struct MDLockCache;
struct MutationImpl : public TrackedOp {
- metareqid_t reqid;
- __u32 attempt = 0; // which attempt for this request
- LogSegment *ls = nullptr; // the log segment i'm committing to
-
-private:
- utime_t mds_stamp; ///< mds-local timestamp (real time)
- utime_t op_stamp; ///< op timestamp (client provided)
-
public:
- // flag mutation as slave
- mds_rank_t slave_to_mds = MDS_RANK_NONE; // this is a slave request if >= 0.
-
// -- my pins and auth_pins --
struct ObjectState {
bool pinned = false;
bool auth_pinned = false;
mds_rank_t remote_auth_pinned = MDS_RANK_NONE;
};
- ceph::unordered_map<MDSCacheObject*, ObjectState> object_states;
- int num_pins = 0;
- int num_auth_pins = 0;
- int num_remote_auth_pins = 0;
-
- const ObjectState* find_object_state(MDSCacheObject *obj) const {
- auto it = object_states.find(obj);
- return it != object_states.end() ? &it->second : nullptr;
- }
- bool is_any_remote_auth_pin() const { return num_remote_auth_pins > 0; }
-
- // cache pins (so things don't expire)
- CInode* stickydiri = nullptr;
-
-
// held locks
struct LockOp {
enum {
REMOTE_WRLOCK = 8,
STATE_PIN = 16, // no RW after locked, just pin lock state
};
- SimpleLock* lock;
- mutable unsigned flags;
- mutable mds_rank_t wrlock_target;
+
LockOp(SimpleLock *l, unsigned f=0, mds_rank_t t=MDS_RANK_NONE) :
lock(l), flags(f), wrlock_target(t) {}
+
bool is_rdlock() const { return !!(flags & RDLOCK); }
bool is_xlock() const { return !!(flags & XLOCK); }
bool is_wrlock() const { return !!(flags & WRLOCK); }
bool operator<(const LockOp& r) const {
return lock < r.lock;
}
+
+ SimpleLock* lock;
+ mutable unsigned flags;
+ mutable mds_rank_t wrlock_target;
};
struct LockOpVec : public vector<LockOp> {
+ LockOpVec() {
+ reserve(32);
+ }
+
void add_rdlock(SimpleLock *lock) {
emplace_back(lock, LockOp::RDLOCK);
}
emplace_back(lock, LockOp::WRLOCK | LockOp::STATE_PIN);
}
void sort_and_merge();
-
- LockOpVec() {
- reserve(32);
- }
};
+
using lock_set = set<LockOp>;
using lock_iterator = lock_set::iterator;
- lock_set locks; // full ordering
- MDLockCache* lock_cache = nullptr;
- bool lock_cache_disabled = false;
+ // keep our default values synced with MDRequestParam's
+ MutationImpl() : TrackedOp(nullptr, utime_t()) {}
+ MutationImpl(OpTracker *tracker, utime_t initiated,
+ const metareqid_t &ri, __u32 att=0, mds_rank_t slave_to=MDS_RANK_NONE)
+ : TrackedOp(tracker, initiated),
+ reqid(ri), attempt(att),
+ slave_to_mds(slave_to) {}
+ ~MutationImpl() override {
+ ceph_assert(!locking);
+ ceph_assert(!lock_cache);
+ ceph_assert(num_pins == 0);
+ ceph_assert(num_auth_pins == 0);
+ }
+
+ const ObjectState* find_object_state(MDSCacheObject *obj) const {
+ auto it = object_states.find(obj);
+ return it != object_states.end() ? &it->second : nullptr;
+ }
+
+ bool is_any_remote_auth_pin() const { return num_remote_auth_pins > 0; }
void disable_lock_cache() {
lock_cache_disabled = true;
return lock == last_locked;
}
- SimpleLock *last_locked = nullptr;
- // lock we are currently trying to acquire. if we give up for some reason,
- // be sure to eval() this.
- SimpleLock *locking = nullptr;
- mds_rank_t locking_target_mds = -1;
-
- // if this flag is set, do not attempt to acquire further locks.
