diff mbox series

[v6,22/26] rust: alloc: implement `Cmalloc` in module allocator_test

Message ID 20240816001216.26575-23-dakr@kernel.org (mailing list archive)
State New
Headers show
Series Generic `Allocator` support for Rust | expand

Commit Message

Danilo Krummrich Aug. 16, 2024, 12:11 a.m. UTC
So far the kernel's `Box` and `Vec` types can't be used by userspace
test cases, since all users of those types (e.g. `CString`) use kernel
allocators for instantiation.

In order to allow userspace test cases to make use of such types as
well, implement the `Cmalloc` allocator within the allocator_test module
and type alias all kernel allocators to `Cmalloc`. The `Cmalloc`
allocator uses libc's realloc() function as allocator backend.

Signed-off-by: Danilo Krummrich <dakr@kernel.org>
---
 rust/kernel/alloc/allocator_test.rs | 178 ++++++++++++++++++++++++++--
 1 file changed, 171 insertions(+), 7 deletions(-)

Comments

Benno Lossin Aug. 29, 2024, 7:14 p.m. UTC | #1
On 16.08.24 02:11, Danilo Krummrich wrote:
> So far the kernel's `Box` and `Vec` types can't be used by userspace
> test cases, since all users of those types (e.g. `CString`) use kernel
> allocators for instantiation.
> 
> In order to allow userspace test cases to make use of such types as
> well, implement the `Cmalloc` allocator within the allocator_test module
> and type alias all kernel allocators to `Cmalloc`. The `Cmalloc`
> allocator uses libc's realloc() function as allocator backend.
> 
> Signed-off-by: Danilo Krummrich <dakr@kernel.org>
> ---
>  rust/kernel/alloc/allocator_test.rs | 178 ++++++++++++++++++++++++++--
>  1 file changed, 171 insertions(+), 7 deletions(-)
> 
> diff --git a/rust/kernel/alloc/allocator_test.rs b/rust/kernel/alloc/allocator_test.rs
> index 1b2642c547ec..7fff308d02dc 100644
> --- a/rust/kernel/alloc/allocator_test.rs
> +++ b/rust/kernel/alloc/allocator_test.rs
> @@ -2,20 +2,184 @@
> 

Could add a short paragraph as the module description why this module
exists? Would probably be enough to paste the commit message.

>  #![allow(missing_docs)]
> 
> -use super::{AllocError, Allocator, Flags};
> +use super::{flags::*, AllocError, Allocator, Flags};
>  use core::alloc::Layout;
> +use core::cmp;
> +use core::mem;
> +use core::ptr;
>  use core::ptr::NonNull;
> 
> -pub struct Kmalloc;
> +pub struct Cmalloc;
> +pub type Kmalloc = Cmalloc;
>  pub type Vmalloc = Kmalloc;
>  pub type KVmalloc = Kmalloc;
> 
> -unsafe impl Allocator for Kmalloc {
> +extern "C" {
> +    #[link_name = "aligned_alloc"]
> +    fn libc_aligned_alloc(align: usize, size: usize) -> *mut core::ffi::c_void;
> +
> +    #[link_name = "free"]
> +    fn libc_free(ptr: *mut core::ffi::c_void);
> +}
> +
> +struct CmallocData {
> +    // The actual size as requested through `Cmalloc::alloc` or `Cmalloc::realloc`.
> +    size: usize,
> +    // The offset from the pointer returned to the caller of `Cmalloc::alloc` or `Cmalloc::realloc`
> +    // to the actual base address of the allocation.
> +    offset: usize,
> +}
> +
> +impl Cmalloc {
> +    /// Adjust the size and alignment such that we can additionally store `CmallocData` right
> +    /// before the actual data described by `layout`.
> +    ///
> +    /// Example:
> +    ///
> +    /// For `CmallocData` assume an alignment of 8 and a size of 16.
> +    /// For `layout` assume and alignment of 16 and a size of 64.

This looks like you want it rendered as bulletpoints (but it won't).

> +    ///
> +    /// 0                16               32                                               96
> +    /// |----------------|----------------|------------------------------------------------|
> +    ///        empty         CmallocData                         data

Can you put this inside of '```'? Then it will render nicely in markdown
(don't forget to specify the type 'text')

> +    ///
> +    /// For this example the returned `Layout` has an alignment of 32 and a size of 96.
> +    fn layout_adjust(layout: Layout) -> Result<Layout, AllocError> {
> +        let layout = layout.pad_to_align();
> +
> +        // Ensure that `CmallocData` fits into half the alignment. Additionally, this guarantees
> +        // that advancing a pointer aligned to `align` by `align / 2` we still satisfy or exceed
> +        // the alignment requested through `layout`.
> +        let align = cmp::max(
> +            layout.align(),
> +            mem::size_of::<CmallocData>().next_power_of_two(),
> +        ) * 2;
> +
> +        // Add the additional space required for `CmallocData`.
> +        let size = layout.size() + mem::size_of::<CmallocData>();
> +
> +        Ok(Layout::from_size_align(size, align)
> +            .map_err(|_| AllocError)?
> +            .pad_to_align())
> +    }
> +
> +    fn alloc_store_data(layout: Layout) -> Result<NonNull<u8>, AllocError> {
> +        let requested_size = layout.size();
> +
> +        let layout = Self::layout_adjust(layout)?;
> +        let min_align = layout.align() / 2;
> +
> +        // SAFETY: Returns either NULL or a pointer to a memory allocation that satisfies or
> +        // exceeds the given size and alignment requirements.
> +        let raw_ptr = unsafe { libc_aligned_alloc(layout.align(), layout.size()) } as *mut u8;
> +
> +        let priv_ptr = NonNull::new(raw_ptr).ok_or(AllocError)?;
> +
> +        // SAFETY: Advance the pointer by `min_align`. The adjustments from `Self::layout_adjust`
> +        // ensure that after this operation the original size and alignment requirements are still
> +        // satisfied or exceeded.

