static int vvp_io_prepare_write(const struct lu_env *env,
const struct cl_io_slice *ios,
const struct cl_page_slice *slice,
unsigned from, unsigned to)
{
struct cl_object *obj = slice->cpl_obj;
struct ccc_page *cp = cl2ccc_page(slice);
struct cl_page *pg = slice->cpl_page;
struct page *vmpage = cp->cpg_page;
int result;
ENTRY;
LINVRNT(cl_page_is_vmlocked(env, pg));
LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
result = 0;
CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
if (!PageUptodate(vmpage)) {
/*
* We're completely overwriting an existing page, so _don't_
* set it up to date until commit_write
*/
if (from == 0 && to == PAGE_CACHE_SIZE) {
CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
POISON_PAGE(page, 0x11);
} else
result = vvp_io_prepare_partial(env, ios->cis_io, obj,
pg, cp, from, to);
} else
CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
RETURN(result);
}
static int osd_oi_iam_refresh(struct osd_thread_info *oti, struct osd_oi *oi,
const struct dt_rec *rec, const struct dt_key *key,
struct thandle *th, bool insert)
{
struct iam_container *bag;
struct iam_path_descr *ipd;
struct osd_thandle *oh;
int rc;
ENTRY;
LASSERT(oi);
LASSERT(oi->oi_inode);
ll_vfs_dq_init(oi->oi_inode);
bag = &oi->oi_dir.od_container;
ipd = osd_idx_ipd_get(oti->oti_env, bag);
if (unlikely(ipd == NULL))
RETURN(-ENOMEM);
oh = container_of0(th, struct osd_thandle, ot_super);
LASSERT(oh->ot_handle != NULL);
LASSERT(oh->ot_handle->h_transaction != NULL);
if (insert)
rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
(const struct iam_rec *)rec, ipd);
else
rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)key,
(const struct iam_rec *)rec, ipd);
osd_ipd_put(oti->oti_env, bag, ipd);
LINVRNT(osd_invariant(obj));
RETURN(rc);
}
static int osd_oi_iam_delete(struct osd_thread_info *oti, struct osd_oi *oi,
const struct dt_key *key, handle_t *th)
{
struct iam_container *bag;
struct iam_path_descr *ipd;
int rc;
ENTRY;
LASSERT(oi);
LASSERT(oi->oi_inode);
ll_vfs_dq_init(oi->oi_inode);
bag = &oi->oi_dir.od_container;
ipd = osd_idx_ipd_get(oti->oti_env, bag);
if (unlikely(ipd == NULL))
RETURN(-ENOMEM);
LASSERT(th != NULL);
LASSERT(th->h_transaction != NULL);
rc = iam_delete(th, bag, (const struct iam_key *)key, ipd);
osd_ipd_put(oti->oti_env, bag, ipd);
LINVRNT(osd_invariant(obj));
RETURN(rc);
}
static int osd_oi_iam_lookup(struct osd_thread_info *oti,
struct osd_oi *oi, struct dt_rec *rec,
const struct dt_key *key)
{
struct iam_container *bag;
struct iam_iterator *it = &oti->oti_idx_it;
struct iam_path_descr *ipd;
int rc;
ENTRY;
LASSERT(oi);
LASSERT(oi->oi_inode);
bag = &oi->oi_dir.od_container;
ipd = osd_idx_ipd_get(oti->oti_env, bag);
if (IS_ERR(ipd))
RETURN(-ENOMEM);
/* got ipd now we can start iterator. */
iam_it_init(it, bag, 0, ipd);
rc = iam_it_get(it, (struct iam_key *)key);
if (rc > 0)
iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)rec);
iam_it_put(it);
iam_it_fini(it);
osd_ipd_put(oti->oti_env, bag, ipd);
LINVRNT(osd_invariant(obj));
RETURN(rc);
}
static int osc_lock_unuse(const struct lu_env *env,
const struct cl_lock_slice *slice)
{
struct osc_lock *ols = cl2osc_lock(slice);
LINVRNT(osc_lock_invariant(ols));
switch (ols->ols_state) {
case OLS_NEW:
LASSERT(!ols->ols_hold);
LASSERT(ols->ols_agl);
return 0;
case OLS_UPCALL_RECEIVED:
osc_lock_unhold(ols);
case OLS_ENQUEUED:
LASSERT(!ols->ols_hold);
osc_lock_detach(env, ols);
ols->ols_state = OLS_NEW;
return 0;
case OLS_GRANTED:
LASSERT(!ols->ols_glimpse);
LASSERT(ols->ols_hold);
/*
* Move lock into OLS_RELEASED state before calling
* osc_cancel_base() so that possible synchronous cancellation
* (that always happens e.g., for liblustre) sees that lock is
* released.
