On Fri, Feb 9, 2018 at 8:33 PM, Michael Clark <m...@sifive.com> wrote:
> > > On Fri, Feb 9, 2018 at 5:35 AM, Richard Henderson < > richard.hender...@linaro.org> wrote: > >> On 02/07/2018 05:28 PM, Michael Clark wrote: >> > +++ b/hw/riscv/riscv_elf.c >> > @@ -0,0 +1,244 @@ >> > +/* >> > + * elf.c - A simple package for manipulating symbol tables in elf >> binaries. >> > + * >> > + * Taken from >> > + * https://www.cs.cmu.edu/afs/cs.cmu.edu/academic/class/15213-f03/www/ >> > + * ftrace/elf.c >> > + * >> > + */ >> > +#include <stdio.h> >> > +#include <stdlib.h> >> > +#include <unistd.h> >> > +#include <string.h> >> > +#include <sys/stat.h> >> > +#include <sys/types.h> >> > +#include <sys/mman.h> >> > +#include <errno.h> >> > +#include <fcntl.h> >> > +#include <glib.h> >> > + >> > +#include "hw/riscv/riscv_elf.h" >> > + >> > +/* >> > + * elf_open - Map a binary into the address space and extract the >> > + * locations of the static and dynamic symbol tables and their string >> > + * tables. Return this information in a Elf object file handle that >> will >> > + * be passed to all of the other elf functions. >> > + */ >> > +Elf_obj64 *elf_open64(const char *filename) >> > +{ >> > + int i, fd; >> > + struct stat sbuf; >> > + Elf_obj64 *ep; >> > + Elf64_Shdr *shdr; >> > + >> > + ep = g_new(Elf_obj64, 1); >> > + >> > + /* Do some consistency checks on the binary */ >> > + fd = open(filename, O_RDONLY); >> > + if (fd == -1) { >> > + fprintf(stderr, "Can't open \"%s\": %s\n", filename, >> strerror(errno)); >> > + exit(1); >> > + } >> > + if (fstat(fd, &sbuf) == -1) { >> > + fprintf(stderr, "Can't stat \"%s\": %s\n", filename, >> strerror(errno)); >> > + exit(1); >> > + } >> > + if (sbuf.st_size < sizeof(Elf64_Ehdr)) { >> > + fprintf(stderr, "\"%s\" is not an ELF binary object\n", >> filename); >> > + exit(1); >> > + } >> > + >> > + /* It looks OK, so map the Elf binary into our address space */ >> > + ep->mlen = sbuf.st_size; >> > + ep->maddr = mmap(NULL, ep->mlen, PROT_READ, MAP_SHARED, fd, 0); >> > + if (ep->maddr == (void *)-1) { >> > + fprintf(stderr, "Can't mmap \"%s\": %s\n", filename, >> strerror(errno)); >> > + exit(1); >> > + } >> > + close(fd); >> > + >> > + /* The Elf binary begins with the Elf header */ >> > + ep->ehdr = ep->maddr; >> > + >> > + /* check we have a 64-bit little-endian RISC-V ELF object */ >> > + if (ep->ehdr->e_ident[EI_MAG0] != ELFMAG0 || >> > + ep->ehdr->e_ident[EI_MAG1] != ELFMAG1 || >> > + ep->ehdr->e_ident[EI_MAG2] != ELFMAG2 || >> > + ep->ehdr->e_ident[EI_MAG3] != ELFMAG3 || >> > + ep->ehdr->e_ident[EI_CLASS] != ELFCLASS64 || >> > + ep->ehdr->e_ident[EI_DATA] != ELFDATA2LSB || >> > + ep->ehdr->e_machine != EM_RISCV) >> > + { >> > + fprintf(stderr, "\"%s\" is not a 64-bit RISC-V ELF object\n", >> filename); >> > + exit(1); >> > + } >> > + >> > + /* >> > + * Find the static and dynamic symbol tables and their string >> > + * tables in the the mapped binary. The sh_link field in symbol >> > + * table section headers gives the section index of the string >> > + * table for that symbol table. >> > + */ >> > + shdr = (Elf64_Shdr *)(ep->maddr + ep->ehdr->e_shoff); >> >> This duplicates, badly, existing code within include/hw/elf_ops.h. >> >> In particular, this fails to bswap these fields. As such, this code will >> only >> work on little-endian hosts. >> > > There is enough information in the structures held in the syminfo table > used by the disassembler however elf_ops.h currently throws away all > symbols that are not STT_FUNC, and we are after STT_OBJECT symbols. We > would need to add an option to load_elf to load the entire symbol table. > Also, it runs qsort on the table by address, so we'd need to do a linear > scan for the HTIF symbols by name. There are only two symbols we need: > 'fromhost' and 'tohost'. > > /* We are only interested in function symbols. > Throw everything else away. */ > if (syms[i].st_shndx == SHN_UNDEF || > syms[i].st_shndx >= SHN_LORESERVE || > ELF_ST_TYPE(syms[i].st_info) != STT_FUNC) { > nsyms--; > if (i < nsyms) { > syms[i] = syms[nsyms]; > } > continue; > } > > It would be really nice if HTIF just specified registers in device-tree; > alas it doesn't; and there is still quite a bit of code that depend on > using these magic HTIF 'fromhost' and 'tohost' symbols. riscv-tests for one > uses the HTIF mechanism and spike the official ISA simulator uses HTIF for > IO. > Luckily elf_ops.h load_elf32 and load_elf64 are wrapped by load_elf, and several load_elf_* variants. We need the symbol name, type, address and size to find and validate the HTIF symbols. The current thought is to pass a callback function pointer to a new load_elf_sym variant, where the new function pointer, in the common case, won't be called by load_elf32 and load_elf64 as it will be NULL for all present users. We could use uint64_t to abstract the caller from the elf32_sym/elf64_sym address and size types. This approach would likely require less code than adding a new /linear/ search by name method, along with a retain all symbols argument to load_symbols and associated plumbing. It's also inherently one pass e.g. typedef void (*symbol_fn_t)(const char *st_name, int st_info, uint64_t st_value, uint64_t st_size); static int glue(load_symbols, SZ)(struct elfhdr *ehdr, int fd, int must_swab, int clear_lsb, symbol_fn_t sym_cb) { .... if (sym_cb) { sym_cb(strtab + syms[i].st_name, syms[i].st_info, syms[i].st_value, syms[i].st_size) } ... } static int glue(load_elf, SZ)(const char *name, int fd, uint64_t (*translate_fn)(void *, uint64_t), void *translate_opaque, int must_swab, uint64_t *pentry, uint64_t *lowaddr, uint64_t *highaddr, int elf_machine, int clear_lsb, int data_swab, AddressSpace *as, bool load_rom, symbol_fn_t sym_cb) { .... glue(load_symbols, SZ)(&ehdr, fd, must_swab, clear_lsb, sym_cb); .... }
