fix: #254 bound receiver queue by aggregate payload bytes

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.

src/server/server.c

@@ -4,6 +4,7 @@
 #include "log.h"
 #include "metadata.h"
 #include "multiprocessing.h"
+#include "protocol.h"
 #include "queue.h"
 #include "receiver.h"
 #include "transport_tcp.h"
@@ -26,6 +27,13 @@ static bool allow_delete;
 static bool allow_unauthenticated;
 static const char* required_client_cn;
 
+/* Aggregate payload bytes the multithreaded receiver may buffer ahead of the
+   slow disk writer.  Receiving one more chunk adds up to ~2 * MAX_CHUNK_SIZE
+   of transient wire/decompression buffers on top of the queued payloads, so
+   this ceiling keeps total per-connection receive memory (decompressed and
+   per-file copied chunk buffers included) within MAX_CONNECTION_MEMORY. */
+#define RECEIVER_QUEUE_MAX_BYTES (MAX_CONNECTION_MEMORY - 2 * MAX_CHUNK_SIZE)
+
 static bool tls_client_identity_allowed(SSL* ssl) {
   if (!ssl || !required_client_cn)
     return false;
@@ -314,6 +322,7 @@ void handler(int file_descriptor) {
     protocol_session_set_max_alloc(&context->session, config->max_alloc);
     atomic_store(&context->session.total_allocated_bytes,
                  atomic_load(&session.total_allocated_bytes));
+    pipeline_context_receiver_set_queue_byte_limit(context, RECEIVER_QUEUE_MAX_BYTES);
     thrd_t receiver, writer;
     bool receiver_created = thrd_create(&receiver, receive_thread, context) == thrd_success;
     bool writer_created = false;

src/shared/multiprocessing.c

@@ -108,6 +108,8 @@ PipelineContextReceiver* pipeline_context_receiver_create(Config* config, Queue*
   protocol_session_init(&context->session, file_descriptor, file_descriptor);
   protocol_session_set_ssl(&context->session, ssl);
   context->receiver_done = false;
+  context->queued_bytes = 0;
+  context->max_queue_bytes = 0;
   atomic_init(&context->cancelled, false);
   int init = 0;
   if (mtx_init(&context->mutex, mtx_plain) != thrd_success)
@@ -143,14 +145,73 @@ void pipeline_context_receiver_destroy(PipelineContextReceiver* context) {
   free(context);
 }
 
+void pipeline_context_receiver_set_queue_byte_limit(PipelineContextReceiver* context,
+                                                    size_t max_bytes) {
+  if (context == NULL)
+    return;
+  mtx_lock(&context->mutex);
+  context->max_queue_bytes = max_bytes;
+  context->queued_bytes = 0;
+  cnd_broadcast(&context->condition_not_full);
+  mtx_unlock(&context->mutex);
+}
+
+void pipeline_context_receiver_note_bytes_released(PipelineContextReceiver* context,
+                                                   size_t released_bytes) {
+  if (context == NULL || context->max_queue_bytes == 0 || released_bytes == 0)
+    return;
+  mtx_lock(&context->mutex);
+  if (released_bytes >= context->queued_bytes)
+    context->queued_bytes = 0;
+  else
+    context->queued_bytes -= released_bytes;
+  cnd_signal(&context->condition_not_full);
+  mtx_unlock(&context->mutex);
+}
+
+bool pipeline_context_receiver_enqueue_file(PipelineContextReceiver* context, File* file) {
+  if (context == NULL || file == NULL)
+    return false;
+  size_t file_bytes = file->data ? file->data->size : 0;
+  mtx_lock(&context->mutex);
+  while (!atomic_load(&context->cancelled)) {
+    bool blocked_by_count = queue_is_full(context->queue);
+    bool blocked_by_budget = false;
+    if (context->max_queue_bytes > 0) {
+      size_t budget = context->max_queue_bytes;
+      size_t used = context->queued_bytes;
+      if (used >= budget) {
+        blocked_by_budget = true;
+      } else if (file_bytes > budget - used) {
+        /* A single payload larger than the whole budget (not possible with
+           the per-file receive cap) is only admitted to an empty pipeline so
+           the wait can never deadlock. */
+        blocked_by_budget = used != 0;
+      }
+    }
+    if (!blocked_by_count && !blocked_by_budget)
+      break;
+    cnd_wait(&context->condition_not_full, &context->mutex);
+  }
+  if (atomic_load(&context->cancelled)) {
+    mtx_unlock(&context->mutex);
+    file_destroy(file);
+    return false;
+  }
+  if (!queue_enqueue(context->queue, file)) {
+    mtx_unlock(&context->mutex);
+    file_destroy(file);
+    return false;
+  }
+  context->queued_bytes += file_bytes;
+  cnd_signal(&context->condition_not_empty);
+  mtx_unlock(&context->mutex);
+  return true;
+}
+
 static bool receiver_enqueue_file(File* file, void* context_pointer) {
-  PipelineContextReceiver* context = context_pointer;
-  if (queue_enqueue_multithreaded_cancel(context->queue, file, &context->mutex,
-                                         &context->condition_not_empty,
-                                         &context->condition_not_full, &context->cancelled))
-    return true;
-  file_destroy(file);
-  return false;
+  PipelineContextReceiver* context = (PipelineContextReceiver*)context_pointer;
+  return pipeline_context_receiver_enqueue_file(context, file);
 }
 
