uefi-ref/binutils-gdb
1
0
Fork 0
mirror of https://sourceware.org/git/binutils-gdb.git synced 2026-08-27 00:26:02 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions
binutils-gdb/libctf/ctf-link.c
Nick Alcock d2f3633152 libctf: v4: symtypetab unifications 
The v4 symtypetab design is quite different from v3. Where v3 has a pair
of possibly indexed symtypetabs per dict, v4 uses a pair of symtypetabs
in two ELF sections, one per ELF object: one has straight ctf_id_t's in
it and relates to the parent dict, the other a pair of (index, ctF_id_t)
relating to other dicts.  The symtypetab is always indexed: the old 1:1
"like the symtab if the symtypetab only contained STT_FUNC and
STT_OBJECT symbols, with some symbols dropped" scheme is gone entirely,
and with it all the arbitrary decisions that went with it (when to use
indexed versus 1:1 representation, which symbols to skip, etc) and its
fragility (if a single symbol is assigned to the wrong table, disaster
results: see commit 211bcd0133 for an example of this biting in real
life).

As a first phase, unify the object and function info sections and drop
the indexed/non-indexed representation. Symtypetab support in the
resulting library doesn't work any more, but it didn't work before now
in v4 anyway, so this is no loss.  Future commits will fix this.

Also drop all the caching machinery for symbol lookups: it's only there
because we might need to consult many symtypetabs to look up types of
symbols in child dicts, which will no longer be true after these
changes.

(API changes introduced by this series will be indicated in the last
commit in the series.)
2026-06-17 16:42:41 +01:00

2154 lines
62 KiB
C
Raw Blame History

/* CTF linking.
Copyright (C) 2019-2025 Free Software Foundation, Inc.
This file is part of libctf.
libctf is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
This program is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; see the file COPYING. If not see
<http://www.gnu.org/licenses/>. */
#include <ctf-impl.h>
#include <string.h>
#include "ctf-util-endian.h"
/* CTF linking consists of adding CTF archives full of content to be merged into
this one to the current file (which must be writable) by calling
ctf_link_add. Once this is done, a call to ctf_link will merge the type
tables together, generating new CTF files as needed, with this one as a
parent, to contain types from the inputs which conflict. ctf_link_add_strtab
takes a callback which provides string/offset pairs to be added to the
external symbol table and deduplicated from all CTF string tables in the
output link; ctf_link_shuffle_syms takes a callback which provides symtab
entries in ascending order, and shuffles the function and data sections to
match; and ctf_link_write emits a CTF file (if there are no conflicts
requiring per-compilation-unit sub-CTF files) or CTF archives (otherwise) and
returns it, suitable for addition in the .ctf section of the output. */
/* The linker inputs look like this. clin_fp is used for short-circuited
CU-mapped links that can entirely avoid the first link phase in some
situations in favour of just passing on the contained ctf_dict_t: it is
always the sole ctf_dict_t inside the corresponding clin_arc. If set, it
gets assigned directly to the final link inputs and freed from there, so it
never gets explicitly freed in the ctf_link_input. */
typedef struct ctf_link_input
{
char *clin_filename;
ctf_archive_t *clin_arc;
ctf_dict_t *clin_fp;
char *cuname_prefix;
int n;
} ctf_link_input_t;
/* Return the name of the compilation unit this CTF dict or its parent applies
to, or a non-null string otherwise: prefer the parent. Used in debugging
output. Sometimes used for outputs too. */
const char *
ctf_link_input_name (ctf_dict_t *fp)
{
if (fp->ctf_parent && fp->ctf_parent->ctf_cu_name)
return fp->ctf_parent->ctf_cu_name;
else if (fp->ctf_cu_name)
return fp->ctf_cu_name;
else
{
void *input;
ctf_next_t *it = NULL;
/* Try looking up the filename of some random input: it's better than
nothing. */
if (ctf_dynhash_next (fp->ctf_link_inputs, &it, NULL, &input) == 0)
{
ctf_next_destroy (it);
if (((ctf_link_input_t *) input)->clin_filename != NULL)
return ((ctf_link_input_t *) input)->clin_filename;
}
}
return "(unnamed)";
}
/* Return the cuname of a dict, or the string "unnamed-CU" if none. */
static const char *
ctf_unnamed_cuname (ctf_dict_t *fp)
{
const char *cuname = ctf_dict_cuname (fp);
if (!cuname)
cuname = "unnamed-CU";
return cuname;
}
static void
ctf_link_input_close (void *input)
{
ctf_link_input_t *i = (ctf_link_input_t *) input;
if (i->clin_arc)
ctf_arc_close (i->clin_arc);
free (i->clin_filename);
free (i->cuname_prefix);
free (i);
}
/* Like ctf_link_add, below, but with no error-checking, so it can be called
in the middle of an ongoing link. */
static ctf_ret_t
ctf_link_add_internal (ctf_dict_t *fp, ctf_archive_t *ctf,
ctf_dict_t *fp_input, const char *name,
const char *cuname_prefix)
{
int existing = 0;
ctf_link_input_t *input;
char *filename, *keyname, *prefix = NULL;
/* Existing: return it, or (if a different dict with the same name
is already there) make up a new unique name. Always use the actual name
for the filename, because that needs to be ctf_open()ed. */
if ((input = ctf_dynhash_lookup (fp->ctf_link_inputs, name)) != NULL)
{
if ((fp_input != NULL && (input->clin_fp == fp_input))
|| (ctf != NULL && (input->clin_arc == ctf)))
return 0;
existing = 1;
}
if ((filename = strdup (name)) == NULL)
goto oom;
if (cuname_prefix && ((prefix = strdup (cuname_prefix)) == NULL))
goto oom1;
if ((input = calloc (1, sizeof (ctf_link_input_t))) == NULL)
goto oom2;
input->clin_arc = ctf;
input->clin_fp = fp_input;
input->clin_filename = filename;
input->cuname_prefix = prefix;
input->n = ctf_dynhash_elements (fp->ctf_link_inputs);
if (existing)
{
if (asprintf (&keyname, "%s#%li", name, (long int)
ctf_dynhash_elements (fp->ctf_link_inputs)) < 0)
goto oom3;
}
else if ((keyname = strdup (name)) == NULL)
goto oom3;
if (ctf_dynhash_insert (fp->ctf_link_inputs, keyname, input) < 0)
goto oom4;
return 0;
oom4:
free (keyname);
oom3:
free (input);
oom2:
free (prefix);
oom1:
free (filename);
oom:
return ctf_set_errno (fp, ENOMEM);
}
/* Add an opened CTF archive or unopened file (by name) to a link.
If CTF is NULL and FILENAME is non-null, an unopened file is meant:
otherwise, the specified archive is assumed to have the given FILENAME (as
its cuname).
If CTF is NULL, the FILENAME is only opened when needed, and is closed when
no longer needed, so that large cu-mapped links will only use memory for
their cu-mapped inputs briefly (compensating for the memory usage of the
smushed-together cu-mapped verion).
The CUNAME_PREFIX, if set, prefixes all input cunames with a given string
(used for archives).
Passed in CTF args are owned by the dictionary and will be freed by it.
The order of calls to this function influences the order of types in the
final link output, but otherwise is not important.
Repeated additions of the same FILENAME have no effect; repeated additions of
different dicts with the same FILENAME add all the dicts with unique
FILENAMEs derived from FILENAME. */
ctf_ret_t
ctf_link_add (ctf_dict_t *fp, ctf_archive_t *ctf, const char *filename,
const char *cuname_prefix)
{
if (!filename)
return (ctf_set_errno (fp, EINVAL));
if (fp->ctf_link_outputs)
return (ctf_set_errno (fp, ECTF_LINKADDEDLATE));
if (fp->ctf_link_inputs == NULL)
fp->ctf_link_inputs = ctf_dynhash_create (ctf_hash_string,
ctf_hash_eq_string, free,
ctf_link_input_close);
if (fp->ctf_link_inputs == NULL)
return (ctf_set_errno (fp, ENOMEM));
/* We can only lazily open files if libctf.so is in use rather than
libctf-nobfd.so. */
#if NOBFD
if (!ctf)
return (ctf_set_errno (fp, ECTF_NEEDSBFD));
#endif
return ctf_link_add_internal (fp, ctf, NULL, filename, cuname_prefix);
}
/* Lazily open a CTF archive for linking, if not already open.
