Skip to contentSkip to conversation

ARM: mach-shmobile: Add memchunk support V2

PATCH In review Magnus Damm 2 participants 2 messages Oct 14, 2010 – Oct 15, 2010

Subsystems (4)
lore
Share & tools
lore

Magnus Damm

author maintainer
#0 lore raw

PATCH 1/1 ARM: mach-shmobile: Add memchunk support V2

From: Magnus Damm <[email protected]>

Add SH-Mobile ARM support for "memchunk" V2. The code is based on the SH implementation but has been slightly modified to use memblock for physical memory reservation.

The function shmobile_memchunk_setup() should be called during boot from the ->reserve() callback in struct machine_desc.

---

 arch/arm/mach-shmobile/Makefile              |    1 
 arch/arm/mach-shmobile/include/mach/common.h |    3 +
 arch/arm/mach-shmobile/memchunk.c            |   76 ++++++++++++++++++++++++++
 3 files changed, 80 insertions(+)
arch/arm/mach-shmobile/Makefile +1 -0
@@ -4,6 +4,7 @@
 
 # Common objects
 obj-y				:= timer.o console.o clock.o pm_runtime.o
+obj-y				+= memchunk.o
 
 # CPU objects
 obj-$(CONFIG_ARCH_AP45)		+= head-ap45.o
arch/arm/mach-shmobile/include/mach/common.h +3 -0
@@ -3,6 +3,9 @@
 
 extern struct sys_timer shmobile_timer;
 extern void shmobile_setup_console(void);
+struct platform_device;
+extern int shmobile_memchunk_setup(struct platform_device *pdev,
+				   char *name, unsigned long memsize);
 struct clk;
 extern int clk_init(void);
 
arch/arm/mach-shmobile/memchunk.c new file +76 -0

76 lines not on this page · raw message

#1 lore raw
Quoted 6 lines · Magnus Damm

From: Magnus Damm <[email protected]>

Add SH-Mobile ARM support for "memchunk" V2. The code is based on the SH implementation but has been slightly modified to use memblock for physical memory reservation.

There are two issues here..

- first:

Is there any reason not to use the same approach on SH? I rather dislike having wholly generic code split between the architectures for no reason, as seems to be the trend lately.

All of this started out abusing the platform_ namespace because it related to platform devices, as it still does. You were told previously to run this by gregkh and make a drivers/base/platform.c push for it the last time you posted these patches, and so far I haven't seen any progress in that department.

Now you've moved on to memblock allocation/reservation, which while not as generic as the DMA allocator, is still common between all the platforms that care about this. The only stuff that really belongs in the architecture subdirectories are the reservation callbacks, which in this case we can quite easily wrap in to a generic callback.

Now that memblock is common enough, the alternative you have is simply to do the reservations from the ->reserve() callback and then have the driver code grovel for a mapping with memblock_find(). You can of course use ATAGs to aid in this if you wish, as OMAP does for its framebuffer memory.

- second, and more fundamental:

My main concern however is that this basically won't work for the common platform device case on ARM given that all of the memblock allocations have valid backing PFNs. The UIO case "works" by virtue of blindsiding the remap path and calling in to remap_pfn_range() directly, which is unaware of any architectural limitations. Any non-UIO driver using this scheme today will just error out on the first ioremap() once the pfn_valid() check returns.

In order to support this properly you will need to punch out memblock region holes, not unlike the ARCH_HAS_HOLES_MEMORYMODEL behaviour, albeit with a bit more flexibility. I vaguely recall having mentioned this before, too.