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
  77 /*
  78  * TSA annotations are not part of function types, so checks are defeated when
  79  * using a function pointer. As a workaround, annotate function pointers with
  80  * this macro that will require that the lock is at least taken while reading
  81  * the pointer. In most cases this is equivalent to actually protecting the
  82  * function call.
  83  */
  84 #define GRAPH_RDLOCK_PTR TSA_GUARDED_BY(graph_lock)
  85 #define GRAPH_WRLOCK_PTR TSA_GUARDED_BY(graph_lock)
  86
  87 /*
  88  * register_aiocontext:
  89  * Add AioContext @ctx to the list of AioContext.
  90  * This list is used to obtain the total number of readers
  91  * currently running the graph.
  92  */
  93 void register_aiocontext(AioContext *ctx);
  94
  95 /*
  96  * unregister_aiocontext:
  97  * Removes AioContext @ctx to the list of AioContext.
  98  */
  99 void unregister_aiocontext(AioContext *ctx);
 100
 101 /*
 102  * bdrv_graph_wrlock:
 103  * Start an exclusive write operation to modify the graph. This means we are
 104  * adding or removing an edge or a node in the block layer graph. Nobody else
 105  * is allowed to access the graph.
 106  *
 107  * Must only be called from outside bdrv_graph_co_rdlock.
 108  *
 109  * The wrlock can only be taken from the main loop, with BQL held, as only the
 110  * main loop is allowed to modify the graph.
 111  *
 112  * This function polls. Callers must not hold the lock of any AioContext other
 113  * than the current one.
 114  */
 115 void bdrv_graph_wrlock(void) TSA_ACQUIRE(graph_lock) TSA_NO_TSA;
 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 bdrv_graph_wrunlock(void) TSA_RELEASE(graph_lock) TSA_NO_TSA;
 123
 124 /*
 125  * bdrv_graph_co_rdlock:
 126  * Read the bs graph. This usually means traversing all nodes in
 127  * the graph, therefore it can't happen while another thread is
 128  * modifying it.
 129  * Increases the reader counter of the current aiocontext,
 130  * and if has_writer is set, it means that the writer is modifying
 131  * the graph, therefore wait in a coroutine queue.
 132  * The writer will then wake this coroutine once it is done.
 133  *
 134  * This lock should be taken from Iothreads (IO_CODE() class of functions)
 135  * because it signals the writer that there are some
 136  * readers currently running, or waits until the current
 137  * write is finished before continuing.
 138  * Calling this function from the Main Loop with BQL held
 139  * is not necessary, since the Main Loop itself is the only
 140  * writer, thus won't be able to read and write at the same time.
 141  * The only exception to that is when we can't take the lock in the
 142  * function/coroutine itself, and need to delegate the caller (usually main
 143  * loop) to take it and wait that the coroutine ends, so that
 144  * we always signal that a reader is running.
 145  */
 146 void coroutine_fn TSA_ACQUIRE_SHARED(graph_lock) TSA_NO_TSA
 147 bdrv_graph_co_rdlock(void);
 148
 149 /*
 150  * bdrv_graph_rdunlock:
 151  * Read terminated, decrease the count of readers in the current aiocontext.
 152  * If the writer is waiting for reads to finish (has_writer == 1), signal
 153  * the writer that we are done via aio_wait_kick() to let it continue.
 154  */
 155 void coroutine_fn TSA_RELEASE_SHARED(graph_lock) TSA_NO_TSA
 156 bdrv_graph_co_rdunlock(void);
 157
 158 /*
 159  * bdrv_graph_rd{un}lock_main_loop:
 160  * Just a placeholder to mark where the graph rdlock should be taken
 161  * in the main loop. It is just asserting that we are not
 162  * in a coroutine and in GLOBAL_STATE_CODE.
 163  */
 164 void TSA_ACQUIRE_SHARED(graph_lock) TSA_NO_TSA
 165 bdrv_graph_rdlock_main_loop(void);
 166
 167 void TSA_RELEASE_SHARED(graph_lock) TSA_NO_TSA
 168 bdrv_graph_rdunlock_main_loop(void);
 169
 170 /*
 171  * assert_bdrv_graph_readable:
 172  * Make sure that the reader is either the main loop,
 173  * or there is at least a reader helding the rdlock.
 174  * In this way an incoming writer is aware of the read and waits.
 175  */
 176 void GRAPH_RDLOCK assert_bdrv_graph_readable(void);
 177
 178 /*
 179  * assert_bdrv_graph_writable:
 180  * Make sure that the writer is the main loop and has set @has_writer,
 181  * so that incoming readers will pause.
