mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: Alice Ryhl <aliceryhl@google.com>
To: Daniel Almeida <daniel.almeida@collabora.com>,
	 Deborah Brouwer <deborah.brouwer@collabora.com>,
	Sami Tolvanen <samitolvanen@google.com>
Cc: "Danilo Krummrich" <dakr@kernel.org>,
	"Matthew Brost" <matthew.brost@intel.com>,
	"Thomas Hellström" <thomas.hellstrom@linux.intel.com>,
	"Miguel Ojeda" <ojeda@kernel.org>,
	"Boqun Feng" <boqun@kernel.org>, "Gary Guo" <gary@garyguo.net>,
	"Björn Roy Baron" <bjorn3_gh@protonmail.com>,
	"Benno Lossin" <lossin@kernel.org>,
	"Andreas Hindborg" <a.hindborg@kernel.org>,
	"Trevor Gross" <tmgross@umich.edu>,
	"Tamir Duberstein" <tamird@kernel.org>,
	"Alexandre Courbot" <acourbot@nvidia.com>,
	"Onur Özkan" <work@onurozkan.dev>,
	"Joerg Roedel (AMD)" <joro@8bytes.org>,
	"Will Deacon" <will@kernel.org>,
	"Robin Murphy" <robin.murphy@arm.com>,
	"Boris Brezillon" <boris.brezillon@collabora.com>,
	dri-devel@lists.freedesktop.org, rust-for-linux@vger.kernel.org,
	linux-kernel@vger.kernel.org, iommu@lists.linux.dev,
	"Alice Ryhl" <aliceryhl@google.com>
Subject: [PATCH 3/4] drm/tyr: vm: expand sub-block unmaps to 2MB boundaries on remap
Date: Tue, 17 Nov 2026 19:32:20 +0000	[thread overview]
Message-ID: <20261117-tyr-remap-2mb-v1-3-2a8efd4c8180@google.com> (raw)
In-Reply-To: <20261117-tyr-remap-2mb-v1-0-2a8efd4c8180@google.com>

From: Sami Tolvanen <samitolvanen@google.com>

Backing GEM shmem with transparent huge pages lets Tyr install 2MB block
entries. arm-lpae refuses to unmap part of a block, so a sparse partial
rebind that splits one in sm_step_remap fails the VM_BIND and poisons
the VM.

Grow such an unmap to the block boundary and rebuild the surviving
fragment, expanding only when contiguous.

Fixes: f2dc32d5a133 ("drm/tyr: add GPU virtual memory (VM) support")
Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
[aliceryhl: tweaks for drm-rust-next and use safe casts instead of 'as']
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
---
 drivers/gpu/drm/tyr/mmu.rs               |  13 +++
 drivers/gpu/drm/tyr/mmu/address_space.rs |  29 ++++++
 drivers/gpu/drm/tyr/vm.rs                | 146 ++++++++++++++++++++++++++++---
 3 files changed, 176 insertions(+), 12 deletions(-)

diff --git a/drivers/gpu/drm/tyr/mmu.rs b/drivers/gpu/drm/tyr/mmu.rs
index 5e88c9e6a4cd..99739a3a8aa1 100644
--- a/drivers/gpu/drm/tyr/mmu.rs
+++ b/drivers/gpu/drm/tyr/mmu.rs
@@ -109,6 +109,19 @@ pub(crate) fn start_vm_update(
         self.as_manager.lock().start_vm_update(vm_as_data, region)
     }
 