*/
ols->ols_state = OLS_RELEASED;
return osc_lock_unhold(ols);
default:
CERROR("Impossible state: %d\n", ols->ols_state);
LBUG();
}
}
static struct osc_io *cl2osc_io(const struct lu_env *env,
const struct cl_io_slice *slice)
{
struct osc_io *oio = container_of0(slice, struct osc_io, oi_cl);
LINVRNT(oio == osc_env_io(env));
return oio;
}
static void lov_key_fini(const struct lu_context *ctx,
struct lu_context_key *key, void *data)
{
struct lov_thread_info *info = data;
LINVRNT(list_empty(&info->lti_closure.clc_list));
OBD_SLAB_FREE_PTR(info, lov_thread_kmem);
}
static void lov_key_fini(const struct lu_context *ctx,
struct lu_context_key *key, void *data)
{
struct lov_thread_info *info = data;
LINVRNT(list_empty(&info->lti_closure.clc_list));
kmem_cache_free(lov_thread_kmem, info);
}
void osc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
{
struct osc_lock *ols = cl2osc_lock(slice);
LINVRNT(osc_lock_invariant(ols));
LASSERT(ols->ols_dlmlock == NULL);
OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
}
static int lov_page_own(const struct lu_env *env,
const struct cl_page_slice *slice, struct cl_io *io,
int nonblock)
{
struct lov_io *lio = lov_env_io(env);
struct lov_io_sub *sub;
LINVRNT(lov_page_invariant(slice));
LINVRNT(!cl2lov_page(slice)->lps_invalid);
sub = lov_page_subio(env, lio, slice);
if (!IS_ERR(sub)) {
lov_sub_page(slice)->cp_owner = sub->sub_io;
lov_sub_put(sub);
} else
LBUG(); /* Arrgh */
return 0;
}
static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
struct lustre_handle *lockh, bool lvb_update)
{
struct ldlm_lock *dlmlock;
dlmlock = ldlm_handle2lock_long(lockh, 0);
LASSERT(dlmlock);
/* lock reference taken by ldlm_handle2lock_long() is
* owned by osc_lock and released in osc_lock_detach()
*/
lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
oscl->ols_has_ref = 1;
LASSERT(!oscl->ols_dlmlock);
oscl->ols_dlmlock = dlmlock;
/* This may be a matched lock for glimpse request, do not hold
* lock reference in that case.
*/
if (!oscl->ols_glimpse) {
/* hold a refc for non glimpse lock which will
* be released in osc_lock_cancel()
*/
lustre_handle_copy(&oscl->ols_handle, lockh);
ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
oscl->ols_hold = 1;
}
/* Lock must have been granted. */
lock_res_and_lock(dlmlock);
if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
/* extend the lock extent, otherwise it will have problem when
* we decide whether to grant a lockless lock.
