grpc 1.0.0.pre1 → 1.0.0.pre2

data/src/core/ext/transport/chttp2/transport/internal.h

@@ -305,6 +305,22 @@ typedef struct grpc_chttp2_executor_action_header {
   void *arg;
 } grpc_chttp2_executor_action_header;
 
+typedef enum {
+  /** no writing activity */
+  GRPC_CHTTP2_WRITING_INACTIVE,
+  /** write has been requested, but not scheduled yet */
+  GRPC_CHTTP2_WRITE_REQUESTED_WITH_POLLER,
+  GRPC_CHTTP2_WRITE_REQUESTED_NO_POLLER,
+  /** write has been requested and scheduled against the workqueue */
+  GRPC_CHTTP2_WRITE_SCHEDULED,
+  /** write has been initiated after being reaped from the workqueue */
+  GRPC_CHTTP2_WRITING,
+  /** write has been initiated, AND another write needs to be started once it's
+      done */
+  GRPC_CHTTP2_WRITING_STALE_WITH_POLLER,
+  GRPC_CHTTP2_WRITING_STALE_NO_POLLER,
+} grpc_chttp2_write_state;
+
 struct grpc_chttp2_transport {
   grpc_transport base; /* must be first */
   gpr_refcount refs;
@@ -319,10 +335,10 @@ struct grpc_chttp2_transport {
 
     /** is a thread currently in the global lock */
     bool global_active;
-    /** is a thread currently writing */
-    bool writing_active;
     /** is a thread currently parsing */
     bool parsing_active;
+    /** write execution state of the transport */
+    grpc_chttp2_write_state write_state;
 
     grpc_chttp2_executor_action_header *pending_actions_head;
     grpc_chttp2_executor_action_header *pending_actions_tail;
@@ -342,7 +358,8 @@ struct grpc_chttp2_transport {
   /** global state for reading/writing */
   grpc_chttp2_transport_global global;
   /** state only accessible by the chain of execution that
-      set writing_active=1 */
+      set writing_state >= GRPC_WRITING, and only by the writing closure
+      chain. */
   grpc_chttp2_transport_writing writing;
   /** state only accessible by the chain of execution that
       set parsing_active=1 */
@@ -363,6 +380,8 @@ struct grpc_chttp2_transport {
   grpc_closure reading_action;
   /** closure to actually do parsing */
   grpc_closure parsing_action;
+  /** closure to initiate writing */
+  grpc_closure initiate_writing;
 
   /** incoming read bytes */
   gpr_slice_buffer read_buffer;
@@ -436,8 +455,10 @@ typedef struct {
   bool seen_error;
   bool exceeded_metadata_size;
 
-  /** the error that resulted in this stream being removed */
-  grpc_error *removal_error;
+  /** the error that resulted in this stream being read-closed */
+  grpc_error *read_closed_error;
+  /** the error that resulted in this stream being write-closed */
+  grpc_error *write_closed_error;
 
   bool published_initial_metadata;
   bool published_trailing_metadata;
@@ -514,15 +535,20 @@ struct grpc_chttp2_stream {
 };
 
 /** Transport writing call flow:
-    chttp2_transport.c calls grpc_chttp2_unlocking_check_writes to see if writes
-   are required;
-    if they are, chttp2_transport.c calls grpc_chttp2_perform_writes to do the
-   writes.
-    Once writes have been completed (meaning another write could potentially be
-   started),
-    grpc_chttp2_terminate_writing is called. This will call
-   grpc_chttp2_cleanup_writing, at which
-    point the write phase is complete. */
+    grpc_chttp2_initiate_write() is called anywhere that we know bytes need to
+    go out on the wire.
+    If no other write has been started, a task is enqueued onto our workqueue.
+    When that task executes, it obtains the global lock, and gathers the data
+    to write.
+    The global lock is dropped and we do the syscall to write.
+    After writing, a follow-up check is made to see if another round of writing
+    should be performed.
+
+    The actual call chain is documented in the implementation of this function.
+    */
+void grpc_chttp2_initiate_write(grpc_exec_ctx *exec_ctx,
+                                grpc_chttp2_transport_global *transport_global,
+                                bool covered_by_poller, const char *reason);
 
