X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;ds=sidebyside;f=lustre%2Fobdfilter%2Ffilter_io.c;h=773baf4abad664f9a34a6fc60759751ffc13e5c9;hb=a888a27ac14736d3df8e730a3909d026d6f40f49;hp=0885a80319549cd5733f4622597e1b7d5ba9421c;hpb=bdd161755db65e1e7382ac4f22df9f574c1ed52c;p=fs%2Flustre-release.git diff --git a/lustre/obdfilter/filter_io.c b/lustre/obdfilter/filter_io.c index 0885a80..773baf4 100644 --- a/lustre/obdfilter/filter_io.c +++ b/lustre/obdfilter/filter_io.c @@ -43,8 +43,10 @@ static int filter_start_page_read(struct inode *inode, struct niobuf_local *lnb) int rc; page = grab_cache_page(mapping, index); /* locked page */ - if (IS_ERR(page)) - return lnb->rc = PTR_ERR(page); + if (page == NULL) + return lnb->rc = -ENOMEM; + + LASSERT(page->mapping == mapping); lnb->page = page; @@ -97,136 +99,171 @@ err_page: return lnb->rc; } -static struct page *lustre_get_page_write(struct inode *inode, - unsigned long index) +/* Grab the dirty and seen grant announcements from the incoming obdo. + * We will later calculate the clients new grant and return it. + * Caller must hold osfs lock */ +static void filter_grant_incoming(struct obd_export *exp, struct obdo *oa) { - struct address_space *mapping = inode->i_mapping; - struct page *page; - int rc; + struct filter_export_data *fed; + struct obd_device *obd = exp->exp_obd; + ENTRY; - page = grab_cache_page(mapping, index); /* locked page */ + LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock); - if (!IS_ERR(page)) { - /* Note: Called with "O" and "PAGE_SIZE" this is essentially - * a no-op for most filesystems, because we write the whole - * page. For partial-page I/O this will read in the page. - */ - rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE); - if (rc) { - CERROR("page index %lu, rc = %d\n", index, rc); - if (rc != -ENOSPC) - LBUG(); - GOTO(err_unlock, rc); - } - /* XXX not sure if we need this if we are overwriting page */ - if (PageError(page)) { - CERROR("error on page index %lu, rc = %d\n", index, rc); - LBUG(); - GOTO(err_unlock, rc = -EIO); - } + if ((oa->o_valid & (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) != + (OBD_MD_FLBLOCKS|OBD_MD_FLGRANT)) { + oa->o_valid &= ~OBD_MD_FLGRANT; + EXIT; + return; } - return page; -err_unlock: - unlock_page(page); - page_cache_release(page); - return ERR_PTR(rc); + fed = &exp->exp_filter_data; + + /* Add some margin, since there is a small race if other RPCs arrive + * out-or-order and have already consumed some grant. We want to + * leave this here in case there is a large error in accounting. */ + CDEBUG(oa->o_grant > fed->fed_grant + FILTER_GRANT_CHUNK ? + D_ERROR : D_CACHE, + "%s: cli %s/%p reports grant: "LPU64" dropped: %u, local: %lu\n", + obd->obd_name, exp->exp_client_uuid.uuid, exp, oa->o_grant, + oa->o_dropped, fed->fed_grant); + + /* Update our accounting now so that statfs takes it into account. + * Note that fed_dirty is only approximate and can become incorrect + * if RPCs arrive out-of-order. No important calculations depend + * on fed_dirty however. */ + obd->u.filter.fo_tot_dirty += oa->o_dirty - fed->fed_dirty; + if (fed->fed_grant < oa->o_dropped) { + CERROR("%s: cli %s/%p reports %u dropped > fed_grant %lu\n", + obd->obd_name, exp->exp_client_uuid.uuid, exp, + oa->o_dropped, fed->fed_grant); + oa->o_dropped = 0; + } + if (obd->u.filter.fo_tot_granted < oa->o_dropped) { + CERROR("%s: cli %s/%p reports %u dropped > tot_grant "LPU64"\n", + obd->obd_name, exp->exp_client_uuid.uuid, exp, + oa->o_dropped, obd->u.filter.fo_tot_granted); + oa->o_dropped = 0; + } + obd->u.filter.fo_tot_granted -= oa->o_dropped; + fed->fed_grant -= oa->o_dropped; + fed->fed_dirty = oa->o_dirty; + EXIT; } -#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)) -int wait_on_page_locked(struct page *page) -{ - waitfor_one_page(page); - return 0; -} +#define GRANT_FOR_LLOG(obd) 16 -/* We should only change the file mtime (and not the ctime, like - * update_inode_times() in generic_file_write()) when we only change data. */ -static inline void inode_update_time(struct inode *inode, int ctime_too) +/* Figure out how much space is available between what we've granted + * and what remains in the filesystem. Compensate for ext3 indirect + * block overhead when computing how much free space is left ungranted. + * + * Caller must hold obd_osfs_lock. */ +obd_size filter_grant_space_left(struct obd_export *exp) { - time_t now = CURRENT_TIME; - if (inode->i_mtime == now && (!ctime_too || inode->i_ctime == now)) - return; - inode->i_mtime = now; - if (ctime_too) - inode->i_ctime = now; - mark_inode_dirty_sync(inode); -} -#endif + struct obd_device *obd = exp->exp_obd; + int blockbits = obd->u.filter.fo_sb->s_blocksize_bits; + obd_size tot_granted = obd->u.filter.fo_tot_granted, avail, left = 0; + int rc, statfs_done = 0; + + LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock); + + if (time_before(obd->obd_osfs_age, jiffies - HZ)) { +restat: + rc = fsfilt_statfs(obd, obd->u.filter.fo_sb, jiffies + 1); + if (rc) /* N.B. statfs can't really fail */ + RETURN(0); + statfs_done = 1; + } -static int lustre_commit_write(struct niobuf_local *lnb) -{ - struct page *page = lnb->page; - unsigned from = lnb->offset & ~PAGE_MASK; - unsigned to = from + lnb->len; - struct inode *inode = page->mapping->host; - int err; - - LASSERT(to <= PAGE_SIZE); - err = page->mapping->a_ops->commit_write(NULL, page, from, to); -#warning 2.4 folks: wait_on_page_locked does NOT return its error here. - if (!err && IS_SYNC(inode)) - wait_on_page_locked(page); - //SetPageUptodate(page); // the client commit_write will do this - - SetPageReferenced(page); - unlock_page(page); - page_cache_release(page); - return err; -} + avail = obd->obd_osfs.os_bavail; + left = avail - (avail >> (blockbits - 3)); /* (d)indirect */ + if (left > GRANT_FOR_LLOG(obd)) { + left = (left - GRANT_FOR_LLOG(obd)) << blockbits; + } else { + left = 0 /* << blockbits */; + } -int filter_get_page_write(struct inode *inode, struct niobuf_local *lnb, - int *pglocked) -{ - unsigned long index = lnb->offset >> PAGE_SHIFT; - struct address_space *mapping = inode->i_mapping; - struct page *page; - int rc; + if (!statfs_done && left < 32 * FILTER_GRANT_CHUNK + tot_granted) { + CDEBUG(D_CACHE, "fs has no space left and statfs too old\n"); + goto restat; + } - //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL)); - if (*pglocked) - page = grab_cache_page_nowait(mapping, index); /* locked page */ - else - page = grab_cache_page(mapping, index); /* locked page */ + if (left >= tot_granted) { + left -= tot_granted; + } else { + static unsigned long next; + if (left < tot_granted - obd->u.filter.fo_tot_pending && + time_after(jiffies, next)) { + spin_unlock(&obd->obd_osfs_lock); + CERROR("%s: cli %s/%p grant "LPU64" > available " + LPU64" and pending "LPU64"\n", obd->obd_name, + exp->exp_client_uuid.uuid, exp, tot_granted, + left, obd->u.filter.fo_tot_pending); + if (next == 0) + portals_debug_dumplog(); + next = jiffies + 20 * HZ; + spin_lock(&obd->obd_osfs_lock); + } + left = 0; + } + CDEBUG(D_CACHE, "%s: cli %s/%p free: "LPU64" avail: "LPU64" grant "LPU64 + " left: "LPU64" pending: "LPU64"\n", obd->obd_name, + exp->exp_client_uuid.uuid, exp, + obd->obd_osfs.os_bfree << blockbits, avail << blockbits, + tot_granted, left, obd->u.filter.fo_tot_pending); - /* This page is currently locked, so get a temporary page instead. */ - if (page == NULL) { - CDEBUG(D_INFO, "ino %lu page %ld locked\n", inode->i_ino,index); - page = alloc_pages(GFP_KERNEL, 0); /* locked page */ - if (page == NULL) { - CERROR("no memory for a temp page\n"); - GOTO(err, rc = -ENOMEM); - } - page->index = index; - lnb->page = page; - lnb->flags |= N_LOCAL_TEMP_PAGE; - } else if (!IS_ERR(page)) { - unsigned from = lnb->offset & ~PAGE_MASK, to = from + lnb->len; - (*pglocked)++; - - rc = mapping->a_ops->prepare_write(NULL, page, from, to); - if (rc) { - if (rc != -ENOSPC) - CERROR("page index %lu, rc = %d\n", index, rc); - GOTO(err_unlock, rc); - } - /* XXX not sure if we need this if we are overwriting page */ - if (PageError(page)) { - CERROR("error on page index %lu, rc = %d\n", index, rc); - LBUG(); - GOTO(err_unlock, rc = -EIO); + return left; +} + +/* Calculate how much grant space to allocate to this client, based on how + * much space is currently free and how much of that is already granted. + * + * Caller must hold obd_osfs_lock. */ +long filter_grant(struct obd_export *exp, obd_size current_grant, + obd_size want, obd_size fs_space_left) +{ + struct obd_device *obd = exp->exp_obd; + struct filter_export_data *fed = &exp->exp_filter_data; + int blockbits = obd->u.filter.fo_sb->s_blocksize_bits; + __u64 grant = 0; + + LASSERT_SPIN_LOCKED(&obd->obd_osfs_lock); + + /* Grant some fraction of the client's requested grant space so that + * they are not always waiting for write credits (not all of it to + * avoid overgranting in face of multiple RPCs in flight). This + * essentially will be able to control the OSC_MAX_RIF for a client. + * + * If we do have a large disparity between what the client thinks it + * has and what we think it has, don't grant very much