grpc 1.42.0 → 1.49.1

data/src/core/lib/iomgr/tcp_posix.cc

@@ -18,6 +18,8 @@
 
 #include <grpc/support/port_platform.h>
 
+#include <grpc/impl/codegen/grpc_types.h>
+
 #include "src/core/lib/iomgr/port.h"
 
 #ifdef GRPC_POSIX_SOCKET_TCP
@@ -54,10 +56,12 @@
 #include "src/core/lib/iomgr/buffer_list.h"
 #include "src/core/lib/iomgr/ev_posix.h"
 #include "src/core/lib/iomgr/executor.h"
-#include "src/core/lib/iomgr/resource_quota.h"
 #include "src/core/lib/iomgr/socket_utils_posix.h"
 #include "src/core/lib/iomgr/tcp_posix.h"
 #include "src/core/lib/profiling/timers.h"
+#include "src/core/lib/resource_quota/api.h"
+#include "src/core/lib/resource_quota/memory_quota.h"
+#include "src/core/lib/resource_quota/trace.h"
 #include "src/core/lib/slice/slice_internal.h"
 #include "src/core/lib/slice/slice_string_helpers.h"
 
@@ -93,6 +97,8 @@ typedef size_t msg_iovlen_type;
 
 extern grpc_core::TraceFlag grpc_tcp_trace;
 
+GPR_GLOBAL_CONFIG_DECLARE_BOOL(grpc_experimental_enable_tcp_frame_size_tuning);
+
 namespace grpc_core {
 
 class TcpZerocopySendRecord {
@@ -228,7 +234,11 @@ class TcpZerocopySendCtx {
   // buffers that were sent with the corresponding call to sendmsg().
   void NoteSend(TcpZerocopySendRecord* record) {
     record->Ref();
-    AssociateSeqWithSendRecord(last_send_, record);
+    {
+      MutexLock guard(&lock_);
+      is_in_write_ = true;
+      AssociateSeqWithSendRecordLocked(last_send_, record);
+    }
     ++last_send_;
   }
 
@@ -249,8 +259,8 @@ class TcpZerocopySendCtx {
 
   // Simply associate this send record (and the underlying sent data buffers)
   // with the implicit sequence number for this zerocopy sendmsg().
-  void AssociateSeqWithSendRecord(uint32_t seq, TcpZerocopySendRecord* record) {
-    MutexLock guard(&lock_);
+  void AssociateSeqWithSendRecordLocked(uint32_t seq,
+                                        TcpZerocopySendRecord* record) {
     ctx_lookup_.emplace(seq, record);
   }
 
@@ -308,7 +318,107 @@ class TcpZerocopySendCtx {
   // zerocopy is not useful for small transfers.
   size_t threshold_bytes() const { return threshold_bytes_; }
 