 /** Someone is unlocking the transport mutex: check to see if writes
     are required, and schedule them if so */
@@ -610,9 +636,8 @@ int grpc_chttp2_list_pop_check_read_ops(
 void grpc_chttp2_list_add_writing_stalled_by_transport(
     grpc_chttp2_transport_writing *transport_writing,
     grpc_chttp2_stream_writing *stream_writing);
-void grpc_chttp2_list_flush_writing_stalled_by_transport(
-    grpc_exec_ctx *exec_ctx, grpc_chttp2_transport_writing *transport_writing,
-    bool is_window_available);
+bool grpc_chttp2_list_flush_writing_stalled_by_transport(
+    grpc_exec_ctx *exec_ctx, grpc_chttp2_transport_writing *transport_writing);
 
 void grpc_chttp2_list_add_stalled_by_transport(
     grpc_chttp2_transport_writing *transport_writing,
@@ -822,7 +847,9 @@ void grpc_chttp2_ack_ping(grpc_exec_ctx *exec_ctx,
 
 /** add a ref to the stream and add it to the writable list;
     ref will be dropped in writing.c */
-void grpc_chttp2_become_writable(grpc_chttp2_transport_global *transport_global,
-                                 grpc_chttp2_stream_global *stream_global);
+void grpc_chttp2_become_writable(grpc_exec_ctx *exec_ctx,
+                                 grpc_chttp2_transport_global *transport_global,
+                                 grpc_chttp2_stream_global *stream_global,
+                                 bool covered_by_poller, const char *reason);
 
 #endif /* GRPC_CORE_EXT_TRANSPORT_CHTTP2_TRANSPORT_INTERNAL_H */

data/src/core/ext/transport/chttp2/transport/parsing.c

@@ -154,10 +154,8 @@ void grpc_chttp2_publish_reads(
                                   transport_parsing, outgoing_window);
   is_zero = transport_global->outgoing_window <= 0;
   if (was_zero && !is_zero) {
-    while (grpc_chttp2_list_pop_stalled_by_transport(transport_global,
-                                                     &stream_global)) {
-      grpc_chttp2_become_writable(transport_global, stream_global);
-    }
+    grpc_chttp2_initiate_write(exec_ctx, transport_global, false,
+                               "new_global_flow_control");
   }
 
   if (transport_parsing->incoming_window <
@@ -168,6 +166,8 @@ void grpc_chttp2_publish_reads(
                                       announce_incoming_window, announce_bytes);
     GRPC_CHTTP2_FLOW_CREDIT_TRANSPORT("parsed", transport_parsing,
                                       incoming_window, announce_bytes);
+    grpc_chttp2_initiate_write(exec_ctx, transport_global, false,
+                               "global incoming window");
   }
 
   /* for each stream that saw an update, fixup global state */
@@ -190,7 +190,8 @@ void grpc_chttp2_publish_reads(
                                  outgoing_window);
     is_zero = stream_global->outgoing_window <= 0;
     if (was_zero && !is_zero) {
-      grpc_chttp2_become_writable(transport_global, stream_global);
+      grpc_chttp2_become_writable(exec_ctx, transport_global, stream_global,
+                                  false, "stream.read_flow_control");
     }
 
     stream_global->max_recv_bytes -= (uint32_t)GPR_MIN(

data/src/core/ext/transport/chttp2/transport/stream_lists.c

@@ -329,6 +329,7 @@ void grpc_chttp2_list_add_writing_stalled_by_transport(
     grpc_chttp2_transport_writing *transport_writing,
     grpc_chttp2_stream_writing *stream_writing) {
   grpc_chttp2_stream *stream = STREAM_FROM_WRITING(stream_writing);
+  gpr_log(GPR_DEBUG, "writing stalled %d", stream->global.id);
   if (!stream->included[GRPC_CHTTP2_LIST_WRITING_STALLED_BY_TRANSPORT]) {
     GRPC_CHTTP2_STREAM_REF(&stream->global, "chttp2_writing_stalled");
   }
@@ -336,27 +337,28 @@ void grpc_chttp2_list_add_writing_stalled_by_transport(
                   GRPC_CHTTP2_LIST_WRITING_STALLED_BY_TRANSPORT);
 }
 
-void grpc_chttp2_list_flush_writing_stalled_by_transport(
-    grpc_exec_ctx *exec_ctx, grpc_chttp2_transport_writing *transport_writing,
-    bool is_window_available) {
+bool grpc_chttp2_list_flush_writing_stalled_by_transport(
+    grpc_exec_ctx *exec_ctx, grpc_chttp2_transport_writing *transport_writing) {
   grpc_chttp2_stream *stream;
+  bool out = false;
   grpc_chttp2_transport *transport = TRANSPORT_FROM_WRITING(transport_writing);
   while (stream_list_pop(transport, &stream,
                          GRPC_CHTTP2_LIST_WRITING_STALLED_BY_TRANSPORT)) {
-    if (is_window_available) {
-      grpc_chttp2_become_writable(&transport->global, &stream->global);
-    } else {
-      grpc_chttp2_list_add_stalled_by_transport(transport_writing,
-                                                &stream->writing);
-    }
+    gpr_log(GPR_DEBUG, "move %d from writing stalled to just stalled",
+            stream->global.id);
+    grpc_chttp2_list_add_stalled_by_transport(transport_writing,
+                                              &stream->writing);
     GRPC_CHTTP2_STREAM_UNREF(exec_ctx, &stream->global,
                              "chttp2_writing_stalled");
+    out = true;
   }
+  return out;
 }
 
