#include <sys/sa_impl.h>
#include <sys/txg.h>
-static char *osd_0copy_tag = "zerocopy";
+static char osd_0copy_tag[] = "zerocopy";
+static void dbuf_set_pending_evict(dmu_buf_t *db)
+{
+ dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
+ dbi->db_pending_evict = TRUE;
+}
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) {
+ 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 {
+ 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);
- }
+
+ lprocfs_oh_tally_log2(&h[BRW_R_PAGES + rw], npages);
+ if (disksize > 0)
+ lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE + rw], disksize);
+ if (elapsed)
+ lprocfs_oh_tally_log2(&h[BRW_R_IO_TIME + rw], elapsed);
}
static ssize_t __osd_read(const struct lu_env *env, struct dt_object *dt,
{
struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
size_t size = buf->lb_len;
- ktime_t start;
+ hrtime_t start = gethrtime();
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);
+ delta_ms = gethrtime() - start;
+ do_div(delta_ms, NSEC_PER_MSEC);
record_end_io(osd, READ, delta_ms, size, size >> PAGE_SHIFT);
return rc;
uint64_t oid;
ENTRY;
- oh = container_of0(th, struct osd_thandle, ot_super);
+ oh = container_of(th, struct osd_thandle, ot_super);
/* in some cases declare can race with creation (e.g. llog)
* and we need to wait till object is initialized. notice
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, int ignore_quota)
+ struct thandle *th)
{
struct osd_object *obj = osd_dt_obj(dt);
struct osd_device *osd = osd_obj2dev(obj);
LASSERT(obj->oo_dn);
LASSERT(th != NULL);
- oh = container_of0(th, struct osd_thandle, ot_super);
+ oh = container_of(th, struct osd_thandle, ot_super);
osd_dmu_write(osd, obj->oo_dn, offset, (uint64_t)buf->lb_len,
buf->lb_buf, oh->ot_tx);
} else if (lnb[i].lnb_data != NULL) {
int j, apages, abufsz;
abufsz = arc_buf_size(lnb[i].lnb_data);
- apages = abufsz / PAGE_SIZE;
+ apages = abufsz >> PAGE_SHIFT;
/* these references to pages must be invalidated
* to prevent access in osd_bufs_put() */
for (j = 0; j < apages; j++)
* \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,
+ int maxlnb)
{
struct osd_device *osd = osd_obj2dev(obj);
- int rc, i, numbufs, npages = 0;
- ktime_t start = ktime_get();
+ int rc, i, numbufs, npages = 0, drop_cache = 0;
+ hrtime_t start = gethrtime();
dmu_buf_t **dbp;
s64 delta_ms;
ENTRY;
record_start_io(osd, READ, 0);
+ if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize)
+ drop_cache = 1;
+
/* 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
off + len > obj->oo_dn->dn_datablksz)
len = obj->oo_dn->dn_datablksz - off;
+ dbp = NULL;
+ if (unlikely(npages >= maxlnb))
+ GOTO(err, rc = -EOVERFLOW);
+
rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db,
off, len, TRUE, osd_0copy_tag,
&numbufs, &dbp);
dbf = (void *) ((unsigned long)dbp[i] | 1);
while (tocpy > 0) {
+ if (unlikely(npages >= maxlnb))
+ GOTO(err, rc = -EOVERFLOW);
+
thispage = PAGE_SIZE;
thispage -= bufoff & (PAGE_SIZE - 1);
thispage = min(tocpy, thispage);
lnb++;
}
+ if (drop_cache)
+ dbuf_set_pending_evict(dbp[i]);
+
/* steal dbuf so dmu_buf_rele_array() can't release
* it */
dbp[i] = NULL;
dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
}
- delta_ms = ktime_ms_delta(ktime_get(), start);
+ delta_ms = gethrtime() - start;
+ do_div(delta_ms, NSEC_PER_MSEC);
record_end_io(osd, READ, delta_ms, npages * PAGE_SIZE, npages);
RETURN(npages);
err:
LASSERT(rc < 0);
+ if (dbp)
+ dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
osd_bufs_put(env, &obj->oo_dt, lnb - npages, npages);
RETURN(rc);
}
}
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,
+ int maxlnb)
{
struct osd_device *osd = osd_obj2dev(obj);
- int plen, off_in_block, sz_in_block;
+ int poff, plen, off_in_block, sz_in_block;
int rc, i = 0, npages = 0;