+  // Expected to be called by handler reading messages from the err queue.
+  // It is used to indicate that some OMem meory is now available. It returns
+  // true to tell the caller to mark the file descriptor as immediately
+  // writable.
+  //
+  // If a write is currently in progress on the socket (ie. we have issued a
+  // sendmsg() and are about to check its return value) then we set omem state
+  // to CHECK to make the sending thread know that some tcp_omem was
+  // concurrently freed even if sendmsg() returns ENOBUFS. In this case, since
+  // there is already an active send thread, we do not need to mark the
+  // socket writeable, so we return false.
+  //
+  // If there was no write in progress on the socket, and the socket was not
+  // marked as FULL, then we need not mark the socket writeable now that some
+  // tcp_omem memory is freed since it was not considered as blocked on
+  // tcp_omem to begin with. So in this case, return false.
+  //
+  // But, if a write was not in progress and the omem state was FULL, then we
+  // need to mark the socket writeable since it is no longer blocked by
+  // tcp_omem. In this case, return true.
+  //
+  // Please refer to the STATE TRANSITION DIAGRAM below for more details.
+  //
+  bool UpdateZeroCopyOMemStateAfterFree() {
+    MutexLock guard(&lock_);
+    if (is_in_write_) {
+      zcopy_enobuf_state_ = OMemState::CHECK;
+      return false;
+    }
+    GPR_DEBUG_ASSERT(zcopy_enobuf_state_ != OMemState::CHECK);
+    if (zcopy_enobuf_state_ == OMemState::FULL) {
+      // A previous sendmsg attempt was blocked by ENOBUFS. Return true to
+      // mark the fd as writable so the next write attempt could be made.
+      zcopy_enobuf_state_ = OMemState::OPEN;
+      return true;
+    } else if (zcopy_enobuf_state_ == OMemState::OPEN) {
+      // No need to mark the fd as writable because the previous write
+      // attempt did not encounter ENOBUFS.
+      return false;
+    } else {
+      // This state should never be reached because it implies that the previous
+      // state was CHECK and is_in_write is false. This means that after the
+      // previous sendmsg returned and set is_in_write to false, it did
+      // not update the z-copy change from CHECK to OPEN.
+      GPR_ASSERT(false && "OMem state error!");
+    }
+  }
+
+  // Expected to be called by the thread calling sendmsg after the syscall
+  // invocation. is complete. If an ENOBUF is seen, it checks if the error
+  // handler (Tx0cp completions) has already run and free'ed up some OMem. It
+  // returns true indicating that the write can be attempted again immediately.
+  // If ENOBUFS was seen but no Tx0cp completions have been received between the
+  // sendmsg() and us taking this lock, then tcp_omem is still full from our
+  // point of view. Therefore, we do not signal that the socket is writeable
+  // with respect to the availability of tcp_omem. Therefore the function
+  // returns false. This indicates that another write should not be attempted
+  // immediately and the calling thread should wait until the socket is writable
+  // again. If ENOBUFS was not seen, then again return false because the next
+  // write should be attempted only when the socket is writable again.
+  //
+  // Please refer to the STATE TRANSITION DIAGRAM below for more details.
+  //
+  bool UpdateZeroCopyOMemStateAfterSend(bool seen_enobuf) {
+    MutexLock guard(&lock_);
+    is_in_write_ = false;
+    if (seen_enobuf) {
+      if (zcopy_enobuf_state_ == OMemState::CHECK) {
+        zcopy_enobuf_state_ = OMemState::OPEN;
+        return true;
+      } else {
+        zcopy_enobuf_state_ = OMemState::FULL;
+      }
+    } else if (zcopy_enobuf_state_ != OMemState::OPEN) {
+      zcopy_enobuf_state_ = OMemState::OPEN;
+    }
+    return false;
+  }
+
  private:
+  //                      STATE TRANSITION DIAGRAM
+  //
+  // sendmsg succeeds       Tx-zero copy succeeds and there is no active sendmsg
+  //      ----<<--+  +------<<-------------------------------------+
+  //      |       |  |                                             |
+  //      |       |  v       sendmsg returns ENOBUFS               |
+  //      +-----> OPEN  ------------->>-------------------------> FULL
+  //                ^                                              |
+  //                |                                              |
+  //                | sendmsg completes                            |
+  //                +----<<---------- CHECK <-------<<-------------+
+  //                                        Tx-zero copy succeeds and there is
+  //                                        an active sendmsg
+  //
+  enum class OMemState : int8_t {
+    OPEN,   // Everything is clear and omem is not full.
+    FULL,   // The last sendmsg() has returned with an errno of ENOBUFS.
+    CHECK,  // Error queue is read while is_in_write_ was true, so we should
+            // check this state after the sendmsg.
+  };
+
   TcpZerocopySendRecord* ReleaseSendRecordLocked(uint32_t seq) {
     auto iter = ctx_lookup_.find(seq);
     GPR_DEBUG_ASSERT(iter != ctx_lookup_.end());
@@ -345,6 +455,8 @@ class TcpZerocopySendCtx {
   size_t threshold_bytes_ = kDefaultSendBytesThreshold;
   std::unordered_map<uint32_t, TcpZerocopySendRecord*> ctx_lookup_;
   bool memory_limited_ = false;
+  bool is_in_write_ = false;
+  OMemState zcopy_enobuf_state_;
 };
 
 }  // namespace grpc_core
@@ -353,6 +465,13 @@ using grpc_core::TcpZerocopySendCtx;
 using grpc_core::TcpZerocopySendRecord;
 
 namespace {
+
+bool ExperimentalTcpFrameSizeTuningEnabled() {
+  static const bool kEnableTcpFrameSizeTuning =
+      GPR_GLOBAL_CONFIG_GET(grpc_experimental_enable_tcp_frame_size_tuning);
+  return kEnableTcpFrameSizeTuning;
+}
+
 struct grpc_tcp {
   grpc_tcp(int max_sends, size_t send_bytes_threshold)
       : tcp_zerocopy_send_ctx(max_sends, send_bytes_threshold) {}
@@ -362,6 +481,7 @@ struct grpc_tcp {
   /* Used by the endpoint read function to distinguish the very first read call
    * from the rest */
   bool is_first_read;
+  bool has_posted_reclaimer;
   double target_length;
   double bytes_read_this_round;
   grpc_core::RefCount refcount;
@@ -373,7 +493,8 @@ struct grpc_tcp {
   /* garbage after the last read */
   grpc_slice_buffer last_read_buffer;
 
-  grpc_slice_buffer* incoming_buffer;
+  grpc_core::Mutex read_mu;
+  grpc_slice_buffer* incoming_buffer ABSL_GUARDED_BY(read_mu) = nullptr;
   int inq;          /* bytes pending on the socket from the last read. */
   bool inq_capable; /* cache whether kernel supports inq */
 
@@ -393,7 +514,8 @@ struct grpc_tcp {
   std::string peer_string;
   std::string local_address;
 
