[PATCH v2 8/9] gpu: nova-core: gsp: support large RPCs via continuation record

Eliot Courtney posted 9 patches 1 month, 2 weeks ago
There is a newer version of this series
[PATCH v2 8/9] gpu: nova-core: gsp: support large RPCs via continuation record
Posted by Eliot Courtney 1 month, 2 weeks ago
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/cmdq.rs     |  43 +++++++++-
 drivers/gpu/nova-core/gsp/commands.rs | 143 +++++++++++++++++++++++++++++++++-
 drivers/gpu/nova-core/gsp/fw.rs       |   5 ++
 3 files changed, 187 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/nova-core/gsp/cmdq.rs b/drivers/gpu/nova-core/gsp/cmdq.rs
index f832252ae45c..402b0b9418ab 100644
--- a/drivers/gpu/nova-core/gsp/cmdq.rs
+++ b/drivers/gpu/nova-core/gsp/cmdq.rs
@@ -29,6 +29,10 @@
 use crate::{
     driver::Bar0,
     gsp::{
+        commands::{
+            ContinuationRecord,
+            SplitState, //
+        },
         fw::{
             GspMsgElement,
             MsgFunction,
@@ -439,7 +443,7 @@ struct GspMessage<'a> {
 
 /// Computes the total size of the command (not including the [`GspMsgElement`] header), including
 /// its variable-length payload.
-fn command_size<M>(command: &M) -> usize
+pub(crate) fn command_size<M>(command: &M) -> usize
 where
     M: CommandToGsp,
 {
@@ -507,7 +511,7 @@ fn notify_gsp(bar: &Bar0) {
             .write(bar);
     }
 
-    /// Sends `command` to the GSP.
+    /// Sends `command` to the GSP, without splitting it.
     ///
     /// # Errors
     ///
@@ -516,7 +520,7 @@ 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`.
@@ -575,6 +579,39 @@ 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>,
+    {
+        let mut state = SplitState::new(&command)?;
+
+        self.send_single_command(bar, state.command(command))?;
+
+        while let Some(continuation) = state.next_continuation_record() {
+            dev_dbg!(
+                &self.dev,
+                "GSP RPC: send continuation: size=0x{:x}\n",
+                command_size(&continuation),
+            );
+            // 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/commands.rs b/drivers/gpu/nova-core/gsp/commands.rs
index 8f270eca33be..e722c6091ec8 100644
--- a/drivers/gpu/nova-core/gsp/commands.rs
+++ b/drivers/gpu/nova-core/gsp/commands.rs
@@ -4,6 +4,7 @@
     array,
     convert::Infallible,
     ffi::FromBytesUntilNulError,
+    marker::PhantomData,
     str::Utf8Error, //
 };
 
