Anforderungen  |   Konzepte  |   Entwurf  |   Entwicklung  |   Qualitätssicherung  |   Lebenszyklus  |   Steuerung
 
 
 
 


Quelle  freelist.rs   Sprache: unbekannt

 
use {
    crate::{
        align_down, align_up,
        error::AllocationError,
        heap::Heap,
        util::{arc_unwrap, is_arc_unique},
        MemoryBounds,
    },
    alloc::{sync::Arc, vec::Vec},
    core::{cmp::Ordering, ptr::NonNull},
    gpu_alloc_types::{AllocationFlags, DeviceMapError, MemoryDevice, MemoryPropertyFlags},
};

unsafe fn opt_ptr_add(ptr: Option<NonNull<u8>>, size: u64) -> Option<NonNull<u8>> {
    ptr.map(|ptr| {
        // Size is within memory region started at `ptr`.
        // size is within `chunk_size` that fits `isize`.
        NonNull::new_unchecked(ptr.as_ptr().offset(size as isize))
    })
}

#[derive(Debug)]
pub(super) struct FreeList<M> {
    array: Vec<FreeListRegion<M>>,
    counter: u64,
}

impl<M> FreeList<M> {
    pub fn new() -> Self {
        FreeList {
            array: Vec::new(),
            counter: 0,
        }
    }

    pub fn get_block_from_new_memory(
        &mut self,
        memory: Arc<M>,
        memory_size: u64,
        ptr: Option<NonNull<u8>>,
        align_mask: u64,
        size: u64,
    ) -> FreeListBlock<M> {
        debug_assert!(size <= memory_size);

        self.counter += 1;
        self.array.push(FreeListRegion {
            memory,
            ptr,
            chunk: self.counter,
            start: 0,
            end: memory_size,
        });
        self.get_block_at(self.array.len() - 1, align_mask, size)
    }

    pub fn get_block(&mut self, align_mask: u64, size: u64) -> Option<FreeListBlock<M>> {
        let (index, _) = self.array.iter().enumerate().rev().find(|(_, region)| {
            match region.end.checked_sub(size) {
                Some(start) => {
                    let aligned_start = align_down(start, align_mask);
                    aligned_start >= region.start
                }
                None => false,
            }
        })?;

        Some(self.get_block_at(index, align_mask, size))
    }

    fn get_block_at(&mut self, index: usize, align_mask: u64, size: u64) -> FreeListBlock<M> {
        let region = &mut self.array[index];

        let start = region.end - size;
        let aligned_start = align_down(start, align_mask);

        if aligned_start > region.start {
            let block = FreeListBlock {
                offset: aligned_start,
                size: region.end - aligned_start,
                chunk: region.chunk,
                ptr: unsafe { opt_ptr_add(region.ptr, aligned_start - region.start) },
                memory: region.memory.clone(),
            };

            region.end = aligned_start;

            block
        } else {
            debug_assert_eq!(aligned_start, region.start);
            let region = self.array.remove(index);
            region.into_block()
        }
    }

    pub fn insert_block(&mut self, block: FreeListBlock<M>) {
        match self.array.binary_search_by(|b| b.cmp(&block)) {
            Ok(_) => {
                panic!("Overlapping block found in free list");
            }
            Err(index) if self.array.len() > index => match &mut self.array[..=index] {
                [] => unreachable!(),
                [next] => {
                    debug_assert!(!next.is_suffix_block(&block));

                    if next.is_prefix_block(&block) {
                        next.merge_prefix_block(block);
                    } else {
                        self.array.insert(0, FreeListRegion::from_block(block));
                    }
                }
                [.., prev, next] => {
                    debug_assert!(!prev.is_prefix_block(&block));
                    debug_assert!(!next.is_suffix_block(&block));

                    if next.is_prefix_block(&block) {
                        next.merge_prefix_block(block);

