Skip to content

asahi: Improve VM bind performance for large BOs #495

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
TrungNguyen1909 wants to merge 1 commit into
base: asahi
Choose a base branch
from
+82 −33
Open

asahi: Improve VM bind performance for large BOs #495

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
115 changes: 82 additions & 33 deletions drivers/gpu/drm/asahi/mmu.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,8 @@ use core::sync::atomic::{
Ordering, //
};

use core::cmp;

use kernel::{
addr::PhysicalAddr,
bindings::drm_gpuvm_flags_DRM_GPUVM_IMMEDIATE_MODE,
Expand Down Expand Up @@ -194,17 +196,25 @@ struct VmBinding {
ttb: u64,
}

struct VmBoInner {
sgt: Option<shmem::SGTable<gem::AsahiObject>>,
sg_vec: Option<KVVec<(usize, Range<usize>)>>,
}

/// Data associated with a VM <=> BO pairing
#[pin_data]
struct VmBo {
#[pin]
sgt: Mutex<Option<shmem::SGTable<gem::AsahiObject>>>,
inner: Mutex<VmBoInner>,
}

impl gpuvm::DriverGpuVmBo for VmBo {
fn new() -> impl PinInit<Self> {
pin_init!(VmBo {
sgt <- new_mutex!(None, "VmBinding"),
inner <- new_mutex!(VmBoInner {
sgt: None,
sg_vec: None,
}, "VmBinding"),
})
}
}
Expand Down Expand Up @@ -236,28 +246,21 @@ impl gpuvm::DriverGpuVm for VmInner {

let one_page = op.flags().contains(gpuvm::GpuVaFlags::REPEAT);

let guard = bo.inner().sgt.lock();
for range in guard.as_ref().expect("step_map with no SGT").iter() {
// TODO: proper DMA address/length handling
let mut addr = range.dma_address() as usize;
let mut len: usize = range.dma_len() as usize;

let mut do_map = |mut addr: usize, mut len: usize, offset: &mut usize| -> Result<bool> {
if left == 0 {
break;
return Ok(false);
}

if offset > 0 {
let skip = len.min(offset);
if *offset > 0 {
let skip = len.min(*offset);
addr += skip;
len -= skip;
offset -= skip;
*offset -= skip;
}

if len == 0 {
continue;
return Ok(true);
}

assert!(offset == 0);
assert!(*offset == 0);

if one_page {
len = left;
Expand All @@ -283,6 +286,39 @@ impl gpuvm::DriverGpuVm for VmInner {

left -= len;
iova += len as u64;
Ok(true)
};

let guard = bo.inner().inner.lock();
if let Some(sg_vec) = guard.sg_vec.as_ref() {
let start_idx = sg_vec.binary_search_by(|range| {
if range.0 > offset {
cmp::Ordering::Greater
} else if (range.0 + range.1.len()) <= offset {
cmp::Ordering::Less
} else {
cmp::Ordering::Equal
}
}).expect("sg_vec does not contain offset???");

offset -= sg_vec[start_idx].0 as usize;

for cur in start_idx..sg_vec.len() {
let addr = sg_vec[cur].1.start as usize;
let len: usize = sg_vec[cur].1.len() as usize;
if do_map(addr, len, &mut offset)? == false {
break;
}
}
} else {
for range in guard.sgt.as_ref().expect("step_map with no SGT").iter() {
// TODO: proper DMA address/length handling
let addr = range.dma_address() as usize;
let len: usize = range.dma_len() as usize;
if do_map(addr, len, &mut offset)? == false {
break;
}
}
}

let gpuva = ctx.new_va.take().expect("Multiple step_map calls");
Expand Down Expand Up @@ -446,8 +482,8 @@ impl VmInner {
/// Map an `mm::Node` representing an mapping in VA space.
fn map_node(&mut self, node: &mm::Node<(), KernelMappingInner>, prot: Prot) -> Result {
let mut iova = node.start();
let guard = node.bo.as_ref().ok_or(EINVAL)?.inner().sgt.lock();
let sgt = guard.as_ref().ok_or(EINVAL)?;
let guard = node.bo.as_ref().ok_or(EINVAL)?.inner().inner.lock();
let sgt = guard.sgt.as_ref().ok_or(EINVAL)?;
let mut offset = node.offset;
let mut left = node.mapped_size;

Expand Down Expand Up @@ -1052,9 +1088,9 @@ impl Vm {
let mut inner = self.inner.exec_lock(Some(gem), false)?;
let vm_bo = self.inner.obtain_bo(gem)?;

let mut vm_bo_guard = vm_bo.inner().sgt.lock();
if vm_bo_guard.is_none() {
vm_bo_guard.replace(sgt);
let mut vm_bo_guard = vm_bo.inner().inner.lock();
if vm_bo_guard.sgt.is_none() {
vm_bo_guard.sgt.replace(sgt);
}
core::mem::drop(vm_bo_guard);

Expand Down Expand Up @@ -1100,9 +1136,9 @@ impl Vm {

let vm_bo = self.inner.obtain_bo(&gem)?;

let mut vm_bo_guard = vm_bo.inner().sgt.lock();
if vm_bo_guard.is_none() {
vm_bo_guard.replace(sgt);
let mut vm_bo_guard = vm_bo.inner().inner.lock();
if vm_bo_guard.sgt.is_none() {
vm_bo_guard.sgt.replace(sgt);
}
core::mem::drop(vm_bo_guard);

Expand Down Expand Up @@ -1150,20 +1186,33 @@ impl Vm {
..Default::default()
};

let sgt = gem.owned_sg_table()?;
let vm_bo = self.inner.obtain_bo(gem)?;
{
let mut vm_bo_guard = vm_bo.inner().inner.lock();
if vm_bo_guard.sgt.is_none() {
let sgt = gem.owned_sg_table()?;

if vm_bo_guard.sg_vec.is_none() {
let mut sg_vec = KVVec::new();
let mut offset = 0;
for range in sgt.iter() {
let addr = range.dma_address() as usize;
let len = range.dma_len() as usize;
sg_vec.push((offset, addr..(addr + len)), GFP_KERNEL)?;
offset += len;
}
vm_bo_guard.sg_vec.replace(sg_vec);
}
vm_bo_guard.sgt.replace(sgt);
}
core::mem::drop(vm_bo_guard);
}

let mut inner = self.inner.exec_lock(Some(gem), true)?;

// Preallocate the page tables, to fail early if we ENOMEM
inner.page_table.alloc_pages(addr..(addr + size))?;

let vm_bo = self.inner.obtain_bo(gem)?;

let mut vm_bo_guard = vm_bo.inner().sgt.lock();
if vm_bo_guard.is_none() {
vm_bo_guard.replace(sgt);
}
core::mem::drop(vm_bo_guard);

ctx.vm_bo = Some(vm_bo);

if (addr | size | offset) & (UAT_PGMSK as u64) != 0 {
Expand Down