 void grpc_chttp2_list_add_stalled_by_transport(
     grpc_chttp2_transport_writing *transport_writing,
     grpc_chttp2_stream_writing *stream_writing) {
+  gpr_log(GPR_DEBUG, "stalled %d", stream_writing->id);
   stream_list_add(TRANSPORT_FROM_WRITING(transport_writing),
                   STREAM_FROM_WRITING(stream_writing),
                   GRPC_CHTTP2_LIST_STALLED_BY_TRANSPORT);

data/src/core/ext/transport/chttp2/transport/writing.c

@@ -75,9 +75,13 @@ int grpc_chttp2_unlocking_check_writes(
 
   GRPC_CHTTP2_FLOW_MOVE_TRANSPORT("write", transport_writing, outgoing_window,
                                   transport_global, outgoing_window);
-  bool is_window_available = transport_writing->outgoing_window > 0;
-  grpc_chttp2_list_flush_writing_stalled_by_transport(
-      exec_ctx, transport_writing, is_window_available);
+  if (transport_writing->outgoing_window > 0) {
+    while (grpc_chttp2_list_pop_stalled_by_transport(transport_global,
+                                                     &stream_global)) {
+      grpc_chttp2_become_writable(exec_ctx, transport_global, stream_global,
+                                  false, "transport.read_flow_control");
+    }
+  }
 
   /* for each grpc_chttp2_stream that's become writable, frame it's data
      (according to available window sizes) and add to the output buffer */
@@ -331,6 +335,12 @@ void grpc_chttp2_cleanup_writing(
   grpc_chttp2_stream_writing *stream_writing;
   grpc_chttp2_stream_global *stream_global;
 
+  if (grpc_chttp2_list_flush_writing_stalled_by_transport(exec_ctx,
+                                                          transport_writing)) {
+    grpc_chttp2_initiate_write(exec_ctx, transport_global, false,
+                               "resume_stalled_stream");
+  }
+
   while (grpc_chttp2_list_pop_written_stream(
       transport_global, transport_writing, &stream_global, &stream_writing)) {
     if (stream_writing->sent_initial_metadata) {

data/src/core/lib/iomgr/endpoint.c

@@ -65,3 +65,7 @@ void grpc_endpoint_destroy(grpc_exec_ctx* exec_ctx, grpc_endpoint* ep) {
 char* grpc_endpoint_get_peer(grpc_endpoint* ep) {
   return ep->vtable->get_peer(ep);
 }
+
+grpc_workqueue* grpc_endpoint_get_workqueue(grpc_endpoint* ep) {
+  return ep->vtable->get_workqueue(ep);
+}

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

@@ -51,6 +51,7 @@ struct grpc_endpoint_vtable {
                gpr_slice_buffer *slices, grpc_closure *cb);
   void (*write)(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
                 gpr_slice_buffer *slices, grpc_closure *cb);
+  grpc_workqueue *(*get_workqueue)(grpc_endpoint *ep);
   void (*add_to_pollset)(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
                          grpc_pollset *pollset);
   void (*add_to_pollset_set)(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
@@ -69,6 +70,9 @@ void grpc_endpoint_read(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
 
 char *grpc_endpoint_get_peer(grpc_endpoint *ep);
 
+/* Retrieve a reference to the workqueue associated with this endpoint */
+grpc_workqueue *grpc_endpoint_get_workqueue(grpc_endpoint *ep);
+
 /* Write slices out to the socket.
 