Returns the number of files contained within the opened archive (0 for none),
or -1 on error, as usual. */
static ssize_t
ctf_link_lazy_open (ctf_dict_t *fp, ctf_link_input_t *input)
{
size_t count;
ctf_error_t err;
if (input->clin_arc)
return ctf_archive_count (input->clin_arc);
if (input->clin_fp)
return 1;
/* See ctf_link_add. */
#if NOBFD
ctf_err (err_locus (fp), ECTF_NEEDSBFD, _("cannot open %s lazily"), input->clin_filename);
return -1;
#else
input->clin_arc = ctf_open (input->clin_filename, NULL, &err);
#endif
/* Having no CTF sections is not an error. We just don't need to do
anything. */
if (!input->clin_arc)
{
if (err == ECTF_NOCTFDATA)
return 0;
ctf_err (err_locus (fp), err, _("opening CTF %s failed"), input->clin_filename);
return -1;
}
if ((count = ctf_archive_count (input->clin_arc)) == 0)
ctf_arc_close (input->clin_arc);
return (ssize_t) count;
}
/* Find a non-clashing unique name for a per-CU output dict, to prevent distinct
members corresponding to inputs with identical cunames from overwriting each
other. The name should be something like NAME. */
static char *
ctf_new_per_cu_name (ctf_dict_t *fp, const char *name)
{
char *dynname;
long int i = 0;
if ((dynname = strdup (name)) == NULL)
return NULL;
while ((ctf_dynhash_lookup (fp->ctf_link_outputs, dynname)) != NULL)
{
free (dynname);
if (asprintf (&dynname, "%s#%li", name, i++) < 0)
return NULL;
}
return dynname;
}
/* Return a per-CU output CTF dictionary suitable for the given INPUT or CU,
creating and interning it if need be. */
static ctf_dict_t *
ctf_create_per_cu (ctf_dict_t *fp, ctf_dict_t *input, const char *cu_name)
{
ctf_dict_t *cu_fp;
const char *ctf_name = NULL;
char *dynname = NULL;
/* Already has a per-CU mapping? Just return it. */
if (input && input->ctf_link_in_out)
return input->ctf_link_in_out;
/* Check the mapping table and translate the per-CU name we use
accordingly. */
if (cu_name == NULL)
cu_name = ctf_unnamed_cuname (input);
if (fp->ctf_link_in_cu_mapping)
{
if ((ctf_name = ctf_dynhash_lookup (fp->ctf_link_in_cu_mapping,
cu_name)) == NULL)
ctf_name = cu_name;
}
if (ctf_name == NULL)
ctf_name = cu_name;
/* Look up the per-CU dict. If we don't know of one, or it is for a different input
CU which just happens to have the same name, create a new one. If we are creating
a dict with no input specified, anything will do. */
if ((cu_fp = ctf_dynhash_lookup (fp->ctf_link_outputs, ctf_name)) == NULL
|| (input && cu_fp->ctf_link_in_out != fp))
{
ctf_error_t err;
/* The parent is unreffed to avoid a circular reference, since fp holds a
reference to cu_fp itself, and ctf_create doesn't provide a way to pass
internal flags to the importing machinery. */
if ((cu_fp = ctf_create_internal (fp, CTF_IMPORT_UNREF, &err)) == NULL)
{
ctf_err (err_locus (fp), err, _("cannot create per-CU CTF archive for "
"input CU %s"), cu_name);
return NULL;
}
/* The deduplicator is ready for strict enumerator value checking, unless
doing dedup-against-first links, in which case no child dicts are
created to put conflicting enums in and duplicate enumerators must be
tolerated. */
if (!(fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST))
cu_fp->ctf_flags |= LCTF_STRICT_NO_DUP_ENUMERATORS;
if ((dynname = ctf_new_per_cu_name (fp, ctf_name)) == NULL)
goto oom;
ctf_dict_set_cuname (cu_fp, cu_name);
cu_fp->ctf_parent_name = _CTF_SECTION;
cu_fp->ctf_link_in_out = fp;
fp->ctf_link_in_out = cu_fp;
if (ctf_dynhash_insert (fp->ctf_link_outputs, dynname, cu_fp) < 0)
goto oom;
}
return cu_fp;
oom:
free (dynname);
ctf_dict_close (cu_fp);
ctf_set_errno (fp, ENOMEM);
return NULL;
}
/* Add a mapping directing that the CU named FROM should have its
conflicting/non-duplicate types (depending on link mode) go into a dict
named TO. Many FROMs can share a TO, but adding the same FROM with
a different TO will replace the old mapping.
We forcibly add a dict named TO in every case, even though it may well
wind up empty, because clients that use this facility usually expect to find
every TO dict present, even if empty, and malfunction otherwise.
The TO mapping named "" is special: types in this are merged directly and
unconditionally into the shared parent dict, and are hidden (marked conflicting) if
they clash, rather than being moved into child dicts. */
ctf_ret_t
ctf_link_add_cu_mapping (ctf_dict_t *fp, const char *from, const char *to)
{
ctf_error_t err;
char *f = NULL, *t = NULL, *existing;
ctf_dynhash_t *one_out;
/* Mappings cannot be set up if per-CU output dicts already exist. */
if (fp->ctf_link_outputs && ctf_dynhash_elements (fp->ctf_link_outputs) != 0)
return (ctf_set_errno (fp, ECTF_LINKADDEDLATE));
if (fp->ctf_link_in_cu_mapping == NULL)
fp->ctf_link_in_cu_mapping = ctf_dynhash_create (ctf_hash_string,
ctf_hash_eq_string, free,
free);
if (fp->ctf_link_in_cu_mapping == NULL)
goto oom;
if (fp->ctf_link_out_cu_mapping == NULL)
fp->ctf_link_out_cu_mapping = ctf_dynhash_create (ctf_hash_string,
ctf_hash_eq_string, free,
(ctf_hash_free_fun)
ctf_dynhash_destroy);
if (fp->ctf_link_out_cu_mapping == NULL)
goto oom;
/* If this FROM already exists, remove the mapping from both the FROM->TO
and the TO->FROM lists: the user wants to change it. */
if ((existing = ctf_dynhash_lookup (fp->ctf_link_in_cu_mapping, from)) != NULL)
{
one_out = ctf_dynhash_lookup (fp->ctf_link_out_cu_mapping, existing);
if (!ctf_assert (fp, one_out))
return -1; /* errno is set for us. */
ctf_dynhash_remove (one_out, from);
ctf_dynhash_remove (fp->ctf_link_in_cu_mapping, from);
}
f = strdup (from);
t = strdup (to);
if (!f || !t)
goto oom;
/* Track both in a list from FROM to TO and in a list from TO to a list of
FROM. The former is used to create TUs with the mapped-to name at need:
the latter is used in deduplicating links to pull in all input CUs
corresponding to a single output CU. */
if ((err = ctf_dynhash_insert (fp->ctf_link_in_cu_mapping, f, t)) < 0)
{
ctf_set_errno (fp, err);
goto oom_noerrno;
}
/* f and t are now owned by the in_cu_mapping: reallocate them. */
f = strdup (from);
t = strdup (to);
if (!f || !t)
goto oom;
if ((one_out = ctf_dynhash_lookup (fp->ctf_link_out_cu_mapping, t)) == NULL)
{
if ((one_out = ctf_dynhash_create (ctf_hash_string, ctf_hash_eq_string,
free, NULL)) == NULL)
goto oom;
if ((err = ctf_dynhash_insert (fp->ctf_link_out_cu_mapping,
t, one_out)) < 0)
{
ctf_dynhash_destroy (one_out);
ctf_set_errno (fp, err);
goto oom_noerrno;
}
}
else
{
free (t);
t = NULL;
}
if (ctf_dynhash_insert (one_out, f, NULL) < 0)
{
ctf_set_errno (fp, err);
goto oom_noerrno;
}
return 0;
oom:
ctf_set_errno (fp, errno);
oom_noerrno:
free (f);
free (t);
return -1;
}
/* Set a function which is called to transform the names of archive members.
This is useful for applying regular transformations to many names, where
ctf_link_add_cu_mapping applies arbitrarily irregular changes to single
names. The member name changer is applied at ctf_link_write time, so it
cannot conflate multiple CUs into one the way ctf_link_add_cu_mapping can.
The changer function accepts a name and should return a new
dynamically-allocated name, or NULL if the name should be left unchanged. */
void
ctf_link_set_memb_name_changer (ctf_dict_t *fp,
ctf_link_memb_name_changer_f *changer,
void *arg)
{
fp->ctf_link_memb_name_changer = changer;
fp->ctf_link_memb_name_changer_arg = arg;
}
/* Set a function which is used to filter out unwanted variables from the link.
