| Submitter | Bao, Zheng |
|---|---|
| Date | 2009-11-03 03:23:17 |
| Message ID | <DD1CC71B621B004FA76856E5129D6B170329FFA7@sbjgexmb1.amd.com> |
| Download | mbox | patch |
| Permalink | /patch/528/ |
| State | Accepted |
| Headers | show |
Comments
Am 03.11.2009 04:23, schrieb Bao, Zheng: > If the coreboot and filo overlap, it will "slice off" a piece at the > beginning or end. In the beginning case, a new segment is inserted > before the current one. The ptr will move forward and doesn't seem to > have any chance to process the "new" segment. > You are aware that your patch only has an effect for non-compressed payloads? Patrick
On Tue, Nov 3, 2009 at 9:12 AM, Patrick Georgi <patrick@georgi-clan.de> wrote: > Am 03.11.2009 04:23, schrieb Bao, Zheng: >> >> If the coreboot and filo overlap, it will "slice off" a piece at the >> beginning or end. In the beginning case, a new segment is inserted >> before the current one. The ptr will move forward and doesn't seem to >> have any chance to process the "new" segment. >> > > You are aware that your patch only has an effect for non-compressed > payloads? > Patrick and Zheng, I'm struggling to understand the bug. If the payload is uncompressed, it can put a segment before coreboot in the bouncebuffer (this seems to be the bug?). Then the loop needs to be re-run on the newly split/added segment. If it is compressed, It will skip all of the coreboot area and not allocate a segment before coreboot (put the entire thing in the bounce buffer?). Does this get back to the CONFIG_RAMBASE=0x00200000 on fam10? Marc
Marc and Patrick, The LZMA compressing way doesn't work on my board. I haven't found any solution to resolve the overlapping in current code. ulzma() doesn't seem to know that overlapping happens. It is a problem that has to be solved. Do you guys agree that my patch anyway fix the bug for non-compressed payloads? Zheng -----Original Message----- From: Marc Jones [mailto:marcj303@gmail.com] Sent: Wednesday, November 04, 2009 6:42 AM To: Patrick Georgi Cc: Bao, Zheng; coreboot@coreboot.org Subject: Re: [coreboot] [PATCH] The filo crashes if the filo and coreboot overlap. On Tue, Nov 3, 2009 at 9:12 AM, Patrick Georgi <patrick@georgi-clan.de> wrote: > Am 03.11.2009 04:23, schrieb Bao, Zheng: >> >> If the coreboot and filo overlap, it will "slice off" a piece at the >> beginning or end. In the beginning case, a new segment is inserted >> before the current one. The ptr will move forward and doesn't seem to >> have any chance to process the "new" segment. >> > > You are aware that your patch only has an effect for non-compressed > payloads? > Patrick and Zheng, I'm struggling to understand the bug. If the payload is uncompressed, it can put a segment before coreboot in the bouncebuffer (this seems to be the bug?). Then the loop needs to be re-run on the newly split/added segment. If it is compressed, It will skip all of the coreboot area and not allocate a segment before coreboot (put the entire thing in the bounce buffer?). Does this get back to the CONFIG_RAMBASE=0x00200000 on fam10? Marc
Am 04.11.2009 03:34, schrieb Bao, Zheng: > Marc and Patrick, > The LZMA compressing way doesn't work on my board. I haven't found any solution to resolve the overlapping in current code. ulzma() doesn't seem to know that overlapping happens. It is a problem that has to be solved. > Thanks for your fix, but I'd like to come back to the ulzma issue. What do you mean, that LZMA compression doesn't work on your board? There is a known problem that decompression takes _very_ long (several minutes for a moderately sized payload such as FILO). If it looks like the boards was stuck in the decompression phase, please try again and wait to see if it moves on eventually (15 minutes should be enough with some safety margin), so we know if you ran into that known issue, or if you found another bug. If it's something entirely different, I'd also like to hear about it, of course :-) As for ulzma(): ulzma really doesn't know about the overlap, but the compression related code compensates for that. The bounce buffer function returns the start address of the bounce buffer. The location that is used for decompression is (segment_start - RAMBASE + bouncebuffer_base). If the segment starts before the rambase, the segment is decompressed to the bounce buffer and the memory region before it. Right after decompression, the memory region before the bounce buffer is copied. That solution has its own share of problems, but the bounce buffer handling code is quite nasty, and I wanted to keep the changes as small as possible. Thanks, Patrick Georgi
Patch
