[PATCH v5 00/30] ARM Scalable Vector Extension (SVE)

Mon Jan 15 08:55:48 PST 2018

On Mon, Jan 08, 2018 at 05:49:05PM +0300, Yury Norov wrote:
> On Tue, Oct 31, 2017 at 03:50:52PM +0000, Dave Martin wrote:
> > This series implements Linux kernel support for the ARM Scalable Vector
> > Extension (SVE). [1]  It supersedes the previous v3: see [3] for link
> > and full cover letter.
> > 
> > This is a minor update to v4, but does contain a couple of important
> > fixes.
> > 
> > As in previous postings, the last two patches (here 29-30) are still
> > RFC and not proposed for merging at this time.
> > 
> > The patches apply on
> > git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux.git
> > for-next/core
> > d7b1d22d3821 ("arm64: uapi: Remove PSR_Q_BIT")
> > 
> > To reduce spam, some people may not been copied on the entire series.
> > For those who did not receive the whole series, it can be found in the
> > linux-arm-kernel archive. [2]
> > 
> > See the individual patches for details of changes.
> > 
> > For reviewer convenience, a git tree is available. [4]
> > 
> > Since there are some changes against already-reviewed patches, I've also
> > pushed an unsquashed fixes tree for people to take a look at if it
> > helps. [5]
> > 
> > 
> > Summary:
> > 
> >  * "regset: Add support for dynamically sized regsets" fixed to avoid
> >    x86 breakage;
> > 
> >  * one trival arm64 patch added to add asmlinkage annotations, and a
> >    corresponding minor change to the Core task context handling patch;
> > 
> >  * one new arm64 fix ("signal: Verify extra data is user-readable in
> >    sys_rt_sigreturn") to ensure that access_ok() checks are done for the
> >    whole extended signal frame, not just the base frame;
> > 
> >  * one minor fix to the SVE sigreturn code to return consistent
> >    intermediate error values (semantic correctness, non-functional
> >    change);
> > 
> >  * one minor change to call __copy_from_user() instead of
> >    copy_from_user() in a situation where there is already an access_ok()
> >    check;
> > 
> > 
> > [1] ARM Scalable Vector Extension
> > https://community.arm.com/groups/processors/blog/2016/08/22/technology-update-the-scalable-vector-extension-sve-for-the-armv8-a-architecture
> > 
> > [2] linux-arm-kernel October 2017 Archives by thread
> > http://lists.infradead.org/pipermail/linux-arm-kernel/2017-October/thread.html
> > 
> > [3] [PATCH v4 00/28] ARM Scalable Vector Extension (SVE)
> > http://lists.infradead.org/pipermail/linux-arm-kernel/2017-October/539414.html
> > 
> > [4] For review and testing only -- **do not pull**
> >     (This branch has review changelogs which should not form part of
> >     the final commits.)
> > 
> >     v5 series (this posting)
> > 
> >     http://linux-arm.org/git?p=linux-dm.git;a=shortlog;h=refs/heads/sve/v5
> >     git://linux-arm.org/linux-dm.git sve/v5
> > 
> > [5] For review and testing only -- **do not pull**
> >     (This branch has review changelogs which should not form part of
> >     the final commits.)
> > 
> >     v4 with unsquashed fixes
> > 
> >     http://linux-arm.org/git?p=linux-dm.git;a=shortlog;h=refs/heads/sve/v4%2Bfixes
> >     git://linux-arm.org/linux-dm.git sve/v4+fixes
> 
> Hi Dave,
> 
> During rebase of my ILP32 series on 4.15 kernel I found that 
> ILP32 needs to be enabled with SVE support, like you do for 
> LP64 in this series.
> 
> I did all rebase work on this draft branch:
> https://github.com/norov/linux/tree/ilp32-4.15-rc7
> 
> But any ILP32 program I tried crash, and the message in dmesg looks
> like this:
> [   39.510667] CPU: 0 PID: 1857 Comm: mytime Not tainted 4.15.0-rc7-00028-g45e0659df4d9 #41
> [   39.510712] Hardware name: linux,dummy-virt (DT)
> [   39.510829] pstate: 00000000 (nzcv daif -PAN -UAO)
> [   39.511101] pc : 0x33488e28
> [   39.511125] lr : 0x33488e28
> [   39.511138] sp : 00000000fffef670
> [   39.511158] x29: 000000005a536c33 x28: 0000000000000000 
> [   39.511211] x27: 0000000000000000 x26: 0000000000000000 
> [   39.511235] x25: 0000000000000000 x24: 0000000000000000 
> [   39.511257] x23: 0000000000466000 x22: 0000000000000000 
> [   39.511278] x21: 0000000000000000 x20: 000000000047f2a8 
> [   39.511300] x19: 0000000000000000 x18: 0000000000000001 
> [   39.511321] x17: 0000000000001000 x16: 0000000000001030 
> [   39.511342] x15: 0000000000554e47 x14: 0000000000000001 
> [   39.511364] x13: 0000000000000004 x12: 000000000000003c 
> [   39.511385] x11: 0000100000000000 x10: 0800000000000000 
> [   39.511406] x9 : 0fffffffffffffff x8 : 000000000000007c 
> [   39.511427] x7 : 0000000000000077 x6 : 0000000000000041 
> [   39.511448] x5 : 0000000000000411 x4 : 00000000fbad2488 
> [   39.511468] x3 : 0000000000000001 x2 : 0000000000497950 
> [   39.511489] x1 : 0000000000497550 x0 : 0000000000000001
> 
> (PC, LR and other registers may differ depending on test)
> 
> I'm pretty sure that this is SVE-related issue because if I disable
> ARM64_SVE in config, everything becomes working.

Really, software that does not use SVE should not see any difference.

Do you see any pattern to the failures?  Do they appear to be signal-
related?

Are you running on the SVE-enabled model, or is this on hardware?

> I didn't get deep enough into this yet, and most probably there's
> some stupid reason for crashing apps on my side, but in mail list
> I found some sve-related patches that I cannot apply both on 4.15-rc
> and next-20180108. And according to patch names, it is important
> fixes, like this one:
> https://www.spinics.net/lists/arm-kernel/msg619548.html
> 
> ILP32 code may spit some kernel issues from time to time, so there's
> minor chance that the problem is not in ILP32 itself - that's why I
> write this email to you.
> 
> What I want to ask you, do you have some branch with the most recent
> SVE code that includes all fixes? Better if it would be 4.15-based
> series, because we don't support ILP32 on next-* kernels now.

The above fix was queued for -rc2.  All pending fixes should be in
mainline by now.

> Also, if you have few minutes to take look at my series, I would
> kindly ask you do this, because I still think that this is my
> misunderstanding of how SVE should work for ILP32, and you as author
> may just catch the bug at a glance.

I'll take a look -- I didn't think that carefully about the ILP32
case so far, because it "should" be straightforward.  But there may
be gotchas I hadn't considered so far.

Cheers
---Dave