Root/Documentation/driver-model/devres.txt

1Devres - Managed Device Resource
2================================
3
4Tejun Heo <teheo@suse.de>
5
6First draft 10 January 2007
7
8
91. Intro : Huh? Devres?
102. Devres : Devres in a nutshell
113. Devres Group : Group devres'es and release them together
124. Details : Life time rules, calling context, ...
135. Overhead : How much do we have to pay for this?
146. List of managed interfaces : Currently implemented managed interfaces
15
16
17  1. Intro
18  --------
19
20devres came up while trying to convert libata to use iomap. Each
21iomapped address should be kept and unmapped on driver detach. For
22example, a plain SFF ATA controller (that is, good old PCI IDE) in
23native mode makes use of 5 PCI BARs and all of them should be
24maintained.
25
26As with many other device drivers, libata low level drivers have
27sufficient bugs in ->remove and ->probe failure path. Well, yes,
28that's probably because libata low level driver developers are lazy
29bunch, but aren't all low level driver developers? After spending a
30day fiddling with braindamaged hardware with no document or
31braindamaged document, if it's finally working, well, it's working.
32
33For one reason or another, low level drivers don't receive as much
34attention or testing as core code, and bugs on driver detach or
35initialization failure don't happen often enough to be noticeable.
36Init failure path is worse because it's much less travelled while
37needs to handle multiple entry points.
38
39So, many low level drivers end up leaking resources on driver detach
40and having half broken failure path implementation in ->probe() which
41would leak resources or even cause oops when failure occurs. iomap
42adds more to this mix. So do msi and msix.
43
44
45  2. Devres
46  ---------
47
48devres is basically linked list of arbitrarily sized memory areas
49associated with a struct device. Each devres entry is associated with
50a release function. A devres can be released in several ways. No
51matter what, all devres entries are released on driver detach. On
52release, the associated release function is invoked and then the
53devres entry is freed.
54
55Managed interface is created for resources commonly used by device
56drivers using devres. For example, coherent DMA memory is acquired
57using dma_alloc_coherent(). The managed version is called
58dmam_alloc_coherent(). It is identical to dma_alloc_coherent() except
59for the DMA memory allocated using it is managed and will be
60automatically released on driver detach. Implementation looks like
61the following.
62
63  struct dma_devres {
64    size_t size;
65    void *vaddr;
66    dma_addr_t dma_handle;
67  };
68
69  static void dmam_coherent_release(struct device *dev, void *res)
70  {
71    struct dma_devres *this = res;
72
73    dma_free_coherent(dev, this->size, this->vaddr, this->dma_handle);
74  }
75
76  dmam_alloc_coherent(dev, size, dma_handle, gfp)
77  {
78    struct dma_devres *dr;
79    void *vaddr;
80
81    dr = devres_alloc(dmam_coherent_release, sizeof(*dr), gfp);
82    ...
83
84    /* alloc DMA memory as usual */
85    vaddr = dma_alloc_coherent(...);
86    ...
87
88    /* record size, vaddr, dma_handle in dr */
89    dr->vaddr = vaddr;
90    ...
91
92    devres_add(dev, dr);
93
94    return vaddr;
95  }
96
97If a driver uses dmam_alloc_coherent(), the area is guaranteed to be
98freed whether initialization fails half-way or the device gets
99detached. If most resources are acquired using managed interface, a
100driver can have much simpler init and exit code. Init path basically
101looks like the following.
102
103  my_init_one()
104  {
105    struct mydev *d;
106
107    d = devm_kzalloc(dev, sizeof(*d), GFP_KERNEL);
108    if (!d)
109        return -ENOMEM;
110
111    d->ring = dmam_alloc_coherent(...);
112    if (!d->ring)
113        return -ENOMEM;
114
115    if (check something)
116        return -EINVAL;
117    ...
118
119    return register_to_upper_layer(d);
120  }
121
122And exit path,
123
124  my_remove_one()
125  {
126    unregister_from_upper_layer(d);
127    shutdown_my_hardware();
128  }
129
130As shown above, low level drivers can be simplified a lot by using
131devres. Complexity is shifted from less maintained low level drivers
