X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdc%2Fmdc_request.c;h=aca18074651b33b66a77ca5121ef359995dd192b;hp=aa8e39d4481010b4425ac7cfd76855b9e9d3c19d;hb=0d8c5ccc4ecfe7c0d10a0a4f92fd291320a97190;hpb=4322e0f99c87bc0412d315d0674d70cc1ffc0bb4 diff --git a/lustre/mdc/mdc_request.c b/lustre/mdc/mdc_request.c index aa8e39d..aca1807 100644 --- a/lustre/mdc/mdc_request.c +++ b/lustre/mdc/mdc_request.c @@ -47,11 +47,15 @@ #endif #include +#include #include +#include #include #include #include #include +#include +#include #include "mdc_internal.h" @@ -92,15 +96,15 @@ static inline int mdc_queue_wait(struct ptlrpc_request *req) struct client_obd *cli = &req->rq_import->imp_obd->u.cli; int rc; - /* mdc_enter_request() ensures that this client has no more + /* obd_get_request_slot() ensures that this client has no more * than cl_max_rpcs_in_flight RPCs simultaneously inf light * against an MDT. */ - rc = mdc_enter_request(cli); + rc = obd_get_request_slot(cli); if (rc != 0) return rc; rc = ptlrpc_queue_wait(req); - mdc_exit_request(cli); + obd_put_request_slot(cli); return rc; } @@ -190,14 +194,13 @@ static int mdc_getattr_common(struct obd_export *exp, CDEBUG(D_NET, "mode: %o\n", body->mode); - if (body->eadatasize != 0) { - mdc_update_max_ea_from_body(exp, body); - - eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD, - body->eadatasize); - if (eadata == NULL) - RETURN(-EPROTO); - } + mdc_update_max_ea_from_body(exp, body); + if (body->eadatasize != 0) { + eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD, + body->eadatasize); + if (eadata == NULL) + RETURN(-EPROTO); + } if (body->valid & OBD_MD_FLRMTPERM) { struct mdt_remote_perm *perm; @@ -371,7 +374,7 @@ static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt, /* Flush local XATTR locks to get rid of a possible cancel RPC */ if (opcode == MDS_REINT && fid_is_sane(fid) && exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) { - CFS_LIST_HEAD(cancels); + struct list_head cancels = LIST_HEAD_INIT(cancels); int count; /* Without that packing would fail */ @@ -403,8 +406,8 @@ static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt, sizeof(struct mdt_rec_reint)); rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT); rec->sx_opcode = REINT_SETXATTR; - rec->sx_fsuid = current_fsuid(); - rec->sx_fsgid = current_fsgid(); + rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid()); + rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid()); rec->sx_cap = cfs_curproc_cap_pack(); rec->sx_suppgid1 = suppgid; rec->sx_suppgid2 = -1; @@ -570,25 +573,25 @@ int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req, "but eadatasize 0\n"); RETURN(-EPROTO); } - if (md->body->valid & OBD_MD_MEA) { - lmvsize = md->body->eadatasize; - lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD, - lmvsize); - if (!lmv) - GOTO(out, rc = -EPROTO); - - rc = obd_unpackmd(md_exp, (void *)&md->mea, lmv, - lmvsize); - if (rc < 0) - GOTO(out, rc); - - if (rc < sizeof(*md->mea)) { - CDEBUG(D_INFO, "size too small: " - "rc < sizeof(*md->mea) (%d < %d)\n", - rc, (int)sizeof(*md->mea)); - GOTO(out, rc = -EPROTO); - } - } + if (md->body->valid & OBD_MD_MEA) { + lmvsize = md->body->eadatasize; + lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD, + lmvsize); + if (!lmv) + GOTO(out, rc = -EPROTO); + + rc = obd_unpackmd(md_exp, (void *)&md->lmv, lmv, + lmvsize); + if (rc < 0) + GOTO(out, rc); + + if (rc < sizeof(*md->lmv)) { + CDEBUG(D_INFO, "size too small: " + "rc < sizeof(*md->lmv) (%d < %d)\n", + rc, (int)sizeof(*md->lmv)); + GOTO(out, rc = -EPROTO); + } + } } rc = 0; @@ -876,7 +879,7 @@ int mdc_close(struct obd_export *exp, struct md_op_data *op_data, req_fmt = &RQF_MDS_RELEASE_CLOSE; /* allocate a FID for volatile file */ - rc = mdc_fid_alloc(exp, &op_data->op_fid2, op_data); + rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data); if (rc < 0) { CERROR("%s: "DFID" failed to allocate FID: %d\n", obd->obd_name, PFID(&op_data->op_fid1), rc); @@ -924,10 +927,10 @@ int mdc_close(struct obd_export *exp, struct md_op_data *op_data, mdc_close_pack(req, op_data); - req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, - obd->u.cli.cl_max_mds_easize); - req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, RCL_SERVER, - obd->u.cli.cl_max_mds_cookiesize); + req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, + obd->u.cli.cl_default_mds_easize); + req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, RCL_SERVER, + obd->u.cli.cl_default_mds_cookiesize); ptlrpc_request_set_replen(req); @@ -1095,8 +1098,10 @@ out: EXPORT_SYMBOL(mdc_sendpage); #endif -int mdc_readpage(struct obd_export *exp, struct md_op_data *op_data, - struct page **pages, struct ptlrpc_request **request) +static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid, + __u64 offset, struct obd_capa *oc, + struct page **pages, int npages, + struct ptlrpc_request **request) { struct ptlrpc_request *req; struct ptlrpc_bulk_desc *desc; @@ -1111,73 +1116,694 @@ int mdc_readpage(struct obd_export *exp, struct md_op_data *op_data, init_waitqueue_head(&waitq); restart_bulk: - req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE); - if (req == NULL) - RETURN(-ENOMEM); + req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE); + if (req == NULL) + RETURN(-ENOMEM); - mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1); + mdc_set_capa_size(req, &RMF_CAPA1, oc); - rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE); - if (rc) { - ptlrpc_request_free(req); - RETURN(rc); - } + rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE); + if (rc) { + ptlrpc_request_free(req); + RETURN(rc); + } - req->rq_request_portal = MDS_READPAGE_PORTAL; - ptlrpc_at_set_req_timeout(req); + req->rq_request_portal = MDS_READPAGE_PORTAL; + ptlrpc_at_set_req_timeout(req); - desc = ptlrpc_prep_bulk_imp(req, op_data->op_npages, 1, BULK_PUT_SINK, + desc = ptlrpc_prep_bulk_imp(req, npages, 1, BULK_PUT_SINK, MDS_BULK_PORTAL); - if (desc == NULL) { - ptlrpc_request_free(req); - RETURN(-ENOMEM); - } + if (desc == NULL) { + ptlrpc_request_free(req); + RETURN(-ENOMEM); + } - /* NB req now owns desc and will free it when it gets freed */ - for (i = 0; i < op_data->op_npages; i++) + /* NB req now owns desc and will free it when it gets freed */ + for (i = 0; i < npages; i++) ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE); - mdc_readdir_pack(req, op_data->op_offset, - PAGE_CACHE_SIZE * op_data->op_npages, - &op_data->op_fid1, op_data->op_capa1); + mdc_readdir_pack(req, offset, PAGE_CACHE_SIZE * npages, fid, oc); - ptlrpc_request_set_replen(req); - rc = ptlrpc_queue_wait(req); - if (rc) { - ptlrpc_req_finished(req); - if (rc != -ETIMEDOUT) - RETURN(rc); + ptlrpc_request_set_replen(req); + rc = ptlrpc_queue_wait(req); + if (rc) { + ptlrpc_req_finished(req); + if (rc != -ETIMEDOUT) + RETURN(rc); - resends++; - if (!client_should_resend(resends, &exp->exp_obd->u.cli)) { - CERROR("too many resend retries, returning error\n"); - RETURN(-EIO); - } - lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, NULL); - l_wait_event(waitq, 0, &lwi); + resends++; + if (!client_should_resend(resends, &exp->exp_obd->u.cli)) { + CERROR("%s: too many resend retries: rc = %d\n", + exp->exp_obd->obd_name, -EIO); + RETURN(-EIO); + } + lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL, + NULL); + l_wait_event(waitq, 0, &lwi); - goto restart_bulk; - } + goto restart_bulk; + } - rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, - req->rq_bulk->bd_nob_transferred); - if (rc < 0) { - ptlrpc_req_finished(req); - RETURN(rc); - } + rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, + req->rq_bulk->bd_nob_transferred); + if (rc < 0) { + ptlrpc_req_finished(req); + RETURN(rc); + } - if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) { - CERROR("Unexpected # bytes transferred: %d (%ld expected)\n", - req->rq_bulk->bd_nob_transferred, - PAGE_CACHE_SIZE * op_data->op_npages); - ptlrpc_req_finished(req); - RETURN(-EPROTO); - } + if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) { + CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n", + exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred, + PAGE_CACHE_SIZE * npages); + ptlrpc_req_finished(req); + RETURN(-EPROTO); + } - *request = req; - RETURN(0); + *request = req; + RETURN(0); +} + +#ifdef __KERNEL__ +static void mdc_release_page(struct page *page, int remove) +{ + if (remove) { + lock_page(page); + if (likely(page->mapping != NULL)) + truncate_complete_page(page->mapping, page); + unlock_page(page); + } + page_cache_release(page); +} + +static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash, + __u64 *start, __u64 *end, int hash64) +{ + /* + * Complement of hash is used as an index so that + * radix_tree_gang_lookup() can be used to find a page with starting + * hash _smaller_ than one we are looking for. + */ + unsigned long offset = hash_x_index(*hash, hash64); + struct page *page; + int found; + + spin_lock_irq(&mapping->tree_lock); + found = radix_tree_gang_lookup(&mapping->page_tree, + (void **)&page, offset, 1); + if (found > 0) { + struct lu_dirpage *dp; + + page_cache_get(page); + spin_unlock_irq(&mapping->tree_lock); + /* + * In contrast to find_lock_page() we are sure that directory + * page cannot be truncated (while DLM lock is held) and, + * hence, can avoid restart. + * + * In fact, page cannot be locked here at all, because + * mdc_read_page_remote does synchronous io. + */ + wait_on_page_locked(page); + if (PageUptodate(page)) { + dp = kmap(page); + if (BITS_PER_LONG == 32 && hash64) { + *start = le64_to_cpu(dp->ldp_hash_start) >> 32; + *end = le64_to_cpu(dp->ldp_hash_end) >> 32; + *hash = *hash >> 32; + } else { + *start = le64_to_cpu(dp->ldp_hash_start); + *end = le64_to_cpu(dp->ldp_hash_end); + } + if (unlikely(*start == 1 && *hash == 0)) + *hash = *start; + else + LASSERTF(*start <= *hash, "start = "LPX64 + ",end = "LPX64",hash = "LPX64"\n", + *start, *end, *hash); + CDEBUG(D_VFSTRACE, "offset %lx ["LPX64" "LPX64"]," + " hash "LPX64"\n", offset, *start, *end, *hash); + if (*hash > *end) { + kunmap(page); + mdc_release_page(page, 0); + page = NULL; + } else if (*end != *start && *hash == *end) { + /* + * upon hash collision, remove this page, + * otherwise put page reference, and + * ll_get_dir_page() will issue RPC to fetch + * the page we want. + */ + kunmap(page); + mdc_release_page(page, + le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE); + page = NULL; + } + } else { + page_cache_release(page); + page = ERR_PTR(-EIO); + } + } else { + spin_unlock_irq(&mapping->tree_lock); + page = NULL; + } + return page; +} + +/* + * Adjust a set of pages, each page containing an array of lu_dirpages, + * so that each page can be used as a single logical lu_dirpage. + * + * A lu_dirpage is laid out as follows, where s = ldp_hash_start, + * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a + * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end + * value is used as a cookie to request the next lu_dirpage in a + * directory listing that spans multiple pages (two in this example): + * ________ + * | | + * .|--------v------- -----. + * |s|e|f|p|ent|ent| ... |ent| + * '--|-------------- -----' Each CFS_PAGE contains a single + * '------. lu_dirpage. + * .---------v------- -----. + * |s|e|f|p|ent| 0 | ... | 0 | + * '----------------- -----' + * + * However, on hosts where the native VM page size (PAGE_CACHE_SIZE) is + * larger than LU_PAGE_SIZE, a single host page may contain multiple + * lu_dirpages. After reading the lu_dirpages from the MDS, the + * ldp_hash_end of the first lu_dirpage refers to the one immediately + * after it in the same CFS_PAGE (arrows simplified for brevity, but + * in general e0==s1, e1==s2, etc.): + * + * .-------------------- -----. + * |s0|e0|f0|p|ent|ent| ... |ent| + * |---v---------------- -----| + * |s1|e1|f1|p|ent|ent| ... |ent| + * |---v---------------- -----| Here, each CFS_PAGE contains + * ... multiple lu_dirpages. + * |---v---------------- -----| + * |s'|e'|f'|p|ent|ent| ... |ent| + * '---|---------------- -----' + * v + * .----------------------------. + * | next CFS_PAGE | + * + * This structure is transformed into a single logical lu_dirpage as follows: + * + * - Replace e0 with e' so the request for the next lu_dirpage gets the page + * labeled 'next CFS_PAGE'. + * + * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether + * a hash collision with the next page exists. + * + * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span + * to the first entry of the next lu_dirpage. + */ +#if PAGE_CACHE_SIZE > LU_PAGE_SIZE +static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs) +{ + int i; + + for (i = 0; i < cfs_pgs; i++) { + struct lu_dirpage *dp = kmap(pages[i]); + struct lu_dirpage *first = dp; + struct lu_dirent *end_dirent = NULL; + struct lu_dirent *ent; + __u64 hash_end = le64_to_cpu(dp->ldp_hash_end); + __u32 flags = le32_to_cpu(dp->ldp_flags); + + while (--lu_pgs > 0) { + ent = lu_dirent_start(dp); + for (end_dirent = ent; ent != NULL; + end_dirent = ent, ent = lu_dirent_next(ent)); + + /* Advance dp to next lu_dirpage. */ + dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE); + + /* Check if we've reached the end of the CFS_PAGE. */ + if (!