    If the connection is ready for more data after the end of the call, it

data/src/core/lib/iomgr/ev_epoll_linux.c

@@ -57,6 +57,7 @@
 #include "src/core/lib/iomgr/ev_posix.h"
 #include "src/core/lib/iomgr/iomgr_internal.h"
 #include "src/core/lib/iomgr/wakeup_fd_posix.h"
+#include "src/core/lib/iomgr/workqueue.h"
 #include "src/core/lib/profiling/timers.h"
 #include "src/core/lib/support/block_annotate.h"
 
@@ -113,9 +114,7 @@ struct grpc_fd {
   grpc_closure *read_closure;
   grpc_closure *write_closure;
 
-  /* The polling island to which this fd belongs to and the mutex protecting the
-     the field */
-  gpr_mu pi_mu;
+  /* The polling island to which this fd belongs to (protected by mu) */
   struct polling_island *polling_island;
 
   struct grpc_fd *freelist_next;
@@ -152,16 +151,17 @@ static void fd_global_shutdown(void);
  * Polling island Declarations
  */
 
-// #define GRPC_PI_REF_COUNT_DEBUG
+//#define GRPC_PI_REF_COUNT_DEBUG
 #ifdef GRPC_PI_REF_COUNT_DEBUG
 
 #define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__)
-#define PI_UNREF(p, r) pi_unref_dbg((p), (r), __FILE__, __LINE__)
+#define PI_UNREF(exec_ctx, p, r) \
+  pi_unref_dbg((exec_ctx), (p), (r), __FILE__, __LINE__)
 
 #else /* defined(GRPC_PI_REF_COUNT_DEBUG) */
 
 #define PI_ADD_REF(p, r) pi_add_ref((p))
-#define PI_UNREF(p, r) pi_unref((p))
+#define PI_UNREF(exec_ctx, p, r) pi_unref((exec_ctx), (p))
 
 #endif /* !defined(GPRC_PI_REF_COUNT_DEBUG) */
 
@@ -172,7 +172,7 @@ typedef struct polling_island {
      Once the ref count becomes zero, this structure is destroyed which means
      we should ensure that there is never a scenario where a PI_ADD_REF() is
      racing with a PI_UNREF() that just made the ref_count zero. */
-  gpr_refcount ref_count;
+  gpr_atm ref_count;
 
   /* Pointer to the polling_island this merged into.
    * merged_to value is only set once in polling_island's lifetime (and that too
@@ -184,6 +184,9 @@ typedef struct polling_island {
    * (except mu and ref_count) are invalid and must be ignored. */
   gpr_atm merged_to;
 
+  /* The workqueue associated with this polling island */
+  grpc_workqueue *workqueue;
+
   /* The fd of the underlying epoll set */
   int epoll_fd;
 
@@ -191,11 +194,6 @@ typedef struct polling_island {
   size_t fd_cnt;
   size_t fd_capacity;
   grpc_fd **fds;
-
-  /* Polling islands that are no longer needed are kept in a freelist so that
-     they can be reused. This field points to the next polling island in the
-     free list */
-  struct polling_island *next_free;
 } polling_island;
 
 /*******************************************************************************
@@ -253,13 +251,14 @@ struct grpc_pollset_set {
  * Common helpers
  */
 
-static void append_error(grpc_error **composite, grpc_error *error,
+static bool append_error(grpc_error **composite, grpc_error *error,
                          const char *desc) {
-  if (error == GRPC_ERROR_NONE) return;
+  if (error == GRPC_ERROR_NONE) return true;
   if (*composite == GRPC_ERROR_NONE) {
     *composite = GRPC_ERROR_CREATE(desc);
   }
   *composite = grpc_error_add_child(*composite, error);
+  return false;
 }
 
 /*******************************************************************************
@@ -275,11 +274,8 @@ static void append_error(grpc_error **composite, grpc_error *error,
    threads that woke up MUST NOT call grpc_wakeup_fd_consume_wakeup() */
 static grpc_wakeup_fd polling_island_wakeup_fd;
 
-/* Polling island freelist */
-static gpr_mu g_pi_freelist_mu;
-static polling_island *g_pi_freelist = NULL;
-
-static void polling_island_delete(); /* Forward declaration */
+/* Forward declaration */
+static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi);
 
 #ifdef GRPC_TSAN
 /* Currently TSAN may incorrectly flag data races between epoll_ctl and
@@ -293,28 +289,35 @@ gpr_atm g_epoll_sync;
 #endif /* defined(GRPC_TSAN) */
 
 #ifdef GRPC_PI_REF_COUNT_DEBUG
-void pi_add_ref(polling_island *pi);
-void pi_unref(polling_island *pi);
+static void pi_add_ref(polling_island *pi);
+static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi);
 
-void pi_add_ref_dbg(polling_island *pi, char *reason, char *file, int line) {
-  long old_cnt = gpr_atm_acq_load(&(pi->ref_count.count));
+static void pi_add_ref_dbg(polling_island *pi, char *reason, char *file,
+                           int line) {
+  long old_cnt = gpr_atm_acq_load(&pi->ref_count);
   pi_add_ref(pi);
   gpr_log(GPR_DEBUG, "Add ref pi: %p, old: %ld -> new:%ld (%s) - (%s, %d)",
           (void *)pi, old_cnt, old_cnt + 1, reason, file, line);
 }
 