132to better maintained higher layer. Also, as init failure path is
133shared with exit path, both can get more testing.
134
135
136  3. Devres group
137  ---------------
138
139Devres entries can be grouped using devres group. When a group is
140released, all contained normal devres entries and properly nested
141groups are released. One usage is to rollback series of acquired
142resources on failure. For example,
143
144  if (!devres_open_group(dev, NULL, GFP_KERNEL))
145    return -ENOMEM;
146
147  acquire A;
148  if (failed)
149    goto err;
150
151  acquire B;
152  if (failed)
153    goto err;
154  ...
155
156  devres_remove_group(dev, NULL);
157  return 0;
158
159 err:
160  devres_release_group(dev, NULL);
161  return err_code;
162
163As resource acquisition failure usually means probe failure, constructs
164like above are usually useful in midlayer driver (e.g. libata core
165layer) where interface function shouldn't have side effect on failure.
166For LLDs, just returning error code suffices in most cases.
167
168Each group is identified by void *id. It can either be explicitly
169specified by @id argument to devres_open_group() or automatically
170created by passing NULL as @id as in the above example. In both
171cases, devres_open_group() returns the group's id. The returned id
172can be passed to other devres functions to select the target group.
173If NULL is given to those functions, the latest open group is
174selected.
175
176For example, you can do something like the following.
177
178  int my_midlayer_create_something()
179  {
180    if (!devres_open_group(dev, my_midlayer_create_something, GFP_KERNEL))
181        return -ENOMEM;
182
183    ...
184
185    devres_close_group(dev, my_midlayer_create_something);
186    return 0;
187  }
188
189  void my_midlayer_destroy_something()
190  {
191    devres_release_group(dev, my_midlayer_create_something);
192  }
193
194
195  4. Details
196  ----------
197
198Lifetime of a devres entry begins on devres allocation and finishes
199when it is released or destroyed (removed and freed) - no reference
200counting.
201
202devres core guarantees atomicity to all basic devres operations and
203has support for single-instance devres types (atomic
204lookup-and-add-if-not-found). Other than that, synchronizing
205concurrent accesses to allocated devres data is caller's
206responsibility. This is usually non-issue because bus ops and
207resource allocations already do the job.
208
209For an example of single-instance devres type, read pcim_iomap_table()
210in lib/devres.c.
211
212All devres interface functions can be called without context if the
213right gfp mask is given.
214
215
216  5. Overhead
217  -----------
218
219Each devres bookkeeping info is allocated together with requested data
220area. With debug option turned off, bookkeeping info occupies 16
221bytes on 32bit machines and 24 bytes on 64bit (three pointers rounded
222up to ull alignment). If singly linked list is used, it can be
223reduced to two pointers (8 bytes on 32bit, 16 bytes on 64bit).
224
225Each devres group occupies 8 pointers. It can be reduced to 6 if
226singly linked list is used.
227
228Memory space overhead on ahci controller with two ports is between 300
229and 400 bytes on 32bit machine after naive conversion (we can
230certainly invest a bit more effort into libata core layer).
231
232
233  6. List of managed interfaces
234  -----------------------------
235
236IO region
237  devm_request_region()
238  devm_request_mem_region()
239  devm_release_region()
240  devm_release_mem_region()
241
242IRQ
243  devm_request_irq()
244  devm_free_irq()
245
246DMA
247  dmam_alloc_coherent()
248  dmam_free_coherent()
249  dmam_alloc_noncoherent()
250  dmam_free_noncoherent()
251  dmam_declare_coherent_memory()
252  dmam_pool_create()
253  dmam_pool_destroy()
254
255PCI
256  pcim_enable_device() : after success, all PCI ops become managed
257  pcim_pin_device() : keep PCI device enabled after release
258
259IOMAP
260  devm_ioport_map()
261  devm_ioport_unmap()
262  devm_ioremap()
263  devm_ioremap_nocache()
264  devm_iounmap()
265  pcim_iomap()
266  pcim_iounmap()
267  pcim_iomap_table() : array of mapped addresses indexed by BAR
268  pcim_iomap_regions() : do request_region() and iomap() on multiple BARs
269

Archive Download this file



interactive