xref: /linux/rust/kernel/dma.rs (revision 6a069876eb1402478900ee0eb7d7fe276bb1f4e3)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 //! Direct memory access (DMA).
4 //!
5 //! C header: [`include/linux/dma-mapping.h`](srctree/include/linux/dma-mapping.h)
6 
7 use crate::{
8     bindings, build_assert, device,
9     device::{Bound, Core},
10     error::{to_result, Result},
11     prelude::*,
12     sync::aref::ARef,
13     transmute::{AsBytes, FromBytes},
14 };
15 use core::ptr::NonNull;
16 
17 /// DMA address type.
18 ///
19 /// Represents a bus address used for Direct Memory Access (DMA) operations.
20 ///
21 /// This is an alias of the kernel's `dma_addr_t`, which may be `u32` or `u64` depending on
22 /// `CONFIG_ARCH_DMA_ADDR_T_64BIT`.
23 ///
24 /// Note that this may be `u64` even on 32-bit architectures.
25 pub type DmaAddress = bindings::dma_addr_t;
26 
27 /// Trait to be implemented by DMA capable bus devices.
28 ///
29 /// The [`dma::Device`](Device) trait should be implemented by bus specific device representations,
30 /// where the underlying bus is DMA capable, such as:
31 #[cfg_attr(CONFIG_PCI, doc = "* [`pci::Device`](kernel::pci::Device)")]
32 /// * [`platform::Device`](::kernel::platform::Device)
33 pub trait Device: AsRef<device::Device<Core>> {
34     /// Set up the device's DMA streaming addressing capabilities.
35     ///
36     /// This method is usually called once from `probe()` as soon as the device capabilities are
37     /// known.
38     ///
39     /// # Safety
40     ///
41     /// This method must not be called concurrently with any DMA allocation or mapping primitives,
42     /// such as [`CoherentAllocation::alloc_attrs`].
43     unsafe fn dma_set_mask(&self, mask: DmaMask) -> Result {
44         // SAFETY:
45         // - By the type invariant of `device::Device`, `self.as_ref().as_raw()` is valid.
46         // - The safety requirement of this function guarantees that there are no concurrent calls
47         //   to DMA allocation and mapping primitives using this mask.
48         to_result(unsafe { bindings::dma_set_mask(self.as_ref().as_raw(), mask.value()) })
49     }
50 
51     /// Set up the device's DMA coherent addressing capabilities.
52     ///
53     /// This method is usually called once from `probe()` as soon as the device capabilities are
54     /// known.
55     ///
56     /// # Safety
57     ///
58     /// This method must not be called concurrently with any DMA allocation or mapping primitives,
59     /// such as [`CoherentAllocation::alloc_attrs`].
60     unsafe fn dma_set_coherent_mask(&self, mask: DmaMask) -> Result {
61         // SAFETY:
62         // - By the type invariant of `device::Device`, `self.as_ref().as_raw()` is valid.
63         // - The safety requirement of this function guarantees that there are no concurrent calls
64         //   to DMA allocation and mapping primitives using this mask.
65         to_result(unsafe { bindings::dma_set_coherent_mask(self.as_ref().as_raw(), mask.value()) })
66     }
67 
68     /// Set up the device's DMA addressing capabilities.
69     ///
70     /// This is a combination of [`Device::dma_set_mask`] and [`Device::dma_set_coherent_mask`].
71     ///
72     /// This method is usually called once from `probe()` as soon as the device capabilities are
73     /// known.
74     ///
75     /// # Safety
76     ///
77     /// This method must not be called concurrently with any DMA allocation or mapping primitives,
78     /// such as [`CoherentAllocation::alloc_attrs`].
79     unsafe fn dma_set_mask_and_coherent(&self, mask: DmaMask) -> Result {
80         // SAFETY:
81         // - By the type invariant of `device::Device`, `self.as_ref().as_raw()` is valid.
82         // - The safety requirement of this function guarantees that there are no concurrent calls
83         //   to DMA allocation and mapping primitives using this mask.
84         to_result(unsafe {
85             bindings::dma_set_mask_and_coherent(self.as_ref().as_raw(), mask.value())
86         })
87     }
88 }
89 
90 /// A DMA mask that holds a bitmask with the lowest `n` bits set.
