#include <linux/obd_lov.h>
#include <linux/lustre_fsfilt.h>
#include <linux/lprocfs_status.h>
+#include <linux/lustre_gs.h>
#include "mds_internal.h"
return mfd;
}
-static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
+struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
{
ENTRY;
LASSERT(handle != NULL);
RETURN(class_handle2object(handle->cookie));
}
-static void mds_mfd_put(struct mds_file_data *mfd)
+void mds_mfd_put(struct mds_file_data *mfd)
{
CDEBUG(D_INFO, "PUTting mfd %p : new refcount %d\n", mfd,
atomic_read(&mfd->mfd_refcount) - 1);
mds_mfd_put(mfd);
}
-#ifdef IFILTERDATA_ACTUALLY_USED
/* Caller must hold mds->mds_epoch_sem */
static int mds_alloc_filterdata(struct inode *inode)
{
- LASSERT(inode->i_filterdata == NULL);
- OBD_ALLOC(inode->i_filterdata, sizeof(struct mds_filter_data));
- if (inode->i_filterdata == NULL)
+ LASSERT(LUSTRE_FILTERDATA(inode) == NULL);
+ OBD_ALLOC(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
+ if (LUSTRE_FILTERDATA(inode) == NULL)
return -ENOMEM;
LASSERT(igrab(inode) == inode);
return 0;
/* Caller must hold mds->mds_epoch_sem */
static void mds_free_filterdata(struct inode *inode)
{
- LASSERT(inode->i_filterdata != NULL);
- OBD_FREE(inode->i_filterdata, sizeof(struct mds_filter_data));
- inode->i_filterdata = NULL;
+ LASSERT(LUSTRE_FILTERDATA(inode) != NULL);
+ OBD_FREE(LUSTRE_FILTERDATA(inode), sizeof(struct mds_filter_data));
+ LUSTRE_FILTERDATA(inode) = NULL;
iput(inode);
}
-#endif /*IFILTERDATA_ACTUALLY_USED*/
/* Write access to a file: executors cause a negative count,
* writers a positive count. The semaphore is needed to perform
RETURN(-ETXTBSY);
}
-#ifdef IFILTERDATA_ACTUALLY_USED
if (MDS_FILTERDATA(inode) && MDS_FILTERDATA(inode)->io_epoch != 0) {
CDEBUG(D_INODE, "continuing MDS epoch "LPU64" for ino %lu/%u\n",
MDS_FILTERDATA(inode)->io_epoch, inode->i_ino,
goto out;
}
- if (inode->i_filterdata == NULL)
+ if (MDS_FILTERDATA(inode) == NULL)
mds_alloc_filterdata(inode);
- if (inode->i_filterdata == NULL) {
+ if (MDS_FILTERDATA(inode) == NULL) {
rc = -ENOMEM;
goto out;
}
-#endif /*IFILTERDATA_ACTUALLY_USED*/
- if (epoch > mds->mds_io_epoch)
+ if (epoch > mds->mds_io_epoch) {
mds->mds_io_epoch = epoch;
- else
+ CDEBUG(D_INODE, "repair MDS epoch "LPU64" for ino %lu/%u\n",
+ mds->mds_io_epoch, inode->i_ino, inode->i_generation);
+ } else {
mds->mds_io_epoch++;
-#ifdef IFILTERDATA_ACTUALLY_USED
+ CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
+ mds->mds_io_epoch, inode->i_ino, inode->i_generation);
+ }
MDS_FILTERDATA(inode)->io_epoch = mds->mds_io_epoch;
- CDEBUG(D_INODE, "starting MDS epoch "LPU64" for ino %lu/%u\n",
- mds->mds_io_epoch, inode->i_ino, inode->i_generation);
out:
-#endif /*IFILTERDATA_ACTUALLY_USED*/
if (rc == 0)
atomic_inc(&inode->i_writecount);
up(&mds->mds_epoch_sem);
if (!unlinking && !(body->valid & OBD_MD_FLSIZE))
GOTO(out, rc = EAGAIN);
#endif
-#ifdef IFILTERDATA_ACTUALLY_USED
mds_free_filterdata(inode);
-#endif
out:
up(&mds->mds_epoch_sem);
return rc;
/* This replaces the VFS dentry_open, it manages mfd and writecount */
static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
struct vfsmount *mnt, int flags,
- struct ptlrpc_request *req)
+ struct ptlrpc_request *req,
+ struct mds_update_record *rec)
{
struct mds_export_data *med = &req->rq_export->exp_mds_data;
struct mds_obd *mds = mds_req2mds(req);
struct mds_file_data *mfd;
struct mds_body *body;
- int error;
+ int rc = 0;
ENTRY;
mfd = mds_mfd_new();
if (mfd == NULL) {
CERROR("mds: out of memory\n");
- GOTO(cleanup_dentry, error = -ENOMEM);
+ GOTO(cleanup_dentry, rc = -ENOMEM);
}
body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
if (flags & FMODE_WRITE) {
/* FIXME: in recovery, need to pass old epoch here */
- error = mds_get_write_access(mds, dentry->d_inode, 0);
- if (error)
- GOTO(cleanup_mfd, error);
-#ifdef IFILTERDATA_ACTUALLY_USED
+ rc = mds_get_write_access(mds, dentry->d_inode, rec->ur_ioepoch);
+ if (rc)
+ GOTO(cleanup_mfd, rc);
body->io_epoch = MDS_FILTERDATA(dentry->d_inode)->io_epoch;
-#endif /*IFILTERDATA_ACTUALLY_USED*/
} else if (flags & FMODE_EXEC) {
- error = mds_deny_write_access(mds, dentry->d_inode);
