[PATCH v3 4/8] gpu: nova-core: add NVKV encoder

[Eliot Courtney <ecourtney@nvidia.com>]

Add an encoder for NVKV, which is the wire format for GMCAPI. The
encoded stream is a sequence of 64-bit values. The first 64-bit value
encodes an op word which describes the function of the next N values.

Essentially, the format encodes a sequence of calls to some function
f(key, index, value), where value is a [u8], u32, u64, [u32], or a
[u64]. The key is a u16 and the index is a 12 bit integer. The
interpretation of these function calls is per GMCAPI.

Add tests for the wire encoding for each primitive.

Co-developed-by: Danilo Krummrich <dakr@kernel.org>
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
---
 drivers/gpu/nova-core/gsp.rs             |   1 +
 drivers/gpu/nova-core/gsp/nvkv.rs        | 122 ++++++++++++++++++
 drivers/gpu/nova-core/gsp/nvkv/encode.rs | 210 +++++++++++++++++++++++++++++++
 3 files changed, 333 insertions(+)

diff --git a/drivers/gpu/nova-core/gsp.rs b/drivers/gpu/nova-core/gsp.rs
index 25ea43f1cbe9..7b1bf74464e2 100644
--- a/drivers/gpu/nova-core/gsp.rs
+++ b/drivers/gpu/nova-core/gsp.rs
@@ -24,6 +24,7 @@
 pub(crate) mod cmdq;
 pub(crate) mod commands;
 mod fw;
+mod nvkv;
 mod regs;
 mod sequencer;
 
