grpc 1.48.0 → 1.50.0

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

@@ -24,13 +24,13 @@
 
 #ifdef GRPC_WINSOCK_SOCKET
 
+#include <grpc/event_engine/endpoint_config.h>
 #include <grpc/slice_buffer.h>
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
 #include <grpc/support/log_windows.h>
 
 #include "src/core/lib/address_utils/sockaddr_utils.h"
-#include "src/core/lib/channel/channel_args.h"
 #include "src/core/lib/iomgr/iocp_windows.h"
 #include "src/core/lib/iomgr/sockaddr.h"
 #include "src/core/lib/iomgr/sockaddr_windows.h"
@@ -38,8 +38,11 @@
 #include "src/core/lib/iomgr/tcp_client.h"
 #include "src/core/lib/iomgr/tcp_windows.h"
 #include "src/core/lib/iomgr/timer.h"
+#include "src/core/lib/resource_quota/api.h"
 #include "src/core/lib/slice/slice_internal.h"
 
+using ::grpc_event_engine::experimental::EndpointConfig;
+
 struct async_connect {
   grpc_closure* on_done;
   gpr_mu mu;
@@ -50,7 +53,6 @@ struct async_connect {
   int refs;
   grpc_closure on_connect;
   grpc_endpoint** endpoint;
-  grpc_channel_args* channel_args;
 };
 
 static void async_connect_unlock_and_cleanup(async_connect* ac,
@@ -58,7 +60,6 @@ static void async_connect_unlock_and_cleanup(async_connect* ac,
   int done = (--ac->refs == 0);
   gpr_mu_unlock(&ac->mu);
   if (done) {
-    grpc_channel_args_destroy(ac->channel_args);
     gpr_mu_destroy(&ac->mu);
     delete ac;
   }
@@ -105,7 +106,7 @@ static void on_connect(void* acp, grpc_error_handle error) {
         error = GRPC_WSA_ERROR(WSAGetLastError(), "ConnectEx");
         closesocket(socket->socket);
       } else {
-        *ep = grpc_tcp_create(socket, ac->channel_args, ac->addr_name);
+        *ep = grpc_tcp_create(socket, ac->addr_name);
         socket = nullptr;
       }
     } else {
@@ -123,7 +124,7 @@ static void on_connect(void* acp, grpc_error_handle error) {
    notification request for the connection, and one timeout alert. */
 static int64_t tcp_connect(grpc_closure* on_done, grpc_endpoint** endpoint,
                            grpc_pollset_set* interested_parties,
-                           const grpc_channel_args* channel_args,
+                           const EndpointConfig& config,
                            const grpc_resolved_address* addr,
                            grpc_core::Timestamp deadline) {
   SOCKET sock = INVALID_SOCKET;
@@ -208,7 +209,6 @@ static int64_t tcp_connect(grpc_closure* on_done, grpc_endpoint** endpoint,
   ac->refs = 2;
   ac->addr_name = addr_uri.value();
   ac->endpoint = endpoint;
-  ac->channel_args = grpc_channel_args_copy(channel_args);
   GRPC_CLOSURE_INIT(&ac->on_connect, on_connect, ac, grpc_schedule_on_exec_ctx);
 
   GRPC_CLOSURE_INIT(&ac->on_alarm, on_alarm, ac, grpc_schedule_on_exec_ctx);

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

@@ -20,6 +20,7 @@
 
 #include <grpc/impl/codegen/grpc_types.h>
 
+#include "src/core/lib/gprpp/global_config_generic.h"
 #include "src/core/lib/iomgr/port.h"
 
 #ifdef GRPC_POSIX_SOCKET_TCP
@@ -47,9 +48,9 @@
 #include <grpc/support/time.h>
 
 #include "src/core/lib/address_utils/sockaddr_utils.h"
-#include "src/core/lib/channel/channel_args.h"
 #include "src/core/lib/debug/stats.h"
 #include "src/core/lib/debug/trace.h"
+#include "src/core/lib/experiments/experiments.h"
 #include "src/core/lib/gpr/string.h"
 #include "src/core/lib/gpr/useful.h"
 #include "src/core/lib/gprpp/sync.h"
@@ -58,7 +59,6 @@
 #include "src/core/lib/iomgr/executor.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"
@@ -97,8 +97,6 @@ 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 {
@@ -234,7 +232,11 @@ 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_;
   }
 
