Manage cloned /dev/fuse fds in sessions

- Store cloned read-worker fds on fuse_session and close them there.
- Set up clone fallback through session helpers used by the loop.
- Carry per-request fds through receive, process, and reply paths.

Tests show significant performance increases in highly concurrent
workloads on a high core count system. Upto 15% better on reads and
27% on writes.

vendored/libfuse/include/fuse_req.hpp

@@ -10,6 +10,7 @@ struct fuse_req_t
 {
   fuse_req_ctx_t ctx;
   struct fuse_session *se;
+  int fd;
   fuse_conn_info_t conn;
   unsigned int ioctl_64bit : 1;
 };

vendored/libfuse/lib/fuse_i.hpp

@@ -15,6 +15,7 @@
 #include "extern_c.h"
 
 #include <atomic>
+#include <vector>
 
 /* Simplified fuse_session - fields inlined from former fuse_chan
  * Since mergerfs only supports one mount, we collapse the hierarchy:
@@ -24,9 +25,11 @@
 struct fuse_session
 {
   int (*receive_buf)(struct fuse_session *se,
+                     int                  fd,
                      fuse_msgbuf_t       *msgbuf);
 
   void (*process_buf)(struct fuse_session *se,
+                      int                  fd,
                       const fuse_msgbuf_t *msgbuf);
 
   void (*destroy)(void *data);
@@ -35,8 +38,10 @@ struct fuse_session
   std::atomic<int> exited;
 
   /* Formerly in fuse_chan - inlined for single-mount simplicity */
-  int fd;            /* /dev/fuse file descriptor */
-  size_t bufsize;    /* Buffer size for I/O operations */
+  int fd;          /* /dev/fuse file descriptor */
+  size_t bufsize;  /* Buffer size for I/O operations */
+
+  std::vector<int> clone_fds; /* cloned /dev/fuse fds for read workers */
 };
 
 struct fuse_notify_req
@@ -50,6 +55,12 @@ struct fuse_notify_req
   struct fuse_notify_req *prev;
 };
 
+bool fuse_session_setup_read_fds(struct fuse_session *se,
+                                 int                  read_thread_count);
+
+int fuse_session_read_fd(struct fuse_session *se,
+                         int                  index);
+
 
 
 EXTERN_C_BEGIN

vendored/libfuse/lib/fuse_loop.cpp

@@ -11,7 +11,6 @@
 #include "maintenance_thread.hpp"
 
 #include "fuse_i.hpp"
-#include "fuse_kernel.h"
 #include "fuse_lowlevel.h"
 
 #include "fuse_cfg.hpp"
@@ -67,13 +66,16 @@ _nproc()
 struct AsyncWorker
 {
   fuse_session *_se;
+  int _fd;
   sem_t *_finished;
   std::shared_ptr<ThreadPool> _process_tp;
 
   AsyncWorker(fuse_session                *se_,
+              int                          fd_,
               sem_t                       *finished_,
               std::shared_ptr<ThreadPool>  process_tp_)
     : _se(se_),
+      _fd(fd_),
       _finished(finished_),
       _process_tp(process_tp_)
   {
@@ -102,7 +104,7 @@ struct AsyncWorker
         while(true)
           {
             pthread_setcancelstate(PTHREAD_CANCEL_ENABLE,NULL);
-            rv = _se->receive_buf(_se,msgbuf);
+            rv = _se->receive_buf(_se,_fd,msgbuf);
             pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,NULL);
 
             if(rv > 0)
@@ -118,9 +120,9 @@ struct AsyncWorker
           }
 
         _process_tp->enqueue_work(ptok,
-                                  [=]()
+                                  [se = _se,fd = _fd,msgbuf]()
                                   {
-                                    _se->process_buf(_se,msgbuf);
+                                    se->process_buf(se,fd,msgbuf);
                                     msgbuf_free(msgbuf);
                                   });
       }
@@ -130,12 +132,15 @@ struct AsyncWorker
 struct SyncWorker
 {
   fuse_session *_se;
+  int _fd;
   sem_t *_finished;
 
   SyncWorker(fuse_session *se_,
+             int           fd_,
              sem_t        *finished_)
 