                        if prev.consecutive(&*next) {
                            let next = self.array.remove(index);
                            let prev = &mut self.array[index - 1];
                            prev.merge(next);
                        }
                    } else if prev.is_suffix_block(&block) {
                        prev.merge_suffix_block(block);
                    } else {
                        self.array.insert(index, FreeListRegion::from_block(block));
                    }
                }
            },
            Err(_) => match &mut self.array[..] {
                [] => self.array.push(FreeListRegion::from_block(block)),
                [.., prev] => {
                    debug_assert!(!prev.is_prefix_block(&block));
                    if prev.is_suffix_block(&block) {
                        prev.merge_suffix_block(block);
                    } else {
                        self.array.push(FreeListRegion::from_block(block));
                    }
                }
            },
        }
    }

    pub fn drain(&mut self, keep_last: bool) -> Option<impl Iterator<Item = (M, u64)> + '_> {
        // Time to deallocate

        let len = self.array.len();

        let mut del = 0;
        {
            let regions = &mut self.array[..];

            for i in 0..len {
                if (i < len - 1 || !keep_last) && is_arc_unique(&mut regions[i].memory) {
                    del += 1;
                } else if del > 0 {
                    regions.swap(i - del, i);
                }
            }
        }

        if del > 0 {
            Some(self.array.drain(len - del..).map(move |region| {
                debug_assert_eq!(region.start, 0);
                (unsafe { arc_unwrap(region.memory) }, region.end)
            }))
        } else {
            None
        }
    }
}

#[derive(Debug)]
struct FreeListRegion<M> {
    memory: Arc<M>,
    ptr: Option<NonNull<u8>>,
    chunk: u64,
    start: u64,
    end: u64,
}

unsafe impl<M> Sync for FreeListRegion<M> where M: Sync {}
unsafe impl<M> Send for FreeListRegion<M> where M: Send {}

impl<M> FreeListRegion<M> {
    pub fn cmp(&self, block: &FreeListBlock<M>) -> Ordering {
        debug_assert_eq!(
            Arc::ptr_eq(&self.memory, &block.memory),
            self.chunk == block.chunk
        );

        if self.chunk == block.chunk {
            debug_assert_eq!(
                Ord::cmp(&self.start, &block.offset),
                Ord::cmp(&self.end, &(block.offset + block.size)),
                "Free region {{ start: {}, end: {} }} overlaps with block {{ offset: {}, size: {} }}",
                self.start,
                self.end,
                block.offset,
                block.size,
            );
        }

        Ord::cmp(&self.chunk, &block.chunk).then(Ord::cmp(&self.start, &block.offset))
    }

    fn from_block(block: FreeListBlock<M>) -> Self {
        FreeListRegion {
            memory: block.memory,
            chunk: block.chunk,
            ptr: block.ptr,
            start: block.offset,
            end: block.offset + block.size,
        }
    }

    fn into_block(self) -> FreeListBlock<M> {
        FreeListBlock {
            memory: self.memory,
            chunk: self.chunk,
            ptr: self.ptr,
            offset: self.start,
            size: self.end - self.start,
        }
    }

    fn consecutive(&self, other: &Self) -> bool {
        if self.chunk != other.chunk {
            return false;
        }

        debug_assert!(Arc::ptr_eq(&self.memory, &other.memory));

        debug_assert_eq!(
            Ord::cmp(&self.start, &other.start),
            Ord::cmp(&self.end, &other.end)
        );

        self.end == other.start
    }

    fn merge(&mut self, next: FreeListRegion<M>) {
        debug_assert!(self.consecutive(&next));
        self.end = next.end;
    }

    fn is_prefix_block(&self, block: &FreeListBlock<M>) -> bool {
        if self.chunk != block.chunk {
            return false;
        }

        debug_assert!(Arc::ptr_eq(&self.memory, &block.memory));

        debug_assert_eq!(
            Ord::cmp(&self.start, &block.offset),
            Ord::cmp(&self.end, &(block.offset + block.size))
        );

        self.start == (block.offset + block.size)
    }

    fn merge_prefix_block(&mut self, block: FreeListBlock<M>) {
        debug_assert!(self.is_prefix_block(&block));
        self.start = block.offset;
        self.ptr = block.ptr;
    }

    fn is_suffix_block(&self, block: &FreeListBlock<M>) -> bool {
        if self.chunk != block.chunk {
            return false;
        }

        debug_assert!(Arc::ptr_eq(&self.memory, &block.memory));

        debug_assert_eq!(
            Ord::cmp(&self.start, &block.offset),
            Ord::cmp(&self.end, &(block.offset + block.size))
        );

        self.end == block.offset
    }

    fn merge_suffix_block(&mut self, block: FreeListBlock<M>) {
        debug_assert!(self.is_suffix_block(&block));
        self.end += block.size;
    }
}