91 ///
92 /// Use [`DmaMask::new`] or [`DmaMask::try_new`] to construct a value. Values
93 /// are guaranteed to never exceed the bit width of `u64`.
94 ///
95 /// This is the Rust equivalent of the C macro `DMA_BIT_MASK()`.
96 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
97 pub struct DmaMask(u64);
98 
99 impl DmaMask {
100     /// Constructs a `DmaMask` with the lowest `n` bits set to `1`.
101     ///
102     /// For `n <= 64`, sets exactly the lowest `n` bits.
103     /// For `n > 64`, results in a build error.
104     ///
105     /// # Examples
106     ///
107     /// ```
108     /// use kernel::dma::DmaMask;
109     ///
110     /// let mask0 = DmaMask::new::<0>();
111     /// assert_eq!(mask0.value(), 0);
112     ///
113     /// let mask1 = DmaMask::new::<1>();
114     /// assert_eq!(mask1.value(), 0b1);
115     ///
116     /// let mask64 = DmaMask::new::<64>();
117     /// assert_eq!(mask64.value(), u64::MAX);
118     ///
119     /// // Build failure.
120     /// // let mask_overflow = DmaMask::new::<100>();
121     /// ```
122     #[inline]
123     pub const fn new<const N: u32>() -> Self {
124         let Ok(mask) = Self::try_new(N) else {
125             build_error!("Invalid DMA Mask.");
126         };
127 
128         mask
129     }
130 
131     /// Constructs a `DmaMask` with the lowest `n` bits set to `1`.
132     ///
133     /// For `n <= 64`, sets exactly the lowest `n` bits.
134     /// For `n > 64`, returns [`EINVAL`].
135     ///
136     /// # Examples
137     ///
138     /// ```
139     /// use kernel::dma::DmaMask;
140     ///
141     /// let mask0 = DmaMask::try_new(0)?;
142     /// assert_eq!(mask0.value(), 0);
143     ///
144     /// let mask1 = DmaMask::try_new(1)?;
145     /// assert_eq!(mask1.value(), 0b1);
146     ///
147     /// let mask64 = DmaMask::try_new(64)?;
148     /// assert_eq!(mask64.value(), u64::MAX);
149     ///
150     /// let mask_overflow = DmaMask::try_new(100);
151     /// assert!(mask_overflow.is_err());
152     /// # Ok::<(), Error>(())
153     /// ```
154     #[inline]
155     pub const fn try_new(n: u32) -> Result<Self> {
156         Ok(Self(match n {
157             0 => 0,
158             1..=64 => u64::MAX >> (64 - n),
159             _ => return Err(EINVAL),
160         }))
161     }
162 
163     /// Returns the underlying `u64` bitmask value.
164     #[inline]
165     pub const fn value(&self) -> u64 {
166         self.0
167     }
168 }
169 
170 /// Possible attributes associated with a DMA mapping.
171 ///
172 /// They can be combined with the operators `|`, `&`, and `!`.
173 ///
174 /// Values can be used from the [`attrs`] module.
175 ///
176 /// # Examples
177 ///
178 /// ```
179 /// # use kernel::device::{Bound, Device};
180 /// use kernel::dma::{attrs::*, CoherentAllocation};
181 ///
182 /// # fn test(dev: &Device<Bound>) -> Result {
183 /// let attribs = DMA_ATTR_FORCE_CONTIGUOUS | DMA_ATTR_NO_WARN;
184 /// let c: CoherentAllocation<u64> =
185 ///     CoherentAllocation::alloc_attrs(dev, 4, GFP_KERNEL, attribs)?;
186 /// # Ok::<(), Error>(()) }
187 /// ```
188 #[derive(Clone, Copy, PartialEq)]
189 #[repr(transparent)]
190 pub struct Attrs(u32);
191 
192 impl Attrs {
193     /// Get the raw representation of this attribute.
194     pub(crate) fn as_raw(self) -> crate::ffi::c_ulong {
195         self.0 as crate::ffi::c_ulong
196     }
197 
198     /// Check whether `flags` is contained in `self`.