-  grpc_slice_allocator* slice_allocator;
+  grpc_core::MemoryOwner memory_owner;
+  grpc_core::MemoryAllocator::Reservation self_reservation;
 
   grpc_core::TracedBuffer* tb_head; /* List of traced buffers */
   gpr_mu tb_mu; /* Lock for access to list of traced buffers */
@@ -419,6 +541,11 @@ struct grpc_tcp {
                                       on errors anymore */
   TcpZerocopySendCtx tcp_zerocopy_send_ctx;
   TcpZerocopySendRecord* current_zerocopy_send = nullptr;
+
+  bool frame_size_tuning_enabled;
+  int min_progress_size; /* A hint from upper layers specifying the minimum
+                            number of bytes that need to be read to make
+                            meaningful progress */
 };
 
 struct backup_poller {
@@ -457,7 +584,8 @@ static void run_poller(void* bp, grpc_error_handle /*error_ignored*/) {
     gpr_log(GPR_INFO, "BACKUP_POLLER:%p run", p);
   }
   gpr_mu_lock(p->pollset_mu);
-  grpc_millis deadline = grpc_core::ExecCtx::Get()->Now() + 10 * GPR_MS_PER_SEC;
+  grpc_core::Timestamp deadline =
+      grpc_core::ExecCtx::Get()->Now() + grpc_core::Duration::Seconds(10);
   GRPC_STATS_INC_TCP_BACKUP_POLLER_POLLS();
   GRPC_LOG_IF_ERROR(
       "backup_poller:pollset_work",
@@ -608,7 +736,6 @@ static void tcp_free(grpc_tcp* tcp) {
   grpc_fd_orphan(tcp->em_fd, tcp->release_fd_cb, tcp->release_fd,
                  "tcp_unref_orphan");
   grpc_slice_buffer_destroy_internal(&tcp->last_read_buffer);
-  grpc_slice_allocator_destroy(tcp->slice_allocator);
   /* The lock is not really necessary here, since all refs have been released */
   gpr_mu_lock(&tcp->tb_mu);
   grpc_core::TracedBuffer::Shutdown(
@@ -657,9 +784,35 @@ static void tcp_destroy(grpc_endpoint* ep) {
   TCP_UNREF(tcp, "destroy");
 }
 
-static void call_read_cb(grpc_tcp* tcp, grpc_error_handle error) {
-  grpc_closure* cb = tcp->read_cb;
+static void perform_reclamation(grpc_tcp* tcp)
+    ABSL_LOCKS_EXCLUDED(tcp->read_mu) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_resource_quota_trace)) {
+    gpr_log(GPR_INFO, "TCP: benign reclamation to free memory");
+  }
+  tcp->read_mu.Lock();
+  if (tcp->incoming_buffer != nullptr) {
+    grpc_slice_buffer_reset_and_unref_internal(tcp->incoming_buffer);
+  }
+  tcp->read_mu.Unlock();
+  tcp->has_posted_reclaimer = false;
+}
+
+static void maybe_post_reclaimer(grpc_tcp* tcp)
+    ABSL_EXCLUSIVE_LOCKS_REQUIRED(tcp->read_mu) {
+  if (!tcp->has_posted_reclaimer) {
+    tcp->has_posted_reclaimer = true;
+    tcp->memory_owner.PostReclaimer(
+        grpc_core::ReclamationPass::kBenign,
+        [tcp](absl::optional<grpc_core::ReclamationSweep> sweep) {
+          if (!sweep.has_value()) return;
+          perform_reclamation(tcp);
+        });
+  }
+}
 
+static void tcp_trace_read(grpc_tcp* tcp, grpc_error_handle error)
+    ABSL_EXCLUSIVE_LOCKS_REQUIRED(tcp->read_mu) {
+  grpc_closure* cb = tcp->read_cb;
   if (GRPC_TRACE_FLAG_ENABLED(grpc_tcp_trace)) {
     gpr_log(GPR_INFO, "TCP:%p call_cb %p %p:%p", tcp, cb, cb->cb, cb->cb_arg);
     size_t i;
@@ -674,15 +827,16 @@ static void call_read_cb(grpc_tcp* tcp, grpc_error_handle error) {
       }
     }
   }
-
-  tcp->read_cb = nullptr;
-  tcp->incoming_buffer = nullptr;
-  grpc_core::Closure::Run(DEBUG_LOCATION, cb, error);
 }
 
+/* Returns true if data available to read or error other than EAGAIN. */
 #define MAX_READ_IOVEC 4
-static void tcp_do_read(grpc_tcp* tcp) {
+static bool tcp_do_read(grpc_tcp* tcp, grpc_error_handle* error)
+    ABSL_EXCLUSIVE_LOCKS_REQUIRED(tcp->read_mu) {
   GPR_TIMER_SCOPE("tcp_do_read", 0);
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_tcp_trace)) {
+    gpr_log(GPR_INFO, "TCP:%p do_read", tcp);
+  }
   struct msghdr msg;
   struct iovec iov[MAX_READ_IOVEC];
   ssize_t read_bytes;
@@ -701,6 +855,9 @@ static void tcp_do_read(grpc_tcp* tcp) {
     iov[i].iov_len = GRPC_SLICE_LENGTH(tcp->incoming_buffer->slices[i]);
   }
 
+  GPR_ASSERT(tcp->incoming_buffer->length != 0);
+  GPR_DEBUG_ASSERT(tcp->min_progress_size > 0);
+
   do {
     /* Assume there is something on the queue. If we receive TCP_INQ from
      * kernel, we will update this value, otherwise, we have to assume there is
@@ -731,7 +888,8 @@ static void tcp_do_read(grpc_tcp* tcp) {
 
