From 88c5beb51d24423d31fbb1f41fa5f7ab501becd2 Mon Sep 17 00:00:00 2001 From: Alex Zhuravlev Date: Tue, 29 May 2012 14:53:35 +0400 Subject: [PATCH] LU-1305 osd: body operations for osd-zfs read/write, get/put bufs, punch Signed-off-by: Alex Zhuravlev Change-Id: Ie2ccd127b52b00904ec0ce5ba143466ca0328f95 Reviewed-on: http://review.whamcloud.com/2968 Reviewed-by: Andreas Dilger Tested-by: Hudson Reviewed-by: Johann Lombardi --- lustre/osd-zfs/osd_io.c | 740 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 740 insertions(+) create mode 100644 lustre/osd-zfs/osd_io.c diff --git a/lustre/osd-zfs/osd_io.c b/lustre/osd-zfs/osd_io.c new file mode 100644 index 0000000..4dafcfb --- /dev/null +++ b/lustre/osd-zfs/osd_io.c @@ -0,0 +1,740 @@ +/* + * GPL HEADER START + * + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). + * + * 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. + * + * GPL HEADER END + */ +/* + * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + */ +/* + * Copyright (c) 2011, 2012 Whamcloud, Inc. + * Use is subject to license terms. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/osd-zfs/osd_io.c + * + * Author: Alex Zhuravlev + * Author: Mike Pershin + */ + +#ifndef EXPORT_SYMTAB +# define EXPORT_SYMTAB +#endif +#define DEBUG_SUBSYSTEM S_OSD + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "osd_internal.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +static char *osd_zerocopy_tag = "zerocopy"; + +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 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; + + LASSERT(dt_object_exists(dt)); + LASSERT(obj->oo_db); + + cfs_read_lock(&obj->oo_attr_lock); + old_size = obj->oo_attr.la_size; + cfs_read_unlock(&obj->oo_attr_lock); + + if (*pos + size > old_size) { + if (old_size < *pos) + return 0; + else + size = old_size - *pos; + } + + rc = -dmu_read(osd->od_objset.os, obj->oo_db->db_object, *pos, size, + buf->lb_buf, DMU_READ_PREFETCH); + 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, + struct thandle *th) +{ + struct osd_object *obj = osd_dt_obj(dt); + struct osd_thandle *oh; + uint64_t oid; + ENTRY; + + oh = container_of0(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 + * 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 { + 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); + + RETURN(0); +} + +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 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(th != NULL); + oh = container_of0(th, struct osd_thandle, ot_super); + + dmu_write(osd->od_objset.os, obj->oo_db->db_object, offset, + (uint64_t)buf->lb_len, buf->lb_buf, oh->ot_tx); + cfs_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; + cfs_write_unlock(&obj->oo_attr_lock); + /* 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); + if (unlikely(rc)) + GOTO(out, rc); + } else { + cfs_write_unlock(&obj->oo_attr_lock); + } + + *pos += buf->lb_len; + rc = buf->lb_len; + +out: + 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) + * copy buffer: .lnb_page->mapping = obj (page we allocate for write) + * + * bzzz, to blame + */ +static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, + struct niobuf_local *lnb, int npages) +{ + struct osd_object *obj = osd_dt_obj(dt); + struct osd_device *osd = osd_obj2dev(obj); + unsigned long ptr; + int i; + + LASSERT(dt_object_exists(dt)); + LASSERT(obj->oo_db); + + for (i = 0; i < npages; i++) { + if (lnb[i].page == NULL) + continue; + if (lnb[i].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); + } else { + /* see comment in osd_bufs_get_read() */ + ptr = (unsigned long)lnb[i].dentry; + 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); + } + } + lnb[i].page = NULL; + lnb[i].dentry = NULL; + } + + return 0; +} + +static struct page *kmem_to_page(void *addr) +{ + struct page *page; + + if (kmem_virt(addr)) + page = vmalloc_to_page(addr); + else + page = virt_to_page(addr); + + return page; +} + +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; + int rc, i, numbufs, npages = 0; + ENTRY; + + /* 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 + * moment DMU doesn't provide us with a method to grab buffers. + * 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); + LASSERT(rc == 0); + + for (i = 0; i < numbufs; i++) { + int