199     pub fn contains(self, flags: Attrs) -> bool {
200         (self & flags) == flags
201     }
202 }
203 
204 impl core::ops::BitOr for Attrs {
205     type Output = Self;
206     fn bitor(self, rhs: Self) -> Self::Output {
207         Self(self.0 | rhs.0)
208     }
209 }
210 
211 impl core::ops::BitAnd for Attrs {
212     type Output = Self;
213     fn bitand(self, rhs: Self) -> Self::Output {
214         Self(self.0 & rhs.0)
215     }
216 }
217 
218 impl core::ops::Not for Attrs {
219     type Output = Self;
220     fn not(self) -> Self::Output {
221         Self(!self.0)
222     }
223 }
224 
225 /// DMA mapping attributes.
226 pub mod attrs {
227     use super::Attrs;
228 
229     /// Specifies that reads and writes to the mapping may be weakly ordered, that is that reads
230     /// and writes may pass each other.
231     pub const DMA_ATTR_WEAK_ORDERING: Attrs = Attrs(bindings::DMA_ATTR_WEAK_ORDERING);
232 
233     /// Specifies that writes to the mapping may be buffered to improve performance.
234     pub const DMA_ATTR_WRITE_COMBINE: Attrs = Attrs(bindings::DMA_ATTR_WRITE_COMBINE);
235 
236     /// Lets the platform to avoid creating a kernel virtual mapping for the allocated buffer.
237     pub const DMA_ATTR_NO_KERNEL_MAPPING: Attrs = Attrs(bindings::DMA_ATTR_NO_KERNEL_MAPPING);
238 
239     /// Allows platform code to skip synchronization of the CPU cache for the given buffer assuming
240     /// that it has been already transferred to 'device' domain.
241     pub const DMA_ATTR_SKIP_CPU_SYNC: Attrs = Attrs(bindings::DMA_ATTR_SKIP_CPU_SYNC);
242 
243     /// Forces contiguous allocation of the buffer in physical memory.
244     pub const DMA_ATTR_FORCE_CONTIGUOUS: Attrs = Attrs(bindings::DMA_ATTR_FORCE_CONTIGUOUS);
245 
246     /// Hints DMA-mapping subsystem that it's probably not worth the time to try
247     /// to allocate memory to in a way that gives better TLB efficiency.
248     pub const DMA_ATTR_ALLOC_SINGLE_PAGES: Attrs = Attrs(bindings::DMA_ATTR_ALLOC_SINGLE_PAGES);
249 
250     /// This tells the DMA-mapping subsystem to suppress allocation failure reports (similarly to
251     /// `__GFP_NOWARN`).
252     pub const DMA_ATTR_NO_WARN: Attrs = Attrs(bindings::DMA_ATTR_NO_WARN);
253 
254     /// Indicates that the buffer is fully accessible at an elevated privilege level (and
255     /// ideally inaccessible or at least read-only at lesser-privileged levels).
256     pub const DMA_ATTR_PRIVILEGED: Attrs = Attrs(bindings::DMA_ATTR_PRIVILEGED);
257 
258     /// Indicates that the buffer is MMIO memory.
259     pub const DMA_ATTR_MMIO: Attrs = Attrs(bindings::DMA_ATTR_MMIO);
260 }
261 
262 /// DMA data direction.
263 ///
264 /// Corresponds to the C [`enum dma_data_direction`].
265 ///
266 /// [`enum dma_data_direction`]: srctree/include/linux/dma-direction.h
267 #[derive(Copy, Clone, PartialEq, Eq, Debug)]
268 #[repr(u32)]
269 pub enum DataDirection {
270     /// The DMA mapping is for bidirectional data transfer.
271     ///
272     /// This is used when the buffer can be both read from and written to by the device.
273     /// The cache for the corresponding memory region is both flushed and invalidated.
274     Bidirectional = Self::const_cast(bindings::dma_data_direction_DMA_BIDIRECTIONAL),
275 
276     /// The DMA mapping is for data transfer from memory to the device (write).
277     ///
278     /// The CPU has prepared data in the buffer, and the device will read it.
279     /// The cache for the corresponding memory region is flushed before device access.
280     ToDevice = Self::const_cast(bindings::dma_data_direction_DMA_TO_DEVICE),
281 
282     /// The DMA mapping is for data transfer from the device to memory (read).
283     ///
284     /// The device will write data into the buffer for the CPU to read.
285     /// The cache for the corresponding memory region is invalidated before CPU access.
