Re: [PATCH] RISC-V: Add CONFIG_PREEMPT=y to defconfig

[Vincent Chen <deanbo422@gmail.com>]

Palmer Dabbelt <palmer@sifive.com> 於 2018年12月27日 週四 上午1:19寫道：
>
> On Tue, 25 Dec 2018 19:23:50 PST (-0800), linux@roeck-us.net wrote:
> > On Wed, Dec 19, 2018 at 09:25:21PM -0800, Guenter Roeck wrote:
> >> On Wed, Dec 19, 2018 at 05:45:11PM -0800, Guenter Roeck wrote:
> >> > On Wed, Nov 21, 2018 at 10:58:05AM -0800, Palmer Dabbelt wrote:
> >> > > I don't remember who suggested this, but I was talking to someone at
> >> > > Plumbers about all the horrible bugs we've had lurking around in our
> >> > > port despite it still getting to userspace and they suggested that I
> >> > > enable CONFIG_PREEMPT in our defconfig as it'll help shake out more
> >> > > issues.
> >> > >
> >> > > This patch enables CONFIG_PREEMPT in the single RISC-V defconfig.  It
> >> > > still passes my tests (ie, booting Fedora on QEMU).  Maybe I'm too
> >> > > paranoid, but I still don't trust our port enough to just pull the
> >> > > trigger on this without at least accumulating some "Tested-by"s on it.
> >> > >
> >> > > Signed-off-by: Palmer Dabbelt <palmer@sifive.com>
> >> >
> >> > I can only give you a
> >> >
> >> > Fails-miserably-for: Guenter Roeck <linux@roeck-us.net>
> >> >
> >> > Not the new defconfig itself, but the new defconfig plus a variety of
> >> > selftest options. I don't know what exactly is failing, but here is
> >> > my list of enabled selftests on top of defconfig:
> >> >
> >> > CONFIG_CRYPTO_MANAGER_DISABLE_TESTS=y
> >> > CONFIG_CRC32_SELFTEST=y
> >> > CONFIG_DEBUG_LOCKING_API_SELFTESTS=y
> >> > CONFIG_DEBUG_NMI_SELFTEST=y
> >> > CONFIG_DEBUG_RODATA_TEST=y
> >> > CONFIG_DEBUG_TLBFLUSH=y
> >> > CONFIG_DMATEST=y
> >> > CONFIG_GLOB_SELFTEST=y
> >> > CONFIG_OF_UNITTEST=y
> >> > CONFIG_PCI_EPF_TEST=y
> >> > CONFIG_PCI_ENDPOINT_TEST=y
> >> > CONFIG_PROVE_LOCKING=y
> >> > CONFIG_PROVE_RCU=y
> >> > CONFIG_RCU_EQS_DEBUG=y
> >> > CONFIG_STATIC_KEYS_SELFTEST=y
> >> > CONFIG_STRING_SELFTEST=y
> >> > CONFIG_TEST_BITMAP=y
> >> > CONFIG_TEST_FIRMWARE=y
> >> > CONFIG_TEST_SORT=y
> >> > CONFIG_TEST_SYSCTL=y
> >> > CONFIG_TEST_UUID=y
> >> > CONFIG_USB_TEST=y
> >> > CONFIG_USB_EHSET_TEST_FIXTURE=y
> >> > CONFIG_USB_LINK_LAYER_TEST=y
> >> > CONFIG_WW_MUTEX_SELFTEST=y
> >> > CONFIG_TORTURE_TEST=y
> >> > CONFIG_LOCK_TORTURE_TEST=y
> >> > CONFIG_RCU_TORTURE_TEST=y
> >> > CONFIG_RBTREE_TEST=y
> >> > CONFIG_INTERVAL_TREE_TEST=y
> >> >
> >>
> >> Update: defconfig+CONFIG_LOCK_TORTURE_TEST=y causes the hang
> >> if CONFIG_PREEMPT=y.
> >>
> >
> > Another update. CONFIG_RCU_TORTURE_TEST=y or CONFIG_WW_MUTEX_SELFTEST=y
> > also result in hangs, only a bit later. I have now disabled all three test
> > options in my riscv boot tests if CONFIG_PREEMPT=y. The lack of error
> > reports in the future does _not_ mean that the problem has magically
> > disappeared. I just find it pointless to keep testing a known broken
> > configuration, even more so since doing so would hide other potential
> > problems.
>
> Thanks.  I agree it's pointless to test more, as there appear to be lots of
> bugs to fix already and this will just mask other bugs.
>

Last week, I encountered some locking issues in the kernel, so I tried
to test the
kernel with some relative selftests. Since I had already set CONFIG_PREEMPT=y,
I encountered the same results in RCU_TORTURE_TEST and
LOCK_TORTURE_TEST. I tried to debug it and found the hang was due to
the lack of _TIF_NEED_RESCHED check for kernel threads in entry.S. I will send a
patch to solve this problem and I hope you can give me some comments
if you find some
mistakes.

Vincent Chen

_______________________________________________
linux-riscv mailing list
linux-riscv@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-riscv