Re: dma_declare_coherent_memory fails for RAM allocated memory

From: Brian Starkey
Date: Wed Jun 01 2016 - 07:19:01 EST

Hi Ivo,

On Sun, May 29, 2016 at 05:56:02PM +0300, Ivaylo Dimitrov wrote:

When trying to declare and use DT reserved memory region on ARM (OMAP3), dma_declare_coherent_memory() fails in memremap(). This is from today's master:

------------[ cut here ]------------
WARNING: CPU: 0 PID: 1 at kernel/memremap.c:111 memremap+0x118/0x194
memremap attempted on ram 0x8f800000 size: 0x700000
Modules linked in:
CPU: 0 PID: 1 Comm: swapper Not tainted 4.6.0+ #15
Hardware name: Nokia RX-51 board
[<c010bc30>] (unwind_backtrace) from [<c0109f70>] (show_stack+0x10/0x14)
[<c0109f70>] (show_stack) from [<c0126d80>] (__warn+0xcc/0xf8)
[<c0126d80>] (__warn) from [<c0126e40>] (warn_slowpath_fmt+0x34/0x44)
[<c0126e40>] (warn_slowpath_fmt) from [<c019c8d4>] (memremap+0x118/0x194)
[<c019c8d4>] (memremap) from [<c03b88fc>] (dma_init_coherent_memory+0x48/0x104)
[<c03b88fc>] (dma_init_coherent_memory) from [<c03b89e4>] (dma_declare_coherent_memory+0x2c/0x68)
[<c03b89e4>] (dma_declare_coherent_memory) from [<c0115c0c>] (rmem_omapfb_device_init+0x34/0x64)
[<c0115c0c>] (rmem_omapfb_device_init) from [<c045b0f8>] (of_reserved_mem_device_init+0x94/0xd8)
[<c045b0f8>] (of_reserved_mem_device_init) from [<c080aaf8>] (omapdss_init_of+0xe4/0x154)
[<c080aaf8>] (omapdss_init_of) from [<c08033f4>] (customize_machine+0x20/0x44)
[<c08033f4>] (customize_machine) from [<c01016b8>] (do_one_initcall+0xac/0x158)
[<c01016b8>] (do_one_initcall) from [<c0800d64>] (kernel_init_freeable+0xf8/0x1c8)
[<c0800d64>] (kernel_init_freeable) from [<c05519dc>] (kernel_init+0x8/0x110)
[<c05519dc>] (kernel_init) from [<c0107258>] (ret_from_fork+0x14/0x3c)
---[ end trace 73a8c076df72166b ]---
omapfb: dma_declare_coherent_memory failed

The failing code looks like:
static int rmem_omapfb_device_init(struct reserved_mem *rmem, struct device *dev)
int dma;

if (rmem->priv)
return 0;

dma = dma_declare_coherent_memory(&, rmem->base,
rmem->base, rmem->size,
if (!(dma & DMA_MEMORY_MAP)) {
pr_err("omapfb: dma_declare_coherent_memory failed\n");
return -ENOMEM;
rmem->priv =;

return 0;

static void rmem_omapfb_device_release(struct reserved_mem *rmem,
struct device *dev)

static const struct reserved_mem_ops rmem_omapfb_ops = {
.device_init = rmem_omapfb_device_init,
.device_release = rmem_omapfb_device_release,

static int __init rmem_omapfb_setup(struct reserved_mem *rmem)
rmem->ops = &rmem_omapfb_ops;
pr_info("omapfb: reserved %d bytes at %pa\n", rmem->size, &rmem->base);

return 0;

RESERVEDMEM_OF_DECLARE(dss, "ti,omapfb-memsize", rmem_omapfb_setup);

It turns out that dma_init_coherent_memory calls memremap with MEMREMAP_WC flag, which is disallowed for RAM IIUC.

Right. If you want to use a memory region as coherent DMA memory, then
that same region can't be System RAM, because System RAM is usually
mapped in a non-coherent fashion.

I quickly hacked some code to fix the issue, but as memremap API is relatively new(esp to me), I wonder if this is the correct way to go:

diff --git a/drivers/base/dma-coherent.c b/drivers/base/dma-coherent.c
index bdf28f7..04b1687 100644
--- a/drivers/base/dma-coherent.c
+++ b/drivers/base/dma-coherent.c
@@ -5,6 +5,7 @@
#include <linux/io.h>
#include <linux/slab.h>
#include <linux/kernel.h>
+#include <linux/memblock.h>
#include <linux/module.h>
#include <linux/dma-mapping.h>

@@ -32,8 +33,12 @@ static bool dma_init_coherent_memory(
if (!size)
goto out;

- if (flags & DMA_MEMORY_MAP)
- mem_base = memremap(phys_addr, size, MEMREMAP_WC);
+ if (flags & DMA_MEMORY_MAP) {
+ unsigned long map_type = memblock_is_map_memory(phys_addr) ?
+ mem_base = memremap(phys_addr, size, map_type);
+ }
mem_base = ioremap(phys_addr, size);
if (!mem_base)

Does the above code looks sane? How to fix the problem if not?

AFAIK dma_declare_coherent_memory() should only be used on physical
addresses which don't already have a CPU virtual mapping. Certainly
returning (potentially) cached memory as DMA coherent is asking for

I think what you want to do is add a "no-map;" property to your
reserved-memory region in DT. That will stop the kernel from using it
as System RAM, leaving it free for your framebuffer.