286     FromDevice = Self::const_cast(bindings::dma_data_direction_DMA_FROM_DEVICE),
287 
288     /// The DMA mapping is not for data transfer.
289     ///
290     /// This is primarily for debugging purposes. With this direction, the DMA mapping API
291     /// will not perform any cache coherency operations.
292     None = Self::const_cast(bindings::dma_data_direction_DMA_NONE),
293 }
294 
295 impl DataDirection {
296     /// Casts the bindgen-generated enum type to a `u32` at compile time.
297     ///
298     /// This function will cause a compile-time error if the underlying value of the
299     /// C enum is out of bounds for `u32`.
300     const fn const_cast(val: bindings::dma_data_direction) -> u32 {
301         // CAST: The C standard allows compilers to choose different integer types for enums.
302         // To safely check the value, we cast it to a wide signed integer type (`i128`)
303         // which can hold any standard C integer enum type without truncation.
304         let wide_val = val as i128;
305 
306         // Check if the value is outside the valid range for the target type `u32`.
307         // CAST: `u32::MAX` is cast to `i128` to match the type of `wide_val` for the comparison.
308         if wide_val < 0 || wide_val > u32::MAX as i128 {
309             // Trigger a compile-time error in a const context.
310             build_error!("C enum value is out of bounds for the target type `u32`.");
311         }
312 
313         // CAST: This cast is valid because the check above guarantees that `wide_val`
314         // is within the representable range of `u32`.
315         wide_val as u32
316     }
317 }
318 
319 impl From<DataDirection> for bindings::dma_data_direction {
320     /// Returns the raw representation of [`enum dma_data_direction`].
321     fn from(direction: DataDirection) -> Self {
322         // CAST: `direction as u32` gets the underlying representation of our `#[repr(u32)]` enum.
323         // The subsequent cast to `Self` (the bindgen type) assumes the C enum is compatible
324         // with the enum variants of `DataDirection`, which is a valid assumption given our
325         // compile-time checks.
326         direction as u32 as Self
327     }
328 }
329 
330 /// An abstraction of the `dma_alloc_coherent` API.
331 ///
332 /// This is an abstraction around the `dma_alloc_coherent` API which is used to allocate and map
333 /// large coherent DMA regions.
334 ///
335 /// A [`CoherentAllocation`] instance contains a pointer to the allocated region (in the
336 /// processor's virtual address space) and the device address which can be given to the device
337 /// as the DMA address base of the region. The region is released once [`CoherentAllocation`]
338 /// is dropped.
339 ///
340 /// # Invariants
341 ///
342 /// - For the lifetime of an instance of [`CoherentAllocation`], the `cpu_addr` is a valid pointer
343 ///   to an allocated region of coherent memory and `dma_handle` is the DMA address base of the
344 ///   region.
345 /// - The size in bytes of the allocation is equal to `size_of::<T> * count`.
346 /// - `size_of::<T> * count` fits into a `usize`.
347 // TODO
348 //
349 // DMA allocations potentially carry device resources (e.g.IOMMU mappings), hence for soundness
350 // reasons DMA allocation would need to be embedded in a `Devres` container, in order to ensure
351 // that device resources can never survive device unbind.
352 //
353 // However, it is neither desirable nor necessary to protect the allocated memory of the DMA
354 // allocation from surviving device unbind; it would require RCU read side critical sections to
355 // access the memory, which may require subsequent unnecessary copies.
356 //
357 // Hence, find a way to revoke the device resources of a `CoherentAllocation`, but not the
358 // entire `CoherentAllocation` including the allocated memory itself.
359 pub struct CoherentAllocation<T: AsBytes + FromBytes> {
360     dev: ARef<device::Device>,
361     dma_handle: DmaAddress,
362     count: usize,
363     cpu_addr: NonNull<T>,
364     dma_attrs: Attrs,
365 }
366 
367 impl<T: AsBytes + FromBytes> CoherentAllocation<T> {
368     /// Allocates a region of `size_of::<T> * count` of coherent memory.