 static void receiver_thread_fail(PipelineContextReceiver* context) {
@@ -211,8 +272,10 @@ int write_thread(void* pipeline_context) {
       protocol_session_unbind();
       return thrd_success;
     }
+    size_t file_bytes = file->data ? file->data->size : 0;
     if (save_to_disk && !file_save_to_disk(root_directory, file, context->config)) {
       file_destroy(file);
+      pipeline_context_receiver_note_bytes_released(context, file_bytes);
       mtx_lock(&context->mutex);
       atomic_store(&context->cancelled, true);
       context->receiver_done = true;
@@ -224,5 +287,6 @@ int write_thread(void* pipeline_context) {
       return thrd_error;
     }
     file_destroy(file);
+    pipeline_context_receiver_note_bytes_released(context, file_bytes);
   }
 }

src/shared/multiprocessing.h

@@ -45,6 +45,14 @@ typedef struct PipelineContextReceiver {
   cnd_t condition_not_empty;
   bool receiver_done;
   atomic_bool cancelled;
+  /* Aggregate payload bytes that have been received but not yet released by
+     the disk writer (queued or in the writer's hand).  Guarded by `mutex`.
+     When `max_queue_bytes` is non-zero the receiver blocks before enqueuing
+     once this total would exceed it, so decompressed/copied file payloads
+     buffered ahead of a slow disk writer respect the per-connection memory
+     budget instead of growing without bound. */
+  size_t queued_bytes;
+  size_t max_queue_bytes;
 } PipelineContextReceiver;
 
 PipelineContextSender* pipeline_context_sender_create(Config* config, Queue* queue_scanner,
@@ -53,6 +61,18 @@ void pipeline_context_sender_destroy(PipelineContextSender* context);
 PipelineContextReceiver* pipeline_context_receiver_create(Config* config, Queue* queue_receiver,
                                                           int file_descriptor, SSL* ssl);
 void pipeline_context_receiver_destroy(PipelineContextReceiver* context);
+/* Bound the bytes buffered ahead of the disk writer (see max_queue_bytes). */
+void pipeline_context_receiver_set_queue_byte_limit(PipelineContextReceiver* context,
+                                                    size_t max_bytes);
+/* Blocking enqueue used by the receive pipeline sink.  Blocks while the queue
+   is full by element count or when adding `file` would push queued_bytes over
+   the configured byte limit; waits until the disk writer releases bytes.
+   Takes ownership of `file` on success and destroys it on failure/cancel. */
+bool pipeline_context_receiver_enqueue_file(PipelineContextReceiver* context, File* file);
+/* Account for `released_bytes` of payload memory that has been freed by the
+   disk writer, unblocking a receiver that is waiting on the byte limit. */
+void pipeline_context_receiver_note_bytes_released(PipelineContextReceiver* context,
+                                                   size_t released_bytes);
 int receive_thread(void* pipeline_context);
 int write_thread(void* pipeline_context);
 #endif

src/shared/protocol.c

@@ -14,7 +14,6 @@
 
 #define RECEIVE_TIMEOUT_SEC 60 /* 60 second per-message timeout */
 #define SEND_TIMEOUT_SEC 60
-#define MAX_CONNECTION_MEMORY (256ULL * 1024 * 1024) /* bounded cumulative receive budget */
 
 static __thread int io_read_fd = -1;
 static __thread int io_write_fd = -1;

src/shared/protocol.h

@@ -22,6 +22,10 @@
 #define DEFAULT_MAX_ALLOC (1ULL * 1024 * 1024 * 1024)
 /* Server policy ceiling for a client-provided allocation limit. */
 #define MAX_SERVER_ALLOC (256ULL * 1024 * 1024)
+/* Bounded cumulative per-connection receive budget.  In-flight wire buffers,
+   decompression buffers and queued (not yet written) file payloads for a
+   connection must stay within this ceiling. */
+#define MAX_CONNECTION_MEMORY (256ULL * 1024 * 1024)
 
 typedef struct ssl_st SSL;
 

tests/test_multiprocessing.c

@@ -250,6 +250,87 @@ static void test_write_thread_done() {
   config_delete(cfg);
 }
 
