Whamcloud - gitweb
LU-6245 libcfs: replace IS_PO2 with is_power_of_2 in server code
[fs/lustre-release.git] / lustre / osd-zfs / osd_io.c
index e315fc1..1d3b230 100644 (file)
  *
  * You should have received a copy of the GNU General Public License
  * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
  *
  * GPL HEADER END
  */
 /*
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
- */
-/*
- * Copyright (c) 2012, 2013, Intel Corporation.
- * Use is subject to license terms.
+ *
+ * Copyright (c) 2012, 2016, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
  * Author: Mike Pershin <tappro@whamcloud.com>
  */
 
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
 #define DEBUG_SUBSYSTEM S_OSD
 
 #include <lustre_ver.h>
 #include <libcfs/libcfs.h>
-#include <lustre_fsfilt.h>
 #include <obd_support.h>
 #include <lustre_net.h>
 #include <obd.h>
 #include <obd_class.h>
 #include <lustre_disk.h>
 #include <lustre_fid.h>
+#include <lustre/lustre_idl.h> /* LLOG_MIN_CHUNK_SIZE definition */
 
 #include "osd_internal.h"
 
 #include <sys/sa_impl.h>
 #include <sys/txg.h>
 
-static char *osd_zerocopy_tag = "zerocopy";
+static char *osd_0copy_tag = "zerocopy";
+
+
+static void record_start_io(struct osd_device *osd, int rw, int discont_pages)
+{
+       struct obd_histogram *h = osd->od_brw_stats.hist;
+
+       if (rw == READ) {
+               atomic_inc(&osd->od_r_in_flight);
+               lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
+                                atomic_read(&osd->od_r_in_flight));
+               lprocfs_oh_tally(&h[BRW_R_DISCONT_PAGES], discont_pages);
+
+       } else {
+               atomic_inc(&osd->od_w_in_flight);
+               lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
+                                atomic_read(&osd->od_w_in_flight));
+               lprocfs_oh_tally(&h[BRW_W_DISCONT_PAGES], discont_pages);
+
+       }
+}
+
+static void record_end_io(struct osd_device *osd, int rw,
+                         unsigned long elapsed, int disksize, int npages)
+{
+       struct obd_histogram *h = osd->od_brw_stats.hist;
+
+       if (rw == READ) {
+               atomic_dec(&osd->od_r_in_flight);
+               lprocfs_oh_tally_log2(&h[BRW_R_PAGES], npages);
+               if (disksize > 0)
+                       lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], disksize);
+               if (elapsed)
+                       lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME], elapsed);
+
+       } else {
+               atomic_dec(&osd->od_w_in_flight);
+               lprocfs_oh_tally_log2(&h[BRW_W_PAGES], npages);
+               if (disksize > 0)
+                       lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], disksize);
+               if (elapsed)
+                       lprocfs_oh_tally_log2(&h[BRW_W_IO_TIME], elapsed);
+       }
+}
 
 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
-                       struct lu_buf *buf, loff_t *pos,
-                       struct lustre_capa *capa)
+                       struct lu_buf *buf, loff_t *pos)
 {
        struct osd_object *obj  = osd_dt_obj(dt);
        struct osd_device *osd = osd_obj2dev(obj);
        uint64_t           old_size;
        int                size = buf->lb_len;
        int                rc;
+       unsigned long      start;
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
+
+       start = cfs_time_current();
 
        read_lock(&obj->oo_attr_lock);
        old_size = obj->oo_attr.la_size;
@@ -97,22 +133,22 @@ static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
                        size = old_size - *pos;
        }
 
-       rc = -dmu_read(osd->od_objset.os, obj->oo_db->db_object, *pos, size,
+       record_start_io(osd, READ, 0);
+
+       rc = -dmu_read(osd->od_os, obj->oo_dn->dn_object, *pos, size,
                        buf->lb_buf, DMU_READ_PREFETCH);
+
+       record_end_io(osd, READ, cfs_time_current() - start, size,
+                     size >> PAGE_SHIFT);
        if (rc == 0) {
                rc = size;
                *pos += size;
-
-               /* XXX: workaround for bug in HEAD: fsfilt_ldiskfs_read() returns
-                * requested number of bytes, not actually read ones */
-               if (S_ISLNK(obj->oo_dt.do_lu.lo_header->loh_attr))
-                       rc = buf->lb_len;
        }
        return rc;
 }
 
 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
-                               const loff_t size, loff_t pos,
+                               const struct lu_buf *buf, loff_t pos,
                                struct thandle *th)
 {
        struct osd_object  *obj  = osd_dt_obj(dt);
@@ -128,20 +164,19 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
         * LOHA_EXISTs is supposed to be the last step in the
         * initialization */
 