-void pi_unref_dbg(polling_island *pi, char *reason, char *file, int line) {
-  long old_cnt = gpr_atm_acq_load(&(pi->ref_count.count));
-  pi_unref(pi);
+static void pi_unref_dbg(grpc_exec_ctx *exec_ctx, polling_island *pi,
+                         char *reason, char *file, int line) {
+  long old_cnt = gpr_atm_acq_load(&pi->ref_count);
+  pi_unref(exec_ctx, pi);
   gpr_log(GPR_DEBUG, "Unref pi: %p, old:%ld -> new:%ld (%s) - (%s, %d)",
           (void *)pi, old_cnt, (old_cnt - 1), reason, file, line);
 }
 #endif
 
-void pi_add_ref(polling_island *pi) { gpr_ref(&pi->ref_count); }
+static void pi_add_ref(polling_island *pi) {
+  gpr_atm_no_barrier_fetch_add(&pi->ref_count, 1);
+}
 
-void pi_unref(polling_island *pi) {
-  /* If ref count went to zero, delete the polling island.
+static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) {
+  /* If ref count went to one, we're back to just the workqueue owning a ref.
+     Unref the workqueue to break the loop.
+
+     If ref count went to zero, delete the polling island.
      Note that this deletion not be done under a lock. Once the ref count goes
      to zero, we are guaranteed that no one else holds a reference to the
      polling island (and that there is no racing pi_add_ref() call either).
@@ -322,12 +325,20 @@ void pi_unref(polling_island *pi) {
      Also, if we are deleting the polling island and the merged_to field is
      non-empty, we should remove a ref to the merged_to polling island
    */
-  if (gpr_unref(&pi->ref_count)) {
-    polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
-    polling_island_delete(pi);
-    if (next != NULL) {
-      PI_UNREF(next, "pi_delete"); /* Recursive call */
+  switch (gpr_atm_full_fetch_add(&pi->ref_count, -1)) {
+    case 2: /* last external ref: the only one now owned is by the workqueue */
+      GRPC_WORKQUEUE_UNREF(exec_ctx, pi->workqueue, "polling_island");
+      break;
+    case 1: {
+      polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
+      polling_island_delete(exec_ctx, pi);
+      if (next != NULL) {
+        PI_UNREF(exec_ctx, next, "pi_delete"); /* Recursive call */
+      }
+      break;
     }
+    case 0:
+      GPR_UNREACHABLE_CODE(return );
   }
 }
 
@@ -462,69 +473,68 @@ static void polling_island_remove_fd_locked(polling_island *pi, grpc_fd *fd,
 }
 
 /* Might return NULL in case of an error */
-static polling_island *polling_island_create(grpc_fd *initial_fd,
+static polling_island *polling_island_create(grpc_exec_ctx *exec_ctx,
+                                             grpc_fd *initial_fd,
                                              grpc_error **error) {
   polling_island *pi = NULL;
-  char *err_msg;
   const char *err_desc = "polling_island_create";
 
-  /* Try to get one from the polling island freelist */
-  gpr_mu_lock(&g_pi_freelist_mu);
-  if (g_pi_freelist != NULL) {
-    pi = g_pi_freelist;
-    g_pi_freelist = g_pi_freelist->next_free;
-    pi->next_free = NULL;
-  }
-  gpr_mu_unlock(&g_pi_freelist_mu);
+  *error = GRPC_ERROR_NONE;
 
-  /* Create new polling island if we could not get one from the free list */
-  if (pi == NULL) {
-    pi = gpr_malloc(sizeof(*pi));
-    gpr_mu_init(&pi->mu);
-    pi->fd_cnt = 0;
-    pi->fd_capacity = 0;
-    pi->fds = NULL;
-  }
+  pi = gpr_malloc(sizeof(*pi));
+  gpr_mu_init(&pi->mu);
+  pi->fd_cnt = 0;
+  pi->fd_capacity = 0;
+  pi->fds = NULL;
+  pi->epoll_fd = -1;
+  pi->workqueue = NULL;
 
-  gpr_ref_init(&pi->ref_count, 0);
+  gpr_atm_rel_store(&pi->ref_count, 0);
   gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL);
 
   pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
 
   if (pi->epoll_fd < 0) {
-    gpr_asprintf(&err_msg, "epoll_create1 failed with error %d (%s)", errno,
-                 strerror(errno));
-    append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
-    gpr_free(err_msg);
-  } else {
-    polling_island_add_wakeup_fd_locked(pi, &grpc_global_wakeup_fd, error);
-    pi->next_free = NULL;
+    append_error(error, GRPC_OS_ERROR(errno, "epoll_create1"), err_desc);
+    goto done;
+  }
 