bufoff, tocpy, thispage; + void *dbf = dbp[i]; + + LASSERT(len > 0); + + cfs_atomic_inc(&osd->od_zerocopy_pin); + + bufoff = off - dbp[i]->db_offset; + tocpy = min_t(int, dbp[i]->db_size - bufoff, len); + + /* kind of trick to differentiate dbuf vs. arcbuf */ + LASSERT(((unsigned long)dbp[i] & 1) == 0); + dbf = (void *) ((unsigned long)dbp[i] | 1); + + while (tocpy > 0) { + thispage = CFS_PAGE_SIZE; + thispage -= bufoff & (CFS_PAGE_SIZE - 1); + thispage = min(tocpy, thispage); + + lnb->rc = 0; + lnb->lnb_file_offset = off; + lnb->offset = bufoff & ~CFS_PAGE_MASK; + lnb->len = thispage; + 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; + dbf = NULL; + + tocpy -= thispage; + len -= thispage; + bufoff += thispage; + off += thispage; + + npages++; + lnb++; + } + + /* steal dbuf so dmu_buf_rele_array() cant release it */ + dbp[i] = NULL; + } + + dmu_buf_rele_array(dbp, numbufs, osd_zerocopy_tag); + } + + RETURN(npages); +} + +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) +{ + 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; + 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 + */ + while (len > 0) { + LASSERT(npages < PTLRPC_MAX_BRW_PAGES); + + off_in_block = off & (bs - 1); + sz_in_block = min_t(int, bs - off_in_block, len); + + if (sz_in_block == bs) { + /* full block, try to use zerocopy */ + + abuf = dmu_request_arcbuf(obj->oo_db, bs); + if (unlikely(abuf == NULL)) + GOTO(out_err, -ENOMEM); + + cfs_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); + + lnb[i].lnb_file_offset = off; + lnb[i].offset = 0; + lnb[i].len = plen; + lnb[i].rc = 0; + if (sz_in_block == bs) + lnb[i].dentry = (void *)abuf; + else + lnb[i].dentry = NULL; + + /* this one is not supposed to fail */ + lnb[i].page = kmem_to_page(abuf->b_data + + off_in_block); + LASSERT(lnb[i].page); + + lprocfs_counter_add(osd->od_stats, + LPROC_OSD_ZEROCOPY_IO, 1); + + sz_in_block -= plen; + len -= plen; + off += plen; + off_in_block += plen; + i++; + npages++; + } + } else { + if (off_in_block == 0 && len < bs && + off + len >= obj->oo_attr.la_size) + lprocfs_counter_add(osd->od_stats, + LPROC_OSD_TAIL_IO, 1); + + /* can't use zerocopy, allocate temp. buffers */ + while (sz_in_block > 0) { + plen = min_t(int, sz_in_block, CFS_PAGE_SIZE); + + lnb[i].lnb_file_offset = off; + lnb[i].offset = 0; + lnb[i].len = plen; + lnb[i].rc = 0; + lnb[i].dentry = NULL; + + lnb[i].page = alloc_page(OSD_GFP_IO); + if (unlikely(lnb[i].page == NULL)) + GOTO(out_err, -ENOMEM); + + LASSERT(lnb[i].page->mapping == NULL); + lnb[i].page->mapping = (void *)obj; + + cfs_atomic_inc(&osd->od_zerocopy_alloc); + lprocfs_counter_add(osd->od_stats, + LPROC_OSD_COPY_IO, 1); + + sz_in_block -= plen; + len -= plen; + off += plen; + i++; + npages++; + } + } + } + + RETURN(npages); + +out_err: + osd_bufs_put(env, &obj->oo_dt, lnb, npages); + RETURN(-ENOMEM); +} + +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) +{ + struct osd_object *obj = osd_dt_obj(dt); + int rc; + + LASSERT(dt_object_exists(dt)); + LASSERT(obj->oo_db); + + if (rw == 0) + rc = osd_bufs_get_read(env, obj, offset, len, lnb); + else + rc = osd_bufs_get_write(env, obj, offset, len, lnb); + + return rc; +} + +static int osd_write_prep(const struct lu_env *env, struct dt_object *dt, + struct niobuf_local *lnb, int npages) +{ + struct osd_object *obj = osd_dt_obj(dt); + + LASSERT(dt_object_exists(dt)); + LASSERT(obj->oo_db); + + return 0; +} + +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 osd_object *obj = osd_dt_obj(dt); + struct osd_thandle *oh; + uint64_t offset = 0; + uint32_t size = 0; + int i; + ENTRY; + + LASSERT(dt_object_exists(dt)); + LASSERT(obj->oo_db); + + LASSERT(lnb); + LASSERT(npages > 0); + + oh = container_of0(th, struct osd_thandle, ot_super); + + for (i = 0; i < npages; i++) { + if (lnb[i].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 + * with lnb_rc != 0 here too */ + continue; + if (size == 0) { + /* first valid lnb */ + offset = lnb[i].lnb_file_offset; + size = lnb[i].len; + continue; + } + if (offset + size == lnb[i].lnb_file_offset) { + /* this lnb is contiguous to the previous one */ + size += lnb[i].len; + continue; + } + + dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size); + + offset = lnb->lnb_file_offset; + size = lnb->len; + } + + if (size) + dmu_tx_hold_write(oh->ot_tx, obj->oo_db->db_object, offset,size); + + 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 */ + + RETURN(0); +} + +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; + ENTRY; + + LASSERT(dt_object_exists(dt)); + LASSERT(obj->oo_db); + + LASSERT(th != NULL); + oh = container_of0(th, struct osd_thandle, ot_super); + + for (i = 0; i < npages; i++) { + CDEBUG(D_INODE, "write %u bytes at %u\n", + (unsigned) lnb[i].len, + (unsigned) lnb[i].lnb_file_offset); + + if (lnb[i].