This SAFETY comment should address why it's OK to call `add`. You
justify something different, namely why the allocation still satisfies
the requirements of `layout`. That is something that this function
should probably guarantee.

> +        let ptr = unsafe { priv_ptr.as_ptr().add(min_align) };
> +
> +        // SAFETY: `min_align` is greater than or equal to the size of `CmallocData`, hence we
> +        // don't exceed the allocation boundaries.
> +        let data_ptr: *mut CmallocData = unsafe { ptr.sub(mem::size_of::<CmallocData>()) }.cast();
> +
> +        let data = CmallocData {
> +            size: requested_size,
> +            offset: min_align,
> +        };
> +
> +        // SAFETY: `data_ptr` is properly aligned and within the allocation boundaries reserved for
> +        // `CmallocData`.
> +        unsafe { data_ptr.write(data) };
> +
> +        NonNull::new(ptr).ok_or(AllocError)
> +    }
> +
> +    /// # Safety
> +    ///
> +    /// `ptr` must have been previously allocated with `Self::alloc_store_data`.

You additionally need that you have shared access to the pointee.

> +    unsafe fn data<'a>(ptr: NonNull<u8>) -> &'a CmallocData {
> +        // SAFETY: `Self::alloc_store_data` stores the `CmallocData` right before the address
> +        // returned to callers of `Self::alloc_store_data`.
> +        let data_ptr: *mut CmallocData =
> +            unsafe { ptr.as_ptr().sub(mem::size_of::<CmallocData>()) }.cast();
> +
> +        // SAFETY: The `CmallocData` has been previously stored at this offset with
> +        // `Self::alloc_store_data`.
> +        unsafe { &*data_ptr }
> +    }
> +
> +    /// # Safety
> +    ///
> +    /// This function must not be called more than once for the same allocation.
> +    ///
> +    /// `ptr` must have been previously allocated with `Self::alloc_store_data`.

You additionally need that you have exclusive access to the pointee.

> +    unsafe fn free_read_data(ptr: NonNull<u8>) {
> +        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
> +        let data = unsafe { Self::data(ptr) };
> +
> +        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
> +        let priv_ptr = unsafe { ptr.as_ptr().sub(data.offset) };
> +
> +        // SAFETY: `priv_ptr` has previously been allocatored with this `Allocator`.
> +        unsafe { libc_free(priv_ptr.cast()) };
> +    }
> +}
> +
> +unsafe impl Allocator for Cmalloc {
> +    fn alloc(layout: Layout, flags: Flags) -> Result<NonNull<[u8]>, AllocError> {
> +        if layout.size() == 0 {
> +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
> +        }
> +
> +        let ptr = Self::alloc_store_data(layout)?;
> +
> +        if flags.contains(__GFP_ZERO) {
> +            // SAFETY: `Self::alloc_store_data` guarantees that `ptr` points to memory of at least
> +            // `layout.size()` bytes.
> +            unsafe { ptr.as_ptr().write_bytes(0, layout.size()) };
> +        }

This makes me wonder, what other flags should we handle for this
allocator?

> +
> +        Ok(NonNull::slice_from_raw_parts(ptr, layout.size()))
> +    }
> +
>      unsafe fn realloc(
> -        _ptr: Option<NonNull<u8>>,
> -        _layout: Layout,
> -        _flags: Flags,
> +        ptr: Option<NonNull<u8>>,
> +        layout: Layout,
> +        flags: Flags,
>      ) -> Result<NonNull<[u8]>, AllocError> {
> -        panic!();
> +        let src: NonNull<u8> = if let Some(src) = ptr {
> +            src.cast()

Why the cast?

> +        } else {
> +            return Self::alloc(layout, flags);
> +        };

You should be able to write this instead:

    let Some(src) = ptr else {
        return Self::alloc(layout, flags);
    };

> +
> +        if layout.size() == 0 {
> +            // SAFETY: `src` has been created by `Self::alloc_store_data`.

This is not true, consider:

    let ptr = alloc(size = 0);
    free(ptr)

Alloc will return a dangling pointer due to the first if statement and
then this function will pass it to `free_read_data`, even though it
wasn't created by `alloc_store_data`.
This isn't forbidden by the `Allocator` trait function's safety
requirements.

> +            unsafe { Self::free_read_data(src) };
> +
> +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
> +        }
> +
> +        let dst = Self::alloc(layout, flags)?;
> +
> +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
> +        let data = unsafe { Self::data(src) };

Same issue here, if the allocation passed in is zero size. I think you
have no other choice than to allocate even for zero size requests...
Otherwise how would you know that they are zero-sized.

---
Cheers,
Benno

> +
> +        // SAFETY: `src` has previously been allocated with this `Allocator`; `dst` has just been
> +        // newly allocated. Copy up to the smaller of both sizes.
> +        unsafe {
> +            ptr::copy_nonoverlapping(
> +                src.as_ptr(),
> +                dst.as_ptr().cast(),
> +                cmp::min(layout.size(), data.size),
> +            )
> +        };
> +
> +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
> +        unsafe { Self::free_read_data(src) };
> +
> +        Ok(dst)
>      }
>  }
> --
> 2.46.0
>
Danilo Krummrich Aug. 29, 2024, 10:25 p.m. UTC | #2
On Thu, Aug 29, 2024 at 07:14:18PM +0000, Benno Lossin wrote:
> On 16.08.24 02:11, Danilo Krummrich wrote:
> > So far the kernel's `Box` and `Vec` types can't be used by userspace
> > test cases, since all users of those types (e.g. `CString`) use kernel
> > allocators for instantiation.
> > 
> > In order to allow userspace test cases to make use of such types as
> > well, implement the `Cmalloc` allocator within the allocator_test module
> > and type alias all kernel allocators to `Cmalloc`. The `Cmalloc`
> > allocator uses libc's realloc() function as allocator backend.
> > 
> > Signed-off-by: Danilo Krummrich <dakr@kernel.org>
> > ---
> >  rust/kernel/alloc/allocator_test.rs | 178 ++++++++++++++++++++++++++--
> >  1 file changed, 171 insertions(+), 7 deletions(-)
> > 
> > diff --git a/rust/kernel/alloc/allocator_test.rs b/rust/kernel/alloc/allocator_test.rs
> > index 1b2642c547ec..7fff308d02dc 100644
> > --- a/rust/kernel/alloc/allocator_test.rs
> > +++ b/rust/kernel/alloc/allocator_test.rs
> > @@ -2,20 +2,184 @@
> > 
> 
> Could add a short paragraph as the module description why this module
> exists? Would probably be enough to paste the commit message.

