Re: [RFC] On the Current Troubles of Mainlining Loongson Platform Drivers
From: Aaro Koskinen
Date: Mon Jun 10 2019 - 17:54:10 EST
On Sun, May 26, 2019 at 06:19:00AM -0300, Alexandre Oliva wrote:
> On Mar 8, 2019, "Maciej W. Rozycki" <macro@xxxxxxxxxxxxxx> wrote:
> > Anyway I meant: does `war_io_reorder_wmb' expand to `wmb' on your system?
> No, it expands to `barrier' on the yeeloong:
> I've finally managed to do the bisection on object files I mentioned I'd
> do to try to pinpoint where __BUILT_IOPORT_PFX with barrier rather than
> !barrier regressed.
> I found that forcing barrier off for drivers/irqchip/irq-i8259 was
> enough to avoid the problem.
> (I further narrowed it down to byte I/O, which is no surprise
> considering irq-i8259 doesn't seem to use any non-byte I/O.)
> Then I narrowed it down to output only.
> A Loongson2F kernel built with the patch below works at normal speed.
> I've also keyed the -1 barrier selector to compiling the irq-i8259
> driver only, and that got me a functional kernel, but I'm not confident
> that the same issues that affect the interrupt controller, preventing it
> from initializing correctly, is not also affecting other drivers, just
> in less visible ways, so the patch conservatively reverts to the older
> barriers for all I/O (i.e., non-mem) out primitives.
> I've tested this on a yeeloong on top of v5.1.5.
> I'm tempted to start using this patch in my Freeloong builds of GNU
> Linux-libre for gnewsense/yeeloong of 5.0 and 5.1 stable releases, to
> try to make them usable. Can anyone suggest any reason why it might be
> risky to do so, moving on as much as I could to the new barriers,
> sticking to the 4.19-one only for non-mem out? As in, could mixing the
> barriers be riskier than reverting to the 4.19 barriers everywhere?
> Thanks in advance for any insights and recommendations,
Thanks for this work! I'm not yet sure if this completely solves the
issue, but it's surely an improvement.
Testing multiple reboots on Loongson Mini-PC with the libata driver,
/proc/irq/14/spurious still shows tens of thousands spurious interrupts,
last_unhandled 4294673096 ms
where as it with the legacy IDE driver it stays reliably at 0 and the total
count is well under 10000.
However, with your patch the "nobody cared" is never reached so all is
good. I tried 10 boots with the patch and all were successful. Without
the patch 8 out of 10 failed with the "nobody cared" warning.
> diff --git a/arch/mips/include/asm/io.h b/arch/mips/include/asm/io.h
> index 845fbbc7a2e3..04be4758d4ff 100644
> --- a/arch/mips/include/asm/io.h
> +++ b/arch/mips/include/asm/io.h
> @@ -416,7 +416,7 @@ static inline void pfx##out##bwlq##p(type val, unsigned long port) \
> volatile type *__addr; \
> type __val; \
> - if (barrier) \
> + if (barrier > 0) \
> iobarrier_rw(); \
> else \
> war_io_reorder_wmb(); \
> @@ -467,13 +467,22 @@ BUILDIO_MEM(w, u16)
> BUILDIO_MEM(l, u32)
> BUILDIO_MEM(q, u64)
> -#define __BUILD_IOPORT_PFX(bus, bwlq, type) \
> - __BUILD_IOPORT_SINGLE(bus, bwlq, type, 1, 0,) \
> - __BUILD_IOPORT_SINGLE(bus, bwlq, type, 1, 0, _p)
> +#define __BUILD_IOPORT_PFX(bus, bwlq, type, barrier) \
> + __BUILD_IOPORT_SINGLE(bus, bwlq, type, barrier, 0,) \
> + __BUILD_IOPORT_SINGLE(bus, bwlq, type, barrier, 0, _p)
> +/* Choose the kind of barrier used for out in __BUILD_IOPORT_SINGLE in
> + non-__mem_ variants. On Loongson2F, irq-i8259 fails to initialize
> + when this is defined to 1. */
> +#if defined(CONFIG_CPU_LOONGSON2)
> +#define USE_IO_BARRIER_FOR_NON_MEM_OUT -1
> +#define USE_IO_BARRIER_FOR_NON_MEM_OUT 1
> #define BUILDIO_IOPORT(bwlq, type) \
> - __BUILD_IOPORT_PFX(, bwlq, type) \
> - __BUILD_IOPORT_PFX(__mem_, bwlq, type)
> + __BUILD_IOPORT_PFX(, bwlq, type, USE_IO_BARRIER_FOR_NON_MEM_OUT) \
> + __BUILD_IOPORT_PFX(__mem_, bwlq, type, 2)
> BUILDIO_IOPORT(b, u8)
> BUILDIO_IOPORT(w, u16)
> Alexandre Oliva, freedom fighter he/him https://FSFLA.org/blogs/lxo
> Be the change, be Free! FSF Latin America board member
> GNU Toolchain Engineer Free Software Evangelist
> Hay que enGNUrecerse, pero sin perder la terGNUra jamás - Che GNUevara