Splits large RPCs if necessary and sends the remaining parts using
continuation records. RPCs that do not need continuation records
continue to write directly into the command buffer. Ones that do write
into a staging buffer first, so there is one copy.
Continuation record for receive is not necessary to support at the
moment because those replies do not need to be read and are currently
drained by retrying `receive_msg` on ERANGE.
Signed-off-by: Eliot Courtney <ecourtney@nvidia.com>
---
drivers/gpu/nova-core/gsp.rs | 1 +
drivers/gpu/nova-core/gsp/cmdq.rs | 41 +++++++-
drivers/gpu/nova-core/gsp/continuation.rs | 167 ++++++++++++++++++++++++++++++
drivers/gpu/nova-core/gsp/fw.rs | 4 +
4 files changed, 210 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/nova-core/gsp.rs b/drivers/gpu/nova-core/gsp.rs
index 174feaca0a6b..ccf56f1ad246 100644
--- a/drivers/gpu/nova-core/gsp.rs
+++ b/drivers/gpu/nova-core/gsp.rs
@@ -16,6 +16,7 @@
pub(crate) mod cmdq;
pub(crate) mod commands;
+mod continuation;
mod fw;
mod sequencer;
diff --git a/drivers/gpu/nova-core/gsp/cmdq.rs b/drivers/gpu/nova-core/gsp/cmdq.rs
index 4a663a5b3437..4caf5917de64 100644
--- a/drivers/gpu/nova-core/gsp/cmdq.rs
+++ b/drivers/gpu/nova-core/gsp/cmdq.rs
@@ -28,6 +28,10 @@
use crate::{
driver::Bar0,
gsp::{
+ continuation::{
+ ContinuationRecord,
+ SplitState, //
+ },
fw::{
GspMsgElement,
MsgFunction,
@@ -452,7 +456,7 @@ struct GspMessage<'a> {
/// Computes the total size of the command (including its variable-length payload) without the
/// [`GspMsgElement`] header.
-fn command_size<M>(command: &M) -> usize
+pub(crate) fn command_size<M>(command: &M) -> usize
where
M: CommandToGsp,
{
@@ -520,7 +524,7 @@ fn notify_gsp(bar: &Bar0) {
.write(bar);
}
- /// Sends `command` to the GSP.
+ /// Sends `command` to the GSP, without splitting it.
///
/// # Errors
///
@@ -529,7 +533,7 @@ fn notify_gsp(bar: &Bar0) {
/// written to by its [`CommandToGsp::init_variable_payload`] method.
///
/// Error codes returned by the command initializers are propagated as-is.
- pub(crate) fn send_command<M>(&mut self, bar: &Bar0, command: M) -> Result
+ fn send_single_command<M>(&mut self, bar: &Bar0, command: M) -> Result
where
M: CommandToGsp,
// This allows all error types, including `Infallible`, to be used for `M::InitError`.
@@ -588,6 +592,37 @@ pub(crate) fn send_command<M>(&mut self, bar: &Bar0, command: M) -> Result
Ok(())
}
+ /// Sends `command` to the GSP.
+ ///
+ /// The command may be split into multiple messages if it is large.
+ ///
+ /// # Errors
+ ///
+ /// - `ETIMEDOUT` if space does not become available within the timeout.
+ /// - `EIO` if the variable payload requested by the command has not been entirely
+ /// written to by its [`CommandToGsp::init_variable_payload`] method.
+ ///
+ /// Error codes returned by the command initializers are propagated as-is.
+ pub(crate) fn send_command<M>(&mut self, bar: &Bar0, command: M) -> Result
+ where
+ M: CommandToGsp,
+ Error: From<M::InitError>,
+ {
+ match SplitState::new(command)? {
+ SplitState::Single(command) => self.send_single_command(bar, command),
+ SplitState::Split(command, mut continuations) => {
+ self.send_single_command(bar, command)?;
+
+ while let Some(continuation) = continuations.next() {
+ // Turbofish needed because the compiler cannot infer M here.
+ self.send_single_command::<ContinuationRecord<'_>>(bar, continuation)?;
+ }
+
+ Ok(())
+ }
+ }
+ }
+
/// Wait for a message to become available on the message queue.