and let the + * client consume its grant first. Either it just has lots of RPCs + * in flight, or it was evicted and its grants will soon be used up. */ + if (current_grant < want) { + if (current_grant > fed->fed_grant + FILTER_GRANT_CHUNK) + want = 65536; + grant = min((want >> blockbits) / 2, + (fs_space_left >> blockbits) / 8); + grant <<= blockbits; + + if (grant) { + if (grant > FILTER_GRANT_CHUNK) + grant = FILTER_GRANT_CHUNK; + + obd->u.filter.fo_tot_granted += grant; + fed->fed_grant += grant; } - lnb->page = page; } - return 0; + CDEBUG(D_CACHE,"%s: cli %s/%p wants: "LPU64" granting: "LPU64"\n", + obd->obd_name, exp->exp_client_uuid.uuid, exp, want, grant); + CDEBUG(D_CACHE, + "%s: cli %s/%p tot cached:"LPU64" granted:"LPU64 + " num_exports: %d\n", obd->obd_name, exp->exp_client_uuid.uuid, + exp, obd->u.filter.fo_tot_dirty, + obd->u.filter.fo_tot_granted, obd->obd_num_exports); -err_unlock: - unlock_page(page); - page_cache_release(page); -err: - return lnb->rc = rc; + return grant; } static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa, @@ -235,19 +272,36 @@ static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa, struct niobuf_local *res, struct obd_trans_info *oti) { + struct obd_device *obd = exp->exp_obd; struct obd_run_ctxt saved; struct obd_ioobj *o; struct niobuf_remote *rnb; - struct niobuf_local *lnb; + struct niobuf_local *lnb = NULL; struct fsfilt_objinfo *fso; struct dentry *dentry; struct inode *inode; - int rc = 0, i, j, tot_bytes = 0; + int rc = 0, i, j, tot_bytes = 0, cleanup_phase = 0; unsigned long now = jiffies; ENTRY; - /* We are currently not supporting multi-obj BRW_READ RPCS at all */ + /* We are currently not supporting multi-obj BRW_READ RPCS at all. + * When we do this function's dentry cleanup will need to be fixed */ LASSERT(objcount == 1); + LASSERT(obj->ioo_bufcnt > 0); + + if (oa && oa->o_valid & OBD_MD_FLGRANT) { + spin_lock(&obd->obd_osfs_lock); + filter_grant_incoming(exp, oa); + +#if 0 + /* Reads do not increase grants */ + oa->o_grant = filter_grant(exp, oa->o_grant, oa->o_undirty, + filter_grant_space_left(exp)); +#else + oa->o_grant = 0; +#endif + spin_unlock(&obd->obd_osfs_lock); + } OBD_ALLOC(fso, objcount * sizeof(*fso)); if (fso == NULL) @@ -255,55 +309,45 @@ static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa, memset(res, 0, niocount * sizeof(*res)); - push_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL); + push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL); for (i = 0, o = obj; i < objcount; i++, o++) { - struct filter_dentry_data *fdd; LASSERT(o->ioo_bufcnt); - dentry = filter_oa2dentry(exp->exp_obd, oa); + dentry = filter_oa2dentry(obd, oa); if (IS_ERR(dentry)) - GOTO(out_objinfo, rc = PTR_ERR(dentry)); + GOTO(cleanup, rc = PTR_ERR(dentry)); if (dentry->d_inode == NULL) { CERROR("trying to BRW to non-existent file "LPU64"\n", o->ioo_id); f_dput(dentry); - GOTO(out_objinfo, rc = -ENOENT); + GOTO(cleanup, rc = -ENOENT); } fso[i].fso_dentry = dentry; fso[i].fso_bufcnt = o->ioo_bufcnt; - - fdd = dentry->d_fsdata; - if (fdd == NULL || !atomic_read(&fdd->fdd_open_count)) - CDEBUG(D_PAGE, "I/O to unopened object "LPU64"\n", - o->ioo_id); } if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow prep setup %lus\n", (jiffies - now) / HZ); + CERROR("slow preprw_read setup %lus\n", (jiffies - now) / HZ); + else + CDEBUG(D_INFO, "preprw_read setup: %lu jiffies\n", + (jiffies - now)); for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) { dentry = fso[i].fso_dentry; inode = dentry->d_inode; for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) { - if (j == 0) - lnb->dentry = dentry; - else - lnb->dentry = dget(dentry); - + lnb->dentry = dentry; lnb->offset = rnb->offset; lnb->len = rnb->len; lnb->flags = rnb->flags; - lnb->start = jiffies; if (inode->i_size <= rnb->offset) { /* If there's no more data, abort early. * lnb->page == NULL and lnb->rc == 0, so it's * easy to detect later. */ - f_dput(dentry); - lnb->dentry = NULL; break; } else { rc = filter_start_page_read(inode, lnb); @@ -314,90 +358,188 @@ static int filter_preprw_read(int cmd, struct obd_export *exp, struct obdo *oa, "page err %u@"LPU64" %u/%u %p: rc %d\n", lnb->len, lnb->offset, j, o->ioo_bufcnt, dentry, rc); - f_dput(dentry); - GOTO(out_pages, rc); + cleanup_phase = 1; + GOTO(cleanup, rc); } tot_bytes += lnb->rc; - if (lnb->rc < lnb->len) - break; /* short read */ + if (lnb->rc < lnb->len) { + /* short read, be sure to wait on it */ + lnb++; + break; + } } } if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow prep get page %lus\n", (jiffies - now) / HZ); + CERROR("slow start_page_read %lus\n", (jiffies - now) / HZ); + else + CDEBUG(D_INFO, "start_page_read: %lu jiffies\n", + (jiffies - now)); - lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_READ_BYTES, - tot_bytes); + lprocfs_counter_add(obd->obd_stats, LPROC_FILTER_READ_BYTES, tot_bytes); while (lnb-- > res) { rc = filter_finish_page_read(lnb); if (rc) { CERROR("error page %u@"LPU64" %u %p: rc %d\n", lnb->len, lnb->offset, (int)(lnb - res), lnb->dentry, rc); - f_dput(lnb->dentry); - GOTO(out_pages, rc); + cleanup_phase = 1; + GOTO(cleanup, rc); } } if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow prep finish page %lus\n", (jiffies - now) / HZ); + CERROR("slow finish_page_read %lus\n", (jiffies - now) / HZ); + else + CDEBUG(D_INFO, "finish_page_read: %lu jiffies\n", + (jiffies - now)); + + filter_tally_read(&exp->exp_obd->u.filter, res, niocount); EXIT; -out: - OBD_FREE(fso, objcount * sizeof(*fso)); - /* we saved the journal handle into oti->oti_handle instead */ - current->journal_info = NULL; - pop_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL); - return rc; -out_pages: - while (lnb-- > res) { - page_cache_release(lnb->page); - f_dput(lnb->dentry); + cleanup: + switch (cleanup_phase) { + case 1: + for (lnb = res; lnb < (res + niocount); lnb++) { + if (lnb->page) + page_cache_release(lnb->page); + } + if (res->dentry != NULL) + f_dput(res->dentry); + else + CERROR("NULL dentry in cleanup -- tell CFS\n"); + case 0: + OBD_FREE(fso, objcount * sizeof(*fso)); + pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL); } - goto out; /* dropped the dentry refs already (one per page) */ - -out_objinfo: - for (i = 0; i < objcount && fso[i].fso_dentry; i++) - f_dput(fso[i].fso_dentry); - goto out; + return rc; } -/* We need to balance prepare_write() calls with commit_write() calls. - * If the page has been prepared, but we have no data for it, we don't - * want to overwrite valid data on disk, but we still need to zero out - * data for space which was newly allocated. Like part of what happens - * in __block_prepare_write() for newly allocated blocks. +/* When clients have dirtied as much space as they've been granted they + * fall through to sync writes. These sync writes haven't been expressed + * in grants and need to error with ENOSPC when there isn't room in the + * filesystem for them after grants are taken into account. However, + * writeback of the dirty data that was already granted space can write + * right on through. * - * XXX currently __block_prepare_write() creates buffers for all the - * pages, and the filesystems mark these buffers as BH_New if they - * were newly allocated from disk. We use the BH_New flag similarly. */ -static int filter_commit_write(struct niobuf_local *lnb, int err) + * Caller must hold obd_osfs_lock. */ +static int filter_grant_check(struct obd_export *exp, int objcount, + struct fsfilt_objinfo *fso, int niocount, + struct niobuf_remote *rnb, + struct niobuf_local *lnb, obd_size *left, + struct inode *inode) { -#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)) - if (err) { - unsigned block_start, block_end; - struct buffer_head *bh, *head = lnb->page->buffers; - unsigned blocksize = head->b_size; - - /* debugging: just seeing if this ever happens */ - CDEBUG(err == -ENOSPC ? D_INODE : D_ERROR, - "called for ino %lu:%lu on err %d\n", - lnb->page->mapping->host->i_ino, lnb->page->index, err); - - /* Currently one buffer per page, but in the future... */ - for (bh = head, block_start = 0; bh != head || !block_start; - block_start = block_end, bh = bh->b_this_page) { - block_end = block_start + blocksize; - if (buffer_new(bh)) { - memset(kmap(lnb->page) + block_start, 0, - blocksize); - kunmap(lnb->page); + struct filter_export_data *fed = &exp->exp_filter_data; + int blocksize = exp->exp_obd->u.filter.fo_sb->s_blocksize; + unsigned long used = 0, ungranted = 0, using; + int i, rc = -ENOSPC, obj, n = 0, mask = D_CACHE; + + LASSERT_SPIN_LOCKED(&exp->exp_obd->obd_osfs_lock); + + for (obj = 0; obj < objcount; obj++) { + for (i = 0; i < fso[obj].fso_bufcnt; i++, n++) { + int tmp, bytes; + + /* FIXME: this is calculated with PAGE_SIZE on client */ + bytes = rnb[n].len; + bytes += rnb[n].offset & (blocksize - 1); + tmp = (rnb[n].offset + rnb[n].len) & (blocksize - 1); + if (tmp) + bytes += blocksize - tmp; + + if (rnb[n].flags & OBD_BRW_FROM_GRANT) { + if (fed->fed_grant < used + bytes) { + CDEBUG(D_CACHE, + "%s: cli %s/%p claims %ld+%d " + "GRANT, real grant %lu idx %d\n", + exp->exp_obd->obd_name, + exp->exp_client_uuid.uuid, exp, + used, bytes, fed->fed_grant, n); + mask = D_ERROR; + } else { + used += bytes; + rnb[n].flags |= OBD_BRW_GRANTED; + lnb[n].lnb_grant_used = bytes; + CDEBUG(0, "idx %d used=%lu\n", n, used); + rc = 0; + continue; + } + } + if (*left > ungranted) { + /* if enough space, pretend it was granted */ + ungranted += bytes; + rnb[n].flags |= OBD_BRW_GRANTED; + CDEBUG(0, "idx %d ungranted=%lu\n",n,ungranted); + rc = 0; + continue; } + + /* We can't check for already-mapped blocks here, as + * it requires dropping the osfs lock to do the bmap. + * Instead, we return ENOSPC and in that case we need + * to go through and verify if all of the blocks not + * marked BRW_GRANTED are already mapped and we can + * ignore this error. */ + lnb[n].rc = -ENOSPC; + rnb[n].flags &= OBD_BRW_GRANTED; + CDEBUG(D_CACHE,"%s: cli %s/%p idx %d no space for %d\n", + exp->exp_obd->obd_name, + exp->exp_client_uuid.uuid, exp, n, bytes); } } -#endif - return lustre_commit_write(lnb); + + /* Now substract what client have used already. We don't subtract + * this from the tot_granted yet, so that other client's can't grab + * that space before we have actually allocated our blocks. That + * happens in filter_grant_commit() after the writes are done. */ + *left -= ungranted; + fed->fed_grant -= used; + fed->fed_pending += used; + exp->exp_obd->u.filter.fo_tot_pending += used; + + CDEBUG(mask, + "%s: cli %s/%p used: %lu ungranted: %lu grant: %lu dirty: %lu\n", + exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, used, + ungranted, fed->fed_grant, fed->fed_dirty); + + /* Rough calc in case we don't refresh cached statfs data */ + using = (used + ungranted + 1 ) >> + exp->exp_obd->u.filter.fo_sb->s_blocksize_bits; + if (exp->exp_obd->obd_osfs.os_bavail > using) + exp->exp_obd->obd_osfs.os_bavail -= using; + else + exp->exp_obd->obd_osfs.os_bavail = 0; + + if (fed->fed_dirty < used) { + CERROR("%s: cli %s/%p claims used %lu > fed_dirty %lu\n", + exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, + used, fed->fed_dirty); + used = fed->fed_dirty; + } + exp->exp_obd->u.filter.fo_tot_dirty -= used; + fed->fed_dirty -= used; + + return rc; +} + +static int filter_start_page_write(struct inode *inode, + struct niobuf_local *lnb) +{ + struct page *page = alloc_pages(GFP_HIGHUSER, 0); + if (page == NULL) { + CERROR("no memory for a temp page\n"); + RETURN(lnb->rc = -ENOMEM); + } + POISON_PAGE(page, 0xf1); + if (lnb->len != PAGE_SIZE) { + memset(kmap(page) + lnb->len, 0, PAGE_SIZE - lnb->len); + kunmap(page); + } + page->index = lnb->offset >> PAGE_SHIFT; + lnb->page = page; + + return 0; } /* If we ever start to support multi-object BRW RPCs, we will need to get locks @@ -417,124 +559,104 @@ static int filter_preprw_write(int cmd, struct obd_export *exp, struct obdo *oa, struct obd_trans_info *oti) { struct obd_run_ctxt saved; - struct obd_ioobj *o; struct niobuf_remote *rnb; struct niobuf_local *lnb; - struct fsfilt_objinfo *fso; + struct fsfilt_objinfo fso; struct dentry *dentry; - int pglocked = 0, rc = 0, i, j, tot_bytes = 0; + obd_size left; unsigned long now = jiffies; + int rc = 0, i, tot_bytes = 0, cleanup_phase = 1; ENTRY; LASSERT(objcount == 1); - - OBD_ALLOC(fso, objcount * sizeof(*fso)); - if (fso == NULL) - RETURN(-ENOMEM); + LASSERT(obj->ioo_bufcnt > 0); memset(res, 0, niocount * sizeof(*res)); - push_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL); - for (i = 0, o = obj; i < objcount; i++, o++) { - struct filter_dentry_data *fdd; - LASSERT(o->ioo_bufcnt); - - dentry = filter_oa2dentry(exp->exp_obd, oa); - if (IS_ERR(dentry)) - GOTO(out_objinfo, rc = PTR_ERR(dentry)); + push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL); + dentry = filter_fid2dentry(exp->exp_obd, NULL, obj->ioo_gr, + obj->ioo_id); + if (IS_ERR(dentry)) + GOTO(cleanup, rc = PTR_ERR(dentry)); + + if (dentry->d_inode == NULL) { + CERROR("trying to BRW to non-existent file "LPU64"\n", + obj->ioo_id); + f_dput(dentry); + GOTO(cleanup, rc = -ENOENT); + } - if (dentry->d_inode == NULL) { - CERROR("trying to BRW to non-existent file "LPU64"\n", - o->ioo_id); - f_dput(dentry); - GOTO(out_objinfo, rc = -ENOENT); - } + fso.fso_dentry = dentry; + fso.fso_bufcnt = obj->ioo_bufcnt; - fso[i].fso_dentry = dentry; - fso[i].fso_bufcnt = o->ioo_bufcnt; + if (time_after(jiffies, now + 15 * HZ)) + CERROR("slow preprw_write setup %lus\n", (jiffies - now) / HZ); + else + CDEBUG(D_INFO, "preprw_write setup: %lu jiffies\n", + (jiffies - now)); - down(&dentry->d_inode->i_sem); - fdd = dentry->d_fsdata; - if (fdd == NULL || !atomic_read(&fdd->fdd_open_count)) - CDEBUG(D_PAGE, "I/O to unopened object "LPU64"\n", - o->ioo_id); - } + spin_lock(&exp->exp_obd->obd_osfs_lock); + if (oa) + filter_grant_incoming(exp, oa); + cleanup_phase = 0; - if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow prep setup %lus\n", (jiffies - now) / HZ); - - LASSERT(oti != NULL); - oti->oti_handle = fsfilt_brw_start(exp->exp_obd, objcount, fso, - niocount, oti); - if (IS_ERR(oti->oti_handle)) { - rc = PTR_ERR(oti->oti_handle); - CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR, - "error starting transaction: rc = %d\n", rc); - oti->oti_handle = NULL; - GOTO(out_objinfo, rc); - } + left = filter_grant_space_left(exp); - for (i = 0, o = obj, rnb = nb, lnb = res; i < objcount; i++, o++) { - dentry = fso[i].fso_dentry; - for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) { - if (j == 0) - lnb->dentry = dentry; - else - lnb->dentry = dget(dentry); + rc = filter_grant_check(exp, objcount, &fso, niocount, nb, res, + &left, dentry->d_inode); + if (oa && oa->o_valid & OBD_MD_FLGRANT) + oa->o_grant = filter_grant(exp,oa->o_grant,oa->o_undirty,left); - lnb->offset = rnb->offset; - lnb->len = rnb->len; - lnb->flags = rnb->flags; - lnb->start = jiffies; + spin_unlock(&exp->exp_obd->obd_osfs_lock); - rc = filter_get_page_write(dentry->d_inode, lnb, - &pglocked); - if (rc) - up(&dentry->d_inode->i_sem); + if (rc) { + f_dput(dentry); + GOTO(cleanup, rc); + } - if (rc) { - CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR, - "page err %u@"LPU64" %u/%u %p: rc %d\n", - lnb->len, lnb->offset, j, o->ioo_bufcnt, - dentry, rc); - f_dput(dentry); - GOTO(out_pages, rc); - } - tot_bytes += lnb->len; + for (i = 0, rnb = nb, lnb = res; i < obj->ioo_bufcnt; + i++, lnb++, rnb++) { + /* We still set up for ungranted pages so that granted pages + * can be written to disk as they were promised, and portals + * needs to keep the pages all aligned properly. */ + lnb->dentry = dentry; + lnb->offset = rnb->offset; + lnb->len = rnb->len; + lnb->flags = rnb->flags; + + rc = filter_start_page_write(dentry->d_inode, lnb); + if (rc) { + CDEBUG(D_ERROR, "page err %u@"LPU64" %u/%u %p: rc %d\n", + lnb->len, lnb->offset, + i, obj->ioo_bufcnt, dentry, rc); + while (lnb-- > res) + __free_pages(lnb->page, 0); + f_dput(dentry); + GOTO(cleanup, rc); } + if (lnb->rc == 0) + tot_bytes += lnb->len; } if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow prep get page %lus\n", (jiffies - now) / HZ); + CERROR("slow start_page_write %lus\n", (jiffies - now) / HZ); + else + CDEBUG(D_INFO, "start_page_write: %lu jiffies\n", + (jiffies - now)); lprocfs_counter_add(exp->exp_obd->obd_stats, LPROC_FILTER_WRITE_BYTES, tot_bytes); - EXIT; -out: - OBD_FREE(fso, objcount * sizeof(*fso)); - /* we saved the journal handle into oti->oti_handle instead */ - current->journal_info = NULL; - pop_ctxt(&saved, &exp->exp_obd->u.filter.fo_ctxt, NULL); - return rc; - -out_pages: - while (lnb-- > res) { - filter_commit_write(lnb, rc); - up(&lnb->dentry->d_inode->i_sem); - f_dput(lnb->dentry); - } - filter_finish_transno(exp, oti, rc); - fsfilt_commit(exp->exp_obd, - filter_parent(exp->exp_obd,S_IFREG,obj->ioo_id)->d_inode, - oti->oti_handle, 0); - goto out; /* dropped the dentry refs already (one per page) */ - -out_objinfo: - for (i = 0; i < objcount && fso[i].fso_dentry; i++) { - up(&fso[i].fso_dentry->d_inode->i_sem); - f_dput(fso[i].fso_dentry); +cleanup: + switch(cleanup_phase) { + case 1: + spin_lock(&exp->exp_obd->obd_osfs_lock); + if (oa) + filter_grant_incoming(exp, oa); + spin_unlock(&exp->exp_obd->obd_osfs_lock); + default: ; } - goto out; + pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL); + return rc; } int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa, @@ -551,199 +673,130 @@ int filter_preprw(int cmd, struct obd_export *exp, struct obdo *oa, niocount, nb, res, oti); LBUG(); - return -EPROTO; } -/* It is highly unlikely that we would ever get an error here. The page we want - * to get was previously locked, so it had to have already allocated the space, - * and we were just writing over the same data, so there would be no hole in the - * file. - * - * XXX: possibility of a race with truncate could exist, need to check that. - * There are no guarantees w.r.t. write order even on a local filesystem, - * although the normal response would be to return the number of bytes - * successfully written and leave the rest to the app. */ -static int filter_write_locked_page(struct niobuf_local *lnb) -{ - struct page *lpage; - void *lpage_addr, *lnb_addr; - int