((unsigned long)dp & ~CFS_PAGE_MASK)) + break; + + /* Save the hash and flags of this lu_dirpage. */ + hash_end = le64_to_cpu(dp->ldp_hash_end); + flags = le32_to_cpu(dp->ldp_flags); + + /* Check if lu_dirpage contains no entries. */ + if (end_dirent == NULL) + break; + + /* Enlarge the end entry lde_reclen from 0 to + * first entry of next lu_dirpage. */ + LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0); + end_dirent->lde_reclen = + cpu_to_le16((char *)(dp->ldp_entries) - + (char *)end_dirent); + } + + first->ldp_hash_end = hash_end; + first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE); + first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE); + + kunmap(pages[i]); + } + LASSERTF(lu_pgs == 0, "left = %d", lu_pgs); +} +#else +#define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0) +#endif /* PAGE_CACHE_SIZE > LU_PAGE_SIZE */ + +/* parameters for readdir page */ +struct readpage_param { + struct md_op_data *rp_mod; + __u64 rp_off; + int rp_hash64; + struct obd_export *rp_exp; + struct md_callback *rp_cb; +}; + +/** + * Read pages from server. + * + * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains + * a header lu_dirpage which describes the start/end hash, and whether this + * page is empty (contains no dir entry) or hash collide with next page. + * After client receives reply, several pages will be integrated into dir page + * in CFS_PAGE_SIZE (if CFS_PAGE_SIZE greater than LU_PAGE_SIZE), and the + * lu_dirpage for this integrated page will be adjusted. + **/ +static int mdc_read_page_remote(void *data, struct page *page0) +{ + struct readpage_param *rp = data; + struct page **page_pool; + struct page *page; + struct lu_dirpage *dp; + int rd_pgs = 0; /* number of pages read actually */ + int npages; + struct md_op_data *op_data = rp->rp_mod; + struct ptlrpc_request *req; + int max_pages = op_data->op_max_pages; + struct inode *inode; + struct lu_fid *fid; + int i; + int rc; + ENTRY; + + LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES); + if (op_data->op_mea1 != NULL) { + __u32 index = op_data->op_stripe_offset; + + inode = op_data->op_mea1->lsm_md_oinfo[index].lmo_root; + fid = &op_data->op_mea1->lsm_md_oinfo[index].lmo_fid; + } else { + inode = op_data->op_data; + fid = &op_data->op_fid1; + } + LASSERT(inode != NULL); + + OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages); + if (page_pool != NULL) { + page_pool[0] = page0; + } else { + page_pool = &page0; + max_pages = 1; + } + + for (npages = 1; npages < max_pages; npages++) { + page = page_cache_alloc_cold(inode->i_mapping); + if (page == NULL) + break; + page_pool[npages] = page; + } + + rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, op_data->op_capa1, + page_pool, npages, &req); + if (rc == 0) { + int lu_pgs; + + rd_pgs = (req->rq_bulk->bd_nob_transferred + + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; + lu_pgs = req->rq_bulk->bd_nob_transferred >> + LU_PAGE_SHIFT; + LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK)); + + CDEBUG(D_INODE, "read %d(%d)/%d pages\n", rd_pgs, lu_pgs, + op_data->op_npages); + + mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs); + + SetPageUptodate(page0); + } + + unlock_page(page0); + ptlrpc_req_finished(req); + CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages); + for (i = 1; i < npages; i++) { + unsigned long offset; + __u64 hash; + int ret; + + page = page_pool[i]; + + if (rc < 0 || i >= rd_pgs) { + page_cache_release(page); + continue; + } + + SetPageUptodate(page); + + dp = kmap(page); + hash = le64_to_cpu(dp->ldp_hash_start); + kunmap(page); + + offset = hash_x_index(hash, rp->rp_hash64); + + prefetchw(&page->flags); + ret = add_to_page_cache_lru(page, inode->i_mapping, offset, + GFP_KERNEL); + if (ret == 0) + unlock_page(page); + else + CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:" + " rc = %d\n", offset, ret); + page_cache_release(page); + } + + if (page_pool != &page0) + OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages); + + RETURN(rc); +} + +/** + * Read dir page from cache first, if it can not find it, read it from + * server and add into the cache. + */ +static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data, + struct md_callback *cb_op, struct page **ppage) +{ + struct lookup_intent it = { .it_op = IT_READDIR }; + struct page *page; + struct inode *dir = op_data->op_data; + struct address_space *mapping; + struct lu_dirpage *dp; + __u64 start = 0; + __u64 end = 0; + struct lustre_handle lockh; + struct ptlrpc_request *enq_req = NULL; + struct readpage_param rp_param; + int rc; + + ENTRY; + + *ppage = NULL; + + LASSERT(dir != NULL); + mapping = dir->i_mapping; + + rc = mdc_intent_lock(exp, op_data, &it, &enq_req, + cb_op->md_blocking_ast, 0); + if (enq_req != NULL) + ptlrpc_req_finished(enq_req); + + if (rc < 0) { + CERROR("%s: "DFID" lock enqueue fails: rc = %d\n", + exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc); + RETURN(rc); + } + + rc = 0; + mdc_set_lock_data(exp, &it.d.lustre.it_lock_handle, dir, NULL); + + rp_param.rp_off = op_data->op_hash_offset; + rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64; + page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end, + rp_param.rp_hash64); + if (IS_ERR(page)) { + CERROR("%s: dir page locate: "DFID" at "LPU64": rc %ld\n", + exp->exp_obd->obd_name, PFID(&op_data->op_fid1), + rp_param.rp_off, PTR_ERR(page)); + GOTO(out_unlock, rc = PTR_ERR(page)); + } else if (page != NULL) { + /* + * XXX nikita: not entirely correct handling of a corner case: + * suppose hash chain of entries with hash value HASH crosses + * border between pages P0 and P1. First both P0 and P1 are + * cached, seekdir() is called for some entry from the P0 part + * of the chain. Later P0 goes out of cache. telldir(HASH) + * happens and finds P1, as it starts with matching hash + * value. Remaining entries from P0 part of the chain are + * skipped. (Is that really a bug?) + * + * Possible solutions: 0. don't cache P1 is such case, handle + * it as an "overflow" page. 1. invalidate all pages at + * once. 2. use HASH|1 as an index for P1. + */ + GOTO(hash_collision, page); + } + + rp_param.rp_exp = exp; + rp_param.rp_mod = op_data; + page = read_cache_page(mapping, + hash_x_index(rp_param.rp_off, + rp_param.rp_hash64), + mdc_read_page_remote, &rp_param); + if (IS_ERR(page)) { + CERROR("%s: read cache page: "DFID" at "LPU64": rc %ld\n", + exp->exp_obd->obd_name, PFID(&op_data->op_fid1), + rp_param.rp_off, PTR_ERR(page)); + GOTO(out_unlock, rc = PTR_ERR(page)); + } + + wait_on_page_locked(page); + (void)kmap(page); + if (!PageUptodate(page)) { + CERROR("%s: page not updated: "DFID" at "LPU64": rc %d\n", + exp->exp_obd->obd_name, PFID(&op_data->op_fid1), + rp_param.rp_off, -5); + goto fail; + } + if (!PageChecked(page)) + SetPageChecked(page); + if (PageError(page)) { + CERROR("%s: page error: "DFID" at "LPU64": rc %d\n", + exp->exp_obd->obd_name, PFID(&op_data->op_fid1), + rp_param.rp_off, -5); + goto fail; + } + +hash_collision: + dp = page_address(page); + if (BITS_PER_LONG == 32 && rp_param.rp_hash64) { + start = le64_to_cpu(dp->ldp_hash_start) >> 32; + end = le64_to_cpu(dp->ldp_hash_end) >> 32; + rp_param.rp_off = op_data->op_hash_offset >> 32; + } else { + start = le64_to_cpu(dp->ldp_hash_start); + end = le64_to_cpu(dp->ldp_hash_end); + rp_param.rp_off = op_data->op_hash_offset; + } + if (end == start) { + LASSERT(start == rp_param.rp_off); + CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end); +#if BITS_PER_LONG == 32 + CWARN("Real page-wide hash collision at ["LPU64" "LPU64"] with " + "hash "LPU64"\n", le64_to_cpu(dp->ldp_hash_start), + le64_to_cpu(dp->ldp_hash_end), op_data->op_hash_offset); +#endif + + /* + * Fetch whole overflow chain... + * + * XXX not yet. + */ + goto fail; + } + *ppage = page; +out_unlock: + lockh.cookie = it.d.lustre.it_lock_handle; + ldlm_lock_decref(&lockh, it.d.lustre.it_lock_mode); + it.d.lustre.it_lock_handle = 0; + return rc; +fail: + kunmap(page); + mdc_release_page(page, 1); + rc = -EIO; + goto out_unlock; } +/** + * Read one directory entry from the cache. + */ +int mdc_read_entry(struct obd_export *exp, struct md_op_data *op_data, + struct md_callback *cb_op, struct lu_dirent **entp, + struct page **ppage) +{ + struct page *page = NULL; + struct lu_dirpage *dp; + struct lu_dirent *ent; + int rc = 0; + __u32 same_hash_count; + __u64 hash_offset = op_data->op_hash_offset; + ENTRY; + + CDEBUG(D_INFO, DFID " offset = "LPU64", flags %#x\n", + PFID(&op_data->op_fid1), op_data->op_hash_offset, + op_data->op_cli_flags); + + *ppage = NULL; + *entp = NULL; + + if (op_data->op_hash_offset == MDS_DIR_END_OFF) + RETURN(0); + + rc = mdc_read_page(exp, op_data, cb_op, &page); + if (rc != 0) + RETURN(rc); + + /* same_hash_count means how many entries with this + * hash value has been read */ + same_hash_count = op_data->op_same_hash_offset + 1; + dp = page_address(page); + for (ent = lu_dirent_start(dp); ent != NULL; + ent = lu_dirent_next(ent)) { + /* Skip dummy entry */ + if (le16_to_cpu(ent->lde_namelen) == 0) + continue; + + if (le64_to_cpu(ent->lde_hash) < + op_data->op_hash_offset) + continue; + + if (unlikely(le64_to_cpu(ent->lde_hash) == + op_data->op_hash_offset)) { + /* If it is not for next entry, which usually from + * ll_dir_entry_start, return this entry. */ + if (!