Merge pull request #1209 from trapexit/o_direct

Align msg buffer memory to allow O_DIRECT to work

libfuse/lib/fuse.c

@@ -2337,7 +2337,7 @@ fuse_lib_read(fuse_req_t             req,
 
   f = req_fuse_prepare(req);
 
-  msgbuf = msgbuf_alloc();
+  msgbuf = msgbuf_alloc_page_aligned();
 
   res = f->fs->op.read(&ffi,msgbuf->mem,arg->size,arg->offset);
 

libfuse/lib/fuse_msgbuf.cpp

@@ -17,6 +17,7 @@
 
 #include "fuse_msgbuf.hpp"
 #include "fuse.h"
+#include "fuse_kernel.h"
 
 #include <unistd.h>
 
@@ -35,34 +36,52 @@ static std::atomic<std::uint_fast64_t> g_MSGBUF_ALLOC_COUNT;
 static std::mutex g_MUTEX;
 static std::vector<fuse_msgbuf_t*> g_MSGBUF_STACK;
 
-static
-__attribute__((constructor))
+uint64_t
+msgbuf_get_bufsize()
+{
+  return g_BUFSIZE;
+}
+
 void
-msgbuf_constructor()
+msgbuf_set_bufsize(const uint32_t size_in_pages_)
 {
-  g_PAGESIZE = sysconf(_SC_PAGESIZE);
-  // +2 because to do O_DIRECT we need to offset the buffer to align
-  g_BUFSIZE  = (g_PAGESIZE * (FUSE_MAX_MAX_PAGES + 2));
+  g_BUFSIZE = ((size_in_pages_ + 1) * g_PAGESIZE);
 }
 
-static
-__attribute__((destructor))
 void
-msgbuf_destroy()
+msgbuf_page_align(fuse_msgbuf_t *msgbuf_)
 {
+  msgbuf_->mem   = (char*)msgbuf_;
+  msgbuf_->mem  += g_PAGESIZE;
+  msgbuf_->size  = (g_BUFSIZE - g_PAGESIZE);
+}
 
+void
+msgbuf_write_align(fuse_msgbuf_t *msgbuf_)
+{
+  msgbuf_->mem   = (char*)msgbuf_;
+  msgbuf_->mem  += g_PAGESIZE;
+  msgbuf_->mem  -= sizeof(struct fuse_in_header);
+  msgbuf_->mem  -= sizeof(struct fuse_write_in);
+  msgbuf_->size  = (g_BUFSIZE - g_PAGESIZE);
 }
 
-uint64_t
-msgbuf_get_bufsize()
+static
+__attribute__((constructor))
+void
+msgbuf_constructor()
 {
-  return g_BUFSIZE;
+  g_PAGESIZE = sysconf(_SC_PAGESIZE);
+  // FUSE_MAX_MAX_PAGES for payload + 1 for message header
+  msgbuf_set_bufsize(FUSE_MAX_MAX_PAGES + 1);
 }
 
+static
+__attribute__((destructor))
 void
-msgbuf_set_bufsize(const uint32_t size_in_pages_)
+msgbuf_destroy()
 {
-  g_BUFSIZE = ((size_in_pages_ + 1) * g_PAGESIZE);
+
 }
 
 static
@@ -79,8 +98,11 @@ page_aligned_malloc(const uint64_t size_)
   return buf;
 }
 
+typedef void (*msgbuf_setup_func_t)(fuse_msgbuf_t*);
+
+static
 fuse_msgbuf_t*
-msgbuf_alloc()
+_msgbuf_alloc(msgbuf_setup_func_t setup_func_)
 {
   fuse_msgbuf_t *msgbuf;
 
@@ -93,9 +115,6 @@ msgbuf_alloc()
       if(msgbuf == NULL)
         return NULL;
 
-      msgbuf->mem = (((char*)msgbuf) + g_PAGESIZE);
-
-      msgbuf->size = g_BUFSIZE - g_PAGESIZE;
       g_MSGBUF_ALLOC_COUNT++;
     }
   else
@@ -105,9 +124,26 @@ msgbuf_alloc()
       g_MUTEX.unlock();
     }
 
+  setup_func_(msgbuf);
+
   return msgbuf;
 }
 
+// Offset the memory so write request payload will be placed on page
+// boundry so O_DIRECT can work. No impact on other message types
+// except for `read` which will require using `msgbuf_page_align`.
+fuse_msgbuf_t*
+msgbuf_alloc()
+{
+  return _msgbuf_alloc(msgbuf_write_align);
+}
+
+fuse_msgbuf_t*
+msgbuf_alloc_page_aligned()
+{
+  return _msgbuf_alloc(msgbuf_page_align);
+}
+
 static
 void
 msgbuf_destroy(fuse_msgbuf_t *msgbuf_)

libfuse/lib/fuse_msgbuf.hpp

@@ -26,7 +26,8 @@ EXTERN_C_BEGIN
 void     msgbuf_set_bufsize(const uint32_t size);
 uint64_t msgbuf_get_bufsize();
 
-fuse_msgbuf_t* msgbuf_alloc();
+fuse_msgbuf_t *msgbuf_alloc();
+fuse_msgbuf_t *msgbuf_alloc_page_aligned();
 void           msgbuf_free(fuse_msgbuf_t *msgbuf);
 
 void           msgbuf_gc();
@@ -35,4 +36,7 @@ void           msgbuf_gc_10percent();
 uint64_t       msgbuf_alloc_count();
 uint64_t       msgbuf_avail_count();
 
+void           msgbuf_page_align(fuse_msgbuf_t *msgbuf);
+void           msgbuf_write_align(fuse_msgbuf_t *msgbuf);
+
 EXTERN_C_END

tests/TEST_o_direct

@@ -0,0 +1,25 @@
+#!/usr/bin/env python3
+
+import os
+import sys
+import tempfile
+import mmap
+import resource
+
+(fd,filepath) = tempfile.mkstemp(dir=sys.argv[1])
+
+os.close(fd)
+
+fd = os.open(filepath,os.O_RDWR|os.O_DIRECT|os.O_TRUNC)
+
+os.unlink(filepath)
+
+data = mmap.mmap(-1, resource.getpagesize())
+
+os.write(fd,data)
+
+os.lseek(fd,0,os.SEEK_SET)
+
+data = os.read(fd,resource.getpagesize())
+
+os.close(fd)