From: Arnaldo Carvalho de Melo <acme@infradead.org>
To: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com>
Cc: Dave Martin <dave.martin@linaro.org>,
peterz@infradead.org, akpm@linux-foundation.org,
torvalds@linux-foundation.org, ananth@in.ibm.com,
jkenisto@linux.vnet.ibm.com, linux-kernel@vger.kernel.org,
linux-mm@kvack.org, oleg@redhat.com, andi@firstfloor.org,
hch@infradead.org, rostedt@goodmis.org,
masami.hiramatsu.pt@hitachi.com, tglx@linutronix.de,
anton@redhat.com, srikar@linux.vnet.ibm.com,
linux-perf-users@vger.kernel.org, mingo@elte.hu,
Roland McGrath <roland@hack.frob.com>
Subject: Re: Fwd: [PATCH-V2] perf symbols: fix symbol offset breakage with separated debug info
Date: Mon, 11 Jun 2012 10:53:52 -0300 [thread overview]
Message-ID: <20120611135352.GA2202@infradead.org> (raw)
In-Reply-To: <4FD5D3CE.2010307@linux.vnet.ibm.com>
Adding Roland to the CC list, perhaps he can help us with these
debuginfo details :-)
Em Mon, Jun 11, 2012 at 04:47:34PM +0530, Prashanth Nageshappa escreveu:
> Dave/Arnaldo,
> Let me know if you have any comments on this patch so that I can address them.
> -------- Original Message --------
> Subject: [PATCH-V2] perf symbols: fix symbol offset breakage with separated debug info
> Date: Wed, 02 May 2012 12:32:06 +0530
> From: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com>
> To: peterz@infradead.org, mingo@elte.hu
> CC: akpm@linux-foundation.org, torvalds@linux-foundation.org, ananth@in.ibm.com, jkenisto@linux.vnet.ibm.com, linux-kernel@vger.kernel.org, linux-mm@kvack.org, oleg@redhat.com, andi@firstfloor.org, hch@infradead.org, rostedt@goodmis.org, acme@infradead.org, masami.hiramatsu.pt@hitachi.com, tglx@linutronix.de, anton@redhat.com, srikar@linux.vnet.ibm.com, dave.martin@linaro.org, linux-perf-users@vger.kernel.org
>
> From: Dave Martin <dave.martin@linaro.org>
>
> perf resolves symbols to wrong offsets when debug info is separated
> from the lib/executable.
>
> This patch is based on Dave Martin's initial work first published at
> https://lkml.org/lkml/2010/11/22/137
>
> This patch loads the ELF section headers from a separate file if
> necessary, to avoid getting confused by the different section file
> offsets seen in debug images. Invalid section headers are detected by
> checking for the presence of non-writable SHT_NOBITS sections, which
> don't make sense under normal circumstances.
>
> In particular, this allows symbols in ET_EXEC images to get fixed up
> correctly in the presence of separated debug images.
>
> v2 addresses review comments (to remove asserts) from Dave Martin.
>
> Signed-off-by: Dave Martin <dave.martin@linaro.org>
> Signed-off-by: Prashanth Nageshappa <prashanth@linux.vnet.ibm.com>
> ---
>
> tools/perf/util/symbol.c | 168 +++++++++++++++++++++++++++++++++++++++++++++-
> 1 files changed, 164 insertions(+), 4 deletions(-)
>
> diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c
> index caaf75a..a6ad4b1 100644
> --- a/tools/perf/util/symbol.c
> +++ b/tools/perf/util/symbol.c
> @@ -1158,8 +1158,105 @@ static size_t elf_addr_to_index(Elf *elf, GElf_Addr addr)
> return -1;
> }
>
> +/**
> + * Read all section headers, copying them into a separate array so they survive
> + * elf_end.
> + *
> + * @elf: the libelf instance to operate on.
> + * @ehdr: the elf header: this must already have been read with gelf_getehdr().
> + * @count: the number of headers read is assigned to *count on successful
> + * return. count must not be NULL.
> + *
> + * Returns a pointer to the allocated headers, which should be deallocated with
> + * free() when no longer needed.
> + */
> +static GElf_Shdr *elf_get_all_shdrs(Elf *elf, GElf_Ehdr const *ehdr,
> + unsigned *count)
> +{
> + GElf_Shdr *shdrs;
> + Elf_Scn *scn;
> + unsigned max_index = 0;
> + unsigned i;
> +
> + shdrs = malloc(ehdr->e_shnum * sizeof *shdrs);
This could be zalloc so that...
