[U-Boot-Users] [PATCH] Fix initrd booting
The device tree needs to be passed to Linux within CFG_BOOTMAPSZ. The current code places the device tree right before the initrd if it exists, and that will usually be closer to the end of memory. Instead, we should always put the device tree right before the bd_info structure, thus ensuring it is within CFG_BOOTMAPSZ.
We do, however, allow for FDTs in uImages to shoot themselves in the foot by requesting a location outside of CFG_BOOTMAPSZ.
Signed-off-by: Andy Fleming afleming@freescale.com --- common/cmd_bootm.c | 20 ++++++++------------ 1 files changed, 8 insertions(+), 12 deletions(-)
diff --git a/common/cmd_bootm.c b/common/cmd_bootm.c index a6499e8..d12ef76 100644 --- a/common/cmd_bootm.c +++ b/common/cmd_bootm.c @@ -753,10 +753,7 @@ do_bootm_linux (cmd_tbl_t *cmdtp, int flag, #else if (*(ulong *)of_flat_tree == OF_DT_HEADER) { #endif -#ifndef CFG_NO_FLASH - if (addr2info((ulong)of_flat_tree) != NULL) - of_data = (ulong)of_flat_tree; -#endif + of_data = (ulong)of_flat_tree; } else if (ntohl(hdr->ih_magic) == IH_MAGIC) { printf("## Flat Device Tree Image at %08lX\n", hdr); print_image_hdr(hdr); @@ -939,11 +936,9 @@ do_bootm_linux (cmd_tbl_t *cmdtp, int flag, ulong of_start, of_len;
of_len = be32_to_cpu(fdt_totalsize(of_data)); + /* position on a 4K boundary before the initrd/kbd */ - if (initrd_start) - of_start = initrd_start - of_len; - else - of_start = (ulong)kbd - of_len; + of_start = (ulong)kbd - of_len; of_start &= ~(4096 - 1); /* align on page */ debug ("## device tree at 0x%08lX ... 0x%08lX (len=%ld=0x%lX)\n", of_data, of_data + of_len - 1, of_len, of_len); @@ -955,6 +950,8 @@ do_bootm_linux (cmd_tbl_t *cmdtp, int flag, if (err != 0) { printf ("libfdt: %s " __FILE__ " %d\n", fdt_strerror(err), __LINE__); } + printf("OK\n"); + /* * Add the chosen node if it doesn't exist, add the env and bd_t * if the user wants it (the logic is in the subroutines). @@ -982,11 +979,9 @@ do_bootm_linux (cmd_tbl_t *cmdtp, int flag, if (of_data) { ulong of_start, of_len; of_len = ((struct boot_param_header *)of_data)->totalsize; + /* provide extra 8k pad */ - if (initrd_start) - of_start = initrd_start - of_len - 8192; - else - of_start = (ulong)kbd - of_len - 8192; + of_start = (ulong)kbd - of_len - 8192; of_start &= ~(4096 - 1); /* align on page */ debug ("## device tree at 0x%08lX ... 0x%08lX (len=%ld=0x%lX)\n", of_data, of_data + of_len - 1, of_len, of_len); @@ -995,6 +990,7 @@ do_bootm_linux (cmd_tbl_t *cmdtp, int flag, printf (" Loading Device Tree to %08lx, end %08lx ... ", of_start, of_start + of_len - 1); memmove ((void *)of_start, (void *)of_data, of_len); + printf ("OK\n"); } #endif
Hi Andy,
I'm not sure if you are aiming this at u-boot-fdt, which would be logical, or if you are aiming this at u-boot-testing (or other). Care to clarify?
Andy Fleming wrote:
The device tree needs to be passed to Linux within CFG_BOOTMAPSZ. The current code places the device tree right before the initrd if it exists, and that will usually be closer to the end of memory. Instead, we should always put the device tree right before the bd_info structure, thus ensuring it is within CFG_BOOTMAPSZ.
We do, however, allow for FDTs in uImages to shoot themselves in the foot by requesting a location outside of CFG_BOOTMAPSZ.
Signed-off-by: Andy Fleming afleming@freescale.com
common/cmd_bootm.c | 20 ++++++++------------ 1 files changed, 8 insertions(+), 12 deletions(-)
diff --git a/common/cmd_bootm.c b/common/cmd_bootm.c index a6499e8..d12ef76 100644 --- a/common/cmd_bootm.c +++ b/common/cmd_bootm.c @@ -753,10 +753,7 @@ do_bootm_linux (cmd_tbl_t *cmdtp, int flag, #else if (*(ulong *)of_flat_tree == OF_DT_HEADER) { #endif -#ifndef CFG_NO_FLASH
if (addr2info((ulong)of_flat_tree) != NULL)of_data = (ulong)of_flat_tree;-#endif
of_data = (ulong)of_flat_tree;
Is this right? The logic in bootm *in general* and specifically here is hard to figure out, but the way I'm puzzling it out... 1) If of_data is *not* NULL, it is used as the indicator that the blob must be relocated. 2) The above is saying that, if there is flash (not defined CFG_NO_FLASH gaak) and if of_flat_tree points into flash (addr2info returns a pointer), then of_data is set which makes #1 true.
As I read it, this causes the blob to be copied to RAM if it is in flash, which is necessary since we cannot add/rewrite nodes and properties in a flash copy.