-       /* declare possible size change. notice we can't check
-        * current size here as another thread can change it */
-
-       if (dt_object_exists(dt)) {
-               LASSERT(obj->oo_db);
-               oid = obj->oo_db->db_object;
-
-               dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
-       } else {
+       /* size change (in dnode) will be declared by dmu_tx_hold_write() */
+       if (dt_object_exists(dt))
+               oid = obj->oo_dn->dn_object;
+       else
                oid = DMU_NEW_OBJECT;
-               dmu_tx_hold_sa_create(oh->ot_tx, ZFS_SA_BASE_ATTR_SIZE);
-       }
 
-       dmu_tx_hold_write(oh->ot_tx, oid, pos, size);
+       /* XXX: we still miss for append declaration support in ZFS
+        *      -1 means append which is used by llog mostly, llog
+        *      can grow upto LLOG_MIN_CHUNK_SIZE*8 records */
+       if (pos == -1)
+               pos = max_t(loff_t, 256 * 8 * LLOG_MIN_CHUNK_SIZE,
+                           obj->oo_attr.la_size + (2 << 20));
+       dmu_tx_hold_write(oh->ot_tx, oid, pos, buf->lb_len);
 
        /* dt_declare_write() is usually called for system objects, such
         * as llog or last_rcvd files. We needn't enforce quota on those
@@ -153,24 +188,23 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
 
 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
                        const struct lu_buf *buf, loff_t *pos,
-                       struct thandle *th, struct lustre_capa *capa,
-                       int ignore_quota)
+                       struct thandle *th, int ignore_quota)
 {
        struct osd_object  *obj  = osd_dt_obj(dt);
        struct osd_device  *osd = osd_obj2dev(obj);
-       udmu_objset_t      *uos = &osd->od_objset;
        struct osd_thandle *oh;
        uint64_t            offset = *pos;
        int                 rc;
+
        ENTRY;
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
 
        LASSERT(th != NULL);
        oh = container_of0(th, struct osd_thandle, ot_super);
 
-       dmu_write(osd->od_objset.os, obj->oo_db->db_object, offset,
+       dmu_write(osd->od_os, obj->oo_dn->dn_object, offset,
                (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
        write_lock(&obj->oo_attr_lock);
        if (obj->oo_attr.la_size < offset + buf->lb_len) {
@@ -179,7 +213,7 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
                /* osd_object_sa_update() will be copying directly from oo_attr
                 * into dbuf.  any update within a single txg will copy the
                 * most actual */
-               rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
+               rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
                                        &obj->oo_attr.la_size, 8, oh);
                if (unlikely(rc))
                        GOTO(out, rc);
@@ -198,8 +232,8 @@ out:
  * XXX: for the moment I don't want to use lnb_flags for osd-internal
  *      purposes as it's not very well defined ...
  *      instead I use the lowest bit of the address so that:
- *        arc buffer:  .lnb_obj = abuf          (arc we loan for write)
- *        dbuf buffer: .lnb_obj = dbuf | 1      (dbuf we get for read)
+ *        arc buffer:  .lnb_data = abuf          (arc we loan for write)
+ *        dbuf buffer: .lnb_data = dbuf | 1      (dbuf we get for read)
  *        copy buffer: .lnb_page->mapping = obj (page we allocate for write)
  *
  *      bzzz, to blame
@@ -213,55 +247,74 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
        int                i;
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
 
        for (i = 0; i < npages; i++) {
-               if (lnb[i].page == NULL)
+               if (lnb[i].lnb_page == NULL)
                        continue;
-               if (lnb[i].page->mapping == (void *)obj) {
+               if (lnb[i].lnb_page->mapping == (void *)obj) {
                        /* this is anonymous page allocated for copy-write */
-                       lnb[i].page->mapping = NULL;
-                       __free_page(lnb[i].page);
-                       cfs_atomic_dec(&osd->od_zerocopy_alloc);
+                       lnb[i].lnb_page->mapping = NULL;
+                       __free_page(lnb[i].lnb_page);
+                       atomic_dec(&osd->od_zerocopy_alloc);
                } else {
                        /* see comment in osd_bufs_get_read() */
-                       ptr = (unsigned long)lnb[i].dentry;
+                       ptr = (unsigned long)lnb[i].lnb_data;
                        if (ptr & 1UL) {
                                ptr &= ~1UL;
-                               dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
-                               cfs_atomic_dec(&osd->od_zerocopy_pin);
-                       } else if (lnb[i].dentry != NULL) {
-                               dmu_return_arcbuf((void *)lnb[i].dentry);
-                               cfs_atomic_dec(&osd->od_zerocopy_loan);
+                               dmu_buf_rele((void *)ptr, osd_0copy_tag);
+                               atomic_dec(&osd->od_zerocopy_pin);
+                       } else if (lnb[i].lnb_data != NULL) {
+                               dmu_return_arcbuf(lnb[i].lnb_data);
+                               atomic_dec(&osd->od_zerocopy_loan);
                        }
                }
-               lnb[i].page = NULL;
-               lnb[i].dentry = NULL;
+               lnb[i].lnb_page = NULL;
+               lnb[i].lnb_data = NULL;
        }
 
        return 0;
 }
 
-static struct page *kmem_to_page(void *addr)
+static inline struct page *kmem_to_page(void *addr)
 {
-       struct page *page;
-
-       if (kmem_virt(addr))
-               page = vmalloc_to_page(addr);
+       if (is_vmalloc_addr(addr))
+               return vmalloc_to_page(addr);
        else
-               page = virt_to_page(addr);
-
-       return page;
+               return virt_to_page(addr);
 }
 
