LU-13669 llite: make readahead aware of hints

Calling madvise(MADV_SEQUENTIAL) and madvise(MADV_RANDOM) sets the
VM_SEQ_READ and VM_RAND_READ hints in vma->vm_flags.  These should
be used to guide the Lustre readahead for better performance.

Disable the kernel readahead for mmap() pages and use the llite
readahead instead.  There was also a bug in ll_fault0() that would
set both VM_SEQ_READ and VM_RAND_READ at the same time, which was
confusing the detection of the VM_SEQ_READ case, since VM_RAND_READ
was being checked first.

This changes the readahead for mmap from submitting mostly 4KB RPCs
to a large number of 1MB RPCs for the application profiled:

  llite.*.read_ahead_stats     before        patched
  ------------------------     ------        -------
  hits                           2408         135924 samples [pages]
  misses                        34160           2384 samples [pages]

  osc.*.rpc_stats           read before    read patched
  ---------------          -------------  --------------
  pages per rpc            rpcs   % cum%   rpcs   % cum%
     1:                    6542  95  95     351  55  55
     2:                     224   3  99      76  12  67
     4:                      32   0  99      28   4  72
     8:                       2   0  99       9   1  73
    16:                      25   0  99      32   5  78
    32:                       0   0  99       8   1  80
    64:                       0   0  99       5   0  80
   128:                       0   0  99      15   2  83
   256:                       2   0  99     102  16  99
   512:                       0   0  99       0   0  99
  1024:                       1   0 100       3   0 100

Readahead hit rate improved from 6% to 98%, and 4KB RPCs dropped from
95% to 55% and 1MB+ RPCs increased from 0% to 16% (79% of all pages).

Add debug to ll_file_mmap(), ll_fault() and ll_fault_io_init() to
allow tracing VMA state functions for future IO optimizations.

Fixes: 62ef9c949753 ("add 2.6.27 kernel support")
Signed-off-by: Wang Shilong <wshilong@ddn.com>
Change-Id: I4bbb028db05b21ae01dafe6a7bea398e9b74d8a4
Reviewed-on: https://review.whamcloud.com/41228
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Wang Shilong <wshilong@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>

diff --git a/lustre/include/cl_object.h b/lustre/include/cl_object.h
index b1bcaaf..05c89eb 100644 (file)
--- a/lustre/include/cl_object.h
+++ b/lustre/include/cl_object.h
@@ -1964,7 +1964,15 @@ struct cl_io {
         * the read IO will check to-be-read OSCs' status, and make fast-switch
         * another mirror if some of the OSTs are not healthy.
         */
-                            ci_tried_all_mirrors:1;
+                            ci_tried_all_mirrors:1,
+       /**
+        * Random read hints, readahead will be disabled.
+        */
+                            ci_rand_read:1,
+       /**
+        * Sequential read hints.
+        */
+                            ci_seq_read:1;
        /**
         * Bypass quota check
         */

diff --git a/lustre/llite/file.c b/lustre/llite/file.c
index 2e78b2d..2e0ca9e 100644 (file)
--- a/lustre/llite/file.c
+++ b/lustre/llite/file.c
@@ -730,6 +730,8 @@ static int ll_local_open(struct file *file, struct lookup_intent *it,
        file->private_data = fd;
        ll_readahead_init(inode, &fd->fd_ras);
        fd->fd_omode = it->it_flags & (FMODE_READ | FMODE_WRITE | FMODE_EXEC);
+       /* turn off the kernel's read-ahead */
+       file->f_ra.ra_pages = 0;
 
        /* ll_cl_context initialize */
        rwlock_init(&fd->fd_lock);

diff --git a/lustre/llite/llite_mmap.c b/lustre/llite/llite_mmap.c
index 242f4f4..8ac6698 100644 (file)
--- a/lustre/llite/llite_mmap.c
+++ b/lustre/llite/llite_mmap.c
@@ -76,13 +76,11 @@ struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
  * \param env - corespondent lu_env to processing
  * \param vma - virtual memory area addressed to page fault
  * \param index - page index corespondent to fault.
- * \parm ra_flags - vma readahead flags.
  *
  * \return error codes from cl_io_init.
  */
 static struct cl_io *
-ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
-                pgoff_t index, unsigned long *ra_flags)
+ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma, pgoff_t index)
 {
        struct file            *file = vma->vm_file;
        struct inode           *inode = file_inode(file);
@@ -91,30 +89,27 @@ ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
        int                     rc;
        ENTRY;
 