- if (error)
- GOTO(cleanup_mfd, error);
+ rc = mds_deny_write_access(mds, dentry->d_inode);
+ if (rc)
+ GOTO(cleanup_mfd, rc);
}
dget(dentry);
- /* Mark the file as open to handle open-unlink. */
+ /* FIXME: invalidate update locks when first open for write comes in */
+ /* mark the file as open to handle open-unlink. */
DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
mds_orphan_open_inc(dentry->d_inode);
UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
mds_mfd_put(mfd);
body->handle.cookie = mfd->mfd_handle.h_cookie;
-
RETURN(mfd);
-
cleanup_mfd:
mds_mfd_put(mfd);
mds_mfd_destroy(mfd);
cleanup_dentry:
- return ERR_PTR(error);
+ return ERR_PTR(rc);
}
-static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
- struct lov_desc *desc)
+static inline void
+mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
+ struct lov_desc *desc)
{
int i;
+
for (i = 0; i < le32_to_cpu(lmm->lmm_stripe_count); i++) {
ids[le32_to_cpu(lmm->lmm_objects[i].l_ost_idx)] =
le64_to_cpu(lmm->lmm_objects[i].l_object_id);
}
}
-/* Must be called with i_sem held */
-static int mds_create_objects(struct ptlrpc_request *req, int offset,
- struct mds_update_record *rec,
- struct mds_obd *mds, struct obd_device *obd,
- struct dentry *dchild, void **handle,
- obd_id **ids)
+/* must be called with i_sem held */
+int
+mds_create_objects(struct obd_device *obd, struct ptlrpc_request *req,
+ int offset, struct mds_update_record *rec,
+ struct dentry *dchild, void **handle,
+ obd_id **ids)
{
- struct obdo *oa = NULL;
+ struct inode *inode = dchild->d_inode;
+ struct mds_obd *mds = &obd->u.mds;
struct obd_trans_info oti = { 0 };
- struct mds_body *body;
struct lov_stripe_md *lsm = NULL;
struct lov_mds_md *lmm = NULL;
- struct inode *inode = dchild->d_inode;
- void *lmm_buf;
int rc, lmm_bufsize, lmm_size;
+ struct obdo *oa = NULL;
+ struct mds_body *body;
+ void *lmm_buf;
ENTRY;
if (rec->ur_flags & MDS_OPEN_DELAY_CREATE ||
if (*ids == NULL)
RETURN(-ENOMEM);
oti.oti_objid = *ids;
-
+
/* replay case */
if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
LASSERT(id_ino(rec->ur_id2));
LASSERT(lmm);
if (*handle == NULL)
- *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
+ *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
if (IS_ERR(*handle)) {
rc = PTR_ERR(*handle);
*handle = NULL;
- GOTO(out_oa, rc);
+ RETURN(rc);
}
mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
lmm_bufsize = req->rq_repmsg->buflens[offset];
- LASSERT(lmm_buf);
+ LASSERT(lmm_buf != NULL);
LASSERT(lmm_bufsize >= lmm_size);
+
memcpy(lmm_buf, lmm, lmm_size);
rc = fsfilt_set_md(obd, inode, *handle, lmm,
lmm_size, EA_LOV);
oa = obdo_alloc();
if (oa == NULL)
- GOTO(out_ids, rc = -ENOMEM);
+ RETURN(-ENOMEM);
oa->o_mode = S_IFREG | 0600;
oa->o_id = inode->i_ino;
oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
oa->o_generation = inode->i_generation;
oa->o_uid = 0; /* must have 0 uid / gid on OST */
oa->o_gid = 0;
- oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
- OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
+ oa->o_valid = OBD_MD_FLID | OBD_MD_FLTYPE | OBD_MD_FLMODE |
+ OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
oa->o_size = 0;
- obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
- OBD_MD_FLCTIME);
+ obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
+ OBD_MD_FLMTIME | OBD_MD_FLCTIME);
if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
/* check if things like lfs setstripe are sending us the ea */
OBD_FREE(lmm, mds->mds_max_mdsize);
if (rc)
GOTO(out_oa, rc);
- }
+ }
+
LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
- rc = obd_create(mds->mds_dt_exp, oa, &lsm, &oti);
+ oti.oti_flags |= OBD_MODE_CROW;
+ rc = obd_create(mds->mds_dt_exp, oa, NULL, 0, &lsm, &oti);
+
if (rc) {
- int level = D_ERROR;
- if (rc == -ENOSPC)
- level = D_INODE;
- CDEBUG(level, "error creating objects for "
- "inode %lu: rc = %d\n",
+ CDEBUG((rc == -ENOSPC ? D_INODE : D_ERROR),
+ "error creating objects for "
+ "inode %lu: rc = %d\n",
inode->i_ino, rc);
if (rc > 0) {
CERROR("obd_create returned invalid "
} else {
rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