Index: src/boot/selfboot.c =================================================================== --- src/boot/selfboot.c (revision 4892) +++ src/boot/selfboot.c (working copy) @@ -211,19 +211,21 @@ return !((end <= lb_start) || (start >= lb_end)); } -static void relocate_segment(unsigned long buffer, struct segment *seg) +static int relocate_segment(unsigned long buffer, struct segment *seg) { /* Modify all segments that want to load onto coreboot * to load onto the bounce buffer instead. */ - unsigned long start, middle, end; + /* ret: 1 : A new segment is inserted before the seg. + * 0 : A new segment is inserted after the seg, or no new one. */ + unsigned long start, middle, end, ret = 0; printk_spew("lb: [0x%016lx, 0x%016lx)\n", lb_start, lb_end); /* I don't conflict with coreboot so get out of here */ if (!overlaps_coreboot(seg)) - return; + return 0; start = seg->s_dstaddr;
If the coreboot and filo overlap, it will "slice off" a piece at the beginning or end. In the beginning case, a new segment is inserted before the current one. The ptr will move forward and doesn't seem to have any chance to process the "new" segment. ptr ---------+ move ---> | V +--------+ +--------+ | | | | | new | <---> |current | <---> ..... | | | | +--------+ +--------+ Now we change the ptr to the previous one and restart the loop. The new and current segment will both be processed. +----------------ptr move ---> | V +--------+ +--------+ +--------+ | | | | | | | prev | <---> | new | <---> |current | <---> ..... | | | | | | +--------+ +--------+ +--------+ It is tested on my Family 10 board. Zheng Signed-off-by: Zheng Bao <zheng.bao@amd.com> middle = start + seg->s_filesz; @@ -270,6 +272,8 @@ new->s_dstaddr, new->s_dstaddr + new->s_filesz, new->s_dstaddr + new->s_memsz); + + ret = 1; } /* Slice off a piece at the end @@ -319,6 +323,8 @@ seg->s_dstaddr, seg->s_dstaddr + seg->s_filesz, seg->s_dstaddr + seg->s_memsz); + + return ret; } @@ -446,7 +452,10 @@ /* Modify the segment to load onto the bounce_buffer if necessary. */ - relocate_segment(bounce_buffer, ptr); + if (relocate_segment(bounce_buffer, ptr)) { + ptr = (ptr->prev)->prev; + continue; + } printk_debug("Post relocation: addr: 0x%016lx memsz: 0x%016lx filesz: 0x%016lx\n", ptr->s_dstaddr, ptr->s_memsz, ptr->s_filesz); -----Original Message----- From: coreboot-bounces@coreboot.org [mailto:coreboot-bounces@coreboot.org] On Behalf Of Bao, Zheng Sent: Monday, November 02, 2009 11:25 AM To: Patrick Georgi Cc: coreboot@coreboot.org Subject: Re: [coreboot] The filo crashes if the filo and coreboot overlap. In relocate_segment(). If the coreboot and filo overlap, it will "slice off" a piece at the beginning or end. A new segment is allocated. If it is inserted before the "seg" that is being processed, is there any chance that the "new" segment will be processed? I am confused about it. On my fam 10 board, it seems that the "new" segment was not processed and an error happens when the code jumps to filo which is actually middle of nowhere. Zheng -----Original Message----- From: coreboot-bounces+zheng.bao=amd.com@coreboot.org [mailto:coreboot-bounces+zheng.bao=amd.com@coreboot.org] On Behalf Of Patrick Georgi Sent: Sunday, November 01, 2009 12:13 AM To: Zheng Bao Cc: coreboot@coreboot.org Subject: Re: [coreboot] The filo crashes if the filo and coreboot overlap. Am Samstag, den 31.10.2009, 15:43 +0000 schrieb Zheng Bao: > The filo crashes if the filo and coreboot overlap. > Since the CBFS is the must-have feature, my family 10 > board crashes when it jumps to filo. I am trying to > find out why. I need help. > Based on current code, the AMD Family 10 will cause the filo > and coreboot overlap in RAM. The overlaps_coreboot() in selfboot.c > will return 1. But I am not sure if it will make the system > crashes. What revision is that? There was an issue like that but I fixed it several weeks ago. > If anybody explains briefly what happens if they > overlap. When coreboot and payload overlap, coreboot uses a bounce buffer. The bounce buffer is twice the size of coreboot. The first half is for the part of the payload that overlaps coreboot, the other half is for coreboot itself. The SELF loader loads data that would overlap coreboot to the bounce buffer, and jumps into jmp_to_elf_entry when it's done with loading. The jmp_to_elf_entry function copies coreboot to the upper half of the bounce buffer, and jumps in there, so the code is out of the way. Then it copies the lower half to the coreboot area and jumps to the entry point. There are some complications to that because of the decompression routine, so the code is not as nice as it should be. But I specifically tested your scenario (payload from 1mb to 2.3mb or so, coreboot starting at 2mb) > The coreboot information: > CONFIG_RAMBASE=0x00200000 Try changing that to 0x100000. Patrick