-        if (ll_file_nolock(file))
-                RETURN(ERR_PTR(-EOPNOTSUPP));
+       if (ll_file_nolock(file))
+               RETURN(ERR_PTR(-EOPNOTSUPP));
 
 restart:
        io = vvp_env_thread_io(env);
-        io->ci_obj = ll_i2info(inode)->lli_clob;
-        LASSERT(io->ci_obj != NULL);
-
-        fio = &io->u.ci_fault;
-        fio->ft_index      = index;
-        fio->ft_executable = vma->vm_flags&VM_EXEC;
-
-        /*
-         * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
-         * the kernel will not read other pages not covered by ldlm in
-         * filemap_nopage. we do our readahead in ll_readpage.
-         */
-        if (ra_flags != NULL)
-                *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
-        vma->vm_flags &= ~VM_SEQ_READ;
-        vma->vm_flags |= VM_RAND_READ;
-
-        CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
-               fio->ft_index, fio->ft_executable);
+       io->ci_obj = ll_i2info(inode)->lli_clob;
+       LASSERT(io->ci_obj != NULL);
+
+       fio = &io->u.ci_fault;
+       fio->ft_index = index;
+       fio->ft_executable = vma->vm_flags & VM_EXEC;
+
+       CDEBUG(D_MMAP,
+              DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
+              PFID(&ll_i2info(inode)->lli_fid), vma, vma->vm_start,
+              vma->vm_end, vma->vm_flags, fio->ft_index);
+
+       if (vma->vm_flags & VM_SEQ_READ)
+               io->ci_seq_read = 1;
+       else if (vma->vm_flags & VM_RAND_READ)
+               io->ci_rand_read = 1;
 
        rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
        if (rc == 0) {
@@ -158,7 +153,7 @@ static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
        if (IS_ERR(env))
                RETURN(PTR_ERR(env));
 
-       io = ll_fault_io_init(env, vma, vmpage->index, NULL);
+       io = ll_fault_io_init(env, vma, vmpage->index);
        if (IS_ERR(io))
                GOTO(out, result = PTR_ERR(io));
 
@@ -268,7 +263,6 @@ static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
        struct cl_io            *io;
        struct vvp_io           *vio = NULL;
        struct page             *vmpage;
-       unsigned long            ra_flags;
        int                      result = 0;
        int                      fault_ret = 0;
        __u16                    refcheck;
@@ -302,7 +296,7 @@ static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
                fault_ret = 0;
        }
 
-       io = ll_fault_io_init(env, vma, vmf->pgoff, &ra_flags);
+       io = ll_fault_io_init(env, vma, vmf->pgoff);
        if (IS_ERR(io))
                GOTO(out, result = PTR_ERR(io));
 
@@ -335,8 +329,6 @@ static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
         }
        cl_io_fini(env, io);
 
-       vma->vm_flags |= ra_flags;
-
 out:
        cl_env_put(env, &refcheck);
        if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
@@ -365,6 +357,10 @@ static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
        if (cached)
                goto out;
 
+       CDEBUG(D_MMAP, DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
+              PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
+              vma, vma->vm_start, vma->vm_end, vma->vm_flags);
+
        /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
         * so that it can be killed by admin but not cause segfault by
         * other signals.
@@ -375,6 +371,7 @@ static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
        /* make sure offset is not a negative number */
        if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
                return VM_FAULT_SIGBUS;
+
 restart:
        result = ll_fault0(vma, vmf);
        if (vmf->page &&
@@ -547,6 +544,10 @@ int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
        int rc;
 
        ENTRY;
+       CDEBUG(D_VFSTRACE | D_MMAP,
+              "VFS_Op: fid="DFID" vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
+              PFID(&ll_i2info(inode)->lli_fid),
+              vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 
        if (ll_file_nolock(file))
                RETURN(-EOPNOTSUPP);

diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c
index daebffa..9202524 100644 (file)
--- a/lustre/llite/rw.c
+++ b/lustre/llite/rw.c
@@ -1249,7 +1249,7 @@ static bool index_in_stride_window(struct ll_readahead_state *ras,
  */
 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
                       struct ll_readahead_state *ras, pgoff_t index,
-                      enum ras_update_flags flags)
+                      enum ras_update_flags flags, struct cl_io *io)
 {
        struct ll_ra_info *ra = &sbi->ll_ra_info;
        bool hit = flags & LL_RAS_HIT;
@@ -1271,6 +1271,18 @@ static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
        if (ras->ras_no_miss_check)
                GOTO(out_unlock, 0);
 
+       if (io && io->ci_rand_read)
+               GOTO(out_unlock, 0);
+
+       if (io && io->ci_seq_read) {
+               if (!hit) {
+                       /* to avoid many small read RPC here */
+                       ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
+                       ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
+               }
+               goto skip;
+       }
+
        if (flags & LL_RAS_MMAP) {
                unsigned long ra_pages;
 
@@ -1589,7 +1601,7 @@ int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
                        flags |= LL_RAS_HIT;
                if (!vio->vui_ra_valid)
                        flags |= LL_RAS_MMAP;
-               ras_update(sbi, inode, ras, vvp_index(vpg), flags);
+               ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
        }
 
        cl_2queue_init(queue);
@@ -1608,7 +1620,7 @@ int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
        io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
        io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
                                io->u.ci_rw.crw_count - 1);
-       if (ll_readahead_enabled(sbi) && ras) {
+       if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
                pgoff_t skip_index = 0;
 
                if (ras->ras_next_readahead_idx < vvp_index(vpg))
@@ -1794,7 +1806,7 @@ int ll_readpage(struct file *file, struct page *vmpage)
                        /* For fast read, it updates read ahead state only
                         * if the page is hit in cache because non cache page
                         * case will be handled by slow read later. */
-                       ras_update(sbi, inode, ras, vvp_index(vpg), flags);
+                       ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
                        /* avoid duplicate ras_update() call */
                        vpg->vpg_ra_updated = 1;