+/**
+ * Prepare buffers for read.
+ *
+ * The function maps the range described by \a off and \a len to \a lnb array.
+ * dmu_buf_hold_array_by_bonus() finds/creates appropriate ARC buffers, then
+ * we fill \a lnb array with the pages storing ARC buffers. Notice the current
+ * implementationt passes TRUE to dmu_buf_hold_array_by_bonus() to fill ARC
+ * buffers with actual data, I/O is done in the conext of osd_bufs_get_read().
+ * A better implementation would just return the buffers (potentially unfilled)
+ * and subsequent osd_read_prep() would do I/O for many ranges concurrently.
+ *
+ * \param[in] env      environment
+ * \param[in] obj      object
+ * \param[in] off      offset in bytes
+ * \param[in] len      the number of bytes to access
+ * \param[out] lnb     array of local niobufs pointing to the buffers with data
+ *
+ * \retval             0 for success
+ * \retval             negative error number of failure
+ */
 static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
                                loff_t off, ssize_t len, struct niobuf_local *lnb)
 {
        struct osd_device *osd = osd_obj2dev(obj);
-       dmu_buf_t        **dbp;
+       unsigned long      start = cfs_time_current();
        int                rc, i, numbufs, npages = 0;
+       dmu_buf_t        **dbp;
        ENTRY;
 
+       record_start_io(osd, READ, 0);
+
        /* grab buffers for read:
         * OSD API let us to grab buffers first, then initiate IO(s)
         * so that all required IOs will be done in parallel, but at the
@@ -270,9 +323,9 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
         * can get own replacement for dmu_buf_hold_array_by_bonus().
         */
        while (len > 0) {
-               rc = -dmu_buf_hold_array_by_bonus(obj->oo_db, off, len, TRUE,
-                                                 osd_zerocopy_tag, &numbufs,
-                                                 &dbp);
+               rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
+                                                 off, len, TRUE, osd_0copy_tag,
+                                                 &numbufs, &dbp);
                if (unlikely(rc))
                        GOTO(err, rc);
 
@@ -282,7 +335,7 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
 
                        LASSERT(len > 0);
 
-                       cfs_atomic_inc(&osd->od_zerocopy_pin);
+                       atomic_inc(&osd->od_zerocopy_pin);
 
                        bufoff = off - dbp[i]->db_offset;
                        tocpy = min_t(int, dbp[i]->db_size - bufoff, len);
@@ -292,19 +345,19 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
                        dbf = (void *) ((unsigned long)dbp[i] | 1);
 
                        while (tocpy > 0) {
-                               thispage = CFS_PAGE_SIZE;
-                               thispage -= bufoff & (CFS_PAGE_SIZE - 1);
+                               thispage = PAGE_SIZE;
+                               thispage -= bufoff & (PAGE_SIZE - 1);
                                thispage = min(tocpy, thispage);
 
-                               lnb->rc = 0;
+                               lnb->lnb_rc = 0;
                                lnb->lnb_file_offset = off;
-                               lnb->lnb_page_offset = bufoff & ~CFS_PAGE_MASK;
-                               lnb->len = thispage;
-                               lnb->page = kmem_to_page(dbp[i]->db_data +
-                                                               bufoff);
+                               lnb->lnb_page_offset = bufoff & ~PAGE_MASK;
+                               lnb->lnb_len = thispage;
+                               lnb->lnb_page = kmem_to_page(dbp[i]->db_data +
+                                                            bufoff);
                                /* mark just a single slot: we need this
-                                * reference to dbuf to be release once */
-                               lnb->dentry = dbf;
+                                * reference to dbuf to be released once */
+                               lnb->lnb_data = dbf;
                                dbf = NULL;
 
                                tocpy -= thispage;
@@ -316,13 +369,17 @@ static int osd_bufs_get_read(const struct lu_env *env, struct osd_object *obj,
                                lnb++;
                        }
 
-                       /* steal dbuf so dmu_buf_rele_array() cant release it */
+                       /* steal dbuf so dmu_buf_rele_array() can't release
+                        * it */
                        dbp[i] = NULL;
                }
 
-               dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag);
+               dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
        }
 