 182  */
 183 void GRAPH_WRLOCK assert_bdrv_graph_writable(void);
 184
 185 /*
 186  * Calling this function tells TSA that we know that the lock is effectively
 187  * taken even though we cannot prove it (yet) with GRAPH_RDLOCK. This can be
 188  * useful in intermediate stages of a conversion to using the GRAPH_RDLOCK
 189  * macro.
 190  */
 191 static inline void TSA_ASSERT_SHARED(graph_lock) TSA_NO_TSA
 192 assume_graph_lock(void)
 193 {
 194 }
 195
 196 typedef struct GraphLockable { } GraphLockable;
 197
 198 /*
 199  * In C, compound literals have the lifetime of an automatic variable.
 200  * In C++ it would be different, but then C++ wouldn't need QemuLockable
 201  * either...
 202  */
 203 #define GML_OBJ_() (&(GraphLockable) { })
 204
 205 /*
 206  * This is not marked as TSA_ACQUIRE() because TSA doesn't understand the
 207  * cleanup attribute and would therefore complain that the graph is never
 208  * unlocked. TSA_ASSERT() makes sure that the following calls know that we
 209  * hold the lock while unlocking is left unchecked.
 210  */
 211 static inline GraphLockable * TSA_ASSERT(graph_lock) TSA_NO_TSA
 212 graph_lockable_auto_lock(GraphLockable *x)
 213 {
 214     bdrv_graph_co_rdlock();
 215     return x;
 216 }
 217
 218 static inline void TSA_NO_TSA
 219 graph_lockable_auto_unlock(GraphLockable *x)
 220 {
 221     bdrv_graph_co_rdunlock();
 222 }
 223
 224 G_DEFINE_AUTOPTR_CLEANUP_FUNC(GraphLockable, graph_lockable_auto_unlock)
 225
 226 #define WITH_GRAPH_RDLOCK_GUARD_(var)                                         \
 227     for (g_autoptr(GraphLockable) var = graph_lockable_auto_lock(GML_OBJ_()); \
 228          var;                                                                 \
 229          graph_lockable_auto_unlock(var), var = NULL)
 230
 231 #define WITH_GRAPH_RDLOCK_GUARD() \
 232     WITH_GRAPH_RDLOCK_GUARD_(glue(graph_lockable_auto, __COUNTER__))
 233
 234 #define GRAPH_RDLOCK_GUARD(x)                                       \
 235     g_autoptr(GraphLockable)                                        \
 236     glue(graph_lockable_auto, __COUNTER__) G_GNUC_UNUSED =          \
 237             graph_lockable_auto_lock(GML_OBJ_())
 238
 239
 240 typedef struct GraphLockableMainloop { } GraphLockableMainloop;
 241
 242 /*
 243  * In C, compound literals have the lifetime of an automatic variable.
 244  * In C++ it would be different, but then C++ wouldn't need QemuLockable
 245  * either...
 246  */
 247 #define GMLML_OBJ_() (&(GraphLockableMainloop) { })
 248
 249 /*
 250  * This is not marked as TSA_ACQUIRE() because TSA doesn't understand the
 251  * cleanup attribute and would therefore complain that the graph is never
 252  * unlocked. TSA_ASSERT() makes sure that the following calls know that we
 253  * hold the lock while unlocking is left unchecked.
 254  */
 255 static inline GraphLockableMainloop * TSA_ASSERT(graph_lock) TSA_NO_TSA
 256 graph_lockable_auto_lock_mainloop(GraphLockableMainloop *x)
 257 {
 258     bdrv_graph_rdlock_main_loop();
 259     return x;
 260 }
 261
 262 static inline void TSA_NO_TSA
 263 graph_lockable_auto_unlock_mainloop(GraphLockableMainloop *x)
 264 {
 265     bdrv_graph_rdunlock_main_loop();
 266 }
 267
 268 G_DEFINE_AUTOPTR_CLEANUP_FUNC(GraphLockableMainloop,
 269                               graph_lockable_auto_unlock_mainloop)
 270
 271 #define GRAPH_RDLOCK_GUARD_MAINLOOP(x)                              \
 272     g_autoptr(GraphLockableMainloop)                                \
 273     glue(graph_lockable_auto, __COUNTER__) G_GNUC_UNUSED =          \
 274             graph_lockable_auto_lock_mainloop(GMLML_OBJ_())
 275
 276 #endif /* GRAPH_LOCK_H */
 277