+typedef struct {
+  PipelineContextReceiver* context;
+  File* file;
+  atomic_bool* done;
+  atomic_bool* result;
+} ByteBudgetEnqueueArg;
+
+static int byte_budget_enqueue_worker(void* arg) {
+  ByteBudgetEnqueueArg* worker = arg;
+  bool ok = pipeline_context_receiver_enqueue_file(worker->context, worker->file);
+  atomic_store(worker->result, ok);
+  atomic_store(worker->done, true);
+  return thrd_success;
+}
+
+/* A receiver must not buffer more decompressed/copied payload bytes ahead of
+   the (slow) disk writer than the configured byte budget: an enqueue that
+   would exceed the budget blocks until the writer releases bytes. */
+static void test_receiver_enqueue_byte_budget() {
+  Config* cfg = config_create();
+  EXPECT_NOT_NULL(cfg);
+  free(cfg->version);
+  cfg->version = str_dup(PROTOCOL_VERSION);
+  cfg->send_directory = str_dup("/src");
+  cfg->receive_root_directory = str_dup("/dst");
+  cfg->save_to_disk = false;
+
+  Queue* q = queue_create(16, file_destroy);
+  EXPECT_NOT_NULL(q);
+  PipelineContextReceiver* ctx = pipeline_context_receiver_create(cfg, q, -1, NULL);
+  EXPECT_NOT_NULL(ctx);
+  pipeline_context_receiver_set_queue_byte_limit(ctx, 3000);
+  ctx->receiver_done = false;
+
+  File* first = file_create("budget_file_1");
+  EXPECT_NOT_NULL(first);
+  first->data->size = 2000;
+  EXPECT_TRUE(pipeline_context_receiver_enqueue_file(ctx, first));
+  EXPECT_EQ_INT((int)ctx->queued_bytes, 2000);
+
+  /* Second 2000-byte payload would push the pipeline to 4000 > 3000 budget,
+     so the enqueue must block until the first payload is released. */
+  File* second = file_create("budget_file_2");
+  EXPECT_NOT_NULL(second);
+  second->data->size = 2000;
+  atomic_bool done;
+  atomic_bool result;
+  atomic_init(&done, false);
+  atomic_init(&result, false);
+  ByteBudgetEnqueueArg arg = {ctx, second, &done, &result};
+  thrd_t enqueuer;
+  EXPECT_EQ_INT(thrd_create(&enqueuer, byte_budget_enqueue_worker, &arg), thrd_success);
+
+  /* Give a broken (unbounded) implementation every chance to enqueue. */
+  struct timespec wait = {0, 200 * 1000000L};
+  thrd_sleep(&wait, NULL);
+  EXPECT_FALSE(atomic_load(&done));
+  EXPECT_EQ_INT((int)ctx->queued_bytes, 2000); /* budget still honored */
+
+  /* Simulate the disk writer: dequeue + destroy + release the first file. */
+  File* drained = queue_dequeue_multithreaded(q, &ctx->mutex, &ctx->condition_not_empty,
+                                              &ctx->condition_not_full, &ctx->receiver_done);
+  EXPECT_NOT_NULL(drained);
+  file_destroy(drained);
+  pipeline_context_receiver_note_bytes_released(ctx, 2000);
+  EXPECT_EQ_INT((int)ctx->queued_bytes, 0);
+
+  EXPECT_EQ_INT(thrd_join(enqueuer, NULL), thrd_success);
+  EXPECT_TRUE(atomic_load(&done));
+  EXPECT_TRUE(atomic_load(&result));
+  EXPECT_EQ_INT((int)ctx->queued_bytes, 2000); /* second payload now in flight */
+
+  /* Tear down: the second file is still queued and is freed by queue_destroy. */
+  mtx_destroy(&ctx->mutex);
+  cnd_destroy(&ctx->condition_not_full);
+  cnd_destroy(&ctx->condition_not_empty);
+  free(ctx);
+  queue_destroy(q);
+  config_delete(cfg);
+}
+
 void test_multiprocessing() {
   test_sender_create_destroy();
   test_receiver_create_destroy();
@@ -261,4 +342,5 @@ void test_multiprocessing() {
     test_receive_thread_failure_wakes_writer();
   }
   test_write_thread_done();
+  test_receiver_enqueue_byte_budget();
 }