UPTODO: call this somewhere (ctf-dedup, probably). */
ctf_bool_t
ctf_link_set_variable_filter (ctf_dict_t *fp, ctf_link_variable_filter_f *filter,
void *arg)
{
fp->ctf_link_variable_filter = filter;
fp->ctf_link_variable_filter_arg = arg;
return 0;
}
typedef struct link_sort_inputs_cb_arg
{
int is_cu_mapped;
ctf_dict_t *fp;
} link_sort_inputs_cb_arg_t;
/* Sort the inputs by N (the link order). For CU-mapped links, this is a
mapping of input to output name, not a mapping of input name to input
ctf_link_input_t: compensate accordingly. */
static int
ctf_link_sort_inputs (const ctf_next_hkv_t *one, const ctf_next_hkv_t *two,
void *arg)
{
ctf_link_input_t *input_1;
ctf_link_input_t *input_2;
link_sort_inputs_cb_arg_t *cu_mapped = (link_sort_inputs_cb_arg_t *) arg;
if (!cu_mapped || !cu_mapped->is_cu_mapped)
{
input_1 = (ctf_link_input_t *) one->hkv_value;
input_2 = (ctf_link_input_t *) two->hkv_value;
}
else
{
const char *name_1 = (const char *) one->hkv_key;
const char *name_2 = (const char *) two->hkv_key;
input_1 = ctf_dynhash_lookup (cu_mapped->fp->ctf_link_inputs, name_1);
input_2 = ctf_dynhash_lookup (cu_mapped->fp->ctf_link_inputs, name_2);
/* There is no guarantee that CU-mappings actually have corresponding
inputs: the relative ordering in that case is unimportant. */
if (!input_1)
return -1;
if (!input_2)
return 1;
}
if (input_1->n < input_2->n)
return -1;
else if (input_1->n > input_2->n)
return 1;
else
return 0;
}
/* Count the number of input dicts in the ctf_link_inputs, or that subset of the
ctf_link_inputs given by CU_NAMES if set: open them if need be. Return the number
of input dicts, and optionally the name and ctf_link_input_t of the single input
archive if only one exists (no matter how many dicts it contains). */
static ssize_t
ctf_link_deduplicating_count_inputs (ctf_dict_t *fp, ctf_dynhash_t *cu_names,
ctf_link_input_t **only_one_input)
{
ctf_dynhash_t *inputs = fp->ctf_link_inputs;
ctf_next_t *i = NULL;
void *name, *input;
ctf_link_input_t *one_input = NULL;
const char *one_name = NULL;
ssize_t count = 0, narcs = 0;
ctf_error_t err;
if (cu_names)
inputs = cu_names;
while ((err = ctf_dynhash_next (inputs, &i, &name, &input)) == 0)
{
ssize_t one_count;
one_name = (const char *) name;
/* If we are processing CU names, get the real input. */
if (cu_names)
one_input = ctf_dynhash_lookup (fp->ctf_link_inputs, one_name);
else
one_input = (ctf_link_input_t *) input;
if (!one_input)
continue;
one_count = ctf_link_lazy_open (fp, one_input);
if (one_count < 0)
{
ctf_next_destroy (i);
return -1; /* errno is set for us. */
}
/* If this is an unnamed, single-dict archive, propagate the input name
into the dict itself, for the sake of deduplicator child dict
naming. */
if (one_input->clin_arc
&& ctf_archive_count (one_input->clin_arc) == 1
&& one_input->clin_arc->ctfi_dict != NULL &&
ctf_dict_cuname (one_input->clin_arc->ctfi_dict) == NULL)
ctf_dict_set_cuname (one_input->clin_arc->ctfi_dict, name);
count += one_count;
narcs++;
}
if (err != ECTF_NEXT_END)
{
ctf_err (err_locus (fp), err, NULL);
return -1;
}
if (!count)
return 0;
if (narcs == 1)
{
if (only_one_input)
*only_one_input = one_input;
}
else if (only_one_input)
*only_one_input = NULL;
return count;
}
static ctf_ret_t
ctf_link_deduplicating_close_inputs (ctf_dict_t *fp, ctf_dynhash_t *cu_names,
ctf_dict_t **inputs, ssize_t ninputs);
/* Allocate and populate an inputs array big enough for a given set of inputs:
either a specific set of CU names (those from that set found in the
ctf_link_inputs), or the entire ctf_link_inputs (if cu_names is not set).
The cunames are optionally prefixed with the cuname_prefix at the same time.
The number of inputs (from ctf_link_deduplicating_count_inputs, above) is
passed in NINPUTS: an array of uint32_t containing parent pointers
(corresponding to those members of the inputs that have parents) is allocated
and returned in PARENTS.
The inputs are *archives*, not files: the archive can have multiple members
if it is the result of a previous incremental link. We want to add every one
in turn, including the shared parent. (The dedup machinery knows that a type
used by a single dictionary and its parent should not be shared in
CTF_LINK_SHARE_DUPLICATED mode.)
If no inputs exist that correspond to these CUs, return NULL with the errno
set to ECTF_NOCTFDATA. */
static ctf_dict_t **
ctf_link_deduplicating_open_inputs (ctf_dict_t *fp, ctf_dynhash_t *cu_names,
ssize_t ninputs, uint32_t **parents)
{
ctf_dynhash_t *inputs = fp->ctf_link_inputs;
ctf_next_t *i = NULL;
void *name, *input;
link_sort_inputs_cb_arg_t sort_arg;
ctf_dict_t **dedup_inputs = NULL;
ctf_dict_t **walk;
uint32_t *parents_ = NULL;
ctf_error_t err;
if (cu_names)
inputs = cu_names;
if ((dedup_inputs = calloc (ninputs, sizeof (ctf_dict_t *))) == NULL)
goto oom;
if ((parents_ = calloc (ninputs, sizeof (uint32_t))) == NULL)
goto oom;
walk = dedup_inputs;
/* Counting done: push every input into the array, in the order they were
passed to ctf_link_add (and ultimately ld). */
sort_arg.is_cu_mapped = (cu_names != NULL);
sort_arg.fp = fp;
while ((err = ctf_dynhash_next_sorted (inputs, &i, &name, &input,
ctf_link_sort_inputs, &sort_arg)) == 0)
{
const char *one_name = (const char *) name;
const char *member_name;
ctf_link_input_t *one_input;
ctf_dict_t *one_fp;
ctf_next_t *j = NULL;
ssize_t parent_i;
int parent_set = 0;
ctf_dict_t *parent_fp = NULL;
/* If we are processing CU names, get the real input. All the inputs
will have been opened, if they contained any CTF at all. */
if (cu_names)
one_input = ctf_dynhash_lookup (fp->ctf_link_inputs, one_name);
else
one_input = (ctf_link_input_t *) input;
if (!one_input || (!one_input->clin_arc && !one_input->clin_fp))
continue;
/* Short-circuit: if clin_fp is set, just use it. */
if (one_input->clin_fp)
{
parents_[walk - dedup_inputs] = walk - dedup_inputs;
*walk = one_input->clin_fp;
walk++;
continue;
}
if (one_input->clin_filename)
ctf_dprintf ("Opening %s\n", one_input->clin_filename);
/* We augument the input cuname with the prefix, if provided, but
otherwise disregard it: either it is the parent (which we can get via
ctf_dict_parent), or we want to put everything into one CU sharing the
cuname anyway (if this is a CU-mapped link), or this is the final phase
of a relink with CU-mapping off (i.e. ld -r) in which case the cuname
is correctly set regardless. */
while ((one_fp = ctf_archive_next (one_input->clin_arc, &j, &member_name,
0, &err)) != NULL)
{
/* Augment the cuname with a prefix, if desired. Set it to the
filename if not set at all. */
if (ctf_dict_cuname (one_fp) == NULL &&
one_input->clin_filename != NULL &&
ctf_dict_set_cuname (one_fp, one_input->clin_filename) < 0)
goto oom;
if (one_input->cuname_prefix)
{
char *new_cuname;
new_cuname = strdup (one_input->cuname_prefix);
new_cuname = ctf_str_append (new_cuname, ctf_dict_cuname (one_fp));
if (!new_cuname)
goto oom;
if (ctf_dict_set_cuname (one_fp, new_cuname) < 0)
goto oom;
free (new_cuname);
}
/* ctf_archive_next either auto-imports the parent, or this *is* the parent.
In both cases, we can set the parent up if it's not already set.