///
/// This works purely at the transport layer and does not interpret or validate the message
diff --git a/drivers/gpu/nova-core/gsp/continuation.rs b/drivers/gpu/nova-core/gsp/continuation.rs
new file mode 100644
index 000000000000..cf6d950a0e3f
--- /dev/null
+++ b/drivers/gpu/nova-core/gsp/continuation.rs
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: GPL-2.0
+
+//! Support for splitting large GSP commands across continuation records.
+
+use core::convert::Infallible;
+
+use kernel::prelude::*;
+
+use crate::{
+ gsp::{
+ cmdq::{
+ command_size,
+ CommandToGsp, //
+ },
+ fw::{
+ GspMsgElement,
+ MsgFunction,
+ GSP_MSG_QUEUE_ELEMENT_SIZE_MAX, //
+ },
+ },
+ sbuffer::SBufferIter,
+};
+
+/// Maximum command size that fits in a single queue element.
+const MAX_CMD_SIZE: usize = GSP_MSG_QUEUE_ELEMENT_SIZE_MAX - size_of::<GspMsgElement>();
+
+/// Acts as an iterator over the continuation records for a split command.
+pub(super) struct ContinuationRecords {
+ payload: KVVec<u8>,
+ offset: usize,
+}
+
+impl ContinuationRecords {
+ /// Creates a new iterator over continuation records for the given payload.
+ fn new(payload: KVVec<u8>) -> Self {
+ Self { payload, offset: 0 }
+ }
+
+ /// Returns the next continuation record, or [`None`] if there are no more.
+ pub(super) fn next(&mut self) -> Option<ContinuationRecord<'_>> {
+ let remaining = self.payload.len() - self.offset;
+
+ if remaining > 0 {
+ let chunk_size = remaining.min(MAX_CMD_SIZE);
+ let record =
+ ContinuationRecord::new(&self.payload[self.offset..(self.offset + chunk_size)]);
+ self.offset += chunk_size;
+ Some(record)
+ } else {
+ None
+ }
+ }
+}
+
+/// The [`ContinuationRecord`] command.
+pub(super) struct ContinuationRecord<'a> {
+ data: &'a [u8],
+}
+
+impl<'a> ContinuationRecord<'a> {
+ /// Creates a new [`ContinuationRecord`] command with the given data.
+ fn new(data: &'a [u8]) -> Self {
+ Self { data }
+ }
+}
+
+impl<'a> CommandToGsp for ContinuationRecord<'a> {
+ const FUNCTION: MsgFunction = MsgFunction::ContinuationRecord;
+ type Command = ();
+ type InitError = Infallible;
+
+ fn init(&self) -> impl Init<Self::Command, Self::InitError> {
+ <()>::init_zeroed()
+ }
+
+ fn variable_payload_len(&self) -> usize {
+ self.data.len()
+ }
+
+ fn init_variable_payload(
+ &self,
+ dst: &mut SBufferIter<core::array::IntoIter<&mut [u8], 2>>,
+ ) -> Result {
+ dst.write_all(self.data)
+ }
+}
+
+/// Whether a command needs to be split across continuation records or not.
+pub(super) enum SplitState<C: CommandToGsp> {
+ /// A command that fits in a single queue element.
+ Single(C),
+ /// A command split across continuation records.
+ Split(SplitCommand<C>, ContinuationRecords),
+}
+
+impl<C: CommandToGsp> SplitState<C> {
+ /// Maximum variable payload size that fits in the first command alongside the command header.
+ const MAX_FIRST_PAYLOAD: usize = MAX_CMD_SIZE - size_of::<C::Command>();
+
+ /// Creates a new [`SplitState`] for the given command.
+ ///
+ /// If the command is too large, it will be split into a main command and some number of
+ /// continuation records.
