For a number of reasons, I&#39;d like to run my kernel in SRAM but allocate data memory (i.e. socket buffers) in SDRAM. This seems possible with a NIOS &#39;Hello World&#39; application, or the like, but these options aren&#39;t readily apparent when creating a kernel project in Eclipse. Is there a config file I can tweak, or other steps to be performed to accomplish putting .DATA in SDRAM? I&#39;ve considered actually rewriting appropriate allocation routines in the kernel, but that feels very dangerous. Any help would be appreciated. Thanks in advance, Ryan

If my understanding is right, you want to put kernel image (text, data and bss) in sram, while make the entire sdram system memory. There is a lot of work to do in order to achieve this, including, 1. change the core to move the reset address to 0x20 off the start of sram (as opposed to adram). 2. define nasys_program_mem as start of sram (as opposed to sdram) 3. Modification of the kernel linker script and setup_arch function. You might be able to bypass this if there is no holes in the address space between sram and sdram (in other words, sdram starts immediately after sram). ... I did not try this. Good luck,

Thanks for the response Wentao, <div class='quotetop'>QUOTE </div> --- Quote Start --- If my understanding is right, you want to put kernel image (text, data and bss) in sram, while make the entire sdram system memory[/b] --- Quote End --- No, this is not quite what I&#39;m going for. Let me explain further. Currently, I have 16MB of onboard SRAM and up to 1GB of SDRAM. My entire kernel and user-level code runs completely in SRAM at the moment. I can explicitly address and store data in SDRAM from my code (but all OS-allocated memory is in SRAM). Basically, the problem here is that I want to be able to DMA data directly from SDRAM to ethernet; with my current setup, this cannot be done, because all kernel data structures are allocated in SRAM. So, calling a socket send() with an SDRAM buffer results in a copy from SDRAM to the SRAM socket buffer. I _think_ this can be mitigated by moving the kernel .DATA and .BSS segments to SDRAM. Another option is to modify the kernel to allocate socket buffers in SDRAM, but I forsee problems with this. I can edit the kernel linker script to put .DATA and .BSS in sdram. But, my SDRAM sits at 0x0 to 0x3FFFFFFF, with SRAM beginning at 0x40000000. I&#39;m guessing this is a problem as well, as my kernel linked this way will not boot. Hope this makes more sense. Ryan

<div class='quotetop'>QUOTE </div> --- Quote Start --- No, this is not quite what I&#39;m going for[/b] --- Quote End --- sorry. <div class='quotetop'>QUOTE </div> --- Quote Start --- I _think_ this can be mitigated by moving the kernel .DATA and .BSS segments to SDRAM[/b] --- Quote End --- sure you can go this way. Please note you also need to modify the bootstrap code (arch/nios2nommu/kernel/head.S) to copy .DATA section into a different place than immediately following .TEXT. <div class='quotetop'>QUOTE </div> --- Quote Start --- Another option is to modify the kernel to allocate socket buffers in SDRAM, but I forsee problems with this[/b] --- Quote End --- Can you tell me what problem you foresee? Currently the DMA zone is empty, you can put all your sdram into that zone and just remember to specify GFP_DMA when your driver wants memory. You don&#39;t need to change the memory management part of the kernel. I think this is easier than the first solution. To add sdram to dma zone, you need to modify function paging_init in file arch/nios2nommu/mm/init.c. Good luck. Let me know if there are problems, wentao Microtronix

Thanks Wentao, your second idea sounds perfect. <div class='quotetop'>QUOTE </div> --- Quote Start --- Can you tell me what problem you foresee?[/b] --- Quote End --- The only real problem I saw with modifying the kernel was the fact that I&#39;m no kernel expert, and it _basically_ works now http://forum.niosforum.com/work2/style_emoticons/<#EMO_DIR#>/smile.gif . But your explanation definitely seems simple enough. <div class='quotetop'>QUOTE </div> --- Quote Start --- Currently the DMA zone is empty, you can put all your sdram into that zone and just remember to specify GFP_DMA when your driver wants memory.[/b] --- Quote End --- From cursory inspection of init.c -- where exactly can I tell the kernel that ZONE_DMA begins at 0x0? I see in paging_init where to specify size -- and it works, but gives me chunks of SRAM right now -- I just can&#39;t find where the zone structure offsets are defined. Thanks again for all the help, Ryan

The function free_area_init assumes the starting address is PAGE_OFFSET. You need to change the definition of PAGE_OFFSET to 0 (the minimum of the starting addresses of all your memory devices) in file include/asm-niosnommu/page.h. Tell me if you have problems

uClinux .data segment in SDRAM? | Altera Community

22 Replies

Altera_Forum
Honored Contributor
21 years ago
If let me guess, I will go to arch/nios2nommu/kernel/setup.c, at the end of setup_arch(). There give all the memory to init_bootmem_node. I am not sure if this is the reason of your problem, but you should be able to find out by stepping through the booting procedures...
Altera_Forum
Honored Contributor
21 years ago
Hi Wentao,

I've been back and forth on this project and another, so sorry for no response last week. I've been through a number of permutations of mappings and kernel changes in attacking this problem, so I'll just post what I have now (and how it's failing) and hopefully you may have some suggestions.

