Coverage Report

Created: 2026-08-14 20:52

next uncovered line (L), next uncovered region (R), next uncovered branch (B)
/build/source/nativelink-util/src/proto_stream_utils.rs
Line
Count
Source
1
// Copyright 2024 The NativeLink Authors. All rights reserved.
2
//
3
// Licensed under the Functional Source License, Version 1.1, Apache 2.0 Future License (the "License");
4
// you may not use this file except in compliance with the License.
5
// You may obtain a copy of the License at
6
//
7
//    See LICENSE file for details
8
//
9
// Unless required by applicable law or agreed to in writing, software
10
// distributed under the License is distributed on an "AS IS" BASIS,
11
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
// See the License for the specific language governing permissions and
13
// limitations under the License.
14
15
use core::fmt::Debug;
16
use core::mem;
17
use core::pin::Pin;
18
use core::task::{Context, Poll};
19
use std::borrow::Cow;
20
use std::sync::Arc;
21
22
use futures::{Stream, StreamExt};
23
use nativelink_error::{Error, ResultExt, error_if, make_input_err};
24
use nativelink_proto::google::bytestream::{ReadResponse, WriteRequest};
25
use parking_lot::Mutex;
26
use tonic::{Status, Streaming};
27
28
use crate::resource_info::ResourceInfo;
29
30
pub struct WriteRequestStreamWrapper<T> {
31
    pub resource_info: ResourceInfo<'static>,
32
    pub bytes_received: usize,
33
    stream: T,
34
    first_msg: Option<WriteRequest>,
35
    pub write_finished: bool,
36
}
37
38
impl<T> Debug for WriteRequestStreamWrapper<T> {
39
19
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
40
19
        f.debug_struct("WriteRequestStreamWrapper")
41
19
            .field("resource_info", &self.resource_info)
42
19
            .field("bytes_received", &self.bytes_received)
43
19
            .field("first_msg", &self.first_msg)
44
19
            .field("write_finished", &self.write_finished)
45
19
            .finish()
46
19
    }
47
}
48
49
impl<T, E> WriteRequestStreamWrapper<T>
50
where
51
    T: Stream<Item = Result<WriteRequest, E>> + Unpin,
52
    E: Into<Error>,
53
{
54
45
    pub async fn from(mut stream: T) -> Result<Self, Error> {
55
45
        let 
first_msg44
= stream
56
45
            .next()
57
45
            .await
58
45
            .err_tip(|| "Error receiving first message in stream")
?0
59
45
            .err_tip(|| "Expected WriteRequest struct in stream (from)")
?1
;
60
61
44
        let resource_info = ResourceInfo::new(&first_msg.resource_name, true)
62
44
            .err_tip(|| 
{0
63
0
                format!(
64
                    "Could not extract resource info from first message of stream: {}",
65
                    first_msg.resource_name
66
                )
67
0
            })?
68
44
            .to_owned();
69
70
44
        Ok(Self {
71
44
            resource_info,
72
44
            bytes_received: 0,
73
44
            stream,
74
44
            first_msg: Some(first_msg),
75
44
            write_finished: false,
76
44
        })
77
45
    }
78
79
53
    pub async fn next(&mut self) -> Option<Result<WriteRequest, Error>> {
80
161
        
futures::future::poll_fn53
(|cx| Pin::new(&mut *self).poll_next(cx)).
await53
81
52
    }
82
83
4
    pub const fn is_first_msg(&self) -> bool {
84
4
        self.first_msg.is_some()
85
4
    }
86
87
    /// Returns whether the first message has `finish_write` set to true.
88
    /// This indicates a single-shot upload where all data is in one message.
89
20
    pub fn is_first_msg_complete(&self) -> bool {
90
20
        self.first_msg.as_ref().is_some_and(|msg| msg.finish_write)
91
20
    }
92
93
90
    fn enforce_wire_size_matches_digest_size(&self) -> bool {
94
32
        matches!(
95
90
            self.resource_info.compressor.as_deref(),
96
33
            None | Some("identity")
97
        )
98
90
    }
99
}
100
101
impl<T, E> Stream for WriteRequestStreamWrapper<T>
102
where
103
    E: Into<Error>,
104
    T: Stream<Item = Result<WriteRequest, E>> + Unpin,
105
{
106
    type Item = Result<WriteRequest, Error>;
107
108
207
    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
109
        // If the stream said that the previous message was the last one, then
110
        // return a stream EOF (i.e. None).