+ pub(super) fn new(command: C) -> Result<Self> {
+ let payload_len = command.variable_payload_len();
+
+ if command_size(&command) > MAX_CMD_SIZE {
+ let mut command_payload =
+ KVVec::<u8>::from_elem(0u8, payload_len.min(Self::MAX_FIRST_PAYLOAD), GFP_KERNEL)?;
+ let mut continuation_payload =
+ KVVec::<u8>::from_elem(0u8, payload_len - command_payload.len(), GFP_KERNEL)?;
+ let mut sbuffer = SBufferIter::new_writer([
+ command_payload.as_mut_slice(),
+ continuation_payload.as_mut_slice(),
+ ]);
+
+ command.init_variable_payload(&mut sbuffer)?;
+ if !sbuffer.is_empty() {
+ return Err(EIO);
+ }
+ drop(sbuffer);
+
+ Ok(Self::Split(
+ SplitCommand::new(command, command_payload),
+ ContinuationRecords::new(continuation_payload),
+ ))
+ } else {
+ Ok(Self::Single(command))
+ }
+ }
+}
+
+/// A command that has been truncated to maximum accepted length of the command queue.
+///
+/// The remainder of its payload is expected to be sent using [`ContinuationRecords`].
+pub(super) struct SplitCommand<C: CommandToGsp> {
+ command: C,
+ payload: KVVec<u8>,
+}
+
+impl<C: CommandToGsp> SplitCommand<C> {
+ /// Creates a new [`SplitCommand`] wrapping `command` with the given truncated payload.
+ fn new(command: C, payload: KVVec<u8>) -> Self {
+ Self { command, payload }
+ }
+}
+
+impl<C: CommandToGsp> CommandToGsp for SplitCommand<C> {
+ const FUNCTION: MsgFunction = C::FUNCTION;
+ type Command = C::Command;
+ type InitError = C::InitError;
+
+ fn init(&self) -> impl Init<Self::Command, Self::InitError> {
+ self.command.init()
+ }
+
+ fn variable_payload_len(&self) -> usize {
+ self.payload.len()
+ }
+
+ fn init_variable_payload(
+ &self,
+ dst: &mut SBufferIter<core::array::IntoIter<&mut [u8], 2>>,
+ ) -> Result {
+ dst.write_all(&self.payload)
+ }
+}
diff --git a/drivers/gpu/nova-core/gsp/fw.rs b/drivers/gpu/nova-core/gsp/fw.rs
index 6005362450cb..25fca1f6db2c 100644
--- a/drivers/gpu/nova-core/gsp/fw.rs
+++ b/drivers/gpu/nova-core/gsp/fw.rs
@@ -202,6 +202,7 @@ pub(crate) enum MsgFunction {
AllocObject = bindings::NV_VGPU_MSG_FUNCTION_ALLOC_OBJECT,
AllocRoot = bindings::NV_VGPU_MSG_FUNCTION_ALLOC_ROOT,
BindCtxDma = bindings::NV_VGPU_MSG_FUNCTION_BIND_CTX_DMA,
+ ContinuationRecord = bindings::NV_VGPU_MSG_FUNCTION_CONTINUATION_RECORD,
Free = bindings::NV_VGPU_MSG_FUNCTION_FREE,
GetGspStaticInfo = bindings::NV_VGPU_MSG_FUNCTION_GET_GSP_STATIC_INFO,
GetStaticInfo = bindings::NV_VGPU_MSG_FUNCTION_GET_STATIC_INFO,
@@ -239,6 +240,9 @@ fn try_from(value: u32) -> Result<MsgFunction> {
bindings::NV_VGPU_MSG_FUNCTION_ALLOC_OBJECT => Ok(MsgFunction::AllocObject),
bindings::NV_VGPU_MSG_FUNCTION_ALLOC_ROOT => Ok(MsgFunction::AllocRoot),
bindings::NV_VGPU_MSG_FUNCTION_BIND_CTX_DMA => Ok(MsgFunction::BindCtxDma),
+ bindings::NV_VGPU_MSG_FUNCTION_CONTINUATION_RECORD => {
+ Ok(MsgFunction::ContinuationRecord)
+ }
bindings::NV_VGPU_MSG_FUNCTION_FREE => Ok(MsgFunction::Free),
bindings::NV_VGPU_MSG_FUNCTION_GET_GSP_STATIC_INFO => Ok(MsgFunction::GetGspStaticInfo),
bindings::NV_VGPU_MSG_FUNCTION_GET_STATIC_INFO => Ok(MsgFunction::GetStaticInfo),
--