Yes, sounds good.

> 
> >  #![allow(missing_docs)]
> > 
> > -use super::{AllocError, Allocator, Flags};
> > +use super::{flags::*, AllocError, Allocator, Flags};
> >  use core::alloc::Layout;
> > +use core::cmp;
> > +use core::mem;
> > +use core::ptr;
> >  use core::ptr::NonNull;
> > 
> > -pub struct Kmalloc;
> > +pub struct Cmalloc;
> > +pub type Kmalloc = Cmalloc;
> >  pub type Vmalloc = Kmalloc;
> >  pub type KVmalloc = Kmalloc;
> > 
> > -unsafe impl Allocator for Kmalloc {
> > +extern "C" {
> > +    #[link_name = "aligned_alloc"]
> > +    fn libc_aligned_alloc(align: usize, size: usize) -> *mut core::ffi::c_void;
> > +
> > +    #[link_name = "free"]
> > +    fn libc_free(ptr: *mut core::ffi::c_void);
> > +}
> > +
> > +struct CmallocData {
> > +    // The actual size as requested through `Cmalloc::alloc` or `Cmalloc::realloc`.
> > +    size: usize,
> > +    // The offset from the pointer returned to the caller of `Cmalloc::alloc` or `Cmalloc::realloc`
> > +    // to the actual base address of the allocation.
> > +    offset: usize,
> > +}
> > +
> > +impl Cmalloc {
> > +    /// Adjust the size and alignment such that we can additionally store `CmallocData` right
> > +    /// before the actual data described by `layout`.
> > +    ///
> > +    /// Example:
> > +    ///
> > +    /// For `CmallocData` assume an alignment of 8 and a size of 16.
> > +    /// For `layout` assume and alignment of 16 and a size of 64.
> 
> This looks like you want it rendered as bulletpoints (but it won't).

Actually, that wasn't my intention, but I'm fine changing that.

> 
> > +    ///
> > +    /// 0                16               32                                               96
> > +    /// |----------------|----------------|------------------------------------------------|
> > +    ///        empty         CmallocData                         data
> 
> Can you put this inside of '```'? Then it will render nicely in markdown
> (don't forget to specify the type 'text')

Sure.

> 
> > +    ///
> > +    /// For this example the returned `Layout` has an alignment of 32 and a size of 96.
> > +    fn layout_adjust(layout: Layout) -> Result<Layout, AllocError> {
> > +        let layout = layout.pad_to_align();
> > +
> > +        // Ensure that `CmallocData` fits into half the alignment. Additionally, this guarantees
> > +        // that advancing a pointer aligned to `align` by `align / 2` we still satisfy or exceed
> > +        // the alignment requested through `layout`.
> > +        let align = cmp::max(
> > +            layout.align(),
> > +            mem::size_of::<CmallocData>().next_power_of_two(),
> > +        ) * 2;
> > +
> > +        // Add the additional space required for `CmallocData`.
> > +        let size = layout.size() + mem::size_of::<CmallocData>();
> > +
> > +        Ok(Layout::from_size_align(size, align)
> > +            .map_err(|_| AllocError)?
> > +            .pad_to_align())
> > +    }
> > +
> > +    fn alloc_store_data(layout: Layout) -> Result<NonNull<u8>, AllocError> {
> > +        let requested_size = layout.size();
> > +
> > +        let layout = Self::layout_adjust(layout)?;
> > +        let min_align = layout.align() / 2;
> > +
> > +        // SAFETY: Returns either NULL or a pointer to a memory allocation that satisfies or
> > +        // exceeds the given size and alignment requirements.
> > +        let raw_ptr = unsafe { libc_aligned_alloc(layout.align(), layout.size()) } as *mut u8;
> > +
> > +        let priv_ptr = NonNull::new(raw_ptr).ok_or(AllocError)?;
> > +
> > +        // SAFETY: Advance the pointer by `min_align`. The adjustments from `Self::layout_adjust`
> > +        // ensure that after this operation the original size and alignment requirements are still
> > +        // satisfied or exceeded.
> 
> This SAFETY comment should address why it's OK to call `add`. You
> justify something different, namely why the allocation still satisfies
> the requirements of `layout`. That is something that this function
> should probably guarantee.

So, I guess you're arguing that instead I should say that, we're still within
the bounds of the same allocated object and don't exceed `isize`?