@@ -255,8 +257,8 @@ 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);
   }
 
@@ -314,7 +316,107 @@ 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());
@@ -351,6 +453,8 @@ 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
@@ -360,22 +464,20 @@ 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) {}
+  explicit grpc_tcp(const grpc_core::PosixTcpOptions& tcp_options)
+      : min_read_chunk_size(tcp_options.tcp_min_read_chunk_size),
+        max_read_chunk_size(tcp_options.tcp_max_read_chunk_size),
+        tcp_zerocopy_send_ctx(
+            tcp_options.tcp_tx_zerocopy_max_simultaneous_sends,
+            tcp_options.tcp_tx_zerocopy_send_bytes_threshold) {}
   grpc_endpoint base;
   grpc_fd* em_fd;
   int fd;
   /* Used by the endpoint read function to distinguish the very first read call
    * from the rest */
   bool is_first_read;
-  bool has_posted_reclaimer;
+  bool has_posted_reclaimer ABSL_GUARDED_BY(read_mu) = false;
   double target_length;
   double bytes_read_this_round;
   grpc_core::RefCount refcount;
@@ -383,6 +485,7 @@ struct grpc_tcp {
 
   int min_read_chunk_size;
   int max_read_chunk_size;
+  int set_rcvlowat = 0;
 
   /* garbage after the last read */
   grpc_slice_buffer last_read_buffer;
@@ -479,8 +582,7 @@ static void run_poller(void* bp, grpc_error_handle /*error_ignored*/) {
   }
   gpr_mu_lock(p->pollset_mu);
   grpc_core::Timestamp deadline =
-      grpc_core::ExecCtx::Get()->Now() + grpc_core::Duration::Seconds(10);
-  GRPC_STATS_INC_TCP_BACKUP_POLLER_POLLS();
+      grpc_core::Timestamp::Now() + grpc_core::Duration::Seconds(10);
   GRPC_LOG_IF_ERROR(
       "backup_poller:pollset_work",
       grpc_pollset_work(BACKUP_POLLER_POLLSET(p), nullptr, deadline));
@@ -541,7 +643,6 @@ static void cover_self(grpc_tcp* tcp) {
     g_backup_poller = p;
     grpc_pollset_init(BACKUP_POLLER_POLLSET(p), &p->pollset_mu);
     g_backup_poller_mu->Unlock();
-    GRPC_STATS_INC_TCP_BACKUP_POLLERS_CREATED();
     if (GRPC_TRACE_FLAG_ENABLED(grpc_tcp_trace)) {
       gpr_log(GPR_INFO, "BACKUP_POLLER:%p create", p);
     }
@@ -687,8 +788,8 @@ static void perform_reclamation(grpc_tcp* tcp)
   if (tcp->incoming_buffer != nullptr) {
     grpc_slice_buffer_reset_and_unref_internal(tcp->incoming_buffer);
   }
-  tcp->read_mu.Unlock();
   tcp->has_posted_reclaimer = false;
+  tcp->read_mu.Unlock();
 }
 
 static void maybe_post_reclaimer(grpc_tcp* tcp)
@@ -723,11 +824,54 @@ static void tcp_trace_read(grpc_tcp* tcp, grpc_error_handle error)
   }
 }
 
