devres: device resource management

Implement device resource management, in short, devres.  A device
driver can allocate arbirary size of devres data which is associated
with a release function.  On driver detach, release function is
invoked on the devres data, then, devres data is freed.

devreses are typed by associated release functions.  Some devreses are
better represented by single instance of the type while others need
multiple instances sharing the same release function.  Both usages are
supported.

devreses can be grouped using devres group such that a device driver
can easily release acquired resources halfway through initialization
or selectively release resources (e.g. resources for port 1 out of 4
ports).

This patch adds devres core including documentation and the following
managed interfaces.

* alloc/free	: devm_kzalloc(), devm_kzfree()
* IO region	: devm_request_region(), devm_release_region()
* IRQ		: devm_request_irq(), devm_free_irq()
* DMA		: dmam_alloc_coherent(), dmam_free_coherent(),
		  dmam_declare_coherent_memory(), dmam_pool_create(),
		  dmam_pool_destroy()
* PCI		: pcim_enable_device(), pcim_pin_device(), pci_is_managed()
* iomap		: devm_ioport_map(), devm_ioport_unmap(), devm_ioremap(),
		  devm_ioremap_nocache(), devm_iounmap(), pcim_iomap_table(),
		  pcim_iomap(), pcim_iounmap()

Signed-off-by: Tejun Heo <[email protected]>
Signed-off-by: Jeff Garzik <[email protected]>