- // (useful for wrlock, which may be a moving auth target)
- enum {
- SNAP_LOCKED = 1,
- SNAP2_LOCKED = 2,
- PATH_LOCKED = 4,
- ALL_LOCKED = 8,
- };
- int locking_state = 0;
-
- bool committing = false;
- bool aborted = false;
- bool killed = false;
-
- // for applying projected inode changes
- list<CInode*> projected_inodes;
- std::vector<CDir*> projected_fnodes;
- list<ScatterLock*> updated_locks;
-
- list<CInode*> dirty_cow_inodes;
- list<pair<CDentry*,version_t> > dirty_cow_dentries;
-
- // keep our default values synced with MDRequestParam's
- MutationImpl() : TrackedOp(nullptr, utime_t()) {}
- MutationImpl(OpTracker *tracker, utime_t initiated,
- const metareqid_t &ri, __u32 att=0, mds_rank_t slave_to=MDS_RANK_NONE)
- : TrackedOp(tracker, initiated),
- reqid(ri), attempt(att),
- slave_to_mds(slave_to) { }
- ~MutationImpl() override {
- ceph_assert(!locking);
- ceph_assert(!lock_cache);
- ceph_assert(num_pins == 0);
- ceph_assert(num_auth_pins == 0);
- }
-
bool is_master() const { return slave_to_mds == MDS_RANK_NONE; }
bool is_slave() const { return slave_to_mds != MDS_RANK_NONE; }
virtual void dump(Formatter *f) const {}
void _dump_op_descriptor_unlocked(ostream& stream) const override;
-};
-
-inline ostream& operator<<(ostream &out, const MutationImpl &mut)
-{
- mut.print(out);
- return out;
-}
-
-typedef boost::intrusive_ptr<MutationImpl> MutationRef;
-
+ metareqid_t reqid;
+ __u32 attempt = 0; // which attempt for this request
+ LogSegment *ls = nullptr; // the log segment i'm committing to
-/**
- * MDRequestImpl: state we track for requests we are currently processing.
- * mostly information about locks held, so that we can drop them all
- * the request is finished or forwarded. see request_*().
- */
-struct MDRequestImpl : public MutationImpl {
- Session *session;
- elist<MDRequestImpl*>::item item_session_request; // if not on list, op is aborted.
-
- // -- i am a client (master) request
- cref_t<MClientRequest> client_request; // client request (if any)
-
- // tree and depth info of path1 and path2
- inodeno_t dir_root[2] = {0, 0};
- int dir_depth[2] = {-1, -1};
- file_layout_t dir_layout;
-
- // store up to two sets of dn vectors, inode pointers, for request path1 and path2.
- vector<CDentry*> dn[2];
- CInode *in[2];
- CDentry *straydn;
- snapid_t snapid;
-
- CInode *tracei;
- CDentry *tracedn;
-
- inodeno_t alloc_ino, used_prealloc_ino;
- interval_set<inodeno_t> prealloc_inos;
+ // flag mutation as slave
+ mds_rank_t slave_to_mds = MDS_RANK_NONE; // this is a slave request if >= 0.
- int snap_caps = 0;
- int getattr_caps = 0; ///< caps requested by getattr
- bool no_early_reply = false;
- bool did_early_reply = false;
- bool o_trunc = false; ///< request is an O_TRUNC mutation
- bool has_completed = false; ///< request has already completed
+ ceph::unordered_map<MDSCacheObject*, ObjectState> object_states;
+ int num_pins = 0;
+ int num_auth_pins = 0;
+ int num_remote_auth_pins = 0;
+ // cache pins (so things don't expire)
+ CInode* stickydiri = nullptr;
- bufferlist reply_extra_bl;
+ lock_set locks; // full ordering
+ MDLockCache* lock_cache = nullptr;
+ bool lock_cache_disabled = false;
+ SimpleLock *last_locked = nullptr;
+ // Lock we are currently trying to acquire. If we give up for some reason,
+ // be sure to eval() this.
+ SimpleLock *locking = nullptr;
+ mds_rank_t locking_target_mds = -1;
- // inos we did a embedded cap release on, and may need to eval if we haven't since reissued
- map<vinodeno_t, ceph_seq_t> cap_releases;
+ // if this flag is set, do not attempt to acquire further locks.
