* Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2015, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define DEBUG_SUBSYSTEM S_OSD
-#include <lustre_ver.h>
#include <libcfs/libcfs.h>
#include <obd_support.h>
#include <lustre_net.h>
#include <obd_class.h>
#include <lustre_disk.h>
#include <lustre_fid.h>
-#include <lustre/lustre_idl.h> /* LLOG_MIN_CHUNK_SIZE definition */
+#include <lustre_quota.h>
#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)
}
}
-static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
- struct lu_buf *buf, loff_t *pos)
+static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
+ struct lu_buf *buf, loff_t *pos, size_t *size)
{
- 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;
+ struct osd_object *obj = osd_dt_obj(dt);
+ uint64_t old_size;
+ int rc;
LASSERT(dt_object_exists(dt));
- LASSERT(obj->oo_db);
-
- start = cfs_time_current();
+ LASSERT(obj->oo_dn);
read_lock(&obj->oo_attr_lock);
old_size = obj->oo_attr.la_size;
read_unlock(&obj->oo_attr_lock);
- if (*pos + size > old_size) {
+ if (*pos + *size > old_size) {
if (old_size < *pos)
return 0;
- else
- size = old_size - *pos;
+
+ *size = old_size - *pos;
}
- record_start_io(osd, READ, 0);
+ rc = osd_dmu_read(osd_obj2dev(obj), obj->oo_dn, *pos, *size,
+ buf->lb_buf, DMU_READ_PREFETCH);
+ if (!rc) {
+ rc = *size;
+ *pos += *size;
+ }
- rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size,
- buf->lb_buf, DMU_READ_PREFETCH);
+ return rc;
+}
+
+static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
+ struct lu_buf *buf, loff_t *pos)
+{
+ struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
+ size_t size = buf->lb_len;
+ ktime_t start;
+ s64 delta_ms;
+ int rc;
+
+ start = ktime_get();
+ record_start_io(osd, READ, 0);
+ rc = __osd_read(env, dt, buf, pos, &size);
+ delta_ms = ktime_ms_delta(ktime_get(), start);
+ record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
- record_end_io(osd, READ, cfs_time_current() - start, size,
- size >> PAGE_SHIFT);
- if (rc == 0) {
- rc = size;
- *pos += size;
- }
return rc;
}
+static inline ssize_t osd_read_no_record(const struct lu_env *env,
+ struct dt_object *dt,
+ struct lu_buf *buf, loff_t *pos)
+{
+ size_t size = buf->lb_len;
+
+ return __osd_read(env, dt, buf, pos, &size);
+}
+
static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
const struct lu_buf *buf, loff_t pos,
struct thandle *th)
/* size change (in dnode) will be declared by dmu_tx_hold_write() */
if (dt_object_exists(dt))
- oid = obj->oo_db->db_object;
+ oid = obj->oo_dn->dn_object;
else
oid = DMU_NEW_OBJECT;
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);
+ osd_tx_hold_write(oh->ot_tx, oid, obj->oo_dn, 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
* objects, so always set the lqi_space as 0. */
RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
- obj->oo_attr.la_gid, 0, oh, true, NULL,
- false));
+ obj->oo_attr.la_gid, obj->oo_attr.la_projid,
+ 0, oh, NULL, OSD_QID_BLK));
}
static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
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_os, obj->oo_db->db_object, offset,
- (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
+ osd_dmu_write(osd, obj->oo_dn, 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) {
obj->oo_attr.la_size = offset + buf->lb_len;
int i;
LASSERT(dt_object_exists(dt));
- LASSERT(obj->oo_db);
+ LASSERT(obj->oo_dn);
for (i = 0; i < npages; i++) {
if (lnb[i].lnb_page == NULL)
ptr = (unsigned long)lnb[i].lnb_data;
if (ptr & 1UL) {
ptr &= ~1UL;
- dmu_buf_rele((void *)ptr, osd_zerocopy_tag);
+ dmu_buf_rele((void *)ptr, osd_0copy_tag);
atomic_dec(&osd->od_zerocopy_pin);
} else if (lnb[i].lnb_data != NULL) {
+ int j, apages, abufsz;
+ abufsz = arc_buf_size(lnb[i].lnb_data);
+ apages = abufsz / PAGE_SIZE;
+ /* these references to pages must be invalidated
+ * to prevent access in osd_bufs_put() */
+ for (j = 0; j < apages; j++)
+ lnb[i + j].lnb_page = NULL;
dmu_return_arcbuf(lnb[i].lnb_data);
atomic_dec(&osd->od_zerocopy_loan);
}
static inline struct page *kmem_to_page(void *addr)