> 
> > +        let ptr = unsafe { priv_ptr.as_ptr().add(min_align) };
> > +
> > +        // SAFETY: `min_align` is greater than or equal to the size of `CmallocData`, hence we
> > +        // don't exceed the allocation boundaries.
> > +        let data_ptr: *mut CmallocData = unsafe { ptr.sub(mem::size_of::<CmallocData>()) }.cast();
> > +
> > +        let data = CmallocData {
> > +            size: requested_size,
> > +            offset: min_align,
> > +        };
> > +
> > +        // SAFETY: `data_ptr` is properly aligned and within the allocation boundaries reserved for
> > +        // `CmallocData`.
> > +        unsafe { data_ptr.write(data) };
> > +
> > +        NonNull::new(ptr).ok_or(AllocError)
> > +    }
> > +
> > +    /// # Safety
> > +    ///
> > +    /// `ptr` must have been previously allocated with `Self::alloc_store_data`.
> 
> You additionally need that you have shared access to the pointee.
> 
> > +    unsafe fn data<'a>(ptr: NonNull<u8>) -> &'a CmallocData {
> > +        // SAFETY: `Self::alloc_store_data` stores the `CmallocData` right before the address
> > +        // returned to callers of `Self::alloc_store_data`.
> > +        let data_ptr: *mut CmallocData =
> > +            unsafe { ptr.as_ptr().sub(mem::size_of::<CmallocData>()) }.cast();
> > +
> > +        // SAFETY: The `CmallocData` has been previously stored at this offset with
> > +        // `Self::alloc_store_data`.
> > +        unsafe { &*data_ptr }
> > +    }
> > +
> > +    /// # Safety
> > +    ///
> > +    /// This function must not be called more than once for the same allocation.
> > +    ///
> > +    /// `ptr` must have been previously allocated with `Self::alloc_store_data`.
> 
> You additionally need that you have exclusive access to the pointee.
> 
> > +    unsafe fn free_read_data(ptr: NonNull<u8>) {
> > +        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
> > +        let data = unsafe { Self::data(ptr) };
> > +
> > +        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
> > +        let priv_ptr = unsafe { ptr.as_ptr().sub(data.offset) };
> > +
> > +        // SAFETY: `priv_ptr` has previously been allocatored with this `Allocator`.
> > +        unsafe { libc_free(priv_ptr.cast()) };
> > +    }
> > +}
> > +
> > +unsafe impl Allocator for Cmalloc {
> > +    fn alloc(layout: Layout, flags: Flags) -> Result<NonNull<[u8]>, AllocError> {
> > +        if layout.size() == 0 {
> > +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
> > +        }
> > +
> > +        let ptr = Self::alloc_store_data(layout)?;
> > +
> > +        if flags.contains(__GFP_ZERO) {
> > +            // SAFETY: `Self::alloc_store_data` guarantees that `ptr` points to memory of at least
> > +            // `layout.size()` bytes.
> > +            unsafe { ptr.as_ptr().write_bytes(0, layout.size()) };
> > +        }
> 
> This makes me wonder, what other flags should we handle for this
> allocator?

I don't think there are any other flags that we can handle. The only other one
that'd make sense is __GFP_NOFAIL, but we can't guarantee that.

If any specific gfp flags are needed, I think it's simply not a candidate for a
userspace test.

If we really want to do something here, we could whitelist the flags we ignore,
since they do not matter (such as __GFP_NOWARN) and panic() for everything else.

But I don't think that's really needed.

> 
> > +
> > +        Ok(NonNull::slice_from_raw_parts(ptr, layout.size()))
> > +    }
> > +
> >      unsafe fn realloc(
> > -        _ptr: Option<NonNull<u8>>,
> > -        _layout: Layout,
> > -        _flags: Flags,
> > +        ptr: Option<NonNull<u8>>,
> > +        layout: Layout,
> > +        flags: Flags,
> >      ) -> Result<NonNull<[u8]>, AllocError> {
> > -        panic!();
> > +        let src: NonNull<u8> = if let Some(src) = ptr {
> > +            src.cast()
> 
> Why the cast?

Probably a copy-paste mistake.

> 
> > +        } else {
> > +            return Self::alloc(layout, flags);
> > +        };
> 
> You should be able to write this instead:
> 
>     let Some(src) = ptr else {
>         return Self::alloc(layout, flags);
>     };

Yes, indeed.

> 
> > +
> > +        if layout.size() == 0 {
> > +            // SAFETY: `src` has been created by `Self::alloc_store_data`.
> 
> This is not true, consider:
> 
>     let ptr = alloc(size = 0);
>     free(ptr)
> 
> Alloc will return a dangling pointer due to the first if statement and
> then this function will pass it to `free_read_data`, even though it
> wasn't created by `alloc_store_data`.
> This isn't forbidden by the `Allocator` trait function's safety
> requirements.
> 
> > +            unsafe { Self::free_read_data(src) };
> > +
> > +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
> > +        }
> > +
> > +        let dst = Self::alloc(layout, flags)?;
> > +
> > +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
> > +        let data = unsafe { Self::data(src) };
> 
> Same issue here, if the allocation passed in is zero size. I think you
> have no other choice than to allocate even for zero size requests...
> Otherwise how would you know that they are zero-sized.

Good catch - gonna fix it.

> 
> ---
> Cheers,
> Benno
> 
> > +
> > +        // SAFETY: `src` has previously been allocated with this `Allocator`; `dst` has just been
> > +        // newly allocated. Copy up to the smaller of both sizes.
> > +        unsafe {
> > +            ptr::copy_nonoverlapping(
> > +                src.as_ptr(),
> > +                dst.as_ptr().cast(),
> > +                cmp::min(layout.size(), data.size),
> > +            )
> > +        };
> > +
> > +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
> > +        unsafe { Self::free_read_data(src) };
> > +
> > +        Ok(dst)
> >      }
> >  }
> > --
> > 2.46.0
> > 
>
Benno Lossin Aug. 30, 2024, 12:56 p.m. UTC | #3
On 30.08.24 00:25, Danilo Krummrich wrote:
> On Thu, Aug 29, 2024 at 07:14:18PM +0000, Benno Lossin wrote:
>> On 16.08.24 02:11, Danilo Krummrich wrote:
>>> +impl Cmalloc {
>>> +    /// Adjust the size and alignment such that we can additionally store `CmallocData` right
>>> +    /// before the actual data described by `layout`.
>>> +    ///
>>> +    /// Example:
>>> +    ///
>>> +    /// For `CmallocData` assume an alignment of 8 and a size of 16.
>>> +    /// For `layout` assume and alignment of 16 and a size of 64.
>>
>> This looks like you want it rendered as bulletpoints (but it won't).
> 
> Actually, that wasn't my intention, but I'm fine changing that.