+static void update_rcvlowat(grpc_tcp* tcp)
+    ABSL_EXCLUSIVE_LOCKS_REQUIRED(tcp->read_mu) {
+  if (!grpc_core::IsTcpRcvLowatEnabled()) return;
+
+  // TODO(ctiller): Check if supported by OS.
+  // TODO(ctiller): Allow some adjustments instead of hardcoding things.
+
+  static constexpr int kRcvLowatMax = 16 * 1024 * 1024;
+  static constexpr int kRcvLowatThreshold = 16 * 1024;
+
+  int remaining = std::min(static_cast<int>(tcp->incoming_buffer->length),
+                           tcp->min_progress_size);
+  remaining = std::min(remaining, kRcvLowatMax);
+
+  // Setting SO_RCVLOWAT for small quantities does not save on CPU.
+  if (remaining < kRcvLowatThreshold) {
+    remaining = 0;
+  }
+
+  // If zerocopy is off, wake shortly before the full RPC is here. More can
+  // show up partway through recvmsg() since it takes a while to copy data.
+  // So an early wakeup aids latency.
+  if (!tcp->tcp_zerocopy_send_ctx.enabled() && remaining > 0) {
+    remaining -= kRcvLowatThreshold;
+  }
+
+  // We still do not know the RPC size. Do not set SO_RCVLOWAT.
+  if (tcp->set_rcvlowat <= 1 && remaining <= 1) return;
+
+  // Previous value is still valid. No change needed in SO_RCVLOWAT.
+  if (tcp->set_rcvlowat == remaining) {
+    return;
+  }
+  if (setsockopt(tcp->fd, SOL_SOCKET, SO_RCVLOWAT, &remaining,
+                 sizeof(remaining)) != 0) {
+    gpr_log(GPR_ERROR, "%s",
+            absl::StrCat("Cannot set SO_RCVLOWAT on fd=", tcp->fd,
+                         " err=", strerror(errno))
+                .c_str());
+    return;
+  }
+  tcp->set_rcvlowat = remaining;
+}
+
 /* Returns true if data available to read or error other than EAGAIN. */
-#define MAX_READ_IOVEC 4
+#define MAX_READ_IOVEC 64
 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);
   }
@@ -775,7 +919,6 @@ static bool tcp_do_read(grpc_tcp* tcp, grpc_error_handle* error)
     GRPC_STATS_INC_TCP_READ_OFFER_IOV_SIZE(tcp->incoming_buffer->count);
 
     do {
-      GPR_TIMER_SCOPE("recvmsg", 0);
       GRPC_STATS_INC_SYSCALL_READ();
       read_bytes = recvmsg(tcp->fd, &msg, 0);
     } while (read_bytes < 0 && errno == EINTR);
@@ -904,31 +1047,66 @@ static bool tcp_do_read(grpc_tcp* tcp, grpc_error_handle* error)
 
 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; 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 (grpc_core::IsTcpReadChunksEnabled()) {
+    static const int kBigAlloc = 64 * 1024;
+    static const int kSmallAlloc = 8 * 1024;
+    if (tcp->incoming_buffer->length <
+        static_cast<size_t>(tcp->min_progress_size)) {
+      size_t allocate_length = tcp->min_progress_size;
+      const size_t target_length = static_cast<size_t>(tcp->target_length);
+      // If memory pressure is low and we think there will be more than
+      // min_progress_size bytes to read, allocate a bit more.
+      const bool low_memory_pressure =
+          tcp->memory_owner.GetPressureInfo().pressure_control_value < 0.8;
+      if (low_memory_pressure && target_length > allocate_length) {
+        allocate_length = target_length;
+      }
+      int extra_wanted =
+          allocate_length - static_cast<int>(tcp->incoming_buffer->length);
+      if (extra_wanted >=
+          (low_memory_pressure ? kSmallAlloc * 3 / 2 : kBigAlloc)) {
+        while (extra_wanted > 0) {
+          extra_wanted -= kBigAlloc;
+          grpc_slice_buffer_add_indexed(tcp->incoming_buffer,
+                                        tcp->memory_owner.MakeSlice(kBigAlloc));
+          GRPC_STATS_INC_TCP_READ_ALLOC_64K();
+        }
+      } else {
+        while (extra_wanted > 0) {
+          extra_wanted -= kSmallAlloc;
+          grpc_slice_buffer_add_indexed(
+              tcp->incoming_buffer, tcp->memory_owner.MakeSlice(kSmallAlloc));
+          GRPC_STATS_INC_TCP_READ_ALLOC_8K();
+        }
+      }
+      maybe_post_reclaimer(tcp);
+    }
+  } else {
+    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; 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);
+      }
+      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 slice = tcp->memory_owner.MakeSlice(grpc_core::MemoryRequest(
+          min_read_chunk_size,
+          grpc_core::Clamp(extra_wanted, min_read_chunk_size,
+                           max_read_chunk_size)));
+      grpc_slice_buffer_add_indexed(tcp->incoming_buffer, slice);
+      maybe_post_reclaimer(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);
   }
 }
 
