]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net> | |
7 | * | |
8 | * This is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License version 2.1, as published by the Free Software | |
11 | * Foundation. See file COPYING. | |
12 | * | |
13 | */ | |
14 | ||
15 | #ifndef CEPH_OSD_SESSION_H | |
16 | #define CEPH_OSD_SESSION_H | |
17 | ||
18 | #include "common/RefCountedObj.h" | |
19 | #include "common/Mutex.h" | |
20 | #include "include/Spinlock.h" | |
21 | #include "OSDCap.h" | |
22 | #include "Watch.h" | |
23 | #include "OSDMap.h" | |
24 | ||
25 | struct Session; | |
26 | typedef boost::intrusive_ptr<Session> SessionRef; | |
27 | struct Backoff; | |
28 | typedef boost::intrusive_ptr<Backoff> BackoffRef; | |
29 | class PG; | |
30 | typedef boost::intrusive_ptr<PG> PGRef; | |
31 | ||
32 | /* | |
33 | * A Backoff represents one instance of either a PG or an OID | |
34 | * being plugged at the client. It's refcounted and linked from | |
35 | * the PG {pg_oid}_backoffs map and from the client Session | |
36 | * object. | |
37 | * | |
38 | * The Backoff has a lock that protects it's internal fields. | |
39 | * | |
40 | * The PG has a backoff_lock that protects it's maps to Backoffs. | |
41 | * This lock is *inside* of Backoff::lock. | |
42 | * | |
43 | * The Session has a backoff_lock that protects it's map of pg and | |
44 | * oid backoffs. This lock is *inside* the Backoff::lock *and* | |
45 | * PG::backoff_lock. | |
46 | * | |
47 | * That's | |
48 | * | |
49 | * Backoff::lock | |
50 | * PG::backoff_lock | |
51 | * Session::backoff_lock | |
52 | * | |
53 | * When the Session goes away, we move our backoff lists aside, | |
54 | * then we lock each of the Backoffs we | |
55 | * previously referenced and clear the Session* pointer. If the PG | |
56 | * is still linked, we unlink it, too. | |
57 | * | |
58 | * When the PG clears the backoff, it will send an unblock message | |
59 | * if the Session* is still non-null, and unlink the session. | |
60 | * | |
61 | */ | |
62 | ||
63 | struct Backoff : public RefCountedObject { | |
64 | enum { | |
65 | STATE_NEW = 1, ///< backoff in flight to client | |
66 | STATE_ACKED = 2, ///< backoff acked | |
67 | STATE_DELETING = 3 ///< backoff deleted, but un-acked | |
68 | }; | |
69 | std::atomic_int state = {STATE_NEW}; | |
70 | spg_t pgid; ///< owning pgid | |
71 | uint64_t id = 0; ///< unique id (within the Session) | |
72 | ||
73 | bool is_new() const { | |
74 | return state.load() == STATE_NEW; | |
75 | } | |
76 | bool is_acked() const { | |
77 | return state.load() == STATE_ACKED; | |
78 | } | |
79 | bool is_deleting() const { | |
80 | return state.load() == STATE_DELETING; | |
81 | } | |
82 | const char *get_state_name() const { | |
83 | switch (state.load()) { | |
84 | case STATE_NEW: return "new"; | |
85 | case STATE_ACKED: return "acked"; | |
86 | case STATE_DELETING: return "deleting"; | |
87 | default: return "???"; | |
88 | } | |
89 | } | |
90 | ||
91 | Mutex lock; | |
92 | // NOTE: the owning PG and session are either | |
93 | // - *both* set, or | |
94 | // - both null (teardown), or | |
95 | // - only session is set (and state == DELETING) | |
96 | PGRef pg; ///< owning pg | |
97 | SessionRef session; ///< owning session | |
98 | hobject_t begin, end; ///< [) range to block, unless ==, then single obj | |
99 | ||
100 | Backoff(spg_t pgid, PGRef pg, SessionRef s, | |
101 | uint64_t i, | |
102 | const hobject_t& b, const hobject_t& e) | |
103 | : RefCountedObject(g_ceph_context, 0), | |
104 | pgid(pgid), | |
105 | id(i), | |
106 | lock("Backoff::lock"), | |
107 | pg(pg), | |
108 | session(s), | |
109 | begin(b), | |
110 | end(e) {} | |
111 | ||
112 | friend ostream& operator<<(ostream& out, const Backoff& b) { | |