If parent importing failed, we must report an error.
We must do it first, so we can set later elements of the parents_ array to
the parent. */
if (one_fp->ctf_flags & LCTF_CHILD && one_fp->ctf_parent == NULL)
{
ctf_next_destroy (i);
ctf_next_destroy (j);
ctf_err (err_locus (fp), ECTF_NOPARENT,
_("archive member %s in CU %s is child with no parent"),
member_name, one_name);
goto reported_err;
}
if (!parent_set)
{
if (one_fp->ctf_flags & LCTF_CHILD)
parent_fp = one_fp->ctf_parent;
else
parent_fp = one_fp;
*walk = parent_fp;
parent_i = walk - dedup_inputs;
parents_[parent_i] = parent_i;
walk++;
parent_set = 1;
/* If this is the parent, we don't need to process the child dict itself
(because there isn't one). */
if (!(one_fp->ctf_flags & LCTF_CHILD))
continue;
}
if (!ctf_assert (fp, one_fp->ctf_parent == parent_fp))
{
ctf_next_destroy (j);
ctf_next_destroy (i);
goto err; /* errno is set for us. */
}
parents_[walk - dedup_inputs] = parent_i;
*walk = one_fp;
walk++;
}
if (err != ECTF_NEXT_END)
{
ctf_next_destroy (i);
goto iterr;
}
}
if (err != ECTF_NEXT_END)
goto iterr;
*parents = parents_;
return dedup_inputs;
oom:
err = ENOMEM;
iterr:
ctf_set_errno (fp, err);
err:
ctf_err (err_locus (fp), 0, NULL); /* The function name is enough. */
reported_err:
ctf_link_deduplicating_close_inputs (fp, cu_names, dedup_inputs, ninputs);
/* UPTODO XXX fix pre-existing leak of the parents on error. */
free (dedup_inputs);
free (parents_);
return NULL;
}
/* Close INPUTS that have already been linked, first the passed array, and then
that subset of the ctf_link_inputs archives they came from cited by the
CU_NAMES. If CU_NAMES is not specified, close all the ctf_link_inputs in one
go, leaving it empty. */
static ctf_ret_t
ctf_link_deduplicating_close_inputs (ctf_dict_t *fp, ctf_dynhash_t *cu_names,
ctf_dict_t **inputs, ssize_t ninputs)
{
ctf_next_t *it = NULL;
void *name;
ctf_error_t err;
ssize_t i;
/* This is the inverse of ctf_link_deduplicating_open_inputs: so first, close
all the individual input dicts, opened by the archive iterator. */
for (i = 0; i < ninputs; i++)
ctf_dict_close (inputs[i]);
/* Now close the archives they are part of. */
if (cu_names)
{
while ((err = ctf_dynhash_next (cu_names, &it, &name, NULL)) == 0)
{
/* Remove the input from the linker inputs, if it exists, which also
closes it. */
ctf_dynhash_remove (fp->ctf_link_inputs, (const char *) name);
}
if (err != ECTF_NEXT_END)
return ctf_err (err_locus (fp), err, NULL);
}
else
ctf_dynhash_empty (fp->ctf_link_inputs);
return 0;
}
/* Check for symbol conflicts during linking. Three possibilities: already exists,
conflicting, or nonexistent. We use a tristate return value: -1: conflicting; 1:
nonexistent: 0: already exists. */
static int
check_sym (ctf_dict_t *fp, const char *name, ctf_id_t type)
{
void *value;
/* Not present at all yet. */
if (!ctf_dynhash_lookup_kv (fp->ctf_symtypehash, name, NULL, &value))
return 1;
/* Already present. */
if ((ctf_id_t) (uintptr_t) value == type)
return 0;
/* Wrong type. */
return -1;
}
/* Do a deduplicating link of the symtypetab to one input dict. */
static ctf_ret_t
ctf_link_deduplicating_one_symtypetab (ctf_dict_t *fp, ctf_dict_t *input,
int cu_mapped)
{
ctf_next_t *it = NULL;
const char *name;
ctf_id_t type;
while ((type = ctf_symbol_next (input, &it, &name)) != CTF_ERR)
{
ctf_id_t dst_type;
ctf_dict_t *per_cu_out_fp;
int sym;
/* Look in the parent first. */
if ((dst_type = ctf_dedup_type_mapping (fp, input, type)) == CTF_ERR)
goto err; /* errno is set for us. */
if (dst_type != 0)
{
if (!ctf_assert (fp, ctf_type_isparent (fp, dst_type)))
goto err; /* errno is set for us. */
sym = check_sym (fp, name, dst_type);
/* Already present: next symbol. */
if (sym == 0)
continue;
/* Not present: add it. */
else if (sym > 0)
{
if (ctf_add_sym (fp, name, dst_type) < 0)
goto err; /* errno is set for us. */
continue;
}
}
/* Can't add to the parent due to a name clash (most unlikely), or because
it references a type only present in the child. Try adding to the
child, creating if need be. If we can't do that, skip it. Don't add
to a child if we're doing a CU-mapped link, since that has only one
output. */
if (cu_mapped)
{
ctf_dprintf ("Symbol %s in input file %s depends on a type %lx "
"hidden due to conflicts: skipped.\n", name,
ctf_unnamed_cuname (input), type);
continue;
}
if ((per_cu_out_fp = ctf_create_per_cu (fp, input, NULL)) == NULL)
goto err; /* errno is set for us. */
/* If the type was not found, check for it in the child too. */
if (dst_type == 0)
{
if ((dst_type = ctf_dedup_type_mapping (per_cu_out_fp,
input, type)) == CTF_ERR)
goto err; /* errno is set for us. */
if (dst_type == 0)
{
ctf_warn (link_type_err_locus (fp, input, -1, type), 1, 0,
_("type for symbol %s not found: skipped"), name);
continue;
}
}
sym = check_sym (per_cu_out_fp, name, dst_type);
/* Already present: next symbol. */
if (sym == 0)
continue;
/* Not present: add it. */
else if (sym > 0)
{
if (ctf_add_sym (per_cu_out_fp, name, dst_type) < 0)
goto err; /* errno is set for us. */
}
else
{
/* Perhaps this should be an assertion failure. */
ctf_next_destroy (it);
return ctf_err (link_type_err_locus (fp, input, -1, type), ECTF_DUPLICATE,
_("symbol %s found conflicting even when trying in per-CU dict."),
name);
}
}
if (ctf_errno (input) != ECTF_NEXT_END)
{
return ctf_err (link_err_locus (fp, input, -1), 0,
_("iterating over symbols"));
}
return 0;
err:
ctf_next_destroy (it);
return -1; /* errno is set for us. */
}
/* Do a deduplicating link of the function info and data objects
in the inputs. */
static ctf_ret_t
ctf_link_deduplicating_syms (ctf_dict_t *fp, ctf_dict_t **inputs,
size_t ninputs, int cu_mapped)
{
size_t i;
for (i = 0; i < ninputs; i++)
{
if (ctf_link_deduplicating_one_symtypetab (fp, inputs[i],
cu_mapped) < 0)
return -1; /* errno is set for us. */
}
return 0;
}
/* Do the per-CU part of a deduplicating link. */
static ctf_ret_t
ctf_link_deduplicating_per_cu (ctf_dict_t *fp)
{
ctf_next_t *i = NULL;
ctf_error_t err;
void *out_cu;
void *in_cus;
/* Links with a per-CU mapping in force get a first pass of deduplication,
dedupping the inputs for a given CU mapping into the output for that
mapping. The outputs from this process get fed back into the final pass
that is carried out even for non-CU links. */
while ((err = ctf_dynhash_next (fp->ctf_link_out_cu_mapping, &i, &out_cu,
&in_cus)) == 0)
{
const char *out_name = (const char *) out_cu;
ctf_dynhash_t *in = (ctf_dynhash_t *) in_cus;
ctf_dict_t *out = NULL;
ctf_dict_t **inputs;
ctf_dict_t **outputs;
ctf_archive_t *in_arc;
ssize_t ninputs;
ctf_link_input_t *only_input;
uint32_t noutputs;
uint32_t *parents;