     /* We have read something in previous reads. We need to deliver those
      * bytes to the upper layer. */
-    if (read_bytes <= 0 && total_read_bytes > 0) {
+    if (read_bytes <= 0 &&
+        total_read_bytes >= static_cast<size_t>(tcp->min_progress_size)) {
       tcp->inq = 1;
       break;
     }
@@ -740,17 +898,17 @@ static void tcp_do_read(grpc_tcp* tcp) {
       /* NB: After calling call_read_cb a parallel call of the read handler may
        * be running. */
       if (errno == EAGAIN) {
+        if (total_read_bytes > 0) {
+          break;
+        }
         finish_estimate(tcp);
         tcp->inq = 0;
-        /* We've consumed the edge, request a new one */
-        notify_on_read(tcp);
+        return false;
       } else {
         grpc_slice_buffer_reset_and_unref_internal(tcp->incoming_buffer);
-        call_read_cb(tcp,
-                     tcp_annotate_error(GRPC_OS_ERROR(errno, "recvmsg"), tcp));
-        TCP_UNREF(tcp, "read");
+        *error = tcp_annotate_error(GRPC_OS_ERROR(errno, "recvmsg"), tcp);
+        return true;
       }
-      return;
     }
     if (read_bytes == 0) {
       /* 0 read size ==> end of stream
@@ -759,11 +917,9 @@ static void tcp_do_read(grpc_tcp* tcp) {
        * since the connection is closed we will drop the data here, because we
        * can't call the callback multiple times. */
       grpc_slice_buffer_reset_and_unref_internal(tcp->incoming_buffer);
-      call_read_cb(
-          tcp, tcp_annotate_error(
-                   GRPC_ERROR_CREATE_FROM_STATIC_STRING("Socket closed"), tcp));
-      TCP_UNREF(tcp, "read");
-      return;
+      *error = tcp_annotate_error(
+          GRPC_ERROR_CREATE_FROM_STATIC_STRING("Socket closed"), tcp);
+      return true;
     }
 
     GRPC_STATS_INC_TCP_READ_SIZE(read_bytes);
@@ -787,7 +943,6 @@ static void tcp_do_read(grpc_tcp* tcp) {
 
     total_read_bytes += read_bytes;
     if (tcp->inq == 0 || total_read_bytes == tcp->incoming_buffer->length) {
-      /* We have filled incoming_buffer, and we cannot read any more. */
       break;
     }
 
@@ -818,50 +973,69 @@ static void tcp_do_read(grpc_tcp* tcp) {
   }
 
   GPR_DEBUG_ASSERT(total_read_bytes > 0);
+  *error = GRPC_ERROR_NONE;
+  if (tcp->frame_size_tuning_enabled) {
+    // Update min progress size based on the total number of bytes read in
+    // this round.
+    tcp->min_progress_size -= total_read_bytes;
+    if (tcp->min_progress_size > 0) {
+      // There is still some bytes left to be read before we can signal
+      // the read as complete. Append the bytes read so far into
+      // last_read_buffer which serves as a staging buffer. Return false
+      // to indicate tcp_handle_read needs to be scheduled again.
+      grpc_slice_buffer_move_first(tcp->incoming_buffer, total_read_bytes,
+                                   &tcp->last_read_buffer);
+      return false;
+    } else {
+      // The required number of bytes have been read. Append the bytes
+      // read in this round into last_read_buffer. Then swap last_read_buffer
+      // and incoming_buffer. Now incoming buffer contains all the bytes
+      // read since the start of the last tcp_read operation. last_read_buffer
+      // would contain any spare space left in the incoming buffer. This
+      // space will be used in the next tcp_read operation.
+      tcp->min_progress_size = 1;
+      grpc_slice_buffer_move_first(tcp->incoming_buffer, total_read_bytes,
+                                   &tcp->last_read_buffer);
+      grpc_slice_buffer_swap(&tcp->last_read_buffer, tcp->incoming_buffer);
+      return true;
+    }
+  }
   if (total_read_bytes < tcp->incoming_buffer->length) {
     grpc_slice_buffer_trim_end(tcp->incoming_buffer,
                                tcp->incoming_buffer->length - total_read_bytes,
                                &tcp->last_read_buffer);
   }
-  call_read_cb(tcp, GRPC_ERROR_NONE);
-  TCP_UNREF(tcp, "read");
-}
-
-static void tcp_read_allocation_done(void* tcpp, grpc_error_handle error) {
-  grpc_tcp* tcp = static_cast<grpc_tcp*>(tcpp);
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_tcp_trace)) {
-    gpr_log(GPR_INFO, "TCP:%p read_allocation_done: %s", tcp,
-            grpc_error_std_string(error).c_str());
-  }
-  if (GPR_UNLIKELY(error != GRPC_ERROR_NONE)) {
-    grpc_slice_buffer_reset_and_unref_internal(tcp->incoming_buffer);
-    grpc_slice_buffer_reset_and_unref_internal(&tcp->last_read_buffer);
-    call_read_cb(tcp, GRPC_ERROR_REF(error));
-    TCP_UNREF(tcp, "read");
-  } else {
-    tcp_do_read(tcp);
-  }
+  return true;
 }
 
