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;
Message-ID: <32fb03bf-be43-d416-4a32-b30a0c339496@redhat.com>
Date:   Tue, 27 Jun 2023 20:19:59 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.12.0
Subject: Re: [PATCH] fs/vboxsf: Replace kmap() with kmap_local_{page, folio}()
From:   Hans de Goede <hdegoede@redhat.com>
To:     Matthew Wilcox <willy@infradead.org>
Cc:     Sumitra Sharma <sumitraartsy@gmail.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>
References: <20230627135115.GA452832@sumitra.com>
 <6a566e51-6288-f782-2fa5-f9b0349b6d7c@redhat.com>
 <ZJsgWQb+tOqtQuKL@casper.infradead.org>
 <3baed0e0-db2c-906c-5256-1d83d59794e9@redhat.com>
Content-Language: en-US, nl
In-Reply-To: <3baed0e0-db2c-906c-5256-1d83d59794e9@redhat.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi,

On 6/27/23 20:10, Hans de Goede wrote:
> Hi Matthew,
> 
> On 6/27/23 19:46, Matthew Wilcox wrote:
>> On Tue, Jun 27, 2023 at 04:34:51PM +0200, Hans de Goede wrote:
>>> Hi,
>>>
>>> On 6/27/23 15:51, Sumitra Sharma wrote:
>>>> kmap() has been deprecated in favor of the kmap_local_page() due to high
>>>> cost, restricted mapping space, the overhead of a global lock for
>>>> synchronization, and making the process sleep in the absence of free
>>>> slots.
>>>>
>>>> kmap_local_{page, folio}() is faster than kmap() and offers thread-local
>>>> and CPU-local mappings, can take pagefaults in a local kmap region and
>>>> preserves preemption by saving the mappings of outgoing tasks and
>>>> restoring those of the incoming one during a context switch.
>>>>
>>>> The difference between kmap_local_page() and kmap_local_folio() consist
>>>> only in the first taking a pointer to a page and the second taking two
>>>> arguments, a pointer to a folio and the byte offset within the folio which
>>>> identifies the page.
>>>>
>>>> The mappings are kept thread local in the functions 'vboxsf_read_folio',
>>>> 'vboxsf_writepage', 'vboxsf_write_end' in file.c
>>>>
>>>> Suggested-by: Ira Weiny <ira.weiny@intel.com>
>>>> Signed-off-by: Sumitra Sharma <sumitraartsy@gmail.com>
>>>
>>> Thanks, patch looks good to me:
>>
>> It doesn't look great to me, tbh.  It's generally an antipattern to map
>> the page/folio up at the top and then pass the virtual address down to
>> the bottom.  Usually we want to work in terms of physical addresses
>> as long as possible.  I see the vmmdev_hgcm_function_parameter can
>> take physical addresses; does it work to simply use the phys_addr
>> instead of the linear_addr?  I see this commentary:
>>
>>        /** Deprecated Doesn't work, use PAGELIST. */
>>         VMMDEV_HGCM_PARM_TYPE_PHYSADDR           = 3,
>>
>> so, um, can we use
>>         /** Physical addresses of locked pages for a buffer. */
>>         VMMDEV_HGCM_PARM_TYPE_PAGELIST           = 10,
>>
>> and convert vboxsf_read_folio() to pass the folio down to vboxsf_read()
>> which converts it to a PAGELIST (however one does that)?
> 
> 
> It has been a long time since I looked at this code in detail. I don't
> think you can just use different types when making virtualbox hypervisor
> calls and then expect the hypervisor to say sure that another way to
> represent a memory buffer, I'll take that instead.

Ok correction to this, drivers/virt/vboxguest/vboxguest_utils.c actually
already translates the VMMDEV_HGCM_PARM_TYPE_LINADDR_KERNEL_IN /
VMMDEV_HGCM_PARM_TYPE_LINADDR_KERNEL_OUT buffers used by vboxsf_write()/
vboxsf_read() to PAGELIST-s before passing them to the hypervisor
using page_to_phys(virt_to_page()) so we map a page and then call
page_to_phys(virt_to_page()) on it which indeed is quite inefficient.

That still leaves the problem that I have very little time to look into
this though ...

Regards,

Hans