111
207
        if self.write_finished {
112
10
            if self.enforce_wire_size_matches_digest_size() {
113
0
                error_if!(
114
7
                    self.bytes_received != self.resource_info.expected_size,
115
                    "Did not send enough data. Expected {}, but so far received {}",
116
0
                    self.resource_info.expected_size,
117
0
                    self.bytes_received
118
                );
119
3
            }
120
10
            return Poll::Ready(None);
121
197
        }
122
123
        // Gets the next message, this is either the cached first or a
124
        // subsequent message from the wrapped Stream.
125
197
        let 
maybe_message84
= if let Some(
first_msg42
) = self.first_msg.take() {
126
42
            Ok(first_msg)
127
        } else {
128
155
            match Pin::new(&mut self.stream).poll_next(cx) {
129
113
                Poll::Pending => return Poll::Pending,
130
38
                Poll::Ready(Some(maybe_message)) => maybe_message
131
38
                    .err_tip(|| 
format!0
("Stream error at byte {}",
self.bytes_received0
)),
132
4
                Poll::Ready(None) => Err(make_input_err!(
133
4
                    "Expected WriteRequest struct in stream (got None)"
134
4
                )),
135
            }
136
        };
137
138
        // If we successfully got a message, update our internal state with the
139
        // message meta data.
140
84
        Poll::Ready(Some(maybe_message.and_then(|message| 
{80
141
80
            self.write_finished = message.finish_write;
142
80
            self.bytes_received += message.data.len();
143
144
            // Check that we haven't read past the expected end.
145
80
            if self.enforce_wire_size_matches_digest_size()
146
51
                && self.bytes_received > self.resource_info.expected_size
147
            {
148
2
                Err(make_input_err!(
149
2
                    "Sent too much data. Expected {}, but so far received {}",
150
2
                    self.resource_info.expected_size,
151
2
                    self.bytes_received
152
2
                ))
153
            } else {
154
78
                Ok(message)
155
            }
156
80
        })))
157
207
    }
158
}
159
160
/// Represents the state of the first response in a `FirstStream`.
161
#[derive(Debug)]
162
pub enum FirstResponseState {
163
    /// Contains an optional first response that hasn't been consumed yet.
164
    /// A `None` value indicates the first response was EOF.
165
    Unused(Option<ReadResponse>),
166
    /// Indicates the first response has been consumed and future reads should
167
    /// come from the underlying stream.
168
    Used,
169
}
170
171
/// This provides a buffer for the first response from GrpcStore.read in order
172
/// to allow the first read to occur within the retry loop.  That means that if
173
/// the connection establishes fine, but reading the first byte of the file
174
/// fails we have the ability to retry before returning to the caller.
175
#[derive(Debug)]
176
pub struct FirstStream {
177
    /// The current state of the first response. When in the `Unused` state,
178
    /// contains an optional response which could be `None` or an EOF.
179
    /// Once consumed, transitions to the `Used` state.
180
    state: FirstResponseState,
181
    /// The stream to get responses from after the first response is consumed.
182
    stream: Streaming<ReadResponse>,
183
}
184
185
impl FirstStream {
186
    /// Creates a new `FirstStream` with the given first response and underlying
187
    /// stream.
188
14
    pub const fn new(
189
14
        first_response: Option<ReadResponse>,
190
14
        stream: Streaming<ReadResponse>,
191
14
    ) -> Self {
192
14
        Self {
193
14
            state: FirstResponseState::Unused(first_response),
194
14
            stream,
195
14
        }
196
14
    }
197
}
198
199
impl Stream for FirstStream {
200
    type Item = Result<ReadResponse, Status>;
201
202
111
    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
203
111
        match mem::replace(&mut self.state, FirstResponseState::Used) {
204
14
            FirstResponseState::Unused(first_response) => Poll::Ready(first_response.map(Ok)),
205
97
            FirstResponseState::Used => Pin::new(&mut self.stream).poll_next(cx),
206
        }
207
111
    }
208
}
209
210
/// Largest `data` payload in a single `ByteStream` `WriteRequest`, kept under
211
/// tonic's 4MiB default `max_decoding_message_size` with room to spare.
212
pub const MAX_WRITE_REQUEST_DATA_BYTES: usize = 2 * 1024 * 1024;
213
214
/// This structure wraps all of the information required to perform a write
215
/// request on the `GrpcStore`.  It stores the last message retrieved which allows
216
/// the write to resume since the UUID allows upload resume at the server.
