The per-connection memory budget (MAX_CONNECTION_MEMORY, 256 MiB) only
charged wire buffers via receive_data_limited. Decompression buffers and
per-file chunk copies were not accounted for, and the multithreaded
receiver could enqueue up to 100 files (each up to 64 MiB uncompressed)
ahead of a slow disk writer, retaining ~6.4 GiB per connection. A client
sending highly compressible chunks with little bandwidth could OOM the
host while the reserve never tripped.
Bound the receive pipeline by aggregate payload bytes instead of item
count alone:
- Export MAX_CONNECTION_MEMORY from protocol.h.
- PipelineContextReceiver tracks queued_bytes (payload bytes received but
not yet released by the disk writer, i.e. queued or in the writer's
hand) under the existing mutex.
- receiver enqueue now blocks while the queue is full by count OR when
adding the file would push queued_bytes over the configured byte limit,
applying backpressure to the sender instead of failing the transfer.
- The disk writer releases the byte budget after each file is freed and
signals the not-full condition.
- The server sets the byte ceiling to
MAX_CONNECTION_MEMORY - 2*MAX_CHUNK_SIZE so that the queued payloads
plus the transient wire/decompression buffers of the one in-flight
chunk stay within the per-connection budget.
The single-threaded receive path is already bounded: it writes files to
disk before reading the next chunk, so its transient is at most one
chunk's wire + decompressed + copied payload (~3 * MAX_CHUNK_SIZE, below
the budget). Wire buffers remain charged exactly once by
receive_data_limited; this change does not double charge them.
Adds a deterministic unit test in test_multiprocessing.c proving that an
enqueue which would exceed the byte budget blocks until the writer
releases bytes.
TapTap
98120fc722