@@ -944,6 +1122,7 @@ static void tcp_handle_read(void* arg /* grpc_tcp */, grpc_error_handle error) {
     maybe_make_read_slices(tcp);
     if (!tcp_do_read(tcp, &tcp_read_error)) {
       /* We've consumed the edge, request a new one */
+      update_rcvlowat(tcp);
       tcp->read_mu.Unlock();
       notify_on_read(tcp);
       return;
@@ -973,19 +1152,23 @@ static void tcp_read(grpc_endpoint* ep, grpc_slice_buffer* incoming_buffer,
       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) {
+    update_rcvlowat(tcp);
+    tcp->read_mu.Unlock();
     /* Endpoint read called for the very first time. Register read callback with
      * the polling engine */
     tcp->is_first_read = false;
     notify_on_read(tcp);
   } else if (!urgent && tcp->inq == 0) {
+    update_rcvlowat(tcp);
+    tcp->read_mu.Unlock();
     /* Upper layer asked to read more but we know there is no pending data
      * to read from previous reads. So, wait for POLLIN.
      */
     notify_on_read(tcp);
   } else {
+    tcp->read_mu.Unlock();
     /* Not the first time. We may or may not have more bytes available. In any
      * case call tcp->read_done_closure (i.e tcp_handle_read()) which does the
      * right thing (i.e calls tcp_do_read() which either reads the available
@@ -1000,7 +1183,6 @@ static void tcp_read(grpc_endpoint* ep, grpc_slice_buffer* incoming_buffer,
  * of bytes sent. */
 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 */
@@ -1126,6 +1308,9 @@ 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.
@@ -1436,10 +1621,14 @@ static bool do_tcp_flush_zerocopy(grpc_tcp* tcp, TcpZerocopySendRecord* record,
       GRPC_STATS_INC_TCP_WRITE_IOV_SIZE(iov_size);
       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 (saved_errno == EAGAIN) {
+      if (saved_errno == EAGAIN || saved_errno == ENOBUFS) {
         record->UnwindIfThrottled(unwind_slice_idx, unwind_byte_idx);
         return false;
       } else if (saved_errno == EPIPE) {
@@ -1547,7 +1736,7 @@ static bool tcp_flush(grpc_tcp* tcp, grpc_error_handle* error) {
     }
 
     if (sent_length < 0) {
-      if (saved_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
@@ -1609,7 +1798,6 @@ 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)
@@ -1636,7 +1824,6 @@ 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, 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;
   TcpZerocopySendRecord* zerocopy_send_record = nullptr;
@@ -1757,72 +1944,16 @@ static const grpc_endpoint_vtable vtable = {tcp_read,
                                             tcp_get_fd,
                                             tcp_can_track_err};
 