-    if (initial_fd != NULL) {
-      /* Lock the polling island here just in case we got this structure from
-         the freelist and the polling island lock was not released yet (by the
-         code that adds the polling island to the freelist) */
-      gpr_mu_lock(&pi->mu);
-      polling_island_add_fds_locked(pi, &initial_fd, 1, true, error);
-      gpr_mu_unlock(&pi->mu);
-    }
+  polling_island_add_wakeup_fd_locked(pi, &grpc_global_wakeup_fd, error);
+
+  if (initial_fd != NULL) {
+    polling_island_add_fds_locked(pi, &initial_fd, 1, true, error);
+  }
+
+  if (append_error(error, grpc_workqueue_create(exec_ctx, &pi->workqueue),
+                   err_desc) &&
+      *error == GRPC_ERROR_NONE) {
+    polling_island_add_fds_locked(pi, &pi->workqueue->wakeup_read_fd, 1, true,
+                                  error);
+    GPR_ASSERT(pi->workqueue->wakeup_read_fd->polling_island == NULL);
+    pi->workqueue->wakeup_read_fd->polling_island = pi;
+    PI_ADD_REF(pi, "fd");
   }
 
+done:
+  if (*error != GRPC_ERROR_NONE) {
+    if (pi->workqueue != NULL) {
+      GRPC_WORKQUEUE_UNREF(exec_ctx, pi->workqueue, "polling_island");
+    }
+    polling_island_delete(exec_ctx, pi);
+    pi = NULL;
+  }
   return pi;
 }
 
-static void polling_island_delete(polling_island *pi) {
+static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) {
   GPR_ASSERT(pi->fd_cnt == 0);
 
-  gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL);
-
-  close(pi->epoll_fd);
-  pi->epoll_fd = -1;
-
-  gpr_mu_lock(&g_pi_freelist_mu);
-  pi->next_free = g_pi_freelist;
-  g_pi_freelist = pi;
-  gpr_mu_unlock(&g_pi_freelist_mu);
+  if (pi->epoll_fd >= 0) {
+    close(pi->epoll_fd);
+  }
+  gpr_mu_destroy(&pi->mu);
+  gpr_free(pi->fds);
+  gpr_free(pi);
 }
 
 /* Attempts to gets the last polling island in the linked list (liked by the
@@ -704,9 +714,6 @@ static polling_island *polling_island_merge(polling_island *p,
 static grpc_error *polling_island_global_init() {
   grpc_error *error = GRPC_ERROR_NONE;
 
-  gpr_mu_init(&g_pi_freelist_mu);
-  g_pi_freelist = NULL;
-
   error = grpc_wakeup_fd_init(&polling_island_wakeup_fd);
   if (error == GRPC_ERROR_NONE) {
     error = grpc_wakeup_fd_wakeup(&polling_island_wakeup_fd);
@@ -716,18 +723,6 @@ static grpc_error *polling_island_global_init() {
 }
 
 static void polling_island_global_shutdown() {
-  polling_island *next;
-  gpr_mu_lock(&g_pi_freelist_mu);
-  gpr_mu_unlock(&g_pi_freelist_mu);
-  while (g_pi_freelist != NULL) {
-    next = g_pi_freelist->next_free;
-    gpr_mu_destroy(&g_pi_freelist->mu);
-    gpr_free(g_pi_freelist->fds);
-    gpr_free(g_pi_freelist);
-    g_pi_freelist = next;
-  }
-  gpr_mu_destroy(&g_pi_freelist_mu);
-
   grpc_wakeup_fd_destroy(&polling_island_wakeup_fd);
 }
 
@@ -845,7 +840,6 @@ static grpc_fd *fd_create(int fd, const char *name) {
   if (new_fd == NULL) {
     new_fd = gpr_malloc(sizeof(grpc_fd));
     gpr_mu_init(&new_fd->mu);
-    gpr_mu_init(&new_fd->pi_mu);
   }
 
   /* Note: It is not really needed to get the new_fd->mu lock here. If this is a
@@ -896,6 +890,7 @@ static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                       const char *reason) {
   bool is_fd_closed = false;
   grpc_error *error = GRPC_ERROR_NONE;
+  polling_island *unref_pi = NULL;
 
   gpr_mu_lock(&fd->mu);
   fd->on_done_closure = on_done;
@@ -923,21 +918,26 @@ static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
      - Unlock the latest polling island
      - Set fd->polling_island to NULL (but remove the ref on the polling island
        before doing this.) */
-  gpr_mu_lock(&fd->pi_mu);
   if (fd->polling_island != NULL) {
     polling_island *pi_latest = polling_island_lock(fd->polling_island);
     polling_island_remove_fd_locked(pi_latest, fd, is_fd_closed, &error);
     gpr_mu_unlock(&pi_latest->mu);
 