dnode_t *dn = obj->oo_dn;
arc_buf_t *abuf;
* so that we're sure nobody is trying to update the same block
*/
while (len > 0) {
- LASSERT(npages < PTLRPC_MAX_BRW_PAGES);
+ if (unlikely(npages >= maxlnb))
+ GOTO(out_err, rc = -EOVERFLOW);
off_in_block = off & (bs - 1);
sz_in_block = min_t(int, bs - off_in_block, len);
while (sz_in_block > 0) {
plen = min_t(int, sz_in_block, PAGE_SIZE);
+ if (unlikely(npages >= maxlnb))
+ GOTO(out_err, rc = -EOVERFLOW);
+
lnb[i].lnb_file_offset = off;
lnb[i].lnb_page_offset = 0;
lnb[i].lnb_len = plen;
LPROC_OSD_TAIL_IO, 1);
/* can't use zerocopy, allocate temp. buffers */
+ poff = off & (PAGE_SIZE - 1);
while (sz_in_block > 0) {
- plen = min_t(int, sz_in_block, PAGE_SIZE);
+ plen = min_t(int, poff + sz_in_block,
+ PAGE_SIZE);
+ plen -= poff;
+
+ if (unlikely(npages >= maxlnb))
+ GOTO(out_err, rc = -EOVERFLOW);
lnb[i].lnb_file_offset = off;
- lnb[i].lnb_page_offset = 0;
+ lnb[i].lnb_page_offset = poff;
+ poff = 0;
+
lnb[i].lnb_len = plen;
lnb[i].lnb_rc = 0;
lnb[i].lnb_data = NULL;
static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
loff_t offset, ssize_t len, struct niobuf_local *lnb,
- enum dt_bufs_type rw)
+ int maxlnb, enum dt_bufs_type rw)
{
struct osd_object *obj = osd_dt_obj(dt);
int rc;
LASSERT(obj->oo_dn);
if (rw & DT_BUFS_TYPE_WRITE)
- rc = osd_bufs_get_write(env, obj, offset, len, lnb);
+ rc = osd_bufs_get_write(env, obj, offset, len, lnb, maxlnb);
else
- rc = osd_bufs_get_read(env, obj, offset, len, lnb);
+ rc = osd_bufs_get_read(env, obj, offset, len, lnb, maxlnb);
return rc;
}
size += offset % blksz;
if (likely(is_power_of_2(blksz)))
- return PO2_ROUNDUP_TYPED(size, blksz, uint64_t);
-
- size += blksz - 1;
- do_div(size, blksz);
- return size * blksz;
+ return round_up(size, blksz);
+ else
+ return DIV_ROUND_UP_ULL(size, blksz) * blksz;
}
static int osd_declare_write_commit(const struct lu_env *env,
uint64_t offset = 0;
uint32_t size = 0;
uint32_t blksz = obj->oo_dn->dn_datablksz;
- int i, rc, flags = 0;
+ int i, rc;
bool synced = false;
long long space = 0;
struct page *last_page = NULL;
unsigned long discont_pages = 0;
+ enum osd_quota_local_flags local_flags = 0;
enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
ENTRY;
LASSERT(lnb);
LASSERT(npages > 0);
- oh = container_of0(th, struct osd_thandle, ot_super);
+ oh = container_of(th, struct osd_thandle, ot_super);
for (i = 0; i < npages; i++) {
if (last_page && lnb[i].lnb_page->index != (last_page->index + 1))
space += osd_roundup2blocksz(size, offset, blksz);
}
- oh->ot_write_commit = 1; /* used in osd_trans_start() for fail_loc */
-
/* backend zfs filesystem might be configured to store multiple data
* copies */
space *= osd->od_os->os_copies;
/* acquire quota space if needed */
rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
obj->oo_attr.la_gid, obj->oo_attr.la_projid,
- space, oh, &flags, declare_flags);
+ space, oh, &local_flags, declare_flags);
- if (!synced && rc == -EDQUOT && (flags & QUOTA_FL_SYNC) != 0) {
+ if (!synced && rc == -EDQUOT &&
+ (local_flags & QUOTA_FL_SYNC) != 0) {
dt_sync(env, th->th_dev);
synced = true;
CDEBUG(D_QUOTA, "retry after sync\n");
- flags = 0;
+ local_flags = 0;
goto retry;
}
/* we need only to store the overquota flags in the first lnb for
* now, once we support multiple objects BRW, this code needs be
* revised. */
- if (flags & QUOTA_FL_OVER_USRQUOTA)
+ if (local_flags & QUOTA_FL_OVER_USRQUOTA)
lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
- if (flags & QUOTA_FL_OVER_GRPQUOTA)
+ if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
#ifdef ZFS_PROJINHERIT
- if (flags & QUOTA_FL_OVER_PRJQUOTA)
+ if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
#endif
return rc;
}
+static void osd_evict_dbufs_after_write(struct osd_object *obj,