     : _se(se_),
+      _fd(fd_),
       _finished(finished_)
   {
   }
@@ -162,7 +167,7 @@ struct SyncWorker
         while(true)
           {
             pthread_setcancelstate(PTHREAD_CANCEL_ENABLE,NULL);
-            rv = _se->receive_buf(_se,msgbuf);
+            rv = _se->receive_buf(_se,_fd,msgbuf);
             pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,NULL);
             if(rv > 0)
               break;
@@ -176,7 +181,7 @@ struct SyncWorker
             return ::_print_error(rv);
           }
 
-        _se->process_buf(_se,msgbuf);
+        _se->process_buf(_se,_fd,msgbuf);
 
         msgbuf_free(msgbuf);
       }
@@ -255,15 +260,18 @@ fuse_session_loop_mt(struct fuse_session *se_,
                      const std::string    pin_threads_type_)
 {
   sem_t finished;
+  bool cloned_read_fds;
   int read_thread_count;
   int process_thread_count;
   int process_thread_queue_depth;
   std::vector<pthread_t> read_threads;
   std::vector<pthread_t> process_threads;
-  std::unique_ptr<ThreadPool> read_tp;
-  std::shared_ptr<ThreadPool> process_tp;
 
   sem_init(&finished,0,0);
+  DEFER{ sem_destroy(&finished); };
+
+  std::unique_ptr<ThreadPool> read_tp;
+  std::shared_ptr<ThreadPool> process_tp;
 
   read_thread_count          = raw_read_thread_count_;
   process_thread_count       = raw_process_thread_count_;
@@ -274,21 +282,28 @@ fuse_session_loop_mt(struct fuse_session *se_,
 
   if(process_thread_count > 0)
     process_tp = std::make_shared<ThreadPool>(process_thread_count,
-                                              process_thread_queue_depth,
-                                              "fuse.process");
+                                               process_thread_queue_depth,
+                                               "fuse.process");
+
+  cloned_read_fds = fuse_session_setup_read_fds(se_,read_thread_count);
 
   read_tp = std::make_unique<ThreadPool>(read_thread_count,
-                                         read_thread_count,
-                                         "fuse.read");
+                                          read_thread_count,
+                                          "fuse.read");
   if(process_tp)
     {
       for(auto i = 0; i < read_thread_count; i++)
-        read_tp->enqueue_work(AsyncWorker(se_,&finished,process_tp));
+        read_tp->enqueue_work(AsyncWorker(se_,
+                                          fuse_session_read_fd(se_,i),
+                                          &finished,
+                                          process_tp));
     }
   else
     {
       for(auto i = 0; i < read_thread_count; i++)
-        read_tp->enqueue_work(SyncWorker(se_,&finished));
+        read_tp->enqueue_work(SyncWorker(se_,
+                                         fuse_session_read_fd(se_,i),
+                                         &finished));
     }
 
   if(read_tp)
@@ -301,10 +316,12 @@ fuse_session_loop_mt(struct fuse_session *se_,
   SysLog::info("read-thread-count={}; "
                "process-thread-count={}; "
                "process-thread-queue-depth={}; "
+               "fuse-dev-ioc-clone={}; "
                "pin-threads={};",
                read_thread_count,
                process_thread_count,
                process_thread_queue_depth,
+               cloned_read_fds,
                pin_threads_type_);
 
   while(!fuse_session_exited(se_))
@@ -313,7 +330,8 @@ fuse_session_loop_mt(struct fuse_session *se_,
   for(auto t : read_threads)
     pthread_cancel(t);
 
-  sem_destroy(&finished);
+  read_tp.reset();
+  process_tp.reset();
 
   return 0;
 }

vendored/libfuse/lib/fuse_lowlevel.cpp

@@ -103,16 +103,16 @@ iov_length(const struct iovec *iov,
 
 static
 int
-fuse_send_msg(struct fuse_session *se,
-              struct iovec        *iov,
-              int                  count)
+fuse_send_msg(const int     fd_,
+              struct iovec *iov,
+              int           count)
 {
   int rv;
   struct fuse_out_header *out = (fuse_out_header*)iov[0].iov_base;
 
   out->len = iov_length(iov, count);
 
-  rv = writev(se->fd,iov,count);
+  rv = writev(fd_,iov,count);
   if(rv == -1)
     return -errno;
 
@@ -144,7 +144,7 @@ fuse_send_reply_iov_nofree(fuse_req_t   *req,
   iov[0].iov_base = &out;
   iov[0].iov_len  = sizeof(struct fuse_out_header);
 
