If we failed to read /proc/meminfo, just run the shrinkers.

[bcachefs-tools-debian] / linux / blkdev.c

diff --git a/linux/blkdev.c b/linux/blkdev.c
index 8139e18d3b95999d8604936b3b31c0471c9d75d5..709e770345f925e31101a82ea3bd5366525bb848 100644 (file)
--- a/linux/blkdev.c
+++ b/linux/blkdev.c
@@ -10,7 +10,9 @@
 
 #include <libaio.h>
 
+#ifdef CONFIG_VALGRIND
 #include <valgrind/memcheck.h>
+#endif
 
 #include <linux/bio.h>
 #include <linux/blkdev.h>
@@ -21,6 +23,7 @@
 #include "tools-util.h"
 
 static io_context_t aio_ctx;
+static atomic_t running_requests;
 
 void generic_make_request(struct bio *bio)
 {
@@ -56,9 +59,11 @@ void generic_make_request(struct bio *bio)
                        .iov_len = len,
                };
 
+#ifdef CONFIG_VALGRIND
                /* To be pedantic it should only be on IO completion. */
                if (bio_op(bio) == REQ_OP_READ)
                        VALGRIND_MAKE_MEM_DEFINED(start, len);
+#endif
        }
 
        struct iocb iocb = {
@@ -75,6 +80,7 @@ void generic_make_request(struct bio *bio)
                iocb.u.v.nr             = i;
                iocb.u.v.offset         = bio->bi_iter.bi_sector << 9;
 
+               atomic_inc(&running_requests);
                ret = io_submit(aio_ctx, 1, &iocbp);
                if (ret != 1)
                        die("io_submit err: %s", strerror(-ret));
@@ -85,6 +91,7 @@ void generic_make_request(struct bio *bio)
                iocb.u.v.nr             = i;
                iocb.u.v.offset         = bio->bi_iter.bi_sector << 9;
 
+               atomic_inc(&running_requests);
                ret = io_submit(aio_ctx, 1, &iocbp);
                if (ret != 1)
                        die("io_submit err: %s", strerror(-ret));
@@ -232,8 +239,9 @@ static int aio_completion_thread(void *arg)
 {
        struct io_event events[8], *ev;
        int ret;
+       bool stop = false;
 
-       while (1) {
+       while (!stop) {
                ret = io_getevents(aio_ctx, 1, ARRAY_SIZE(events),
                                   events, NULL);
 
@@ -245,16 +253,26 @@ static int aio_completion_thread(void *arg)
                for (ev = events; ev < events + ret; ev++) {
                        struct bio *bio = (struct bio *) ev->data;
 
+                       /* This should only happen during blkdev_cleanup() */
+                       if (!bio) {
+                               BUG_ON(atomic_read(&running_requests) != 0);
+                               stop = true;
+                               continue;
+                       }
+
                        if (ev->res != bio->bi_iter.bi_size)
                                bio->bi_status = BLK_STS_IOERR;
 
                        bio_endio(bio);
+                       atomic_dec(&running_requests);
                }
        }
 
        return 0;
 }
 
+static struct task_struct *aio_task = NULL;
+
 __attribute__((constructor(102)))
 static void blkdev_init(void)
 {
@@ -265,4 +283,41 @@ static void blkdev_init(void)
 
        p = kthread_run(aio_completion_thread, NULL, "aio_completion");
        BUG_ON(IS_ERR(p));
+
+       aio_task = p;
+}
+
+__attribute__((destructor(102)))
+static void blkdev_cleanup(void)
+{
+       struct task_struct *p = NULL;
+       swap(aio_task, p);
+       get_task_struct(p);
+
+       /* I mean, really?! IO_CMD_NOOP is even defined, but not implemented. */
+       int fds[2];
+       int ret = pipe(fds);
+       if (ret != 0)
+               die("pipe err: %s", strerror(ret));
+
+       /* Wake up the completion thread with spurious work. */
+       int junk = 0;
+       struct iocb iocb = {
+               .aio_lio_opcode = IO_CMD_PWRITE,
+               .data = NULL, /* Signal to stop */
+               .aio_fildes = fds[1],
+               .u.c.buf = &junk,
+               .u.c.nbytes = 1,
+       }, *iocbp = &iocb;
+       ret = io_submit(aio_ctx, 1, &iocbp);
+       if (ret != 1)
+               die("io_submit cleanup err: %s", strerror(-ret));
+
+       ret = kthread_stop(p);
+       BUG_ON(ret);
+
+       put_task_struct(p);
+
+       close(fds[0]);
+       close(fds[1]);
 }