@@ -22,13 +23,16 @@
     driver::Bar0,
     gsp::{
         cmdq::{
+            command_size,
             Cmdq,
             CommandToGsp,
             MessageFromGsp, //
         },
         fw::{
             commands::*,
-            MsgFunction, //
+            GspMsgElement,
+            MsgFunction,
+            GSP_MSG_QUEUE_ELEMENT_SIZE_MAX, //
         },
     },
     sbuffer::SBufferIter,
@@ -242,3 +246,140 @@ pub(crate) fn get_gsp_info(cmdq: &mut Cmdq, bar: &Bar0) -> Result<GetGspStaticIn
         }
     }
 }
+
+/// The `ContinuationRecord` command.
+pub(crate) struct ContinuationRecord<'a> {
+    data: &'a [u8],
+}
+
+impl<'a> ContinuationRecord<'a> {
+    /// Creates a new `ContinuationRecord` command with the given data.
+    pub(crate) 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)
+    }
+}
+
+/// Wrapper that splits a command across continuation records if needed.
+pub(crate) struct SplitState<C: CommandToGsp> {
+    state: Option<(KVVec<u8>, usize)>,
+    _phantom: PhantomData<C>,
+}
+
+impl<C: CommandToGsp> SplitState<C> {
+    /// Maximum command size that fits in a single queue element.
+    const MAX_CMD_SIZE: usize = GSP_MSG_QUEUE_ELEMENT_SIZE_MAX - size_of::<GspMsgElement>();
+
+    /// Maximum size of the variable payload that can be sent in the main command.
+    const MAX_FIRST_PAYLOAD_SIZE: usize = Self::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(crate) fn new(inner: &C) -> Result<Self> {
+        if command_size(inner) > Self::MAX_CMD_SIZE {
+            let mut staging =
+                KVVec::<u8>::from_elem(0u8, inner.variable_payload_len(), GFP_KERNEL)?;
+            let mut sbuffer = SBufferIter::new_writer([staging.as_mut_slice(), &mut []]);
+            inner.init_variable_payload(&mut sbuffer)?;
+            if !sbuffer.is_empty() {
+                return Err(EIO);
+            }
+            drop(sbuffer);
+
+            Ok(Self {
+                state: Some((staging, Self::MAX_FIRST_PAYLOAD_SIZE)),
+                _phantom: PhantomData,
+            })
+        } else {
+            Ok(Self {
+                state: None,
+                _phantom: PhantomData,
+            })
+        }
+    }
+
+    /// Returns the main command.
+    pub(crate) fn command(&self, inner: C) -> SplitCommand<'_, C> {
+        if let Some((staging, _)) = &self.state {
+            SplitCommand::Split(inner, staging)
+        } else {
+            SplitCommand::Single(inner)
+        }
+    }
+
+    /// Returns the next continuation record, or `None` if there are no more.
+    pub(crate) fn next_continuation_record(&mut self) -> Option<ContinuationRecord<'_>> {
+        let (staging, offset) = self.state.as_mut()?;
+
+        let remaining = staging.len() - *offset;
+        if remaining > 0 {
+            let chunk_size = remaining.min(Self::MAX_CMD_SIZE);
+            let record = ContinuationRecord::new(&staging[*offset..(*offset + chunk_size)]);
+            *offset += chunk_size;
+            Some(record)
+        } else {
+            None
+        }
+    }
+}
+
+/// Wrapper enum that represents either a single command or a split command with its staging buffer.
+pub(crate) enum SplitCommand<'a, C: CommandToGsp> {
+    Single(C),
+    Split(C, &'a [u8]),
+}
+
+impl<'a, C: CommandToGsp> CommandToGsp for SplitCommand<'a, C> {
+    const FUNCTION: MsgFunction = C::FUNCTION;
+    type Command = C::Command;
+    type InitError = C::InitError;
+
+    fn init(&self) -> impl Init<Self::Command, Self::InitError> {
+        match self {
+            SplitCommand::Single(cmd) => cmd.init(),
+            SplitCommand::Split(cmd, _) => cmd.init(),
+        }
+    }
+
+    fn variable_payload_len(&self) -> usize {
+        match self {
+            SplitCommand::Single(cmd) => cmd.variable_payload_len(),
+            SplitCommand::Split(_, _) => SplitState::<C>::MAX_FIRST_PAYLOAD_SIZE,
+        }
+    }
+
+    fn init_variable_payload(
+        &self,
+        dst: &mut SBufferIter<core::array::IntoIter<&mut [u8], 2>>,
+    ) -> Result {
+        match self {
+            SplitCommand::Single(cmd) => cmd.init_variable_payload(dst),
+            SplitCommand::Split(_, staging) => {
+                dst.write_all(&staging[..self.variable_payload_len()])
+            }
+        }
+    }
+}
+
diff --git a/drivers/gpu/nova-core/gsp/fw.rs b/drivers/gpu/nova-core/gsp/fw.rs
index c604f423fff3..1a317dfb0bc9 100644
--- a/drivers/gpu/nova-core/gsp/fw.rs
+++ b/drivers/gpu/nova-core/gsp/fw.rs
@@ -203,6 +203,7 @@ 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,
@@ -238,6 +239,7 @@ fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
             MsgFunction::AllocObject => write!(f, "ALLOC_OBJECT"),
             MsgFunction::AllocRoot => write!(f, "ALLOC_ROOT"),
             MsgFunction::BindCtxDma => write!(f, "BIND_CTX_DMA"),
+            MsgFunction::ContinuationRecord => write!(f, "CONTINUATION_RECORD"),
             MsgFunction::Free => write!(f, "FREE"),
             MsgFunction::GetGspStaticInfo => write!(f, "GET_GSP_STATIC_INFO"),
             MsgFunction::GetStaticInfo => write!(f, "GET_STATIC_INFO"),
@@ -277,6 +279,9 @@ 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),

