coredump: extend core dump note section to contain file names of mapped files
Subsystems (2)
PATCH v2 1/1 coredump: extend core dump note section to contain file names of mapped files
This note has the following format:
long count -- how many files are mapped
long page_size -- units for file_ofs
array of [COUNT] elements of
long start
long end
long file_ofs
followed by COUNT filenames in ASCII: "FILE1" NUL "FILE2" NUL...
If file name is not available, "" string is encoded.
Changes since previous version: rediffed on top of NT_SIGINFO patches.
--- fs/binfmt_elf.c | 85 ++++++++++++++++++++++++++++++++++++++++++++++++-- include/linux/elf.h | 1 + 2 files changed, 82 insertions(+), 4 deletions(-)
fs/binfmt_elf.c +81 -4
@@ -1380,6 +1380,72 @@ static void fill_siginfo_note(struct memelfnote *note, siginfo_t *csigdata, sigifill_note (note,"CORE" ,NT_SIGINFO ,sizeof (*csigdata), csigdata); } +#define MAX_FILE_NOTE_SIZE (4*1024*1024) + +static void fill_files_note (struct memelfnote *note) +{ +struct vm_area_struct *vma; +struct file *file; +unsigned count, word_count, size, remaining; +long *data; +long *start_end_ofs; +char *name; + + count =0 ; +for (vma = current->mm->mmap; vma !=NULL ; vma = vma->vm_next) { + file = vma->vm_file; +if (!file) +continue ; + count++; + } + + size = count *64 ; + word_count =2 +3 * count; + alloc: +if (size >=MAX_FILE_NOTE_SIZE )/* paranoia check */ + goto err; + size =round_up (size,PAGE_SIZE ); + data =vmalloc (size); +if (!data) + goto err; + + start_end_ofs = data; + name = (void *)&start_end_ofs[word_count]; + remaining = size - word_count *sizeof (long ); + + *start_end_ofs++ = count; + *start_end_ofs++ =PAGE_SIZE ; +for (vma = current->mm->mmap; vma !=NULL ; vma = vma->vm_next) { +const char *filename; + + file = vma->vm_file; +if (!file) +continue ; + filename =d_path (&file->f_path, name, remaining); +if (IS_ERR (filename)) { +if (PTR_ERR (filename) == -ENAMETOOLONG ) { +vfree (data); + size = size *5 /4 ; + goto alloc; + } +/* continue; - WRONG, we must have COUNT elements */ + filename ="" ; + } +/* d_path() fills at the end, move it to front */ +do + remaining--; +while ((*name++ = *filename++) !='\0' ); + + *start_end_ofs++ = vma->vm_start; + *start_end_ofs++ = vma->vm_end; + *start_end_ofs++ = vma->vm_pgoff; + } + + size = name - (char *)data; +fill_note (note,"CORE" ,NT_FILE , size, data); + err: ; +} +#ifdef CORE_DUMP_USE_REGSET #include <linux/regset.h>
@@ -1395,6 +1461,7 @@ struct elf_note_info {struct memelfnote psinfo;struct memelfnote signote;struct memelfnote auxv; +struct memelfnote files;siginfo_t csigdata;size_t size;int thread_notes;
@@ -1575,6 +1642,9 @@ static int fill_note_info(struct elfhdr *elf, int phdrs,fill_auxv_note (&info->auxv, current->mm); info->size +=notesize (&info->auxv); +fill_files_note (&info->files); + info->size +=notesize (&info->files); +return 1 ; }
@@ -1605,6 +1675,8 @@ static int write_note_info(struct elf_note_info *info,return 0 ;if (first && !writenote (&info->auxv, file, foffset))return 0 ; +if (first && !writenote (&info->files, file, foffset)) +return 0 ;for (i =1 ; i < info->thread_notes; ++i) if (t->notes[i].data &&
@@ -1631,6 +1703,7 @@ static void free_note_info(struct elf_note_info *info) kfree(t); }kfree (info->psinfo.data); +vfree (info->files.data); }#else