I see, in that case not putting a newline there is also fine with me.
But I think bulletpoints are probably easier to read.

>>> +    fn alloc_store_data(layout: Layout) -> Result<NonNull<u8>, AllocError> {
>>> +        let requested_size = layout.size();
>>> +
>>> +        let layout = Self::layout_adjust(layout)?;
>>> +        let min_align = layout.align() / 2;
>>> +
>>> +        // SAFETY: Returns either NULL or a pointer to a memory allocation that satisfies or
>>> +        // exceeds the given size and alignment requirements.
>>> +        let raw_ptr = unsafe { libc_aligned_alloc(layout.align(), layout.size()) } as *mut u8;
>>> +
>>> +        let priv_ptr = NonNull::new(raw_ptr).ok_or(AllocError)?;
>>> +
>>> +        // SAFETY: Advance the pointer by `min_align`. The adjustments from `Self::layout_adjust`
>>> +        // ensure that after this operation the original size and alignment requirements are still
>>> +        // satisfied or exceeded.
>>
>> This SAFETY comment should address why it's OK to call `add`. You
>> justify something different, namely why the allocation still satisfies
>> the requirements of `layout`. That is something that this function
>> should probably guarantee.
> 
> So, I guess you're arguing that instead I should say that, we're still within
> the bounds of the same allocated object and don't exceed `isize`?

Yes.

>>> +    unsafe fn free_read_data(ptr: NonNull<u8>) {
>>> +        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
>>> +        let data = unsafe { Self::data(ptr) };
>>> +
>>> +        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
>>> +        let priv_ptr = unsafe { ptr.as_ptr().sub(data.offset) };
>>> +
>>> +        // SAFETY: `priv_ptr` has previously been allocatored with this `Allocator`.
>>> +        unsafe { libc_free(priv_ptr.cast()) };
>>> +    }
>>> +}
>>> +
>>> +unsafe impl Allocator for Cmalloc {
>>> +    fn alloc(layout: Layout, flags: Flags) -> Result<NonNull<[u8]>, AllocError> {
>>> +        if layout.size() == 0 {
>>> +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
>>> +        }
>>> +
>>> +        let ptr = Self::alloc_store_data(layout)?;
>>> +
>>> +        if flags.contains(__GFP_ZERO) {
>>> +            // SAFETY: `Self::alloc_store_data` guarantees that `ptr` points to memory of at least
>>> +            // `layout.size()` bytes.
>>> +            unsafe { ptr.as_ptr().write_bytes(0, layout.size()) };
>>> +        }
>>
>> This makes me wonder, what other flags should we handle for this
>> allocator?
> 
> I don't think there are any other flags that we can handle. The only other one
> that'd make sense is __GFP_NOFAIL, but we can't guarantee that.
> 
> If any specific gfp flags are needed, I think it's simply not a candidate for a
> userspace test.
> 
> If we really want to do something here, we could whitelist the flags we ignore,
> since they do not matter (such as __GFP_NOWARN) and panic() for everything else.
> 
> But I don't think that's really needed.

Makes sense, just wanted to check that this has been accounted for.

---
Cheers,
Benno
Danilo Krummrich Sept. 11, 2024, 12:31 p.m. UTC | #4
On Fri, Aug 30, 2024 at 12:25:27AM +0200, Danilo Krummrich wrote:
> On Thu, Aug 29, 2024 at 07:14:18PM +0000, Benno Lossin wrote:
> > On 16.08.24 02:11, Danilo Krummrich wrote:
> > 
> > > +
> > > +        if layout.size() == 0 {
> > > +            // SAFETY: `src` has been created by `Self::alloc_store_data`.
> > 
> > This is not true, consider:
> > 
> >     let ptr = alloc(size = 0);
> >     free(ptr)
> > 
> > Alloc will return a dangling pointer due to the first if statement and
> > then this function will pass it to `free_read_data`, even though it
> > wasn't created by `alloc_store_data`.
> > This isn't forbidden by the `Allocator` trait function's safety
> > requirements.
> > 
> > > +            unsafe { Self::free_read_data(src) };
> > > +
> > > +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
> > > +        }
> > > +
> > > +        let dst = Self::alloc(layout, flags)?;
> > > +
> > > +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
> > > +        let data = unsafe { Self::data(src) };
> > 
> > Same issue here, if the allocation passed in is zero size. I think you
> > have no other choice than to allocate even for zero size requests...
> > Otherwise how would you know that they are zero-sized.
> 
> Good catch - gonna fix it.

Almost got me. :) I think the code is fine, callers are not allowed to pass
pointers to `realloc` and `free`, which haven't been allocated with the same
corresponding allocator or are dangling.

