On Sun, Mar 31, 2013 at 06:53:00PM +0300, Mikolaj Golub wrote: > On Sun, Mar 31, 2013 at 04:40:47PM +0300, Konstantin Belousov wrote: > > > I inspected imgact_elf.c:parse_note(), imgact_elf.c:putnote() and > > rtld.c:digest_notes(). Only putnote() uses 8-byte alignment. > > Every other OS and our !coredump code assumes 4-byte alignment. > > Thanks! > > > Does changing the putnote() to align on the 4-byte boundary cause > > real change in the core file notes layout ? > > Currently, we store only 4 types of notes in a core file: > > #define NT_PRSTATUS 1 /* Process status. */ > #define NT_FPREGSET 2 /* Floating point registers. */ > #define NT_PRPSINFO 3 /* Process state info. */ > #define NT_THRMISC 7 /* Thread miscellaneous info. */ > > I checked the sizes of structures inserted into the notes, and on amd64 > they all are multiple of 8: > > (kgdb) p sizeof(prpsinfo_t) % 8 > $1 = 0 > (kgdb) p sizeof(prstatus_t) % 8 > $2 = 0 > (kgdb) p sizeof(prfpregset_t) % 8 > $3 = 0 > (kgdb) p sizeof(thrmisc_t) % 8 > $4 = 0 > > so both 4-byte and 8-byte aligned. Well, FreeBSD supports some more 64bit architectures, besides amd64. At least on powerpc64, I get prpsinfo 120 0 prstatus_t 344 0 prfpregset_t 264 0 thrmisc_t 24 0
Second column is sizeof(), third is sizeof() % 8. This is in fact not much surprising, since all ABIs define the alignment of the structure as the alignment of the most demanding member. And, because 64bit architectures have 8-byte registers, it is indeed expected that the size % 8 == 0. > > I believe that the patch below will not change the current core file > notes layout, will make things consistent in our tree, and will make > adding my procstat notes easier, if I use 4-byte alignment. > > Are you ok if I commit it before introducing my changes? Yes, I believe this is the right thing to do. > > Index: sys/kern/imgact_elf.c > =================================================================== > --- sys/kern/imgact_elf.c (revision 248706) > +++ sys/kern/imgact_elf.c (working copy) > @@ -1538,10 +1538,10 @@ __elfN(putnote)(void *dst, size_t *off, const char > *off += sizeof note; > if (dst != NULL) > bcopy(name, (char *)dst + *off, note.n_namesz); > - *off += roundup2(note.n_namesz, sizeof(Elf_Size)); > + *off += roundup2(note.n_namesz, sizeof(Elf32_Size)); > if (dst != NULL) > bcopy(desc, (char *)dst + *off, note.n_descsz); > - *off += roundup2(note.n_descsz, sizeof(Elf_Size)); > + *off += roundup2(note.n_descsz, sizeof(Elf32_Size)); > } > > static boolean_t > > Also, shouldn't we update then the following comment in sys/elf_common.h? > > /* > * Note header. The ".note" section contains an array of notes. Each > * begins with this header, aligned to a word boundary. Immediately > * following the note header is n_namesz bytes of name, padded to the > * next word boundary. Then comes n_descsz bytes of descriptor, again > * padded to a word boundary. The values of n_namesz and n_descsz do > * not include the padding. > */ > > -- > Mikolaj Golub
pgpt3ztmYjZwj.pgp
Description: PGP signature