rc; - ENTRY; - - lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index); - if (IS_ERR(lpage)) { - rc = PTR_ERR(lpage); - CERROR("error getting locked page index %ld: rc = %d\n", - lnb->page->index, rc); - LBUG(); - lustre_commit_write(lnb); - RETURN(rc); - } - - /* 2 kmaps == vanishingly small deadlock opportunity */ - lpage_addr = kmap(lpage); - lnb_addr = kmap(lnb->page); - - memcpy(lpage_addr, lnb_addr, PAGE_SIZE); - - kunmap(lnb->page); - kunmap(lpage); - - page_cache_release(lnb->page); - - lnb->page = lpage; - rc = lustre_commit_write(lnb); - if (rc) - CERROR("error committing locked page %ld: rc = %d\n", - lnb->page->index, rc); - RETURN(rc); -} - -int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa, - int objcount, struct obd_ioobj *obj, int niocount, - struct niobuf_local *res, struct obd_trans_info *oti) +static int filter_commitrw_read(struct obd_export *exp, struct obdo *oa, + int objcount, struct obd_ioobj *obj, + int niocount, struct niobuf_local *res, + struct obd_trans_info *oti, int rc) { - struct obd_run_ctxt saved; struct obd_ioobj *o; struct niobuf_local *lnb; - struct obd_device *obd = exp->exp_obd; - int found_locked = 0, rc = 0, i; - int nested_trans = current->journal_info != NULL; - unsigned long now = jiffies; /* DEBUGGING OST TIMEOUTS */ + int i, j, drop = 0; ENTRY; - push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL); - - if (cmd & OBD_BRW_WRITE) { - LASSERT(oti); - LASSERT(current->journal_info == NULL || - current->journal_info == oti->oti_handle); - current->journal_info = oti->oti_handle; - } + if (res->dentry != NULL) + drop = (res->dentry->d_inode->i_size > + exp->exp_obd->u.filter.fo_readcache_max_filesize); for (i = 0, o = obj, lnb = res; i < objcount; i++, o++) { - struct inode *inode; - int j; - - /* If all of the page reads were beyond EOF, let's pretend - * this read didn't really happen at all. */ - if (lnb->dentry == NULL) { - oa->o_valid = OBD_MD_FLID|(oa->o_valid&OBD_MD_FLCKSUM); - continue; - } - - inode = igrab(lnb->dentry->d_inode); - - if (cmd & OBD_BRW_WRITE) { - /* FIXME: MULTI OBJECT BRW */ - if (oa && oa->o_valid & (OBD_MD_FLMTIME|OBD_MD_FLCTIME)) - obdo_refresh_inode(inode, oa, OBD_MD_FLATIME | - OBD_MD_FLMTIME | - OBD_MD_FLCTIME); - else - inode_update_time(lnb->dentry->d_inode, 1); - } else if (oa && oa->o_valid & OBD_MD_FLATIME) { - /* Note that we don't necessarily write this to disk */ - obdo_refresh_inode(inode, oa, OBD_MD_FLATIME); - } - for (j = 0 ; j < o->ioo_bufcnt ; j++, lnb++) { - if (lnb->page == NULL) { + if (lnb->page == NULL) continue; + /* drop from cache like truncate_list_pages() */ + if (drop && !TryLockPage(lnb->page)) { + if (lnb->page->mapping) + ll_truncate_complete_page(lnb->page); + unlock_page(lnb->page); } - - if (lnb->flags & N_LOCAL_TEMP_PAGE) { - found_locked++; - continue; - } - - if (time_after(jiffies, lnb->start + 15 * HZ)) - CERROR("slow commitrw %lusi (%lus)\n", - (jiffies - lnb->start) / HZ, - (jiffies - now) / HZ); - - if (cmd & OBD_BRW_WRITE) { - int err = filter_commit_write(lnb, 0); - - if (!rc) - rc = err; - } else { - page_cache_release(lnb->page); - } - - f_dput(lnb->dentry); - if (time_after(jiffies, lnb->start + 15 * HZ)) - CERROR("slow commit_write %lus (%lus)\n", - (jiffies - lnb->start) / HZ, - (jiffies - now) / HZ); - } - - /* FIXME: MULTI OBJECT BRW */ - if (oa) { - oa->o_valid = OBD_MD_FLID|(oa->o_valid&OBD_MD_FLCKSUM); - obdo_from_inode(oa, inode, FILTER_VALID_FLAGS); + page_cache_release(lnb->page); } - - if (cmd & OBD_BRW_WRITE) - up(&inode->i_sem); - - iput(inode); } - for (i = 0, o = obj, lnb = res; found_locked > 0 && i < objcount; - i++, o++) { - int j; + if (res->dentry != NULL) + f_dput(res->dentry); + RETURN(rc); +} - for (j = 0 ; j < o->ioo_bufcnt ; j++, lnb++) { - int err; - if (!(lnb->flags & N_LOCAL_TEMP_PAGE)) - continue; +void flip_into_page_cache(struct inode *inode, struct page *new_page) +{ + struct page *old_page; + int rc; - if (time_after(jiffies, lnb->start + 15 * HZ)) - CERROR("slow commitrw locked %lus (%lus)\n", - (jiffies - lnb->start) / HZ, - (jiffies - now) / HZ); - - err = filter_write_locked_page(lnb); - if (!rc) - rc = err; - f_dput(lnb->dentry); - found_locked--; - - if (time_after(jiffies, lnb->start + 15 * HZ)) - CERROR("slow commit_write locked %lus (%lus)\n", - (jiffies - lnb->start) / HZ, - (jiffies - now) / HZ); + do { + /* the dlm is protecting us from read/write concurrency, so we + * expect this find_lock_page to return quickly. even if we + * race with another writer it won't be doing much work with + * the page locked. we do this 'cause