block/graph-lock.c

   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
  21 #include "qemu/osdep.h"
  22 #include "qemu/main-loop.h"
  23 #include "block/graph-lock.h"
  24 #include "block/block.h"
  25 #include "block/block_int.h"
  26
  27 /* Dummy lock object to use for Thread Safety Analysis (TSA) */
  28 BdrvGraphLock graph_lock;
  29
  30 /* Protects the list of aiocontext and orphaned_reader_count */
  31 static QemuMutex aio_context_list_lock;
  32
  33 /* Written and read with atomic operations. */
  34 static int has_writer;
  35
  36 /*
  37  * A reader coroutine could move from an AioContext to another.
  38  * If this happens, there is no problem from the point of view of
  39  * counters. The problem is that the total count becomes
  40  * unbalanced if one of the two AioContexts gets deleted.
  41  * The count of readers must remain correct, so the AioContext's
  42  * balance is transferred to this glboal variable.
  43  * Protected by aio_context_list_lock.
  44  */
  45 static uint32_t orphaned_reader_count;
  46
  47 /* Queue of readers waiting for the writer to finish */
  48 static CoQueue reader_queue;
  49
  50 struct BdrvGraphRWlock {
  51     /* How many readers are currently reading the graph. */
  52     uint32_t reader_count;
  53
  54     /*
  55      * List of BdrvGraphRWlock kept in graph-lock.c
  56      * Protected by aio_context_list_lock
  57      */
  58     QTAILQ_ENTRY(BdrvGraphRWlock) next_aio;
  59 };
  60
  61 /*
  62  * List of BdrvGraphRWlock. This list ensures that each BdrvGraphRWlock
  63  * can safely modify only its own counter, avoid reading/writing
  64  * others and thus improving performances by avoiding cacheline bounces.
  65  */
  66 static QTAILQ_HEAD(, BdrvGraphRWlock) aio_context_list =
  67     QTAILQ_HEAD_INITIALIZER(aio_context_list);
  68
  69 static void __attribute__((__constructor__)) bdrv_init_graph_lock(void)
  70 {
  71     qemu_mutex_init(&aio_context_list_lock);
  72     qemu_co_queue_init(&reader_queue);
  73 }
  74
  75 void register_aiocontext(AioContext *ctx)
  76 {
  77     ctx->bdrv_graph = g_new0(BdrvGraphRWlock, 1);
  78     QEMU_LOCK_GUARD(&aio_context_list_lock);
  79     assert(ctx->bdrv_graph->reader_count == 0);
  80     QTAILQ_INSERT_TAIL(&aio_context_list, ctx->bdrv_graph, next_aio);
  81 }
  82
  83 void unregister_aiocontext(AioContext *ctx)
  84 {
  85     QEMU_LOCK_GUARD(&aio_context_list_lock);
  86     orphaned_reader_count += ctx->bdrv_graph->reader_count;
  87     QTAILQ_REMOVE(&aio_context_list, ctx->bdrv_graph, next_aio);
  88     g_free(ctx->bdrv_graph);
  89 }
  90
  91 static uint32_t reader_count(void)
  92 {
  93     BdrvGraphRWlock *brdv_graph;
  94     uint32_t rd;
  95
  96     QEMU_LOCK_GUARD(&aio_context_list_lock);
  97
  98     /* rd can temporarily be negative, but the total will *always* be >= 0 */
  99     rd = orphaned_reader_count;
 100     QTAILQ_FOREACH(brdv_graph, &aio_context_list, next_aio) {
 101         rd += qatomic_read(&brdv_graph->reader_count);
 102     }
 103
 104     /* shouldn't overflow unless there are 2^31 readers */
 105     assert((int32_t)rd >= 0);
 106     return rd;
 107 }
 108
 109 void no_coroutine_fn bdrv_graph_wrlock(void)
 110 {
 111     GLOBAL_STATE_CODE();
 112     assert(!qatomic_read(&has_writer));
 113     assert(!qemu_in_coroutine());
 114
 115     /* Make sure that constantly arriving new I/O doesn't cause starvation */
 116     bdrv_drain_all_begin_nopoll();
 117
 118     /*
 119      * reader_count == 0: this means writer will read has_reader as 1
 120      * reader_count >= 1: we don't know if writer read has_writer == 0 or 1,
 121      *                    but we need to wait.
 122      * Wait by allowing other coroutine (and possible readers) to continue.
 123      */
 124     do {
 125         /*
 126          * has_writer must be 0 while polling, otherwise we get a deadlock if
 127          * any callback involved during AIO_WAIT_WHILE() tries to acquire the
 128          * reader lock.
 129          */
 130         qatomic_set(&has_writer, 0);
 131         AIO_WAIT_WHILE_UNLOCKED(NULL, reader_count() >= 1);
 132         qatomic_set(&has_writer, 1);
 133
 134         /*
 135          * We want to only check reader_count() after has_writer = 1 is visible
 136          * to other threads. That way no more readers can sneak in after we've
 137          * determined reader_count() == 0.
