Hi Simon,
On 16/8/20 00:06, Simon Glass wrote:
Hi Walter,
On Sun, 26 Jul 2020 at 20:45, Walter Lozano <walter.loz...@collabora.com> wrote:
Hi Simon,
On 10/7/20 01:12, Walter Lozano wrote:
Hi Simon,
On 2/7/20 18:10, Simon Glass wrote:
This series provides a proposed enhancement to driver model to reduce
overhead in SPL.
These patches should not be reviewed other than to comment on the
approach. The code is all lumped together in a few patches and so cannot
be applied as is.
For now, the source tree is available at:
https://gitlab.denx.de/u-boot/custodians/u-boot-dm/-/tree/dtoc-working
Comments welcome!
Benefits (good news)
--------------------
As an example of the impact of tiny-dm, chromebook_jerry is converted to
use it. This shows approximately a 30% reduction in code and data
size and
a 85% reduction in malloc() space over of-platdata:
text data bss dec hex filename
25248 1836 12 27096 69d8 spl/u-boot-spl
(original with DT)
19727 3436 12 23175 5a87 spl/u-boot-spl
(OF_PLATDATA)
78% 187% 100% 86% as %age of original
13784 1408 12 15204 3b64 spl/u-boot-spl
(SPL_TINY)
70% 41% 100% 66% as %age of platdata
55% 77% 100% 56% as %age of original
SPL malloc() usage drops from 944 bytes (OF_PLATDATA) to 116 (SPL_TINY).
Overall the 'overhead' of tiny-dm is much less than the full driver
model. Code size is currently about 600 bytes for these functions on
Thumb2:
00000054 T tiny_dev_probe
00000034 T tiny_dev_get_by_drvdata
00000024 T tiny_dev_find
0000001a T tiny_dev_get
0000003c T tinydev_alloc_data
0000002a t tinydev_lookup_data
00000022 T tinydev_ensure_data
00000014 T tinydev_get_data
00000004 T tinydev_get_parent
Effort (bad news)
-----------------
Unfortunately it is quite a bit of work to convert drivers over to
tiny-dm. First, the of-platdata conversion must be done. But on top of
that, tiny-dm needs entirely separate code for dealing with devices.
This
means that instead of 'struct udevice' and 'struct uclass' there is just
'struct tinydev'. Each driver and uclass must be modified to support
both, pulling common code into internal static functions.
Another option
--------------
Note: It is assumed that any board that is space-contrained should use
of-platdata in SPL (see doc/driver-model/of-plat.rst). This is shown to
reduce device-tree overhead by approximately 4KB.
Designing tiny-dm has suggested a number of things that could be changed
in the current driver model to make it more space-efficient for TPL and
SPL. The ones with least impact on driver code are (CS=reduces code
size,
DS=reduces data size):
CS - drop driver_bind() and create devices (struct udevice) at
build-time
CS - allocate all device- and uclass-private data at build-time
CS - remove all but the basic operations for each uclass (e.g. SPI
flash only supports reading)
DS - use 8-bit indexes instead of 32/64-bit pointers for device
pointers possible since these are created at build-time)
DS - use singly-linked lists
DS - use 16-bit offsets to private data, instead of 32/64-bit
pointers
(possible since it is all in SRAM relative to malloc() base,
presumably word-aligned and < 256KB)
DS - move private pointers into a separate data structure so that
NULLs
are not stored
CS / DS - Combine req_seq and seq and calculate the new value at
build-time
More difficult are:
DS - drop some of the lesser-used driver and uclass methods
DS - drop all uclass methods except init()
DS - drop all driver methods except probe()
CS / DS - drop uclasses and require drivers to manually call uclass
functions
Even with all of this we would not reach tiny-dm and it would muddy
up the
driver-model datas structures. But we might come close to tiny-dm on
size
and there are some advantages:
- much of the benefit would apply to all boards that use of-platdata
(i.e.
with very little effort on behalf of board maintainers)
- the impact on the code base is much less (we keep a single, unified
driver mode in SPL and U-Boot proper)
Overall I think it is worth looking at this option. While it doesn't
have
the 'nuclear' impact of tiny-dm, neither does it mess with the U-Boot
driver code as much and it is easier to learn.
Thanks for your hard work on this topic.
I think that there is great value in this research and in this
conclusion. It is clear that there two different approaches, but I
feel that the improvement to the current DM implementation would have
a higher impact in the community.
Since the first version of this proposal I have been thinking in a
solution that takes some of the advantages of tiny-dm idea but that
does not require so much effort. This seems to be aligned with what
you have been explaining in this section.