-static void tcp_continue_read(grpc_tcp* tcp) {
-  /* Wait for allocation only when there is no buffer left. */
-  if (tcp->incoming_buffer->length == 0 &&
+static void maybe_make_read_slices(grpc_tcp* tcp)
+    ABSL_EXCLUSIVE_LOCKS_REQUIRED(tcp->read_mu) {
+  if (tcp->incoming_buffer->length <
+          static_cast<size_t>(tcp->min_progress_size) &&
       tcp->incoming_buffer->count < MAX_READ_IOVEC) {
     if (GRPC_TRACE_FLAG_ENABLED(grpc_tcp_trace)) {
-      gpr_log(GPR_INFO, "TCP:%p alloc_slices", tcp);
+      gpr_log(GPR_INFO,
+              "TCP:%p alloc_slices; min_chunk=%d max_chunk=%d target=%lf "
+              "buf_len=%" PRIdPTR,
+              tcp, tcp->min_read_chunk_size, tcp->max_read_chunk_size,
+              tcp->target_length, tcp->incoming_buffer->length);
     }
-    if (GPR_UNLIKELY(!grpc_slice_allocator_allocate(
-            tcp->slice_allocator, tcp->target_length, 1,
-            grpc_slice_allocator_intent::kReadBuffer, tcp->incoming_buffer,
-            tcp_read_allocation_done, tcp))) {
-      // Wait for allocation.
-      return;
-    }
-  }
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_tcp_trace)) {
-    gpr_log(GPR_INFO, "TCP:%p do_read", tcp);
+    int target_length =
+        std::max(static_cast<int>(tcp->target_length), tcp->min_progress_size);
+    int extra_wanted =
+        target_length - static_cast<int>(tcp->incoming_buffer->length);
+    int min_read_chunk_size =
+        std::max(tcp->min_read_chunk_size, tcp->min_progress_size);
+    int max_read_chunk_size =
+        std::max(tcp->max_read_chunk_size, tcp->min_progress_size);
+    grpc_slice_buffer_add_indexed(
+        tcp->incoming_buffer,
+        tcp->memory_owner.MakeSlice(grpc_core::MemoryRequest(
+            min_read_chunk_size,
+            grpc_core::Clamp(extra_wanted, min_read_chunk_size,
+                             max_read_chunk_size))));
+    maybe_post_reclaimer(tcp);
   }
-  tcp_do_read(tcp);
 }
 
 static void tcp_handle_read(void* arg /* grpc_tcp */, grpc_error_handle error) {
@@ -870,25 +1044,42 @@ static void tcp_handle_read(void* arg /* grpc_tcp */, grpc_error_handle error) {
     gpr_log(GPR_INFO, "TCP:%p got_read: %s", tcp,
             grpc_error_std_string(error).c_str());
   }
-
-  if (GPR_UNLIKELY(error != GRPC_ERROR_NONE)) {
+  tcp->read_mu.Lock();
+  grpc_error_handle tcp_read_error;
+  if (GPR_LIKELY(GRPC_ERROR_IS_NONE(error))) {
+    maybe_make_read_slices(tcp);
+    if (!tcp_do_read(tcp, &tcp_read_error)) {
+      /* We've consumed the edge, request a new one */
+      tcp->read_mu.Unlock();
+      notify_on_read(tcp);
+      return;
+    }
+    tcp_trace_read(tcp, tcp_read_error);
+  } else {
+    tcp_read_error = GRPC_ERROR_REF(error);
     grpc_slice_buffer_reset_and_unref_internal(tcp->incoming_buffer);
     grpc_slice_buffer_reset_and_unref_internal(&tcp->last_read_buffer);
-    call_read_cb(tcp, GRPC_ERROR_REF(error));
-    TCP_UNREF(tcp, "read");
-  } else {
-    tcp_continue_read(tcp);
   }
+  grpc_closure* cb = tcp->read_cb;
+  tcp->read_cb = nullptr;
+  tcp->incoming_buffer = nullptr;
+  tcp->read_mu.Unlock();
+  grpc_core::Closure::Run(DEBUG_LOCATION, cb, tcp_read_error);
+  TCP_UNREF(tcp, "read");
 }
 
 static void tcp_read(grpc_endpoint* ep, grpc_slice_buffer* incoming_buffer,
-                     grpc_closure* cb, bool urgent) {
+                     grpc_closure* cb, bool urgent, int min_progress_size) {
   grpc_tcp* tcp = reinterpret_cast<grpc_tcp*>(ep);
   GPR_ASSERT(tcp->read_cb == nullptr);
   tcp->read_cb = cb;
+  tcp->read_mu.Lock();
   tcp->incoming_buffer = incoming_buffer;
+  tcp->min_progress_size =
+      tcp->frame_size_tuning_enabled ? min_progress_size : 1;
   grpc_slice_buffer_reset_and_unref_internal(incoming_buffer);
   grpc_slice_buffer_swap(incoming_buffer, &tcp->last_read_buffer);
+  tcp->read_mu.Unlock();
   TCP_REF(tcp, "read");
   if (tcp->is_first_read) {
     /* Endpoint read called for the very first time. Register read callback with
@@ -913,14 +1104,15 @@ static void tcp_read(grpc_endpoint* ep, grpc_slice_buffer* incoming_buffer,
 
