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 temporarly 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 bdrv_graph_wrlock(void)
 110 {
 111     GLOBAL_STATE_CODE();
 112     assert(!qatomic_read(&has_writer));
 113
 114     /* Make sure that constantly arriving new I/O doesn't cause starvation */
 115     bdrv_drain_all_begin_nopoll();
 116
 117     /*
 118      * reader_count == 0: this means writer will read has_reader as 1
 119      * reader_count >= 1: we don't know if writer read has_writer == 0 or 1,
 120      *                    but we need to wait.
 121      * Wait by allowing other coroutine (and possible readers) to continue.
 122      */
 123     do {
 124         /*
 125          * has_writer must be 0 while polling, otherwise we get a deadlock if
 126          * any callback involved during AIO_WAIT_WHILE() tries to acquire the
 127          * reader lock.
 128          */
 129         qatomic_set(&has_writer, 0);
 130         AIO_WAIT_WHILE(qemu_get_aio_context(), reader_count() >= 1);
 131         qatomic_set(&has_writer, 1);
 132
 133         /*
 134          * We want to only check reader_count() after has_writer = 1 is visible
 135          * to other threads. That way no more readers can sneak in after we've
 136          * determined reader_count() == 0.
 137          */
 138         smp_mb();
 139     } while (reader_count() >= 1);
 140
 141     bdrv_drain_all_end();
 142 }
 143
 144 void bdrv_graph_wrunlock(void)
 145 {
 146     GLOBAL_STATE_CODE();
 147     QEMU_LOCK_GUARD(&aio_context_list_lock);
 148     assert(qatomic_read(&has_writer));
 149
 150     /*
 151      * No need for memory barriers, this works in pair with
 152      * the slow path of rdlock() and both take the lock.
 153      */
 154     qatomic_store_release(&has_writer, 0);
 155
 156     /* Wake up all coroutine that are waiting to read the graph */
 157     qemu_co_enter_all(&reader_queue, &aio_context_list_lock);
 158 }
 159
 160 void coroutine_fn bdrv_graph_co_rdlock(void)
 161 {
 162     BdrvGraphRWlock *bdrv_graph;
 163     bdrv_graph = qemu_get_current_aio_context()->bdrv_graph;
 164
 165     /* Do not lock if in main thread */
 166     if (qemu_in_main_thread()) {
 167         return;
 168     }
 169
 170     for (;;) {
 171         qatomic_set(&bdrv_graph->reader_count,
 172                     bdrv_graph->reader_count + 1);
 173         /* make sure writer sees reader_count before we check has_writer */
 174         smp_mb();
 175
 176         /*
 177          * has_writer == 0: this means writer will read reader_count as >= 1
 178          * has_writer == 1: we don't know if writer read reader_count == 0
 179          *                  or > 0, but we need to wait anyways because
 180          *                  it will write.
 181          */
 182         if (!qatomic_read(&has_writer)) {
 183             break;
 184         }
 185
 186         /*
 187          * Synchronize access with reader_count() in bdrv_graph_wrlock().
 188          * Case 1:
 189          * If this critical section gets executed first, reader_count will
 190          * decrease and the reader will go to sleep.
 191          * Then the writer will read reader_count that does not take into
 192          * account this reader, and if there's no other reader it will
 193          * enter the write section.
 194          * Case 2:
 195          * If reader_count() critical section gets executed first,
 196          * then writer will read reader_count >= 1.
 197          * It will wait in AIO_WAIT_WHILE(), but once it releases the lock
 198          * we will enter this critical section and call aio_wait_kick().
 199          */
 200         WITH_QEMU_LOCK_GUARD(&aio_context_list_lock) {
 201             /*
 202              * Additional check when we use the above lock to synchronize
 203              * with bdrv_graph_wrunlock().
 204              * Case 1:
 205              * If this gets executed first, has_writer is still 1, so we reduce
 206              * reader_count and go to sleep.
 207              * Then the writer will set has_writer to 0 and wake up all readers,
 208              * us included.
 209              * Case 2:
 210              * If bdrv_graph_wrunlock() critical section gets executed first,
 211              * then it will set has_writer to 0 and wake up all other readers.
 212              * Then we execute this critical section, and therefore must check
 213              * again for has_writer, otherwise we sleep without any writer
 214              * actually running.
 215              */
 216             if (!qatomic_read(&has_writer)) {
 217                 return;
 218             }
 219
 220             /* slow path where reader sleeps */
 221             bdrv_graph->reader_count--;
 222             aio_wait_kick();
 223             qemu_co_queue_wait(&reader_queue, &aio_context_list_lock);
 224         }
 225     }
 226 }
 227
 228 void coroutine_fn bdrv_graph_co_rdunlock(void)
 229 {
 230     BdrvGraphRWlock *bdrv_graph;
 231     bdrv_graph = qemu_get_current_aio_context()->bdrv_graph;
 232
 233     /* Do not lock if in main thread */
 234     if (qemu_in_main_thread()) {
 235         return;
 236     }
 237
 238     qatomic_store_release(&bdrv_graph->reader_count,
 239                           bdrv_graph->reader_count - 1);
 240     /* make sure writer sees reader_count before we check has_writer */
 241     smp_mb();
 242
 243     /*
 244      * has_writer == 0: this means reader will read reader_count decreased
 245      * has_writer == 1: we don't know if writer read reader_count old or
 246      *                  new. Therefore, kick again so on next iteration
 247      *                  writer will for sure read the updated value.
 248      */
 249     if (qatomic_read(&has_writer)) {
 250         aio_wait_kick();
 251     }
 252 }
 253
 254 void bdrv_graph_rdlock_main_loop(void)
 255 {
 256     GLOBAL_STATE_CODE();
 257     assert(!qemu_in_coroutine());
 258 }
 259
 260 void bdrv_graph_rdunlock_main_loop(void)
 261 {
 262     GLOBAL_STATE_CODE();
 263     assert(!qemu_in_coroutine());
 264 }
 265
 266 void assert_bdrv_graph_readable(void)
 267 {
 268     assert(qemu_in_main_thread() || reader_count());
 269 }
 270
 271 void assert_bdrv_graph_writable(void)
 272 {
 273     assert(qemu_in_main_thread());
 274     assert(qatomic_read(&has_writer));
 275 }