(op_data->op_cli_flags & CLI_NEXT_ENTRY)) + break; + + /* Keep reading until all of entries being read are + * skipped. */ + if (same_hash_count > 0) { + same_hash_count--; + continue; + } + } + break; + } + + /* If it can not find entry in current page, try next page. */ + if (ent == NULL) { + if (le64_to_cpu(dp->ldp_hash_end) == MDS_DIR_END_OFF) { + op_data->op_same_hash_offset = 0; + mdc_release_page(page, + le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE); + RETURN(0); + } + + op_data->op_hash_offset = le64_to_cpu(dp->ldp_hash_end); + mdc_release_page(page, + le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE); + rc = mdc_read_page(exp, op_data, cb_op, &page); + if (rc != 0) + RETURN(rc); + + if (page != NULL) { + dp = page_address(page); + ent = lu_dirent_start(dp); + } + } + + /* If the next hash is the same as the current hash, increase + * the op_same_hash_offset to resolve the same hash conflict */ + if (ent != NULL && op_data->op_cli_flags & CLI_NEXT_ENTRY) { + if (unlikely(le64_to_cpu(ent->lde_hash) == hash_offset)) + op_data->op_same_hash_offset++; + else + op_data->op_same_hash_offset = 0; + } + + *ppage = page; + *entp = ent; + RETURN(rc); +} + +#else /* __KERNEL__ */ + +static struct page +*mdc_read_page_remote(struct obd_export *exp, const struct lmv_oinfo *lmo, + const __u64 hash, struct obd_capa *oc) +{ + struct ptlrpc_request *req = NULL; + struct page *page; + int rc; + + OBD_PAGE_ALLOC(page, 0); + if (page == NULL) + return ERR_PTR(-ENOMEM); + + rc = mdc_getpage(exp, &lmo->lmo_fid, hash, oc, &page, 1, &req); + if (req != NULL) + ptlrpc_req_finished(req); + + if (unlikely(rc)) { + OBD_PAGE_FREE(page); + return ERR_PTR(rc); + } + return page; +} + + +static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data, + struct md_callback *cb_op, + struct page **ppage) +{ + struct page *page; + struct lmv_oinfo *lmo; + int rc = 0; + + /* No local cache for liblustre, always read entry remotely */ + lmo = &op_data->op_mea1->lsm_md_oinfo[op_data->op_stripe_offset]; + page = mdc_read_page_remote(exp, lmo, op_data->op_hash_offset, + op_data->op_capa1); + if (IS_ERR(page)) + return PTR_ERR(page); + + *ppage = page; + + return rc; +} + +int mdc_read_entry(struct obd_export *exp, struct md_op_data *op_data, + struct md_callback *cb_op, struct lu_dirent **entp, + struct page **ppage) +{ + struct page *page = NULL; + struct lu_dirpage *dp; + struct lu_dirent *ent; + int rc; + ENTRY; + + rc = mdc_read_page(exp, op_data, cb_op, &page); + if (rc != 0) + RETURN(rc); + + dp = page_address(page); + if (le64_to_cpu(dp->ldp_hash_end) < op_data->op_hash_offset) + GOTO(out, *entp = NULL); + + for (ent = lu_dirent_start(dp); ent != NULL; + ent = lu_dirent_next(ent)) + if (le64_to_cpu(ent->lde_hash) >= op_data->op_hash_offset) + break; + *entp = ent; +out: + + OBD_PAGE_FREE(page); + RETURN(rc); +} + +#endif + static int mdc_statfs(const struct lu_env *env, struct obd_export *exp, struct obd_statfs *osfs, __u64 max_age, __u32 flags) @@ -1677,8 +2303,9 @@ out: static int mdc_ioc_changelog_send(struct obd_device *obd, struct ioc_changelog *icc) { - struct changelog_show *cs; - int rc; + struct changelog_show *cs; + struct task_struct *task; + int rc; /* Freed in mdc_changelog_send_thread */ OBD_ALLOC_PTR(cs); @@ -1695,16 +2322,20 @@ static int mdc_ioc_changelog_send(struct obd_device *obd, * New thread because we should return to user app before * writing into our pipe */ - rc = PTR_ERR(kthread_run(mdc_changelog_send_thread, cs, - "mdc_clg_send_thread")); - if (!IS_ERR_VALUE(rc)) { - CDEBUG(D_CHANGELOG, "start changelog thread\n"); - return 0; + task = kthread_run(mdc_changelog_send_thread, cs, + "mdc_clg_send_thread"); + if (IS_ERR(task)) { + rc = PTR_ERR(task); + CERROR("%s: cannot start changelog thread: rc = %d\n", + obd->obd_name, rc); + OBD_FREE_PTR(cs); + } else { + rc = 0; + CDEBUG(D_CHANGELOG, "%s: started changelog thread\n", + obd->obd_name); } - CERROR("Failed to start changelog thread: %d\n", rc); - OBD_FREE_PTR(cs); - return rc; + return rc; } static int mdc_ioc_hsm_ct_start(struct obd_export *exp, @@ -1797,7 +2428,7 @@ static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp, static int mdc_ioc_swap_layouts(struct obd_export *exp, struct md_op_data *op_data) { - CFS_LIST_HEAD(cancels); + struct list_head cancels = LIST_HEAD_INIT(cancels); struct ptlrpc_request *req; int rc, count; struct mdc_swap_layouts *msl, *payload; @@ -1812,9 +2443,11 @@ static int mdc_ioc_swap_layouts(struct obd_export *exp, * with the request RPC to avoid extra RPC round trips */ count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels, - LCK_CR, MDS_INODELOCK_LAYOUT); + LCK_EX, MDS_INODELOCK_LAYOUT | + MDS_INODELOCK_XATTR); count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels, - LCK_CR, MDS_INODELOCK_LAYOUT); + LCK_EX, MDS_INODELOCK_LAYOUT | + MDS_INODELOCK_XATTR); req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SWAP_LAYOUTS); @@ -1857,7 +2490,6 @@ static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len, struct obd_device *obd = exp->exp_obd; struct obd_ioctl_data *data = karg; struct obd_import *imp = obd->u.cli.cl_import; - struct llog_ctxt *ctxt; int rc; ENTRY; @@ -1910,22 +2542,6 @@ static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len, case IOC_OSC_SET_ACTIVE: rc = ptlrpc_set_import_active(imp, data->ioc_offset); GOTO(out, rc); - case OBD_IOC_PARSE: { - ctxt = llog_get_context(exp->exp_obd, LLOG_CONFIG_REPL_CTXT); - rc = class_config_parse_llog(NULL, ctxt, data->ioc_inlbuf1, - NULL); - llog_ctxt_put(ctxt); - GOTO(out, rc); - } -#ifdef __KERNEL__ - case OBD_IOC_LLOG_INFO: - case OBD_IOC_LLOG_PRINT: { - ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT); - rc = llog_ioctl(NULL, ctxt, cmd, data); - llog_ctxt_put(ctxt); - GOTO(out, rc); - } -#endif case OBD_IOC_POLL_QUOTACHECK: rc = mdc_quota_poll_check(exp, (struct if_quotacheck *)karg); GOTO(out, rc); @@ -2237,22 +2853,50 @@ int mdc_set_info_async(const struct lu_env *env, } int mdc_get_info(const struct lu_env *env, struct obd_export *exp, - __u32 keylen, void *key, __u32 *vallen, void *val, - struct lov_stripe_md *lsm) + __u32 keylen, void *key, __u32 *vallen, void *val, + struct lov_stripe_md *lsm) { - int rc = -EINVAL; + int rc = -EINVAL; - if (KEY_IS(KEY_MAX_EASIZE)) { - int mdsize, *max_easize; + if (KEY_IS(KEY_MAX_EASIZE)) { + int mdsize, *max_easize; - if (*vallen != sizeof(int)) - RETURN(-EINVAL); - mdsize = *(int*)val; - if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize) - exp->exp_obd->u.cli.cl_max_mds_easize = mdsize; - max_easize = val; - *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize; - RETURN(0); + if (*vallen != sizeof(int)) + RETURN(-EINVAL); + mdsize = *(int *)val; + if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize) + exp->exp_obd->u.cli.cl_max_mds_easize = mdsize; + max_easize = val; + *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize; + RETURN(0); + } else if (KEY_IS(KEY_DEFAULT_EASIZE)) { + int *default_easize; + + if (*vallen != sizeof(int)) + RETURN(-EINVAL); + default_easize = val; + *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize; + RETURN(0); + } else if (KEY_IS(KEY_MAX_COOKIESIZE)) { + int mdsize, *max_cookiesize; + + if (*vallen != sizeof(int)) + RETURN(-EINVAL); + mdsize = *(int *)val; + if (mdsize > exp->exp_obd->u.cli.cl_max_mds_cookiesize) + exp->exp_obd->u.cli.cl_max_mds_cookiesize = mdsize; + max_cookiesize = val; + *max_cookiesize = exp->exp_obd->u.cli.cl_max_mds_cookiesize; + RETURN(0); + } else if (KEY_IS(KEY_DEFAULT_COOKIESIZE)) { + int *default_cookiesize; + + if (*vallen != sizeof(int)) + RETURN(-EINVAL); + default_cookiesize = val; + *default_cookiesize = + exp->exp_obd->u.cli.cl_default_mds_cookiesize; + RETURN(0); } else if (KEY_IS(KEY_CONN_DATA)) { struct obd_import *imp = class_exp2cliimp(exp); struct obd_connect_data *data = val; @@ -2444,13 +3088,13 @@ static int mdc_import_event(struct obd_device *obd, struct obd_import *imp, RETURN(rc); } -int mdc_fid_alloc(struct obd_export *exp, struct lu_fid *fid, - struct md_op_data *op_data) +int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp, + struct lu_fid *fid, struct md_op_data *op_data) { - struct client_obd *cli = &exp->exp_obd->u.cli; - struct lu_client_seq *seq = cli->cl_seq; - ENTRY; - RETURN(seq_client_alloc_fid(NULL, seq, fid)); + struct client_obd *cli = &exp->exp_obd->u.cli; + struct lu_client_seq *seq = cli->cl_seq; + ENTRY; + RETURN(seq_client_alloc_fid(env, seq, fid)); } struct obd_uuid *mdc_get_uuid(struct obd_export *exp) { @@ -2463,18 +3107,18 @@ struct obd_uuid *mdc_get_uuid(struct obd_export *exp) { * recovery, non zero value will be return if the lock can be canceled, * or zero returned for not */ -static int mdc_cancel_for_recovery(struct ldlm_lock *lock) +static int mdc_cancel_weight(struct ldlm_lock *lock) { - if (lock->l_resource->lr_type != LDLM_IBITS) - RETURN(0); + if (lock->l_resource->lr_type != LDLM_IBITS) + RETURN(0); - /* FIXME: if we ever get into a situation where there are too many - * opened files with open locks on a single node, then we really - * should replay these open locks to reget it */ - if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN) - RETURN(0); + /* FIXME: if we ever get into a