On 2019-08-09 1:00 a.m., James Bottomley wrote:
On Thu, 2019-08-08 at 23:08 +0200, Douglas Gilbert wrote:
On 2019-08-08 9:10 p.m., James Bottomley wrote:
On Wed, 2019-08-07 at 13:42 +0200, Douglas Gilbert wrote:
This patchset extends the SCSI generic (sg) driver found in
lk 5.3 .  The sg driver has a version number which is visible
via ioctl(SG_GET_VERSION_NUM) and is bumped from 3.5.36 to
4.0.03 by this patchset. The additions and changes are
described in some detail in this long webpage:
      http://sg.danny.cz/sg/sg_v40.html

Most new features described in the above webpage are not
implemented in this patchset.

Since this will be an extension of something that exists both in
your sg driver and in the block bsg interface (and thus needs an
implementation there), I added both linux-block and linux-api to
the cc (the latter because you're adding to an API).

The SG_IO ioctl has been the synchronous SCSI pass-through interface
for over 15 years. Its quirk is that it takes two different formats
depending on the character device its file descriptor belongs to:
     - sg device file descriptor: sg v3 interface (struct sg_io_hdr)
     - bsg device file descriptor: sg v4 interface (struct sg_io_v4)

I'm only proposing one change in the synchronous interface based
on the SG_IO ioctl: to additionally accept the sg v4 interface in
the sg driver.

Right, as I said below, adding v4 to sg looks to be fine.  We don't
need v3 in BSG since v4 is a superset.

Arnd Bergmann considers two interface structures through one ioctl
as undesirable but is prepared to accept SG_IO. His POV is as the
maintainer of 32/64 bit compatibility ioctls. The case of SG_IO
is a well established exception to his rules (i.e. a known evil).

I don't believe extending ioctl SG_IO for asynchronous work is a
good idea. As pointed out above, it is already overloaded too
much. Additionally it would need further flags to differentiate
these cases:
     - sync/async
     - if async: submission or reception
     - and optionally if async: abort (an inflight request)
     - and do you want to add multiple requests in there too?

So are you looking at reducing the number of ioctl to the absolute
minimum? If so I don't think the SG_IO ioctl is the correct vehicle
for that. It doesn't use the _IOR(W) macros, instead it is hard
coded at 0x2285 ***. And the size checking offered by the _IOR(W)
macro (on the object pointed to by the 3rd argument) is useless with
SG_IO because it takes two different sized objects. Worse, one of
those objects changes size between 32 and 64 bits, while the other
does not.

OK, so this is where interface design is important.  It's perfectly
possible to design an async interface using the current v4 BSG SG_IO.
You simply open the device O_NONBLOCK and send the ioctl which will
return immediately.  To find out when it returns you poll (we'd
obviously need to add polling in bsg, but that's fairly simple).  Since
  poll/select/epoll is our standard async event handling mechanism, we
should at least try it and have a good reason why it doesn't work
before inventing our own equivalent.

I don't see much new interface design here over what was done by
Lawrence Foard 27 years ago. Linux cleverness has broken the
classic use of read() and write() in char drivers. Hard to hold that
against the original design. The design works if the last 27 years
is any guide and maps reasonably closely to how the mid-level
(or the block layer) issues SCSI commands. All other general
purpose OSes that I am aware of, with the exception of Darwin,
have a similar SCSI pass-through, at least on the sync side.

Stepping back, this started about 18 months ago when security
janitors got upset about the bsg driver's use of write()/read() for
its async interface. Linus Torvalds suggested SG_IOSUBMIT and
SG_IORECEIVE to replace write() and read() respectively in bsg.
Instead the write()/read() interface was removed from the bsg driver.
With it went the ability to submit multiple requests in one write()
call (by passing an array of sg_io_v4 objects rather than just one).
My intention is to re-add that capability in the sg driver, using the
ioctls that Linus suggested.

What I'm saying is let's discuss the interface design before we add it.
  the submit/receive interface is essentially a packet one.  The problem
with a packet one is that we're request/response where the response
likely altered some user memory area, so each response has to be
matched with the outbound request to find out what happened ... how is
that done?

