[PATCH v2 2/3] gpu: nova-core: Hopper: use correct sysmem flush registers

[John Hubbard <jhubbard@nvidia.com>]

Hopper has its own FBHUB sysmem flush page registers, but the Hopper
framebuffer HAL delegates to the Ampere NISO path, which encodes the
address with an 8-bit right-shift. That programs the wrong value into
the wrong registers, so the GPU's sysmembar flush targets the wrong
system memory address.

Add Hopper's FBHUB flush registers and program them directly from the
Hopper HAL.

This has not yet been tested on real Hopper hardware (that's true for
nova-core in general).

Signed-off-by: John Hubbard <jhubbard@nvidia.com>
---
 drivers/gpu/nova-core/fb/hal/gh100.rs | 31 ++++++++++++++++++++++++---
 drivers/gpu/nova-core/regs.rs         | 19 ++++++++++++++++
 2 files changed, 47 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/nova-core/fb/hal/gh100.rs b/drivers/gpu/nova-core/fb/hal/gh100.rs
index 5450c7254dad..d39fe99537ed 100644
--- a/drivers/gpu/nova-core/fb/hal/gh100.rs
+++ b/drivers/gpu/nova-core/fb/hal/gh100.rs
@@ -2,24 +2,49 @@
 // SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 
 use kernel::{
+    io::Io,
+    num::Bounded,
     prelude::*,
     sizes::SizeConstants, //
 };
 
 use crate::{
     driver::Bar0,
-    fb::hal::FbHal, //
+    fb::hal::FbHal,
+    regs, //
 };
 
 struct Gh100;
 
+fn read_sysmem_flush_page_gh100(bar: Bar0<'_>) -> u64 {
+    let lo = u64::from(bar.read(regs::NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_LO).adr());
+    let hi = u64::from(bar.read(regs::NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_HI).adr());
+
+    (hi << 32) | lo
+}
+
+/// Write the sysmem flush page address through the Hopper FBHUB registers.
+fn write_sysmem_flush_page_gh100(bar: Bar0<'_>, addr: Bounded<u64, 52>) {
+    // Write HI first. The hardware will trigger the flush on the LO write.
+    bar.write_reg(
+        regs::NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_HI::zeroed()
+            .with_adr(addr.shr::<32, 20>().cast::<u32>()),
+    );
+    bar.write_reg(
+        // CAST: lower 32 bits. Hardware ignores bits 7:0.
+        regs::NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_LO::zeroed().with_adr(*addr as u32),
+    );
+}
+
 impl FbHal for Gh100 {
     fn read_sysmem_flush_page(&self, bar: Bar0<'_>) -> u64 {
-        super::ga100::read_sysmem_flush_page_ga100(bar)
+        read_sysmem_flush_page_gh100(bar)
     }
 
     fn write_sysmem_flush_page(&self, bar: Bar0<'_>, addr: u64) -> Result {
-        super::ga100::write_sysmem_flush_page_ga100(bar, addr);
+        let addr = Bounded::<u64, 52>::try_new(addr).ok_or(EINVAL)?;
+
+        write_sysmem_flush_page_gh100(bar, addr);
 
         Ok(())
     }
diff --git a/drivers/gpu/nova-core/regs.rs b/drivers/gpu/nova-core/regs.rs
index 5ab7ccfb9855..7982778fd6cb 100644
--- a/drivers/gpu/nova-core/regs.rs
+++ b/drivers/gpu/nova-core/regs.rs
@@ -189,6 +189,25 @@ fn fmt(&self, f: &mut kernel::fmt::Formatter<'_>) -> kernel::fmt::Result {
     }
 }
 
+register! {
+    /// Low bits of the physical system memory address used by the GPU to perform
+    /// sysmembar operations on Hopper.
+    ///
+    /// Like the GB20x FBHUB0 registers, and unlike the Ampere
+    /// `NV_PFB_NISO_FLUSH_SYSMEM_ADDR` registers (which encode the address with an
+    /// 8-bit right-shift), these take the raw address split into lower and upper
+    /// halves. Hardware ignores bits 7:0 of the LO register.
+    pub(crate) NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_LO(u32) @ 0x00100a34 {
+        31:0    adr => u32;
+    }
+
+    /// High bits of the physical system memory address used by the GPU to perform
+    /// sysmembar operations on Hopper.
+    pub(crate) NV_PFB_FBHUB_PCIE_FLUSH_SYSMEM_ADDR_HI(u32) @ 0x00100a38 {
+        19:0    adr;
+    }
+}
+
 impl NV_PFB_PRI_MMU_LOCAL_MEMORY_RANGE {
     /// Returns the usable framebuffer size, in bytes.
     pub(crate) fn usable_fb_size(self) -> u64 {
-- 
2.54.0