#[derive(Debug)]
pub struct FreeListBlock<M> {
    pub memory: Arc<M>,
    pub ptr: Option<NonNull<u8>>,
    pub chunk: u64,
    pub offset: u64,
    pub size: u64,
}

unsafe impl<M> Sync for FreeListBlock<M> where M: Sync {}
unsafe impl<M> Send for FreeListBlock<M> where M: Send {}

#[derive(Debug)]
pub(crate) struct FreeListAllocator<M> {
    freelist: FreeList<M>,
    chunk_size: u64,
    final_chunk_size: u64,
    memory_type: u32,
    props: MemoryPropertyFlags,
    atom_mask: u64,

    total_allocations: u64,
    total_deallocations: u64,
}

impl<M> Drop for FreeListAllocator<M> {
    fn drop(&mut self) {
        match Ord::cmp(&self.total_allocations, &self.total_deallocations) {
            Ordering::Equal => {}
            Ordering::Greater => {
                report_error_on_drop!("Not all blocks were deallocated")
            }
            Ordering::Less => {
                report_error_on_drop!("More blocks deallocated than allocated")
            }
        }

        if !self.freelist.array.is_empty() {
            report_error_on_drop!(
                "FreeListAllocator has free blocks on drop. Allocator should be cleaned"
            );
        }
    }
}

impl<M> FreeListAllocator<M>
where
    M: MemoryBounds + 'static,
{
    pub fn new(
        starting_chunk_size: u64,
        final_chunk_size: u64,
        memory_type: u32,
        props: MemoryPropertyFlags,
        atom_mask: u64,
    ) -> Self {
        debug_assert_eq!(
            align_down(starting_chunk_size, atom_mask),
            starting_chunk_size
        );

        let starting_chunk_size = min(starting_chunk_size, isize::max_value());

        debug_assert_eq!(align_down(final_chunk_size, atom_mask), final_chunk_size);
        let final_chunk_size = min(final_chunk_size, isize::max_value());

        FreeListAllocator {
            freelist: FreeList::new(),
            chunk_size: starting_chunk_size,
            final_chunk_size,
            memory_type,
            props,
            atom_mask,

            total_allocations: 0,
            total_deallocations: 0,
        }
    }

    #[cfg_attr(feature = "tracing", tracing::instrument(skip(self, device)))]
    pub unsafe fn alloc(
        &mut self,
        device: &impl MemoryDevice<M>,
        size: u64,
        align_mask: u64,
        flags: AllocationFlags,
        heap: &mut Heap,
        allocations_remains: &mut u32,
    ) -> Result<FreeListBlock<M>, AllocationError> {
        debug_assert!(
            self.final_chunk_size >= size,
            "GpuAllocator must not request allocations equal or greater to chunks size"
        );

        let size = align_up(size, self.atom_mask).expect(
            "Any value not greater than final chunk size (which is aligned) has to fit for alignment",
        );

        let align_mask = align_mask | self.atom_mask;
        let host_visible = self.host_visible();

        if size <= self.chunk_size {
            // Otherwise there can't be any sufficiently large free blocks
            if let Some(block) = self.freelist.get_block(align_mask, size) {
                self.total_allocations += 1;
                return Ok(block);
            }
        }

        // New allocation is required.
        if *allocations_remains == 0 {
            return Err(AllocationError::TooManyObjects);
        }

        if size > self.chunk_size {
            let multiple = (size - 1) / self.chunk_size + 1;
            let multiple = multiple.next_power_of_two();

            self.chunk_size = (self.chunk_size * multiple).min(self.final_chunk_size);
        }

        let mut memory = device.allocate_memory(self.chunk_size, self.memory_type, flags)?;
        *allocations_remains -= 1;
        heap.alloc(self.chunk_size);