+ // (useful for wrlock, which may be a moving auth target)
+ enum {
+ SNAP_LOCKED = 1,
+ SNAP2_LOCKED = 2,
+ PATH_LOCKED = 4,
+ ALL_LOCKED = 8,
+ };
+ int locking_state = 0;
- // -- i am a slave request
- cref_t<MMDSSlaveRequest> slave_request; // slave request (if one is pending; implies slave == true)
+ bool committing = false;
+ bool aborted = false;
+ bool killed = false;
- // -- i am an internal op
- int internal_op;
- Context *internal_op_finish;
- void *internal_op_private;
+ // for applying projected inode changes
+ list<CInode*> projected_inodes;
+ std::vector<CDir*> projected_fnodes;
+ list<ScatterLock*> updated_locks;
- // indicates how may retries of request have been made
- int retry;
+ list<CInode*> dirty_cow_inodes;
+ list<pair<CDentry*,version_t> > dirty_cow_dentries;
- bool is_batch_head = false;
+private:
+ utime_t mds_stamp; ///< mds-local timestamp (real time)
+ utime_t op_stamp; ///< op timestamp (client provided)
+};
- // indicator for vxattr osdmap update
- bool waited_for_osdmap;
+/**
+ * MDRequestImpl: state we track for requests we are currently processing.
+ * mostly information about locks held, so that we can drop them all
+ * the request is finished or forwarded. see request_*().
+ */
+struct MDRequestImpl : public MutationImpl {
+ // TrackedOp stuff
+ typedef boost::intrusive_ptr<MDRequestImpl> Ref;
// break rarely-used fields into a separately allocated structure
// to save memory for most ops
struct More {
+ More() {}
+
int slave_error = 0;
set<mds_rank_t> slaves; // mds nodes that have slave requests to me (implies client_request)
set<mds_rank_t> waiting_on_slave; // peers i'm waiting for slavereq replies from.
// for internal ops doing lookup
filepath filepath1;
filepath filepath2;
-
- More() {}
- } *_more;
-
+ } *_more = nullptr;
// ---------------------------------------------------
struct Params {
- metareqid_t reqid;
- __u32 attempt;
- cref_t<MClientRequest> client_req;
- cref_t<Message> triggering_slave_req;
- mds_rank_t slave_to;
- utime_t initiated;
- utime_t throttled, all_read, dispatched;
- int internal_op;
// keep these default values synced to MutationImpl's
- Params() : attempt(0), slave_to(MDS_RANK_NONE), internal_op(-1) {}
+ Params() {}
const utime_t& get_recv_stamp() const {
return initiated;
}
const utime_t& get_dispatch_stamp() const {
return dispatched;
}
+ metareqid_t reqid;
+ __u32 attempt = 0;
+ cref_t<MClientRequest> client_req;
+ cref_t<Message> triggering_slave_req;
+ mds_rank_t slave_to = MDS_RANK_NONE;
+ utime_t initiated;
+ utime_t throttled, all_read, dispatched;
+ int internal_op = -1;
};
MDRequestImpl(const Params* params, OpTracker *tracker) :
MutationImpl(tracker, params->initiated,
params->reqid, params->attempt, params->slave_to),
- session(NULL), item_session_request(this),
- client_request(params->client_req), straydn(NULL), snapid(CEPH_NOSNAP),
- tracei(NULL), tracedn(NULL), alloc_ino(0), used_prealloc_ino(0),
- internal_op(params->internal_op), internal_op_finish(NULL),
- internal_op_private(NULL),
- retry(0),
- waited_for_osdmap(false), _more(NULL) {
- in[0] = in[1] = NULL;
- }
+ item_session_request(this), client_request(params->client_req),
+ internal_op(params->internal_op) {}
~MDRequestImpl() override;
More* more();
cref_t<MClientRequest> release_client_request();
void reset_slave_request(const cref_t<MMDSSlaveRequest>& req=nullptr);
- // TrackedOp stuff
- typedef boost::intrusive_ptr<MDRequestImpl> Ref;
+ Session *session = nullptr;
+ elist<MDRequestImpl*>::item item_session_request; // if not on list, op is aborted.
+
+ // -- i am a client (master) request
+ cref_t<MClientRequest> client_request; // client request (if any)
+
+ // tree and depth info of path1 and path2
+ inodeno_t dir_root[2] = {0, 0};
+ int dir_depth[2] = {-1, -1};
+ file_layout_t dir_layout;
+ // store up to two sets of dn vectors, inode pointers, for request path1 and path2.
