On Wed, May 07, 2025 at 03:49:41PM -0600, Uday Shankar wrote:
> Add support in kublk for decoupled ublk_queues and ublk server threads.
> kublk now has two modes of operation:
>
> - (preexisting mode) threads and queues are paired 1:1, and each thread
> services all the I/Os of one queue
> - (new mode) thread and queue counts are independently configurable.
> threads service I/Os in a way that balances load across threads even
> if load is not balanced over queues. requires passing --round_robin
>
> The new mode of operation is exercised by the new test_generic_08, which
> issues I/O against a single queue and verifies that each of the 8 ublk
> server threads handles exactly 1/8 of the total I/O count. Under the old
> mode of operation (i.e. without --round_robin), all I/O goes to one ublk
> server thread, and the test fails.
>
> Signed-off-by: Uday Shankar <ushan...@purestorage.com>
> ---
> tools/testing/selftests/ublk/Makefile | 1 +
> tools/testing/selftests/ublk/file_backed.c | 4 +-
> tools/testing/selftests/ublk/kublk.c | 96
> ++++++++++++++++++----
> tools/testing/selftests/ublk/kublk.h | 3 +
> tools/testing/selftests/ublk/null.c | 4 +-
> tools/testing/selftests/ublk/stripe.c | 4 +-
> tools/testing/selftests/ublk/test_generic_08.sh | 61 ++++++++++++++
> .../selftests/ublk/trace/count_ios_per_tid.bt | 9 ++
> 8 files changed, 160 insertions(+), 22 deletions(-)
>
> diff --git a/tools/testing/selftests/ublk/Makefile
> b/tools/testing/selftests/ublk/Makefile
> index
> e2e7b1e52a06252f626df5606221d17e3106d0d3..f8579019f0d1f9185df098440611c3b75087073b
> 100644
> --- a/tools/testing/selftests/ublk/Makefile
> +++ b/tools/testing/selftests/ublk/Makefile
> @@ -14,6 +14,7 @@ TEST_PROGS += test_generic_04.sh
> TEST_PROGS += test_generic_05.sh
> TEST_PROGS += test_generic_06.sh
> TEST_PROGS += test_generic_07.sh
> +TEST_PROGS += test_generic_08.sh
>
> TEST_PROGS += test_null_01.sh
> TEST_PROGS += test_null_02.sh
> diff --git a/tools/testing/selftests/ublk/file_backed.c
> b/tools/testing/selftests/ublk/file_backed.c
> index
> 563f11a21604bbf5b9531f69f806d09cdd785960..72e22f54f7fe60d26096ace4eb4001987fbe7b15
> 100644
> --- a/tools/testing/selftests/ublk/file_backed.c
> +++ b/tools/testing/selftests/ublk/file_backed.c
> @@ -50,7 +50,7 @@ static int loop_queue_tgt_rw_io(struct ublk_queue *q, const
> struct ublksrv_io_de
>
> ublk_io_alloc_sqes(ublk_get_io(q, tag), sqe, 3);
>
> - io_uring_prep_buf_register(sqe[0], 0, tag, q->q_id, tag);
> + io_uring_prep_buf_register(sqe[0], 0, tag, q->q_id, ublk_get_io(q,
> tag)->buf_index);
->buf_index can be calculated runtime by adding helper of ublk_io_buf_idx(io),
so
the extra field can be avoided, CPU is fast than memory.