{
+ LASSERT(!((unsigned long)addr & ~PAGE_MASK));
if (is_vmalloc_addr(addr))
return vmalloc_to_page(addr);
else
* \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)
+ loff_t off, ssize_t len, struct niobuf_local *lnb)
{
struct osd_device *osd = osd_obj2dev(obj);
- unsigned long start = cfs_time_current();
- int rc, i, numbufs, npages = 0;
- dmu_buf_t **dbp;
- ENTRY;
+ int rc, i, numbufs, npages = 0;
+ ktime_t start = ktime_get();
+ dmu_buf_t **dbp;
+ s64 delta_ms;
+ ENTRY;
record_start_io(osd, READ, 0);
/* grab buffers for read:
* If we discover this is a vital for good performance we
* 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);
+ while (len > 0 &&
+ (obj->oo_dn->dn_datablkshift != 0 ||
+ off < obj->oo_dn->dn_datablksz)) {
+ if (obj->oo_dn->dn_datablkshift == 0 &&
+ off + len > obj->oo_dn->dn_datablksz)
+ len = obj->oo_dn->dn_datablksz - off;
+
+ 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);
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);
+ delta_ms = ktime_ms_delta(ktime_get(), start);
+ record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
RETURN(npages);
RETURN(rc);
}
+static inline arc_buf_t *osd_request_arcbuf(dnode_t *dn, size_t bs)
+{
+ arc_buf_t *abuf;
+
+ abuf = dmu_request_arcbuf(&dn->dn_bonus->db, bs);
+ if (unlikely(!abuf))
+ return ERR_PTR(-ENOMEM);
+
+#if ZFS_VERSION_CODE < OBD_OCD_VERSION(0, 7, 0, 0)
+ /**
+ * ZFS prior to 0.7.0 doesn't guarantee PAGE_SIZE alignment for zio
+ * blocks smaller than (PAGE_SIZE << 2). This poses a problem of
+ * setting up page array for RDMA transfer. See LU-9305.
+ */
+ if ((unsigned long)abuf->b_data & ~PAGE_MASK) {
+ dmu_return_arcbuf(abuf);
+ return NULL;
+ }
+#endif
+
+ return abuf;
+}
+
static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj,
- loff_t off, ssize_t len, struct niobuf_local *lnb)
+ loff_t off, ssize_t len, struct niobuf_local *lnb)
{
struct osd_device *osd = osd_obj2dev(obj);
int plen, off_in_block, sz_in_block;
int rc, i = 0, npages = 0;
- arc_buf_t *abuf;
- uint32_t bs;
- uint64_t dummy;
+ 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
off_in_block = off & (bs - 1);
sz_in_block = min_t(int, bs - off_in_block, len);
+ abuf = NULL;
if (sz_in_block == bs) {
/* full block, try to use zerocopy */
+ abuf = osd_request_arcbuf(dn, bs);
+ if (unlikely(IS_ERR(abuf)))
+ GOTO(out_err, rc = PTR_ERR(abuf));
+ }
- abuf = dmu_request_arcbuf(obj->oo_db, bs);
- if (unlikely(abuf == NULL))
- GOTO(out_err, rc = -ENOMEM);
-
+ if (abuf != NULL) {
atomic_inc(&osd->od_zerocopy_loan);
/* go over pages arcbuf contains, put them as
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)
+ enum dt_bufs_type 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);
- else
+ if (rw & DT_BUFS_TYPE_WRITE)
rc = osd_bufs_get_write(env, obj, offset, len, lnb);
+ else
+ rc = osd_bufs_get_read(env, obj, offset, len, lnb);
return rc;
}
struct osd_object *obj = osd_dt_obj(dt);
LASSERT(dt_object_exists(dt));
- LASSERT(obj->oo_db);
+ LASSERT(obj->oo_dn);
return 0;
}
-static inline uint32_t osd_get_blocksz(struct osd_object *obj)
-{
- uint32_t blksz;
- u_longlong_t unused;
-
- LASSERT(obj->oo_db);
-
- dmu_object_size_from_db(obj->oo_db, &blksz, &unused);
- return blksz;
-}
-
static inline uint64_t osd_roundup2blocksz(uint64_t size,
uint64_t offset,
uint32_t blksz)
size += offset % blksz;
- if (likely(IS_PO2(blksz)))
+ if (likely(is_power_of_2(blksz)))
return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
size += blksz - 1;
struct osd_thandle *oh;
uint64_t offset = 0;
uint32_t size = 0;
- uint32_t blksz = osd_get_blocksz(obj);
+ uint32_t blksz = obj->oo_dn->dn_datablksz;
int i, rc, flags = 0;
- bool ignore_quota = false, synced = false;
+ bool synced = false;
long long space = 0;
struct page *last_page = NULL;
unsigned long discont_pages = 0;
+ enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
ENTRY;
LASSERT(dt_object_exists(dt));
- LASSERT(obj->oo_db);
+ LASSERT(obj->oo_dn);
LASSERT(lnb);
LASSERT(npages > 0);
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;
+ declare_flags |= OSD_QID_FORCE;