Author: htejun

Documentation/driver-model/devres.txt

@@ -0,0 +1,268 @@
+Devres - Managed Device Resource
+================================
+
+Tejun Heo	<[email protected]>
+
+First draft	10 January 2007
+
+
+1. Intro			: Huh? Devres?
+2. Devres			: Devres in a nutshell
+3. Devres Group			: Group devres'es and release them together
+4. Details			: Life time rules, calling context, ...
+5. Overhead			: How much do we have to pay for this?
+6. List of managed interfaces	: Currently implemented managed interfaces
+
+
+  1. Intro
+  --------
+
+devres came up while trying to convert libata to use iomap.  Each
+iomapped address should be kept and unmapped on driver detach.  For
+example, a plain SFF ATA controller (that is, good old PCI IDE) in
+native mode makes use of 5 PCI BARs and all of them should be
+maintained.
+
+As with many other device drivers, libata low level drivers have
+sufficient bugs in ->remove and ->probe failure path.  Well, yes,
+that's probably because libata low level driver developers are lazy
+bunch, but aren't all low level driver developers?  After spending a
+day fiddling with braindamaged hardware with no document or
+braindamaged document, if it's finally working, well, it's working.
+
+For one reason or another, low level drivers don't receive as much
+attention or testing as core code, and bugs on driver detach or
+initilaization failure doesn't happen often enough to be noticeable.
+Init failure path is worse because it's much less travelled while
+needs to handle multiple entry points.
+
+So, many low level drivers end up leaking resources on driver detach
+and having half broken failure path implementation in ->probe() which
+would leak resources or even cause oops when failure occurs.  iomap
+adds more to this mix.  So do msi and msix.
+
+
+  2. Devres
+  ---------
+
+devres is basically linked list of arbitrarily sized memory areas
+associated with a struct device.  Each devres entry is associated with
+a release function.  A devres can be released in several ways.  No
+matter what, all devres entries are released on driver detach.  On
+release, the associated release function is invoked and then the
+devres entry is freed.
+
+Managed interface is created for resources commonly used by device
+drivers using devres.  For example, coherent DMA memory is acquired
+using dma_alloc_coherent().  The managed version is called
+dmam_alloc_coherent().  It is identical to dma_alloc_coherent() except
+for the DMA memory allocated using it is managed and will be
+automatically released on driver detach.  Implementation looks like
+the following.
+
+  struct dma_devres {
+	size_t		size;
+	void		*vaddr;
+	dma_addr_t	dma_handle;
+  };
+
+  static void dmam_coherent_release(struct device *dev, void *res)
+  {
+	struct dma_devres *this = res;
+
+	dma_free_coherent(dev, this->size, this->vaddr, this->dma_handle);
+  }
+
+  dmam_alloc_coherent(dev, size, dma_handle, gfp)
+  {
+	struct dma_devres *dr;
+	void *vaddr;
+
+	dr = devres_alloc(dmam_coherent_release, sizeof(*dr), gfp);
+	...
+
+	/* alloc DMA memory as usual */
+	vaddr = dma_alloc_coherent(...);
+	...
+
+	/* record size, vaddr, dma_handle in dr */
+	dr->vaddr = vaddr;
+	...
+
+	devres_add(dev, dr);
+
+	return vaddr;
+  }
+
+If a driver uses dmam_alloc_coherent(), the area is guaranteed to be
+freed whether initialization fails half-way or the device gets
+detached.  If most resources are acquired using managed interface, a
+driver can have much simpler init and exit code.  Init path basically
+looks like the following.
+
+  my_init_one()
+  {
+	struct mydev *d;
+
+	d = devm_kzalloc(dev, sizeof(*d), GFP_KERNEL);
+	if (!d)
+		return -ENOMEM;
+
+	d->ring = dmam_alloc_coherent(...);
+	if (!d->ring)
+		return -ENOMEM;
+
+	if (check something)
+		return -EINVAL;
+	...
+
+	return register_to_upper_layer(d);
+  }
+
+And exit path,
+
+  my_remove_one()
+  {
+	unregister_from_upper_layer(d);
+	shutdown_my_hardware();
+  }
+
+As shown above, low level drivers can be simplified a lot by using
+devres.  Complexity is shifted from less maintained low level drivers
+to better maintained higher layer.  Also, as init failure path is
+shared with exit path, both can get more testing.
+
+
+  3. Devres group
+  ---------------
+
+Devres entries can be grouped using devres group.  When a group is
+released, all contained normal devres entries and properly nested
+groups are released.  One usage is to rollback series of acquired
+resources on failure.  For example,
+
+  if (!devres_open_group(dev, NULL, GFP_KERNEL))
+	return -ENOMEM;
+
+  acquire A;
+  if (failed)
+	goto err;
+
+  acquire B;
+  if (failed)
+	goto err;
+  ...
+
+  devres_remove_group(dev, NULL);
+  return 0;
+
+ err:
+  devres_release_group(dev, NULL);
+  return err_code;
+
+As resource acquision failure usually means probe failure, constructs
+like above are usually useful in midlayer driver (e.g. libata core
+layer) where interface function shouldn't have side effect on failure.
+For LLDs, just returning error code suffices in most cases.
+
+Each group is identified by void *id.  It can either be explicitly
+specified by @id argument to devres_open_group() or automatically
+created by passing NULL as @id as in the above example.  In both
+cases, devres_open_group() returns the group's id.  The returned id
+can be passed to other devres functions to select the target group.
+If NULL is given to those functions, the latest open group is
+selected.
+
+For example, you can do something like the following.
+
+  int my_midlayer_create_something()
+  {
+	if (!devres_open_group(dev, my_midlayer_create_something, GFP_KERNEL))
+		return -ENOMEM;
+
+	...
+
+	devres_close_group(dev, my_midlayer_something);
+	return 0;
+  }
+
+  void my_midlayer_destroy_something()
+  {
+	devres_release_group(dev, my_midlayer_create_soemthing);
+  }
+
+
+  4. Details
+  ----------
+
+Lifetime of a devres entry begins on devres allocation and finishes
+when it is released or destroyed (removed and freed) - no reference
+counting.
+
+devres core guarantees atomicity to all basic devres operations and
+has support for single-instance devres types (atomic
+lookup-and-add-if-not-found).  Other than that, synchronizing
+concurrent accesses to allocated devres data is caller's
+responsibility.  This is usually non-issue because bus ops and
+resource allocations already do the job.
+
+For an example of single-instance devres type, read pcim_iomap_table()
+in lib/iomap.c.
+
+All devres interface functions can be called without context if the
+right gfp mask is given.
+
+
+  5. Overhead
+  -----------
+
+Each devres bookkeeping info is allocated together with requested data
+area.  With debug option turned off, bookkeeping info occupies 16
+bytes on 32bit machines and 24 bytes on 64bit (three pointers rounded
+up to ull alignment).  If singly linked list is used, it can be
+reduced to two pointers (8 bytes on 32bit, 16 bytes on 64bit).
+
+Each devres group occupies 8 pointers.  It can be reduced to 6 if
+singly linked list is used.
+
+Memory space overhead on ahci controller with two ports is between 300
+and 400 bytes on 32bit machine after naive conversion (we can
+certainly invest a bit more effort into libata core layer).
+
+
+  6. List of managed interfaces
+  -----------------------------
+
+IO region
+  devm_request_region()
+  devm_request_mem_region()
+  devm_release_region()
+  devm_release_mem_region()
+
+IRQ
+  devm_request_irq()
+  devm_free_irq()
+
+DMA
+  dmam_alloc_coherent()
+  dmam_free_coherent()
+  dmam_alloc_noncoherent()
+  dmam_free_noncoherent()
+  dmam_declare_coherent_memory()
+  dmam_pool_create()
+  dmam_pool_destroy()
+
+PCI
+  pcim_enable_device()	: after success, all PCI ops become managed
+  pcim_pin_device()	: keep PCI device enabled after release
+
+IOMAP
+  devm_ioport_map()
+  devm_ioport_unmap()
+  devm_ioremap()
+  devm_ioremap_nocache()
+  devm_iounmap()
+  pcim_iomap()
+  pcim_iounmap()
+  pcim_iomap_table()	: array of mapped addresses indexed by BAR
+  pcim_iomap_regions()	: do request_region() and iomap() on multiple BARs

