyone who wishes to bypass libaio.

As for ioctl that describes what kind of PI data the kernel will
accept, I'd like to make it generic enough that someone could
implement a device with any kind of 'checksum' (crc32c, sha1, or maybe
even a cryptographic signature), while allowing for different
geometrical requirements, or none, as in the case of byte streams over
NFS.  It's been suggested to use unique integer values and assume that
programs know what the values mean, but given the potential for
variety I wonder if it should be more descriptive:

{
	name: "NFS-FOO-NONSTANDARD",
	granularity: 0,
	alignment: 0,
	profile: "tag,checksum",
	tag-width: u32,
	checksum-alg: sha256,
	checksum-width: u8[32],
}
or
{
	name: "tag16-crc16-block32",
	granularity: 512,
	alignment: 512,
	profile: "tag,checksum,reftag",
	tag-width: u16,
	checksum-alg: crc16,
	checksum-width: u16,
	reftag-alg: blocknum,
	reftag-width: u32,
}

Now, for the actual mechanics of modifying the kernel, here's my idea:
First, enhance the block_integrity API so that we can ask it about
supported data formats, algorithms, etc. (everything we need to supply
the schema described in the previous section).

For buffered mode, each struct page would point to a buffer that is
big enough to hold all the PI data for all the blocks represented by
the page, as well as descriptors for the PI data.  This gets much
harder for the case of arbitrary byte streams instead of disk sectors.
Perhaps we'd have to have a descriptor that looks like this:

struct {
  u16 start, end;
  int flags;
  void *buffer;
  char[16] pi_profile;
};

In the case of byte stream PI, I'm not sure how the NFS protocols
would handle overlapping ranges -- send one page with the set of PIs
that cover that page?

Anyway, when a buffered write comes in, we simply copy the user's
buffer into the thing hanging off struct page.  When
bio_integrity_prep is called (during submit_bio), it will either find
no buffer and generate the PI data on its own like it does now, or
it'll find a buffer, attach it to the bio->bip, then ask the integrity
provider to fill in whatever's missing.  A directio write would take
the PI data and attach it directly to the bio it submits.

For buffered reads, bio_integrity_endio can allocate the buffer and
attach it to struct page, then fill in the fields that the disk
returned.  The actual userland read function of course can then copy
the data out of the thing hanging off struct page into the user's
buffer, and then userland can do whatever it wants.  A directio read
simply copies the data from the bio->bip into the userland buffer.

As for the GET_ERROR thing, my first (and probably only) thought was
to find a way to attach a buffer and a description of what's in the
buffer to a bio, so that GET_ERROR can return the buffer contents.
A tricky part is to help out userspace by mapping an error code back
to the iocb.  I need to think harder about this piece.  Right now I'm
only thinking about disk storage; is anyone else interested enough in
returning rich error data to userland to help me bikeshed? :)

--D

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>