0, &lsm, rec->ur_eadata);
- if (rc) {
+ if (rc)
GOTO(out_oa, rc);
- }
+
lsm->lsm_object_id = oa->o_id;
lsm->lsm_object_gr = oa->o_gr;
}
if (inode->i_size) {
oa->o_size = inode->i_size;
- obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
- OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
- rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
+ obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLATIME |
+ OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLSIZE);
+
+ /* pack lustre id to oss */
+ *(obdo_id(oa)) = body->id1;
+ oa->o_valid |= OBD_MD_FLIFID;
+
+ rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti, NULL);
if (rc) {
CERROR("error setting attrs for inode %lu: rc %d\n",
inode->i_ino, rc);
rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
if (!id_ino(rec->ur_id2))
obd_free_memmd(mds->mds_dt_exp, &lsm);
- LASSERT(rc >= 0);
+ if (rc < 0) {
+ CERROR("cannot pack lsm, err = %d\n", rc);
+ GOTO(out_oa, rc);
+ }
+
lmm_size = rc;
body->eadatasize = rc;
if (IS_ERR(*handle)) {
rc = PTR_ERR(*handle);
*handle = NULL;
- GOTO(out_ids, rc);
+ GOTO(out_oa, rc);
}
rc = fsfilt_set_md(obd, inode, *handle, lmm,
memcpy(lmm_buf, lmm, lmm_size);
obd_free_diskmd(mds->mds_dt_exp, &lmm);
- out_oa:
+out_oa:
oti_free_cookies(&oti);
obdo_free(oa);
- out_ids:
+out_ids:
if (rc) {
OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
*ids = NULL;
}
- if(lsm)
+ if (lsm)
obd_free_memmd(mds->mds_dt_exp, &lsm);
RETURN(rc);
}
+int
+mds_destroy_object(struct obd_device *obd,
+ struct inode *inode, int async)
+{
+ struct mds_obd *mds = &obd->u.mds;
+ struct lov_mds_md *lmm = NULL;
+ int rc, lmm_size;
+ ENTRY;
+
+ LASSERT(inode != NULL);
+
+ if (inode->i_nlink != 0) {
+ CDEBUG(D_INODE, "attempt to destroy OSS object when "
+ "i_nlink == %d\n", (int)inode->i_nlink);
+ RETURN(0);
+ }
+
+ OBD_ALLOC(lmm, mds->mds_max_mdsize);
+ if (lmm == NULL)
+ RETURN(-ENOMEM);
+
+ lmm_size = mds->mds_max_mdsize;
+ rc = mds_get_md(obd, inode, lmm, &lmm_size, 1, 0);
+ if (rc < 0) {
+ CERROR("no stripe info for %lu/%lu inode\n",
+ (unsigned long)inode->i_ino,
+ (unsigned long)inode->i_generation);
+ GOTO(out_free_lmm, rc);
+ }
+
+ if (rc > 0) {
+ /* asynchronously unlink objecect on OSS */
+ rc = mds_unlink_object(mds, inode, lmm, lmm_size,
+ NULL, 0, async);
+ if (rc) {
+ CERROR("error unlinking object on OSS, "
+ "err %d\n", rc);
+ GOTO(out_free_lmm, rc);
+ }
+ } else {
+ CDEBUG(D_INODE, "no stripping info found for inode "
+ "%lu/%lu\n", (unsigned long)inode->i_ino,
+ (unsigned long)inode->i_generation);
+ }
+ EXIT;
+out_free_lmm:
+ OBD_FREE(lmm, mds->mds_max_mdsize);
+ return rc;
+}
+
static void reconstruct_open(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req,
struct lustre_handle *child_lockh)
struct mds_obd *mds = mds_req2mds(req);
struct mds_file_data *mfd;
struct obd_device *obd = req->rq_export->exp_obd;
- struct dentry *parent, *dchild;
+ struct dentry *parent = NULL, *dchild;
struct ldlm_reply *rep;
struct mds_body *body;
struct list_head *t;
return; /* error looking up parent or child */
}
- parent = mds_id2dentry(obd, rec->ur_id1, NULL);
- LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
- OLID4(rec->ur_id1), PTR_ERR(parent));
-
- dchild = ll_lookup_one_len(rec->ur_name, parent,
- rec->ur_namelen - 1);
- LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
- OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
+ /* first, we try to open the file by fid. by the time of this
+ * request, inode can be an orphan and parent can disappear */
+ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+ CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT|REPLAY)\n",
+ OLID4(rec->ur_id2));
+ dchild = mds_id2dentry(obd, rec->ur_id2, NULL);
+ } else if (rec->ur_namelen == 1) {
+ CDEBUG(D_HA, "OPEN by fid "DLID4" (RESENT)\n",
+ OLID4(rec->ur_id1));
+ dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
+ } else {
+ parent = mds_id2dentry(obd, rec->ur_id1, NULL);
+ LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n",
+ OLID4(rec->ur_id1), PTR_ERR(parent));
+ dchild = ll_lookup_one_len(rec->ur_name, parent,
+ rec->ur_namelen - 1);
+ LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
+ OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
+ }
if (!dchild->d_inode)
GOTO(out_dput, 0); /* child not present to open */
if (!(body->valid & OBD_MD_FLEASIZE))
body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
OBD_MD_FLATIME | OBD_MD_FLMTIME);
+ DEBUG_REQ(D_ERROR, req, "no capa for "DLID4, OLID4(&body->id1));
}
/* If we have -EEXIST as the status, and we were asked to create
mntget(mds->mds_vfsmnt);
CERROR("Re-opened file \n");
mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
- rec->ur_flags & ~MDS_OPEN_TRUNC, req);
+ rec->ur_flags & ~MDS_OPEN_TRUNC, req, rec);
if (!mfd) {
CERROR("mds: out of memory\n");
GOTO(out_dput, req->rq_status = -ENOMEM);
out_dput:
if (put_child)
l_dput(dchild);
- l_dput(parent);
+ if (parent)
+ l_dput(parent);
EXIT;
}
/* do NOT or the MAY_*'s, you'll get the weakest */
-static int accmode(int flags)
+int accmode(int flags)
{
int res = 0;
struct mds_body *body, int flags, void **handle,
struct mds_update_record *rec, struct ldlm_reply *rep)
{
- struct mds_obd *mds = mds_req2mds(req);
struct obd_device *obd = req->rq_export->exp_obd;
+ struct mds_obd *mds = mds_req2mds(req);
+ struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
struct mds_file_data *mfd = NULL;
- obd_id *ids = NULL; /* object IDs created */
+ obd_id *ids = NULL;
unsigned mode;
- int rc = 0;
+ int rc = 0, reply_off;
ENTRY;
/* atomically create objects if necessary */
down(&dchild->d_inode->i_sem);
mode = dchild->d_inode->i_mode;
+
if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) ||
(S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
RETURN(rc);
}
}
+
if (rec != NULL) {
- /* no EA: create objects */
if ((body->valid & OBD_MD_FLEASIZE) &&
(rec->ur_flags & MDS_OPEN_HAS_EA)) {
up(&dchild->d_inode->i_sem);
RETURN(-EEXIST);
}
+
if (!(body->valid & OBD_MD_FLEASIZE)) {
/* no EA: create objects */
- rc = mds_create_objects(req, 2, rec, mds, obd,
+ rc = mds_create_objects(obd, req, 2, rec,
dchild, handle, &ids);
if (rc) {
- CERROR("mds_create_objects: rc = %d\n", rc);
+ CERROR("mds_create_object: rc = %d\n", rc);
up(&dchild->d_inode->i_sem);
RETURN(rc);
}
}
+
if (S_ISREG(dchild->d_inode->i_mode) &&
(body->valid & OBD_MD_FLEASIZE)) {
rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
}
}
}
- rc = mds_pack_acl(obd, req->rq_repmsg, 3, body, dchild->d_inode);
+
+ reply_off = 3;
+ rc = mds_pack_acl(req, reply_off, body, dchild->d_inode);
+ reply_off += 2;
+
if (rc < 0) {
- CERROR("mds_pack_acl: rc = %d\n", rc);
+ CERROR("pack posix acl: rc = %d\n", rc);
up(&dchild->d_inode->i_sem);
RETURN(rc);
}
+ rc = mds_pack_gskey(obd, req->rq_repmsg, &reply_off, body,
+ dchild->d_inode);
+ if (rc < 0) {
+ CERROR("mds_pack_gskey: rc = %d\n", rc);
+ up(&dchild->d_inode->i_sem);
+ RETURN(rc);
+ }
+
+ if (S_ISREG(mode)) {
+ struct lustre_capa capa = {
+ .lc_uid = rec->ur_uc.luc_uid,
+ .lc_op = capa_op(rec->ur_flags),
+ .lc_ino = dchild->d_inode->i_ino,
+ .lc_igen = dchild->d_inode->i_generation,
+ .lc_mdsid = mds->mds_num,
+ };
+
+ rc = mds_pack_capa(obd, med, NULL, &capa, req,
+ &reply_off, body);
+ if (rc < 0) {
+ CERROR("mds_pack_capa: rc = %d\n", rc);
+ up(&dchild->d_inode->i_sem);
+ RETURN(rc);
+ }
+ } else {
+ reply_off++;
+ }
+
/* If the inode has no EA data, then MDSs hold size, mtime */
- if (S_ISREG(dchild->d_inode->i_mode) &&
- !(body->valid & OBD_MD_FLEASIZE)) {
+ if (S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) {
body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
OBD_MD_FLATIME | OBD_MD_FLMTIME);
}
+
up(&dchild->d_inode->i_sem);
intent_set_disposition(rep, DISP_OPEN_OPEN);
- mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req);
+ mfd = mds_dentry_open(dchild, mds->mds_vfsmnt, flags, req, rec);
if (IS_ERR(mfd))
RETURN(PTR_ERR(mfd));
CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
mfd->mfd_handle.h_cookie);
+
if (ids != NULL) {
mds_dt_update_objids(obd, ids);
OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
}
- //if (rc)
- // mds_mfd_destroy(mfd);
+
RETURN(rc);
}
}
l_dput(dchild);
- /*
- * we didn't find it in PENDING so it isn't an orphan. See if it was a
- * regular inode that was previously created.