diff --git a/drivers/gpu/nova-core/gsp/nvkv.rs b/drivers/gpu/nova-core/gsp/nvkv.rs
new file mode 100644
index 000000000000..0957dce92f96
--- /dev/null
+++ b/drivers/gpu/nova-core/gsp/nvkv.rs
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: GPL-2.0
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+
+//! Codec for NVKV, the binary key-value format of GMCAPI.
+//!
+//! Essentially, the format encodes a sequence of calls to some function f(key, index, value),
+//! where value is a [u8], u32, u64, [u32], or a [u64]. The key is a u16 and the index is a 12 bit
+//! integer. The interpretation of these function calls is per GMCAPI. Generally speaking, the
+//! function calls will map to some struct - for example, f(GPU_NAME_STRING_KEY, 0, b"some gpu")
+//! naturally maps to storing a &str with the GPU name.
+
+#![expect(unused_imports)]
+
+use core::ops::Deref;
+
+use kernel::{
+    alloc::{
+        allocator::KVmalloc,
+        Allocator, //
+    },
+    bitfield,
+    num::Bounded,
+    prelude::*, //
+};
+use zerocopy::Immutable;
+
+use crate::bounded_enum;
+
+mod encode;
+pub(crate) use encode::*;
+
+/// The allocator backing [`EncodedStream`].
+type StreamAllocator = KVmalloc;
+
+/// An encoded NVKV byte stream.
+///
+/// # Invariants
+///
+/// The byte length is always a multiple of `size_of::<u64>()`.
+pub(crate) struct EncodedStream(Vec<u8, StreamAllocator>);
+
+impl EncodedStream {
+    /// Creates an empty stream.
+    fn new() -> Self {
+        // INVARIANT: An empty stream's byte length is 0, a multiple of `size_of::<u64>()`.
+        Self(Vec::new())
+    }
+
+    /// Appends a single `u64` to the stream.
+    fn push_u64(&mut self, value: u64) -> Result {
+        // INVARIANT: Appending `size_of::<u64>()` bytes keeps the byte length a multiple of
+        // `size_of::<u64>()`.
+        Ok(self.0.extend_from_slice(&value.to_ne_bytes(), GFP_KERNEL)?)
+    }
+
+    /// Appends `data` as bytes to the stream, zero-padded to a `u64` boundary.
+    fn extend_with_padding<T: IntoBytes + Immutable + ?Sized>(&mut self, data: &T) -> Result {
+        let bytes = data.as_bytes();
+        let padded = bytes.len().next_multiple_of(size_of::<u64>());
+        // Reserve so that a failed allocation can't leave the invariant violated.
+        self.0.reserve(padded, GFP_KERNEL)?;
+        self.0.extend_from_slice(bytes, GFP_KERNEL)?;
+        // INVARIANT: The padding ensures the total length remains a multiple of
+        // `size_of::<u64>()`.
+        Ok(self.0.extend_with(padded - bytes.len(), 0u8, GFP_KERNEL)?)
+    }
+}
+
+// The Deref to `&[u64]` relies on this alignment guarantee.
+static_assert!(align_of::<u64>() <= StreamAllocator::MIN_ALIGN);
+
+impl Deref for EncodedStream {
+    type Target = [u64];
+
+    fn deref(&self) -> &Self::Target {
+        // An empty `Vec`'s pointer isn't necessarily aligned by `StreamAllocator::MIN_ALIGN`.
+        if self.0.is_empty() {
+            return &[];
+        }
+
+        // PANIC: By the type invariants the byte length is a multiple of `size_of::<u64>()`, and
+        // the backing buffer of a non-empty vector has at least `u64` alignment per
+        // `StreamAllocator`'s minimum alignment.
+        <[u64]>::ref_from_bytes(&self.0).expect("EncodedStream invariant violated")
+    }
+}
+
+/// The identifier of an NVKV key.
+pub(crate) type KeyId = u16;
+
+/// The index of an NVKV value.
+pub(crate) type Index = Bounded<u64, 12>;
+
+bitfield! {
+    /// The op word that starts each NVKV operation.
+    struct Op(u64) {
+        15:0 key;
+        27:16 index => Index;
+        31:28 opcode ?=> Opcode;
+        63:32 value;
+    }
+}
+
+bounded_enum! {
+    /// Describes the format of the following NVKV operation.
+    #[derive(Debug, Copy, Clone, PartialEq, Eq)]
+    #[repr(u8)]
+    enum Opcode with TryFrom<Bounded<u64, 4>> {
+        /// A 32-bit value in the op word.
+        Imm32 = 0,
+        /// 32-bit values for consecutive keys, starting at the op word's key.
+        Seq32 = 1,
+        /// 64-bit values for consecutive keys, starting at the op word's key.
+        Seq64 = 2,
+        /// An array of bytes.
+        Array8 = 3,
+        /// An array of 32-bit elements.
+        Array32 = 4,
+        /// An array of 64-bit elements.
+        Array64 = 5,
+    }
+}
diff --git a/drivers/gpu/nova-core/gsp/nvkv/encode.rs b/drivers/gpu/nova-core/gsp/nvkv/encode.rs
new file mode 100644
index 000000000000..8b66ddb15189
--- /dev/null
+++ b/drivers/gpu/nova-core/gsp/nvkv/encode.rs
@@ -0,0 +1,210 @@
+// SPDX-License-Identifier: GPL-2.0
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+
+#![cfg_attr(not(CONFIG_KUNIT), expect(dead_code))]
+
+use kernel::prelude::*;
+
+use super::{
+    EncodedStream,
+    Index,
+    KeyId,
+    Op,
+    Opcode, //
+};
+
+/// An encoder for an NVKV stream.
+pub(crate) struct Encoder {
+    stream: EncodedStream,
+}
+
+impl Encoder {
+    /// Creates an empty encoder.
+    pub(crate) fn new() -> Self {
+        Self {
+            stream: EncodedStream::new(),
+        }
+    }
+
+    /// Returns the encoded data.
+    #[must_use = "encoded stream must be consumed"]
+    pub(crate) fn finish(self) -> EncodedStream {
+        self.stream
+    }
+
+    #[inline]
+    fn encode_op(&mut self, op: Op) -> Result {
+        self.stream.push_u64(op.into_raw())
+    }
+
+    /// Encodes a 32-bit value as an IMM32 pair, with the value in the op word.
+    #[inline]
+    pub(crate) fn encode_u32(&mut self, key: KeyId, index: Index, value: u32) -> Result {
+        // TODO: Consider automatically merging sequential keys.
+        self.encode_op(
+            Op::zeroed()
+                .with_key(key)
+                .with_index(index)
+                .with_opcode(Opcode::Imm32)
+                .with_value(value),
+        )