+ loff_t off, ssize_t len)
+{
+ dmu_buf_t **dbp;
+ int i, rc, numbufs;
+
+ rc = -dmu_buf_hold_array_by_bonus(&obj->oo_dn->dn_bonus->db, off, len,
+ TRUE, osd_0copy_tag, &numbufs, &dbp);
+ if (unlikely(rc))
+ return;
+
+ for (i = 0; i < numbufs; i++)
+ dbuf_set_pending_evict(dbp[i]);
+
+ dmu_buf_rele_array(dbp, numbufs, osd_0copy_tag);
+}
+
static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
struct niobuf_local *lnb, int npages,
- struct thandle *th)
+ struct thandle *th, __u64 user_size)
{
struct osd_object *obj = osd_dt_obj(dt);
struct osd_device *osd = osd_obj2dev(obj);
struct osd_thandle *oh;
uint64_t new_size = 0;
- int i, rc = 0;
+ int i, abufsz, rc = 0, drop_cache = 0;
unsigned long iosize = 0;
ENTRY;
LASSERT(obj->oo_dn);
LASSERT(th != NULL);
- oh = container_of0(th, struct osd_thandle, ot_super);
+ oh = container_of(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);
+ if (obj->oo_attr.la_size >= osd->od_readcache_max_filesize ||
+ lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len >=
+ osd->od_readcache_max_filesize)
+ drop_cache = 1;
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
+ RETURN(-ENOSPC);
+
+ /* if la_size is already bigger than specified user_size,
+ * ignore user_size
+ */
+ if (obj->oo_attr.la_size > user_size)
+ user_size = 0;
+
/* LU-8791: take oo_guard to avoid the deadlock that changing block
* size and assigning arcbuf take place at the same time.
*
if (lnb[i].lnb_page->mapping == (void *)obj) {
osd_dmu_write(osd, obj->oo_dn, lnb[i].lnb_file_offset,
- lnb[i].lnb_len, kmap(lnb[i].lnb_page),
- oh->ot_tx);
+ lnb[i].lnb_len, kmap(lnb[i].lnb_page) +
+ lnb[i].lnb_page_offset, oh->ot_tx);
kunmap(lnb[i].lnb_page);
iosize += lnb[i].lnb_len;
+ abufsz = lnb[i].lnb_len; /* to drop cache below */
} else if (lnb[i].lnb_data) {
- int j, apages, abufsz;
+ int j, apages;
LASSERT(((unsigned long)lnb[i].lnb_data & 1) == 0);
/* buffer loaned for zerocopy, try to use it.
* notice that dmu_assign_arcbuf() is smart
* in this case it fallbacks to dmu_write() */
abufsz = arc_buf_size(lnb[i].lnb_data);
LASSERT(abufsz & PAGE_MASK);
- apages = abufsz / PAGE_SIZE;
+ apages = abufsz >> PAGE_SHIFT;
LASSERT(i + apages <= npages);
/* these references to pages must be invalidated
* to prevent access in osd_bufs_put() */
lnb[i].lnb_data = NULL;
atomic_dec(&osd->od_zerocopy_loan);
iosize += abufsz;
+ } else {
+ /* we don't want to deal with cache if nothing
+ * has been send to ZFS at this step */
+ continue;
}
+ if (!drop_cache)
+ continue;
+
+ /* we have to mark dbufs for eviction here because
+ * dmu_assign_arcbuf() may create a new dbuf for
+ * loaned abuf */
+ osd_evict_dbufs_after_write(obj, lnb[i].lnb_file_offset,
+ abufsz);
}
up_read(&obj->oo_guard);
RETURN(0);
}
+ /* if file has grown, take user_size into account */
+ if (user_size && new_size > user_size)
+ new_size = user_size;
write_lock(&obj->oo_attr_lock);
if (obj->oo_attr.la_size < new_size) {
obj->oo_attr.la_size = new_size;
lnb[i].lnb_rc = lnb[i].lnb_len;
if (lnb[i].lnb_file_offset + lnb[i].lnb_len >= eof) {
+ /* send complete pages all the time */
if (eof <= lnb[i].lnb_file_offset)
lnb[i].lnb_rc = 0;
- else
- lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
/* all subsequent rc should be 0 */
while (++i < npages)
* 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, dnode_t *dn, dmu_tx_t *tx,
- uint64_t size, uint64_t off, uint64_t len)
+static int __osd_object_punch(struct osd_object *obj, objset_t *os,
+ dmu_tx_t *tx, uint64_t off, uint64_t len)
{
+ dnode_t *dn = obj->oo_dn;
+ uint64_t size = obj->oo_attr.la_size;
int rc = 0;
/* Assert that the transaction has been assigned to a
if (len == DMU_OBJECT_END && size == off)
return 0;
+ /* if object holds encrypted content, we need to make sure we truncate