> sqe[0]->flags |= IOSQE_CQE_SKIP_SUCCESS | IOSQE_IO_HARDLINK;
> sqe[0]->user_data = build_user_data(tag,
> ublk_cmd_op_nr(sqe[0]->cmd_op), 0, q->q_id, 1);
> @@ -62,7 +62,7 @@ static int loop_queue_tgt_rw_io(struct ublk_queue *q, const
> struct ublksrv_io_de
> sqe[1]->flags |= IOSQE_FIXED_FILE | IOSQE_IO_HARDLINK;
> sqe[1]->user_data = build_user_data(tag, ublk_op, 0, q->q_id, 1);
>
> - io_uring_prep_buf_unregister(sqe[2], 0, tag, q->q_id, tag);
> + io_uring_prep_buf_unregister(sqe[2], 0, tag, q->q_id, ublk_get_io(q,
> tag)->buf_index);
> sqe[2]->user_data = build_user_data(tag,
> ublk_cmd_op_nr(sqe[2]->cmd_op), 0, q->q_id, 1);
>
> return 2;
> diff --git a/tools/testing/selftests/ublk/kublk.c
> b/tools/testing/selftests/ublk/kublk.c
> index
> 313689f94cd6361a9a0f4b9257085b2a62bc8b8c..27046bb6a13b99879ad164ff8eaabeba57e17387
> 100644
> --- a/tools/testing/selftests/ublk/kublk.c
> +++ b/tools/testing/selftests/ublk/kublk.c
> @@ -478,8 +478,11 @@ static int ublk_thread_init(struct ublk_thread *t)
> }
>
> if (dev->dev_info.flags & UBLK_F_SUPPORT_ZERO_COPY) {
> + unsigned nr_ios = dev->dev_info.queue_depth *
> dev->dev_info.nr_hw_queues;
> + unsigned max_nr_ios_per_thread = nr_ios / dev->nthreads;
> + max_nr_ios_per_thread += !!(nr_ios % dev->nthreads);
> ret = io_uring_register_buffers_sparse(
> - &t->ring, dev->dev_info.queue_depth);
> + &t->ring, max_nr_ios_per_thread);
> if (ret) {
> ublk_err("ublk dev %d thread %d register spare buffers
> failed %d",
> dev->dev_info.dev_id, t->idx, ret);
> @@ -612,18 +615,42 @@ int ublk_queue_io_cmd(struct ublk_io *io)
>
> static void ublk_submit_fetch_commands(struct ublk_thread *t)
> {
> - /*
> - * Service exclusively the queue whose q_id matches our thread
> - * index. This may change in the future.
> - */
> - struct ublk_queue *q = &t->dev->q[t->idx];
> + struct ublk_queue *q;
> struct ublk_io *io;
> - int i = 0;
> + int i = 0, j = 0;
>
> - for (i = 0; i < q->q_depth; i++) {
> - io = &q->ios[i];
> - io->t = t;
> - ublk_queue_io_cmd(io);
> + if (t->dev->dev_info.flags & UBLK_F_RR_TAGS) {
You shouldn't depend on the generic feature of UBLK_F_RR_TAGS, which can be
used for non-io-task too.
Here you need one helper of ublk_use_io_task(), or sort of flags.
> + /*
> + * Lexicographically order all the (qid,tag) pairs, with
> + * qid taking priority, and give this thread every Nth
> + * entry, where N is the total number of threads. The
> + * offset is controlled by the thread index. This takes
> + * load which may be imbalanced across the queues and
> + * balances it across the threads.
> + */
> + const struct ublksrv_ctrl_dev_info *dinfo = &t->dev->dev_info;
> + int nr_ios = dinfo->nr_hw_queues * dinfo->queue_depth;
> + for (i = t->idx; i < nr_ios; i += t->dev->nthreads, j++) {
> + int q_id = i / dinfo->queue_depth;
> + int tag = i % dinfo->queue_depth;
I understand UBLK_F_RR_TAGS means that IO with adjacent tag should be
handled locally, but the above actually does the opposite, only IOs with
non-adjacent tags are handled in same pthread, can you explain a bit
why UBLK_F_RR_TAGS helps for this way?
> + q = &t->dev->q[q_id];
> + io = &q->ios[tag];
> + io->t = t;
> + io->buf_index = j;
> + ublk_queue_io_cmd(io);
> + }
> + } else {
> + /*
> + * Service exclusively the queue whose q_id matches our
> + * thread index.