t_c_p expects a + * locked page, and it wants to grab the pagecache lock + * as well. */ + old_page = find_lock_page(inode->i_mapping, new_page->index); + if (old_page) { + ll_truncate_complete_page(old_page); + unlock_page(old_page); + page_cache_release(old_page); } - } - if (cmd & OBD_BRW_WRITE) { - /* We just want any dentry for the commit, for now */ - struct dentry *dparent = filter_parent(obd, S_IFREG, 0); - int err; - - rc = filter_finish_transno(exp, oti, rc); - err = fsfilt_commit(obd, dparent->d_inode, oti->oti_handle, - obd_sync_filter); - if (err) - rc = err; - if (obd_sync_filter) - LASSERT(oti->oti_transno <= obd->obd_last_committed); - if (time_after(jiffies, now + 15 * HZ)) - CERROR("slow commitrw commit %lus\n", (jiffies-now)/HZ); - } +#if 0 /* this should be a /proc tunable someday */ + /* racing o_directs (no locking ioctl) could race adding + * their pages, so we repeat the page invalidation unless + * we successfully added our new page */ + rc = add_to_page_cache_unique(new_page, inode->i_mapping, + new_page->index, + page_hash(inode->i_mapping, + new_page->index)); + if (rc == 0) { + /* add_to_page_cache clears uptodate|dirty and locks + * the page */ + SetPageUptodate(new_page); + unlock_page(new_page); + } +#else + rc = 0; +#endif + } while (rc != 0); +} - LASSERT(nested_trans || current->journal_info == NULL); - pop_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL); - RETURN(rc); +void filter_grant_commit(struct obd_export *exp, int niocount, + struct niobuf_local *res) +{ + struct filter_obd *filter = &exp->exp_obd->u.filter; + struct niobuf_local *lnb = res; + unsigned long pending = 0; + int i; + + spin_lock(&exp->exp_obd->obd_osfs_lock); + for (i = 0, lnb = res; i < niocount; i++, lnb++) + pending += lnb->lnb_grant_used; + + LASSERTF(exp->exp_filter_data.fed_pending >= pending, + "%s: cli %s/%p fed_pending: %lu grant_used: %lu\n", + exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, + exp->exp_filter_data.fed_pending, pending); + exp->exp_filter_data.fed_pending -= pending; + LASSERTF(filter->fo_tot_granted >= pending, + "%s: cli %s/%p tot_granted: "LPU64" grant_used: %lu\n", + exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, + exp->exp_obd->u.filter.fo_tot_granted, pending); + filter->fo_tot_granted -= pending; + LASSERTF(filter->fo_tot_pending >= pending, + "%s: cli %s/%p tot_pending: "LPU64" grant_used: %lu\n", + exp->exp_obd->obd_name, exp->exp_client_uuid.uuid, exp, + filter->fo_tot_pending, pending); + filter->fo_tot_pending -= pending; + + spin_unlock(&exp->exp_obd->obd_osfs_lock); +} + +int filter_commitrw(int cmd, struct obd_export *exp, struct obdo *oa, + int objcount, struct obd_ioobj *obj, int niocount, + struct niobuf_local *res, struct obd_trans_info *oti,int rc) +{ + if (cmd == OBD_BRW_WRITE) + return filter_commitrw_write(exp, oa, objcount, obj, niocount, + res, oti, rc); + if (cmd == OBD_BRW_READ) + return filter_commitrw_read(exp, oa, objcount, obj, niocount, + res, oti, rc); + LBUG(); + return -EPROTO; } -int filter_brw(int cmd, struct lustre_handle *conn, struct obdo *oa, +int filter_brw(int cmd, struct obd_export *exp, struct obdo *oa, struct lov_stripe_md *lsm, obd_count oa_bufs, struct brw_page *pga, struct obd_trans_info *oti) { - struct obd_export *exp; struct obd_ioobj ioo; struct niobuf_local *lnb; struct niobuf_remote *rnb; @@ -751,12 +804,6 @@ int filter_brw(int cmd, struct lustre_handle *conn, struct obdo *oa, int ret = 0; ENTRY; - exp = class_conn2export(conn); - if (exp == NULL) { - CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",conn->cookie); - RETURN(-EINVAL); - } - OBD_ALLOC(lnb, oa_bufs * sizeof(struct niobuf_local)); OBD_ALLOC(rnb, oa_bufs * sizeof(struct niobuf_remote)); @@ -768,9 +815,7 @@ int filter_brw(int cmd, struct lustre_handle *conn, struct obdo *oa, rnb[i].len = pga[i].count; } - ioo.ioo_id = oa->o_id; - ioo.ioo_gr = 0; - ioo.ioo_type = oa->o_mode & S_IFMT; + obdo_to_ioobj(oa, &ioo); ioo.ioo_bufcnt = oa_bufs; ret = filter_preprw(cmd, exp, oa, 1, &ioo, oa_bufs, rnb, lnb, oti); @@ -789,17 +834,16 @@ int filter_brw(int cmd, struct lustre_handle *conn, struct obdo *oa, else memcpy(virt + off, addr + off, pga[i].count); - kunmap(addr); - kunmap(virt); + kunmap(lnb[i].page); + kunmap(pga[i].pg); } - ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti); + ret = filter_commitrw(cmd, exp, oa, 1, &ioo, oa_bufs, lnb, oti, ret); out: if (lnb) OBD_FREE(lnb, oa_bufs * sizeof(struct niobuf_local)); if (rnb) OBD_FREE(rnb, oa_bufs * sizeof(struct niobuf_remote)); - class_export_put(exp); RETURN(ret); }