+ vector<CDentry*> dn[2];
+ CInode *in[2] = {};
+ CDentry *straydn = nullptr;
+ snapid_t snapid = CEPH_NOSNAP;
+
+ CInode *tracei = nullptr;
+ CDentry *tracedn = nullptr;
+
+ inodeno_t alloc_ino = 0, used_prealloc_ino = 0;
+ interval_set<inodeno_t> prealloc_inos;
+
+ int snap_caps = 0;
+ int getattr_caps = 0; ///< caps requested by getattr
+ bool no_early_reply = false;
+ bool did_early_reply = false;
+ bool o_trunc = false; ///< request is an O_TRUNC mutation
+ bool has_completed = false; ///< request has already completed
+
+ bufferlist reply_extra_bl;
+
+ // inos we did a embedded cap release on, and may need to eval if we haven't since reissued
+ map<vinodeno_t, ceph_seq_t> cap_releases;
+
+ // -- i am a slave request
+ cref_t<MMDSSlaveRequest> slave_request; // slave request (if one is pending; implies slave == true)
+
+ // -- i am an internal op
+ int internal_op;
+ Context *internal_op_finish = nullptr;
+ void *internal_op_private = nullptr;
+
+ // indicates how may retries of request have been made
+ int retry = 0;
+
+ bool is_batch_head = false;
+
+ // indicator for vxattr osdmap update
+ bool waited_for_osdmap = false;
+
std::vector<Ref> batch_reqs;
protected:
void _dump(Formatter *f) const override;
mutable ceph::spinlock msg_lock;
};
-typedef boost::intrusive_ptr<MDRequestImpl> MDRequestRef;
-
-
struct MDSlaveUpdate {
- int origop;
- bufferlist rollback;
- elist<MDSlaveUpdate*>::item item;
- Context *waiter;
- set<CInode*> olddirs;
- set<CInode*> unlinked;
MDSlaveUpdate(int oo, bufferlist &rbl, elist<MDSlaveUpdate*> &list) :
origop(oo),
- item(this),
- waiter(0) {
+ item(this) {
rollback.claim(rbl);
list.push_back(&item);
}
if (waiter)
waiter->complete(0);
}
+ int origop;
+ bufferlist rollback;
+ elist<MDSlaveUpdate*>::item item;
+ Context *waiter = nullptr;
+ set<CInode*> olddirs;
+ set<CInode*> unlinked;
};
struct MDLockCacheItem {
};
struct MDLockCache : public MutationImpl {
- CInode *diri;
- Capability *client_cap;
- int opcode;
- file_layout_t dir_layout;
-
- elist<MDLockCache*>::item item_cap_lock_cache;
-
using LockItem = MDLockCacheItem;
- // link myself to locked locks
- std::unique_ptr<LockItem[]> items_lock;
struct DirItem {
MDLockCache *parent = nullptr;
elist<DirItem*>::item item_dir;
};
- // link myself to auth-pinned dirfrags
- std::unique_ptr<DirItem[]> items_dir;
- std::vector<CDir*> auth_pinned_dirfrags;
-
- int ref = 1;
- bool invalidating = false;
MDLockCache(Capability *cap, int op) :
MutationImpl(), diri(cap->get_inode()), client_cap(cap), opcode(op) {
void attach_locks();
void attach_dirfrags(std::vector<CDir*>&& dfv);
void detach_all();
+
+ CInode *diri;
+ Capability *client_cap;
+ int opcode;
+ file_layout_t dir_layout;
+
+ elist<MDLockCache*>::item item_cap_lock_cache;
+
+ // link myself to locked locks
+ std::unique_ptr<LockItem[]> items_lock;
+
+ // link myself to auth-pinned dirfrags
+ std::unique_ptr<DirItem[]> items_dir;
+ std::vector<CDir*> auth_pinned_dirfrags;
+
+ int ref = 1;
+ bool invalidating = false;
};
+typedef boost::intrusive_ptr<MutationImpl> MutationRef;
+typedef boost::intrusive_ptr<MDRequestImpl> MDRequestRef;
+
+inline ostream& operator<<(ostream &out, const MutationImpl &mut)
+{
+ mut.print(out);
+ return out;
+}
#endif