+    /// Widens the active lock to cover a larger region.
+    ///
+    /// `region` must contain the currently locked range so the GPU stays
+    /// stalled over it while the update completes. If the region is empty or
+    /// the VM is not resident, this is a NOP.
+    pub(crate) fn extend_vm_update(
+        &self,
+        vm_as_data: &VmAsData<'drm>,
+        region: &Range<u64>,
+    ) -> Result {
+        self.as_manager.lock().extend_vm_update(vm_as_data, region)
+    }
+
     /// Flags the end of a VM update.
     ///
     /// If the VM is resident, this will let GPU accesses on the updated
diff --git a/drivers/gpu/drm/tyr/mmu/address_space.rs b/drivers/gpu/drm/tyr/mmu/address_space.rs
index 76e3d2df6afd..666af82a37b7 100644
--- a/drivers/gpu/drm/tyr/mmu/address_space.rs
+++ b/drivers/gpu/drm/tyr/mmu/address_space.rs
@@ -475,6 +475,35 @@ pub(super) fn start_vm_update(
         }
     }
 
+    /// Widens the active lock to cover a larger region.
+    ///
+    /// Re-issues the LOCKADDR and Lock command over `region`, which the
+    /// caller has computed as the union of the currently locked region and
+    /// the wider range it is about to rebuild. The hardware lock is replaced
+    /// in place, so `region` must contain the previously locked range to keep
+    /// the GPU stalled over it.
+    ///
+    /// If the region is empty or the VM is not resident in a hardware slot,
+    /// this is a no-op.
+    pub(super) fn extend_vm_update(
+        &mut self,
+        vm_as_data: &VmAsData<'drm>,
+        region: &Range<u64>,
+    ) -> Result {
+        if region.is_empty() {
+            return Ok(());
+        }
+
+        let seat = vm_as_data.as_seat.access(self);
+        match seat.slot() {
+            Some(slot) => {
+                let as_nr = usize::from(slot);
+                self.as_start_update(as_nr, region)
+            }
+            _ => Ok(()),
+        }
+    }
+
     /// Completes translation table updates and unlocks the region.
     pub(super) fn end_vm_update(&mut self, vm_as_data: &VmAsData<'drm>) -> Result {
         let seat = vm_as_data.as_seat.access(self);
diff --git a/drivers/gpu/drm/tyr/vm.rs b/drivers/gpu/drm/tyr/vm.rs
index 1dd11b2b1253..bcc864b17428 100644
--- a/drivers/gpu/drm/tyr/vm.rs
+++ b/drivers/gpu/drm/tyr/vm.rs
@@ -42,7 +42,12 @@
     },
     new_mutex,
     prelude::*,
+    ptr::{
+        Alignable,
+        Alignment, //
+    },
     sizes::{
+        SizeConstants,
         SZ_1G,
         SZ_2M,
         SZ_4K, //
@@ -290,6 +295,22 @@ fn return_preallocated_gpuva(&mut self, gpuva: GpuVaAlloc<GpuVmData<'drm>>) {
             *slot = Some(gpuva);
         }
     }
+
+    /// Widens the AS lock to also cover `region`.
+    ///
+    /// A remap can rebuild fragments beyond the originally locked request,
+    /// so the lock is grown to the union of the two before those fragments
+    /// are torn down. The locked region only ever grows, so several remaps
+    /// in one unmap union onto a single widening lock.
+    fn extend_lock(&mut self, region: Range<u64>) -> Result {
+        let union =
+            u64::min(self.region.start, region.start)..u64::max(self.region.end, region.end);
+        if union != self.region {
+            self.mmu.extend_vm_update(self.as_data, &union)?;
+            self.region = union;
+        }
+        Ok(())
+    }
 }
 
 impl Drop for PtUpdateContext<'_, '_> {
@@ -312,6 +333,16 @@ pub(crate) struct GpuVmData<'drm> {
     _phantom: PhantomData<&'drm ()>,
 }
 