217
#[derive(Debug)]
218
pub struct WriteState<T, E>
219
where
220
    T: Stream<Item = Result<WriteRequest, E>> + Unpin + Send + 'static,
221
    E: Into<Error> + 'static,
222
{
223
    instance_name: String,
224
    read_stream_error: Option<Error>,
225
    read_stream: WriteRequestStreamWrapper<T>,
226
    // Tonic doesn't appear to report an error until it has taken two messages,
227
    // therefore we are required to buffer the last two messages.
228
    cached_messages: [Option<WriteRequest>; 2],
229
    // When resuming after an error, the previous messages are cloned into this
230
    // queue upfront to allow them to be served back.
231
    resume_queue: [Option<WriteRequest>; 2],
232
    // An optimisation to avoid having to manage resume_queue when it's empty.
233
    is_resumed: bool,
234
    // Remainder of an incoming frame too large to forward in one message.
235
    // Drained before the next frame is read so ordering is preserved.
236
    pending: Option<WriteRequest>,
237
    // When false, a partially-consumed stream never reports `can_resume()`:
238
    // uploads whose server-side protocol cannot accept a replay from a
239
    // nonzero offset (REAPI compressed-blobs writes) must fail fast instead
240
    // of burning retries on guaranteed-rejected resumes. A stream that has
241
    // not yet been consumed can always be retried from the start.
242
    resumable: bool,
243
}
244
245
impl<T, E> WriteState<T, E>
246
where
247
    T: Stream<Item = Result<WriteRequest, E>> + Unpin + Send + 'static,
248
    E: Into<Error> + 'static,
249
{
250
11
    pub const fn new(instance_name: String, read_stream: WriteRequestStreamWrapper<T>) -> Self {
251
11
        Self {
252
11
            instance_name,
253
11
            read_stream_error: None,
254
11
            read_stream,
255
11
            cached_messages: [None, None],
256
11
            resume_queue: [None, None],
257
11
            is_resumed: false,
258
11
            pending: None,
259
11
            resumable: true,
260
11
        }
261
11
    }
262
263
    /// Caps the payload of an outgoing message, stashing anything over the
264
    /// limit to be sent as the following message.
265
    ///
266
    /// A caller can hand us a whole blob in one frame, and forwarding that
267
    /// verbatim produces a message the receiver rejects at its 4MiB default.
268
    /// Only the final piece carries `finish_write`, and each piece advances
269
    /// `write_offset` by what came before it.
270
34
    fn split_oversized(&mut self, mut message: WriteRequest) -> WriteRequest {
271
34
        if message.data.len() <= MAX_WRITE_REQUEST_DATA_BYTES {
272
32
            return message;
273
2
        }
274
2
        let rest = message.data.split_off(MAX_WRITE_REQUEST_DATA_BYTES);
275
2
        let sent_len = i64::try_from(message.data.len()).unwrap_or(i64::MAX);
276
2
        self.pending = Some(WriteRequest {
277
2
            resource_name: message.resource_name.clone(),
278
2
            write_offset: message.write_offset.saturating_add(sent_len),
279
2
            finish_write: message.finish_write,
280
2
            data: rest,
281
2
        });
282
2
        message.finish_write = false;
283
2
        message
284
34
    }
285
286
    /// Marks this write as non-resumable: once the stream has been partially
287
    /// consumed, `can_resume()` reports false so the caller fails fast with
288
    /// the original error instead of replaying messages the server-side
289
    /// protocol is guaranteed to reject. Retrying an unconsumed stream from
290
    /// the start remains allowed.
291
4
    pub const fn set_non_resumable(&mut self) {
292
4
        self.resumable = false;
293
4
    }
294
295
34
    pub(crate) const fn is_resumable(&self) -> bool {
296
34
        self.resumable
297
34
    }
298
299
20
    fn push_message(&mut self, message: WriteRequest) {
300
20
        self.cached_messages.swap(0, 1);
301
20
        self.cached_messages[0] = Some(message);
302
20
    }
303
304
48
    const fn resumed_message(&mut self) -> Option<WriteRequest> {
305
48
        if self.is_resumed {
306
            // The resume_queue is a circular buffer, that we have to shift,
307
            // since its only got two elements its a trivial swap.