+       record_end_io(osd, READ, cfs_time_current() - start,
+                     npages * PAGE_SIZE, npages);
+
        RETURN(npages);
 
 err:
@@ -336,14 +393,12 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 {
        struct osd_device *osd = osd_obj2dev(obj);
        int                plen, off_in_block, sz_in_block;
-       int                i = 0, npages = 0;
-       arc_buf_t         *abuf;
-       uint32_t           bs;
-       uint64_t           dummy;
+       int                rc, i = 0, npages = 0;
+       dnode_t *dn = obj->oo_dn;
+       arc_buf_t *abuf;
+       uint32_t bs = dn->dn_datablksz;
        ENTRY;
 
-       dmu_object_size_from_db(obj->oo_db, &bs, &dummy);
-
        /*
         * currently only full blocks are subject to zerocopy approach:
         * so that we're sure nobody is trying to update the same block
@@ -357,30 +412,30 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
                if (sz_in_block == bs) {
                        /* full block, try to use zerocopy */
 
-                       abuf = dmu_request_arcbuf(obj->oo_db, bs);
+                       abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
                        if (unlikely(abuf == NULL))
-                               GOTO(out_err, -ENOMEM);
+                               GOTO(out_err, rc = -ENOMEM);
 
-                       cfs_atomic_inc(&osd->od_zerocopy_loan);
+                       atomic_inc(&osd->od_zerocopy_loan);
 
                        /* go over pages arcbuf contains, put them as
                         * local niobufs for ptlrpc's bulks */
                        while (sz_in_block > 0) {
-                               plen = min_t(int, sz_in_block, CFS_PAGE_SIZE);
+                               plen = min_t(int, sz_in_block, PAGE_SIZE);
 
                                lnb[i].lnb_file_offset = off;
                                lnb[i].lnb_page_offset = 0;
-                               lnb[i].len = plen;
-                               lnb[i].rc = 0;
+                               lnb[i].lnb_len = plen;
+                               lnb[i].lnb_rc = 0;
                                if (sz_in_block == bs)
-                                       lnb[i].dentry = (void *)abuf;
+                                       lnb[i].lnb_data = abuf;
                                else
-                                       lnb[i].dentry = NULL;
+                                       lnb[i].lnb_data = NULL;
 
                                /* this one is not supposed to fail */
-                               lnb[i].page = kmem_to_page(abuf->b_data +
+                               lnb[i].lnb_page = kmem_to_page(abuf->b_data +
                                                        off_in_block);
-                               LASSERT(lnb[i].page);
+                               LASSERT(lnb[i].lnb_page);
 
                                lprocfs_counter_add(osd->od_stats,
                                                LPROC_OSD_ZEROCOPY_IO, 1);
@@ -400,22 +455,22 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 
                        /* can't use zerocopy, allocate temp. buffers */
                        while (sz_in_block > 0) {
-                               plen = min_t(int, sz_in_block, CFS_PAGE_SIZE);
+                               plen = min_t(int, sz_in_block, PAGE_SIZE);
 
                                lnb[i].lnb_file_offset = off;
                                lnb[i].lnb_page_offset = 0;
-                               lnb[i].len = plen;
-                               lnb[i].rc = 0;
-                               lnb[i].dentry = NULL;
+                               lnb[i].lnb_len = plen;
+                               lnb[i].lnb_rc = 0;
+                               lnb[i].lnb_data = NULL;
 
-                               lnb[i].page = alloc_page(OSD_GFP_IO);
-                               if (unlikely(lnb[i].page == NULL))
-                                       GOTO(out_err, -ENOMEM);
+                               lnb[i].lnb_page = alloc_page(OSD_GFP_IO);
+                               if (unlikely(lnb[i].lnb_page == NULL))
+                                       GOTO(out_err, rc = -ENOMEM);
 
-                               LASSERT(lnb[i].page->mapping == NULL);
-                               lnb[i].page->mapping = (void *)obj;
+                               LASSERT(lnb[i].lnb_page->mapping == NULL);
+                               lnb[i].lnb_page->mapping = (void *)obj;
 
-                               cfs_atomic_inc(&osd->od_zerocopy_alloc);
+                               atomic_inc(&osd->od_zerocopy_alloc);
                                lprocfs_counter_add(osd->od_stats,
                                                LPROC_OSD_COPY_IO, 1);
 
@@ -432,18 +487,18 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
 
 out_err:
        osd_bufs_put(env, &obj->oo_dt, lnb, npages);
-       RETURN(-ENOMEM);
+       RETURN(rc);
 }
 