- */
+ /* we didn't find it in PENDING so it isn't an orphan. See if it was a
+ * regular inode that was previously created. */
dchild = mds_id2dentry(req2obd(req), id, NULL);
if (IS_ERR(dchild))
RETURN(PTR_ERR(dchild));
{
struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
struct lustre_handle lockh;
- int lock_flags = 0;
+ int lock_flags = LDLM_FL_ATOMIC_CB;
int rc;
ENTRY;
struct mea *mea = NULL;
int mea_size, update_mode;
int child_mode = LCK_PR;
- /* Always returning LOOKUP lock if open succesful to guard
- dentry on client. */
- ldlm_policy_data_t policy = {.l_inodebits={MDS_INODELOCK_LOOKUP}};
- struct ldlm_res_id child_res_id = { .name = {0}};
- int lock_flags = 0;
+ struct lustre_id sid;
+ __u64 fid = 0;
ENTRY;
DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
LASSERT(id_ino(rec->ur_id2));
- rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
- rec, rep);
+ rc = mds_open_by_id(req, rec->ur_id2, body,
+ rec->ur_flags, rec, rep);
if (rc != -ENOENT) {
mds_body_do_reverse_map(med, body);
RETURN(rc);
if (rc)
GOTO(cleanup, rc);
- if (mea != NULL) {
+ if (mea != NULL && mea->mea_count) {
/*
* dir is already splitted, check is requested filename should *
* live at this MDS or at another one.
if (dchild->d_flags & DCACHE_CROSS_REF) {
CDEBUG(D_OTHER, "cross reference: "DLID4"\n", OLID4(rec->ur_id1));
+ LASSERT(rec->ur_namelen > 1);
/* we're gonna acquire LOOKUP lock on the child,
* but we have already locked parent and our order
* may conflict with enqueue_order_locks(). so,
* drop parent lock and acquire both the locks in
* common order. bug 6190 */
+#ifdef S_PDIROPS
+ if (parent_lockh[1].cookie != 0)
+ ldlm_lock_decref(parent_lockh + 1, update_mode);
+#endif
ldlm_lock_decref(parent_lockh, parent_mode);
l_dput(dchild);
l_dput(dparent);
if (rc)
GOTO(cleanup, rc);
+#ifdef S_PDIROPS
+ if (parent_lockh[1].cookie != 0)
+ ldlm_lock_decref(parent_lockh + 1, update_mode);
+#endif
+ ldlm_lock_decref(parent_lockh, LCK_PR);
+
if (dchild->d_inode || !(dchild->d_flags & DCACHE_CROSS_REF)) {
- /* wow! someone unlink and create new one yet */
- CDEBUG(D_OTHER, "nice race, repeat lookup\n");
- ldlm_lock_decref(parent_lockh, parent_mode);
- ldlm_lock_decref(child_lockh, LCK_PR);
+ CDEBUG(D_OTHER, "race: name changed (%p)\n",
+ dchild->d_inode);
+ if (dchild->d_inode)
+ ldlm_lock_decref(child_lockh, LCK_PR);
l_dput(dchild);
l_dput(dparent);
- LASSERT(rec->ur_namelen > 1);
GOTO(restart, rc);
}
intent_set_disposition(rep, DISP_LOOKUP_POS);
intent_set_disposition(rep, DISP_LOOKUP_EXECD);
-#ifdef S_PDIROPS
- if (parent_lockh[1].cookie != 0)
- ldlm_lock_decref(parent_lockh + 1, update_mode);
-#endif
- ldlm_lock_decref(parent_lockh, parent_mode);
if (mea)
OBD_FREE(mea, mea_size);
l_dput(dchild);
struct iattr iattr;
struct inode *inode;
+ /* get parent id: ldlm lock on the parent protects ea */
+ rc = mds_read_inode_sid(obd, dparent->d_inode, &sid);
+ if (rc) {
+ CERROR("can't read parent inode id. ino(%lu) rc(%d)\n",
+ dparent->d_inode->i_ino, rc);
+ GOTO(cleanup, rc);
+ }
+
ino = (unsigned long)id_ino(rec->ur_id2);
rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
handle = NULL;
GOTO(cleanup, rc);
}
+ if (id_fid(rec->ur_id2))
+ fid = id_fid(rec->ur_id2);
+ else
+ fid = mds_alloc_fid(obd);
+
dchild->d_fsdata = (void *) &dp;
dp.p_ptr = req;
dp.p_inum = ino;
+
+ dp.p_fid = fid;
+ dp.p_group = mds->mds_num;
rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
if (dchild->d_fsdata == (void *)(unsigned long)ino)
CDEBUG(D_INODE, "error during create: %d\n", rc);
GOTO(cleanup, rc);
}
+
inode = dchild->d_inode;
+
if (ino) {
LASSERT(ino == inode->i_ino);
MD_COUNTER_INCREMENT(obd, create);
}
- down(&dchild->d_inode->i_sem);
- if (ino) {
- rc = mds_update_inode_sid(obd, dchild->d_inode,
- handle, rec->ur_id2);
- if (rc) {
- CERROR("mds_update_inode_sid() failed, "
- "rc = %d\n", rc);
- }
- id_assign_fid(&body->id1, rec->ur_id2);
+ mds_inode2id(obd, &body->id1, inode, fid);
+ mds_update_inode_ids(obd, dchild->d_inode, handle, &body->id1, &sid);
- /*
- * make sure, that fid is up-to-date.