+    }
+
+    /// Encodes a 64-bit value as a single-element SEQ64 pair.
+    #[inline]
+    pub(crate) fn encode_u64(&mut self, key: KeyId, index: Index, value: u64) -> Result {
+        // TODO: Consider automatically merging sequential keys.
+        const KEY_COUNT: u32 = 1;
+        self.encode_op(
+            Op::zeroed()
+                .with_key(key)
+                .with_index(index)
+                .with_opcode(Opcode::Seq64)
+                .with_value(KEY_COUNT),
+        )?;
+        self.stream.push_u64(value)
+    }
+
+    /// Encodes a byte array as an ARRAY8 pair, zero-padded to a multiple of 8 bytes.
+    #[inline]
+    pub(crate) fn encode_array8(&mut self, key: KeyId, index: Index, array: &[u8]) -> Result {
+        let value_count = u32::try_from(array.len()).map_err(|_| EMSGSIZE)?;
+        self.encode_op(
+            Op::zeroed()
+                .with_key(key)
+                .with_index(index)
+                .with_opcode(Opcode::Array8)
+                .with_value(value_count),
+        )?;
+        self.stream.extend_with_padding(array)
+    }
+
+    /// Encodes a 32-bit array as an ARRAY32 pair, zero-padded to a multiple of 8 bytes.
+    #[inline]
+    pub(crate) fn encode_array32(&mut self, key: KeyId, index: Index, array: &[u32]) -> Result {
+        let value_count = u32::try_from(array.len()).map_err(|_| EMSGSIZE)?;
+        self.encode_op(
+            Op::zeroed()
+                .with_key(key)
+                .with_index(index)
+                .with_opcode(Opcode::Array32)
+                .with_value(value_count),
+        )?;
+        self.stream.extend_with_padding(array)
+    }
+
+    /// Encodes a 64-bit array as an ARRAY64 pair.
+    #[inline]
+    pub(crate) fn encode_array64(&mut self, key: KeyId, index: Index, array: &[u64]) -> Result {
+        let value_count = u32::try_from(array.len()).map_err(|_| EMSGSIZE)?;
+        self.encode_op(
+            Op::zeroed()
+                .with_key(key)
+                .with_index(index)
+                .with_opcode(Opcode::Array64)
+                .with_value(value_count),
+        )?;
+        self.stream.extend_with_padding(array)
+    }
+}
+
+#[kunit_tests(nova_core_nvkv_encode)]
+mod tests {
+    use super::*;
+
+    // Tests that each kind of value is encoded to NVKV wire format properly.
+    #[test]
+    fn encode_all_value_kinds() -> Result {
+        // All keys, indexes, and values are distinct but arbitrary values to make it easier for the
+        // test to catch bugs in the encoded output.
+        const U32_KEY: KeyId = 0x1001;
+        const U64_KEY: KeyId = 0x1002;
+        const ARRAY8_KEY: KeyId = 0x1003;
+        const ARRAY32_KEY: KeyId = 0x1004;
+        const ARRAY64_KEY: KeyId = 0x1005;
+
+        const U32_VALUE: u32 = 0x1111_2222;
+        const U64_VALUE: u64 = 0x3333_4444_5555_6666;
+        const ARRAY8_VALUE: &[u8] = &[0xaa, 0xbb, 0xcc];
+        const ARRAY32_VALUE: &[u32] = &[0xbbbb_cccc, 0xdddd_eeee];
+        const ARRAY64_VALUE: &[u64] = &[0x0123_4567_89ab_cdef, 0xfedc_ba98_7654_3210];
+
+        let mut encoder = Encoder::new();
+        encoder.encode_u32(U32_KEY, Index::new::<0>(), U32_VALUE)?;
+        encoder.encode_u64(U64_KEY, Index::new::<1>(), U64_VALUE)?;
+        encoder.encode_array8(ARRAY8_KEY, Index::new::<2>(), ARRAY8_VALUE)?;
+        encoder.encode_array32(ARRAY32_KEY, Index::new::<3>(), ARRAY32_VALUE)?;
+        encoder.encode_array64(ARRAY64_KEY, Index::new::<4>(), ARRAY64_VALUE)?;
+
+        let encoded = encoder.finish();
+        assert_eq!(encoded.len(), 10);
+
+        // IMM32 has its value in the op word.
+        assert_eq!(
+            encoded[0],
+            Op::zeroed()
+                .with_key(U32_KEY)
+                .with_index(Index::new::<0>())
+                .with_opcode(Opcode::Imm32)
+                .with_value(U32_VALUE)
+                .into_raw()
+        );
+
+        // The SEQ64 op word followed by the value.
+        assert_eq!(
+            encoded[1],
+            Op::zeroed()
+                .with_key(U64_KEY)
+                .with_index(Index::new::<1>())
+                .with_opcode(Opcode::Seq64)
+                .with_value(1u32)
+                .into_raw()
+        );
+        assert_eq!(encoded[2], U64_VALUE);
+
+        // The ARRAY8 op word has the byte count. The bytes follow, padded out to a whole word.
+        assert_eq!(
+            encoded[3],
+            Op::zeroed()
+                .with_key(ARRAY8_KEY)
+                .with_index(Index::new::<2>())
+                .with_opcode(Opcode::Array8)
+                .with_value(3u32)
+                .into_raw()
+        );
+        assert_eq!(
+            encoded[4],
+            u64::from_le_bytes([0xaa, 0xbb, 0xcc, 0, 0, 0, 0, 0])
+        );
+
+        // The ARRAY32 op word has the element count. The two elements follow in little endian.
+        assert_eq!(
+            encoded[5],
+            Op::zeroed()
+                .with_key(ARRAY32_KEY)
+                .with_index(Index::new::<3>())
+                .with_opcode(Opcode::Array32)
+                .with_value(2u32)
+                .into_raw()
+        );
+        assert_eq!(
+            encoded[6],
+            u64::from(ARRAY32_VALUE[1]) << 32 | u64::from(ARRAY32_VALUE[0])
+        );
+
+        // The ARRAY64 op word has the element count with the two elements after.
+        assert_eq!(
+            encoded[7],
+            Op::zeroed()
+                .with_key(ARRAY64_KEY)
+                .with_index(Index::new::<4>())
+                .with_opcode(Opcode::Array64)
+                .with_value(2u32)
+                .into_raw()
+        );
+        assert_eq!(encoded[8], ARRAY64_VALUE[0]);
+        assert_eq!(encoded[9], ARRAY64_VALUE[1]);
+
+        Ok(())
+    }
+}

-- 
2.55.0