-#define MAX_CHUNK_SIZE (32 * 1024 * 1024)
-
 grpc_endpoint* grpc_tcp_create(grpc_fd* em_fd,
-                               const grpc_channel_args* channel_args,
+                               const grpc_core::PosixTcpOptions& options,
                                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;
-  int tcp_min_read_chunk_size = 256;
-  bool tcp_tx_zerocopy_enabled = kZerocpTxEnabledDefault;
-  int tcp_tx_zerocopy_send_bytes_thresh =
-      grpc_core::TcpZerocopySendCtx::kDefaultSendBytesThreshold;
-  int tcp_tx_zerocopy_max_simult_sends =
-      grpc_core::TcpZerocopySendCtx::kDefaultMaxSends;
-  if (channel_args != nullptr) {
-    for (size_t i = 0; i < channel_args->num_args; i++) {
-      if (0 ==
-          strcmp(channel_args->args[i].key, GRPC_ARG_TCP_READ_CHUNK_SIZE)) {
-        grpc_integer_options options = {tcp_read_chunk_size, 1, MAX_CHUNK_SIZE};
-        tcp_read_chunk_size =
-            grpc_channel_arg_get_integer(&channel_args->args[i], options);
-      } else if (0 == strcmp(channel_args->args[i].key,
-                             GRPC_ARG_TCP_MIN_READ_CHUNK_SIZE)) {
-        grpc_integer_options options = {tcp_read_chunk_size, 1, MAX_CHUNK_SIZE};
-        tcp_min_read_chunk_size =
-            grpc_channel_arg_get_integer(&channel_args->args[i], options);
-      } else if (0 == strcmp(channel_args->args[i].key,
-                             GRPC_ARG_TCP_MAX_READ_CHUNK_SIZE)) {
-        grpc_integer_options options = {tcp_read_chunk_size, 1, MAX_CHUNK_SIZE};
-        tcp_max_read_chunk_size =
-            grpc_channel_arg_get_integer(&channel_args->args[i], options);
-      } else if (0 == strcmp(channel_args->args[i].key,
-                             GRPC_ARG_TCP_TX_ZEROCOPY_ENABLED)) {
-        tcp_tx_zerocopy_enabled = grpc_channel_arg_get_bool(
-            &channel_args->args[i], kZerocpTxEnabledDefault);
-      } else if (0 == strcmp(channel_args->args[i].key,
-                             GRPC_ARG_TCP_TX_ZEROCOPY_SEND_BYTES_THRESHOLD)) {
-        grpc_integer_options options = {
-            grpc_core::TcpZerocopySendCtx::kDefaultSendBytesThreshold, 0,
-            INT_MAX};
-        tcp_tx_zerocopy_send_bytes_thresh =
-            grpc_channel_arg_get_integer(&channel_args->args[i], options);
-      } else if (0 == strcmp(channel_args->args[i].key,
-                             GRPC_ARG_TCP_TX_ZEROCOPY_MAX_SIMULT_SENDS)) {
-        grpc_integer_options options = {
-            grpc_core::TcpZerocopySendCtx::kDefaultMaxSends, 0, INT_MAX};
-        tcp_tx_zerocopy_max_simult_sends =
-            grpc_channel_arg_get_integer(&channel_args->args[i], options);
-      }
-    }
-  }
-
-  if (tcp_min_read_chunk_size > tcp_max_read_chunk_size) {
-    tcp_min_read_chunk_size = tcp_max_read_chunk_size;
-  }
-  tcp_read_chunk_size = grpc_core::Clamp(
-      tcp_read_chunk_size, tcp_min_read_chunk_size, tcp_max_read_chunk_size);
-
-  grpc_tcp* tcp = new grpc_tcp(tcp_tx_zerocopy_max_simult_sends,
-                               tcp_tx_zerocopy_send_bytes_thresh);
+  grpc_tcp* tcp = new grpc_tcp(options);
   tcp->base.vtable = &vtable;
   tcp->peer_string = std::string(peer_string);
   tcp->fd = grpc_fd_wrapped_fd(em_fd);
-  tcp->memory_owner = grpc_core::ResourceQuotaFromChannelArgs(channel_args)
-                          ->memory_quota()
-                          ->CreateMemoryOwner(peer_string);
+  GPR_ASSERT(options.resource_quota != nullptr);
+  tcp->memory_owner =
+      options.resource_quota->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));
@@ -1841,20 +1972,18 @@ grpc_endpoint* grpc_tcp_create(grpc_fd* em_fd,
   tcp->current_zerocopy_send = nullptr;
   tcp->release_fd_cb = nullptr;
   tcp->release_fd = 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->target_length = static_cast<double>(options.tcp_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->frame_size_tuning_enabled = grpc_core::IsTcpFrameSizeTuningEnabled();
   tcp->min_progress_size = 1;
-  if (tcp_tx_zerocopy_enabled && !tcp->tcp_zerocopy_send_ctx.memory_limited()) {
+  if (options.tcp_tx_zero_copy_enabled &&
+      !tcp->tcp_zerocopy_send_ctx.memory_limited()) {
 #ifdef GRPC_LINUX_ERRQUEUE
     const int enable = 1;
     auto err =

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

@@ -36,12 +36,14 @@
 #include "src/core/lib/iomgr/endpoint.h"
 #include "src/core/lib/iomgr/ev_posix.h"
 #include "src/core/lib/iomgr/port.h"
+#include "src/core/lib/iomgr/socket_utils_posix.h"
 
 extern grpc_core::TraceFlag grpc_tcp_trace;
 
 /// Create a tcp endpoint given a file desciptor and a read slice size.
 /// Takes ownership of \a fd. Takes ownership of the \a slice_allocator.
-grpc_endpoint* grpc_tcp_create(grpc_fd* fd, const grpc_channel_args* args,
+grpc_endpoint* grpc_tcp_create(grpc_fd* fd,
+                               const grpc_core::PosixTcpOptions& options,
                                absl::string_view peer_string);
 
 /// Return the tcp endpoint's fd, or -1 if this is not available. Does not