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