+
if (size == 0) {
/* first valid lnb */
offset = lnb[i].lnb_file_offset;
continue;
}
- dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
- offset, size);
+ osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object,
+ obj->oo_dn, 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 just use as a rough estimate the worse
}
if (size) {
- dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object,
+ osd_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_object, obj->oo_dn,
offset, size);
space += osd_roundup2blocksz(size, offset, blksz);
}
- dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
-
oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
/* backend zfs filesystem might be configured to store multiple data
retry:
/* acquire quota space if needed */
rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
- obj->oo_attr.la_gid, space, oh, true, &flags,
- ignore_quota);
+ obj->oo_attr.la_gid, obj->oo_attr.la_projid,
+ space, oh, &flags, declare_flags);
if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
dt_sync(env, th->th_dev);
lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
if (flags & QUOTA_FL_OVER_GRPQUOTA)
lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
+#ifdef ZFS_PROJINHERIT
+ if (flags & QUOTA_FL_OVER_PRJQUOTA)
+ lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
+#endif
RETURN(rc);
}
uint64_t start, uint64_t end)
{
struct osd_device *osd = osd_obj2dev(obj);
- dmu_buf_impl_t *db = (dmu_buf_impl_t *)obj->oo_db;
- dnode_t *dn;
+ dnode_t *dn = obj->oo_dn;
uint32_t blksz;
int rc = 0;
ENTRY;
- DB_DNODE_ENTER(db);
- dn = DB_DNODE(db);
-
if (dn->dn_maxblkid > 0) /* can't change block size */
GOTO(out, rc);
out_unlock:
up_write(&obj->oo_guard);
out:
- DB_DNODE_EXIT(db);
return 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);
(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].lnb_len,
continue;
}
+ if (new_size < lnb[i].lnb_file_offset + lnb[i].lnb_len)
+ new_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
+ if (lnb[i].lnb_page == NULL)
+ continue;
+
if (lnb[i].lnb_page->mapping == (void *)obj) {
- dmu_write(osd->od_os, obj->oo_db->db_object,
- lnb[i].lnb_file_offset, lnb[i].lnb_len,
- kmap(lnb[i].lnb_page), oh->ot_tx);
+ osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
+ lnb[i].lnb_len, kmap(lnb[i].lnb_page),
+ oh->ot_tx);
kunmap(lnb[i].lnb_page);
+ iosize += lnb[i].lnb_len;
} else if (lnb[i].lnb_data) {
+ int j, apages, abufsz;
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,
+ abufsz = arc_buf_size(lnb[i].lnb_data);
+ LASSERT(abufsz & PAGE_MASK);
+ apages = abufsz / PAGE_SIZE;
+ LASSERT(i + apages <= npages);
+ /* these references to pages must be invalidated
+ * to prevent access in osd_bufs_put() */
+ for (j = 0; j < apages; j++)
+ lnb[i + j].lnb_page = NULL;
+ 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].lnb_data = NULL;
atomic_dec(&osd->od_zerocopy_loan);
+ iosize += abufsz;
}
- 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 */
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;
* 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;
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;
}
len = end - start;
write_unlock(&obj->oo_attr_lock);
- rc = __osd_object_punch(osd->od_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) {
/* 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_mark_netfree(oh->ot_tx);
+ dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object, start, len);
} else {
read_unlock(&obj->oo_attr_lock);
}
- /* ... and we'll modify size attribute */
- dmu_tx_hold_sa(oh->ot_tx, obj->oo_sa_hdl, 0);
-
RETURN(osd_declare_quota(env, osd, obj->oo_attr.la_uid,
- obj->oo_attr.la_gid, 0, oh, true, NULL,
- false));
+ obj->oo_attr.la_gid, obj->oo_attr.la_projid,
+ 0, oh, NULL, OSD_QID_BLK));
}
static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
.dbo_punch = osd_punch,
.dbo_ladvise = osd_ladvise,
};
+
+struct dt_body_operations osd_body_scrub_ops = {
+ .dbo_read = osd_read_no_record,
+ .dbo_declare_write = osd_declare_write,
+ .dbo_write = osd_write,
+};