 /* A wrapper around sendmsg. It sends \a msg over \a fd and returns the number
  * of bytes sent. */
-ssize_t tcp_send(int fd, const struct msghdr* msg, int additional_flags = 0) {
+ssize_t tcp_send(int fd, const struct msghdr* msg, int* saved_errno,
+                 int additional_flags = 0) {
   GPR_TIMER_SCOPE("sendmsg", 1);
   ssize_t sent_length;
   do {
     /* TODO(klempner): Cork if this is a partial write */
     GRPC_STATS_INC_SYSCALL_WRITE();
     sent_length = sendmsg(fd, msg, SENDMSG_FLAGS | additional_flags);
-  } while (sent_length < 0 && errno == EINTR);
+  } while (sent_length < 0 && (*saved_errno = errno) == EINTR);
   return sent_length;
 }
 
@@ -932,7 +1124,7 @@ ssize_t tcp_send(int fd, const struct msghdr* msg, int additional_flags = 0) {
  */
 static bool tcp_write_with_timestamps(grpc_tcp* tcp, struct msghdr* msg,
                                       size_t sending_length,
-                                      ssize_t* sent_length,
+                                      ssize_t* sent_length, int* saved_errno,
                                       int additional_flags = 0);
 
 /** The callback function to be invoked when we get an error on the socket. */
@@ -976,7 +1168,7 @@ static void ZerocopyDisableAndWaitForRemaining(grpc_tcp* tcp) {
 
 static bool tcp_write_with_timestamps(grpc_tcp* tcp, struct msghdr* msg,
                                       size_t sending_length,
-                                      ssize_t* sent_length,
+                                      ssize_t* sent_length, int* saved_errno,
                                       int additional_flags) {
   if (!tcp->socket_ts_enabled) {
     uint32_t opt = grpc_core::kTimestampingSocketOptions;
@@ -1005,7 +1197,7 @@ static bool tcp_write_with_timestamps(grpc_tcp* tcp, struct msghdr* msg,
   msg->msg_controllen = CMSG_SPACE(sizeof(uint32_t));
 
   /* If there was an error on sendmsg the logic in tcp_flush will handle it. */
-  ssize_t length = tcp_send(tcp->fd, msg, additional_flags);
+  ssize_t length = tcp_send(tcp->fd, msg, saved_errno, additional_flags);
   *sent_length = length;
   /* Only save timestamps if all the bytes were taken by sendmsg. */
   if (sending_length == static_cast<size_t>(length)) {
@@ -1040,6 +1232,9 @@ static void process_zerocopy(grpc_tcp* tcp, struct cmsghdr* cmsg) {
     GPR_DEBUG_ASSERT(record);
     UnrefMaybePutZerocopySendRecord(tcp, record, seq, "CALLBACK RCVD");
   }
+  if (tcp->tcp_zerocopy_send_ctx.UpdateZeroCopyOMemStateAfterFree()) {
+    grpc_fd_set_writable(tcp->em_fd);
+  }
 }
 
 // Whether the cmsg received from error queue is of the IPv4 or IPv6 levels.
@@ -1199,7 +1394,7 @@ static void tcp_handle_error(void* arg /* grpc_tcp */,
             grpc_error_std_string(error).c_str());
   }
 
