Whamcloud - gitweb
LU-12616 obclass: fix MDS start/stop race
[fs/lustre-release.git] / lustre / osc / osc_io.c
index b97750f..68cc041 100644 (file)
@@ -121,6 +121,9 @@ int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
        int result = 0;
        int brw_flags;
        unsigned int max_pages;
+       unsigned int ppc_bits; /* pages per chunk bits */
+       unsigned int ppc;
+       bool sync_queue = false;
 
        LASSERT(qin->pl_nr > 0);
 
@@ -129,6 +132,8 @@ int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
        osc = cl2osc(ios->cis_obj);
        cli = osc_cli(osc);
        max_pages = cli->cl_max_pages_per_rpc;
+       ppc_bits = cli->cl_chunkbits - PAGE_SHIFT;
+       ppc = 1 << ppc_bits;
 
        brw_flags = osc_io_srvlock(cl2osc_io(env, ios)) ? OBD_BRW_SRVLOCK : 0;
        brw_flags |= crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
@@ -185,12 +190,30 @@ int osc_io_submit(const struct lu_env *env, const struct cl_io_slice *ios,
                else /* async IO */
                        cl_page_list_del(env, qin, page);
 
-               if (++queued == max_pages) {
-                       queued = 0;
+               queued++;
+               if (queued == max_pages) {
+                       sync_queue = true;
+               } else if (crt == CRT_WRITE) {
+                       unsigned int chunks;
+                       unsigned int next_chunks;
+
+                       chunks = (queued + ppc - 1) >> ppc_bits;
+                       /* chunk number if add another page */
+                       next_chunks = (queued + ppc) >> ppc_bits;
+
+                       /* next page will excceed write chunk limit */
+                       if (chunks == osc_max_write_chunks(cli) &&
+                           next_chunks > chunks)
+                               sync_queue = true;
+               }
+
+               if (sync_queue) {
                        result = osc_queue_sync_pages(env, io, osc, &list,
                                                      brw_flags);
                        if (result < 0)
                                break;
+                       queued = 0;
+                       sync_queue = false;
                }
        }
 
@@ -334,24 +357,26 @@ int osc_io_iter_init(const struct lu_env *env, const struct cl_io_slice *ios)
 {
        struct osc_object *osc = cl2osc(ios->cis_obj);
        struct obd_import *imp = osc_cli(osc)->cl_import;
+       struct osc_io *oio = osc_env_io(env);
        int rc = -EIO;
 
        spin_lock(&imp->imp_lock);
        if (likely(!imp->imp_invalid)) {
-               struct osc_io *oio = osc_env_io(env);
-
                atomic_inc(&osc->oo_nr_ios);
                oio->oi_is_active = 1;
                rc = 0;
        }
        spin_unlock(&imp->imp_lock);
 
+       if (cfs_capable(CFS_CAP_SYS_RESOURCE))
+               oio->oi_cap_sys_resource = 1;
+
        return rc;
 }
 EXPORT_SYMBOL(osc_io_iter_init);
 
-int osc_io_write_iter_init(const struct lu_env *env,
-                          const struct cl_io_slice *ios)
+int osc_io_rw_iter_init(const struct lu_env *env,
+                       const struct cl_io_slice *ios)
 {
        struct cl_io *io = ios->cis_io;
        struct osc_io *oio = osc_env_io(env);
@@ -362,15 +387,15 @@ int osc_io_write_iter_init(const struct lu_env *env,
        if (cl_io_is_append(io))
                RETURN(osc_io_iter_init(env, ios));
 
-       npages = io->u.ci_rw.rw_range.cir_count >> PAGE_SHIFT;
-       if (io->u.ci_rw.rw_range.cir_pos & ~PAGE_MASK)
+       npages = io->u.ci_rw.crw_count >> PAGE_SHIFT;
+       if (io->u.ci_rw.crw_pos & ~PAGE_MASK)
                ++npages;
 
        oio->oi_lru_reserved = osc_lru_reserve(osc_cli(osc), npages);
 
        RETURN(osc_io_iter_init(env, ios));
 }
-EXPORT_SYMBOL(osc_io_write_iter_init);
+EXPORT_SYMBOL(osc_io_rw_iter_init);
 
 void osc_io_iter_fini(const struct lu_env *env,
                      const struct cl_io_slice *ios)
@@ -388,8 +413,8 @@ void osc_io_iter_fini(const struct lu_env *env,
 }
 EXPORT_SYMBOL(osc_io_iter_fini);
 
-void osc_io_write_iter_fini(const struct lu_env *env,
-                           const struct cl_io_slice *ios)
+void osc_io_rw_iter_fini(const struct lu_env *env,
+                        const struct cl_io_slice *ios)
 {
        struct osc_io *oio = osc_env_io(env);
        struct osc_object *osc = cl2osc(ios->cis_obj);
@@ -402,7 +427,7 @@ void osc_io_write_iter_fini(const struct lu_env *env,
 
        osc_io_iter_fini(env, ios);
 }
-EXPORT_SYMBOL(osc_io_write_iter_fini);
+EXPORT_SYMBOL(osc_io_rw_iter_fini);
 
 int osc_io_fault_start(const struct lu_env *env, const struct cl_io_slice *ios)
 {
@@ -489,10 +514,11 @@ static int osc_io_setattr_start(const struct lu_env *env,
         struct obdo             *oa     = &oio->oi_oa;
        struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
        __u64                    size   = io->u.ci_setattr.sa_attr.lvb_size;
-       unsigned int             ia_valid = io->u.ci_setattr.sa_valid;
-       int                      result = 0;
-       ENTRY;
+       unsigned int ia_avalid = io->u.ci_setattr.sa_avalid;
+       enum op_xvalid ia_xvalid = io->u.ci_setattr.sa_xvalid;
+       int result = 0;
 