> + */
> + struct ublk_queue *q = &t->dev->q[t->idx];
> + for (i = 0; i < q->q_depth; i++) {
> + io = &q->ios[i];
> + io->t = t;
> + io->buf_index = i;
> + ublk_queue_io_cmd(io);
> + }
> }
> }
>
> @@ -778,7 +805,8 @@ static void *ublk_io_handler_fn(void *data)
> return NULL;
> }
> /* IO perf is sensitive with queue pthread affinity on NUMA machine*/
> - ublk_thread_set_sched_affinity(t, info->affinity);
> + if (info->affinity)
> + ublk_thread_set_sched_affinity(t, info->affinity);
> sem_post(info->ready);
>
> ublk_dbg(UBLK_DBG_THREAD, "tid %d: ublk dev %d thread %u started\n",
> @@ -844,7 +872,7 @@ static int ublk_start_daemon(const struct dev_ctx *ctx,
> struct ublk_dev *dev)
>
> ublk_dbg(UBLK_DBG_DEV, "%s enter\n", __func__);
>
> - tinfo = calloc(sizeof(struct ublk_thread_info), dinfo->nr_hw_queues);
> + tinfo = calloc(sizeof(struct ublk_thread_info), dev->nthreads);
> if (!tinfo)
> return -ENOMEM;
>
> @@ -867,17 +895,24 @@ static int ublk_start_daemon(const struct dev_ctx *ctx,
> struct ublk_dev *dev)
> dinfo->dev_id, i);
> goto fail;
> }
> + }
>
> + for (i = 0; i < dev->nthreads; i++) {
> tinfo[i].dev = dev;
> tinfo[i].idx = i;
> tinfo[i].ready = &ready;
> - tinfo[i].affinity = &affinity_buf[i];
> + /*
> + * If threads are not tied to queues, setting thread
> + * affinity based on queue affinity makes no sense.
> + */
> + if (!(dinfo->flags & UBLK_F_RR_TAGS))
> + tinfo[i].affinity = &affinity_buf[i];
> pthread_create(&dev->threads[i].thread, NULL,
> ublk_io_handler_fn,
> &tinfo[i]);
> }
>
> - for (i = 0; i < dinfo->nr_hw_queues; i++)
> + for (i = 0; i < dev->nthreads; i++)
> sem_wait(&ready);
> free(tinfo);
> free(affinity_buf);
> @@ -901,7 +936,7 @@ static int ublk_start_daemon(const struct dev_ctx *ctx,
> struct ublk_dev *dev)
> ublk_send_dev_event(ctx, dev, dev->dev_info.dev_id);
>
> /* wait until we are terminated */
> - for (i = 0; i < dinfo->nr_hw_queues; i++)
> + for (i = 0; i < dev->nthreads; i++)
> pthread_join(dev->threads[i].thread, &thread_ret);
> fail:
> for (i = 0; i < dinfo->nr_hw_queues; i++)
> @@ -1011,6 +1046,7 @@ static int ublk_stop_io_daemon(const struct ublk_dev
> *dev)
>
> static int __cmd_dev_add(const struct dev_ctx *ctx)
> {
> + unsigned nthreads = ctx->nthreads;
> unsigned nr_queues = ctx->nr_hw_queues;
> const char *tgt_type = ctx->tgt_type;
> unsigned depth = ctx->queue_depth;
> @@ -1034,6 +1070,23 @@ static int __cmd_dev_add(const struct dev_ctx *ctx)
> return -EINVAL;
> }
>
> + /* default to 1:1 threads:queues if nthreads is unspecified */
> + if (nthreads == -1)
> + nthreads = nr_queues;
Maybe we can start 1:1 for nrthreads == 0, which looks more readable,
and you needn't to set -1 default.