Well this has been fundamental to the sg driver design. There are
now three mechanisms, described below with the oldest first:
  1) pack_id: an identifying integer supplied by the user in the
     submit that can be given as input to the receive call so
     it either yields the (first) matching response, or waits (sync
     action) or returns saying nothing was found (async action).
     pack_id was present in the sg driver in version 1.0 of Linux
     in 1992. It still works fine.
  2) usr_ptr: sometimes described as a "closure pointer". User
     provided, untouched by the driver, and placed in the
     corresponding response. Present in sg v3 and v4 interfaces.
     In place over 15 years.
  3) tag: like pack_id but issued by the block layer so it is
     _output_ by the submit call. In practice it does not work
     reliably *** due to questionable design in the block layer. After
     blk_execute_rq_nowait() is called the tag is picked out of
     its data structure. However that command may not have started
     (so no valid tag) or could have already finished (a stale tag).
     Tag is a new feature in the proposed patchsets.

One way to do it simply with existing SG_IO is to allow only one
outstanding request per FD, so if you want five requests, you open five
FDs and then poll all of them to see which one returns.  I'm not saying
it's the best because FDs are a somewhat precious commodity, but it is
possible and it's a simple extension of what exists today.

I believe this is adequately covered by the mechanisms above.

Initially I had both the sg v3 and v4 interfaces passing through the
two ioctls. Arnd Bergmann preferred that a separate pair of ioctls
be used for each interface. Hence SG_IOSUBMIT_V3 and SG_IORECEIVE_V3
were added for the v3 interface. And thus SG_IOSUBMIT and
SG_IORECEIVE only use the v4 interface. This cleaned up my code and
documentation. As a bonus, all four ioctls use the _IORW macros and
can check the fixed size of the third argument to each ioctl
invocation.

Can we talk first about how the interface should work.  We can get into
assigning ioctls and what extensions are needed later.

I have been trying for 9 months to get some feedback on
the current design. The attitude seems to be present some
patches; no, those ones are too big, etc. As for my design
document at:
    http://sg.danny.cz/sg/sg_v40.html

tl;dr ?

Evidently write() and read() system calls have special features
that could be exploited by an attack via the bsg or sg drivers.
ioctl(SG_IOSUBMIT) is just write(sg_fd, ...) that is less
exploitable. There is a similar relationship between
ioctl(SG_IORECEIVE) an read(sg_fd, ...). Otherwise the async
interface design remains the same in the sg driver as it was
when introduced 27 years ago. ^^^

Bart Van Assche hinted at a better API design but didn't present
it. If he did, that would be the first time an alternate API
design was presented for async usage in the 20 years that I have
been associated with the driver.

Simply extending sg to use the v4 header protocol in
uapi/linux/bsg.h is fine modulo the code being in the right
form.  The problems are the new ioctls you want to add that would
need to be present there as well.

Why? The bsg driver has never accepted the sg v3 interface. It has
also removed functionality that I'm trying to re-add in this, and a
follow-on patchset. The bsg driver has reduced its functionality as a
generalized SCSI pass-through, but it has more than enough other
roles to justify its existence, for example as a SMP (SAS) pass-
through and a driver/transport specific pass-through to various LLDs.
I don't see the need for the sg and bsg driver to move forward in
lockstep.

You quote Linus a lot above, but he also said "I wonder if we could at

twice

least try to unify the bsg/sg code - possibly by making sg use the
prettier bsg code (but definitely have to add all the security
measures)".  It's in the interests of unity that we need to make the
code paths look the same as possible, so eventually one could call the
other.  ideally sg would call bsg for v4 and we'd add async to bsg so
it would work for both.

The bsg driver needs to be rewritten. Ah, no need since by removing the
async (including multiple requests) part, the broken implementation is
no more. Its implementation is sufficient to issue synchronous
SCSI commands, not much more +++. The bsg v4 interface implementation
was only a superset of the sg driver's v3 interface while the Linux
kernel supported bidi. That has unwisely been (completely) removed
from the kernel. So now I would regard the bsg v4 implementation
as a subset of the functionality in the sg v3 implementation which
itself is a subset of the proposed sg v4 implementation. So I see
no reason to use the bsg v4 implementation or handicap the proposed
sg v4 implementation.

   The specific question being how we support async or non-blocking
I/O on the sg and bsg interfaces.  The standard way we add
asynchronous I/O is supposed to be via .poll on the file
descriptor.  you already use read and write in sg and bsg doesn't
have a polling interface,

It is hard to parse that last sentence; the sg driver has always
supported select/poll/epoll (and SIGIO plus RT signals). sg_poll()
will work just as well irrespective of whether a request is submitted
by write(), ioctl(SG_IOSUBMIT) or ioctl(SG_IOSUBMIT_V3).

The way I read how the interface works, it only works when the fd is
ready for read (i.e. a packet submitted with read has returned).  We
need to extend it to make it work with ioctl.  Since ioctl returns
isn't either a 'ready to read' or 'ready to write' event on the fd I
was suggesting using 'msg ready' event for it.

