include/block/graph-lock.h

   1 /*
   2  * Graph lock: rwlock to protect block layer graph manipulations (add/remove
   3  * edges and nodes)
   4  *
   5  *  Copyright (c) 2022 Red Hat
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
  19  */
  20 #ifndef GRAPH_LOCK_H
  21 #define GRAPH_LOCK_H
  22
  23 #include "qemu/clang-tsa.h"
  24
  25 /**
  26  * Graph Lock API
  27  * This API provides a rwlock used to protect block layer
  28  * graph modifications like edge (BdrvChild) and node (BlockDriverState)
  29  * addition and removal.
  30  * Currently we have 1 writer only, the Main loop, and many
  31  * readers, mostly coroutines running in other AioContext thus other threads.
  32  *
  33  * We distinguish between writer (main loop, under BQL) that modifies the
  34  * graph, and readers (all other coroutines running in various AioContext),
  35  * that go through the graph edges, reading
  36  * BlockDriverState ->parents and->children.
  37  *
  38  * The writer (main loop)  has an "exclusive" access, so it first waits for
  39  * current read to finish, and then prevents incoming ones from
  40  * entering while it has the exclusive access.
  41  *
  42  * The readers (coroutines in multiple AioContext) are free to
  43  * access the graph as long the writer is not modifying the graph.
  44  * In case it is, they go in a CoQueue and sleep until the writer
  45  * is done.
  46  *
  47  * If a coroutine changes AioContext, the counter in the original and new
  48  * AioContext are left intact, since the writer does not care where is the
  49  * reader, but only if there is one.
  50  * As a result, some AioContexts might have a negative reader count, to
  51  * balance the positive count of the AioContext that took the lock.
  52  * This also means that when an AioContext is deleted it may have a nonzero
  53  * reader count. In that case we transfer the count to a global shared counter
  54  * so that the writer is always aware of all readers.
  55  */
  56 typedef struct BdrvGraphRWlock BdrvGraphRWlock;
  57
  58 /* Dummy lock object to use for Thread Safety Analysis (TSA) */
  59 typedef struct TSA_CAPABILITY("mutex") BdrvGraphLock {
  60 } BdrvGraphLock;
  61
  62 extern BdrvGraphLock graph_lock;
  63
  64 /*
  65  * clang doesn't check consistency in locking annotations between forward
  66  * declarations and the function definition. Having the annotation on the
  67  * definition, but not the declaration in a header file, may give the reader
  68  * a false sense of security because the condition actually remains unchecked
  69  * for callers in other source files.
  70  *
  71  * Therefore, as a convention, for public functions, GRAPH_RDLOCK and
  72  * GRAPH_WRLOCK annotations should be present only in the header file.
  73  */
  74 #define GRAPH_WRLOCK TSA_REQUIRES(graph_lock)
  75 #define GRAPH_RDLOCK TSA_REQUIRES_SHARED(graph_lock)
  76 #define GRAPH_UNLOCKED TSA_EXCLUDES(graph_lock)
  77
  78 /*
  79  * TSA annotations are not part of function types, so checks are defeated when
  80  * using a function pointer. As a workaround, annotate function pointers with
  81  * this macro that will require that the lock is at least taken while reading
  82  * the pointer. In most cases this is equivalent to actually protecting the
  83  * function call.
  84  */
  85 #define GRAPH_RDLOCK_PTR TSA_GUARDED_BY(graph_lock)
  86 #define GRAPH_WRLOCK_PTR TSA_GUARDED_BY(graph_lock)
  87 #define GRAPH_UNLOCKED_PTR
  88
  89 /*
  90  * register_aiocontext:
  91  * Add AioContext @ctx to the list of AioContext.
  92  * This list is used to obtain the total number of readers
  93  * currently running the graph.
  94  */
  95 void register_aiocontext(AioContext *ctx);
  96
  97 /*
  98  * unregister_aiocontext:
  99  * Removes AioContext @ctx to the list of AioContext.
 100  */
 101 void unregister_aiocontext(AioContext *ctx);
 102
 103 /*
 104  * bdrv_graph_wrlock:
 105  * Start an exclusive write operation to modify the graph. This means we are
 106  * adding or removing an edge or a node in the block layer graph. Nobody else
 107  * is allowed to access the graph.