*/
descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
descr->cld_start = cl_index(descr->cld_obj, ext->start);
descr->cld_end = cl_index(descr->cld_obj, ext->end);
descr->cld_gid = ext->gid;
/* no lvb update for matched lock */
if (lvb_update) {
LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
dlmlock, NULL);
}
LINVRNT(osc_lock_invariant(oscl));
}
unlock_res_and_lock(dlmlock);
LASSERT(oscl->ols_state != OLS_GRANTED);
oscl->ols_state = OLS_GRANTED;
}
static void osc_lock_fini(const struct lu_env *env,
struct cl_lock_slice *slice)
{
struct osc_lock *ols = cl2osc_lock(slice);
LINVRNT(osc_lock_invariant(ols));
LASSERT(!ols->ols_dlmlock);
kmem_cache_free(osc_lock_kmem, ols);
}
static int lov_page_cache_add(const struct lu_env *env,
const struct cl_page_slice *slice,
struct cl_io *io)
{
struct lov_io *lio = lov_env_io(env);
struct lov_io_sub *sub;
int rc = 0;
LINVRNT(lov_page_invariant(slice));
LINVRNT(!cl2lov_page(slice)->lps_invalid);
sub = lov_page_subio(env, lio, slice);
if (!IS_ERR(sub)) {
rc = cl_page_cache_add(sub->sub_env, sub->sub_io,
slice->cpl_page->cp_child, CRT_WRITE);
lov_sub_put(sub);
} else {
rc = PTR_ERR(sub);
CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page, "rc = %d\n", rc);
}
return rc;
}
static void lov_page_fini(const struct lu_env *env,
struct cl_page_slice *slice)
{
struct cl_page *sub = lov_sub_page(slice);
LINVRNT(lov_page_invariant(slice));
if (sub != NULL) {
LASSERT(sub->cp_state == CPS_FREEING);
lu_ref_del(&sub->cp_reference, "lov", sub->cp_parent);
sub->cp_parent = NULL;
slice->cpl_page->cp_child = NULL;
cl_page_put(env, sub);
}
}
static void osc_lock_fini(const struct lu_env *env,
struct cl_lock_slice *slice)
{
struct osc_lock *ols = cl2osc_lock(slice);
LINVRNT(osc_lock_invariant(ols));
/*
* ->ols_hold can still be true at this point if, for example, a
* thread that requested a lock was killed (and released a reference
* to the lock), before reply from a server was received. In this case
* lock is destroyed immediately after upcall.
*/
osc_lock_unhold(ols);
LASSERT(ols->ols_lock == NULL);
LASSERT(atomic_read(&ols->ols_pageref) == 0 ||
atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC);
OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
}
/**
* Called when a lock is granted, from an upcall (when server returned a
* granted lock), or from completion AST, when server returned a blocked lock.
*
* Called under lock and resource spin-locks, that are released temporarily
* here.
*/
static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
struct ldlm_lock *dlmlock, int rc)
{
struct ldlm_extent *ext;
struct cl_lock *lock;
struct cl_lock_descr *descr;
LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
ENTRY;
if (olck->ols_state < OLS_GRANTED) {
lock = olck->ols_cl.cls_lock;
ext = &dlmlock->l_policy_data.l_extent;
descr = &osc_env_info(env)->oti_descr;
descr->cld_obj = lock->cll_descr.cld_obj;
/* XXX check that ->l_granted_mode is valid. */
descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
descr->cld_start = cl_index(descr->cld_obj, ext->start);
descr->cld_end = cl_index(descr->cld_obj, ext->end);
descr->cld_gid = ext->gid;
/*
* tell upper layers the extent of the lock that was actually
* granted
*/
olck->ols_state = OLS_GRANTED;
osc_lock_lvb_update(env, olck, rc);
/* release DLM spin-locks to allow cl_lock_{modify,signal}()
* to take a semaphore on a parent lock. This is safe, because
* spin-locks are needed to protect consistency of
* dlmlock->l_*_mode and LVB, and we have finished processing
* them. */
unlock_res_and_lock(dlmlock);
cl_lock_modify(env, lock, descr);
cl_lock_signal(env, lock);
LINVRNT(osc_lock_invariant(olck));
lock_res_and_lock(dlmlock);
}
EXIT;
}
static struct osc_object *lu2osc(const struct lu_object *obj)
{
LINVRNT(osc_is_object(obj));
return container_of0(obj, struct osc_object, oo_cl.co_lu);
}
static struct osc_req *cl2osc_req(const struct cl_req_slice *slice)
{
LINVRNT(slice->crs_dev->cd_lu_dev.ld_type == &osc_device_type);
return container_of0(slice, struct osc_req, or_cl);
}
static int vvp_io_commit_write(const struct lu_env *env,
const struct cl_io_slice *ios,
const struct cl_page_slice *slice,
unsigned from, unsigned to)
{
struct cl_object *obj = slice->cpl_obj;
struct cl_io *io = ios->cis_io;
struct ccc_page *cp = cl2ccc_page(slice);
struct cl_page *pg = slice->cpl_page;
struct inode *inode = ccc_object_inode(obj);
struct ll_sb_info *sbi = ll_i2sbi(inode);
struct ll_inode_info *lli = ll_i2info(inode);
struct page *vmpage = cp->cpg_page;
int result;
int tallyop;
loff_t size;
ENTRY;
LINVRNT(cl_page_is_vmlocked(env, pg));
LASSERT(vmpage->mapping->host == inode);
LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
/*
* queue a write for some time in the future the first time we
* dirty the page.