369     ///
370     /// # Examples
371     ///
372     /// ```
373     /// # use kernel::device::{Bound, Device};
374     /// use kernel::dma::{attrs::*, CoherentAllocation};
375     ///
376     /// # fn test(dev: &Device<Bound>) -> Result {
377     /// let c: CoherentAllocation<u64> =
378     ///     CoherentAllocation::alloc_attrs(dev, 4, GFP_KERNEL, DMA_ATTR_NO_WARN)?;
379     /// # Ok::<(), Error>(()) }
380     /// ```
381     pub fn alloc_attrs(
382         dev: &device::Device<Bound>,
383         count: usize,
384         gfp_flags: kernel::alloc::Flags,
385         dma_attrs: Attrs,
386     ) -> Result<CoherentAllocation<T>> {
387         build_assert!(
388             core::mem::size_of::<T>() > 0,
389             "It doesn't make sense for the allocated type to be a ZST"
390         );
391 
392         let size = count
393             .checked_mul(core::mem::size_of::<T>())
394             .ok_or(EOVERFLOW)?;
395         let mut dma_handle = 0;
396         // SAFETY: Device pointer is guaranteed as valid by the type invariant on `Device`.
397         let addr = unsafe {
398             bindings::dma_alloc_attrs(
399                 dev.as_raw(),
400                 size,
401                 &mut dma_handle,
402                 gfp_flags.as_raw(),
403                 dma_attrs.as_raw(),
404             )
405         };
406         let addr = NonNull::new(addr).ok_or(ENOMEM)?;
407         // INVARIANT:
408         // - We just successfully allocated a coherent region which is accessible for
409         //   `count` elements, hence the cpu address is valid. We also hold a refcounted reference
410         //   to the device.
411         // - The allocated `size` is equal to `size_of::<T> * count`.
412         // - The allocated `size` fits into a `usize`.
413         Ok(Self {
414             dev: dev.into(),
415             dma_handle,
416             count,
417             cpu_addr: addr.cast(),
418             dma_attrs,
419         })
420     }
421 
422     /// Performs the same functionality as [`CoherentAllocation::alloc_attrs`], except the
423     /// `dma_attrs` is 0 by default.
424     pub fn alloc_coherent(
425         dev: &device::Device<Bound>,
426         count: usize,
427         gfp_flags: kernel::alloc::Flags,
428     ) -> Result<CoherentAllocation<T>> {
429         CoherentAllocation::alloc_attrs(dev, count, gfp_flags, Attrs(0))
430     }
431 
432     /// Returns the number of elements `T` in this allocation.
433     ///
434     /// Note that this is not the size of the allocation in bytes, which is provided by
435     /// [`Self::size`].
436     pub fn count(&self) -> usize {
437         self.count
438     }
439 
440     /// Returns the size in bytes of this allocation.
441     pub fn size(&self) -> usize {
442         // INVARIANT: The type invariant of `Self` guarantees that `size_of::<T> * count` fits into
443         // a `usize`.
444         self.count * core::mem::size_of::<T>()
445     }
446 
447     /// Returns the base address to the allocated region in the CPU's virtual address space.
448     pub fn start_ptr(&self) -> *const T {
449         self.cpu_addr.as_ptr()
450     }
451 
452     /// Returns the base address to the allocated region in the CPU's virtual address space as
453     /// a mutable pointer.
454     pub fn start_ptr_mut(&mut self) -> *mut T {
455         self.cpu_addr.as_ptr()
456     }
457 
458     /// Returns a DMA handle which may be given to the device as the DMA address base of
459     /// the region.
460     pub fn dma_handle(&self) -> DmaAddress {
461         self.dma_handle
462     }
463 
464     /// Returns a DMA handle starting at `offset` (in units of `T`) which may be given to the
465     /// device as the DMA address base of the region.
466     ///
467     /// Returns `EINVAL` if `offset` is not within the bounds of the allocation.
468     pub fn dma_handle_with_offset(&self, offset: usize) -> Result<DmaAddress> {
469         if offset >= self.count {
470             Err(EINVAL)
471         } else {
472             // INVARIANT: The type invariant of `Self` guarantees that `size_of::<T> * count` fits
473             // into a `usize`, and `offset` is inferior to `count`.
474             Ok(self.dma_handle + (offset * core::mem::size_of::<T>()) as DmaAddress)
475         }
476     }
477 
478     /// Common helper to validate a range applied from the allocated region in the CPU's virtual
479     /// address space.