drivers/base/Kconfig

@@ -37,6 +37,18 @@ config DEBUG_DRIVER
 
 	  If you are unsure about this, say N here.
 
+config DEBUG_DEVRES
+	bool "Managed device resources verbose debug messages"
+	depends on DEBUG_KERNEL
+	help
+	  This option enables kernel parameter devres.log. If set to
+	  non-zero, devres debug messages are printed. Select this if
+	  you are having a problem with devres or want to debug
+	  resource management for a managed device. devres.log can be
+	  switched on and off from sysfs node.
+
+	  If you are unsure about this, Say N here.
+
 config SYS_HYPERVISOR
 	bool
 	default n

drivers/base/Makefile

@@ -3,6 +3,7 @@
 obj-y			:= core.o sys.o bus.o dd.o \
 			   driver.o class.o platform.o \
 			   cpu.o firmware.o init.o map.o dmapool.o \
+			   dma-mapping.o devres.o \
 			   attribute_container.o transport_class.o
 obj-y			+= power/
 obj-$(CONFIG_ISA)	+= isa.o

drivers/base/base.h

@@ -44,3 +44,4 @@ struct class_device_attribute *to_class_dev_attr(struct attribute *_attr)
 
 extern char *make_class_name(const char *name, struct kobject *kobj);
 
+extern void devres_release_all(struct device *dev);

drivers/base/core.c

@@ -428,6 +428,8 @@ void device_initialize(struct device *dev)
 	INIT_LIST_HEAD(&dev->dma_pools);
 	INIT_LIST_HEAD(&dev->node);
 	init_MUTEX(&dev->sem);
+	spin_lock_init(&dev->devres_lock);
+	INIT_LIST_HEAD(&dev->devres_head);
 	device_init_wakeup(dev, 0);
 	set_dev_node(dev, -1);
 }

drivers/base/dd.c

@@ -112,6 +112,7 @@ static int really_probe(void *void_data)
 	atomic_inc(&probe_count);
 	pr_debug("%s: Probing driver %s with device %s\n",
 		 drv->bus->name, drv->name, dev->bus_id);
+	WARN_ON(!list_empty(&dev->devres_head));
 
 	dev->driver = drv;
 	if (driver_sysfs_add(dev)) {
@@ -137,6 +138,7 @@ static int really_probe(void *void_data)
 	goto done;
 
 probe_failed:
+	devres_release_all(dev);
 	driver_sysfs_remove(dev);
 	dev->driver = NULL;
 
@@ -327,6 +329,7 @@ static void __device_release_driver(struct device * dev)
 			dev->bus->remove(dev);
 		else if (drv->remove)
 			drv->remove(dev);
+		devres_release_all(dev);
 		dev->driver = NULL;
 		put_driver(drv);
 	}