I had our board image remapped to eliminate the 'gap' between SDRAM and SRAM. So, now the map looks like this:

0x0-0x07ffffff -- 128MB of SDRAM
0x08000000-0x09000000 -- 16 MB of SRAM

So, again, I'd like to run code in SRAM but allocate kernel memory in SDRAM. I edited paging_init to give all of SRAM to ZONE_DMA:

zones_size = (0x08000000-PAGE_OFFSET) >> PAGE_SHIFT;

and I changed ZONE_NORMAL accordingly:

zones_size =(end_mem-0x08000000) >> PAGE_SHIFT;

(note: here end_mem is 0x09000000)
Then I edited page.h to# define PAGE_OFFSET as 0. These changes build fine and boot on the new image. However, when I try:

char * str = kmalloc(2048,GFP_DMA)

...I get the following 'page allocation failure':

swapper: page allocation failure. order:0, mode:0x1 Stack from 08327eb8:<0> <0> 00000000<0> 0802dc14<0> 00000000<0> 00000000<0> 00000001<0> 00000003<0> 00000000<0> 00000001<0> <0> 00000000<0> 0817b8ec<0> 0817b8e0<0> 082f2060<0> 0817b8e0<0> deadbeef<0> 0802dc44<0> 08031a54<0> <0> 00000000<0> 08032300<0> deadbeef<0> deadbeef<0> deadbeef<0> deadbeef<0> fffffc18<0> 00000001<0> <0> 00000001<0> 00000001<0> 08031fb8<0> 08c0e000<0> 08169f90<0> deadbeef<0> fffffc18<0> 00000001<0> <0> 00000000<0> 0816c608<0> 081698f8<0> 0816c30c<0> 0816c8d0<0> 00000001<0> 00000001<0> 08169a0c<0> <0> 00000000<0> 0816c890<0> 08000274<0> deadbeef<0> deadbeef<0> deadbeef<0> 081756c0<0> 080044c4<0> Call Trace:<0> <0> <0> <0> <0> <0> <0> <0> <0>

Just wondering if there is anything you could suggest I do. My options are fairly open (that is, I can remap SDRAM somewhere else if necessary). Let me know if anything needs clarified.

Thanks in advance,

Ryan
Quantapoint, Inc.
Altera_Forum
Honored Contributor
21 years ago
go to arch/nios2nommu/kernel/setup.c, after the call of init_bootmem_node , put a free_bootmem for your SDRAM. This will label your SDRAM as free memory.

wentao
Microtronix Datacom Ltd.
Altera_Forum
Honored Contributor
21 years ago
<div class='quotetop'>QUOTE </div>
--- Quote Start ---
go to arch/nios2nommu/kernel/setup.c, after the call of init_bootmem_node , put a free_bootmem for your SDRAM. This will label your SDRAM as free memory[/b]
--- Quote End ---

I did this (after the first free_bootmem call) as below:

free_bootmem(0x0,0x08000000);

...and got the same result as before (page allocation error on kmalloc(2048,GFP_DMA)). Did I miss something?

Ryan
Altera_Forum
Honored Contributor
21 years ago
sorry, that is all I can imagine about this problem. It could be something else causing this problem...
Altera_Forum
Honored Contributor
21 years ago
I occured to me that your problem might be caused by the address you gaven to SDRAM. It is not a good idea to give address 0 to anyone. It will be intepreted as NULL. When you want to allocated a DAM buffer through kmalloc, the slab allocator will try to grab pages, and the first page (with address 0) will be returned, making the slab think there is no memory available...

<div class='quotetop'>QUOTE </div>
--- Quote Start ---
0x0-0x07ffffff -- 128MB of SDRAM
0x08000000-0x09000000 -- 16 MB of SRAM[/b]
--- Quote End ---
Altera_Forum
Honored Contributor
21 years ago
Hi Wentao,

I'm back on this problem after some more immediate issues. Your last suggestion seems to make no difference. At this point, I'm looking for the simplest solution to this problem. It seems like there is a general problem with allocating memory below where the kernel lives.

My question is, what would be the ideal physical mapping for what I am trying to do? I can move SRAM, SDRAM, FLASH, etc wherever you recommend. I'm thinking the easier solution might be to remap the system in a way that the kernel can deal with, rather than hacking at the kernel to get it to use my mapping. Any thoughts? Again, my requirement is that the kernel runs from SRAM but allocates memory from SDRAM.

Thanks again,

Ryan
Altera_Forum
Honored Contributor
21 years ago
actually my system has sram in a lower address than sdram. And there is a hole between them. I tried to add sram to the paging pool and did not find problems. I can allocate buffer from sram. All the mods needed were discussed in this thread.
Altera_Forum
Honored Contributor
21 years ago
Hi Wentao,

I went back and started from scratch on this, and I got the kernel to use SDRAM as GFP_DMA. I can allocate a buffer using 'kmalloc' in SDRAM. However, I have since encountered a couple new problems:

1) I can't seem to kmalloc more than 4MB -- not sure if this is normal behavior or not.
2) I get Perror: Bad Address when I try to send an SDRAM buffer over a TCP socket, regardless of the size of the buffer. Googling for this perror I see that it's caused by an 'invalid address'.

Thanks again, I think I'm getting close,

Ryan
Altera_Forum
Honored Contributor
21 years ago
1). normal
2). I have no idea. Maybe the way you handle the buffer is not right.

Forum Discussion

uClinux .data segment in SDRAM?

22 Replies

Recent Discussions

Nios V/g load-writeback hazard

Run-time errors using EMIF in Agilex 5 NIOS-V design

Ashling RISC Free IDE fails to download ELF file

Dual Port Nios V BSP error

set_output_delay explained for dummies