LU-9749 llite: Reduce overhead for ll_do_fast_read

In ll_do_fast_read, looking up a cl_env adds some overhead,
and can also cause spinlock contention on older kernels.

Fast read can safely use the preallocated percpu cl_env, so
do that to reduce overhead.

SLES numbers on recent Xeon, CentOS numbers on VMs on
older hardware.  SLES has queued spinlocks and scales
perfectly with multiple threads, with or without this
patch.  CentOS scales poorly at small I/O sizes without
this patch.

SLES is SLES12SP2, CentOS is CentOS 7.3.

SLES:
1 thread
         8b   1K   1M
Without: 23   2200 6800
With:    27.5 2500 7200

4 threads
         8b   1K    1M
Without: 90   8700  27000
With:    108  10000 28000

Earlier kernel (CentOS 7.3):
1 thread
         8b  1K   1M
Without: 9   1000 5100
with:    12  1300 5800

4 threads
         8b  1K   1M
Without: 22  2400 17000
With:    48  4900 20000

Signed-off-by: Patrick Farrell <paf@cray.com>
Change-Id: Ice5d653ace5ce76bc8911501a9b15c11b7a3234a
Reviewed-on: https://review.whamcloud.com/27970
Tested-by: Jenkins
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Reviewed-by: Jinshan Xiong <jinshan.xiong@intel.com>
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Dmitry Eremin <dmitry.eremin@intel.com>

diff --git a/lustre/llite/file.c b/lustre/llite/file.c
index b14fcfe..3a66cab 100644 (file)
--- a/lustre/llite/file.c
+++ b/lustre/llite/file.c
@@ -1385,8 +1385,7 @@ out:
  * \retval - number of bytes have been read, or error code if error occurred.
  */
 static ssize_t
-ll_do_fast_read(const struct lu_env *env, struct kiocb *iocb,
-               struct iov_iter *iter)
+ll_do_fast_read(struct kiocb *iocb, struct iov_iter *iter)
 {
        ssize_t result;
 
@@ -1398,9 +1397,7 @@ ll_do_fast_read(const struct lu_env *env, struct kiocb *iocb,
        if (iocb->ki_filp->f_flags & O_DIRECT)
                return 0;
 
-       ll_cl_add(iocb->ki_filp, env, NULL, LCC_RW);
        result = generic_file_read_iter(iocb, iter);
-       ll_cl_remove(iocb->ki_filp, env);
 
        /* If the first page is not in cache, generic_file_aio_read() will be
         * returned with -ENODATA.
@@ -1426,14 +1423,14 @@ static ssize_t ll_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
        ssize_t rc2;
        __u16 refcheck;
 
+       result = ll_do_fast_read(iocb, to);
+       if (result < 0 || iov_iter_count(to) == 0)
+               GOTO(out, result);
+
        env = cl_env_get(&refcheck);
        if (IS_ERR(env))
                return PTR_ERR(env);
 
-       result = ll_do_fast_read(env, iocb, to);
-       if (result < 0 || iov_iter_count(to) == 0)
-               GOTO(out, result);
-
        args = ll_env_args(env, IO_NORMAL);
        args->u.normal.via_iter = to;
        args->u.normal.via_iocb = iocb;
@@ -1445,8 +1442,8 @@ static ssize_t ll_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
        else if (result == 0)
                result = rc2;
 
-out:
        cl_env_put(env, &refcheck);
+out:
        return result;
 }
 
@@ -1532,30 +1529,22 @@ static ssize_t ll_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 static ssize_t ll_file_read(struct file *file, char __user *buf, size_t count,
                            loff_t *ppos)
 {
-       struct lu_env *env;
        struct iovec   iov = { .iov_base = buf, .iov_len = count };
-       struct kiocb  *kiocb;
+       struct kiocb   kiocb;
        ssize_t        result;
-       __u16          refcheck;
        ENTRY;
 
-       env = cl_env_get(&refcheck);
-       if (IS_ERR(env))
-               RETURN(PTR_ERR(env));
-
-       kiocb = &ll_env_info(env)->lti_kiocb;
-        init_sync_kiocb(kiocb, file);
-        kiocb->ki_pos = *ppos;
+       init_sync_kiocb(&kiocb, file);
+       kiocb.ki_pos = *ppos;
 #ifdef HAVE_KIOCB_KI_LEFT
-       kiocb->ki_left = count;
+       kiocb.ki_left = count;
 #elif defined(HAVE_KI_NBYTES)
-       kiocb->ki_nbytes = count;
+       kiocb.i_nbytes = count;
 #endif
 
-       result = ll_file_aio_read(kiocb, &iov, 1, kiocb->ki_pos);
-       *ppos = kiocb->ki_pos;
+       result = ll_file_aio_read(&kiocb, &iov, 1, kiocb.ki_pos);
+       *ppos = kiocb.ki_pos;
 
-       cl_env_put(env, &refcheck);
        RETURN(result);
 }
 

diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c
index c8c63a5..e359b60 100644 (file)
--- a/lustre/llite/rw.c
+++ b/lustre/llite/rw.c
@@ -1142,24 +1142,23 @@ int ll_readpage(struct file *file, struct page *vmpage)
        struct inode *inode = file_inode(file);
        struct cl_object *clob = ll_i2info(inode)->lli_clob;
        struct ll_cl_context *lcc;
-       const struct lu_env  *env;
-       struct cl_io   *io;
+       const struct lu_env  *env = NULL;
+       struct cl_io   *io = NULL;
        struct cl_page *page;
        int result;
        ENTRY;
 
        lcc = ll_cl_find(file);
-       if (lcc == NULL) {
-               unlock_page(vmpage);
-               RETURN(-EIO);
+       if (lcc != NULL) {
+               env = lcc->lcc_env;
+               io  = lcc->lcc_io;
        }
 
-       env = lcc->lcc_env;
-       io  = lcc->lcc_io;
        if (io == NULL) { /* fast read */
                struct inode *inode = file_inode(file);
                struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
                struct ll_readahead_state *ras = &fd->fd_ras;
+               struct lu_env  *local_env = NULL;
                struct vvp_page *vpg;
 
                result = -ENODATA;
@@ -1172,11 +1171,16 @@ int ll_readpage(struct file *file, struct page *vmpage)
                        RETURN(result);
                }
 
+               if (!env) {
+                       local_env = cl_env_percpu_get();
+                       env = local_env;
+               }
+
                vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
                if (vpg->vpg_defer_uptodate) {
                        enum ras_update_flags flags = LL_RAS_HIT;
 
-                       if (lcc->lcc_type == LCC_MMAP)
+                       if (lcc && lcc->lcc_type == LCC_MMAP)
                                flags |= LL_RAS_MMAP;
 
                        /* For fast read, it updates read ahead state only
@@ -1199,8 +1203,14 @@ int ll_readpage(struct file *file, struct page *vmpage)
                        }
                }
 
-               unlock_page(vmpage);
+               /* release page refcount before unlocking the page to ensure
+                * the object won't be destroyed in the calling path of
+                * cl_page_put(). Please see comment in ll_releasepage(). */
                cl_page_put(env, page);
+               unlock_page(vmpage);
+               if (local_env)
+                       cl_env_percpu_put(local_env);
+
                RETURN(result);
        }