@@ -1707,7 +1780,7 @@ static int elf_note_info_init(struct elf_note_info *info)INIT_LIST_HEAD (&info->thread_list);/* Allocate space for ELF notes */ - info->notes =kmalloc (7 *sizeof (struct memelfnote ),GFP_KERNEL ); + info->notes =kmalloc (8 *sizeof (struct memelfnote ),GFP_KERNEL );if (!info->notes)return 0 ; info->psinfo =kmalloc (sizeof (*info->psinfo),GFP_KERNEL );
@@ -1777,10 +1850,11 @@ static int fill_note_info(struct elfhdr *elf, int phdrs,fill_note (info->notes +1 ,"CORE" ,NT_PRPSINFO ,sizeof (*info->psinfo), info->psinfo); - info->numnote =2 ; +fill_siginfo_note (info->notes +2 , &info->csigdata, siginfo); +fill_auxv_note (info->notes +3 , current->mm); +fill_files_note (info->notes +4 ); -fill_siginfo_note (&info->notes[info->numnote++], &info->csigdata, siginfo); -fill_auxv_note (&info->notes[info->numnote++], current->mm); + info->numnote =5 ;/* Try to dump the FPU. */ info->prstatus->pr_fpvalid = elf_core_copy_task_fpregs(current, regs,
@@ -1842,6 +1916,9 @@ static void free_note_info(struct elf_note_info *info)kfree (list_entry(tmp ,struct elf_thread_status ,list )); } +/* Free data allocated by fill_files_note(): */ +vfree (info->notes[4 ].data); +kfree (info->prstatus);kfree (info->psinfo);kfree (info->notes);
include/linux/elf.h +1 -0
@@ -377,6 +377,7 @@ typedef struct elf64_shdr {* in the future to accomodate more fields, don't assume it is fixed! */ #define NT_SIGINFO 0x53494749 +#define NT_FILE 0x46494c45#define NT_PRXFPREG 0x46e62b7f/* copied from gdb5.1/include/elf/common.h */ #define NT_PPC_VMX 0x100/* PowerPC Altivec/VMX registers */ #define NT_PPC_SPE 0x101/* PowerPC SPE/EVR registers */
Signature · 1 line
1.7.7.6
Quoted 2 lines · Denys Vlasenko
Changes since previous version: rediffed on top of
NT_SIGINFOpatches.
Assuming this is the only change:
Again, I can't judge whether we need this feature or not, although I tend to trust Denys who actually works with the user-space tools which play with coredumps.
But, from the correctness pov
Jonathan M. Foote
Reviewed-by Oleg NesterovThis change will be beneficial for automated vulnerability triage as well. It will allow tools (like http://www.cert.org/vuls/discovery/triage.html) to gather more information about where an application crashed and the addresses in the backtrace. This information can then be used to help determine the severity of the underlying bug.
Jon
-----Original Message-----
From: Oleg Nesterov [mailto:[email protected]]
Sent: Tuesday, September 18, 2012 11:37 AM
To: Denys Vlasenko
Cc: [email protected]; Andrew Morton; Amerigo Wang; Jonathan M. Foote; Roland McGrath; Pedro Alves
Subject: Re: [PATCH -mm v2] coredump: extend core dump note section to contain file names of mapped files
Quoted 2 lines · Denys Vlasenko
Changes since previous version: rediffed on top of
NT_SIGINFOpatches.
Assuming this is the only change:
Again, I can't judge whether we need this feature or not, although I tend to trust Denys who actually works with the user-space tools which play with coredumps.
But, from the correctness pov
The code needs to be macroized a bit so that compat_binfmt_elf.c will produce a version that encodes 32-bit values correctly so as to be compatible with the output of a native 32-bit kernel.
It's doubtful that rolling your own single loop actually performs better than just calling strlen and memcpy.
