/*
* 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, 2014, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
static char *osd_zerocopy_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);
+ start = cfs_time_current();
+
read_lock(&obj->oo_attr_lock);
old_size = obj->oo_attr.la_size;
read_unlock(&obj->oo_attr_lock);
size = old_size - *pos;
}
+ record_start_io(osd, READ, 0);
+
rc = -dmu_read(osd->od_os, obj->oo_db->db_object, *pos, size,
buf->lb_buf, DMU_READ_PREFETCH);
+
+ record_end_io(osd, READ, cfs_time_current() - start, size,
+ size >> PAGE_CACHE_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_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);
struct osd_thandle *oh;
uint64_t offset = *pos;
int rc;
+
ENTRY;
LASSERT(dt_object_exists(dt));
LASSERT(th != NULL);
oh = container_of0(th, struct osd_thandle, ot_super);
+ record_start_io(osd, WRITE, 0);
+
dmu_write(osd->od_os, obj->oo_db->db_object, offset,
(uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx);
write_lock(&obj->oo_attr_lock);
rc = buf->lb_len;
out:
+ record_end_io(osd, WRITE, 0, buf->lb_len,
+ buf->lb_len >> PAGE_CACHE_SHIFT);
+
RETURN(rc);
}
* 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
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
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 */
+ * reference to dbuf to be released once */
lnb->lnb_data = dbf;
dbf = NULL;
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);
}
+ record_end_io(osd, READ, cfs_time_current() - start,
+ npages * PAGE_SIZE, npages);
+
RETURN(npages);
err:
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;
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));
oh = container_of0(th, struct osd_thandle, ot_super);
for (i = 0; i < npages; i++) {
+ 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
CDEBUG(D_QUOTA, "writting %d pages, reserving "LPD64"K 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,
struct osd_thandle *oh;
uint64_t new_size = 0;
int i, rc = 0;
+ unsigned long iosize = 0;
ENTRY;
LASSERT(dt_object_exists(dt));
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;
}
if (unlikely(new_size == 0)) {
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);
}
write_unlock(&obj->oo_attr_lock);
}
+ record_end_io(osd, WRITE, 0, iosize, npages);
+
RETURN(rc);
}
struct niobuf_local *lnb, int npages)
{
struct osd_object *obj = osd_dt_obj(dt);
- struct lu_buf buf;
- loff_t offset;
int i;
+ unsigned long size = 0;
+ loff_t eof;
LASSERT(dt_object_exists(dt));
LASSERT(obj->oo_db);
+ 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].lnb_page);
- buf.lb_len = lnb[i].lnb_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].lnb_len,
- (unsigned) lnb[i].lnb_file_offset);
- lnb[i].lnb_rc = osd_read(env, dt, &buf, &offset, NULL);
- kunmap(lnb[i].lnb_page);
+ lnb[i].lnb_rc = lnb[i].lnb_len;
+ size += lnb[i].lnb_rc;
+
+ if (lnb[i].lnb_file_offset + lnb[i].lnb_len > eof) {
+ lnb[i].lnb_rc = eof - lnb[i].lnb_file_offset;
+ if (lnb[i].lnb_rc < 0)
+ lnb[i].lnb_rc = 0;
- if (lnb[i].lnb_rc < buf.lb_len) {
/* all subsequent rc should be 0 */
while (++i < npages)
lnb[i].lnb_rc = 0;
}
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);
.dbo_declare_punch = osd_declare_punch,
.dbo_punch = osd_punch,
};
-