I found interesting your proposal about simplification some data
structures. In this sense one of my ideas, a bit biased by some of the
improvements in dtoc, is to change the the definition of struct driver
based on if OF_PLATDATA is enabled, and in this case remove some
properties.
struct driver {
#if !CONFIG_IS_ENABLED(OF_PLATDATA)
char *name;
#endif
enum uclass_id id;
#if !CONFIG_IS_ENABLED(OF_PLATDATA)
const struct udevice_id *of_match;
#endif
int (*bind)(struct udevice *dev);
int (*probe)(struct udevice *dev);
int (*remove)(struct udevice *dev);
int (*unbind)(struct udevice *dev);
int (*ofdata_to_platdata)(struct udevice *dev);
int (*child_post_bind)(struct udevice *dev);
int (*child_pre_probe)(struct udevice *dev);
int (*child_post_remove)(struct udevice *dev);
int priv_auto_alloc_size;
int platdata_auto_alloc_size;
int per_child_auto_alloc_size;
int per_child_platdata_auto_alloc_size;
const void *ops; /* driver-specific operations */
uint32_t flags;
#if CONFIG_IS_ENABLED(ACPIGEN)
struct acpi_ops *acpi_ops;
#endif
};
By just removing those properties, we save some bytes as we get rid of
several strings. Also maybe it would be nice to add some macros to
make it cleaner in drivers to use or not those properties, instead of
adding lots of #if.
I feel, as you clearly expressed, that some additional refactotring
can be made to make the logic be more similar to the tiny-dm one. I
also found interesting that several of your proposals will have impact
in U-Boot, not only in TPL/SPL.
Just to be a bit more clear, I was thinking in something like
diff --git a/include/dm/device.h b/include/dm/device.h
index f5738a0cee..0ee239be8f 100644
--- a/include/dm/device.h
+++ b/include/dm/device.h
@@ -203,6 +203,16 @@ struct udevice_id {
#define of_match_ptr(_ptr) NULL
#endif /* CONFIG_IS_ENABLED(OF_CONTROL) */
+#if !CONFIG_IS_ENABLED(OF_PLATDATA)
+#undef OF_PLATDATA_TINY
+#define STRUCT_FIELD(x) x;
+#define STRUCT_VALUE(x) x,
+#else
+#define OF_PLATDATA_TINY
+#define STRUCT_FIELD(x)
+#define STRUCT_VALUE(x)
+#endif
+
/**
* struct driver - A driver for a feature or peripheral
*
@@ -252,9 +262,9 @@ struct udevice_id {
* allowing the device to add things to the ACPI tables passed to Linux
*/
struct driver {
- char *name;
+ STRUCT_FIELD(char *name)
enum uclass_id id;
- const struct udevice_id *of_match;
+ STRUCT_FIELD(const struct udevice_id *of_match)
I didn't actually reply to this, but I think this makes sense. Perhaps
have a DM_ prefix on your #defines?
(
Thanks for your reply and paying attention to my comments. I totally
agree that if you implement this idea you should rename #defines to make
them more readable.
int (*bind)(struct udevice *dev);
int (*probe)(struct udevice *dev);
int (*remove)(struct udevice *dev);
diff --git a/drivers/core/simple-bus.c b/drivers/core/simple-bus.c
index 7cc1d46009..e303d59daf 100644
--- a/drivers/core/simple-bus.c
+++ b/drivers/core/simple-bus.c
@@ -57,8 +57,8 @@ static const struct udevice_id generic_simple_bus_ids[] = {
};
U_BOOT_DRIVER(simple_bus) = {
- .name = "simple_bus",
+ STRUCT_VALUE(.name = "simple_bus")
.id = UCLASS_SIMPLE_BUS,
- .of_match = generic_simple_bus_ids,
+ STRUCT_VALUE(.of_match = generic_simple_bus_ids)
.flags = DM_FLAG_PRE_RELOC,
};
Please don't pay attention to how OF_PLATDATA_TINY is implemented, it is
just part of the test. I think it should be a configuration option that
removes some functionality from OF_PLATDATA, like having the overhead of
the strings. On top of this you could add additional improvements, like
the binding implementation in tiny-dm which uses DM_REF_TINY_DRIVER.
Yes OK. I am thinking of starting with just adding a parent. But time
is not my friend at present so it will probably be a month or so.
Thanks for sharing your plans. Time is always an issue, but I think we
are in the right track, and that all the research you have done with
tiny-dm will give great benefits.
Regards,
Walter