ctf_dprintf ("ctf_link_deduplicating_per_cu: deduplicating into %s\n",
out_name);
if ((ninputs = ctf_link_deduplicating_count_inputs (fp, in,
&only_input)) == -1)
goto err_open_inputs;
/* CU mapping with no inputs? Skip. */
if (ninputs == 0)
continue;
if (labs ((long int) ninputs) > 0xfffffffe)
{
ctf_err (err_locus (fp), EDOM, _("too many inputs: %li"), (long int) ninputs);
goto err_open_inputs;
}
/* Short-circuit: a cu-mapped link with only one input archive with
unconflicting contents is a do-nothing, and we can just leave the input
in place: we do have to change the cuname, though, so we unwrap it,
change the cuname, then stuff it back in the linker input again, via
the clin_fp short-circuit member. ctf_link_deduplicating_open_inputs
will spot this member and jam it straight into the next link phase,
ignoring the corresponding archive. */
if (only_input && ninputs == 1)
{
ctf_next_t *ai = NULL;
ctf_error_t err;
/* We can abuse an archive iterator to get the only member cheaply, no
matter what its name. */
only_input->clin_fp = ctf_archive_next (only_input->clin_arc,
&ai, NULL, 0, &err);
if (!only_input->clin_fp)
{
ctf_err (err_locus (fp), err, _("cannot open archive %s"),
only_input->clin_filename);
goto err_open_inputs;
}
ctf_next_destroy (ai);
if (strcmp (only_input->clin_filename, out_name) != 0)
{
/* Renaming. We need to add a new input, then null out the
clin_arc and clin_fp of the old one to stop it being
auto-closed on removal. The new input needs its cuname changed
to out_name, which is doable only because the cuname is a
dynamic property which can be changed even in readonly
dicts. */
ctf_dict_set_cuname (only_input->clin_fp, out_name);
if (ctf_link_add_internal (fp, only_input->clin_arc,
only_input->clin_fp,
out_name, only_input->cuname_prefix) < 0)
{
ctf_err (err_locus (fp), 0, _("cannot add intermediate files to link"));
goto err_open_inputs;
}
only_input->clin_arc = NULL;
only_input->clin_fp = NULL;
ctf_dynhash_remove (fp->ctf_link_inputs,
only_input->clin_filename);
}
continue;
}
/* This is a real CU many-to-one mapping: we must dedup the inputs into
a new output to be used in the final link phase. */
if ((inputs = ctf_link_deduplicating_open_inputs (fp, in, ninputs,
&parents)) == NULL)
{
ctf_next_destroy (i);
goto err_open_inputs;
}
if ((out = ctf_create (NULL, &err)) == NULL)
{
ctf_err (err_locus (fp), err,
_("cannot create per-CU CTF archive for %s"), out_name);
goto err_inputs;
}
/* The deduplicator is ready for strict enumerator value checking, unless
doing dedup-against-first links, in which case no child dicts are
created to put conflicting enums in and duplicate enumerators must be
tolerated. */
if (!(fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST))
out->ctf_flags |= LCTF_STRICT_NO_DUP_ENUMERATORS;
/* Share the atoms table to reduce memory usage. */
out->ctf_dedup_atoms = fp->ctf_dedup_atoms_alloc;
/* Per-CU dictionaries have no parents at this stage. Parent/child
deduplication happens in the link's final pass. However, the cuname
*is* important, as it is propagated into the final dictionary. */
ctf_dict_set_cuname (out, out_name);
if (ctf_dedup (out, inputs, ninputs, 1) < 0)
{
ctf_err (link_err_locus (fp, out, -1), 0,
_("CU-mapped deduplication failed for %s"), out_name);
goto err_inputs;
}
if ((outputs = ctf_dedup_emit (out, inputs, ninputs, parents,
&noutputs, 1)) == NULL)
{
ctf_err (link_err_locus (fp, out, -1), 0,
_("CU-mapped deduplicating link type emission failed for %s"),
out_name);
goto err_inputs;
}
if (!ctf_assert (fp, noutputs == 1))
{
size_t j;
for (j = 1; j < noutputs; j++)
ctf_dict_close (outputs[j]);
goto err_inputs_outputs;
}
ctf_dedup_fini (out, outputs, noutputs);
/* For now, we omit symbol section linking for CU-mapped links, until it
is clear how to unify the symbol table across such links. (Perhaps we
should emit an unconditionally indexed symtab, like the compiler
does.) */
if (ctf_link_deduplicating_close_inputs (fp, in, inputs, ninputs) < 0)
{
free (inputs);
free (parents);
goto err_outputs;
}
free (inputs);
free (parents);
/* Splice any errors or warnings created during this link back into the
dict that the caller knows about. */
ctf_list_splice (&fp->ctf_errs_warnings, &outputs[0]->ctf_errs_warnings);
/* This output now becomes an input to the next link phase, with a name
equal to the CU name. We have to wrap it in an archive wrapper
first. The dict is closed when the archive is. */
if ((in_arc = ctf_new_archive_internal (NULL, outputs[0], 0,
FREE_ARCHIVE_ONLY_DICT, 0,
NULL, &err)) == NULL)
{
ctf_set_errno (fp, err);
goto err_outputs;
}
if (ctf_link_add_internal (fp, in_arc, NULL,
ctf_dict_cuname (outputs[0]), NULL) < 0)
{
ctf_err (err_locus (fp), 0, _("cannot add intermediate files to link"));
goto err_outputs;
}
ctf_dict_close (out);
free (outputs);
continue;
err_inputs_outputs:
ctf_list_splice (&fp->ctf_errs_warnings, &outputs[0]->ctf_errs_warnings);
ctf_dict_close (outputs[0]);
free (outputs);
err_inputs:
ctf_link_deduplicating_close_inputs (fp, in, inputs, ninputs);
ctf_dict_close (out);
free (inputs);
free (parents);
err_open_inputs:
ctf_next_destroy (i);
return -1;
err_outputs:
ctf_list_splice (&fp->ctf_errs_warnings, &outputs[0]->ctf_errs_warnings);
ctf_dict_close (outputs[0]);
free (outputs);
ctf_next_destroy (i);
return -1; /* Errno is set for us. */
}
if (err != ECTF_NEXT_END)
{
return ctf_err (err_locus (fp), err, _("iteration error"));
}
return 0;
}
/* Empty all the ctf_link_outputs. This is often a nonessential error-path
cleanup operation, so problems are reported as warnings: use
ctf_err_upgrade() to turn into an error if fatal. */
static ctf_ret_t
ctf_link_empty_outputs (ctf_dict_t *fp)
{
ctf_next_t *i = NULL;
void *v;
ctf_error_t err;
ctf_dynhash_empty (fp->ctf_link_outputs);
while ((err = ctf_dynhash_next (fp->ctf_link_inputs, &i, NULL, &v)) == 0)
{
ctf_dict_t *in = (ctf_dict_t *) v;
in->ctf_link_in_out = NULL;
}
if (err != ECTF_NEXT_END)
{
fp->ctf_flags &= ~LCTF_LINKING;
ctf_warn (err_locus (fp), err, _("iteration error"));
return -1;
}
return 0;
}
/* Do a deduplicating link using the ctf-dedup machinery. */
static void
ctf_link_deduplicating (ctf_dict_t *fp)
{
size_t i;
ctf_dict_t **inputs, **outputs = NULL;
ssize_t ninputs;
uint32_t noutputs;
uint32_t *parents;
int cu_phase = 0;
if (ctf_dedup_atoms_init (fp) < 0)
{
ctf_err (err_locus (fp), 0, _("allocating CTF dedup atoms table"));
return; /* Errno is set for us. */
}
/* Trigger a CU-mapped link if need be: one pass of dedups squashing inputs into
single child dicts corresponding to each CU mapping, and one pass that
treats those as if they are ordinary inputs and links them together.
