Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Tue, 27 Feb 2018 02:04:49 +0200
From:   "Michael S. Tsirkin" <mst@redhat.com>
To:     =?iso-8859-1?Q?Marc-Andr=E9?= Lureau <marcandre.lureau@redhat.com>
Cc:     linux-kernel@vger.kernel.org, slp@redhat.com, bhe@redhat.com,
        somlo@cmu.edu, xiaolong.ye@intel.com
Subject: Re: [PATCH v15 11/11] RFC: fw_cfg: do DMA read operation
Message-ID: <20180227020104-mutt-send-email-mst@kernel.org>
References: <20180215213312.29234-1-marcandre.lureau@redhat.com>
 <20180215213312.29234-12-marcandre.lureau@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <20180215213312.29234-12-marcandre.lureau@redhat.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Feb 15, 2018 at 10:33:12PM +0100, Marc-Andr? Lureau wrote:
> Modify fw_cfg_read_blob() to use DMA if the device supports it.
> Return errors, because the operation may fail.
> 
> So far, only one call in fw_cfg_register_dir_entries() is using
> kmalloc'ed buf and is thus clearly eligible to DMA read.
> 
> Initially, I didn't implement DMA read to speed up boot time, but as a
> first step before introducing DMA write (since read operations were
> already presents). Even more, I didn't realize fw-cfg entries were
> being read by the kernel during boot by default. But actally fw-cfg
> entries are being populated during module probe. I knew DMA improved a
> lot bios boot time (the main reason the DMA interface was added
> afaik). Let see the time it would take to read the whole ACPI
> tables (128kb allocated)
> 
>  # time cat /sys/firmware/qemu_fw_cfg/by_name/etc/acpi/tables/raw
>   - with DMA: sys 0m0.003s
>   - without DMA (-global fw_cfg.dma_enabled=off): sys 0m7.674s
> 
> FW_CFG_FILE_DIR (0x19) is the only "file" that is read during kernel
> boot to populate sysfs qemu_fw_cfg directory, and it is quite
> small (1-2kb). Since it does not expose itself, in order to measure
> the time it takes to read such small file, I took a comparable sized
> file of 2048 bytes and exposed it (-fw_cfg test,file=file with a
> modified read_raw enabling DMA)
> 
>  # perf stat -r 100 cat /sys/firmware/qemu_fw_cfg/by_name/test/raw >/dev/null
>   - with DMA:
>           0.636037      task-clock (msec)         #    0.141 CPUs utilized            ( +-  1.19% )
>   - without DMA:
>           6.430128      task-clock (msec)         #    0.622 CPUs utilized            ( +-  0.22% )
> 
> That's a few msec saved during boot by enabling DMA read (the gain
> would be more substantial if other & bigger fw-cfg entries are read by
> others from sysfs, unfortunately, it's not clear if we can always
> enable DMA there)
> 
> Signed-off-by: Marc-Andr? Lureau <marcandre.lureau@redhat.com>
> ---
>  drivers/firmware/qemu_fw_cfg.c | 61 ++++++++++++++++++++++++++++++++++--------
>  1 file changed, 50 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/firmware/qemu_fw_cfg.c b/drivers/firmware/qemu_fw_cfg.c
> index 3015e77aebca..94df57e9be66 100644
> --- a/drivers/firmware/qemu_fw_cfg.c
> +++ b/drivers/firmware/qemu_fw_cfg.c
> @@ -124,12 +124,47 @@ static ssize_t fw_cfg_dma_transfer(void *address, u32 length, u32 control)
>  	return ret;
>  }
>  
> +/* with acpi & dev locks taken */
> +static ssize_t fw_cfg_read_blob_dma(u16 key,
> +				void *buf, loff_t pos, size_t count)
> +{
> +	ssize_t ret;
> +
> +	if (pos == 0) {
> +		ret = fw_cfg_dma_transfer(buf, count, key << 16
> +					| FW_CFG_DMA_CTL_SELECT
> +					| FW_CFG_DMA_CTL_READ);
> +	} else {
> +		fw_cfg_sel_endianness(key);
> +		ret = fw_cfg_dma_transfer(NULL, pos, FW_CFG_DMA_CTL_SKIP);
> +		if (ret < 0)
> +			return ret;
> +		ret = fw_cfg_dma_transfer(buf, count,
> +					FW_CFG_DMA_CTL_READ);
> +	}
> +
> +	return ret;
> +}
> +
> +/* with acpi & dev locks taken */
> +static ssize_t fw_cfg_read_blob_io(u16 key,
> +				void *buf, loff_t pos, size_t count)
> +{
> +	fw_cfg_sel_endianness(key);
> +	while (pos-- > 0)
> +		ioread8(fw_cfg_reg_data);
> +	ioread8_rep(fw_cfg_reg_data, buf, count);
> +	return count;
> +}
> +
>  /* read chunk of given fw_cfg blob (caller responsible for sanity-check) */
>  static ssize_t fw_cfg_read_blob(u16 key,
> -				void *buf, loff_t pos, size_t count)
> +				void *buf, loff_t pos, size_t count,
> +				bool dma)
>  {
>  	u32 glk = -1U;
>  	acpi_status status;
> +	ssize_t ret;
>  
>  	/* If we have ACPI, ensure mutual exclusion against any potential
>  	 * device access by the firmware, e.g. via AML methods:

so this adds a dma flag to fw_cfg_read_blob.


> @@ -143,14 +178,17 @@ static ssize_t fw_cfg_read_blob(u16 key,
>  	}
>  
>  	mutex_lock(&fw_cfg_dev_lock);
> -	fw_cfg_sel_endianness(key);
> -	while (pos-- > 0)
> -		ioread8(fw_cfg_reg_data);
> -	ioread8_rep(fw_cfg_reg_data, buf, count);
> +	if (dma && fw_cfg_dma_enabled()) {
> +		ret = fw_cfg_read_blob_dma(key, buf, pos, count);
> +	} else {
> +		ret = fw_cfg_read_blob_io(key, buf, pos, count);
> +	}
> +
>  	mutex_unlock(&fw_cfg_dev_lock);
>  
>  	acpi_release_global_lock(glk);
> -	return count;
> +
> +	return ret;
>  }
>  
>  #ifdef CONFIG_CRASH_CORE

If set to false it does io, if set to true it does dma.

I would prefer passing an accessor function pointer
since that's clearer than true/false.


> @@ -284,7 +322,7 @@ static int fw_cfg_do_platform_probe(struct platform_device *pdev)
>  
>  	/* verify fw_cfg device signature */
>  	if (fw_cfg_read_blob(FW_CFG_SIGNATURE, sig,
> -				0, FW_CFG_SIG_SIZE) < 0 ||
> +				0, FW_CFG_SIG_SIZE, false) < 0 ||
>  		memcmp(sig, "QEMU", FW_CFG_SIG_SIZE) != 0) {
>  		fw_cfg_io_cleanup();
>  		return -ENODEV;
> @@ -468,7 +506,8 @@ static ssize_t fw_cfg_sysfs_read_raw(struct file *filp, struct kobject *kobj,
>  	if (count > entry->size - pos)
>  		count = entry->size - pos;
>  
> -	return fw_cfg_read_blob(entry->select, buf, pos, count);
> +	/* do not use DMA, virt_to_phys(buf) might not be ok */
> +	return fw_cfg_read_blob(entry->select, buf, pos, count, false);
>  }
>  
>  static struct bin_attribute fw_cfg_sysfs_attr_raw = {
> @@ -634,7 +673,7 @@ static int fw_cfg_register_dir_entries(void)
>  	size_t dir_size;
>  
>  	ret = fw_cfg_read_blob(FW_CFG_FILE_DIR, &files_count,
> -			0, sizeof(files_count));
> +			0, sizeof(files_count), false);
>  	if (ret < 0)
>  		return ret;
>  
> @@ -646,7 +685,7 @@ static int fw_cfg_register_dir_entries(void)
>  		return -ENOMEM;
>  
>  	ret = fw_cfg_read_blob(FW_CFG_FILE_DIR, dir,
> -			sizeof(files_count), dir_size);
> +			sizeof(files_count), dir_size, false);
>  	if (ret < 0)
>  		goto end;
>  
> @@ -697,7 +736,7 @@ static int fw_cfg_sysfs_probe(struct platform_device *pdev)
>  		goto err_probe;
>  
>  	/* get revision number, add matching top-level attribute */
> -	err = fw_cfg_read_blob(FW_CFG_ID, &rev, 0, sizeof(rev));
> +	err = fw_cfg_read_blob(FW_CFG_ID, &rev, 0, sizeof(rev), false);
>  	if (err < 0)
>  		goto err_probe;


Looks like all callers pass in false as parameter.
Given this, how can this speed up any operations?

Are you sure you tested this properly?

> -- 
> 2.16.1.73.g5832b7e9f2