 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
                        loff_t offset, ssize_t len, struct niobuf_local *lnb,
-                       int rw, struct lustre_capa *capa)
+                       int rw)
 {
        struct osd_object *obj  = osd_dt_obj(dt);
        int                rc;
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
 
        if (rw == 0)
                rc = osd_bufs_get_read(env, obj, offset, len, lnb);
@@ -459,95 +514,47 @@ static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
        struct osd_object *obj = osd_dt_obj(dt);
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
 
        return 0;
 }
 
-/* Return number of blocks that aren't mapped in the [start, start + size]
- * region */
-static int osd_count_not_mapped(struct osd_object *obj, uint64_t start,
-                               uint32_t size)
+static inline uint64_t osd_roundup2blocksz(uint64_t size,
+                                          uint64_t offset,
+                                          uint32_t blksz)
 {
-       dmu_buf_impl_t  *dbi = (dmu_buf_impl_t *)obj->oo_db;
-       dmu_buf_impl_t  *db;
-       dnode_t         *dn;
-       uint32_t         blkshift;
-       uint64_t         end, blkid;
-       int              rc;
-       ENTRY;
-
-       DB_DNODE_ENTER(dbi);
-       dn = DB_DNODE(dbi);
-
-       if (dn->dn_maxblkid == 0) {
-               if (start + size <= dn->dn_datablksz)
-                       GOTO(out, size = 0);
-               if (start < dn->dn_datablksz)
-                       start = dn->dn_datablksz;
-               /* assume largest block size */
-               blkshift = SPA_MAXBLOCKSHIFT;
-       } else {
-               /* blocksize can't change */
-               blkshift = dn->dn_datablkshift;
-       }
-
-       /* compute address of last block */
-       end = (start + size - 1) >> blkshift;
-       /* align start on block boundaries */
-       start >>= blkshift;
+       LASSERT(blksz > 0);
 
-       /* size is null, can't be mapped */
-       if (obj->oo_attr.la_size == 0 || dn->dn_maxblkid == 0)
-               GOTO(out, size = (end - start + 1) << blkshift);
+       size += offset % blksz;
 
-       /* beyond EOF, can't be mapped */
-       if (start > dn->dn_maxblkid)
-               GOTO(out, size = (end - start + 1) << blkshift);
+       if (likely(is_power_of_2(blksz)))
+               return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
 
-       size = 0;
-       for (blkid = start; blkid <= end; blkid++) {
-               if (blkid == dn->dn_maxblkid)
-                       /* this one is mapped for sure */
-                       continue;
-               if (blkid > dn->dn_maxblkid) {
-                       size += (end - blkid + 1) << blkshift;
-                       GOTO(out, size);
-               }
-
-               rc = dbuf_hold_impl(dn, 0, blkid, TRUE, FTAG, &db);
-               if (rc) {
-                       /* for ENOENT (block not mapped) and any other errors,
-                        * assume the block isn't mapped */
-                       size += 1 << blkshift;
-                       continue;
-               }
-               dbuf_rele(db, FTAG);
-       }
-
-       GOTO(out, size);
-out:
-       DB_DNODE_EXIT(dbi);
-       return size;
+       size += blksz - 1;
+       do_div(size, blksz);
+       return size * blksz;
 }
 
 static int osd_declare_write_commit(const struct lu_env *env,
-                               struct dt_object *dt,
-                               struct niobuf_local *lnb, int npages,
-                               struct thandle *th)
+                                   struct dt_object *dt,
+                                   struct niobuf_local *lnb, int npages,
+                                   struct thandle *th)
 {
        struct osd_object  *obj = osd_dt_obj(dt);
        struct osd_device  *osd = osd_obj2dev(obj);
        struct osd_thandle *oh;
        uint64_t            offset = 0;
        uint32_t            size = 0;
+       uint32_t blksz = obj->oo_dn->dn_datablksz;
        int                 i, rc, flags = 0;
        bool                ignore_quota = false, synced = false;
        long long           space = 0;
+       struct page        *last_page = NULL;
+       unsigned long       discont_pages = 0;
        ENTRY;
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
 
        LASSERT(lnb);
        LASSERT(npages > 0);
@@ -555,7 +562,10 @@ static int osd_declare_write_commit(const struct lu_env *env,
        oh = container_of0(th, struct osd_thandle, ot_super);
 
        for (i = 0; i < npages; i++) {
-               if (lnb[i].rc)
+               if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
+                       ++discont_pages;
+               last_page = lnb[i].lnb_page;
+               if (lnb[i].lnb_rc)
                        /* ENOSPC, network RPC error, etc.
                         * We don't want to book space for pages which will be
                         * skipped in osd_write_commit(). Hence we skip pages
@@ -570,37 +580,40 @@ static int osd_declare_write_commit(const struct lu_env *env,
                 *
                 * XXX we could handle this on per-lnb basis as done by
                 * grant. */
-               if ((lnb[i].flags & OBD_BRW_NOQUOTA) ||
-                   (lnb[i].flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
+               if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
+                   (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
                    OBD_BRW_FROM_GRANT)
                        ignore_quota = true;
                if (size == 0) {
                        /* first valid lnb */
                        offset = lnb[i].lnb_file_offset;
-                       size = lnb[i].len;
+                       size = lnb[i].lnb_len;
                        continue;
                }
                if (offset + size == lnb[i].lnb_file_offset) {
                        /* this lnb is contiguous to the previous one */
-                       size += lnb[i].len;
+                       size += lnb[i].lnb_len;
                        continue;
                }
 
