firewire: core: allocate workqueue for AR/AT request/response contexts

Some tasklets (softIRQs) are still used as bottom-halves to handle
events for 1394 OHCI AR/AT contexts. However, using softIRQs for IRQ
bottom halves is generally discouraged today.

This commit adds a per-fw_card workqueue to accommodate the behaviour
specified by the 1394 OHCI specification.

According to the 1394 OHCI specification, system memory pages are
reserved for each asynchronous DMA context. This allows concurrent
operation across contexts. In the 1394 OHCI PCI driver implementation,
the hardware generates IRQs either upon receiving asynchronous packets
from other nodes (incoming) or after completing transmission to them
(outgoing). These independent events can occur in the same transmission
cycle, therefore the max_active parameter for the workqueue is set to the
total number of AR/AT contexts (=4). The WQ_UNBOUND flag is used to
allow the work to be scheduled on any available core, since there is
little CPU cache affinity benefit for the data.

Each DMA context uses a circular descriptor list in system memory,
allowing deferred data processing in software as long as buffer overrun
are avoided. Since the overall operation is sleepable except for small
atomic regions, WQ_BH is not used. As the descriptors contain
timestamps, WQ_HIGHPRI is specified to support semi-real-time
processing.

The asynchronous context is also used by the SCSI over IEEE 1394
protocol implementation (sbp2), which can be part of memory reclaim paths.
Therefore, WQ_MEM_RECLAIM is required.

To allow uses to adjust CPU affinity according to workload, WQ_SYSFS is
specified so that workqueue attributes are exposed to user space.

Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Takashi Sakamoto <[email protected]>

Author: takaswie

drivers/firewire/core-card.c

@@ -574,7 +574,6 @@ EXPORT_SYMBOL(fw_card_initialize);
 int fw_card_add(struct fw_card *card, u32 max_receive, u32 link_speed, u64 guid,
 		unsigned int supported_isoc_contexts)
 {
-	struct workqueue_struct *isoc_wq;
 	int ret;
 
 	// This workqueue should be:
@@ -589,29 +588,48 @@ int fw_card_add(struct fw_card *card, u32 max_receive, u32 link_speed, u64 guid,
 	//  * == WQ_SYSFS		Parameters are available via sysfs.
 	//  * max_active == n_it + n_ir	A hardIRQ could notify events for multiple isochronous
 	//				contexts if they are scheduled to the same cycle.
-	isoc_wq = alloc_workqueue("firewire-isoc-card%u",
-				  WQ_UNBOUND | WQ_FREEZABLE | WQ_HIGHPRI | WQ_SYSFS,
-				  supported_isoc_contexts, card->index);
-	if (!isoc_wq)
+	card->isoc_wq = alloc_workqueue("firewire-isoc-card%u",
+					WQ_UNBOUND | WQ_FREEZABLE | WQ_HIGHPRI | WQ_SYSFS,
+					supported_isoc_contexts, card->index);
+	if (!card->isoc_wq)
 		return -ENOMEM;
 
+	// This workqueue should be:
+	//  * != WQ_BH			Sleepable.
+	//  * == WQ_UNBOUND		Any core can process data for asynchronous context.
+	//  * == WQ_MEM_RECLAIM		Used for any backend of block device.
+	//  * == WQ_FREEZABLE		The target device would not be available when being freezed.
+	//  * == WQ_HIGHPRI		High priority to process semi-realtime timestamped data.
+	//  * == WQ_SYSFS		Parameters are available via sysfs.
+	//  * max_active == 4		A hardIRQ could notify events for a pair of requests and
+	//				response AR/AT contexts.
+	card->async_wq = alloc_workqueue("firewire-async-card%u",
+					 WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_FREEZABLE | WQ_HIGHPRI | WQ_SYSFS,
+					 4, card->index);
+	if (!card->async_wq) {
+		ret = -ENOMEM;
+		goto err_isoc;
+	}
+
 	card->max_receive = max_receive;
 	card->link_speed = link_speed;
 	card->guid = guid;
 
-	guard(mutex)(&card_mutex);
+	scoped_guard(mutex, &card_mutex) {
+		generate_config_rom(card, tmp_config_rom);
+		ret = card->driver->enable(card, tmp_config_rom, config_rom_length);
+		if (ret < 0)
+			goto err_async;
 
-	generate_config_rom(card, tmp_config_rom);
-	ret = card->driver->enable(card, tmp_config_rom, config_rom_length);
-	if (ret < 0) {
-		destroy_workqueue(isoc_wq);
-		return ret;
+		list_add_tail(&card->link, &card_list);
 	}
 
-	card->isoc_wq = isoc_wq;
-	list_add_tail(&card->link, &card_list);
-
 	return 0;
+err_async:
+	destroy_workqueue(card->async_wq);
+err_isoc:
+	destroy_workqueue(card->isoc_wq);
+	return ret;
 }
 EXPORT_SYMBOL(fw_card_add);
 
@@ -744,6 +762,7 @@ void fw_core_remove_card(struct fw_card *card)
 	dummy_driver.stop_iso		= card->driver->stop_iso;
 	card->driver = &dummy_driver;
 	drain_workqueue(card->isoc_wq);
+	drain_workqueue(card->async_wq);
 
 	scoped_guard(spinlock_irqsave, &card->lock)
 		fw_destroy_nodes(card);
@@ -753,6 +772,7 @@ void fw_core_remove_card(struct fw_card *card)
 	wait_for_completion(&card->done);
 
 	destroy_workqueue(card->isoc_wq);
+	destroy_workqueue(card->async_wq);
 
 	WARN_ON(!list_empty(&card->transaction_list));
 }

include/linux/firewire.h

@@ -136,6 +136,7 @@ struct fw_card {
 	__be32 maint_utility_register;
 
 	struct workqueue_struct *isoc_wq;
+	struct workqueue_struct *async_wq;
 };
 
 static inline struct fw_card *fw_card_get(struct fw_card *card)