bdev_module: add async fini_start()

fini_start() is called for each bdev module before
iterating over all unclaimed bdevs to unregister them.
This allows bdev modules to behave differently during
each such unregister. Ex. unloading lvol store when
all lvol bdevs on it are unregistered.

Another use of this callback is to unclaim all bdevs
that can be at that point. Especially ones that will
not receive callback due to no bdev registered.
Ex. offline raid bdev, when some underlying bdevs are missing.

fini_start() being synchronous does not help in cases
where to release claim on the bdev, an asynchronous operation
is required. Ex. lvol store with no bdevs present, requires
async lvs unload to be called.

This patch adds async_fini_start flag for the bdev modules,
to be used when async fini_start is required. When done,
bdev module has to call spdk_bdev_module_finish_start_done().

Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Change-Id: I63438b325d4cc53fd236bf9ff143abf6bdd81c49
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/9094
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Aleksey Marchuk <alexeymar@mellanox.com>

CHANGELOG.md

@@ -11,6 +11,9 @@ These functions accept `spdk_bdev_ext_io_opts` structure with extended IO reques
 options, e.g. DMA memory domain which describes data that may belong to another memory domain and
 can't be accessed directly.
 
+Added `async_fini_start` to allow bdev modules to complete the `fini_start` asynchronously,
+with new `spdk_bdev_module_fini_start_done` API.
+
 ### dma
 
 A new library, lib/dma, has been added. This library provides the necessary infrastructure for

include/spdk/bdev_module.h

@@ -151,6 +151,13 @@ struct spdk_bdev_module {
 	 */
 	bool async_fini;
 
+	/**
+	 * Denotes if the fini_start function may complete asynchronously.
+	 * If set to true finishing has to be explicitly completed by calling
+	 * spdk_bdev_module_fini_start_done().
+	 */
+	bool async_fini_start;
+
 	/**
 	 * Fields that are used by the internal bdev subsystem. Bdev modules
 	 *  must not read or write to these fields.
@@ -832,6 +839,15 @@ void spdk_bdev_module_init_done(struct spdk_bdev_module *module);
  */
 void spdk_bdev_module_finish_done(void);
 
+/**
+ * Indicate that the module fini start has completed.
+ *
+ * To be called in response to the fini_start, only if async_fini_start is set.
+ * May be called during fini_start or asynchronously.
+ *
+ */
+void spdk_bdev_module_fini_start_done(void);
+
 /**
  * Add alias to block device names list.
  * Aliases can be add only to registered bdev.

lib/bdev/bdev.c

@@ -1619,11 +1619,56 @@ bdev_finish_unregister_bdevs_iter(void *cb_arg, int bdeverrno)
 	}
 }
 
+static void
+bdev_module_fini_start_iter(void *arg)
+{
+	struct spdk_bdev_module *bdev_module;
+
+	if (!g_resume_bdev_module) {
+		bdev_module = TAILQ_LAST(&g_bdev_mgr.bdev_modules, bdev_module_list);
+	} else {
+		bdev_module = TAILQ_PREV(g_resume_bdev_module, bdev_module_list, internal.tailq);
+	}
+
+	while (bdev_module) {
+		if (bdev_module->async_fini_start) {
+			/* Save our place so we can resume later. We must
+			 * save the variable here, before calling fini_start()
+			 * below, because in some cases the module may immediately
+			 * call spdk_bdev_module_fini_start_done() and re-enter
+			 * this function to continue iterating. */
+			g_resume_bdev_module = bdev_module;
+		}
+
+		if (bdev_module->fini_start) {
+			bdev_module->fini_start();
+		}
+
+		if (bdev_module->async_fini_start) {
+			return;
+		}
+
+		bdev_module = TAILQ_PREV(bdev_module, bdev_module_list, internal.tailq);
+	}
+
+	g_resume_bdev_module = NULL;
+
+	bdev_finish_unregister_bdevs_iter(NULL, 0);
+}
+
+void
+spdk_bdev_module_fini_start_done(void)
+{
+	if (spdk_get_thread() != g_fini_thread) {
+		spdk_thread_send_msg(g_fini_thread, bdev_module_fini_start_iter, NULL);
+	} else {
+		bdev_module_fini_start_iter(NULL);
+	}
+}
+
 void
 spdk_bdev_finish(spdk_bdev_fini_cb cb_fn, void *cb_arg)
 {
-	struct spdk_bdev_module *m;
-
 	assert(cb_fn != NULL);
 
 	g_fini_thread = spdk_get_thread();
@@ -1631,13 +1676,7 @@ spdk_bdev_finish(spdk_bdev_fini_cb cb_fn, void *cb_arg)
 	g_fini_cb_fn = cb_fn;
 	g_fini_cb_arg = cb_arg;
 
-	TAILQ_FOREACH(m, &g_bdev_mgr.bdev_modules, internal.tailq) {
-		if (m->fini_start) {
-			m->fini_start();
-		}
-	}
-
-	bdev_finish_unregister_bdevs_iter(NULL, 0);
+	bdev_module_fini_start_iter(NULL);
 }
 
 struct spdk_bdev_io *

lib/bdev/spdk_bdev.map

@@ -105,6 +105,7 @@
 	spdk_bdev_module_examine_done;
 	spdk_bdev_module_init_done;
 	spdk_bdev_module_finish_done;
+	spdk_bdev_module_fini_start_done;
 	spdk_bdev_module_claim_bdev;
 	spdk_bdev_module_release_bdev;
 	spdk_bdev_alias_add;

module/blob/bdev/Makefile

@@ -34,7 +34,7 @@
 SPDK_ROOT_DIR := $(abspath $(CURDIR)/../../..)
 include $(SPDK_ROOT_DIR)/mk/spdk.common.mk
 
-SO_VER := 5
+SO_VER := 6
 SO_MINOR := 0
 
 C_SRCS = blob_bdev.c