我测试发现，如果cp一个文件，然后direct io读这个文件，会消耗很长时间。

     我猜测dio不能用page cache，而这个文件cp以后都在cache里面，要强制刷到磁盘，才能读？

     我cp这个文件很大，超过256M

    由于数据文件默认是用bufferedio方式打开的，也就是说它的数据是先缓冲在pagecache里面的，写入的数据会导致大量的脏页，而且这部分数据如果内核内存不紧张的话，是一直放在内存里面的的。我们知道directio是直接旁路掉pagecache直接发起设备IO的，也就是说在发起IO之前要保证数据是先落地到介质去，所以如果文件比较大的话，这个时间会比较长。从pagecahce的回写行为我们可以知道，只要脏页的数量不超过总内存的10%, 我们的机器有4G的内存，所以2个100M的文件总共才200M，不会导致writeback发生，我们可以很顺利的观察到这个现象。

    有了上面的分析，下面我们来重现下这个问题。以下是我的步骤：

$ uname -a
Linux rds064075.sqa.cm4 2.6.32-131.21.1.tb477.el6.x86_64 #1 SMP Thu Feb 23 14:24:55 CST 2012 x86_64 x86_64 x86_64 GNU/Linux
$ sudo sysctl vm.drop_caches=3
vm.drop_caches = 3
$ free -m && cat /proc/meminfo |grep -i dirty && time dd if=/dev/urandom of=test.dat count=6144 bs=16384 && free -m && cat /proc/meminfo |grep -i dirty && time dd if=test.dat of=/dev/null count=6144 bs=16384 && free -m && cat /proc/meminfo |grep -i dirty && time dd if=test.dat of=/dev/null count=6144 bs=16384  iflag=direct && free -m && cat /proc/meminfo |grep -i dirty
$ free -m && cat /proc/meminfo |grep -i dirty && time dd if=/dev/urandom of=test.dat count=6144 bs=16384 && free -m && cat /proc/meminfo |grep -i dirty && time dd if=test.dat of=/dev/null count=6144 bs=16384 && free -m && cat /proc/meminfo |grep -i dirty && time dd if=test.dat of=/dev/null count=6144 bs=16384  iflag=direct && free -m && cat /proc/meminfo |grep -i dirty
             total       used       free     shared    buffers     cached
Mem:         48262      22800      25461          0          3         42
-/+ buffers/cache:      22755      25507
Swap:         2047       2047          0
Dirty:               344 kB
6144+0 records in
6144+0 records out
100663296 bytes (101 MB) copied, 15.2308 s, 6.6 MB/s

real	0m15.249s
user	0m0.001s
sys	0m15.228s
             total       used       free     shared    buffers     cached
Mem:         48262      22912      25350          0          3        139
-/+ buffers/cache:      22768      25493
Swap:         2047       2047          0
Dirty:             98556 kB
6144+0 records in
6144+0 records out
100663296 bytes (101 MB) copied, 0.028041 s, 3.6 GB/s

real	0m0.029s
user	0m0.000s
sys	0m0.029s
             total       used       free     shared    buffers     cached
Mem:         48262      22912      25350          0          3        139
-/+ buffers/cache:      22768      25493
Swap:         2047       2047          0
Dirty:             98556 kB
6144+0 records in
6144+0 records out
100663296 bytes (101 MB) copied, 0.466601 s, 216 MB/s

real	0m0.468s
user	0m0.002s
sys	0m0.101s
             total       used       free     shared    buffers     cached
Mem:         48262      22906      25356          0          3        140
-/+ buffers/cache:      22762      25500
Swap:         2047       2047          0
Dirty:               896 kB

    从上面的实验，我们可以看出来我们的文件是101MB左右，脏页用了98544KB内存，在direct方式读后，文件占用的脏页被清洗掉了，脏页变成了80K, 但是这块数据还是留在了pagecache(140-39), 符合我们的预期。