- */
- mds_set_last_fid(obd, id_fid(rec->ur_id2));
- } else {
- rc = mds_alloc_inode_sid(obd, dchild->d_inode,
- handle, &body->id1);
+ if ((rec->ur_flags & MDS_OPEN_HAS_KEY)) {
+ rc = mds_set_gskey(obd, handle, dchild->d_inode,
+ rec->ur_ea2data, rec->ur_ea2datalen,
+ ATTR_KEY | ATTR_MAC);
if (rc) {
- CERROR("mds_alloc_inode_sid() failed, "
- "rc = %d\n", rc);
+ CERROR("error in set gs key rc %d\n", rc);
}
}
- up(&dchild->d_inode->i_sem);
-
- if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
- rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
- dchild->d_inode, handle,
- req->rq_export->exp_sync);
- handle = NULL;
- }
+ rc = fsfilt_commit(obd, dchild->d_inode->i_sb, dchild->d_inode, handle, 0);
+ handle = NULL;
acc_mode = 0; /* Don't check for permissions */
}
mds_pack_inode2body(obd, body, dchild->d_inode, 1);
if ((rec->ur_flags & MDS_OPEN_DIRECTORY) &&
!S_ISDIR(dchild->d_inode->i_mode))
GOTO(cleanup, rc = -ENOTDIR);
-
+
+ /* check permission even it's special files */
+ if (S_ISCHR(dchild->d_inode->i_mode) ||
+ S_ISBLK(dchild->d_inode->i_mode) ||
+ S_ISFIFO(dchild->d_inode->i_mode) ||
+ S_ISSOCK(dchild->d_inode->i_mode)) {
+ rc = ll_permission(dchild->d_inode, acc_mode, NULL);
+ if (rc != 0)
+ GOTO(cleanup, rc);
+ }
+
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_CREATE)) {
obd_fail_loc = OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE;
GOTO(cleanup, rc = -EAGAIN);
}
-
- /* Obtain OPEN lock as well */
- policy.l_inodebits.bits |= MDS_INODELOCK_OPEN;
+#if 0
/* We cannot use acc_mode here, because it is zeroed in case of
creating a file, so we get wrong lockmode */
if (accmode(rec->ur_flags) & MAY_WRITE)
else
child_mode = LCK_CR;
+ /* Always returning LOOKUP lock if open succesful to guard
+ * dentry on client. In case of replay we do not get a lock
+ * assuming that the caller has it already */
if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
- struct lustre_id sid;
-
- down(&dchild->d_inode->i_sem);
- rc = mds_read_inode_sid(obd, dchild->d_inode, &sid);
- up(&dchild->d_inode->i_sem);
- if (rc) {
- CERROR("Can't read inode self id, "
- "inode %lu, rc %d\n",
- dchild->d_inode->i_ino, rc);
- GOTO(cleanup, rc);
- }
+ struct ldlm_res_id child_res_id = { .name = {0}};
+ ldlm_policy_data_t policy;
+ int lock_flags = LDLM_FL_ATOMIC_CB;
- /* In case of replay we do not get a lock assuming that the
- caller has it already */
- child_res_id.name[0] = id_fid(&sid);
- child_res_id.name[1] = id_group(&sid);
+ /* LOOKUP lock will protect dentry on client -bzzz */
+ policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP |
+ MDS_INODELOCK_OPEN;
+
+ child_res_id.name[0] = id_fid(&body->id1);
+ child_res_id.name[1] = id_group(&body->id1);
rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
child_res_id, LDLM_IBITS, &policy,
cleanup_phase = 3;
}
+#else
+/* re-enable test 24n in sanity.sh: it needs LOOKUP lock on open */
+#endif
/* Step 5: mds_open it */
rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
rec, rep);
+ if (rc)
+ GOTO(cleanup, rc);
+
+ /* reintegration case */
+ if ((rec->ur_flags & MDS_REINT_REQ)) {
+ rc = mds_fidmap_add(obd, &body->id1);
+ if (rc < 0) {
+ CERROR("can't create fid->ino mapping, err %d\n",
+ rc);
+ } else {
+ rc = 0;
+ }
+ }
+
+ /* if this is a writer, we have to invalidate client's
+ * update locks in order to make sure they don't use
+ * isize/iblocks from mds anymore.
+ * FIXME: can cause a deadlock, use mds_get_parent_child_locked()
+ * XXX: optimization is to do this for first writer only */
+ if (accmode(rec->ur_flags) & MAY_WRITE) {
+ struct ldlm_res_id child_res_id = { .name = {0}};
+ ldlm_policy_data_t sz_policy;
+ struct lustre_handle sz_lockh;
+ int lock_flags = LDLM_FL_ATOMIC_CB;
+
+ child_res_id.name[0] = id_fid(&body->id1);
+ child_res_id.name[1] = id_group(&body->id1);
+ sz_policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
+
+ rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
+ child_res_id, LDLM_IBITS, &sz_policy,
+ LCK_PW, &lock_flags, mds_blocking_ast,
+ ldlm_completion_ast, NULL, NULL, NULL,
+ 0, NULL, &sz_lockh);
+ if (rc == ELDLM_OK)
+ ldlm_lock_decref(&sz_lockh, LCK_PW);
+ else
+ CERROR("can't invalidate client's update locks\n");
+ }
EXIT;
cleanup:
} else if (created) {
mds_lock_new_child(obd, dchild->d_inode, NULL);
}
+ /* audit stuff for OPEN */
+ if (offset == 3) {
+ mds_audit(req, dchild, rec->ur_name,
+ rec->ur_namelen - 1, AUDIT_OPEN, rc);
+ }
+
l_dput(dchild);
case 1:
if (dparent == NULL)
reply_body = lustre_msg_buf(req->rq_repmsg, 0,
sizeof(*reply_body));
+ if (request_body && (request_body->valid & OBD_MD_FLSIZE)) {
+ /* we set i_size/i_blocks here, nobody will see
+ * them until all write references are dropped.