+/// Per-mapping private data stored on each `GpuVa`.
+///
+/// A remap splits an existing mapping and may have to rebuild page-table
+/// entries for the surviving fragments. Those fragments keep the protection
+/// of the original mapping, so the protection flags are recorded here.
+pub(crate) struct GpuVaData {
+    /// Memory protection flags of this mapping, as passed to `pt_map`.
+    prot: u32,
+}
+
 /// GPU virtual address space.
 ///
 /// Each VM can be mapped into a hardware address space slot.
@@ -599,7 +630,7 @@ impl<'drm> DriverGpuVm for GpuVmData<'drm> {
     type Driver = TyrDrmDriver;
     type Object = Bo;
     type VmBoData = ();
-    type VaData = ();
+    type VaData = GpuVaData;
     type SmContext<'ctx>
         = PtUpdateContext<'ctx, 'drm>
     where
@@ -701,7 +732,7 @@ fn sm_step_map<'op>(
         }
 
         let gpuva = context.preallocated_gpuva()?;
-        let op = op.insert(gpuva, pin_init::init_zeroed());
+        let op = op.insert(gpuva, GpuVaData { prot });
 
         Ok(op)
     }
@@ -749,10 +780,50 @@ fn sm_step_remap<'op>(
             op.va_to_unmap().addr() + op.va_to_unmap().length()
         };
 
-        let unmap_length = unmap_end - unmap_start;
+        // The surviving fragments inherit the protection of the mapping being
+        // split.
+        let prot = op.va_to_unmap().data_ref().prot;
+
+        let block = Alignment::SZ_2M;
+        let aligned_start = unmap_start.align_down(block);
+        let aligned_end = unmap_end.align_up(block).ok_or(EINVAL)?;
+
+        // An end may only be expanded if the fragment the expansion sweeps and
+        // rebuilds is fully mapped and physically contiguous, because the
+        // rebuild maps it linearly from a single recovered address. A genuine
+        // 2MB block is always contiguous, so this never misses one and the
+        // partial-block unmap can never reach arm-lpae. A fragmented region is
+        // never expanded, so its sub-range unmap stays correct.
+        let head_paddr = match op.prev() {
+            Some(prev) if aligned_start < unmap_start && prev.addr() <= aligned_start => {
+                contiguous_phys(context.pt, aligned_start..unmap_start)
+            }
+            _ => None,
+        };
+        let tail_paddr = match op.next() {
+            Some(next) if aligned_end > unmap_end && next.addr() + next.length() >= aligned_end => {
+                contiguous_phys(context.pt, unmap_end..aligned_end)
+            }
+            _ => None,
+        };
+
+        let region_start = if head_paddr.is_some() {
+            aligned_start
+        } else {
+            unmap_start
+        };
+        let region_end = if tail_paddr.is_some() {
+            aligned_end
+        } else {
+            unmap_end
+        };
+
+        let prev_va = context.preallocated_gpuva()?;
+        let next_va = context.preallocated_gpuva()?;
 
-        if unmap_length > 0 {
-            let region = unmap_start..(unmap_start + unmap_length);
+        if region_end > region_start {
+            let region = region_start..region_end;
+            context.extend_lock(region.clone())?;
             pt_unmap(context.dev, context.pt, region.clone()).inspect_err(|e| {
                 dev_err!(
                     context.dev,
@@ -764,14 +835,29 @@ fn sm_step_remap<'op>(
             })?;
         }
 
-        let prev_va = context.preallocated_gpuva()?;
-        let next_va = context.preallocated_gpuva()?;
+        if let Some(head_paddr) = head_paddr {
+            pt_map(
+                context.dev,
+                context.pt,
+                aligned_start,
+                head_paddr,
+                unmap_start - aligned_start,
+                prot,
+            )?;
+        }
+        if let Some(tail_paddr) = tail_paddr {
+            pt_map(
+                context.dev,
+                context.pt,
+                unmap_end,
+                tail_paddr,
+                aligned_end - unmap_end,
+                prot,
+            )?;
+        }
 
-        let (op_remapped, remap_ret) = op.remap(
-            [prev_va, next_va],
-            pin_init::init_zeroed(),
-            pin_init::init_zeroed(),
-        );
+        let (op_remapped, remap_ret) =
+            op.remap([prev_va, next_va], GpuVaData { prot }, GpuVaData { prot });
 
         if let Some(unused_va) = remap_ret.unused_va {
             context.return_preallocated_gpuva(unused_va);
@@ -781,6 +867,42 @@ fn sm_step_remap<'op>(
     }
 }
 
