Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Thu, 29 Jun 2023 02:28:44 -0700
From:   Sumitra Sharma <sumitraartsy@gmail.com>
To:     Matthew Wilcox <willy@infradead.org>
Cc:     Hans de Goede <hdegoede@redhat.com>, linux-fsdevel@vger.kernel.org,
        linux-kernel@vger.kernel.org, Ira Weiny <ira.weiny@intel.com>,
        Fabio <fmdefrancesco@gmail.com>, Deepak R Varma <drv@mailo.com>,
        Sumitra Sharma <sumitraartsy@gmail.com>
Subject: Re: [PATCH] fs/vboxsf: Replace kmap() with kmap_local_{page, folio}()
Message-ID: <20230629092844.GA456505@sumitra.com>
References: <20230627135115.GA452832@sumitra.com>
 <ZJxqmEVKoxxftfXM@casper.infradead.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: <ZJxqmEVKoxxftfXM@casper.infradead.org>
Precedence: bulk

On Wed, Jun 28, 2023 at 06:15:04PM +0100, Matthew Wilcox wrote:
> Here's a more comprehensive read_folio patch.  It's not at all
> efficient, but then if we wanted an efficient vboxsf, we'd implement
> vboxsf_readahead() and actually do an async call with deferred setting
> of the uptodate flag.  I can consult with anyone who wants to do all
> this work.
> 
> I haven't even compiled this, just trying to show the direction this
> should take.
> 
> diff --git a/fs/vboxsf/file.c b/fs/vboxsf/file.c
> index 2307f8037efc..f1af9a7bd3d8 100644
> --- a/fs/vboxsf/file.c
> +++ b/fs/vboxsf/file.c
> @@ -227,26 +227,31 @@ const struct inode_operations vboxsf_reg_iops = {
>  
>  static int vboxsf_read_folio(struct file *file, struct folio *folio)
>  {
> -	struct page *page = &folio->page;
>  	struct vboxsf_handle *sf_handle = file->private_data;
> -	loff_t off = page_offset(page);
> -	u32 nread = PAGE_SIZE;
> -	u8 *buf;
> +	loff_t pos = folio_pos(folio);
> +	size_t offset = 0;
>  	int err;
>  
> -	buf = kmap(page);
> +	do {
> +		u8 *buf = kmap_local_folio(folio, offset);
> +		u32 nread = PAGE_SIZE;
>  
> -	err = vboxsf_read(sf_handle->root, sf_handle->handle, off, &nread, buf);
> -	if (err == 0) {
> -		memset(&buf[nread], 0, PAGE_SIZE - nread);
> -		flush_dcache_page(page);
> -		SetPageUptodate(page);
> -	} else {
> -		SetPageError(page);
> -	}
> +		err = vboxsf_read(sf_handle->root, sf_handle->handle, pos,
> +				&nread, buf);
> +		if (nread < PAGE_SIZE)
> +			memset(&buf[nread], 0, PAGE_SIZE - nread);
> +		kunmap_local(buf);
> +		if (err)
> +			break;
> +		offset += PAGE_SIZE;
> +		pos += PAGE_SIZE;
> +	} while (offset < folio_size(folio);
>  
> -	kunmap(page);
> -	unlock_page(page);
> +	if (!err) {
> +		flush_dcache_folio(folio);
> +		folio_mark_uptodate(folio);
> +	}
> +	folio_unlock(folio);
>  	return err;
>  }
>

Hi 

So, after reading the comments, I understood that the problem presented 
by Hans and Matthew is as follows:

1) In the current code, the buffers used by vboxsf_write()/vboxsf_read() are 
translated to PAGELIST-s before passing to the hypervisor, 
but inefficiently— it first maps a page in vboxsf_read_folio() and then 
calls page_to_phys(virt_to_page()) in the function hgcm_call_init_linaddr(). 

The inefficiency in the current implementation arises due to the unnecessary 
mapping of a page in vboxsf_read_folio() because the mapping output, i.e. the 
linear address, is used deep down in file 'drivers/virt/vboxguest/vboxguest_utils.c'. 
Hence, the mapping must be done in this file; to do so, the folio must be passed 
until this point. It can be done by adding a new member, 'struct folio *folio', 
in the 'struct vmmdev_hgcm_function_parameter64'. 

The unused member 'phys_addr' in this struct can also be removed.

2) Expanding the vboxsf_read_folio so that it can handle larger folios.
Matthew already has suggested the changes, I have to read more on this.

Parallelly I will be setting up the testing environment to test the changes. 


Please let me know if I am wrong anywhere.

Thanks & regards
Sumitra