+ * btw, we hold one reference */
+ LASSERT(mfd->mfd_mode & FMODE_WRITE);
+ LASSERT(request_body->valid & OBD_MD_FLEPOCH);
+ LASSERT(MDS_FILTERDATA(inode));
+ if (MDS_FILTERDATA(inode)->io_epoch != request_body->io_epoch)
+ CDEBUG(D_ERROR, "try to update attr. for old epoch "
+ LPD64" while current "LPD64"\n",
+ MDS_FILTERDATA(inode)->io_epoch,
+ request_body->io_epoch);
+ i_size_write(inode, request_body->size);
+ inode->i_blocks = request_body->blocks;
+ LTIME_S(inode->i_mtime) = (request_body->mtime);
+
+ LTIME_S(iattr.ia_mtime) = request_body->mtime;
+ iattr.ia_size = inode->i_size;
+ iattr.ia_valid |= ATTR_SIZE|ATTR_MTIME;
+ mds_inode_unset_attrs_old(inode);
+ }
+
idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode,
idname, inode->i_nlink, mds_orphan_open_count(inode));
int stripe_count = 0;
LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
- if (obd->obd_recovering) {
- CDEBUG(D_HA, "not remove orphan %s until recovery"
- " is over\n", idname);
- GOTO(out, rc);
- }
-
CDEBUG(D_HA, "destroying orphan object %s\n", idname);
if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
idlen);
if (IS_ERR(pending_child))
GOTO(cleanup, rc = PTR_ERR(pending_child));
- LASSERT(pending_child->d_inode != NULL);
+ if (pending_child->d_inode == NULL) {
+ CERROR("orphan %s has been removed\n", idname);
+ GOTO(cleanup, rc = 0);
+ }
cleanup_phase = 2; /* dput(pending_child) when finished */
if (S_ISDIR(pending_child->d_inode->i_mode)) {
req->rq_repmsg->buflens[2], &lcl) > 0) {
reply_body->valid |= OBD_MD_FLCOOKIE;
}
+
+ rc = mds_destroy_object(obd, inode, 1);
+ if (rc) {
+ CERROR("cannot destroy OSS object on close, err %d\n",
+ rc);
+ rc = 0;
+ }
goto out; /* Don't bother updating attrs on unlinked inode */
+ } else if ((mfd->mfd_mode & FMODE_WRITE) && rc == 0) {
+ /* last writer closed file - let's update i_size/i_blocks */
+ mds_validate_size(obd, inode, request_body, &iattr);
}
#if 0
}
if (iattr.ia_valid != 0) {
- handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
+ if (handle == NULL)
+ handle = fsfilt_start(obd,inode,FSFILT_OP_SETATTR,NULL);
if (IS_ERR(handle))
GOTO(cleanup, rc = PTR_ERR(handle));
rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
/* If other clients have this file open for write, rc will be > 0 */
if (rc > 0)
rc = 0;
+ if (!obd->obd_recovering && mds_inode_has_old_attrs(inode)
+ && !mds_inode_is_orphan(inode)
+ && atomic_read(&inode->i_writecount) == 0
+ && inode->i_nlink != 0) {
+ CERROR("leave inode %lu/%u with old attributes (nlink = %d)\n",
+ inode->i_ino, inode->i_generation, inode->i_nlink);
+ }
l_dput(mfd->mfd_dentry);
mds_mfd_destroy(mfd);
RETURN(rc);
}
+static int mds_extent_lock_callback(struct ldlm_lock *lock,
+ struct ldlm_lock_desc *new, void *data,
+ int flag)
+{
+ struct lustre_handle lockh = { 0 };
+ int rc;
+ ENTRY;
+
+ switch (flag) {
+ case LDLM_CB_BLOCKING:
+ ldlm_lock2handle(lock, &lockh);
+ rc = ldlm_cli_cancel(&lockh);
+ if (rc != ELDLM_OK)
+ CERROR("ldlm_cli_cancel failed: %d\n", rc);
+ break;
+ case LDLM_CB_CANCELING: {
+ break;
+ }
+ default:
+ LBUG();
+ }
+
+ RETURN(0);
+}
+__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
+__u64 lov_merge_blocks(struct lov_stripe_md *lsm);
+__u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time);
+
+int mds_validate_size(struct obd_device *obd, struct inode *inode,
+ struct mds_body *body, struct iattr *iattr)
+{
+ ldlm_policy_data_t policy = { .l_extent = { 0, OBD_OBJECT_EOF } };
+ struct lustre_handle lockh = { 0 };
+ struct lov_stripe_md *lsm = NULL;
+ int rc, len, flags;
+ void *lmm = NULL;
+ ENTRY;
+
+ /* we update i_size/i_blocks only for regular files */
+ if (!S_ISREG(inode->i_mode))