308
0
            self.resume_queue.swap(0, 1);
309
0
            let message = self.resume_queue[0].take();
310
0
            if message.is_none() {
311
0
                self.is_resumed = false;
312
0
            }
313
0
            message
314
        } else {
315
48
            None
316
        }
317
48
    }
318
319
4
    pub const fn can_resume(&self) -> bool {
320
4
        self.read_stream_error.is_none()
321
4
            && ((self.resumable && 
self.cached_messages[0]0
.
is_some0
())
322
4
                || self.read_stream.is_first_msg())
323
4
    }
324
325
0
    pub fn resume(&mut self) {
326
0
        self.resume_queue.clone_from(&self.cached_messages);
327
0
        self.is_resumed = true;
328
0
    }
329
330
11
    pub const fn take_read_stream_error(&mut self) -> Option<Error> {
331
11
        self.read_stream_error.take()
332
11
    }
333
}
334
335
/// A wrapper around `WriteState` to allow it to be reclaimed from the underlying
336
/// write call in the case of failure.
337
#[derive(Debug)]
338
pub struct WriteStateWrapper<T, E>
339
where
340
    T: Stream<Item = Result<WriteRequest, E>> + Unpin + Send + 'static,
341
    E: Into<Error> + 'static,
342
{
343
    shared_state: Arc<Mutex<WriteState<T, E>>>,
344
}
345
346
impl<T, E> WriteStateWrapper<T, E>
347
where
348
    T: Stream<Item = Result<WriteRequest, E>> + Unpin + Send + 'static,
349
    E: Into<Error> + 'static,
350
{
351
11
    pub const fn new(shared_state: Arc<Mutex<WriteState<T, E>>>) -> Self {
352
11
        Self { shared_state }
353
11
    }
354
}
355
356
impl<T, E> Stream for WriteStateWrapper<T, E>
357
where
358
    T: Stream<Item = Result<WriteRequest, E>> + Unpin + Send + 'static,
359
    E: Into<Error> + 'static,
360
{
361
    type Item = WriteRequest;
362
363
48
    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
364
        const IS_UPLOAD_TRUE: bool = true;
365
366
        // This should be an uncontended lock since write was called.
367
48
        let mut local_state = self.shared_state.lock();
368
        // If this is the first or second call after a failure and we have
369
        // cached messages, then use the cached write requests.
370
48
        let cached_message = local_state.resumed_message();
371
48
        if cached_message.is_some() {
372
0
            return Poll::Ready(cached_message);
373
48
        }
374
        // Finish draining an oversized frame before reading the next one.
375
48
        if let Some(
pending2
) = local_state.pending.take() {
376
2
            let message = local_state.split_oversized(pending);
377
2
            if local_state.is_resumable() {
378
2
                local_state.push_message(message.clone());
379
2
            
}0
380
2
            return Poll::Ready(Some(message));
381
46
        }
382
        // Read a new write request from the downstream.
383
46
        let Poll::Ready(
maybe_message42
) = Pin::new(&mut local_state.read_stream).poll_next(cx)
384
        else {
385
4
            return Poll::Pending;
386
        };
387
        // Update the instance name in the write request and forward it on.
388
42
        let result = match 
maybe_message32
{
389
32
            Some(Ok(mut message)) => {
390
32
                if !message.resource_name.is_empty() {
391
                    // Replace the instance name in the resource name if it is
392
                    // different from the instance name in the write state.
393
30
                    match ResourceInfo::new(&message.resource_name, IS_UPLOAD_TRUE) {
394
30
                        Ok(mut resource_name) => {
395
30
                            if resource_name.instance_name != local_state.instance_name {
396
1
                                resource_name.instance_name =
397
1
                                    Cow::Borrowed(&local_state.instance_name);
398
1
                                message.resource_name = resource_name.to_string(IS_UPLOAD_TRUE);
399
29
                            }
400
                        }
401
0
                        Err(err) => {
402
0
                            local_state.read_stream_error = Some(err);
403
0
                            return Poll::Ready(None);
404
                        }
405
                    }
406
2
                }
407
32
                let message = local_state.split_oversized(message);
408
                // Cache the last request in case there is an error to allow
409
                // the upload to be resumed. Non-resumable writes skip the
410
                // clone: cached messages would never be replayed.
411
32
                if local_state.is_resumable() {
412
18
                    local_state.push_message(message.clone());
413
18
                
}14
414
32
                Some(message)
415
            }
416
0
            Some(Err(err)) => {
417
0
                local_state.read_stream_error = Some(err);
418
0
                None
419
            }
420
10
            None => None,
421
        };
422
42
        Poll::Ready(result)
423
48
    }
424
}