480     fn validate_range(&self, offset: usize, count: usize) -> Result {
481         if offset.checked_add(count).ok_or(EOVERFLOW)? > self.count {
482             return Err(EINVAL);
483         }
484         Ok(())
485     }
486 
487     /// Returns the data from the region starting from `offset` as a slice.
488     /// `offset` and `count` are in units of `T`, not the number of bytes.
489     ///
490     /// For ringbuffer type of r/w access or use-cases where the pointer to the live data is needed,
491     /// [`CoherentAllocation::start_ptr`] or [`CoherentAllocation::start_ptr_mut`] could be used
492     /// instead.
493     ///
494     /// # Safety
495     ///
496     /// * Callers must ensure that the device does not read/write to/from memory while the returned
497     ///   slice is live.
498     /// * Callers must ensure that this call does not race with a write to the same region while
499     ///   the returned slice is live.
500     pub unsafe fn as_slice(&self, offset: usize, count: usize) -> Result<&[T]> {
501         self.validate_range(offset, count)?;
502         // SAFETY:
503         // - The pointer is valid due to type invariant on `CoherentAllocation`,
504         //   we've just checked that the range and index is within bounds. The immutability of the
505         //   data is also guaranteed by the safety requirements of the function.
506         // - `offset + count` can't overflow since it is smaller than `self.count` and we've checked
507         //   that `self.count` won't overflow early in the constructor.
508         Ok(unsafe { core::slice::from_raw_parts(self.start_ptr().add(offset), count) })
509     }
510 
511     /// Performs the same functionality as [`CoherentAllocation::as_slice`], except that a mutable
512     /// slice is returned.
513     ///
514     /// # Safety
515     ///
516     /// * Callers must ensure that the device does not read/write to/from memory while the returned
517     ///   slice is live.
518     /// * Callers must ensure that this call does not race with a read or write to the same region
519     ///   while the returned slice is live.
520     pub unsafe fn as_slice_mut(&mut self, offset: usize, count: usize) -> Result<&mut [T]> {
521         self.validate_range(offset, count)?;
522         // SAFETY:
523         // - The pointer is valid due to type invariant on `CoherentAllocation`,
524         //   we've just checked that the range and index is within bounds. The immutability of the
525         //   data is also guaranteed by the safety requirements of the function.
526         // - `offset + count` can't overflow since it is smaller than `self.count` and we've checked
527         //   that `self.count` won't overflow early in the constructor.
528         Ok(unsafe { core::slice::from_raw_parts_mut(self.start_ptr_mut().add(offset), count) })
529     }
530 
531     /// Writes data to the region starting from `offset`. `offset` is in units of `T`, not the
532     /// number of bytes.
533     ///
534     /// # Safety
535     ///
536     /// * Callers must ensure that this call does not race with a read or write to the same region
537     ///   that overlaps with this write.
538     ///
539     /// # Examples
540     ///
541     /// ```
542     /// # fn test(alloc: &mut kernel::dma::CoherentAllocation<u8>) -> Result {
543     /// let somedata: [u8; 4] = [0xf; 4];
544     /// let buf: &[u8] = &somedata;
545     /// // SAFETY: There is no concurrent HW operation on the device and no other R/W access to the
546     /// // region.
547     /// unsafe { alloc.write(buf, 0)?; }
548     /// # Ok::<(), Error>(()) }
549     /// ```
550     pub unsafe fn write(&mut self, src: &[T], offset: usize) -> Result {
551         self.validate_range(offset, src.len())?;
552         // SAFETY:
553         // - The pointer is valid due to type invariant on `CoherentAllocation`
554         //   and we've just checked that the range and index is within bounds.
555         // - `offset + count` can't overflow since it is smaller than `self.count` and we've checked
556         //   that `self.count` won't overflow early in the constructor.
557         unsafe {
558             core::ptr::copy_nonoverlapping(
559                 src.as_ptr(),
560                 self.start_ptr_mut().add(offset),
561                 src.len(),
562             )
563         };
564         Ok(())
565     }
566 
567     /// Returns a pointer to an element from the region with bounds checking. `offset` is in
568     /// units of `T`, not the number of bytes.
569     ///
570     /// Public but hidden since it should only be used from [`dma_read`] and [`dma_write`] macros.