 108  *
 109  * Must only be called from outside bdrv_graph_co_rdlock.
 110  *
 111  * The wrlock can only be taken from the main loop, with BQL held, as only the
 112  * main loop is allowed to modify the graph.
 113  */
 114 void no_coroutine_fn TSA_ACQUIRE(graph_lock) TSA_NO_TSA
 115 bdrv_graph_wrlock(void);
 116
 117 /*
 118  * bdrv_graph_wrunlock:
 119  * Write finished, reset global has_writer to 0 and restart
 120  * all readers that are waiting.
 121  */
 122 void no_coroutine_fn TSA_RELEASE(graph_lock) TSA_NO_TSA
 123 bdrv_graph_wrunlock(void);
 124
 125 /*
 126  * bdrv_graph_co_rdlock:
 127  * Read the bs graph. This usually means traversing all nodes in
 128  * the graph, therefore it can't happen while another thread is
 129  * modifying it.
 130  * Increases the reader counter of the current aiocontext,
 131  * and if has_writer is set, it means that the writer is modifying
 132  * the graph, therefore wait in a coroutine queue.
 133  * The writer will then wake this coroutine once it is done.
 134  *
 135  * This lock should be taken from Iothreads (IO_CODE() class of functions)
 136  * because it signals the writer that there are some
 137  * readers currently running, or waits until the current
 138  * write is finished before continuing.
 139  * Calling this function from the Main Loop with BQL held
 140  * is not necessary, since the Main Loop itself is the only
 141  * writer, thus won't be able to read and write at the same time.
 142  * The only exception to that is when we can't take the lock in the
 143  * function/coroutine itself, and need to delegate the caller (usually main
 144  * loop) to take it and wait that the coroutine ends, so that
 145  * we always signal that a reader is running.
 146  */
 147 void coroutine_fn TSA_ACQUIRE_SHARED(graph_lock) TSA_NO_TSA
 148 bdrv_graph_co_rdlock(void);
 149
 150 /*
 151  * bdrv_graph_rdunlock:
 152  * Read terminated, decrease the count of readers in the current aiocontext.
 153  * If the writer is waiting for reads to finish (has_writer == 1), signal
 154  * the writer that we are done via aio_wait_kick() to let it continue.
 155  */
 156 void coroutine_fn TSA_RELEASE_SHARED(graph_lock) TSA_NO_TSA
 157 bdrv_graph_co_rdunlock(void);
 158
 159 /*
 160  * bdrv_graph_rd{un}lock_main_loop:
 161  * Just a placeholder to mark where the graph rdlock should be taken
 162  * in the main loop. It is just asserting that we are not
 163  * in a coroutine and in GLOBAL_STATE_CODE.
 164  */
 165 void TSA_ACQUIRE_SHARED(graph_lock) TSA_NO_TSA
 166 bdrv_graph_rdlock_main_loop(void);
 167
 168 void TSA_RELEASE_SHARED(graph_lock) TSA_NO_TSA
 169 bdrv_graph_rdunlock_main_loop(void);
 170
 171 /*
 172  * assert_bdrv_graph_readable:
 173  * Make sure that the reader is either the main loop,
 174  * or there is at least a reader helding the rdlock.
 175  * In this way an incoming writer is aware of the read and waits.
 176  */
 177 void GRAPH_RDLOCK assert_bdrv_graph_readable(void);
 178
 179 /*
 180  * assert_bdrv_graph_writable:
 181  * Make sure that the writer is the main loop and has set @has_writer,
 182  * so that incoming readers will pause.
 183  */
 184 void GRAPH_WRLOCK assert_bdrv_graph_writable(void);
 185
 186 /*
 187  * Calling this function tells TSA that we know that the lock is effectively
 188  * taken even though we cannot prove it (yet) with GRAPH_RDLOCK. This can be
 189  * useful in intermediate stages of a conversion to using the GRAPH_RDLOCK
 190  * macro.
 191  */
 192 static inline void TSA_ASSERT_SHARED(graph_lock) TSA_NO_TSA
 193 assume_graph_lock(void)
 194 {
 195 }
 196
 197 typedef struct GraphLockable { } GraphLockable;
 198
 199 /*
 200  * In C, compound literals have the lifetime of an automatic variable.