 138          */
 139         smp_mb();
 140     } while (reader_count() >= 1);
 141
 142     bdrv_drain_all_end();
 143 }
 144
 145 void no_coroutine_fn bdrv_graph_wrunlock(void)
 146 {
 147     GLOBAL_STATE_CODE();
 148     assert(qatomic_read(&has_writer));
 149
 150     WITH_QEMU_LOCK_GUARD(&aio_context_list_lock) {
 151         /*
 152          * No need for memory barriers, this works in pair with
 153          * the slow path of rdlock() and both take the lock.
 154          */
 155         qatomic_store_release(&has_writer, 0);
 156
 157         /* Wake up all coroutines that are waiting to read the graph */
 158         qemu_co_enter_all(&reader_queue, &aio_context_list_lock);
 159     }
 160
 161     /*
 162      * Run any BHs that were scheduled during the wrlock section and that
 163      * callers might expect to have finished (in particular, this is important
 164      * for bdrv_schedule_unref()).
 165      *
 166      * Do this only after restarting coroutines so that nested event loops in
 167      * BHs don't deadlock if their condition relies on the coroutine making
 168      * progress.
 169      */
 170     aio_bh_poll(qemu_get_aio_context());
 171 }
 172
 173 void coroutine_fn bdrv_graph_co_rdlock(void)
 174 {
 175     BdrvGraphRWlock *bdrv_graph;
 176     bdrv_graph = qemu_get_current_aio_context()->bdrv_graph;
 177
 178     for (;;) {
 179         qatomic_set(&bdrv_graph->reader_count,
 180                     bdrv_graph->reader_count + 1);
 181         /* make sure writer sees reader_count before we check has_writer */
 182         smp_mb();
 183
 184         /*
 185          * has_writer == 0: this means writer will read reader_count as >= 1
 186          * has_writer == 1: we don't know if writer read reader_count == 0
 187          *                  or > 0, but we need to wait anyways because
 188          *                  it will write.
 189          */
 190         if (!qatomic_read(&has_writer)) {
 191             break;
 192         }
 193
 194         /*
 195          * Synchronize access with reader_count() in bdrv_graph_wrlock().
 196          * Case 1:
 197          * If this critical section gets executed first, reader_count will
 198          * decrease and the reader will go to sleep.
 199          * Then the writer will read reader_count that does not take into
 200          * account this reader, and if there's no other reader it will
 201          * enter the write section.
 202          * Case 2:
 203          * If reader_count() critical section gets executed first,
 204          * then writer will read reader_count >= 1.
 205          * It will wait in AIO_WAIT_WHILE(), but once it releases the lock
 206          * we will enter this critical section and call aio_wait_kick().
 207          */
 208         WITH_QEMU_LOCK_GUARD(&aio_context_list_lock) {
 209             /*
 210              * Additional check when we use the above lock to synchronize
 211              * with bdrv_graph_wrunlock().
 212              * Case 1:
 213              * If this gets executed first, has_writer is still 1, so we reduce
 214              * reader_count and go to sleep.
 215              * Then the writer will set has_writer to 0 and wake up all readers,
 216              * us included.
 217              * Case 2:
 218              * If bdrv_graph_wrunlock() critical section gets executed first,
 219              * then it will set has_writer to 0 and wake up all other readers.
 220              * Then we execute this critical section, and therefore must check
 221              * again for has_writer, otherwise we sleep without any writer
 222              * actually running.
 223              */
 224             if (!qatomic_read(&has_writer)) {
 225                 return;
 226             }
 227
 228             /* slow path where reader sleeps */
 229             bdrv_graph->reader_count--;
 230             aio_wait_kick();
 231             qemu_co_queue_wait(&reader_queue, &aio_context_list_lock);
 232         }
 233     }
 234 }
 235
 236 void coroutine_fn bdrv_graph_co_rdunlock(void)
 237 {
 238     BdrvGraphRWlock *bdrv_graph;
 239     bdrv_graph = qemu_get_current_aio_context()->bdrv_graph;
 240
 241     qatomic_store_release(&bdrv_graph->reader_count,
 242                           bdrv_graph->reader_count - 1);
 243     /* make sure writer sees reader_count before we check has_writer */
 244     smp_mb();
 245
 246     /*
 247      * has_writer == 0: this means reader will read reader_count decreased
 248      * has_writer == 1: we don't know if writer read reader_count old or
 249      *                  new. Therefore, kick again so on next iteration
 250      *                  writer will for sure read the updated value.
 251      */
 252     if (qatomic_read(&has_writer)) {
 253         aio_wait_kick();
 254     }
 255 }
 256
 257 void bdrv_graph_rdlock_main_loop(void)
 258 {
 259     GLOBAL_STATE_CODE();
 260     assert(!qemu_in_coroutine());
 261 }
 262
 263 void bdrv_graph_rdunlock_main_loop(void)
 264 {
 265     GLOBAL_STATE_CODE();
 266     assert(!qemu_in_coroutine());
 267 }
 268
 269 void assert_bdrv_graph_readable(void)
 270 {
 271     /* reader_count() is slow due to aio_context_list_lock lock contention */
 272 #ifdef CONFIG_DEBUG_GRAPH_LOCK
 273     assert(qemu_in_main_thread() || reader_count());
 274 #endif
 275 }
 276
 277 void assert_bdrv_graph_writable(void)
 278 {
 279     assert(qemu_in_main_thread());
 280     assert(qatomic_read(&has_writer));
 281 }