571     #[doc(hidden)]
572     pub fn item_from_index(&self, offset: usize) -> Result<*mut T> {
573         if offset >= self.count {
574             return Err(EINVAL);
575         }
576         // SAFETY:
577         // - The pointer is valid due to type invariant on `CoherentAllocation`
578         // and we've just checked that the range and index is within bounds.
579         // - `offset` can't overflow since it is smaller than `self.count` and we've checked
580         // that `self.count` won't overflow early in the constructor.
581         Ok(unsafe { self.cpu_addr.as_ptr().add(offset) })
582     }
583 
584     /// Reads the value of `field` and ensures that its type is [`FromBytes`].
585     ///
586     /// # Safety
587     ///
588     /// This must be called from the [`dma_read`] macro which ensures that the `field` pointer is
589     /// validated beforehand.
590     ///
591     /// Public but hidden since it should only be used from [`dma_read`] macro.
592     #[doc(hidden)]
593     pub unsafe fn field_read<F: FromBytes>(&self, field: *const F) -> F {
594         // SAFETY:
595         // - By the safety requirements field is valid.
596         // - Using read_volatile() here is not sound as per the usual rules, the usage here is
597         // a special exception with the following notes in place. When dealing with a potential
598         // race from a hardware or code outside kernel (e.g. user-space program), we need that
599         // read on a valid memory is not UB. Currently read_volatile() is used for this, and the
600         // rationale behind is that it should generate the same code as READ_ONCE() which the
601         // kernel already relies on to avoid UB on data races. Note that the usage of
602         // read_volatile() is limited to this particular case, it cannot be used to prevent
603         // the UB caused by racing between two kernel functions nor do they provide atomicity.
604         unsafe { field.read_volatile() }
605     }
606 
607     /// Writes a value to `field` and ensures that its type is [`AsBytes`].
608     ///
609     /// # Safety
610     ///
611     /// This must be called from the [`dma_write`] macro which ensures that the `field` pointer is
612     /// validated beforehand.
613     ///
614     /// Public but hidden since it should only be used from [`dma_write`] macro.
615     #[doc(hidden)]
616     pub unsafe fn field_write<F: AsBytes>(&self, field: *mut F, val: F) {
617         // SAFETY:
618         // - By the safety requirements field is valid.
619         // - Using write_volatile() here is not sound as per the usual rules, the usage here is
620         // a special exception with the following notes in place. When dealing with a potential
621         // race from a hardware or code outside kernel (e.g. user-space program), we need that
622         // write on a valid memory is not UB. Currently write_volatile() is used for this, and the
623         // rationale behind is that it should generate the same code as WRITE_ONCE() which the
624         // kernel already relies on to avoid UB on data races. Note that the usage of
625         // write_volatile() is limited to this particular case, it cannot be used to prevent
626         // the UB caused by racing between two kernel functions nor do they provide atomicity.
627         unsafe { field.write_volatile(val) }
628     }
629 }
630 
631 /// Note that the device configured to do DMA must be halted before this object is dropped.
632 impl<T: AsBytes + FromBytes> Drop for CoherentAllocation<T> {
633     fn drop(&mut self) {
634         let size = self.count * core::mem::size_of::<T>();
635         // SAFETY: Device pointer is guaranteed as valid by the type invariant on `Device`.
636         // The cpu address, and the dma handle are valid due to the type invariants on
637         // `CoherentAllocation`.
638         unsafe {
639             bindings::dma_free_attrs(
640                 self.dev.as_raw(),
641                 size,
642                 self.start_ptr_mut().cast(),
643                 self.dma_handle,
644                 self.dma_attrs.as_raw(),
645             )
646         }
647     }
648 }
649 
650 // SAFETY: It is safe to send a `CoherentAllocation` to another thread if `T`
651 // can be sent to another thread.
652 unsafe impl<T: AsBytes + FromBytes + Send> Send for CoherentAllocation<T> {}
653 
654 /// Reads a field of an item from an allocated region of structs.
655 ///
656 /// # Examples
657 ///
658 /// ```
659 /// use kernel::device::Device;
660 /// use kernel::dma::{attrs::*, CoherentAllocation};
661 ///
662 /// struct MyStruct { field: u32, }
663 ///
664 /// // SAFETY: All bit patterns are acceptable values for `MyStruct`.