> +
> + if (nthreads > UBLK_MAX_THREADS) {
> + ublk_err("%s: %u is too many threads (max %u)\n",
> + __func__, nthreads, UBLK_MAX_THREADS);
> + return -EINVAL;
> + }
> +
> + if (nthreads != nr_queues && !(ctx->flags & UBLK_F_RR_TAGS)) {
> + ublk_err("%s: threads %u must be same as queues %u if "
> + "not using round robin\n",
> + __func__, nthreads, nr_queues);
> + return -EINVAL;
> + }
> +
> dev = ublk_ctrl_init();
> if (!dev) {
> ublk_err("%s: can't alloc dev id %d, type %s\n",
> @@ -1054,6 +1107,7 @@ static int __cmd_dev_add(const struct dev_ctx *ctx)
> info->nr_hw_queues = nr_queues;
> info->queue_depth = depth;
> info->flags = ctx->flags;
> + dev->nthreads = nthreads;
> dev->tgt.ops = ops;
> dev->tgt.sq_depth = depth;
> dev->tgt.cq_depth = depth;
> @@ -1249,6 +1303,7 @@ static int cmd_dev_get_features(void)
> [const_ilog2(UBLK_F_USER_COPY)] = "USER_COPY",
> [const_ilog2(UBLK_F_ZONED)] = "ZONED",
> [const_ilog2(UBLK_F_USER_RECOVERY_FAIL_IO)] =
> "RECOVERY_FAIL_IO",
> + [const_ilog2(UBLK_F_RR_TAGS)] = "RR_TAGS",
> };
> struct ublk_dev *dev;
> __u64 features = 0;
> @@ -1290,8 +1345,10 @@ static void __cmd_create_help(char *exe, bool recovery)
> exe, recovery ? "recover" : "add");
> printf("\t[--foreground] [--quiet] [-z] [--debug_mask mask] [-r 0|1 ]
> [-g]\n");
> printf("\t[-e 0|1 ] [-i 0|1]\n");
> + printf("\t[--nthreads threads] [--round_robin]\n");
> printf("\t[target options] [backfile1] [backfile2] ...\n");
> printf("\tdefault: nr_queues=2(max 32), depth=128(max 1024),
> dev_id=-1(auto allocation)\n");
> + printf("\tdefault: nthreads=nr_queues");
>
> for (i = 0; i < sizeof(tgt_ops_list) / sizeof(tgt_ops_list[0]); i++) {
> const struct ublk_tgt_ops *ops = tgt_ops_list[i];
> @@ -1343,6 +1400,8 @@ int main(int argc, char *argv[])
> { "recovery_fail_io", 1, NULL, 'e'},
> { "recovery_reissue", 1, NULL, 'i'},
> { "get_data", 1, NULL, 'g'},
> + { "nthreads", 1, NULL, 0 },
> + { "round_robin", 0, NULL, 0 },
> { 0, 0, 0, 0 }
> };
> const struct ublk_tgt_ops *ops = NULL;
> @@ -1351,6 +1410,7 @@ int main(int argc, char *argv[])
> struct dev_ctx ctx = {
> .queue_depth = 128,
> .nr_hw_queues = 2,
> + .nthreads = -1,
> .dev_id = -1,
> .tgt_type = "unknown",
> };
> @@ -1411,6 +1471,10 @@ int main(int argc, char *argv[])
> ublk_dbg_mask = 0;
> if (!strcmp(longopts[option_idx].name, "foreground"))
> ctx.fg = 1;
> + if (!strcmp(longopts[option_idx].name, "nthreads"))
> + ctx.nthreads = strtol(optarg, NULL, 10);
> + if (!strcmp(longopts[option_idx].name, "round_robin"))
> + ctx.flags |= UBLK_F_RR_TAGS;
maybe `--io_task` is more readable, and you can always enable UBLK_F_RR_TAGS
for this '--io_task' ublk server feature.
Thanks,
Ming
