Feedback-ID: i56a14606:Fastmail
User-Agent: Cyrus-JMAP/3.9.0-alpha0-106-gfe3ab13a96-fm-20230124.001-gfe3ab13a
Mime-Version: 1.0
Message-Id: <131ceefe-a205-44e7-99be-275858b3a2df@app.fastmail.com>
In-Reply-To: <20230130171250.5ve3clop4tjv73ri@mobilestation>
References: <20230130130826.596171-1-arnd@kernel.org>
 <20230130171250.5ve3clop4tjv73ri@mobilestation>
Date:   Mon, 30 Jan 2023 19:50:37 +0100
From:   "Arnd Bergmann" <arnd@arndb.de>
To:     "Serge Semin" <fancer.lancer@gmail.com>,
        "Arnd Bergmann" <arnd@kernel.org>
Cc:     "Gustavo Pimentel" <gustavo.pimentel@synopsys.com>,
        "Vinod Koul" <vkoul@kernel.org>,
        "Serge Semin" <Sergey.Semin@baikalelectronics.ru>,
        "Bjorn Helgaas" <bhelgaas@google.com>,
        "Manivannan Sadhasivam" <mani@kernel.org>,
        "Lorenzo Pieralisi" <lpieralisi@kernel.org>,
        "Frank Li" <Frank.Li@nxp.com>, dmaengine@vger.kernel.org,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH] dmaengine: dw-edma: reduce stack usage after debugfs rework
Content-Type: text/plain
Precedence: bulk

On Mon, Jan 30, 2023, at 18:12, Serge Semin wrote:
> Hi Arnd
>
> On Mon, Jan 30, 2023 at 02:08:10PM +0100, Arnd Bergmann wrote:
>> From: Arnd Bergmann <arnd@arndb.de>
>> 
>> After the dw_edma_debugfs_entry arrays are no longer compile-time
>> constant, they take up space on the stack, which exceeds the
>> warning limit after inlining:
>> 
>> drivers/dma/dw-edma/dw-edma-v0-debugfs.c:280:6: error: stack frame size (1784) exceeds limit (1400) in 'dw_edma_v0_debugfs_on' [-Werror,-Wframe-larger-than]
>> void dw_edma_v0_debugfs_on(struct dw_edma *dw)
>> 
>
>> Work around this by marking the functions with the largest arrays
>> as 'noinline_for_stack' to make them not use up space on the same
>> stack together.
>
> You mean the same stack !frame!, right?

Right.
  
>> -static void dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
>> +static noinline_for_stack void
>> +dw_edma_debugfs_regs_ch(struct dw_edma *dw, enum dw_edma_dir dir,
>>  				    u16 ch, struct dentry *dent)
>
> This doesn't seem like required. Does it? This method is called from
> two functions: dw_edma_debugfs_regs_wr() and dw_edma_debugfs_regs_rd()
> and if I am not mistaken will unlikely be inlined. Even if compiler will
> inline it the stack memory consumption won't change much since the
> functions aren't called from each other.

I did this mainly for consistency, marking all the functions that
have an array of dw_edma_debugfs_entry. I have retested without this
change now and adapted the changelog accordingly.

>> @@ -159,7 +161,7 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
>>  		WR_REGISTER(dw, ch67_imwr_data),
>>  		WR_REGISTER(dw, linked_list_err_en),
>>  	};
>> -	static const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {
>> +	const struct dw_edma_debugfs_entry debugfs_unroll_regs[] = {
>
> I am confused how come these locals've turned to be static const?
> Neither my patches nor the Bjorn repo provide such changes:
> https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git/tree/drivers/dma/dw-edma/dw-edma-v0-debugfs.c?h=next#n139
> they were supposed to be just const.
>
> What repo is this patch based on?

My mistake, that was a revert from an earlier broken attempt to
fix the issue. It's fixed now after reverting the other patch
and rebasing.

>>  		/* eDMA channel context grouping */
>>  		WR_REGISTER_UNROLL(dw, engine_chgroup),
>>  		WR_REGISTER_UNROLL(dw, engine_hshake_cnt.lsb),
>> @@ -197,7 +199,8 @@ static void dw_edma_debugfs_regs_wr(struct dw_edma *dw, struct dentry *dent)
>>  	}
>>  }
>>  
>
>> -static void dw_edma_debugfs_regs_rd(struct dw_edma *dw, struct dentry *dent)
>> +static noinline void
>
> noinline_for_stack, right?

Fixed.

     Arnd