+       ENTRY;
        /* truncate cache dirty pages first */
        if (cl_io_is_trunc(io))
                result = osc_cache_truncate_start(env, cl2osc(obj), size,
@@ -505,19 +531,20 @@ static int osc_io_setattr_start(const struct lu_env *env,
                        struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
                        unsigned int cl_valid = 0;
 
-                       if (ia_valid & ATTR_SIZE) {
-                               attr->cat_size = attr->cat_kms = size;
+                       if (ia_avalid & ATTR_SIZE) {
+                               attr->cat_size = size;
+                               attr->cat_kms = size;
                                cl_valid = (CAT_SIZE | CAT_KMS);
                        }
-                       if (ia_valid & ATTR_MTIME_SET) {
+                       if (ia_avalid & ATTR_MTIME_SET) {
                                attr->cat_mtime = lvb->lvb_mtime;
                                cl_valid |= CAT_MTIME;
                        }
-                       if (ia_valid & ATTR_ATIME_SET) {
+                       if (ia_avalid & ATTR_ATIME_SET) {
                                attr->cat_atime = lvb->lvb_atime;
                                cl_valid |= CAT_ATIME;
                        }
-                       if (ia_valid & ATTR_CTIME_SET) {
+                       if (ia_xvalid & OP_XVALID_CTIME_SET) {
                                attr->cat_ctime = lvb->lvb_ctime;
                                cl_valid |= CAT_CTIME;
                        }
@@ -534,45 +561,45 @@ static int osc_io_setattr_start(const struct lu_env *env,
                oa->o_layout = io->u.ci_setattr.sa_layout;
                oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP |
                        OBD_MD_FLOSTLAYOUT;
-               if (ia_valid & ATTR_CTIME) {
+               if (ia_avalid & ATTR_CTIME) {
                        oa->o_valid |= OBD_MD_FLCTIME;
                        oa->o_ctime = attr->cat_ctime;
                }
-               if (ia_valid & ATTR_ATIME) {
+               if (ia_avalid & ATTR_ATIME) {
                        oa->o_valid |= OBD_MD_FLATIME;
                        oa->o_atime = attr->cat_atime;
                }
-               if (ia_valid & ATTR_MTIME) {
+               if (ia_avalid & ATTR_MTIME) {
                        oa->o_valid |= OBD_MD_FLMTIME;
                        oa->o_mtime = attr->cat_mtime;
                }
-                if (ia_valid & ATTR_SIZE) {
-                        oa->o_size = size;
-                        oa->o_blocks = OBD_OBJECT_EOF;
-                        oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
-
-                        if (oio->oi_lockless) {
-                                oa->o_flags = OBD_FL_SRVLOCK;
-                                oa->o_valid |= OBD_MD_FLFLAGS;
-                        }
+               if (ia_avalid & ATTR_SIZE) {
+                       oa->o_size = size;
+                       oa->o_blocks = OBD_OBJECT_EOF;
+                       oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
+
+                       if (oio->oi_lockless) {
+                               oa->o_flags = OBD_FL_SRVLOCK;
+                               oa->o_valid |= OBD_MD_FLFLAGS;
+                       }
 
                        if (io->ci_layout_version > 0) {
                                /* verify layout version */
                                oa->o_valid |= OBD_MD_LAYOUT_VERSION;
                                oa->o_layout_version = io->ci_layout_version;
                        }
-                } else {
-                        LASSERT(oio->oi_lockless == 0);
-                }
+               } else {
+                       LASSERT(oio->oi_lockless == 0);
+               }
 
-               if (ia_valid & ATTR_ATTR_FLAG) {
+               if (ia_xvalid & OP_XVALID_FLAGS) {
                        oa->o_flags = io->u.ci_setattr.sa_attr_flags;
                        oa->o_valid |= OBD_MD_FLFLAGS;
                }
 
                init_completion(&cbargs->opc_sync);
 
-               if (ia_valid & ATTR_SIZE)
+               if (ia_avalid & ATTR_SIZE)
                        result = osc_punch_send(osc_export(cl2osc(obj)),
                                                oa, osc_async_upcall, cbargs);
                else
@@ -637,9 +664,9 @@ struct osc_data_version_args {
 
 static int
 osc_data_version_interpret(const struct lu_env *env, struct ptlrpc_request *req,
-                          void *arg, int rc)
+                          void *args, int rc)
 {
-       struct osc_data_version_args *dva = arg;
+       struct osc_data_version_args *dva = args;
        struct osc_io *oio = dva->dva_oio;
        const struct ost_body *body;
 
@@ -953,14 +980,14 @@ EXPORT_SYMBOL(osc_io_end);
 static const struct cl_io_operations osc_io_ops = {
        .op = {
                [CIT_READ] = {
-                       .cio_iter_init = osc_io_iter_init,
-                       .cio_iter_fini = osc_io_iter_fini,
+                       .cio_iter_init = osc_io_rw_iter_init,
+                       .cio_iter_fini = osc_io_rw_iter_fini,
                        .cio_start  = osc_io_read_start,
                        .cio_fini   = osc_io_fini
                },
                [CIT_WRITE] = {
-                       .cio_iter_init = osc_io_write_iter_init,
-                       .cio_iter_fini = osc_io_write_iter_fini,
+                       .cio_iter_init = osc_io_rw_iter_init,
+                       .cio_iter_fini = osc_io_rw_iter_fini,
                        .cio_start  = osc_io_write_start,
                        .cio_end    = osc_io_end,
                        .cio_fini   = osc_io_fini