-  if (error != GRPC_ERROR_NONE ||
+  if (!GRPC_ERROR_IS_NONE(error) ||
       static_cast<bool>(gpr_atm_acq_load(&tcp->stop_error_notification))) {
     /* We aren't going to register to hear on error anymore, so it is safe to
      * unref. */
@@ -1230,6 +1425,7 @@ static void ZerocopyDisableAndWaitForRemaining(grpc_tcp* /*tcp*/) {}
 static bool tcp_write_with_timestamps(grpc_tcp* /*tcp*/, struct msghdr* /*msg*/,
                                       size_t /*sending_length*/,
                                       ssize_t* /*sent_length*/,
+                                      int* /* saved_errno */,
                                       int /*additional_flags*/) {
   gpr_log(GPR_ERROR, "Write with timestamps not supported for this platform");
   GPR_ASSERT(0);
@@ -1310,6 +1506,7 @@ static bool do_tcp_flush_zerocopy(grpc_tcp* tcp, TcpZerocopySendRecord* record,
   size_t unwind_slice_idx;
   size_t unwind_byte_idx;
   bool tried_sending_message;
+  int saved_errno;
   msghdr msg;
   // iov consumes a large space. Keep it as the last item on the stack to
   // improve locality. After all, we expect only the first elements of it being
@@ -1328,10 +1525,11 @@ static bool do_tcp_flush_zerocopy(grpc_tcp* tcp, TcpZerocopySendRecord* record,
     // Before calling sendmsg (with or without timestamps): we
     // take a single ref on the zerocopy send record.
     tcp->tcp_zerocopy_send_ctx.NoteSend(record);
+    saved_errno = 0;
     if (tcp->outgoing_buffer_arg != nullptr) {
       if (!tcp->ts_capable ||
           !tcp_write_with_timestamps(tcp, &msg, sending_length, &sent_length,
-                                     MSG_ZEROCOPY)) {
+                                     &saved_errno, MSG_ZEROCOPY)) {
         /* We could not set socket options to collect Fathom timestamps.
          * Fallback on writing without timestamps. */
         tcp->ts_capable = false;
@@ -1345,20 +1543,24 @@ static bool do_tcp_flush_zerocopy(grpc_tcp* tcp, TcpZerocopySendRecord* record,
       msg.msg_controllen = 0;
       GRPC_STATS_INC_TCP_WRITE_SIZE(sending_length);
       GRPC_STATS_INC_TCP_WRITE_IOV_SIZE(iov_size);
-      sent_length = tcp_send(tcp->fd, &msg, MSG_ZEROCOPY);
+      sent_length = tcp_send(tcp->fd, &msg, &saved_errno, MSG_ZEROCOPY);
+    }
+    if (tcp->tcp_zerocopy_send_ctx.UpdateZeroCopyOMemStateAfterSend(
+            saved_errno == ENOBUFS)) {
+      grpc_fd_set_writable(tcp->em_fd);
     }
     if (sent_length < 0) {
       // If this particular send failed, drop ref taken earlier in this method.
       tcp->tcp_zerocopy_send_ctx.UndoSend();
-      if (errno == EAGAIN) {
+      if (saved_errno == EAGAIN || saved_errno == ENOBUFS) {
         record->UnwindIfThrottled(unwind_slice_idx, unwind_byte_idx);
         return false;
-      } else if (errno == EPIPE) {
-        *error = tcp_annotate_error(GRPC_OS_ERROR(errno, "sendmsg"), tcp);
+      } else if (saved_errno == EPIPE) {
+        *error = tcp_annotate_error(GRPC_OS_ERROR(saved_errno, "sendmsg"), tcp);
         tcp_shutdown_buffer_list(tcp);
         return true;
       } else {
-        *error = tcp_annotate_error(GRPC_OS_ERROR(errno, "sendmsg"), tcp);
+        *error = tcp_annotate_error(GRPC_OS_ERROR(saved_errno, "sendmsg"), tcp);
         tcp_shutdown_buffer_list(tcp);
         return true;
       }
@@ -1402,6 +1604,7 @@ static bool tcp_flush(grpc_tcp* tcp, grpc_error_handle* error) {
   size_t trailing;
   size_t unwind_slice_idx;
   size_t unwind_byte_idx;
+  int saved_errno;
 
   // We always start at zero, because we eagerly unref and trim the slice
   // buffer as we write
@@ -1433,9 +1636,11 @@ static bool tcp_flush(grpc_tcp* tcp, grpc_error_handle* error) {
     msg.msg_iovlen = iov_size;
     msg.msg_flags = 0;
     bool tried_sending_message = false;
+    saved_errno = 0;
     if (tcp->outgoing_buffer_arg != nullptr) {
       if (!tcp->ts_capable ||
-          !tcp_write_with_timestamps(tcp, &msg, sending_length, &sent_length)) {
+          !tcp_write_with_timestamps(tcp, &msg, sending_length, &sent_length,
+                                     &saved_errno)) {
         /* We could not set socket options to collect Fathom timestamps.
          * Fallback on writing without timestamps. */
         tcp->ts_capable = false;
@@ -1451,11 +1656,11 @@ static bool tcp_flush(grpc_tcp* tcp, grpc_error_handle* error) {
       GRPC_STATS_INC_TCP_WRITE_SIZE(sending_length);
       GRPC_STATS_INC_TCP_WRITE_IOV_SIZE(iov_size);
 
-      sent_length = tcp_send(tcp->fd, &msg);
+      sent_length = tcp_send(tcp->fd, &msg, &saved_errno);
     }
 
     if (sent_length < 0) {
-      if (errno == EAGAIN) {
+      if (saved_errno == EAGAIN || saved_errno == ENOBUFS) {
         tcp->outgoing_byte_idx = unwind_byte_idx;
         // unref all and forget about all slices that have been written to this
         // point
@@ -1463,13 +1668,13 @@ static bool tcp_flush(grpc_tcp* tcp, grpc_error_handle* error) {
           grpc_slice_buffer_remove_first(tcp->outgoing_buffer);
         }
         return false;
-      } else if (errno == EPIPE) {
-        *error = tcp_annotate_error(GRPC_OS_ERROR(errno, "sendmsg"), tcp);
+      } else if (saved_errno == EPIPE) {
+        *error = tcp_annotate_error(GRPC_OS_ERROR(saved_errno, "sendmsg"), tcp);
         grpc_slice_buffer_reset_and_unref_internal(tcp->outgoing_buffer);
         tcp_shutdown_buffer_list(tcp);
         return true;
       } else {
-        *error = tcp_annotate_error(GRPC_OS_ERROR(errno, "sendmsg"), tcp);
+        *error = tcp_annotate_error(GRPC_OS_ERROR(saved_errno, "sendmsg"), tcp);
         grpc_slice_buffer_reset_and_unref_internal(tcp->outgoing_buffer);
         tcp_shutdown_buffer_list(tcp);
         return true;
@@ -1505,7 +1710,7 @@ static void tcp_handle_write(void* arg /* grpc_tcp */,
   grpc_tcp* tcp = static_cast<grpc_tcp*>(arg);
   grpc_closure* cb;
 