This latter pass does need to act very slightly differently from normal, so we
keep track of the "CU phase", with 0 being a normal link, 1 being the
squash-together phase, and 2 being the final act-as-if-it-were-normal pass. */
if (fp->ctf_link_out_cu_mapping)
{
if (ctf_link_deduplicating_per_cu (fp) < 0)
return; /* Errno is set for us. */
cu_phase = 2;
}
else if (fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST)
{
/* Ban non-cu-mapped against-first links because various structures that
are needed by non-cu-mapped links are not populated in this mode,
notably cd_nonroot_consistency. */
ctf_err (err_locus (fp), EINVAL,
_("against-first deduplication requires cu-mapping mode"));
return;
}
if ((ninputs = ctf_link_deduplicating_count_inputs (fp, NULL, NULL)) < 0)
return; /* Errno is set for us. */
if ((inputs = ctf_link_deduplicating_open_inputs (fp, NULL, ninputs,
&parents)) == NULL)
return; /* Errno is set for us. */
if (ninputs == 1 && ctf_dict_cuname (inputs[0]) != NULL)
ctf_dict_set_cuname (fp, ctf_dict_cuname (inputs[0]));
if (ctf_dedup (fp, inputs, ninputs, cu_phase) < 0)
{
ctf_err (err_locus (fp), 0, _("deduplication failed for %s"),
ctf_link_input_name (fp));
goto err;
}
if ((outputs = ctf_dedup_emit (fp, inputs, ninputs, parents, &noutputs,
cu_phase)) == NULL)
{
ctf_err (err_locus (fp), 0, _("deduplicating link type emission failed "
"for %s"), ctf_link_input_name (fp));
goto err;
}
if (!ctf_assert (fp, outputs[0] == fp))
{
for (i = 1; i < noutputs; i++)
ctf_dict_close (outputs[i]);
goto err;
}
for (i = 0; i < noutputs; i++)
{
char *dynname;
/* We already have access to this one. Close the duplicate. */
if (i == 0)
{
ctf_dict_close (outputs[0]);
continue;
}
if ((dynname = ctf_new_per_cu_name (fp, ctf_dict_cuname (outputs[i]))) == NULL)
goto oom_one_output;
if (ctf_dynhash_insert (fp->ctf_link_outputs, dynname, outputs[i]) < 0)
goto oom_one_output;
continue;
oom_one_output:
ctf_err (err_locus (fp), ENOMEM, _("allocating link outputs"));
free (dynname);
for (; i < noutputs; i++)
ctf_dict_close (outputs[i]);
goto err;
}
/* UPTODO: variable section omission, possible integration of symtypetabs (?) et
al */
if (ctf_link_deduplicating_syms (fp, inputs, ninputs, 0) < 0)
{
ctf_err (err_locus (fp), 0, _("deduplicating link symbol emission failed for "
"%s"), ctf_link_input_name (fp));
goto err_clean_outputs;
}
ctf_dedup_fini (fp, outputs, noutputs);
/* Now close all the inputs, including per-CU intermediates. */
if (ctf_link_deduplicating_close_inputs (fp, NULL, inputs, ninputs) < 0)
return; /* errno is set for us. */
ninputs = 0; /* Prevent double-close. */
ctf_set_errno (fp, 0);
/* Fall through. */
err:
for (i = 0; i < (size_t) ninputs; i++)
ctf_dict_close (inputs[i]);
free (inputs);
free (parents);
free (outputs);
return;
err_clean_outputs:
ctf_link_empty_outputs (fp);
}
/* Merge types and variable sections in all dicts added to the link together.
The result of any previous link is discarded. */
ctf_ret_t
ctf_link (ctf_dict_t *fp, ctf_link_flags_t flags)
{
ctf_error_t err;
int oldflags = fp->ctf_flags;
fp->ctf_link_flags = flags;
if (fp->ctf_link_inputs == NULL)
return 0; /* Nothing to do. */
/* Validation for against-first link mode. */
if (fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST)
{
/* If the parent is not already set or has more than one member, fail:
only dicts are permitted (since you can only have a parent dict, not a
parent archive). */
if (fp->ctf_link_parent == NULL)
{
ctf_err (err_locus (fp), EINVAL,
_("parent is not set in against-types mode: use ctf_link_against"));
return -1;
}
if (ctf_dynhash_elements (fp->ctf_link_inputs) != 2)
{
ctf_err (err_locus (fp), ECTF_NOTYET,
_("dedup-against-first link mode may only link two inputs at a time"));
return -1;
}
/* Always use share-duplicated mode. We cannot promote types found only
in the child into the parent, since we cannot modify the parent. */
fp->ctf_link_flags |= CTF_LINK_SHARE_DUPLICATED;
}
if (fp->ctf_link_outputs != NULL)
{
if (ctf_link_empty_outputs (fp) < 0)
{
ctf_err_upgrade (fp);
return -1; /* errno is set for us. */
}
}
else
fp->ctf_link_outputs = ctf_dynhash_create (ctf_hash_string,
ctf_hash_eq_string, free,
(ctf_hash_free_fun)
ctf_dict_close);
if (fp->ctf_link_outputs == NULL)
return ctf_set_errno (fp, ENOMEM);
if (!(fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST))
fp->ctf_flags |= LCTF_LINKING & LCTF_STRICT_NO_DUP_ENUMERATORS;
ctf_link_deduplicating (fp);
fp->ctf_flags = oldflags;
if ((ctf_errno (fp) != 0) && (ctf_errno (fp) != ECTF_NOCTFDATA))
return -1;
/* Create empty CUs if requested. We do not currently claim that multiple
links in succession with CTF_LINK_EMPTY_CU_MAPPINGS set in some calls and
not set in others will do anything especially sensible. */
if (fp->ctf_link_out_cu_mapping && (flags & CTF_LINK_EMPTY_CU_MAPPINGS))
{
ctf_next_t *i = NULL;
void *k;
while ((err = ctf_dynhash_next (fp->ctf_link_out_cu_mapping, &i, &k,
NULL)) == 0)
{
const char *to = (const char *) k;
if (ctf_create_per_cu (fp, NULL, to) == NULL)
{
ctf_next_destroy (i);
return -1; /* Errno is set for us. */
}
}
if (err != ECTF_NEXT_END)
{
ctf_warn (err_locus (fp), err, _("iteration error creating empty CUs"));
return ctf_set_errno (fp, err);
}
}
return 0;
}
/* Link DICT against AGAINST, without changing AGAINST. */
ctf_ret_t
ctf_link_against (ctf_dict_t *fp, ctf_archive_t *against, ctf_archive_t *dict,
const char *name, ctf_link_flags_t flags)
{
ctf_error_t err;
if (ctf_archive_count (against) != 1)
return ctf_err (err_locus (fp), ECTF_NOTYET,
_("parent must have exactly one member in against-types mode"));
/* Allow repeated calls without leaks (with some inefficiency for repeated
reopens if AGAINST is the same). */
ctf_dict_close (fp->ctf_link_parent);
if (fp->ctf_link_outputs)
ctf_link_empty_outputs (fp);
if (fp->ctf_link_inputs)
ctf_dynhash_empty (fp->ctf_link_inputs);
if (fp->ctf_link_out_cu_mapping)
ctf_dynhash_empty (fp->ctf_link_out_cu_mapping);
if (fp->ctf_link_in_cu_mapping)
ctf_dynhash_empty (fp->ctf_link_in_cu_mapping);
if ((fp->ctf_link_parent = ctf_dict_open_by_index (against, 0, &err))
== NULL)
return ctf_err (err_locus (fp), err, _("cannot open against-types parent dict"));
/* Add the sole parent, then the single child. The parent is never written
out or CU-mapped, so give it a non-colliding name. */
if (ctf_link_add (fp, against, "//@@parent@name@irrelevant//", NULL) < 0)
return -1; /* errno is set for us. */
if (ctf_link_add (fp, dict, name, NULL) < 0)
return -1; /* errno is set for us. */
/* Add a mapping from the child to itself, to force a per-CU link and
squashing into a single dict. */
if (ctf_link_add_cu_mapping (fp, name, name) < 0)
return -1; /* errno is set for us. */
flags |= CTF_LINK_DEDUP_AGAINST_FIRST;
if (ctf_link (fp, flags) < 0)
return -1; /* errno is set for us. */
if (ctf_dynhash_elements (fp->ctf_link_outputs) == 0)
{
ctf_dict_t *empty;
char *dynname;
/* No output is perfectly valid -- there are simply no types here. Create
a new dict, import the parent, and stick it in ctf_link_outputs so that
we can return something. */
if ((empty = ctf_create (fp->ctf_link_parent, &err)) == NULL)
return ctf_err (err_locus (fp), err, _("creating empty against-types link dict"));
ctf_dict_set_cuname (empty, name);
empty->ctf_parent_name = ctf_dict_cuname (fp->ctf_link_parent);
if ((dynname = ctf_new_per_cu_name (fp, name)) == NULL
|| ctf_dynhash_insert (fp->ctf_link_outputs, dynname, empty) < 0)
{
free (dynname);
ctf_dict_close (empty);
return ctf_err (err_locus (fp), ENOMEM, _("allocating link outputs"));
}
}
return 0;
}
typedef struct ctf_link_out_string_cb_arg
{
const char *str;
uint32_t offset;
ctf_error_t err;
} ctf_link_out_string_cb_arg_t;
/* Intern a string in the string table of an output per-CU CTF file. */
static void
ctf_link_intern_extern_string (void *key _libctf_unused_, void *value,
void *arg_)
{
ctf_dict_t *fp = (ctf_dict_t *) value;
ctf_link_out_string_cb_arg_t *arg = (ctf_link_out_string_cb_arg_t *) arg_;
if (!ctf_str_add_external (fp, arg->str, arg->offset))
arg->err = ENOMEM;
}
/* Repeatedly call ADD_STRING to acquire strings from the external string table,
adding them to the atoms table for this CU and all subsidiary CUs.
