[PATCH 7/13] devres: add Documentation/driver-model/devres.txt

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Add Documentation/driver-model/drvres.txt.

Signed-off-by: Tejun Heo <[email protected]>
---
 Documentation/driver-model/devres.txt |  268 +++++++++++++++++++++++++++++++++
 1 files changed, 268 insertions(+), 0 deletions(-)

diff --git a/Documentation/driver-model/devres.txt b/Documentation/driver-model/devres.txt
new file mode 100644
index 0000000..5163b85
--- /dev/null
+++ b/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
-- 
1.4.4.3


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[Index of Archives]     [Kernel Newbies]     [Netfilter]     [Bugtraq]     [Photo]     [Stuff]     [Gimp]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Video 4 Linux]     [Linux for the blind]     [Linux Resources]
  Powered by Linux