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); + 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); + /* 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); + /* drop the reference, otherwise osd_put_bufs() + * will be releasing it - bad! */ + lnb[i].dentry = NULL; + cfs_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 (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 */ + RETURN(0); + } + + cfs_write_lock(&obj->oo_attr_lock); + if (obj->oo_attr.la_size < new_size) { + obj->oo_attr.la_size = new_size; + cfs_write_unlock(&obj->oo_attr_lock); + /* 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); + } else { + cfs_write_unlock(&obj->oo_attr_lock); + } + + RETURN(rc); +} + +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; + + LASSERT(dt_object_exists(dt)); + LASSERT(obj->oo_db); + + 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; + + 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); + + if (lnb[i].rc < buf.lb_len) { + /* all subsequent rc should be 0 */ + while (++i < npages) + lnb[i].rc = 0; + break; + } + } + + return 0; +} + +/* + * Punch/truncate an object + * + * IN: db - dmu_buf of the object to free data in. + * off - start of section to free. + * len - length of section to free (DMU_OBJECT_END => to EOF). + * + * RETURN: 0 if success + * error code if failure + * + * The transaction passed to this routine must have + * 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, + uint64_t size, uint64_t off, uint64_t len) +{ + int rc = 0; + + /* Assert that the transaction has been assigned to a + transaction group. */ + LASSERT(tx->tx_txg != 0); + /* + * Nothing to do if file already at desired length. + */ + if (len == DMU_OBJECT_END && size == off) + return 0; + + if (off < size) + rc = -dmu_free_range(os, db->db_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) +{ + 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; + ENTRY; + + LASSERT(dt_object_exists(dt)); + LASSERT(osd_invariant(obj)); + + LASSERT(th != NULL); + oh = container_of0(th, struct osd_thandle, ot_super); + + cfs_write_lock(&obj->oo_attr_lock); + /* truncate */ + if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size) + len = DMU_OBJECT_END; + else + len = end - start; + cfs_write_unlock(&obj->oo_attr_lock); + + rc = __osd_object_punch(osd->od_objset.os, obj->oo_db, oh->ot_tx, + obj->oo_attr.la_size, start, len); + /* set new size */ + if (len == DMU_OBJECT_END) { + cfs_write_lock(&obj->oo_attr_lock); + obj->oo_attr.la_size = start; + cfs_write_unlock(&obj->oo_attr_lock); + rc = osd_object_sa_update(obj, SA_ZPL_SIZE(uos), + &obj->oo_attr.la_size, 8, oh); + } + RETURN(rc); +} + +static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, + __u64 start, __u64 end, struct thandle *handle) +{ + struct osd_object *obj = osd_dt_obj(dt); + struct osd_thandle *oh; + __u64 len; + ENTRY; + + oh = container_of0(handle, struct osd_thandle, ot_super); + + cfs_read_lock(&obj->oo_attr_lock); + if (end == OBD_OBJECT_EOF || end >= obj->oo_attr.la_size) + len = DMU_OBJECT_END; + else + len = end - start; + + /* declare we'll free some blocks ... */ + if (start < obj->oo_attr.la_size) { + cfs_read_unlock(&obj->oo_attr_lock); + dmu_tx_hold_free(oh->ot_tx, obj->oo_db->db_object, start, len); + } else { + cfs_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(0); +} + + +struct dt_body_operations osd_body_ops = { + .dbo_read = osd_read, + .dbo_declare_write = osd_declare_write, + .dbo_write = osd_write, + .dbo_bufs_get = osd_bufs_get, + .dbo_bufs_put = osd_bufs_put, + .dbo_write_prep = osd_write_prep, + .dbo_declare_write_commit = osd_declare_write_commit, + .dbo_write_commit = osd_write_commit, + .dbo_read_prep = osd_read_prep, + .do_declare_punch = osd_declare_punch, + .do_punch = osd_punch, +}; + -- 1.8.3.1