-  return fuse_send_msg(req->se, iov, count);
+  return fuse_send_msg(req->fd, iov, count);
 }
 
 static
@@ -440,7 +440,7 @@ fuse_reply_data(fuse_req_t   *req,
   if(fuse_cfg.debug)
     fuse_debug_data_out(req->ctx.unique,bufsize_);
 
-  res = fuse_send_msg(req->se,iov,2);
+  res = fuse_send_msg(req->fd,iov,2);
   fuse_req_free(req);
 
   return res;
@@ -1310,7 +1310,7 @@ send_notify_iov(struct fuse_session *se,
   iov[0].iov_base = &out;
   iov[0].iov_len = sizeof(struct fuse_out_header);
 
-  return fuse_send_msg(se, iov, count);
+  return fuse_send_msg(se->fd, iov, count);
 }
 
 int
@@ -1558,7 +1558,7 @@ fuse_ll_destroy(void *data)
 
 static
 void
-fuse_send_errno(struct fuse_session *se_,
+fuse_send_errno(const int            fd_,
                 const int            errno_,
                 const uint64_t       unique_id_)
 {
@@ -1574,25 +1574,26 @@ fuse_send_errno(struct fuse_session *se_,
   if(fuse_cfg.debug)
     fuse_debug_out_header(&out);
 
-  fuse_send_msg(se_,&iov,1);
+  fuse_send_msg(fd_,&iov,1);
 }
 
 static
 void
-fuse_send_enomem(struct fuse_session *se_,
+fuse_send_enomem(const int            fd_,
                  const uint64_t       unique_id_)
 {
-  fuse_send_errno(se_,ENOMEM,unique_id_);
+  fuse_send_errno(fd_,ENOMEM,unique_id_);
 }
 
 static
 int
 fuse_ll_buf_receive_read(struct fuse_session *se_,
+                         int                  fd_,
                          fuse_msgbuf_t       *msgbuf_)
 {
   int rv;
 
-  rv = read(se_->fd,msgbuf_->mem,msgbuf_->size);
+  rv = read(fd_,msgbuf_->mem,msgbuf_->size);
   if(rv == -1)
     return -errno;
 
@@ -1608,6 +1609,7 @@ fuse_ll_buf_receive_read(struct fuse_session *se_,
 static
 void
 fuse_ll_buf_process_read(struct fuse_session *se_,
+                         int                  fd_,
                          const fuse_msgbuf_t *msgbuf_)
 {
   int err;
@@ -1621,7 +1623,7 @@ fuse_ll_buf_process_read(struct fuse_session *se_,
 
   req = fuse_req_alloc();
   if(req == NULL)
-    return fuse_send_enomem(se_,in->unique);
+    return fuse_send_enomem(fd_,in->unique);
 
   req->ctx.len    = in->len;
   req->ctx.opcode = in->opcode;
@@ -1633,6 +1635,7 @@ fuse_ll_buf_process_read(struct fuse_session *se_,
   req->ctx.umask  = 0;
   req->conn       = f.conn;
   req->se         = se_;
+  req->fd         = fd_;
   req->ioctl_64bit = 0;
 
   err = ENOSYS;
@@ -1653,6 +1656,7 @@ fuse_ll_buf_process_read(struct fuse_session *se_,
 static
 void
 fuse_ll_buf_process_read_init(struct fuse_session *se_,
+                              int                  fd_,
                               const fuse_msgbuf_t *msgbuf_)
 {
   int err;
@@ -1666,7 +1670,7 @@ fuse_ll_buf_process_read_init(struct fuse_session *se_,
 
   req = fuse_req_alloc();
   if(req == NULL)
-    return fuse_send_enomem(se_,in->unique);
+    return fuse_send_enomem(fd_,in->unique);
 
   req->ctx.len    = in->len;
   req->ctx.opcode = in->opcode;
@@ -1678,6 +1682,7 @@ fuse_ll_buf_process_read_init(struct fuse_session *se_,
   req->ctx.pid    = in->pid;
   req->ctx.umask  = 0;
   req->se         = se_;
+  req->fd         = fd_;
 
   err = EIO;
   if(in->opcode != FUSE_INIT)

vendored/libfuse/lib/fuse_session.cpp

@@ -8,16 +8,24 @@
 
 #include "fuse_i.hpp"
 #include "fuse_kernel.h"
+#include "syslog.hpp"
 