-- 
2.53.0
Re: [PATCH v2 8/9] gpu: nova-core: gsp: support large RPCs via continuation record
Posted by Alexandre Courbot 1 month, 1 week ago
On Thu Feb 19, 2026 at 4:30 PM JST, Eliot Courtney wrote:
<snip>
> @@ -575,6 +579,39 @@ 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>,
> +    {
> +        let mut state = SplitState::new(&command)?;
> +
> +        self.send_single_command(bar, state.command(command))?;

I think it would be nice to send the actual command (and not the
`SplitCommand`) when commands don't need to be split. Split commands are
supposed to be the exception, but the current code makes us use them
unconditionally.

So something like:

    if command_size(&command) <= MAX_CMD_SIZE {
          self.send_single_command(bar, command)
      } else {
      // Split and send the parts
      ...
      }

would read better IMHO. You can have a
`SplitCommand::command_needs_splitting` method if you want to avoid
comparing against the const in `send_command`, but honestly I think this
is self-documenting already - you split the command because it is larger
than the maximum supported size of the queue.

It also would have the benefit of simplifying `SplitState` and
`SplitCommand` since they wouldn't have to handle the non-split case at
all. Actually you could just have a method that consumes the command and
returns a tuple with the `SplitCommand` wrapping it, and its
continuation records. That way I suspect you also remove the need to
pass `command` twice.
Re: [PATCH v2 8/9] gpu: nova-core: gsp: support large RPCs via continuation record
Posted by Alexandre Courbot 1 month, 1 week ago
On Thu Feb 26, 2026 at 1:40 PM JST, Alexandre Courbot wrote:
> On Thu Feb 19, 2026 at 4:30 PM JST, Eliot Courtney wrote:
> <snip>
>> @@ -575,6 +579,39 @@ 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>,
>> +    {
>> +        let mut state = SplitState::new(&command)?;
>> +
>> +        self.send_single_command(bar, state.command(command))?;
>
> I think it would be nice to send the actual command (and not the
> `SplitCommand`) when commands don't need to be split. Split commands are
> supposed to be the exception, but the current code makes us use them
> unconditionally.
>
> So something like:
>
>     if command_size(&command) <= MAX_CMD_SIZE {
>           self.send_single_command(bar, command)
>       } else {
>       // Split and send the parts
>       ...
>       }
>
> would read better IMHO. You can have a
> `SplitCommand::command_needs_splitting` method if you want to avoid
> comparing against the const in `send_command`, but honestly I think this
> is self-documenting already - you split the command because it is larger
> than the maximum supported size of the queue.
>
> It also would have the benefit of simplifying `SplitState` and
> `SplitCommand` since they wouldn't have to handle the non-split case at
> all. Actually you could just have a method that consumes the command and
> returns a tuple with the `SplitCommand` wrapping it, and its
> continuation records. That way I suspect you also remove the need to
> pass `command` twice.

Mmm or maybe it's not that simple. I'll experiment on the rebased v3
once it it sent, please ignore this comment for now.