> 
> > 
> > ---
> > Cheers,
> > Benno
> > 
> > > +
> > > +        // SAFETY: `src` has previously been allocated with this `Allocator`; `dst` has just been
> > > +        // newly allocated. Copy up to the smaller of both sizes.
> > > +        unsafe {
> > > +            ptr::copy_nonoverlapping(
> > > +                src.as_ptr(),
> > > +                dst.as_ptr().cast(),
> > > +                cmp::min(layout.size(), data.size),
> > > +            )
> > > +        };
> > > +
> > > +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
> > > +        unsafe { Self::free_read_data(src) };
> > > +
> > > +        Ok(dst)
> > >      }
> > >  }
> > > --
> > > 2.46.0
> > > 
> >
Benno Lossin Sept. 11, 2024, 1:32 p.m. UTC | #5
On 11.09.24 14:31, Danilo Krummrich wrote:
> On Fri, Aug 30, 2024 at 12:25:27AM +0200, Danilo Krummrich wrote:
>> On Thu, Aug 29, 2024 at 07:14:18PM +0000, Benno Lossin wrote:
>>> On 16.08.24 02:11, Danilo Krummrich wrote:
>>>> +
>>>> +        if layout.size() == 0 {
>>>> +            // SAFETY: `src` has been created by `Self::alloc_store_data`.
>>>
>>> This is not true, consider:
>>>
>>>     let ptr = alloc(size = 0);
>>>     free(ptr)
>>>
>>> Alloc will return a dangling pointer due to the first if statement and
>>> then this function will pass it to `free_read_data`, even though it
>>> wasn't created by `alloc_store_data`.
>>> This isn't forbidden by the `Allocator` trait function's safety
>>> requirements.
>>>
>>>> +            unsafe { Self::free_read_data(src) };
>>>> +
>>>> +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
>>>> +        }
>>>> +
>>>> +        let dst = Self::alloc(layout, flags)?;
>>>> +
>>>> +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
>>>> +        let data = unsafe { Self::data(src) };
>>>
>>> Same issue here, if the allocation passed in is zero size. I think you
>>> have no other choice than to allocate even for zero size requests...
>>> Otherwise how would you know that they are zero-sized.
>>
>> Good catch - gonna fix it.
> 
> Almost got me. :) I think the code is fine, callers are not allowed to pass
> pointers to `realloc` and `free`, which haven't been allocated with the same
> corresponding allocator or are dangling.

But what about the example above (ie the `alloc(size = 0)` and then
`free`)? I guess this all depends on how one interprets the term
"existing, valid memory allocation". To me that describes anything an
`Allocator` returns via `alloc` and `realloc`, including zero-sized
allocations.
But if you argue that those are not valid allocations from that
allocator, then that is not properly documented in the safety
requirements of `Allocator`. 

---
Cheers,
Benno
Danilo Krummrich Sept. 11, 2024, 2:37 p.m. UTC | #6
On Wed, Sep 11, 2024 at 01:32:31PM +0000, Benno Lossin wrote:
> On 11.09.24 14:31, Danilo Krummrich wrote:
> > On Fri, Aug 30, 2024 at 12:25:27AM +0200, Danilo Krummrich wrote:
> >> On Thu, Aug 29, 2024 at 07:14:18PM +0000, Benno Lossin wrote:
> >>> On 16.08.24 02:11, Danilo Krummrich wrote:
> >>>> +
> >>>> +        if layout.size() == 0 {
> >>>> +            // SAFETY: `src` has been created by `Self::alloc_store_data`.
> >>>
> >>> This is not true, consider:
> >>>
> >>>     let ptr = alloc(size = 0);
> >>>     free(ptr)
> >>>
> >>> Alloc will return a dangling pointer due to the first if statement and
> >>> then this function will pass it to `free_read_data`, even though it
> >>> wasn't created by `alloc_store_data`.
> >>> This isn't forbidden by the `Allocator` trait function's safety
> >>> requirements.
> >>>
> >>>> +            unsafe { Self::free_read_data(src) };
> >>>> +
> >>>> +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
> >>>> +        }
> >>>> +
> >>>> +        let dst = Self::alloc(layout, flags)?;
> >>>> +
> >>>> +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
> >>>> +        let data = unsafe { Self::data(src) };
> >>>
> >>> Same issue here, if the allocation passed in is zero size. I think you
> >>> have no other choice than to allocate even for zero size requests...
> >>> Otherwise how would you know that they are zero-sized.
> >>
> >> Good catch - gonna fix it.
> > 
> > Almost got me. :) I think the code is fine, callers are not allowed to pass
> > pointers to `realloc` and `free`, which haven't been allocated with the same
> > corresponding allocator or are dangling.
> 
> But what about the example above (ie the `alloc(size = 0)` and then
> `free`)?

This never has been valid for the `Allocator` trait. Look at `Kmalloc`,
`Vmalloc` and `KVmalloc`, they don't allow this either.

We've discussed this already in previous versions of this series, where for this
purpose, you asked for `old_layout` for `free`. Such that `free` can check if
the `size` was zero and therefore return without doing anything.

> I guess this all depends on how one interprets the term
> "existing, valid memory allocation". To me that describes anything an
> `Allocator` returns via `alloc` and `realloc`, including zero-sized
> allocations.

I argue that the dangling pointer returned for `size == 0` does not point to any
allocation in the sense of those allocators. It's just a dangling `[u8]`
pointer.

> But if you argue that those are not valid allocations from that
> allocator, then that is not properly documented in the safety
> requirements of `Allocator`. 

The safety requirements of `Allocator` where proposed by you and I thought they
consider this aspect?

`realloc` has:

"If `ptr == Some(p)`, then `p` must point to an existing and valid memory
allocation created by this allocator."

`free` has:

"`ptr` must point to an existing and valid memory allocation created by this
`Allocator` and must not be a dangling pointer."

We can add the part about the dangling pointer to `realloc` if you want.