-    PI_UNREF(fd->polling_island, "fd_orphan");
+    unref_pi = fd->polling_island;
     fd->polling_island = NULL;
   }
-  gpr_mu_unlock(&fd->pi_mu);
 
   grpc_exec_ctx_sched(exec_ctx, fd->on_done_closure, error, NULL);
 
   gpr_mu_unlock(&fd->mu);
   UNREF_BY(fd, 2, reason); /* Drop the reference */
+  if (unref_pi != NULL) {
+    /* Unref stale polling island here, outside the fd lock above.
+       The polling island owns a workqueue which owns an fd, and unreffing
+       inside the lock can cause an eventual lock loop that makes TSAN very
+       unhappy. */
+    PI_UNREF(exec_ctx, unref_pi, "fd_orphan");
+  }
   GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error));
 }
 
@@ -1037,6 +1037,17 @@ static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
   gpr_mu_unlock(&fd->mu);
 }
 
+static grpc_workqueue *fd_get_workqueue(grpc_fd *fd) {
+  gpr_mu_lock(&fd->mu);
+  grpc_workqueue *workqueue = NULL;
+  if (fd->polling_island != NULL) {
+    workqueue =
+        GRPC_WORKQUEUE_REF(fd->polling_island->workqueue, "get_workqueue");
+  }
+  gpr_mu_unlock(&fd->mu);
+  return workqueue;
+}
+
 /*******************************************************************************
  * Pollset Definitions
  */
@@ -1227,9 +1238,10 @@ static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
   gpr_mu_unlock(&fd->mu);
 }
 
-static void pollset_release_polling_island(grpc_pollset *ps, char *reason) {
+static void pollset_release_polling_island(grpc_exec_ctx *exec_ctx,
+                                           grpc_pollset *ps, char *reason) {
   if (ps->polling_island != NULL) {
-    PI_UNREF(ps->polling_island, reason);
+    PI_UNREF(exec_ctx, ps->polling_island, reason);
   }
   ps->polling_island = NULL;
 }
@@ -1242,7 +1254,7 @@ static void finish_shutdown_locked(grpc_exec_ctx *exec_ctx,
   pollset->finish_shutdown_called = true;
 
   /* Release the ref and set pollset->polling_island to NULL */
-  pollset_release_polling_island(pollset, "ps_shutdown");
+  pollset_release_polling_island(exec_ctx, pollset, "ps_shutdown");
   grpc_exec_ctx_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE, NULL);
 }
 
@@ -1281,7 +1293,7 @@ static void pollset_reset(grpc_pollset *pollset) {
   pollset->finish_shutdown_called = false;
   pollset->kicked_without_pollers = false;
   pollset->shutdown_done = NULL;
-  pollset_release_polling_island(pollset, "ps_reset");
+  GPR_ASSERT(pollset->polling_island == NULL);
 }
 
 #define GRPC_EPOLL_MAX_EVENTS 1000
@@ -1309,7 +1321,7 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx,
      this function (i.e pollset_work_and_unlock()) is called */
 
   if (pollset->polling_island == NULL) {
-    pollset->polling_island = polling_island_create(NULL, error);
+    pollset->polling_island = polling_island_create(exec_ctx, NULL, error);
     if (pollset->polling_island == NULL) {
       GPR_TIMER_END("pollset_work_and_unlock", 0);
       return; /* Fatal error. We cannot continue */
@@ -1329,7 +1341,7 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx,
     /* Always do PI_ADD_REF before PI_UNREF because PI_UNREF may cause the
        polling island to be deleted */
     PI_ADD_REF(pi, "ps");
-    PI_UNREF(pollset->polling_island, "ps");
+    PI_UNREF(exec_ctx, pollset->polling_island, "ps");
     pollset->polling_island = pi;
   }
 
@@ -1400,7 +1412,7 @@ static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx,
      that we got before releasing the polling island lock). This is because
      pollset->polling_island pointer might get udpated in other parts of the
      code when there is an island merge while we are doing epoll_wait() above */
-  PI_UNREF(pi, "ps_work");
+  PI_UNREF(exec_ctx, pi, "ps_work");
 
   GPR_TIMER_END("pollset_work_and_unlock", 0);
 }
@@ -1517,10 +1529,11 @@ static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   grpc_error *error = GRPC_ERROR_NONE;
 
   gpr_mu_lock(&pollset->mu);
-  gpr_mu_lock(&fd->pi_mu);
+  gpr_mu_lock(&fd->mu);
 
   polling_island *pi_new = NULL;
 