> + if (!shdrs)
> + return NULL;
> +
> + for (i = 0; i < ehdr->e_shnum; i++)
> + shdrs[i].sh_type = SHT_NULL;
The loop above can be removed
> + for (scn = NULL; (scn = elf_nextscn(elf, scn)); ) {
> + size_t j;
> +
> + /*
> + * Just assuming we get section 0, 1, ... in sequence may lead
> + * to wrong section indices. Check the index explicitly:
> + */
> + j = elf_ndxscn(scn);
> + if (j > max_index)
> + max_index = j;
> +
> + if (!gelf_getshdr(scn, &shdrs[j]))
> + goto error;
> + }
> +
> + *count = max_index + 1;
> + return shdrs;
> +
> +error:
> + free(shdrs);
> + return NULL;
> +}
> +
> +/**
> + * Check that the section headers @shdrs reflect accurately the file data
> + * layout of the image that was loaded during perf record. This is generally
> + * not true for separated debug images generated with e.g.,
> + * objcopy --only-keep-debug.
> + *
> + * We identify invalid files by checking for non-empty sections which are
> + * declared as having no file data (SHT_NOBITS) but are not writable.
> + *
> + * @shdrs: the full set of section headers, as loaded by elf_get_all_shdrs().
> + * @count: the number of headers present in @shdrs.
> + *
> + * Returns 1 for valid headers, 0 otherwise.
> + */
Roland, could you take a look at these ELF/debuginfo aspects?
> +static int elf_check_shdrs_valid(GElf_Shdr const *shdrs, unsigned count)
> +{
> + unsigned i;
> +
> + for (i = 0; i < count; i++) {
> + if (shdrs[i].sh_type == SHT_NOBITS &&
> + !(shdrs[i].sh_flags & SHF_WRITE) &&
> + shdrs[i].sh_size != 0)
> + return 0;
> + }
> +
> + return 1;
> +}
> +
> +/*
> + * Notes:
> + *
> + * If saved_shdrs is non-NULL, the section headers will be read if found, and
> + * will be used for address fixups. saved_shdrs_count must also be non-NULL in
> + * this case. This may be needed for separated debug images, since the section
> + * headers and symbols may need to come from separate images in that case.
> + *
> + * Note: irrespective of whether this function returns successfully,
> + * *saved_shdrs may get initialised if saved_shdrs is non-NULL. It is the
> + * caller's responsibility to free() it when non longer needed.
I kept deferring looking at this patch because of the added complexity,
couldn't we try to simplify it somehow?
> + * If want_symtab == 1, this function will only load symbols from .symtab
> + * sections. Otherwise (want_symtab == 0), .dynsym or .symtab symbols are
> + * loaded. This feature is used by dso__load() to search for the best image
> + * to load.
> + */
> +
> static int dso__load_sym(struct dso *dso, struct map *map, const char *name,
> int fd, symbol_filter_t filter, int kmodule,
> + GElf_Shdr **saved_shdrs, unsigned *saved_shdrs_count,
> int want_symtab)
> {
> struct kmap *kmap = dso->kernel ? map__kmap(map) : NULL;
> @@ -1178,6 +1275,17 @@ static int dso__load_sym(struct dso *dso, struct map *map, const char *name,
> int nr = 0;
> size_t opdidx = 0;
>
> + if (saved_shdrs != NULL && saved_shdrs_count == NULL) {
> + /*
> + * If you trigger this check, you're calling this function
> + * incorrectly. Refer to the notes above for details.
> + */
> + pr_debug("%s: warning: saved_shdrs_count == NULL: "
> + "ignoring the saved section headers.\n",
> + __func__);
> + saved_shdrs = NULL;
> + }
Why check if one of the parameters is null? Perhaps we should have
something like:
struct saved_shdrs {
GElf_Shdr **saved_shdrs;
unsigned *saved_shdrs_count;
}
And pass just one parameter?
> elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
> if (elf == NULL) {
> pr_debug("%s: cannot read %s ELF file.\n", __func__, name);
> @@ -1200,6 +1308,36 @@ static int dso__load_sym(struct dso *dso, struct map *map, const char *name,
> goto out_elf_end;
> }
>
> + /*
> + * Copy all section headers from the image if requested and if not
> + * already loaded.
> + */
> + if (saved_shdrs != NULL && *saved_shdrs == NULL) {
> + GElf_Shdr *shdrs;
> + unsigned count;
> +
> + shdrs = elf_get_all_shdrs(elf, &ehdr, &count);
> + if (shdrs == NULL)
> + goto out_elf_end;
> +
> + /*
> + * Only keep the headers if they reflect the actual run-time
> + * image's file layout:
> + */
> + if (elf_check_shdrs_valid(shdrs, count)) {
> + *saved_shdrs = shdrs;
> + *saved_shdrs_count = count;
> + } else
> + free(shdrs);
> + }
> +
> + /*
> + * If no genuine ELF headers are available yet, give up: we can't
> + * adjust symbols correctly in that case:
> + */
> + if (saved_shdrs != NULL && *saved_shdrs == NULL)
> + goto out_elf_end;
> sec = elf_section_by_name(elf, &ehdr, &shdr, ".symtab", NULL);
> if (sec == NULL) {
> if (want_symtab)
> @@ -1344,12 +1482,25 @@ static int dso__load_sym(struct dso *dso, struct map *map, const char *name,
> goto new_symbol;
> }
>
> + /*
> + * Currently, symbols for shared objects and PIE executables
> + * (i.e., ET_DYN) do not seem to get adjusted. This might need
> + * to change if file offset == virtual address is not actually
> + * guaranteed for these images. ELF doesn't provide this
> + * guarantee natively.
