On 4/21/20 2:42 AM, Peng Fan wrote:
Subject: Re: [PATCH 1/2] remoteproc: drop memset when loading elf
segments
On 4/13/20 4:05 AM, Peng Fan wrote:
Subject: Re: [PATCH 1/2] remoteproc: drop memset when loading elf
segments
On Thu 09 Apr 18:29 PDT 2020, Peng Fan wrote:
Hi Bjorn,
Subject: Re: [PATCH 1/2] remoteproc: drop memset when loading elf
segments
On Thu 09 Apr 01:22 PDT 2020, Peng Fan wrote:
To arm64, "dc zva, dst" is used in memset.
Per ARM DDI 0487A.j, chapter C5.3.8 DC ZVA, Data Cache Zero by VA,
"If the memory region being zeroed is any type of Device memory,
this instruction can give an alignment fault which is prioritized
in the same way as other alignment faults that are determined by
the memory type."
On i.MX platforms, when elf is loaded to onchip TCM area, the
region is ioremapped, so "dc zva, dst" will trigger abort.
Since memset is not strictly required, let's drop it.
This would imply that we trust that the firmware doesn't expect
remoteproc to zero out the memory, which we've always done. So I
don't think we can say that it's not required.
Saying an image runs on a remote core needs Linux to help zero out
BSS section, this not make sense to me.
Maybe not, but it has always done it, so if there's firmware out
there that depends on it such change would break them..
Got it.
My case is as following, I need to load section 7 data.
I no need to let remoteproc to memset section 8/9/10/11/12, the
firmware itself could handle that. Just because the memsz is larger
than filesz, remoreproc must memset?
By having a PT_LOAD segment covering these I think it's reasonable to
assume that the ELF loader should be able to touch the associated
memory.
Section Headers:
[Nr] Name Type Addr Off
Size
ES Flg Lk Inf Al
[ 0] NULL 00000000 000000
000000 00 0 0 0
[ 1] .interrupts PROGBITS 1ffe0000 010000 000240
00
A 0 0 4
[ 2] .resource_table PROGBITS 1ffe0240 010240 000058
00
A 0 0 1
[ 3] .text PROGBITS 1ffe02a0 0102a0
009ccc 00
AX 0 0 16
[ 4] .ARM ARM_EXIDX 1ffe9f6c 019f6c
000008
00 AL 3 0 4
[ 5] .init_array INIT_ARRAY 1ffe9f74 019f74 000004
04
WA 0 0 4
[ 6] .fini_array FINI_ARRAY 1ffe9f78 019f78 000004
04
WA 0 0 4
[ 7] .data PROGBITS 1fff9240 029240
000084
00 WA 0 0 4
[ 8] .ncache.init PROGBITS 1fff92c4 0292c4 000000
00
W 0 0 1
[ 9] .ncache NOBITS 1fff92c4 0292c4
000a80
00 WA 0 0 4
[10] .bss NOBITS 1fff9d44 0292c4
01f5c0
00 WA 0 0 4
[11] .heap NOBITS 20019304 0292c4
000404
00 WA 0 0 1
[12] .stack NOBITS 20019708 0292c4
000400
00 WA 0 0 1
Signed-off-by: Peng Fan <peng.fan@xxxxxxx>
---
drivers/remoteproc/remoteproc_elf_loader.c | 7 ++-----
1 file changed, 2 insertions(+), 5 deletions(-)
diff --git a/drivers/remoteproc/remoteproc_elf_loader.c
b/drivers/remoteproc/remoteproc_elf_loader.c
index 16e2c496fd45..cc50fe70d50c 100644
--- a/drivers/remoteproc/remoteproc_elf_loader.c
+++ b/drivers/remoteproc/remoteproc_elf_loader.c
@@ -238,14 +238,11 @@ int rproc_elf_load_segments(struct rproc
*rproc,
const struct firmware *fw)
memcpy(ptr, elf_data + offset, filesz);
/*
- * Zero out remaining memory for this segment.
+ * No need zero out remaining memory for this segment.
*
* This isn't strictly required since dma_alloc_coherent
already
- * did this for us. albeit harmless, we may consider removing
- * this.
+ * did this for us.
In the case of recovery this comment is wrong, we do not
dma_alloc_coherent() the carveout during a recovery.
Isn't the it the firmware's job to memset the region?
I'm not aware of this being a documented requirement, we've always
done it here for them, so removing this call would be a change in behavior.
And in your case you ioremapped existing TCM, so it's never true.
*/
- if (memsz > filesz)
- memset(ptr + filesz, 0, memsz - filesz);
So I think you do want to zero out this region. Question is how we do it...
I have contacted our M4 owners, we no need clear it from Linux side.
And I think _most_ firmware out there, like yours, does clear BSS etc
during initialization.
We also support booting m4 before booting Linux, at that case, Linux
has noting to do with memset. It is just I try loading m4 image with
Linux, and met the issue that memset trigger abort.
Please see the proposal for attaching to already running remoteproc's
from Mathieu. I don't expect that you want to load your PROGBITS
either in this case?
No need to load for early boot case. It is just userspace load trigger
kernel panic, because memset arm64 could not work for ioremapped
memory.
How about adding ops hooks for memset and memcpy to let driver have
their own implementation?
Hi Peng,
The trick is to use the ioremap_wc() variant instead of ioremap() in your
platform driver while mapping the TCMs. I know multiple folks have run into
this issue. This is what most of the remoteproc drivers use, and mmio-sram
driver also uses the same.
TCM is different from OCRAM in i.MX chips.
ioremap_wc not work for TCM memory, I just tried that.
What ARM core is this? Is it a standard ARM TCM memory? The TCM
interfaces from standard ARM cores in general are treated as normal
memory, so write combine should be ok on them. When you say it doesn't
work, whats not working - the memcpy/memset or the remoteproc doesn't boot?
I am thinking we change memset/memcpy to use
memset_io/memcpy_fromio/memcpy_toio for remoteproc common code,
This has other implications. Not everything is IO memory to make this
change in the common core.
If this is a custom memory interface requiring specific handling, then
one option is to provide and use your own ops->load function within the
driver. This is what I do for one of our remoteprocs that requires a
specific memcpy handling semantics.
regards
Suman