+/// Returns the base physical address of `range` if every 4KB page in it is
+/// mapped and the whole range is physically contiguous, otherwise `None`.
+///
+/// `range` is a sub-block fragment, so its length is below 2MB and the walk
+/// visits at most 511 pages.
+fn contiguous_phys(pt: &IoPageTable<'_, ARM64LPAES1>, range: Range<u64>) -> Option<u64> {
+    let start_usize = usize::try_from(range.start).ok()?;
+
+    // SAFETY: The page table is exclusively accessed through the
+    // `&mut UniqueRefGpuVm` held under the `gpuvm_unique` mutex for the duration of
+    // the VM update, so no other io-pgtable operation runs concurrently.
+    let base = unsafe { pt.iova_to_phys(start_usize) }?;
+    // Expressly convert to u64 to work with arm 32-bit builds.
+    #[allow(clippy::useless_conversion)]
+    let base = u64::from(base);
+
+    let mut iova = range.start + u64::SZ_4K;
+    while iova < range.end {
+        let offset = iova - range.start;
+        let iova_usize = usize::try_from(iova).ok()?;
+        // SAFETY: The page table is exclusively accessed through the
+        // `&mut UniqueRefGpuVm` held under the `gpuvm_unique` mutex, so no
+        // concurrent io-pgtable operation runs.
+        let paddr = unsafe { pt.iova_to_phys(iova_usize) }?;
+        // Expressly convert to u64 to work with arm 32-bit builds.
+        #[allow(clippy::useless_conversion)]
+        let paddr = u64::from(paddr);
+        if paddr != base + offset {
+            return None;
+        }
+        iova += u64::SZ_4K;
+    }
+
+    Some(base)
+}
+
 /// This function selects the largest supported block size (currently 4KB or 2MB)
 /// that can be used for a mapping at the given address and size, respecting alignment constraints.
 ///

-- 
2.56.1.703.gf848ed807c-goog


  parent reply	other threads:[~2026-11-17 19:32 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-17 19:32 [PATCH 0/4] drm/tyr: fix partial 2MB block unmaps on GPUVM remap Alice Ryhl
2026-11-17 19:32 ` [PATCH 1/4] rust: drm/gpuvm: add GpuVa::data_ref accessor Alice Ryhl
2026-11-17 19:40   ` Danilo Krummrich
2026-11-17 19:32 ` [PATCH 2/4] rust: iommu: pgtable: add iova_to_phys Alice Ryhl
2026-11-17 19:32 ` Alice Ryhl [this message]
2026-11-17 19:32 ` [PATCH 4/4] drm/tyr: vm: check the remaining length in pt_unmap() Alice Ryhl

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20261117-tyr-remap-2mb-v1-3-2a8efd4c8180@google.com \
    --to=aliceryhl@google.com \
    --cc=a.hindborg@kernel.org \
    --cc=acourbot@nvidia.com \
    --cc=bjorn3_gh@protonmail.com \
    --cc=boqun@kernel.org \
    --cc=boris.brezillon@collabora.com \
    --cc=dakr@kernel.org \
    --cc=daniel.almeida@collabora.com \
    --cc=deborah.brouwer@collabora.com \
    --cc=dri-devel@lists.freedesktop.org \
    --cc=gary@garyguo.net \
    --cc=iommu@lists.linux.dev \
    --cc=joro@8bytes.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=lossin@kernel.org \
    --cc=matthew.brost@intel.com \
    --cc=ojeda@kernel.org \
    --cc=robin.murphy@arm.com \
    --cc=rust-for-linux@vger.kernel.org \
    --cc=samitolvanen@google.com \
    --cc=tamird@kernel.org \
    --cc=thomas.hellstrom@linux.intel.com \
    --cc=tmgross@umich.edu \
    --cc=will@kernel.org \
    --cc=work@onurozkan.dev \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®