 201  * In C++ it would be different, but then C++ wouldn't need QemuLockable
 202  * either...
 203  */
 204 #define GML_OBJ_() (&(GraphLockable) { })
 205
 206 /*
 207  * This is not marked as TSA_ACQUIRE_SHARED() because TSA doesn't understand the
 208  * cleanup attribute and would therefore complain that the graph is never
 209  * unlocked. TSA_ASSERT_SHARED() makes sure that the following calls know that
 210  * we hold the lock while unlocking is left unchecked.
 211  */
 212 static inline GraphLockable * TSA_ACQUIRE_SHARED(graph_lock) coroutine_fn
 213 graph_lockable_auto_lock(GraphLockable *x)
 214 {
 215     bdrv_graph_co_rdlock();
 216     return x;
 217 }
 218
 219 static inline void TSA_RELEASE_SHARED(graph_lock) coroutine_fn
 220 graph_lockable_auto_unlock(GraphLockable **x)
 221 {
 222     bdrv_graph_co_rdunlock();
 223 }
 224
 225 #define GRAPH_AUTO_UNLOCK __attribute__((cleanup(graph_lockable_auto_unlock)))
 226
 227 /*
 228  * @var is only used to break the loop after the first iteration.
 229  * @unlock_var can't be unlocked and then set to NULL because TSA wants the lock
 230  * to be held at the start of every iteration of the loop.
 231  */
 232 #define WITH_GRAPH_RDLOCK_GUARD_(var)                                         \
 233     for (GraphLockable *unlock_var GRAPH_AUTO_UNLOCK =                        \
 234             graph_lockable_auto_lock(GML_OBJ_()),                             \
 235             *var = unlock_var;                                                \
 236          var;                                                                 \
 237          var = NULL)
 238
 239 #define WITH_GRAPH_RDLOCK_GUARD() \
 240     WITH_GRAPH_RDLOCK_GUARD_(glue(graph_lockable_auto, __COUNTER__))
 241
 242 #define GRAPH_RDLOCK_GUARD(x)                                       \
 243     GraphLockable * GRAPH_AUTO_UNLOCK                               \
 244     glue(graph_lockable_auto, __COUNTER__) G_GNUC_UNUSED =          \
 245             graph_lockable_auto_lock(GML_OBJ_())
 246
 247
 248 typedef struct GraphLockableMainloop { } GraphLockableMainloop;
 249
 250 /*
 251  * In C, compound literals have the lifetime of an automatic variable.
 252  * In C++ it would be different, but then C++ wouldn't need QemuLockable
 253  * either...
 254  */
 255 #define GMLML_OBJ_() (&(GraphLockableMainloop) { })
 256
 257 /*
 258  * This is not marked as TSA_ACQUIRE_SHARED() because TSA doesn't understand the
 259  * cleanup attribute and would therefore complain that the graph is never
 260  * unlocked. TSA_ASSERT_SHARED() makes sure that the following calls know that
 261  * we hold the lock while unlocking is left unchecked.
 262  */
 263 static inline GraphLockableMainloop * TSA_ASSERT_SHARED(graph_lock) TSA_NO_TSA
 264 graph_lockable_auto_lock_mainloop(GraphLockableMainloop *x)
 265 {
 266     bdrv_graph_rdlock_main_loop();
 267     return x;
 268 }
 269
 270 static inline void TSA_NO_TSA
 271 graph_lockable_auto_unlock_mainloop(GraphLockableMainloop *x)
 272 {
 273     bdrv_graph_rdunlock_main_loop();
 274 }
 275
 276 G_DEFINE_AUTOPTR_CLEANUP_FUNC(GraphLockableMainloop,
 277                               graph_lockable_auto_unlock_mainloop)
 278
 279 #define GRAPH_RDLOCK_GUARD_MAINLOOP(x)                              \
 280     g_autoptr(GraphLockableMainloop)                                \
 281     glue(graph_lockable_auto, __COUNTER__) G_GNUC_UNUSED =          \
 282             graph_lockable_auto_lock_mainloop(GMLML_OBJ_())
 283
 284 #endif /* GRAPH_LOCK_H */
 285