Your change causes the blob to _always_ be relocated. Not necessarily bad, but is it always good? I'm not a bootm expert by any means, and I suspect the complexity grew substantially over time, but in the mucking about I did, I did not dare to change the logic. :-/
} else if (ntohl(hdr->ih_magic) == IH_MAGIC) { printf("## Flat Device Tree Image at %08lX\n", hdr); print_image_hdr(hdr);
[snip]
Thanks, gvb
In message 46B879B1.4020207@gmail.com you wrote:
I'm not sure if you are aiming this at u-boot-fdt, which would be logical, or if you are aiming this at u-boot-testing (or other). Care to clarify?
My understanding was it's for u-boot-fdt (i. e. I didn't pick it up for u-boot-testing).
As I read it, this causes the blob to be copied to RAM if it is in flash, which is necessary since we cannot add/rewrite nodes and properties in a flash copy.
That's my understanding, too.
Your change causes the blob to _always_ be relocated. Not necessarily
...which IMHO is NOT a good idea.
Best regards,
Wolfgang Denk
On 8/7/07, Jerry Van Baren gvb.uboot@gmail.com wrote:
Hi Andy,
I'm not sure if you are aiming this at u-boot-fdt, which would be logical, or if you are aiming this at u-boot-testing (or other). Care to clarify?
To be honest, I don't care where it goes, but without this patch, anyone booting on e500 with a Ramdisk will run into problems unless they are lucky.
#else if (*(ulong *)of_flat_tree == OF_DT_HEADER) { #endif -#ifndef CFG_NO_FLASH
if (addr2info((ulong)of_flat_tree) != NULL)of_data = (ulong)of_flat_tree;-#endif
of_data = (ulong)of_flat_tree;Is this right? The logic in bootm *in general* and specifically here is hard to figure out, but the way I'm puzzling it out...
- If of_data is *not* NULL, it is used as the indicator that the blob must be relocated.
- The above is saying that, if there is flash (not defined CFG_NO_FLASH gaak) and if of_flat_tree points into flash (addr2info returns a pointer), then of_data is set which makes #1 true.
As I read it, this causes the blob to be copied to RAM if it is in flash, which is necessary since we cannot add/rewrite nodes and properties in a flash copy.
Your change causes the blob to _always_ be relocated. Not necessarily bad, but is it always good? I'm not a bootm expert by any means, and I suspect the complexity grew substantially over time, but in the mucking about I did, I did not dare to change the logic. :-/
On e500, the blob *must* be in the low 16M of memory. MUST. The low 16M is all that is mapped, and the kernel will not map more than that until it reads the blob. Certainly, it doesn't *always* need to be relocated. It's been a while since I created this patch, but I suppose it's possible to modify the logic to properly only relocate it if it's in Flash OR if it's not in the low 16M. I think that, at the time, I determined that this was the least intrusive change which would result in everything working.
I probably need to respin it, though. I haven't updated it since I first created it, and you may have changed the fdt code, since.
Andy
Dear Andy,
in message 2acbd3e40708071212y6396de53l49b8b48e6aab9a5f@mail.gmail.com you wrote:
On e500, the blob *must* be in the low 16M of memory. MUST. The low 16M is all that is mapped, and the kernel will not map more than that until it reads the blob. Certainly, it doesn't *always* need to be relocated. It's been a while since I created this patch, but I
I think the blob should only be copied (to me "relocate" involves more complex operations than just copying) when necessary.
And I guess this restriction is also true for ramdisk images, or am I wrong?
In any case, I think it wouldbe a good idea if the 16M limit was not hard wired, but could be overwritten using the "initrd_high" variable like we can do for ramdisks.
What do you think?
I probably need to respin it, though. I haven't updated it since I first created it, and you may have changed the fdt code, since.
Thanks.
Best regards,
Wolfgang Denk
Wolfgang Denk wrote:
Dear Andy,
in message 2acbd3e40708071212y6396de53l49b8b48e6aab9a5f@mail.gmail.com you wrote:
On e500, the blob *must* be in the low 16M of memory. MUST. The low 16M is all that is mapped, and the kernel will not map more than that until it reads the blob. Certainly, it doesn't *always* need to be relocated. It's been a while since I created this patch, but I
I think the blob should only be copied (to me "relocate" involves more complex operations than just copying) when necessary.
And I guess this restriction is also true for ramdisk images, or am I wrong?
In any case, I think it wouldbe a good idea if the 16M limit was not hard wired, but could be overwritten using the "initrd_high" variable like we can do for ramdisks.
What do you think?
I probably need to respin it, though. I haven't updated it since I first created it, and you may have changed the fdt code, since.
Thanks.
Best regards,
Wolfgang Denk
Hi Andy, Wolfgang,
The chunk I questioned was:
@@ -753,10 +753,7 @@ do_bootm_linux (cmd_tbl_t *cmdtp, int flag, #else if (*(ulong *)of_flat_tree == OF_DT_HEADER) { #endif -#ifndef CFG_NO_FLASH
if (addr2info((ulong)of_flat_tree) != NULL)of_data = (ulong)of_flat_tree;-#endif
of_data = (ulong)of_flat_tree;
Hi Andy,
IMHO, you are abusing unrelated logic to force the blob to be copied. Why not directly add the logic you want rather than removing the flash logic (warning, untested code):
#ifdef CFG_BOOTMAPSZ /* * The blob MUST be within CFG_BOOTMAPSZ, flag it to * be copied if not. */ if (of_flat_tree >= CFG_BOOTMAPSZ) of_data = of_flat_tree; #endif
Best regards, gvb
P.S. The way of_data is used makes my head hurt - combined boolean/address masquerading as an int used by various unrelated snippets of logic (all uncommented) to achieve a single goal.
participants (4)
-
Andy Fleming -
Andy Fleming -
Jerry Van Baren -
Wolfgang Denk