Must be called on a dict that has not yet been serialized.
If ctf_link is also called, it must be called first if you want the new CTF
files ctf_link can create to get their strings dedupped against the ELF
strtab properly. */
ctf_ret_t
ctf_link_add_strtab (ctf_dict_t *fp, ctf_link_strtab_string_f *add_string,
void *arg)
{
const char *str;
uint32_t offset;
ctf_error_t err = 0;
if (fp->ctf_stypes > 0)
return ctf_set_errno (fp, ECTF_RDONLY);
/* If emitting BTF, there is no external string table. */
if (ctf_serialize_output_dict_is_btf (fp))
return 0;
while ((str = add_string (&offset, arg)) != NULL)
{
ctf_link_out_string_cb_arg_t iter_arg = { str, offset, 0 };
if (!ctf_str_add_external (fp, str, offset))
err = ENOMEM;
ctf_dynhash_iter (fp->ctf_link_outputs, ctf_link_intern_extern_string,
&iter_arg);
if (iter_arg.err)
err = iter_arg.err;
}
fp->ctf_serialize.cs_initialized = 0;
if (err)
{
ctf_set_errno (fp, err);
return -1;
}
return 0;
}
/* Inform the ctf-link machinery of a new symbol in the target symbol table
(which must be some symtab that is not usually stripped, and which
is in agreement with ctf_bfdopen_ctfsect). May be called either before or
after ctf_link_add_strtab. As with that function, must be called on a dict which
has not yet been serialized. */
ctf_ret_t
ctf_link_add_linker_symbol (ctf_dict_t *fp, ctf_link_sym_t *sym)
{
ctf_in_flight_dynsym_t *cid;
/* Cheat a little: if there is already an ENOMEM error code recorded against
this dict, we shouldn't even try to add symbols because there will be no
memory to do so: probably we failed to add some previous symbol. This
makes out-of-memory exits 'sticky' across calls to this function, so the
caller doesn't need to worry about error conditions. */
if (ctf_errno (fp) == ENOMEM)
return -ENOMEM; /* errno is set for us. */
if (fp->ctf_stypes > 0)
return ctf_set_errno (fp, ECTF_RDONLY);
if (ctf_symtab_skippable (sym))
return 0;
if (sym->st_type != STT_OBJECT && sym->st_type != STT_FUNC)
return 0;
/* If emitting BTF, there is no symtypetab so linker symbols are ignored. */
if (ctf_serialize_output_dict_is_btf (fp))
return 0;
/* Add the symbol to the in-flight list. */
if ((cid = malloc (sizeof (ctf_in_flight_dynsym_t))) == NULL)
goto oom;
cid->cid_sym = *sym;
ctf_list_append (&fp->ctf_in_flight_dynsyms, cid);
return 0;
oom:
ctf_dynhash_destroy (fp->ctf_dynsyms);
fp->ctf_dynsyms = NULL;
ctf_set_errno (fp, ENOMEM);
return -ENOMEM;
}
/* Impose an ordering on symbols. The ordering takes effect immediately, but
since the ordering info does not include type IDs, lookups may return nothing
until such IDs are added by calls to ctf_add_*_sym. Must be called after
ctf_link_add_strtab and ctf_link_add_linker_symbol. */
ctf_ret_t
ctf_link_shuffle_syms (ctf_dict_t *fp)
{
ctf_in_flight_dynsym_t *did, *nid;
ctf_next_t *i = NULL;
ctf_error_t err = ENOMEM;
void *name_, *sym_;
if (fp->ctf_stypes > 0)
return ctf_set_errno (fp, ECTF_RDONLY);
/* If emitting BTF, there is no symtypetab to shuffle. */
if (ctf_serialize_output_dict_is_btf (fp))
return 0;
if (!fp->ctf_dynsyms)
{
fp->ctf_dynsyms = ctf_dynhash_create (ctf_hash_string,
ctf_hash_eq_string,
NULL, free);
if (!fp->ctf_dynsyms)
{
ctf_set_errno (fp, ENOMEM);
return -1;
}
}
/* Add all the symbols, excluding only those we already know are prohibited
from appearing in symtypetabs. */
for (did = ctf_list_next (&fp->ctf_in_flight_dynsyms); did != NULL; did = nid)
{
ctf_link_sym_t *new_sym;
nid = ctf_list_next (did);
ctf_list_delete (&fp->ctf_in_flight_dynsyms, did);
/* We might get a name or an external strtab offset. The strtab offset is
guaranteed resolvable at this point, so turn it into a string. */
if (did->cid_sym.st_name == NULL)
{
uint32_t off = CTF_SET_STID (did->cid_sym.st_nameidx, CTF_STRTAB_1);
did->cid_sym.st_name = ctf_strraw (fp, off);
did->cid_sym.st_nameidx_set = 0;
if (!ctf_assert (fp, did->cid_sym.st_name != NULL))
return -1; /* errno is set for us. */
}
/* The symbol might have turned out to be nameless, so we have to recheck
for skippability here. */
if (!ctf_symtab_skippable (&did->cid_sym))
{
ctf_dprintf ("symbol from linker: %s (%x)\n", did->cid_sym.st_name,
did->cid_sym.st_symidx);
if ((new_sym = malloc (sizeof (ctf_link_sym_t))) == NULL)
goto local_oom;
memcpy (new_sym, &did->cid_sym, sizeof (ctf_link_sym_t));
if (ctf_dynhash_cinsert (fp->ctf_dynsyms, new_sym->st_name, new_sym) < 0)
goto local_oom;
if (fp->ctf_dynsymmax < new_sym->st_symidx)
fp->ctf_dynsymmax = new_sym->st_symidx;
}
free (did);
continue;
local_oom:
free (did);
free (new_sym);
goto err;
}
/* If no symbols are reported, unwind what we have done and return. This
makes it a bit easier for the serializer to tell that no symbols have been
reported and that it should look elsewhere for reported symbols. */
if (!ctf_dynhash_elements (fp->ctf_dynsyms))
{
ctf_dprintf ("No symbols: not a final link.\n");
ctf_dynhash_destroy (fp->ctf_dynsyms);
fp->ctf_dynsyms = NULL;
return 0;
}
/* Construct a mapping from shndx to the symbol info. */
free (fp->ctf_dynsymidx);
if ((fp->ctf_dynsymidx = calloc (fp->ctf_dynsymmax + 1,
sizeof (ctf_link_sym_t *))) == NULL)
goto err;
while ((err = ctf_dynhash_next (fp->ctf_dynsyms, &i, &name_, &sym_)) == 0)
{
const char *name = (const char *) name;
ctf_link_sym_t *symp = (ctf_link_sym_t *) sym_;
if (!ctf_assert (fp, symp->st_symidx <= fp->ctf_dynsymmax))
{
ctf_next_destroy (i);
err = ctf_errno (fp);
goto err;
}
fp->ctf_dynsymidx[symp->st_symidx] = symp;
}
if (err != ECTF_NEXT_END)
{
ctf_err (err_locus (fp), err, _("iteration error"));
goto err;
}
fp->ctf_serialize.cs_initialized = 0;
return 0;
err:
/* Leave the in-flight symbols around: they'll be freed at
dict close time regardless. */
ctf_dynhash_destroy (fp->ctf_dynsyms);
fp->ctf_dynsyms = NULL;
free (fp->ctf_dynsymidx);
fp->ctf_dynsymidx = NULL;
fp->ctf_dynsymmax = 0;
ctf_set_errno (fp, err);
return -1;
}
/* Convert a 32-bit ELF symbol to a ctf_link_sym_t. */
ctf_link_sym_t *
ctf_elf32_to_link_sym (ctf_dict_t *fp, ctf_link_sym_t *dst, const Elf32_Sym *src,
uint32_t symidx)
{
Elf32_Sym tmp;
int needs_flipping = 0;
#ifdef WORDS_BIGENDIAN
if (fp->ctf_symsect_little_endian)
needs_flipping = 1;
#else
if (!fp->ctf_symsect_little_endian)
needs_flipping = 1;
#endif
memcpy (&tmp, src, sizeof (Elf32_Sym));
if (needs_flipping)
{
swap_thing (tmp.st_name);
swap_thing (tmp.st_size);
swap_thing (tmp.st_shndx);
swap_thing (tmp.st_value);
}
/* The name must be in the external string table. */
if (tmp.st_name < fp->ctf_str[CTF_STRTAB_1].cts_len)