> 
> ---
> Cheers,
> Benno
>
Benno Lossin Sept. 12, 2024, 8:18 a.m. UTC | #7
On 11.09.24 16:37, Danilo Krummrich wrote:
> On Wed, Sep 11, 2024 at 01:32:31PM +0000, Benno Lossin wrote:
>> On 11.09.24 14:31, Danilo Krummrich wrote:
>>> On Fri, Aug 30, 2024 at 12:25:27AM +0200, Danilo Krummrich wrote:
>>>> On Thu, Aug 29, 2024 at 07:14:18PM +0000, Benno Lossin wrote:
>>>>> On 16.08.24 02:11, Danilo Krummrich wrote:
>>>>>> +
>>>>>> +        if layout.size() == 0 {
>>>>>> +            // SAFETY: `src` has been created by `Self::alloc_store_data`.
>>>>>
>>>>> This is not true, consider:
>>>>>
>>>>>     let ptr = alloc(size = 0);
>>>>>     free(ptr)
>>>>>
>>>>> Alloc will return a dangling pointer due to the first if statement and
>>>>> then this function will pass it to `free_read_data`, even though it
>>>>> wasn't created by `alloc_store_data`.
>>>>> This isn't forbidden by the `Allocator` trait function's safety
>>>>> requirements.
>>>>>
>>>>>> +            unsafe { Self::free_read_data(src) };
>>>>>> +
>>>>>> +            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
>>>>>> +        }
>>>>>> +
>>>>>> +        let dst = Self::alloc(layout, flags)?;
>>>>>> +
>>>>>> +        // SAFETY: `src` has been created by `Self::alloc_store_data`.
>>>>>> +        let data = unsafe { Self::data(src) };
>>>>>
>>>>> Same issue here, if the allocation passed in is zero size. I think you
>>>>> have no other choice than to allocate even for zero size requests...
>>>>> Otherwise how would you know that they are zero-sized.
>>>>
>>>> Good catch - gonna fix it.
>>>
>>> Almost got me. :) I think the code is fine, callers are not allowed to pass
>>> pointers to `realloc` and `free`, which haven't been allocated with the same
>>> corresponding allocator or are dangling.
>>
>> But what about the example above (ie the `alloc(size = 0)` and then
>> `free`)?
> 
> This never has been valid for the `Allocator` trait. Look at `Kmalloc`,
> `Vmalloc` and `KVmalloc`, they don't allow this either.

That is true.

> We've discussed this already in previous versions of this series, where for this
> purpose, you asked for `old_layout` for `free`. Such that `free` can check if
> the `size` was zero and therefore return without doing anything.

Yes, but that was only about the old_layout parameter (at least that's
what I thought).

>> I guess this all depends on how one interprets the term
>> "existing, valid memory allocation". To me that describes anything an
>> `Allocator` returns via `alloc` and `realloc`, including zero-sized
>> allocations.
> 
> I argue that the dangling pointer returned for `size == 0` does not point to any
> allocation in the sense of those allocators. It's just a dangling `[u8]`
> pointer.

Sure, but to me the concept of zero-sized allocations does exist.

>> But if you argue that those are not valid allocations from that
>> allocator, then that is not properly documented in the safety
>> requirements of `Allocator`.
> 
> The safety requirements of `Allocator` where proposed by you and I thought they
> consider this aspect?

No, they did not consider this aspect. I was under the impression, that
we would still allow zero-sized allocations (in retrospect, this is
stupid, since dangling pointers shouldn't be passed to `krealloc` etc.).

> `realloc` has:
> 
> "If `ptr == Some(p)`, then `p` must point to an existing and valid memory
> allocation created by this allocator."
> 
> `free` has:
> 
> "`ptr` must point to an existing and valid memory allocation created by this
> `Allocator` and must not be a dangling pointer."
> 
> We can add the part about the dangling pointer to `realloc` if you want.

So I think we should do the following: 
(1) Add a paragraph to the `Allocator` trait that explains that
    zero-sized allocations are not supported.
(2) Add a check to `realloc` for zero-sized allocations + null pointer
    (ie a new allocation request) that prints a warning and returns an
    error
(3) Instead of writing "existing and valid memory allocation created by
    this allocator", I think "valid non-zero-sized memory allocation
    created by this allocator" fits better.

---
Cheers,
Benno
diff mbox series

Patch

diff --git a/rust/kernel/alloc/allocator_test.rs b/rust/kernel/alloc/allocator_test.rs
index 1b2642c547ec..7fff308d02dc 100644
--- a/rust/kernel/alloc/allocator_test.rs
+++ b/rust/kernel/alloc/allocator_test.rs
@@ -2,20 +2,184 @@ 
 
 #![allow(missing_docs)]
 
-use super::{AllocError, Allocator, Flags};
+use super::{flags::*, AllocError, Allocator, Flags};
 use core::alloc::Layout;
+use core::cmp;
+use core::mem;
+use core::ptr;
 use core::ptr::NonNull;
 
-pub struct Kmalloc;
+pub struct Cmalloc;
+pub type Kmalloc = Cmalloc;
 pub type Vmalloc = Kmalloc;
 pub type KVmalloc = Kmalloc;
 