*
* This is different from what other file systems do: they usually
* just mark page (and some of its buffers) dirty and rely on
* balance_dirty_pages() to start a write-back. Lustre wants write-back
* to be started earlier for the following reasons:
*
* (1) with a large number of clients we need to limit the amount
* of cached data on the clients a lot;
*
* (2) large compute jobs generally want compute-only then io-only
* and the IO should complete as quickly as possible;
*
* (3) IO is batched up to the RPC size and is async until the
* client max cache is hit
* (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
*
*/
if (!PageDirty(vmpage)) {
tallyop = LPROC_LL_DIRTY_MISSES;
result = cl_page_cache_add(env, io, pg, CRT_WRITE);
if (result == 0) {
/* page was added into cache successfully. */
set_page_dirty(vmpage);
vvp_write_pending(cl2ccc(obj), cp);
} else if (result == -EDQUOT) {
pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
bool need_clip = true;
/*
* Client ran out of disk space grant. Possible
* strategies are:
*
* (a) do a sync write, renewing grant;
*
* (b) stop writing on this stripe, switch to the
* next one.
*
* (b) is a part of "parallel io" design that is the
* ultimate goal. (a) is what "old" client did, and
* what the new code continues to do for the time
* being.
*/
if (last_index > pg->cp_index) {
to = PAGE_CACHE_SIZE;
need_clip = false;
} else if (last_index == pg->cp_index) {
int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
if (to < size_to)
to = size_to;
}
if (need_clip)
cl_page_clip(env, pg, 0, to);
result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
if (result)
CERROR("Write page %lu of inode %p failed %d\n",
pg->cp_index, inode, result);
}
/**
* Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
* and ldlm_lock caches.
*/
static int osc_dlm_blocking_ast0(const struct lu_env *env,
struct ldlm_lock *dlmlock,
void *data, int flag)
{
struct osc_lock *olck;
struct cl_lock *lock;
int result;
int cancel;
LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
cancel = 0;
olck = osc_ast_data_get(dlmlock);
if (olck != NULL) {
lock = olck->ols_cl.cls_lock;
cl_lock_mutex_get(env, lock);
LINVRNT(osc_lock_invariant(olck));
if (olck->ols_ast_wait) {
/* wake up osc_lock_use() */
cl_lock_signal(env, lock);
olck->ols_ast_wait = 0;
}
/*
* Lock might have been canceled while this thread was
* sleeping for lock mutex, but olck is pinned in memory.
*/
if (olck == dlmlock->l_ast_data) {
/*
* NOTE: DLM sends blocking AST's for failed locks
* (that are still in pre-OLS_GRANTED state)
* too, and they have to be canceled otherwise
* DLM lock is never destroyed and stuck in
* the memory.
*
* Alternatively, ldlm_cli_cancel() can be
* called here directly for osc_locks with
* ols_state < OLS_GRANTED to maintain an
* invariant that ->clo_cancel() is only called
* for locks that were granted.
*/
LASSERT(data == olck);
osc_lock_blocking(env, dlmlock,
olck, flag == LDLM_CB_BLOCKING);
} else
cancel = 1;
cl_lock_mutex_put(env, lock);
osc_ast_data_put(env, olck);
} else
/*
* DLM lock exists, but there is no cl_lock attached to it.
* This is a `normal' race. cl_object and its cl_lock's can be
* removed by memory pressure, together with all pages.
*/
cancel = (flag == LDLM_CB_BLOCKING);
if (cancel) {
struct lustre_handle *lockh;
lockh = &osc_env_info(env)->oti_handle;
ldlm_lock2handle(dlmlock, lockh);
result = ldlm_cli_cancel(lockh, LCF_ASYNC);
} else
result = 0;
return result;
}