+ RETURN(0);
+
+ /* we shouldn't fetch size from OSTes during recovery - deadlock */
+ if (obd->obd_recovering) {
+ CERROR("size-on-mds has no support on OST yet\n");
+ RETURN(0);
+ }
+
+ /* if nobody modified attrs. we're lucky */
+ if (!mds_inode_has_old_attrs(inode))
+ RETURN(0);
+
+ /* 1: client didn't send actual i_size/i_blocks
+ * 2: we seem to be last writer
+ * 3: the file doesn't look like to be disappeared
+ * conclusion: we're gonna fetch them from OSSes */
+
+ down(&inode->i_sem);
+ len = fsfilt_get_md(obd, inode, NULL, 0, EA_LOV);
+ up(&inode->i_sem);
+
+ if (len < 0) {
+ CERROR("error getting inode %lu MD: %d\n", inode->i_ino, len);
+ GOTO(cleanup, rc = len);
+ } else if (len == 0) {
+ CDEBUG(D_INODE, "no LOV in inode %lu\n", inode->i_ino);
+ GOTO(cleanup, rc = 0);
+ }
+
+ OBD_ALLOC(lmm, len);
+ if (lmm == NULL) {
+ CERROR("can't allocate memory\n");
+ GOTO(cleanup, rc = -ENOMEM);
+ }
+
+ down(&inode->i_sem);
+ rc = fsfilt_get_md(obd, inode, lmm, len, EA_LOV);
+ up(&inode->i_sem);
+
+ if (rc < 0) {
+ CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
+ GOTO(cleanup, rc);
+ }
+
+ rc = obd_unpackmd(obd->u.mds.mds_dt_exp, &lsm, lmm, len);
+ if (rc < 0) {
+ CERROR("error getting inode %lu MD: %d\n", inode->i_ino, rc);
+ GOTO(cleanup, rc);
+ }
+
+ CDEBUG(D_DLMTRACE, "Glimpsing inode %lu\n", inode->i_ino);
+
+ flags = LDLM_FL_HAS_INTENT;
+ rc = obd_enqueue(obd->u.mds.mds_dt_exp, lsm, LDLM_EXTENT, &policy,
+ LCK_PR, &flags, mds_extent_lock_callback,
+ ldlm_completion_ast, NULL, NULL,
+ sizeof(struct ost_lvb), lustre_swab_ost_lvb, &lockh);
+ if (rc == -ENOENT) {
+ /* while we were enqueueing lock on OST, another thread
+ * unlinked the file and started OST object destoying.
+ * it's safe to return 0 here */
+ GOTO(cleanup, rc = 0);
+ } else if (rc != 0) {
+ CERROR("obd_enqueue returned rc %d, returning -EIO\n", rc);
+ GOTO(cleanup, rc);
+ }
+
+ CDEBUG(D_INODE, "LOV reports "LPD64"/%lu for "DLID4" [%s%s%s]\n",
+ inode->i_size, inode->i_blocks, OLID4(&body->id1),
+ atomic_read(&inode->i_writecount) > 1 ? "U" : "",
+ mds_inode_has_old_attrs(inode) ? "D" : "",
+ mds_inode_is_orphan(inode) ? "O" : "");
+
+ i_size_write(inode, lov_merge_size(lsm, 0));
+ inode->i_blocks = lov_merge_blocks(lsm);
+ LTIME_S(inode->i_mtime) = lov_merge_mtime(lsm, LTIME_S(inode->i_mtime));
+ iattr->ia_size = inode->i_size;
+ LTIME_S(iattr->ia_mtime) = LTIME_S(inode->i_mtime);
+ iattr->ia_valid |= ATTR_SIZE | ATTR_MTIME;
+
+ DOWN_WRITE_I_ALLOC_SEM(inode);
+ mds_inode_unset_attrs_old(inode);
+ UP_WRITE_I_ALLOC_SEM(inode);
+
+ obd_cancel(obd->u.mds.mds_dt_exp, lsm, LCK_PR, &lockh);
+
+cleanup:
+ if (lsm != NULL)
+ obd_free_memmd(obd->u.mds.mds_dt_exp, &lsm);
+ if (lmm != NULL)
+ OBD_FREE(lmm, len);
+ RETURN(rc);
+}
+
int mds_close(struct ptlrpc_request *req, int offset)
{
struct mds_export_data *med = &req->rq_export->exp_mds_data;
}
if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
- DEBUG_REQ(D_HA, req, "close replay\n");
+ DEBUG_REQ(D_HA, req, "close replay");
memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
req->rq_repmsg->buflens[2]);
}
inode = mfd->mfd_dentry->d_inode;
+
/* child i_alloc_sem protects orphan_dec_test && is_orphan race */
DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
+
+ if (body->flags & MDS_BFLAG_DIRTY_EPOCH) {
+ /* the client modified data through the handle
+ * we need to care about attrs. -bzzz */
+ mds_inode_set_attrs_old(inode);
+ }
+
if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
struct mds_body *rep_body;
RETURN(0);
}
+
int mds_done_writing(struct ptlrpc_request *req, int offset)
{
struct mds_body *body;
git://git.whamcloud.com / fs / lustre-release.git / blobdiff