        // Map host visible allocations
        let ptr = if host_visible {
            match device.map_memory(&mut memory, 0, self.chunk_size) {
                Ok(ptr) => Some(ptr),
                Err(DeviceMapError::MapFailed) => {
                    #[cfg(feature = "tracing")]
                    tracing::error!("Failed to map host-visible memory in linear allocator");
                    device.deallocate_memory(memory);
                    *allocations_remains += 1;
                    heap.dealloc(self.chunk_size);

                    return Err(AllocationError::OutOfHostMemory);
                }
                Err(DeviceMapError::OutOfDeviceMemory) => {
                    return Err(AllocationError::OutOfDeviceMemory);
                }
                Err(DeviceMapError::OutOfHostMemory) => {
                    return Err(AllocationError::OutOfHostMemory);
                }
            }
        } else {
            None
        };

        let memory = Arc::new(memory);
        let block =
            self.freelist
                .get_block_from_new_memory(memory, self.chunk_size, ptr, align_mask, size);

        if self.chunk_size < self.final_chunk_size {
            // Double next chunk size
            // Limit to final value.
            self.chunk_size = (self.chunk_size * 2).min(self.final_chunk_size);
        }

        self.total_allocations += 1;
        Ok(block)
    }

    #[cfg_attr(feature = "tracing", tracing::instrument(skip(self, device)))]
    pub unsafe fn dealloc(
        &mut self,
        device: &impl MemoryDevice<M>,
        block: FreeListBlock<M>,
        heap: &mut Heap,
        allocations_remains: &mut u32,
    ) {
        debug_assert!(block.size < self.chunk_size);
        debug_assert_ne!(block.size, 0);
        self.freelist.insert_block(block);
        self.total_deallocations += 1;

        if let Some(memory) = self.freelist.drain(true) {
            memory.for_each(|(memory, size)| {
                device.deallocate_memory(memory);
                *allocations_remains += 1;
                heap.dealloc(size);
            });
        }
    }

    /// Deallocates leftover memory objects.
    /// Should be used before dropping.
    ///
    /// # Safety
    ///
    /// * `device` must be one with `DeviceProperties` that were provided to create this `GpuAllocator` instance
    /// * Same `device` instance must be used for all interactions with one `GpuAllocator` instance
    ///   and memory blocks allocated from it
    #[cfg_attr(feature = "tracing", tracing::instrument(skip(self, device)))]
    pub unsafe fn cleanup(
        &mut self,
        device: &impl MemoryDevice<M>,
        heap: &mut Heap,
        allocations_remains: &mut u32,
    ) {
        if let Some(memory) = self.freelist.drain(false) {
            memory.for_each(|(memory, size)| {
                device.deallocate_memory(memory);
                *allocations_remains += 1;
                heap.dealloc(size);
            });
        }

        #[cfg(feature = "tracing")]
        {
            if self.total_allocations == self.total_deallocations && !self.freelist.array.is_empty()
            {
                tracing::error!(
                    "Some regions were not deallocated on cleanup, although all blocks are free.
                        This is a bug in `FreeBlockAllocator`.
                        See array of free blocks left:
                        {:#?}",
                    self.freelist.array,
                );
            }
        }
    }

    fn host_visible(&self) -> bool {
        self.props.contains(MemoryPropertyFlags::HOST_VISIBLE)
    }
}

fn min<L, R>(l: L, r: R) -> L
where
    R: core::convert::TryInto<L>,
    L: Ord,
{
    match r.try_into() {
        Ok(r) => core::cmp::min(l, r),
        Err(_) => l,
    }
}

[ Dauer der Verarbeitung: 0.13 Sekunden  (vorverarbeitet)  ]

                                                                                                                                                                                                                                                                                                                                                                                                     


Neuigkeiten

     Aktuelles
     Motto des Tages

Software

     Produkte
     Quellcodebibliothek

Aktivitäten

     Artikel über Sicherheit
     Anleitung zur Aktivierung von SSL

Muße

     Gedichte
     Musik
     Bilder

Jenseits des Üblichen ....

Besucherstatistik

Besucherstatistik

Monitoring

Montastic status badge