A ready response causes a POLL_IN event. POLL_OUT is only cleared when
the command queueing on a fd is turned off and a request is inflight or
awaiting a ioctl(SG_IORECEIVE) (a.k.a. read()). So the existing POLL
events map quite well. As a bonus, if there is a surprise removal of
a sg device, then a SIG_HUP is generated.

As I noted, if an app doesn't need to monitor multiple file
descriptors, then ioctl(SG_GET_NUM_WAITING) is an option. It been in
place for 20 years and is sped up by this patchset. I clocked it at
500 nanoconds per call (i.e. 2 million could be issued in one thread
in 1 second) on my laptop. And since you then find out how many are
waiting to be processed, the user could issue a multiple requests
ioctl(SG_IORECEIVE) and pick them all up in one system call invocation.

but it looks like we could use MSG to signal an ioctl is ready to
be serviced for both.  Would shifting to a non-blocking poll based
interface for ioctls remove the need to add these
SG_IOSUBMIT/SG_IORECEIVE ioctls since we could now do everything
over blocking or non-blocking SG_IO?

Not sure what the MSG is you refer to. The sg driver has had
ioctl(SG_GET_NUM_WAITING) for a long time. And it is made even
more lightweight in this patchset: it takes no locks, just reads one
atomic (a counter obviously) and returns. My guess is its faster the
select()/poll()/epoll() but doesn't have the ability to monitor
multiple file descriptors.

the poll interfaces don't tell you how many outstanding request you
have, they tell you when some event happened on the fd.  The event
we're looking for with async packets is an indication of what
completed.

As well as ioctl(SG_GET_NUM_WAITING) described above there is also
ioctl(SG_GET_PACK_ID) which will return the pack_id (or tag) of the
response that has been waiting the longest to be processed (or -1
if nothing is waiting).

In the area of async polling the existing sg driver supports both
standard Unix techniques and driver-specific techniques that convey
extra information.

Here is the full set of extra ioctls I have, or will be proposing:
     SG_IOSUBMIT
     SG_IOSUBMIT_V3
     SG_IORECEIVE
     SG_IORECEIVE_V3
     SG_IOABORT
     SG_SG_SET_GET_EXTENDED

Well, i/o cancellation is another huge can of worms, but let's get the
submit interface sorted out first before worrying about how
cancellation works.

Yes, and its a great way to test code. The sgh_dd utility (sg3_utils
package, testing folder) has a ae=AEN option for "abort every n
commands". When it decides to abort a command it sets up another
(killer) thread that waits a random amount of time (within bounds)
then issues ioctl(SG_IOABORT). Great fun. For even more fun there
is ae=AEN,MAEN where the extra MAEN parameter is for "abort every
n multiple requests invocations". There are a lot more corner cases
when aborting a multiple requests invocation. Testing with these
found several problems with my code.

It seems to me we are not actually on the same page when it comes
to where this project stands.

They are all new style ioctls using the _IOR(W) macros with fixed
size objects referenced by the ioctl's third argument. ioctls have
been referred to as the "garbage bin of Unix". Well that last one is
a garbage bin within a garbage bin :-) On the plus side, it keeps
that list relatively short.

Doug Gilbert


*** Tony Battersby is a sg driver power user. He has lamented wading
through very large logs looking for some hint of why the sg driver is
playing up he has stated the strong preference for more, not less,
ioctls.

I'm not really bothered about the number of ioctls; I'm bothered about
getting the interface right.

Great. Then please leave ioctl(SG_IO) as is (i.e. sync only).

BTW the write()/read() interface still remains in the sg driver after
these patchsets. It will continue to only support the sg v3
interface. Perhaps calling it should cause a "deprecated" log message
once for each kernel run to annoy maintainers of old code.

That would be ideal given all the security complaints we have about it.

That is easily done.

Doug Gilbert


*** at least with scsi_debug with a command delay set at 5000 nanoseconds.
    With real storage devices with latencies around 50 microseconds or higher
    it may be more reliable. Some sort of handle or the tag itself back from
    blk_execute_rq_nowait() is the real solution.

+++ but the bsg driver can do much more than issue SCSI commands. Question
    is, given the removal of its SCSI async functionality, should it be
    generating SCSI commands at all? Just the general ioctl(SG_IO) as
    supported by sd, sr and st devices should suffice.

^^^ another approach to the write()/read() security problems would be for
    their implementation to check if the given file descriptor was to a
    char (and maybe block) device and if so reduce its exploitable behaviour
    accordingly.

Reply via email to