> + */
> if (curr_dso->adjust_symbols) {
> pr_debug4("%s: adjusting symbol: st_value: %#" PRIx64 " "
> "sh_addr: %#" PRIx64 " sh_offset: %#" PRIx64 "\n", __func__,
> (u64)sym.st_value, (u64)shdr.sh_addr,
> (u64)shdr.sh_offset);
> - sym.st_value -= shdr.sh_addr - shdr.sh_offset;
> + if (saved_shdrs && *saved_shdrs &&
> + sym.st_shndx < *saved_shdrs_count)
> + sym.st_value -=
> + (*saved_shdrs)[sym.st_shndx].sh_addr -
> + (*saved_shdrs)[sym.st_shndx].sh_offset;
> + else
> + sym.st_value -= shdr.sh_addr - shdr.sh_offset;
All this couldn't be on a separate helper routine?
> }
> /*
> * We need to figure out if the object was created from C++ sources
> @@ -1590,6 +1741,8 @@ int dso__load(struct dso *dso, struct map *map, symbol_filter_t filter)
> struct machine *machine;
> const char *root_dir;
> int want_symtab;
> + GElf_Shdr *saved_shdrs = NULL;
> + unsigned saved_shdrs_count;o
Couldn't saved_shdrs be a member of struct dso so that we don't have to
be passing parameters back and forth.
These questions are more trying to simplify the patch, I haven't deeply
looked at how to reimplement it to reduce complexity, hope you can try
to do it.
The symbols code already is complicated, I know there is intrinsic
complexity here, but we really need to try to avoid making it even more
complicaed.
- Arnaldo
>
> dso__set_loaded(dso, map->type);
>
> @@ -1692,6 +1845,7 @@ restart:
> continue;
>
> ret = dso__load_sym(dso, map, name, fd, filter, 0,
> + &saved_shdrs, &saved_shdrs_count,
> want_symtab);
> close(fd);
>
> @@ -1713,14 +1867,19 @@ restart:
>
> /*
> * If we wanted a full symtab but no image had one,
> - * relax our requirements and repeat the search.
> + * relax our requirements and repeat the search,
> + * provided we saw some valid section headers:
> */
> - if (ret <= 0 && want_symtab) {
> + if (ret <= 0 && want_symtab && saved_shdrs != NULL) {
> want_symtab = 0;
> goto restart;
> }
>
> free(name);
> +
> + if (saved_shdrs)
> + free(saved_shdrs);
> +
> if (ret < 0 && strstr(dso->name, " (deleted)") != NULL)
> return 0;
> return ret;
> @@ -1989,7 +2148,8 @@ int dso__load_vmlinux(struct dso *dso, struct map *map,
>
> dso__set_long_name(dso, (char *)vmlinux);
> dso__set_loaded(dso, map->type);
> - err = dso__load_sym(dso, map, symfs_vmlinux, fd, filter, 0, 0);
> + err = dso__load_sym(dso, map, symfs_vmlinux, fd, filter, 0,
> + NULL, NULL, 0);
> close(fd);
>
> if (err > 0)
>
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>
next prev parent reply other threads:[~2012-06-11 13:54 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2012-05-02 7:02 [PATCH-V2] perf symbols: fix symbol offset breakage with separated debug info Prashanth Nageshappa
2012-06-11 11:17 ` Fwd: " Prashanth Nageshappa
2012-06-11 13:53 ` Arnaldo Carvalho de Melo [this message]
2012-06-13 10:14 ` Dave Martin
2012-06-25 23:49 ` Roland McGrath
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20120611135352.GA2202@infradead.org \
--to=acme@infradead.org \
--cc=akpm@linux-foundation.org \
--cc=ananth@in.ibm.com \
--cc=andi@firstfloor.org \
--cc=anton@redhat.com \
--cc=dave.martin@linaro.org \
--cc=hch@infradead.org \
--cc=jkenisto@linux.vnet.ibm.com \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=linux-perf-users@vger.kernel.org \
--cc=masami.hiramatsu.pt@hitachi.com \
--cc=mingo@elte.hu \
--cc=oleg@redhat.com \
--cc=peterz@infradead.org \
--cc=prashanth@linux.vnet.ibm.com \
--cc=roland@hack.frob.com \
--cc=rostedt@goodmis.org \
--cc=srikar@linux.vnet.ibm.com \
--cc=tglx@linutronix.de \
--cc=torvalds@linux-foundation.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).