+ * on an encryption unit boundary, or subsequent reads will get
+ * corrupted content
+ */
+ if (len != DMU_OBJECT_END)
+ len -= LUSTRE_ENCRYPTION_UNIT_SIZE -
+ (off & ~LUSTRE_ENCRYPTION_MASK);
+ if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
+ off & ~LUSTRE_ENCRYPTION_MASK)
+ off = (off & LUSTRE_ENCRYPTION_MASK) +
+ LUSTRE_ENCRYPTION_UNIT_SIZE;
+
+
/* XXX: dnode_free_range() can be used to save on dnode lookup */
if (off < size)
dmu_free_range(os, dn->dn_object, off, len, tx);
LASSERT(osd_invariant(obj));
LASSERT(th != NULL);
- oh = container_of0(th, struct osd_thandle, ot_super);
+ oh = container_of(th, struct osd_thandle, ot_super);
write_lock(&obj->oo_attr_lock);
/* truncate */
len = end - start;
write_unlock(&obj->oo_attr_lock);
- rc = __osd_object_punch(osd->od_os, obj->oo_dn, oh->ot_tx,
- obj->oo_attr.la_size, start, len);
+ rc = __osd_object_punch(obj, osd->od_os, oh->ot_tx, start, len);
+
/* set new size */
if (len == DMU_OBJECT_END) {
write_lock(&obj->oo_attr_lock);
__u64 len;
ENTRY;
- oh = container_of0(handle, struct osd_thandle, ot_super);
+ oh = container_of(handle, struct osd_thandle, ot_super);
read_lock(&obj->oo_attr_lock);
if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size)
len = end - start;
/* declare we'll free some blocks ... */
+ /* if object holds encrypted content, we need to make sure we truncate
+ * on an encryption unit boundary, or subsequent reads will get
+ * corrupted content
+ */
+ if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
+ start & ~LUSTRE_ENCRYPTION_MASK)
+ start = (start & LUSTRE_ENCRYPTION_MASK) +
+ LUSTRE_ENCRYPTION_UNIT_SIZE;
if (start < obj->oo_attr.la_size) {
read_unlock(&obj->oo_attr_lock);
dmu_tx_mark_netfree(oh->ot_tx);
RETURN(rc);
}
+static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
+ __u64 start, __u64 end, int mode, struct thandle *th)
+{
+ int rc = -EOPNOTSUPP;
+ ENTRY;
+
+ /*
+ * space preallocation is not supported for ZFS
+ * Returns -EOPNOTSUPP for now
+ */
+ RETURN(rc);
+}
+
+static int osd_declare_fallocate(const struct lu_env *env,
+ struct dt_object *dt, __u64 start, __u64 end,
+ int mode, struct thandle *th)
+{
+ int rc = -EOPNOTSUPP;
+ ENTRY;
+
+ /*
+ * space preallocation is not supported for ZFS
+ * Returns -EOPNOTSUPP for now
+ */
+ RETURN(rc);
+}
+
+static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
+ loff_t offset, int whence)
+{
+ struct osd_object *obj = osd_dt_obj(dt);
+ uint64_t size = obj->oo_attr.la_size;
+ uint64_t result = offset;
+ int rc;
+ boolean_t hole = whence == SEEK_HOLE;
+
+ ENTRY;
+
+ LASSERT(dt_object_exists(dt));
+ LASSERT(osd_invariant(obj));
+ LASSERT(offset >= 0);
+
+ /* for SEEK_HOLE treat 'offset' beyond the end of file as in real
+ * hole. LOV to decide after all if that real hole or not.
+ */
+ if (offset >= size)
+ RETURN(hole ? offset : -ENXIO);
+
+ rc = osd_dmu_offset_next(osd_obj2dev(obj)->od_os,
+ obj->oo_dn->dn_object, hole, &result);
+ if (rc == ESRCH)
+ RETURN(-ENXIO);
+
+ /* file was dirty, so fall back to using generic logic:
+ * For HOLE return file size, for DATA the result is set
+ * already to the 'offset' parameter value.
+ */
+ if (rc == EBUSY && hole)
+ result = size;
+
+ /* dmu_offset_next() only works on whole blocks so may return SEEK_HOLE
+ * result as end of the last block instead of logical EOF which we need
+ */
+ if (result > size)
+ result = size;
+
+ RETURN(result);
+}
+
struct dt_body_operations osd_body_ops = {
.dbo_read = osd_read,
.dbo_declare_write = osd_declare_write,
.dbo_declare_punch = osd_declare_punch,
.dbo_punch = osd_punch,
.dbo_ladvise = osd_ladvise,
+ .dbo_declare_fallocate = osd_declare_fallocate,
+ .dbo_fallocate = osd_fallocate,
+ .dbo_lseek = osd_lseek,
};
struct dt_body_operations osd_body_scrub_ops = {