Since you're just counting to estimate the size of a temporary buffer, you could skip the initial loop and just estimate based on mm->map_count. Then collect the count in the main loop and write the first word of the buffer last.
Thanks,
Roland
Quoted 3 lines · Roland McGrath
It's doubtful that rolling your own single loop actually performs better than just calling strlen and memcpy.
This is needed to advance 'name' if d_path() fails with !ENAMETOOLONG. Probably needs a comment.
Or perhaps 'filename = ""' case can handle this case itself. In this case we do not even need strlen(), remaining = name - filename.
Oleg.
Quoted 5 lines · Oleg Nesterov
Quoted 3 lines · Roland McGrath
It's doubtful that rolling your own single loop actually performs better than just calling strlen and memcpy.
Ah, I misread your message, sorry...
Quoted 5 lines
This is needed to advance 'name' if
d_path()fails with !ENAMETOOLONG. Probably needs a comment.Or perhaps 'filename = ""' case can handle this case itself. In this case we do not even need
strlen(), remaining = name - filename.
Yes, but you strlen + memcpy as you suggested will work too.
Oleg.
Quoted 3 lines · Oleg Nesterov
Or perhaps 'filename = ""' case can handle this case itself. In this case we do not even need
strlen(), remaining = name - filename.
IOW. Denys, this is up to you and _iirc_ this was already discussed (just I don't remember the result), but perhaps
if (IS_ERR(filename)) {
if (-ENAMETOOLONG) {
...
goto retry;
}
remaining--;
*name++ = 0;
} else {
remaining = name - filename;
strcpy(name, filename);
}Hmm. And I just noticed that 'filename = ""' case does not verify there is a room to write a single '\0' :/
So, afaics you also need
- if (-ENAMETOOLONG)
+ if (-ENAMETOOLONG || !remaining)
or I missed something.
Oleg.
Quoted 8 lines · Oleg Nesterov
Hmm. And I just noticed that 'filename = ""' case does not verify there is a room to write a single '\0' :/
So, afaics you also need
- if (
-ENAMETOOLONG)
+ if (-ENAMETOOLONG|| !remaining)
No, please ignore.
d_path() always starts with prepend("\0", 1), it should always return ENAMETOOLONG if !remaining.
Oleg.
Quoted 3 lines · Roland McGrath
The code needs to be macroized a bit so that
compat_binfmt_elf.cwill produce a version that encodes 32-bit values correctly so as to be compatible with the output of a native 32-bit kernel.
Just did it...
Quoted 2 lines
It's doubtful that rolling your own single loop actually performs better than just calling strlen and memcpy.
I didn't do it for speed, rather for simplicity.
Quoted 4 lines
Since you're just counting to estimate the size of a temporary buffer, you could skip the initial loop and just estimate based on
mm->map_count. Then collect the count in the main loop and write the first word of the buffer last.
The format is:
long count -- how many files are mapped
long page_size -- units for file_ofs
array of [COUNT] elements of
long start
long end
long file_ofs
followed by COUNT filenames in ASCII: "FILE1" NUL "FILE2" NUL...
In order to fill in file names, we need to know exact number of array elements, since file names immediately follow that array. IOW:
+ for (vma = current->mm->mmap; vma != NULL; vma = vma->vm_next) {
+ if (it is a file-backed vma)
+ count++;
+ }
+
+ size = count * 64;
+ word_count = 2 + 3 * count; <=====
...
+ name = (void*)&start_end_ofs[word_count]; <=====
The marked statements are crucial. We can't use estimated count, we need exact one here in order for 'name' pointer to be correct.
Otherwise (if we would use count = mm->map_count), we'd need to memmove block of file names at the end. And we might end up allocating significant amounts of memory we don't need if there are tons of anon mappings. IMO this is worse than the counting loop.
Anyway, I will shortly send a patch which implements your logic.
Signature · 1 line
vda