2.53.0On Mon Mar 2, 2026 at 8:42 PM JST, Eliot Courtney wrote: > Splits large RPCs if necessary and sends the remaining parts using > continuation records. RPCs that do not need continuation records > continue to write directly into the command buffer. Ones that do write > into a staging buffer first, so there is one copy. > > Continuation record for receive is not necessary to support at the > moment because those replies do not need to be read and are currently > drained by retrying `receive_msg` on ERANGE. > > Signed-off-by: Eliot Courtney <ecourtney@nvidia.com> > --- > drivers/gpu/nova-core/gsp.rs | 1 + > drivers/gpu/nova-core/gsp/cmdq.rs | 41 +++++++- > drivers/gpu/nova-core/gsp/continuation.rs | 167 ++++++++++++++++++++++++++++++ > drivers/gpu/nova-core/gsp/fw.rs | 4 + > 4 files changed, 210 insertions(+), 3 deletions(-) > > diff --git a/drivers/gpu/nova-core/gsp.rs b/drivers/gpu/nova-core/gsp.rs > index 174feaca0a6b..ccf56f1ad246 100644 > --- a/drivers/gpu/nova-core/gsp.rs > +++ b/drivers/gpu/nova-core/gsp.rs > @@ -16,6 +16,7 @@ > > pub(crate) mod cmdq; > pub(crate) mod commands; > +mod continuation; Looking at this series it seems more logical to have `continuation` under `cmdq` than just `gsp`. Nothing in `gsp` makes use of it, as it is an implementation detail of the command queue. And that way the `pub(super)` exports would be perfectly scoped to their user. Re-reading my review of v3, I am the one who suggested to put it under `gsp` - so this is my mistake. No need to resend just for that, I can fix when applying and this should have no impact on the patches that come above it. Giving it a few days to gather reviews, but after a first skim this looks ready to me.
On Mon Mar 2, 2026 at 10:06 PM JST, Alexandre Courbot wrote: > On Mon Mar 2, 2026 at 8:42 PM JST, Eliot Courtney wrote: >> Splits large RPCs if necessary and sends the remaining parts using >> continuation records. RPCs that do not need continuation records >> continue to write directly into the command buffer. Ones that do write >> into a staging buffer first, so there is one copy. >> >> Continuation record for receive is not necessary to support at the >> moment because those replies do not need to be read and are currently >> drained by retrying `receive_msg` on ERANGE. >> >> Signed-off-by: Eliot Courtney <ecourtney@nvidia.com> >> --- >> drivers/gpu/nova-core/gsp.rs | 1 + >> drivers/gpu/nova-core/gsp/cmdq.rs | 41 +++++++- >> drivers/gpu/nova-core/gsp/continuation.rs | 167 ++++++++++++++++++++++++++++++ >> drivers/gpu/nova-core/gsp/fw.rs | 4 + >> 4 files changed, 210 insertions(+), 3 deletions(-) >> >> diff --git a/drivers/gpu/nova-core/gsp.rs b/drivers/gpu/nova-core/gsp.rs >> index 174feaca0a6b..ccf56f1ad246 100644 >> --- a/drivers/gpu/nova-core/gsp.rs >> +++ b/drivers/gpu/nova-core/gsp.rs >> @@ -16,6 +16,7 @@ >> >> pub(crate) mod cmdq; >> pub(crate) mod commands; >> +mod continuation; > > Looking at this series it seems more logical to have `continuation` > under `cmdq` than just `gsp`. > > Nothing in `gsp` makes use of it, as it is an implementation detail of > the command queue. And that way the `pub(super)` exports would be > perfectly scoped to their user. > > Re-reading my review of v3, I am the one who suggested to put it under > `gsp` - so this is my mistake. > > No need to resend just for that, I can fix when applying and this should > have no impact on the patches that come above it. I think we might have one more round on this so for now I have made this change locally and will be in the next version.
© 2016 - 2026 Red Hat, Inc.