113 | return out << "Backoff(" << &b << " " << b.pgid << " " << b.id | |
114 | << " " << b.get_state_name() | |
115 | << " [" << b.begin << "," << b.end << ") " | |
116 | << " session " << b.session | |
117 | << " pg " << b.pg << ")"; | |
118 | } | |
119 | }; | |
120 | ||
121 | ||
122 | ||
123 | struct Session : public RefCountedObject { | |
124 | EntityName entity_name; | |
125 | OSDCap caps; | |
126 | int64_t auid; | |
127 | ConnectionRef con; | |
128 | WatchConState wstate; | |
129 | ||
130 | Mutex session_dispatch_lock; | |
131 | boost::intrusive::list<OpRequest> waiting_on_map; | |
132 | ||
133 | Spinlock sent_epoch_lock; | |
134 | epoch_t last_sent_epoch; | |
135 | Spinlock received_map_lock; | |
136 | epoch_t received_map_epoch; // largest epoch seen in MOSDMap from here | |
137 | ||
138 | /// protects backoffs; orders inside Backoff::lock *and* PG::backoff_lock | |
139 | Mutex backoff_lock; | |
140 | std::atomic_int backoff_count= {0}; ///< simple count of backoffs | |
141 | map<spg_t,map<hobject_t,set<BackoffRef>>> backoffs; | |
142 | ||
143 | std::atomic<uint64_t> backoff_seq = {0}; | |
144 | ||
145 | explicit Session(CephContext *cct) : | |
146 | RefCountedObject(cct), | |
147 | auid(-1), con(0), | |
148 | wstate(cct), | |
149 | session_dispatch_lock("Session::session_dispatch_lock"), | |
150 | last_sent_epoch(0), received_map_epoch(0), | |
151 | backoff_lock("Session::backoff_lock") | |
152 | {} | |
153 | ||
154 | void ack_backoff( | |
155 | CephContext *cct, | |
156 | spg_t pgid, | |
157 | uint64_t id, | |
158 | const hobject_t& start, | |
159 | const hobject_t& end); | |
160 | ||
161 | BackoffRef have_backoff(spg_t pgid, const hobject_t& oid) { | |
162 | if (!backoff_count.load()) { | |
163 | return nullptr; | |
164 | } | |
165 | Mutex::Locker l(backoff_lock); | |
166 | assert(!backoff_count == backoffs.empty()); | |
167 | auto i = backoffs.find(pgid); | |
168 | if (i == backoffs.end()) { | |
169 | return nullptr; | |
170 | } | |
171 | auto p = i->second.lower_bound(oid); | |
172 | if (p != i->second.begin() && | |
173 | p->first > oid) { | |
174 | --p; | |
175 | } | |
176 | if (p != i->second.end()) { | |
177 | int r = cmp(oid, p->first); | |
178 | if (r == 0 || r > 0) { | |
179 | for (auto& q : p->second) { | |
180 | if (r == 0 || oid < q->end) { | |
181 | return &(*q); | |
182 | } | |
183 | } | |
184 | } | |
185 | } | |
186 | return nullptr; | |
187 | } | |
188 | ||
189 | bool check_backoff( | |
190 | CephContext *cct, spg_t pgid, const hobject_t& oid, const Message *m); | |
191 | ||
192 | void add_backoff(BackoffRef b) { | |
193 | Mutex::Locker l(backoff_lock); | |
194 | assert(!backoff_count == backoffs.empty()); | |
195 | backoffs[b->pgid][b->begin].insert(b); | |
196 | ++backoff_count; | |
197 | } | |
198 | ||
199 | // called by PG::release_*_backoffs and PG::clear_backoffs() | |
200 | void rm_backoff(BackoffRef b) { | |
201 | Mutex::Locker l(backoff_lock); | |
202 | assert(b->lock.is_locked_by_me()); | |
203 | assert(b->session == this); | |
204 | auto i = backoffs.find(b->pgid); | |
205 | if (i != backoffs.end()) { | |
206 | // may race with clear_backoffs() | |
207 | auto p = i->second.find(b->begin); | |
208 | if (p != i->second.end()) { | |
209 | auto q = p->second.find(b); | |
210 | if (q != p->second.end()) { | |
211 | p->second.erase(q); | |
212 | --backoff_count; | |
213 | if (p->second.empty()) { | |
214 | i->second.erase(p); | |
215 | if (i->second.empty()) { | |
216 | backoffs.erase(i); | |
217 | } | |
218 | } | |
219 | } | |
220 | } | |
221 | } | |
222 | assert(!backoff_count == backoffs.empty()); | |
223 | } | |
224 | void clear_backoffs(); | |
225 | }; | |
226 | ||
227 | #endif |