-               dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
-
-               /* estimating space that will be consumed by a write is rather
+               dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
+                                 offset, size);
+               /* Estimating space to be consumed by a write is rather
                 * complicated with ZFS. As a consequence, we don't account for
-                * indirect blocks and quota overrun will be adjusted once the
-                * operation is committed, if required. */
-               space += osd_count_not_mapped(obj, offset, size);
+                * indirect blocks and just use as a rough estimate the worse
+                * case where the old space is being held by a snapshot. Quota
+                * overrun will be adjusted once the operation is committed, if
+                * required. */
+               space += osd_roundup2blocksz(size, offset, blksz);
 
-               offset = lnb->lnb_file_offset;
-               size = lnb->len;
+               offset = lnb[i].lnb_file_offset;
+               size = lnb[i].lnb_len;
        }
 
        if (size) {
-               dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size);
-               space += osd_count_not_mapped(obj, offset, size);
+               dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
+                                 offset, size);
+               space += osd_roundup2blocksz(size, offset, blksz);
        }
 
        dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
@@ -609,11 +622,12 @@ static int osd_declare_write_commit(const struct lu_env *env,
 
        /* backend zfs filesystem might be configured to store multiple data
         * copies */
-       space  *= osd->od_objset.os->os_copies;
+       space  *= osd->od_os->os_copies;
        space   = toqb(space);
-       CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K of quota "
-              "space\n", npages, space);
+       CDEBUG(D_QUOTA, "writing %d pages, reserving %lldK of quota space\n",
+              npages, space);
 
+       record_start_io(osd, WRITE, discont_pages);
 retry:
        /* acquire quota space if needed */
        rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
@@ -632,74 +646,165 @@ retry:
         * now, once we support multiple objects BRW, this code needs be
         * revised. */
        if (flags & QUOTA_FL_OVER_USRQUOTA)
-               lnb[0].flags |= OBD_BRW_OVER_USRQUOTA;
+               lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
        if (flags & QUOTA_FL_OVER_GRPQUOTA)
-               lnb[0].flags |= OBD_BRW_OVER_GRPQUOTA;
+               lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
 
        RETURN(rc);
 }
 
+/**
+ * Policy to grow ZFS block size by write pattern.
+ * For sequential write, it grows block size gradually until it reaches the
+ * maximum blocksize the dataset can support. Otherwise, it will pick a
+ * a block size by the writing region of this I/O.
+ */
+static int osd_grow_blocksize(struct osd_object *obj, struct osd_thandle *oh,
+                             uint64_t start, uint64_t end)
+{
+       struct osd_device       *osd = osd_obj2dev(obj);
+       dnode_t *dn = obj->oo_dn;
+       uint32_t                 blksz;
+       int                      rc = 0;
+
+       ENTRY;
+
+       if (dn->dn_maxblkid > 0) /* can't change block size */
+               GOTO(out, rc);
+
+       if (dn->dn_datablksz >= osd->od_max_blksz)
+               GOTO(out, rc);
+
+       down_write(&obj->oo_guard);
+
+       blksz = dn->dn_datablksz;
+       if (blksz >= osd->od_max_blksz) /* check again after grabbing lock */
+               GOTO(out_unlock, rc);
+
+       /* now ZFS can support up to 16MB block size, and if the write
+        * is sequential, it just increases the block size gradually */
+       if (start <= blksz) { /* sequential */
+               blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz, end);
+       } else { /* sparse, pick a block size by write region */
+               blksz = (uint32_t)min_t(uint64_t, osd->od_max_blksz,
+                                       end - start);
+       }
+
+       if (!is_power_of_2(blksz))
+               blksz = size_roundup_power2(blksz);
+
+       if (blksz > dn->dn_datablksz) {
+               rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
+                                              blksz, 0, oh->ot_tx);
+               LASSERT(ergo(rc == 0, dn->dn_datablksz >= blksz));
+               if (rc < 0)
+                       CDEBUG(D_INODE, "object "DFID": change block size"
+                              "%u -> %u error rc = %d\n",
+                              PFID(lu_object_fid(&obj->oo_dt.do_lu)),
+                              dn->dn_datablksz, blksz, rc);
+       }
+       EXIT;
+out_unlock:
+       up_write(&obj->oo_guard);
+out:
+       return rc;
+}
+
 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
                        struct niobuf_local *lnb, int npages,
                        struct thandle *th)
 {
        struct osd_object  *obj  = osd_dt_obj(dt);
        struct osd_device  *osd = osd_obj2dev(obj);
-       udmu_objset_t      *uos = &osd->od_objset;
        struct osd_thandle *oh;
        uint64_t            new_size = 0;
        int                 i, rc = 0;
+       unsigned long      iosize = 0;
        ENTRY;
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
 
        LASSERT(th != NULL);
        oh = container_of0(th, struct osd_thandle, ot_super);
 