665 /// unsafe impl kernel::transmute::FromBytes for MyStruct{};
666 /// // SAFETY: Instances of `MyStruct` have no uninitialized portions.
667 /// unsafe impl kernel::transmute::AsBytes for MyStruct{};
668 ///
669 /// # fn test(alloc: &kernel::dma::CoherentAllocation<MyStruct>) -> Result {
670 /// let whole = kernel::dma_read!(alloc[2]);
671 /// let field = kernel::dma_read!(alloc[1].field);
672 /// # Ok::<(), Error>(()) }
673 /// ```
674 #[macro_export]
675 macro_rules! dma_read {
676     ($dma:expr, $idx: expr, $($field:tt)*) => {{
677         (|| -> ::core::result::Result<_, $crate::error::Error> {
678             let item = $crate::dma::CoherentAllocation::item_from_index(&$dma, $idx)?;
679             // SAFETY: `item_from_index` ensures that `item` is always a valid pointer and can be
680             // dereferenced. The compiler also further validates the expression on whether `field`
681             // is a member of `item` when expanded by the macro.
682             unsafe {
683                 let ptr_field = ::core::ptr::addr_of!((*item) $($field)*);
684                 ::core::result::Result::Ok(
685                     $crate::dma::CoherentAllocation::field_read(&$dma, ptr_field)
686                 )
687             }
688         })()
689     }};
690     ($dma:ident [ $idx:expr ] $($field:tt)* ) => {
691         $crate::dma_read!($dma, $idx, $($field)*)
692     };
693     ($($dma:ident).* [ $idx:expr ] $($field:tt)* ) => {
694         $crate::dma_read!($($dma).*, $idx, $($field)*)
695     };
696 }
697 
698 /// Writes to a field of an item from an allocated region of structs.
699 ///
700 /// # Examples
701 ///
702 /// ```
703 /// use kernel::device::Device;
704 /// use kernel::dma::{attrs::*, CoherentAllocation};
705 ///
706 /// struct MyStruct { member: u32, }
707 ///
708 /// // SAFETY: All bit patterns are acceptable values for `MyStruct`.
709 /// unsafe impl kernel::transmute::FromBytes for MyStruct{};
710 /// // SAFETY: Instances of `MyStruct` have no uninitialized portions.
711 /// unsafe impl kernel::transmute::AsBytes for MyStruct{};
712 ///
713 /// # fn test(alloc: &kernel::dma::CoherentAllocation<MyStruct>) -> Result {
714 /// kernel::dma_write!(alloc[2].member = 0xf);
715 /// kernel::dma_write!(alloc[1] = MyStruct { member: 0xf });
716 /// # Ok::<(), Error>(()) }
717 /// ```
718 #[macro_export]
719 macro_rules! dma_write {
720     ($dma:ident [ $idx:expr ] $($field:tt)*) => {{
721         $crate::dma_write!($dma, $idx, $($field)*)
722     }};
723     ($($dma:ident).* [ $idx:expr ] $($field:tt)* ) => {{
724         $crate::dma_write!($($dma).*, $idx, $($field)*)
725     }};
726     ($dma:expr, $idx: expr, = $val:expr) => {
727         (|| -> ::core::result::Result<_, $crate::error::Error> {
728             let item = $crate::dma::CoherentAllocation::item_from_index(&$dma, $idx)?;
729             // SAFETY: `item_from_index` ensures that `item` is always a valid item.
730             unsafe { $crate::dma::CoherentAllocation::field_write(&$dma, item, $val) }
731             ::core::result::Result::Ok(())
732         })()
733     };
734     ($dma:expr, $idx: expr, $(.$field:ident)* = $val:expr) => {
735         (|| -> ::core::result::Result<_, $crate::error::Error> {
736             let item = $crate::dma::CoherentAllocation::item_from_index(&$dma, $idx)?;
737             // SAFETY: `item_from_index` ensures that `item` is always a valid pointer and can be
738             // dereferenced. The compiler also further validates the expression on whether `field`
739             // is a member of `item` when expanded by the macro.
740             unsafe {
741                 let ptr_field = ::core::ptr::addr_of_mut!((*item) $(.$field)*);
742                 $crate::dma::CoherentAllocation::field_write(&$dma, ptr_field, $val)
743             }
744             ::core::result::Result::Ok(())
745         })()
746     };
747 }
748