X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_io.c;h=afe987cd4c74021183cf7f305153a9815b362024;hb=a046e879fcadd601c9a19fd906f82ecbd2d4efd5;hp=1d3b23032f7bb5f0dc61a364cd23ca58367124a6;hpb=fc117cc92d224466c1ec45de422f9f6624bbb1da;p=fs%2Flustre-release.git diff --git a/lustre/osd-zfs/osd_io.c b/lustre/osd-zfs/osd_io.c index 1d3b230..afe987c 100644 --- a/lustre/osd-zfs/osd_io.c +++ b/lustre/osd-zfs/osd_io.c @@ -37,7 +37,6 @@ #define DEBUG_SUBSYSTEM S_OSD -#include #include #include #include @@ -45,7 +44,7 @@ #include #include #include -#include /* LLOG_MIN_CHUNK_SIZE definition */ +#include #include "osd_internal.h" @@ -135,8 +134,8 @@ static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt, 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); + rc = osd_dmu_read(osd, obj->oo_dn, *pos, size, buf->lb_buf, + DMU_READ_PREFETCH); record_end_io(osd, READ, cfs_time_current() - start, size, size >> PAGE_SHIFT); @@ -176,14 +175,14 @@ static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt, 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, @@ -204,8 +203,8 @@ static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt, LASSERT(th != NULL); oh = container_of0(th, struct osd_thandle, ot_super); - dmu_write(osd->od_os, obj->oo_dn->dn_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; @@ -265,6 +264,13 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, 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); } @@ -278,6 +284,7 @@ static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt, 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 @@ -305,7 +312,7 @@ static inline struct page *kmem_to_page(void *addr) * \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(); @@ -388,8 +395,31 @@ err: 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; @@ -409,13 +439,15 @@ static int osd_bufs_get_write(const struct lu_env *env, struct osd_object *obj, 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(&dn->dn_bonus->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 @@ -492,7 +524,7 @@ out_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) + enum dt_bufs_type rw) { struct osd_object *obj = osd_dt_obj(dt); int rc; @@ -500,10 +532,10 @@ static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt, LASSERT(dt_object_exists(dt)); 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; } @@ -547,10 +579,11 @@ static int osd_declare_write_commit(const struct lu_env *env, uint32_t size = 0; 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)); @@ -583,7 +616,8 @@ static int osd_declare_write_commit(const struct lu_env *env, 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; @@ -596,8 +630,8 @@ static int osd_declare_write_commit(const struct lu_env *env, continue; } - dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_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 @@ -611,13 +645,11 @@ static int osd_declare_write_commit(const struct lu_env *env, } if (size) { - dmu_tx_hold_write(oh->ot_tx, obj->oo_dn->dn_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 @@ -631,8 +663,8 @@ static int osd_declare_write_commit(const struct lu_env *env, 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); @@ -649,6 +681,10 @@ retry: 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); } @@ -773,17 +809,32 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, 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_dn->dn_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() */ + 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); @@ -791,11 +842,9 @@ static int osd_write_commit(const struct lu_env *env, struct dt_object *dt, * 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); @@ -959,12 +1008,9 @@ static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt, 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,