    我们知道文件的读是从generic_file_aio_read发起的，我们从源码角度来分析下这个现象：

     在文卿和三百的帮助下，我们不费吃灰之力就找到了源码位置：

$ stap -L \'kernel.function("generic_file_aio_read")\' 
kernel.function("generic_file_aio_read@mm/filemap.c:1331") $iocb:struct kiocb* $iov:struct iovec const* $nr_segs:long unsigned int $pos:loff_t $count:size_t

    我们来看下读代码的实现：

     mm/filemap.c:1331

/**
 * generic_file_aio_read - generic filesystem read routine
 * @iocb:       kernel I/O control block
 * @iov:        io vector request
 * @nr_segs:    number of segments in the iovec
 * @pos:        current file position
 *
 * This is the "read()" routine for all filesystems
 * that can use the page cache directly.
 */
ssize_t
generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
                unsigned long nr_segs, loff_t pos)
{
        /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
        if (filp->f_flags & O_DIRECT) {
                loff_t size;
                struct address_space *mapping;
                struct inode *inode;

                mapping = filp->f_mapping;
                inode = mapping->host;
                if (!count)
                        goto out; /* skip atime */
                size = i_size_read(inode);
                if (pos < size) {
                        retval = filemap_write_and_wait_range(mapping, pos,
                                        pos + iov_length(iov, nr_segs) - 1);
                        if (!retval) {
                                retval = mapping->a_ops->direct_IO(READ,
iocb,
                                                        iov, pos, nr_segs);
                        }
                        if (retval > 0) {
                                *ppos = pos + retval;
                                count -= retval;
                        }

                        /*
                         * Btrfs can have a short DIO read if we encounter
                         * compressed extents, so if there was an error,
or if
                         * we\'ve already read everything we wanted to, or if
                         * there was a short read because we hit EOF, go
ahead
                         * and return.  Otherwise fallthrough to
buffered io for
                         * the rest of the read.
                         */
                        if (retval < 0 || !count || *ppos >= size) {
                                file_accessed(filp);
                                goto out;
                        }
                }
        }

    很清楚在directio方式下打开的文件，先要透过filemap_write_and_wait_range回写数据，才开始后面的IO读流程。

     最后一步骤，我们再用stap来确认下：

$ cat dwb.stp
global i;
probe kernel.function("filemap_write_and_wait_range") {
if (execname() != "dd") next;
print_backtrace();
println("===");
if (i++>2) exit();
}

$ sudo stap dwb.stp 
 0xffffffff8110e200 : filemap_write_and_wait_range+0x0/0x90 [kernel]
 0xffffffff8110f278 : generic_file_aio_read+0x498/0x870 [kernel]
 0xffffffff8117323a : do_sync_read+0xfa/0x140 [kernel]
 0xffffffff81173c65 : vfs_read+0xb5/0x1a0 [kernel]
 0xffffffff81173da1 : sys_read+0x51/0x90 [kernel]
 0xffffffff8100b172 : system_call_fastpath+0x16/0x1b [kernel]
===
 0xffffffff8110e200 : filemap_write_and_wait_range+0x0/0x90 [kernel]
 0xffffffff811acbc8 : __blockdev_direct_IO+0x228/0xc40 [kernel]
 0xffffffffa008a24a
===
 0xffffffff8110e200 : filemap_write_and_wait_range+0x0/0x90 [kernel]
 0xffffffff8110f278 : generic_file_aio_read+0x498/0x870 [kernel]
 0xffffffff8117323a : do_sync_read+0xfa/0x140 [kernel]
 0xffffffff81173c65 : vfs_read+0xb5/0x1a0 [kernel]
 0xffffffff81173da1 : sys_read+0x51/0x90 [kernel]
 0xffffffff8100b172 : system_call_fastpath+0x16/0x1b [kernel]
===
 0xffffffff8110e200 : filemap_write_and_wait_range+0x0/0x90 [kernel]
 0xffffffff811acbc8 : __blockdev_direct_IO+0x228/0xc40 [kernel]
 0xffffffffa008a24a
===

    filemap_write_and_wait_range的调用栈很清晰的暴露了一切！

    小结：文件系统比较复杂，最好不要混用bufferedio和directio！