+       /* adjust block size. Assume the buffers are sorted. */
+       (void)osd_grow_blocksize(obj, oh, lnb[0].lnb_file_offset,
+                                lnb[npages - 1].lnb_file_offset +
+                                lnb[npages - 1].lnb_len);
+
+       /* LU-8791: take oo_guard to avoid the deadlock that changing block
+        * size and assigning arcbuf take place at the same time.
+        *
+        * Thread 1:
+        * osd_write_commit()
+        *  -> osd_grow_blocksize() with osd_object::oo_guard held
+        *   -> dmu_object_set_blocksize()
+        *    -> dnode_set_blksz(), with dnode_t::dn_struct_rwlock
+        *       write lock held
+        *     -> dbuf_new_size()
+        *      -> dmu_buf_will_dirty()
+        *       -> dbuf_read()
+        *        -> wait for the dbuf state to change
+        * Thread 2:
+        * osd_write_commit()
+        *  -> dmu_assign_arcbuf()
+        *   -> dbuf_assign_arcbuf(), set dbuf state to DB_FILL
+        *    -> dbuf_dirty()
+        *     -> try to hold the read lock of dnode_t::dn_struct_rwlock
+        *
+        * By taking the read lock, it can avoid thread 2 to enter into the
+        * critical section of assigning the arcbuf, while thread 1 is
+        * changing the block size.
+        */
+       down_read(&obj->oo_guard);
        for (i = 0; i < npages; i++) {
                CDEBUG(D_INODE, "write %u bytes at %u\n",
-                       (unsigned) lnb[i].len,
+                       (unsigned) lnb[i].lnb_len,
                        (unsigned) lnb[i].lnb_file_offset);
 
-               if (lnb[i].rc) {
+               if (lnb[i].lnb_rc) {
                        /* ENOSPC, network RPC error, etc.
                         * Unlike ldiskfs, zfs allocates new blocks on rewrite,
                         * so we skip this page if lnb_rc is set to -ENOSPC */
                        CDEBUG(D_INODE, "obj "DFID": skipping lnb[%u]: rc=%d\n",
                                PFID(lu_object_fid(&dt->do_lu)), i,
-                               lnb[i].rc);
+                               lnb[i].lnb_rc);
                        continue;
                }
 
-               if (lnb[i].page->mapping == (void *)obj) {
-                       dmu_write(osd->od_objset.os, obj->oo_db->db_object,
-                               lnb[i].lnb_file_offset, lnb[i].len,
-                               kmap(lnb[i].page), oh->ot_tx);
-                       kunmap(lnb[i].page);
-               } else if (lnb[i].dentry) {
-                       LASSERT(((unsigned long)lnb[i].dentry & 1) == 0);
+               if (lnb[i].lnb_page->mapping == (void *)obj) {
+                       dmu_write(osd->od_os, obj->oo_dn->dn_object,
+                               lnb[i].lnb_file_offset, lnb[i].lnb_len,
+                               kmap(lnb[i].lnb_page), oh->ot_tx);
+                       kunmap(lnb[i].lnb_page);
+               } else if (lnb[i].lnb_data) {
+                       LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
                        /* buffer loaned for zerocopy, try to use it.
                         * notice that dmu_assign_arcbuf() is smart
                         * enough to recognize changed blocksize
                         * in this case it fallbacks to dmu_write() */
-                       dmu_assign_arcbuf(obj->oo_db, lnb[i].lnb_file_offset,
-                                       (void *)lnb[i].dentry, oh->ot_tx);
+                       dmu_assign_arcbuf(&obj->oo_dn->dn_bonus->db,
+                                         lnb[i].lnb_file_offset,
+                                         lnb[i].lnb_data, oh->ot_tx);
                        /* drop the reference, otherwise osd_put_bufs()
                         * will be releasing it - bad! */
-                       lnb[i].dentry = NULL;
-                       cfs_atomic_dec(&osd->od_zerocopy_loan);
+                       lnb[i].lnb_data = NULL;
+                       atomic_dec(&osd->od_zerocopy_loan);
                }
 
-               if (new_size < lnb[i].lnb_file_offset + lnb[i].len)
-                       new_size = lnb[i].lnb_file_offset + lnb[i].len;
+               if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
+                       new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
+               iosize += lnb[i].lnb_len;
        }
+       up_read(&obj->oo_guard);
 
        if (unlikely(new_size == 0)) {
                /* no pages to write, no transno is needed */
                th->th_local = 1;
                /* it is important to return 0 even when all lnb_rc == -ENOSPC
                 * since ofd_commitrw_write() retries several times on ENOSPC */
+               record_end_io(osd, WRITE, 0, 0, 0);
                RETURN(0);
        }
 
@@ -710,12 +815,14 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
                /* osd_object_sa_update() will be copying directly from
                 * oo_attr into dbuf. any update within a single txg will copy
                 * the most actual */
-               rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
-                                       &obj->oo_attr.la_size, 8, oh);
+               rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
+                                         &obj->oo_attr.la_size, 8, oh);
        } else {
                write_unlock(&obj->oo_attr_lock);
        }
 