situation where there are too many + * opened files with open locks on a single node, then we really + * should replay these open locks to reget it */ + if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN) + RETURN(0); - RETURN(1); + RETURN(1); } static int mdc_resource_inode_free(struct ldlm_resource *res) @@ -2486,13 +3130,12 @@ static int mdc_resource_inode_free(struct ldlm_resource *res) } struct ldlm_valblock_ops inode_lvbo = { - lvbo_free: mdc_resource_inode_free + .lvbo_free = mdc_resource_inode_free }; static int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg) { struct client_obd *cli = &obd->u.cli; - struct lprocfs_static_vars lvars = { 0 }; int rc; ENTRY; @@ -2513,13 +3156,15 @@ static int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg) rc = client_obd_setup(obd, cfg); if (rc) GOTO(err_close_lock, rc); - lprocfs_mdc_init_vars(&lvars); - lprocfs_obd_setup(obd, lvars.obd_vars); +#ifdef LPROCFS + obd->obd_vars = lprocfs_mdc_obd_vars; + lprocfs_seq_obd_setup(obd); lprocfs_alloc_md_stats(obd, 0); - sptlrpc_lprocfs_cliobd_attach(obd); - ptlrpc_lprocfs_register_obd(obd); +#endif + sptlrpc_lprocfs_cliobd_attach(obd); + ptlrpc_lprocfs_register_obd(obd); - ns_register_cancel(obd->obd_namespace, mdc_cancel_for_recovery); + ns_register_cancel(obd->obd_namespace, mdc_cancel_weight); obd->obd_namespace->ns_lvbo = &inode_lvbo; @@ -2541,26 +3186,33 @@ err_rpc_lock: } /* Initialize the default and maximum LOV EA and cookie sizes. This allows - * us to make MDS RPCs with large enough reply buffers to hold the - * maximum-sized (= maximum striped) EA and cookie without having to - * calculate this (via a call into the LOV + OSCs) each time we make an RPC. */ + * us to make MDS RPCs with large enough reply buffers to hold a default + * sized EA and cookie without having to calculate this (via a call into the + * LOV + OSCs) each time we make an RPC. The maximum size is also tracked + * but not used to avoid wastefully vmalloc()'ing large reply buffers when + * a large number of stripes is possible. If a larger reply buffer is + * required it will be reallocated in the ptlrpc layer due to overflow. + */ static int mdc_init_ea_size(struct obd_export *exp, int easize, - int def_easize, int cookiesize) + int def_easize, int cookiesize, int def_cookiesize) { - struct obd_device *obd = exp->exp_obd; - struct client_obd *cli = &obd->u.cli; - ENTRY; + struct obd_device *obd = exp->exp_obd; + struct client_obd *cli = &obd->u.cli; + ENTRY; - if (cli->cl_max_mds_easize < easize) - cli->cl_max_mds_easize = easize; + if (cli->cl_max_mds_easize < easize) + cli->cl_max_mds_easize = easize; - if (cli->cl_default_mds_easize < def_easize) - cli->cl_default_mds_easize = def_easize; + if (cli->cl_default_mds_easize < def_easize) + cli->cl_default_mds_easize = def_easize; - if (cli->cl_max_mds_cookiesize < cookiesize) - cli->cl_max_mds_cookiesize = cookiesize; + if (cli->cl_max_mds_cookiesize < cookiesize) + cli->cl_max_mds_cookiesize = cookiesize; - RETURN(0); + if (cli->cl_default_mds_cookiesize < def_cookiesize) + cli->cl_default_mds_cookiesize = def_cookiesize; + + RETURN(0); } static int mdc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage) @@ -2640,19 +3292,9 @@ static int mdc_llog_finish(struct obd_device *obd, int count) static int mdc_process_config(struct obd_device *obd, obd_count len, void *buf) { struct lustre_cfg *lcfg = buf; - struct lprocfs_static_vars lvars = { 0 }; - int rc = 0; - - lprocfs_mdc_init_vars(&lvars); - switch (lcfg->lcfg_command) { - default: - rc = class_process_proc_param(PARAM_MDC, lvars.obd_vars, - lcfg, obd); - if (rc > 0) - rc = 0; - break; - } - return(rc); + int rc = class_process_proc_seq_param(PARAM_MDC, obd->obd_vars, + lcfg, obd); + return (rc > 0 ? 0: rc); } @@ -2794,8 +3436,8 @@ struct md_ops mdc_md_ops = { .m_setattr = mdc_setattr, .m_setxattr = mdc_setxattr, .m_getxattr = mdc_getxattr, - .m_fsync = mdc_fsync, - .m_readpage = mdc_readpage, + .m_fsync = mdc_fsync, + .m_read_entry = mdc_read_entry, .m_unlink = mdc_unlink, .m_cancel_unused = mdc_cancel_unused, .m_init_ea_size = mdc_init_ea_size, @@ -2814,13 +3456,11 @@ struct md_ops mdc_md_ops = { int __init mdc_init(void) { - int rc; - struct lprocfs_static_vars lvars = { 0 }; - lprocfs_mdc_init_vars(&lvars); - - rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, lvars.module_vars, - LUSTRE_MDC_NAME, NULL); - RETURN(rc); + return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL, +#ifndef HAVE_ONLY_PROCFS_SEQ + NULL, +#endif + LUSTRE_MDC_NAME, NULL); } #ifdef __KERNEL__