linux-kernel - Re: [PATCH v7 2/2] rust: add dma coherent allocator abstraction.

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-Id: <216C2A0A-8393-47DC-9A83-0020461512BF@collabora.com>
Date: Mon, 16 Dec 2024 08:00:24 -0300
From: Daniel Almeida <daniel.almeida@...labora.com>
To: Abdiel Janulgue <abdiel.janulgue@...il.com>
Cc: Robin Murphy <robin.murphy@....com>,
 Alice Ryhl <aliceryhl@...gle.com>,
 rust-for-linux@...r.kernel.org,
 Miguel Ojeda <ojeda@...nel.org>,
 Alex Gaynor <alex.gaynor@...il.com>,
 Boqun Feng <boqun.feng@...il.com>,
 Gary Guo <gary@...yguo.net>,
 Björn Roy Baron <bjorn3_gh@...tonmail.com>,
 Benno Lossin <benno.lossin@...ton.me>,
 Andreas Hindborg <a.hindborg@...nel.org>,
 Trevor Gross <tmgross@...ch.edu>,
 Danilo Krummrich <dakr@...nel.org>,
 Valentin Obst <kernel@...entinobst.de>,
 open list <linux-kernel@...r.kernel.org>,
 Christoph Hellwig <hch@....de>,
 Marek Szyprowski <m.szyprowski@...sung.com>,
 airlied@...hat.com,
 "open list:DMA MAPPING HELPERS" <iommu@...ts.linux.dev>
Subject: Re: [PATCH v7 2/2] rust: add dma coherent allocator abstraction.



> On 16 Dec 2024, at 07:23, Abdiel Janulgue <abdiel.janulgue@...il.com> wrote:
> 
> 
> 
> On 13/12/2024 21:08, Daniel Almeida wrote:
>> Hi Robin,
>>> On 13 Dec 2024, at 12:28, Robin Murphy <robin.murphy@....com> wrote:
>>> 
>>> On 13/12/2024 2:47 pm, Daniel Almeida wrote:
>>> [...]
>>>>>> +    /// Returns the CPU-addressable region as a slice.
>>>>>> +    pub fn cpu_buf(&self) -> &[T]
>>>>>> +    {
>>>>>> +        // SAFETY: The pointer is valid due to type invariant on `CoherentAllocation` and
>>>>>> +        // is valid for reads for `self.count * size_of::<T>` bytes.
>>>>>> +        unsafe { core::slice::from_raw_parts(self.cpu_addr, self.count) }
>>>>> 
>>>>> Immutable slices require that the data does not change while the
>>>>> reference is live. Is that the case? If so, your safety comment should
>>>>> explain that.
>>>>> 
>>>>>> +    }
>>>>>> +
>>>>>> +    /// Performs the same functionality as `cpu_buf`, except that a mutable slice is returned.
>>>>>> +    pub fn cpu_buf_mut(&mut self) -> &mut [T]
>>>>>> +    {
>>>>>> +        // SAFETY: The pointer is valid due to type invariant on `CoherentAllocation` and
>>>>>> +        // is valid for reads for `self.count * size_of::<T>` bytes.
>>>>>> +        unsafe { core::slice::from_raw_parts_mut(self.cpu_addr, self.count) }
>>>>> 
>>>>> Mutable slices require that the data is not written to *or read* by
>>>>> anybody else while the reference is live. Is that the case? If so,
>>>>> your safety comment should explain that.
>>>>> 
>>>> The buffer will probably be shared between the CPU and some hardware device, since this is the
>>>> point of the dma mapping API.
>>>> It’s up to the caller to ensure that no hardware operations that involve the buffer are currently taking
>>>> place while the slices above are alive.
>>> 
>>> Hmm, that sounds troublesome... the nature of coherent allocations is that both CPU and device may access them at any time, and you can definitely expect ringbuffer-style usage models where a CPU is writing to part of the buffer while the device is reading/writing another part, but also cases where a CPU needs to poll for a device write to a particular location.
>>> 
>> Ok, I had based my answer on some other drivers I’ve worked on in the past where the approach I cited would work.
>> I can see it not working for what you described, though.
>> This is a bit unfortunate, because it means we are back to square one, i.e.: back to read() and write() functions and
>> to the bound on `Copy`. That’s because, no matter how you try to dress this, there is no way to give safe and direct access
>> to the underlying memory if you can’t avoid situations where both the CPU and the device will be accessing the memory
>> at the same time.
> 
> This is unfortunate indeed. Thanks Alice for pointing out the limitations of slice.
> 
> Btw, do we have any other concerns in going back to plain old raw pointers instead? i.e.,
> 
>    pub fn read(&self, index: usize) -> Result<T> {
>        if index >= self.count {
>            return Err(EINVAL);
>        }
> 
>        let ptr = self.cpu_addr.wrapping_add(index);
>        // SAFETY: We just checked that the index is within bounds.
>        Ok(unsafe { ptr.read() })
>    }
> 
>    pub fn write(&self, index: usize, value: &T) -> Result
>    where
>        T: Copy,
>    {
>        if index >= self.count {
>            return Err(EINVAL);
>        }
> 
>        let ptr = self.cpu_addr.wrapping_add(index);
>        // SAFETY: We just checked that the index is within bounds.
>        unsafe { ptr.write(*value) };
>        Ok(())
>    }
> 
>> I guess the only improvement that could be made over the approach used for v2 is to at least use copy_nonoverlapping
>> instead, 
> 
> You mean introduce something like read_raw(dst: *mut u8,...) and write_raw(&self, src: *const u8,...)?


What I mean is that you don’t have to read (and bounds-check) a single index per function call.

Using copy_nonoverlapping lets you read/write multiple values with only a single function call and a
single check. It’s basically equivalent to a memcpy.


> 
> Regards,
> Abdiel