[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <Z4JvONDiaYkQLAv/@alpha.franken.de>
Date: Sat, 11 Jan 2025 14:16:40 +0100
From: Thomas Bogendoerfer <tsbogend@...ha.franken.de>
To: Mateusz Jończyk <mat.jonczyk@...pl>
Cc: linux-mips@...r.kernel.org, linux-kernel@...r.kernel.org,
stable@...r.kernel.org, Dengcheng Zhu <dzhu@...ecomp.com>,
Ming Wang <wangming01@...ngson.cn>,
Tiezhu Yang <yangtiezhu@...ngson.cn>
Subject: Re: [PATCH] mips/math-emu: fix emulation of the prefx instruction
On Sun, Jan 05, 2025 at 10:18:06PM +0100, Mateusz Jończyk wrote:
> Currently, installation of Debian 12.8 for mipsel fails on machines
> without an FPU [1]. This is caused by the fact that zstd (which is used
> for initramfs compression) executes the prefx instruction, which is not
> emulated properly by the kernel.
>
> The prefx (Prefetch Indexed) instruction fetches data from memory into
> the cache without any side effects. Though functionally unrelated, it
> requires an FPU [2].
>
> Bytecode format of this instruction ends on "001111" binary:
>
> (prefx instruction format) & 0x0000003f = 0x0000000f
>
> The code in fpux_emu() runs like so:
>
> #define MIPSInst(x) x
> #define MIPSInst_FMA_FFMT(x) (MIPSInst(x) & 0x00000007)
> #define MIPSInst_FUNC(x) (MIPSInst(x) & 0x0000003f)
> enum cop1x_func { ..., pfetch_op = 0x0f, ... };
>
> ...
>
> switch (MIPSInst_FMA_FFMT(ir)) {
> ...
>
> case 0x3:
> if (MIPSInst_FUNC(ir) != pfetch_op)
> return SIGILL;
>
> /* ignore prefx operation */
> break;
>
> default:
> return SIGILL;
> }
>
> That snippet above contains a logic error and the
> if (MIPSInst_FUNC(ir) != pfetch_op)
> comparison always fires.
>
> When MIPSInst_FUNC(ir) is equal to pfetch_op, ir must end on 001111
> binary. In this case, MIPSInst_FMA_FFMT(ir) must be equal to 0x7, which
> does not match that case label.
>
> This causes emulation failure for the prefx instruction. Fix it.
>
> This has been broken by
> commit 919af8b96c89 ("MIPS: Make definitions of MIPSInst_FMA_{FUNC,FMTM} consistent with MIPS64 manual")
> which modified the MIPSInst_FMA_FFMT macro without updating the users.
>
> Signed-off-by: Mateusz Jończyk <mat.jonczyk@...pl>
> Cc: stable@...r.kernel.org # after 3 weeks
> Cc: Dengcheng Zhu <dzhu@...ecomp.com>
> Cc: Thomas Bogendoerfer <tsbogend@...ha.franken.de>
> Cc: Ming Wang <wangming01@...ngson.cn>
> Cc: Tiezhu Yang <yangtiezhu@...ngson.cn>
> Fixes: 919af8b96c89 ("MIPS: Make definitions of MIPSInst_FMA_{FUNC,FMTM} consistent with MIPS64 manual")
>
> [1] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1091858
> [2] MIPS Architecture For Programmers Volume II-A: The MIPS32 Instruction Set
>
> ---
>
> Tested in QEMU for mipsel and mips64el.
> ---
> arch/mips/math-emu/cp1emu.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/arch/mips/math-emu/cp1emu.c b/arch/mips/math-emu/cp1emu.c
> index 265bc57819df..c89e70df43d8 100644
> --- a/arch/mips/math-emu/cp1emu.c
> +++ b/arch/mips/math-emu/cp1emu.c
> @@ -1660,7 +1660,7 @@ static int fpux_emu(struct pt_regs *xcp, struct mips_fpu_struct *ctx,
> break;
> }
>
> - case 0x3:
> + case 0x7:
> if (MIPSInst_FUNC(ir) != pfetch_op)
> return SIGILL;
>
> --
> 2.25.1
applied to mips-next.
Thomas.
--
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea. [ RFC1925, 2.3 ]
Powered by blists - more mailing lists