+retry:
   /* 1) If fd->polling_island and pollset->polling_island are both non-NULL and
    *    equal, do nothing.
    * 2) If fd->polling_island and pollset->polling_island are both NULL, create
@@ -1535,15 +1548,44 @@ static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
    *    polling_island fields in both fd and pollset to point to the merged
    *    polling island.
    */
+
+  if (fd->orphaned) {
+    gpr_mu_unlock(&fd->mu);
+    gpr_mu_unlock(&pollset->mu);
+    /* early out */
+    return;
+  }
+
   if (fd->polling_island == pollset->polling_island) {
     pi_new = fd->polling_island;
     if (pi_new == NULL) {
-      pi_new = polling_island_create(fd, &error);
-
-      GRPC_POLLING_TRACE(
-          "pollset_add_fd: Created new polling island. pi_new: %p (fd: %d, "
-          "pollset: %p)",
-          (void *)pi_new, fd->fd, (void *)pollset);
+      /* Unlock before creating a new polling island: the polling island will
+         create a workqueue which creates a file descriptor, and holding an fd
+         lock here can eventually cause a loop to appear to TSAN (making it
+         unhappy). We don't think it's a real loop (there's an epoch point where
+         that loop possibility disappears), but the advantages of keeping TSAN
+         happy outweigh any performance advantage we might have by keeping the
+         lock held. */
+      gpr_mu_unlock(&fd->mu);
+      pi_new = polling_island_create(exec_ctx, fd, &error);
+      gpr_mu_lock(&fd->mu);
+      /* Need to reverify any assumptions made between the initial lock and
+         getting to this branch: if they've changed, we need to throw away our
+         work and figure things out again. */
+      if (fd->polling_island != NULL) {
+        GRPC_POLLING_TRACE(
+            "pollset_add_fd: Raced creating new polling island. pi_new: %p "
+            "(fd: %d, pollset: %p)",
+            (void *)pi_new, fd->fd, (void *)pollset);
+        PI_ADD_REF(pi_new, "dance_of_destruction");
+        PI_UNREF(exec_ctx, pi_new, "dance_of_destruction");
+        goto retry;
+      } else {
+        GRPC_POLLING_TRACE(
+            "pollset_add_fd: Created new polling island. pi_new: %p (fd: %d, "
+            "pollset: %p)",
+            (void *)pi_new, fd->fd, (void *)pollset);
+      }
     }
   } else if (fd->polling_island == NULL) {
     pi_new = polling_island_lock(pollset->polling_island);
@@ -1579,7 +1621,7 @@ static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   if (fd->polling_island != pi_new) {
     PI_ADD_REF(pi_new, "fd");
     if (fd->polling_island != NULL) {
-      PI_UNREF(fd->polling_island, "fd");
+      PI_UNREF(exec_ctx, fd->polling_island, "fd");
     }
     fd->polling_island = pi_new;
   }
@@ -1587,13 +1629,15 @@ static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   if (pollset->polling_island != pi_new) {
     PI_ADD_REF(pi_new, "ps");
     if (pollset->polling_island != NULL) {
-      PI_UNREF(pollset->polling_island, "ps");
+      PI_UNREF(exec_ctx, pollset->polling_island, "ps");
     }
     pollset->polling_island = pi_new;
   }
 
-  gpr_mu_unlock(&fd->pi_mu);
+  gpr_mu_unlock(&fd->mu);
   gpr_mu_unlock(&pollset->mu);
+
+  GRPC_LOG_IF_ERROR("pollset_add_fd", error);
 }
 
 /*******************************************************************************
@@ -1744,9 +1788,9 @@ static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx,
 void *grpc_fd_get_polling_island(grpc_fd *fd) {
   polling_island *pi;
 
-  gpr_mu_lock(&fd->pi_mu);
+  gpr_mu_lock(&fd->mu);
   pi = fd->polling_island;
-  gpr_mu_unlock(&fd->pi_mu);
+  gpr_mu_unlock(&fd->mu);
 
   return pi;
 }
@@ -1794,6 +1838,7 @@ static const grpc_event_engine_vtable vtable = {
     .fd_notify_on_read = fd_notify_on_read,
     .fd_notify_on_write = fd_notify_on_write,
     .fd_get_read_notifier_pollset = fd_get_read_notifier_pollset,
+    .fd_get_workqueue = fd_get_workqueue,
 
     .pollset_init = pollset_init,
     .pollset_shutdown = pollset_shutdown,