-  if (error != GRPC_ERROR_NONE) {
+  if (!GRPC_ERROR_IS_NONE(error)) {
     cb = tcp->write_cb;
     tcp->write_cb = nullptr;
     if (tcp->current_zerocopy_send != nullptr) {
@@ -1517,7 +1722,6 @@ static void tcp_handle_write(void* arg /* grpc_tcp */,
     TCP_UNREF(tcp, "write");
     return;
   }
-
   bool flush_result =
       tcp->current_zerocopy_send != nullptr
           ? tcp_flush_zerocopy(tcp, tcp->current_zerocopy_send, &error)
@@ -1528,7 +1732,7 @@ static void tcp_handle_write(void* arg /* grpc_tcp */,
     }
     notify_on_write(tcp);
     // tcp_flush does not populate error if it has returned false.
-    GPR_DEBUG_ASSERT(error == GRPC_ERROR_NONE);
+    GPR_DEBUG_ASSERT(GRPC_ERROR_IS_NONE(error));
   } else {
     cb = tcp->write_cb;
     tcp->write_cb = nullptr;
@@ -1543,7 +1747,7 @@ static void tcp_handle_write(void* arg /* grpc_tcp */,
 }
 
 static void tcp_write(grpc_endpoint* ep, grpc_slice_buffer* buf,
-                      grpc_closure* cb, void* arg) {
+                      grpc_closure* cb, void* arg, int /*max_frame_size*/) {
   GPR_TIMER_SCOPE("tcp_write", 0);
   grpc_tcp* tcp = reinterpret_cast<grpc_tcp*>(ep);
   grpc_error_handle error = GRPC_ERROR_NONE;
@@ -1669,8 +1873,7 @@ static const grpc_endpoint_vtable vtable = {tcp_read,
 
 grpc_endpoint* grpc_tcp_create(grpc_fd* em_fd,
                                const grpc_channel_args* channel_args,
-                               const char* peer_string,
-                               grpc_slice_allocator* slice_allocator) {
+                               absl::string_view peer_string) {
   static constexpr bool kZerocpTxEnabledDefault = false;
   int tcp_read_chunk_size = GRPC_TCP_DEFAULT_READ_SLICE_SIZE;
   int tcp_max_read_chunk_size = 4 * 1024 * 1024;
@@ -1727,35 +1930,42 @@ grpc_endpoint* grpc_tcp_create(grpc_fd* em_fd,
   grpc_tcp* tcp = new grpc_tcp(tcp_tx_zerocopy_max_simult_sends,
                                tcp_tx_zerocopy_send_bytes_thresh);
   tcp->base.vtable = &vtable;
-  tcp->peer_string = peer_string;
+  tcp->peer_string = std::string(peer_string);
   tcp->fd = grpc_fd_wrapped_fd(em_fd);
-  tcp->slice_allocator = slice_allocator;
+  tcp->memory_owner = grpc_core::ResourceQuotaFromChannelArgs(channel_args)
+                          ->memory_quota()
+                          ->CreateMemoryOwner(peer_string);
+  tcp->self_reservation = tcp->memory_owner.MakeReservation(sizeof(grpc_tcp));
   grpc_resolved_address resolved_local_addr;
   memset(&resolved_local_addr, 0, sizeof(resolved_local_addr));
   resolved_local_addr.len = sizeof(resolved_local_addr.addr);
+  absl::StatusOr<std::string> addr_uri;
   if (getsockname(tcp->fd,
                   reinterpret_cast<sockaddr*>(resolved_local_addr.addr),
-                  &resolved_local_addr.len) < 0) {
+                  &resolved_local_addr.len) < 0 ||
+      !(addr_uri = grpc_sockaddr_to_uri(&resolved_local_addr)).ok()) {
     tcp->local_address = "";
   } else {
-    tcp->local_address = grpc_sockaddr_to_uri(&resolved_local_addr);
+    tcp->local_address = addr_uri.value();
   }
   tcp->read_cb = nullptr;
   tcp->write_cb = nullptr;
   tcp->current_zerocopy_send = nullptr;
   tcp->release_fd_cb = nullptr;
   tcp->release_fd = nullptr;
-  tcp->incoming_buffer = nullptr;
   tcp->target_length = static_cast<double>(tcp_read_chunk_size);
   tcp->min_read_chunk_size = tcp_min_read_chunk_size;
   tcp->max_read_chunk_size = tcp_max_read_chunk_size;
   tcp->bytes_read_this_round = 0;
   /* Will be set to false by the very first endpoint read function */
   tcp->is_first_read = true;
+  tcp->has_posted_reclaimer = false;
   tcp->bytes_counter = -1;
   tcp->socket_ts_enabled = false;
   tcp->ts_capable = true;
   tcp->outgoing_buffer_arg = nullptr;
+  tcp->frame_size_tuning_enabled = ExperimentalTcpFrameSizeTuningEnabled();
+  tcp->min_progress_size = 1;
   if (tcp_tx_zerocopy_enabled && !tcp->tcp_zerocopy_send_ctx.memory_limited()) {
 #ifdef GRPC_LINUX_ERRQUEUE
     const int enable = 1;