Re: [PATCH V2] LoongArch: Add unaligned access support

From: WANG Xuerui <kernel@xen0n.name>
To: Huacai Chen <chenhuacai@kernel.org>,
	David Laight <David.Laight@aculab.com>
Cc: Huacai Chen <chenhuacai@loongson.cn>,
	"loongarch@lists.linux.dev" <loongarch@lists.linux.dev>,
	Xuefeng Li <lixuefeng@loongson.cn>,
	Tiezhu Yang <yangtiezhu@loongson.cn>, Guo Ren <guoren@kernel.org>,
	Jiaxun Yang <jiaxun.yang@flygoat.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH V2] LoongArch: Add unaligned access support
Date: Tue, 18 Oct 2022 11:29:39 +0800	[thread overview]
Message-ID: <c34a277a-1735-65e6-c97d-fcc2ac3d57e6@xen0n.name> (raw)
In-Reply-To: <CAAhV-H5WM97+Um9ULjv7gT3i5pJ3GfQ8ex6NRNr6uzmMgdVY_Q@mail.gmail.com>

On 2022/10/18 10:24, Huacai Chen wrote:
> Hi, David,
> 
> On Mon, Oct 17, 2022 at 8:58 PM David Laight <David.Laight@aculab.com> wrote:
>>
>> From: Huacai Chen
>>> Sent: 17 October 2022 03:24
>>>
>>> Loongson-2 series (Loongson-2K500, Loongson-2K1000) don't support
>>> unaligned access in hardware, while Loongson-3 series (Loongson-3A5000,
>>> Loongson-3C5000) are configurable whether support unaligned access in
>>> hardware. This patch add unaligned access emulation for those LoongArch
>>> processors without hardware support.
>>>
>> ...
>>> +     /*
>>> +      * This load never faults.
>>> +      */
>>> +     __get_inst(&insn.word, pc, user);
>>
>> On what basis does it never fault?
>> Any user access can fault.
>> If nothing else another thread of the process can unmap
>> the page.
> Yes, this can happen, since __get_inst() handles fault, we can just
> remove the comment.
> 
>>
>>> +     if (user && !access_ok(addr, 8))
>>> +             goto sigbus;
>>
>> Surely that is technically wrong - a two or four byte
>> access is valid right at the end of valid user addreeses.
> Yes, this check should be moved to each case.
> 
>>
>>> +
>>> +     if (insn.reg2i12_format.opcode == ldd_op ||
>>> +             insn.reg2i14_format.opcode == ldptrd_op ||
>>> +             insn.reg3_format.opcode == ldxd_op) {
>>> +             res = unaligned_read(addr, &value, 8, 1);
>>
>> That is the most horrid indentation of long lines I've
>> ever seen.
>> I'd also guess you can common up some of this code
>> by looking at the instruction field that include the
>> transfer width.
>>
>> The long elsif list will generate horrid code.
>> But maybe since you've just taken a fault it really
>> doesn't matter.
>> Indeed just emulating in C using byte accesses
>> it probably fine.
> I want to keep the assembly, because we can use more efficient methods
> with the upcoming alternative mechanism.

What about my more structured approach in another reply that avoids the 
huge else-if conditions? Both the terrible line wraps and codegen could 
be avoided.

-- 
WANG "xen0n" Xuerui

Linux/LoongArch mailing list: https://lore.kernel.org/loongarch/