+       record_end_io(osd, WRITE, 0, iosize, npages);
+
        RETURN(rc);
 }
 
@@ -723,28 +830,31 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
                        struct niobuf_local *lnb, int npages)
 {
        struct osd_object *obj  = osd_dt_obj(dt);
-       struct lu_buf      buf;
-       loff_t             offset;
        int                i;
+       loff_t             eof;
 
        LASSERT(dt_object_exists(dt));
-       LASSERT(obj->oo_db);
+       LASSERT(obj->oo_dn);
+
+       read_lock(&obj->oo_attr_lock);
+       eof = obj->oo_attr.la_size;
+       read_unlock(&obj->oo_attr_lock);
 
        for (i = 0; i < npages; i++) {
-               buf.lb_buf = kmap(lnb[i].page);
-               buf.lb_len = lnb[i].len;
-               offset = lnb[i].lnb_file_offset;
+               if (unlikely(lnb[i].lnb_rc < 0))
+                       continue;
 
-               CDEBUG(D_OTHER, "read %u bytes at %u\n",
-                       (unsigned) lnb[i].len,
-                       (unsigned) lnb[i].lnb_file_offset);
-               lnb[i].rc = osd_read(env, dt, &buf, &offset, NULL);
-               kunmap(lnb[i].page);
+               lnb[i].lnb_rc = lnb[i].lnb_len;
+
+               if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
+                       if (eof <= lnb[i].lnb_file_offset)
+                               lnb[i].lnb_rc = 0;
+                       else
+                               lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
 
-               if (lnb[i].rc < buf.lb_len) {
                        /* all subsequent rc should be 0 */
                        while (++i < npages)
-                               lnb[i].rc = 0;
+                               lnb[i].lnb_rc = 0;
                        break;
                }
        }
@@ -766,7 +876,7 @@ static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
  * dmu_tx_hold_sa() and if off < size, dmu_tx_hold_free()
  * called and then assigned to a transaction group.
  */
-static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
+static int __osd_object_punch(objset_t *os, dnode_t *dn, dmu_tx_t *tx,
                                uint64_t size, uint64_t off, uint64_t len)
 {
        int rc = 0;
@@ -780,19 +890,18 @@ static int __osd_object_punch(objset_t *os, dmu_buf_t *db, dmu_tx_t *tx,
        if (len == DMU_OBJECT_END && size == off)
                return 0;
 
+       /* XXX: dnode_free_range() can be used to save on dnode lookup */
        if (off < size)
-               rc = -dmu_free_range(os, db->db_object, off, len, tx);
+               dmu_free_range(os, dn->dn_object, off, len, tx);
 
        return rc;
 }
 
 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
-                       __u64 start, __u64 end, struct thandle *th,
-                       struct lustre_capa *capa)
+                       __u64 start, __u64 end, struct thandle *th)
 {
        struct osd_object  *obj = osd_dt_obj(dt);
        struct osd_device  *osd = osd_obj2dev(obj);
-       udmu_objset_t      *uos = &osd->od_objset;
        struct osd_thandle *oh;
        __u64               len;
        int                 rc = 0;
@@ -812,15 +921,15 @@ static int osd_punch(const struct lu_env *env, struct dt_object *dt,
                len = end - start;
        write_unlock(&obj->oo_attr_lock);
 
-       rc = __osd_object_punch(osd->od_objset.os, obj->oo_db, oh->ot_tx,
+       rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
                                obj->oo_attr.la_size, start, len);
        /* set new size */
        if (len == DMU_OBJECT_END) {
                write_lock(&obj->oo_attr_lock);
                obj->oo_attr.la_size = start;
                write_unlock(&obj->oo_attr_lock);
-               rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos),
-                                       &obj->oo_attr.la_size, 8, oh);
+               rc = osd_object_sa_update(obj, SA_ZPL_SIZE(osd),
+                                         &obj->oo_attr.la_size, 8, oh);
        }
        RETURN(rc);
 }
@@ -845,7 +954,7 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
        /* declare we'll free some blocks ... */
        if (start < obj->oo_attr.la_size) {
                read_unlock(&obj->oo_attr_lock);
-               dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len);
+               dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
        } else {
                read_unlock(&obj->oo_attr_lock);
        }
@@ -858,6 +967,20 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
                                 false));
 }
 
+static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
+                      __u64 start, __u64 end, enum lu_ladvise_type advice)
+{
+       int     rc;
+       ENTRY;
+
+       switch (advice) {
+       default:
+               rc = -ENOTSUPP;
+               break;
+       }
+
+       RETURN(rc);
+}
 
 struct dt_body_operations osd_body_ops = {
        .dbo_read                       = osd_read,
@@ -871,5 +994,5 @@ struct dt_body_operations osd_body_ops = {
        .dbo_read_prep                  = osd_read_prep,
        .dbo_declare_punch              = osd_declare_punch,
        .dbo_punch                      = osd_punch,
+       .dbo_ladvise                    = osd_ladvise,
 };
-