-unsafe impl Allocator for Kmalloc {
+extern "C" {
+    #[link_name = "aligned_alloc"]
+    fn libc_aligned_alloc(align: usize, size: usize) -> *mut core::ffi::c_void;
+
+    #[link_name = "free"]
+    fn libc_free(ptr: *mut core::ffi::c_void);
+}
+
+struct CmallocData {
+    // The actual size as requested through `Cmalloc::alloc` or `Cmalloc::realloc`.
+    size: usize,
+    // The offset from the pointer returned to the caller of `Cmalloc::alloc` or `Cmalloc::realloc`
+    // to the actual base address of the allocation.
+    offset: usize,
+}
+
+impl Cmalloc {
+    /// Adjust the size and alignment such that we can additionally store `CmallocData` right
+    /// before the actual data described by `layout`.
+    ///
+    /// Example:
+    ///
+    /// For `CmallocData` assume an alignment of 8 and a size of 16.
+    /// For `layout` assume and alignment of 16 and a size of 64.
+    ///
+    /// 0                16               32                                               96
+    /// |----------------|----------------|------------------------------------------------|
+    ///        empty         CmallocData                         data
+    ///
+    /// For this example the returned `Layout` has an alignment of 32 and a size of 96.
+    fn layout_adjust(layout: Layout) -> Result<Layout, AllocError> {
+        let layout = layout.pad_to_align();
+
+        // Ensure that `CmallocData` fits into half the alignment. Additionally, this guarantees
+        // that advancing a pointer aligned to `align` by `align / 2` we still satisfy or exceed
+        // the alignment requested through `layout`.
+        let align = cmp::max(
+            layout.align(),
+            mem::size_of::<CmallocData>().next_power_of_two(),
+        ) * 2;
+
+        // Add the additional space required for `CmallocData`.
+        let size = layout.size() + mem::size_of::<CmallocData>();
+
+        Ok(Layout::from_size_align(size, align)
+            .map_err(|_| AllocError)?
+            .pad_to_align())
+    }
+
+    fn alloc_store_data(layout: Layout) -> Result<NonNull<u8>, AllocError> {
+        let requested_size = layout.size();
+
+        let layout = Self::layout_adjust(layout)?;
+        let min_align = layout.align() / 2;
+
+        // SAFETY: Returns either NULL or a pointer to a memory allocation that satisfies or
+        // exceeds the given size and alignment requirements.
+        let raw_ptr = unsafe { libc_aligned_alloc(layout.align(), layout.size()) } as *mut u8;
+
+        let priv_ptr = NonNull::new(raw_ptr).ok_or(AllocError)?;
+
+        // SAFETY: Advance the pointer by `min_align`. The adjustments from `Self::layout_adjust`
+        // ensure that after this operation the original size and alignment requirements are still
+        // satisfied or exceeded.
+        let ptr = unsafe { priv_ptr.as_ptr().add(min_align) };
+
+        // SAFETY: `min_align` is greater than or equal to the size of `CmallocData`, hence we
+        // don't exceed the allocation boundaries.
+        let data_ptr: *mut CmallocData = unsafe { ptr.sub(mem::size_of::<CmallocData>()) }.cast();
+
+        let data = CmallocData {
+            size: requested_size,
+            offset: min_align,
+        };
+
+        // SAFETY: `data_ptr` is properly aligned and within the allocation boundaries reserved for
+        // `CmallocData`.
+        unsafe { data_ptr.write(data) };
+
+        NonNull::new(ptr).ok_or(AllocError)
+    }
+
+    /// # Safety
+    ///
+    /// `ptr` must have been previously allocated with `Self::alloc_store_data`.
+    unsafe fn data<'a>(ptr: NonNull<u8>) -> &'a CmallocData {
+        // SAFETY: `Self::alloc_store_data` stores the `CmallocData` right before the address
+        // returned to callers of `Self::alloc_store_data`.
+        let data_ptr: *mut CmallocData =
+            unsafe { ptr.as_ptr().sub(mem::size_of::<CmallocData>()) }.cast();
+
+        // SAFETY: The `CmallocData` has been previously stored at this offset with
+        // `Self::alloc_store_data`.
+        unsafe { &*data_ptr }
+    }
+
+    /// # Safety
+    ///
+    /// This function must not be called more than once for the same allocation.
+    ///
+    /// `ptr` must have been previously allocated with `Self::alloc_store_data`.
+    unsafe fn free_read_data(ptr: NonNull<u8>) {
+        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
+        let data = unsafe { Self::data(ptr) };
+
+        // SAFETY: `ptr` has been created by `Self::alloc_store_data`.
+        let priv_ptr = unsafe { ptr.as_ptr().sub(data.offset) };
+
+        // SAFETY: `priv_ptr` has previously been allocatored with this `Allocator`.
+        unsafe { libc_free(priv_ptr.cast()) };
+    }
+}
+
+unsafe impl Allocator for Cmalloc {
+    fn alloc(layout: Layout, flags: Flags) -> Result<NonNull<[u8]>, AllocError> {
+        if layout.size() == 0 {
+            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
+        }
+
+        let ptr = Self::alloc_store_data(layout)?;
+
+        if flags.contains(__GFP_ZERO) {
+            // SAFETY: `Self::alloc_store_data` guarantees that `ptr` points to memory of at least
+            // `layout.size()` bytes.
+            unsafe { ptr.as_ptr().write_bytes(0, layout.size()) };
+        }
+
+        Ok(NonNull::slice_from_raw_parts(ptr, layout.size()))
+    }
+
     unsafe fn realloc(
-        _ptr: Option<NonNull<u8>>,
-        _layout: Layout,
-        _flags: Flags,
+        ptr: Option<NonNull<u8>>,
+        layout: Layout,
+        flags: Flags,
     ) -> Result<NonNull<[u8]>, AllocError> {
-        panic!();
+        let src: NonNull<u8> = if let Some(src) = ptr {
+            src.cast()
+        } else {
+            return Self::alloc(layout, flags);
+        };
+
+        if layout.size() == 0 {
+            // SAFETY: `src` has been created by `Self::alloc_store_data`.
+            unsafe { Self::free_read_data(src) };
+
+            return Ok(NonNull::slice_from_raw_parts(NonNull::dangling(), 0));
+        }
+
+        let dst = Self::alloc(layout, flags)?;
+
+        // SAFETY: `src` has been created by `Self::alloc_store_data`.
+        let data = unsafe { Self::data(src) };
+
+        // SAFETY: `src` has previously been allocated with this `Allocator`; `dst` has just been
+        // newly allocated. Copy up to the smaller of both sizes.
+        unsafe {
+            ptr::copy_nonoverlapping(
+                src.as_ptr(),
+                dst.as_ptr().cast(),
+                cmp::min(layout.size(), data.size),
+            )
+        };
+
+        // SAFETY: `src` has been created by `Self::alloc_store_data`.
+        unsafe { Self::free_read_data(src) };
+
+        Ok(dst)
     }
 }