 #include <cassert>
 #include <cerrno>
+#include <cstdint>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <fcntl.h>
 #include <new>
+#include <sys/ioctl.h>
 #include <unistd.h>
 #include <sys/uio.h>
 
+#ifndef O_CLOEXEC
+#define O_CLOEXEC 0
+#endif
+
 /*
  * Simplified session management for single-mount filesystems.
  * The fuse_chan structure has been removed and its fields inlined
@@ -25,6 +33,42 @@
  * one mount point per process.
  */
 
+static
+int
+_clone_fuse_fd(const int fd_)
+{
+  int rv;
+  int clone_fd;
+  uint32_t source_fd;
+
+  clone_fd = ::open("/dev/fuse",O_RDWR|O_CLOEXEC);
+  if(clone_fd == -1)
+    return -errno;
+
+  source_fd = fd_;
+  rv = ::ioctl(clone_fd,FUSE_DEV_IOC_CLONE,&source_fd);
+  if(rv == -1)
+    {
+      int err = errno;
+      ::close(clone_fd);
+      return -err;
+    }
+
+  return clone_fd;
+}
+
+static
+void
+_close_clone_fds(struct fuse_session *se_)
+{
+  for(auto fd : se_->clone_fds)
+    {
+      if(fd != -1)
+        ::close(fd);
+    }
+  se_->clone_fds.clear();
+}
+
 struct fuse_session*
 fuse_session_new(void *data,
                  void *receive_buf,
@@ -40,8 +84,8 @@ fuse_session_new(void *data,
     }
 
   se->f           = static_cast<struct fuse_ll*>(data);
-  se->receive_buf = reinterpret_cast<int(*)(struct fuse_session*, fuse_msgbuf_t*)>(receive_buf);
-  se->process_buf = reinterpret_cast<void(*)(struct fuse_session*, const fuse_msgbuf_t*)>(process_buf);
+  se->receive_buf = reinterpret_cast<int(*)(struct fuse_session*, int, fuse_msgbuf_t*)>(receive_buf);
+  se->process_buf = reinterpret_cast<void(*)(struct fuse_session*, int, const fuse_msgbuf_t*)>(process_buf);
   se->destroy     = reinterpret_cast<void(*)(void*)>(destroy);
   se->fd          = -1;
 
@@ -57,6 +101,8 @@ fuse_session_destroy(struct fuse_session *se)
   if(se->destroy)
     se->destroy(se->f);
 
+  ::_close_clone_fds(se);
+
   if(se->fd != -1)
     {
       close(se->fd);
@@ -93,6 +139,8 @@ fuse_session_data(struct fuse_session *se)
 int
 fuse_session_clearfd(struct fuse_session *se)
 {
+  ::_close_clone_fds(se);
+
   int fd = se->fd;
   se->fd = -1;
   return fd;
@@ -102,9 +150,55 @@ void
 fuse_session_setfd(struct fuse_session *se,
                    int                   fd)
 {
+  ::_close_clone_fds(se);
+
   se->fd = fd;
 }
 
+bool
+fuse_session_setup_read_fds(struct fuse_session *se_,
+                            const int            read_thread_count_)
+{
+  ::_close_clone_fds(se_);
+
+  if(read_thread_count_ <= 1)
+    return false;
+
+  se_->clone_fds.reserve(read_thread_count_ - 1);
+
+  for(auto i = 1; i < read_thread_count_; i++)
+    {
+      int fd;
+
+      fd = ::_clone_fuse_fd(se_->fd);
+      if(fd < 0)
+        {
+          int err = -fd;
+
+          ::_close_clone_fds(se_);
+          SysLog::warning("failed to clone /dev/fuse fd for read threads; "
+                          "using shared fd: {} ({})",
+                          strerror(err),
+                          err);
+          return false;
+        }
+
+      se_->clone_fds.push_back(fd);
+    }
+
+  return true;
+}
+
+int
+fuse_session_read_fd(struct fuse_session *se_,
+                     const int            index_)
+{
+  if((index_ > 0) && (index_ <= (int)se_->clone_fds.size()))
+    return se_->clone_fds[index_ - 1];
+
+  return se_->fd;
+}
+
 void
 fuse_session_setbufsize(struct fuse_session *se,
                         size_t               bufsize)