dst->st_name = (const char *) fp->ctf_str[CTF_STRTAB_1].cts_strs + tmp.st_name;
else
dst->st_name = _CTF_NULLSTR;
dst->st_nameidx_set = 0;
dst->st_symidx = symidx;
dst->st_shndx = tmp.st_shndx;
dst->st_type = ELF32_ST_TYPE (tmp.st_info);
dst->st_value = tmp.st_value;
return dst;
}
/* Convert a 64-bit ELF symbol to a ctf_link_sym_t. */
ctf_link_sym_t *
ctf_elf64_to_link_sym (ctf_dict_t *fp, ctf_link_sym_t *dst, const Elf64_Sym *src,
uint32_t symidx)
{
Elf64_Sym tmp;
int needs_flipping = 0;
#ifdef WORDS_BIGENDIAN
if (fp->ctf_symsect_little_endian)
needs_flipping = 1;
#else
if (!fp->ctf_symsect_little_endian)
needs_flipping = 1;
#endif
memcpy (&tmp, src, sizeof (Elf64_Sym));
if (needs_flipping)
{
swap_thing (tmp.st_name);
swap_thing (tmp.st_size);
swap_thing (tmp.st_shndx);
swap_thing (tmp.st_value);
}
/* The name must be in the external string table. */
if (tmp.st_name < fp->ctf_str[CTF_STRTAB_1].cts_len)
dst->st_name = (const char *) fp->ctf_str[CTF_STRTAB_1].cts_strs + tmp.st_name;
else
dst->st_name = _CTF_NULLSTR;
dst->st_nameidx_set = 0;
dst->st_symidx = symidx;
dst->st_shndx = tmp.st_shndx;
dst->st_type = ELF32_ST_TYPE (tmp.st_info);
/* We only care if the value is zero, so avoid nonzeroes turning into
zeroes. */
if (_libctf_unlikely_ (tmp.st_value != 0 && ((uint32_t) tmp.st_value == 0)))
dst->st_value = 1;
else
dst->st_value = (uint32_t) tmp.st_value;
return dst;
}
/* Determine whether the archive that will be built from this linked dict is compatible
with pure BTF or would require CTF. (Other things may nonetheless require CTF, in
particular, compression.) */
ctf_ret_t
ctf_link_output_is_btf (ctf_dict_t *fp)
{
ctf_error_t err;
ctf_next_t *i = NULL;
void *out;
/* Can't call when nothing has been linked yet. */
if (!fp->ctf_link_outputs)
return (ctf_set_errno (fp, EINVAL));
if (ctf_serialize_output_format (fp, 0) < 0)
return -1; /* errno is set for us. */
while ((err = ctf_dynhash_next (fp->ctf_link_outputs, &i, NULL, &out)) == 0)
{
ctf_ret_t ret;
ctf_dict_t *dict = (ctf_dict_t *) out;
ret = ctf_serialize_output_dict_is_btf (dict);
/* Errors (<0) mean an error return; any dicts being non-BTF mean the link
as a whole is not BTF-compatible. */
if (ret != 1)
{
ctf_next_destroy (i);
return ret; /* errno is set for us. */
}
}
return fp->ctf_serialize.cs_is_btf;
}
typedef struct ctf_name_list_accum_cb_arg
{
ctf_dict_t *fp;
ctf_dict_t **files;
ssize_t i;
} ctf_name_list_accum_cb_arg_t;
/* Accumulate the dicts in an array, suitable for the archive-writing
machinery. */
static void
ctf_accumulate_archives (void *key, void *value, void *arg_)
{
const char *name = (const char *) key;
ctf_dict_t *fp = (ctf_dict_t *) value;
ctf_dict_t **files;
ctf_name_list_accum_cb_arg_t *arg = (ctf_name_list_accum_cb_arg_t *) arg_;
if ((files = realloc (arg->files, sizeof (ctf_dict_t *) * ++(arg->i))) == NULL)
{
(arg->i)--;
ctf_set_errno (arg->fp, ENOMEM);
return;
}
/* Allow the caller to get in and modify the cuname at the last minute. (The
original name is still used as the key of the ctf_link_outputs hash, since
those can't change, and is freed by the dynhash machinery.) */
if (fp->ctf_link_memb_name_changer)
{
char *dyname;
void *nc_arg = fp->ctf_link_memb_name_changer_arg;
dyname = fp->ctf_link_memb_name_changer (fp, name, nc_arg);
if (dyname != NULL)
{
if ((ctf_dict_set_cuname (fp, dyname)) < 0)
return; /* errno is set for us. */
free (dyname);
}
}
arg->files = files;
arg->files[(arg->i) - 1] = fp;
}
/* Write out a CTF archive (if there are per-CU CTF files) or a CTF file
(otherwise) into a new dynamically-allocated string, and return it.
The optional arg IS_BTF is set to 1 if the written output is valid BTF
(no archives, no CTF-specific types). */
unsigned char *
ctf_link_write (ctf_dict_t *fp, size_t *size, int *is_btf)
{
ctf_name_list_accum_cb_arg_t arg;
char *transformed_name = NULL;
FILE *f = NULL;
ssize_t i;
ctf_error_t err;
long fsize;
const char *errloc;
unsigned char *buf = NULL;
ctf_arc_write_flags_t flags = 0;
memset (&arg, 0, sizeof (ctf_name_list_accum_cb_arg_t));
arg.fp = fp;
fp->ctf_flags |= LCTF_LINKING;
if (fp->ctf_link_outputs)
{
ctf_dynhash_iter (fp->ctf_link_outputs, ctf_accumulate_archives, &arg);
if (ctf_errno (fp) < 0)
{
errloc = "hash creation";
goto err;
}
}
if (is_btf)
*is_btf = 0;
/* Writing an archive. Stick ourselves (the shared repository, parent of all
other archives) on the front of it with the default name, unless
against-first mode is on. (The parent dict must always come first, by
definition -- the first dict in an archive is the parent -- but also doing
so is essential for strings in child dicts shared with the parent to get
their proper offsets.)
If against-first mode is on, rather than linking all input dicts into fp
and some set of children, all dicts after the first are linked into a
single dict in the output. */
if (fp->ctf_link_memb_name_changer)
{
void *nc_arg = fp->ctf_link_memb_name_changer_arg;
transformed_name = fp->ctf_link_memb_name_changer (fp, _CTF_SECTION,
nc_arg);
if (transformed_name != NULL
&& ctf_dict_set_cuname (fp, transformed_name) < 0)
{
errloc = "cuname transformation";
goto err;
}
}
/* Propagate the link flags to all the dicts in this link. Suppress string
deduplication if against-first mode is on. */
for (i = 0; i < arg.i; i++)
{
arg.files[i]->ctf_link_flags = fp->ctf_link_flags;
arg.files[i]->ctf_flags |= LCTF_LINKING;
if (fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST)
arg.files[i]->ctf_flags |= LCTF_NO_STR_DEDUP;
}
if (!(fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST))
{
ctf_dict_t **files;
if ((files = realloc (arg.files,
sizeof (struct ctf_dict *) * (arg.i + 1))) == NULL)
{
errloc = "ctf_dict reallocation";
goto err_no;
}
arg.files = files;
memmove (&(arg.files[1]), arg.files, sizeof (ctf_dict_t *) * (arg.i));
arg.files[0] = fp;
arg.i++;
}
if ((f = tmpfile ()) == NULL)
{
errloc = "tempfile creation";
goto err_no;
}
if ((err = ctf_arc_write_fd (fileno (f), arg.files, arg.i, flags)) != 0)
{
errloc = NULL; /* errno is set for us. */
goto err_set;
}
if (fseek (f, 0, SEEK_END) < 0)
{
errloc = "seeking to end";
goto err_no;
}
if ((fsize = ftell (f)) < 0)
{
errloc = "filesize determination";
goto err_no;
}
if (fseek (f, 0, SEEK_SET) < 0)
{
errloc = "filepos resetting";
goto err_no;
}
if ((buf = malloc (fsize)) == NULL)
{
errloc = "CTF archive buffer allocation";
goto err_no;
}
while (!feof (f) && fread (buf, fsize, 1, f) == 0)
if (ferror (f))
{
errloc = "reading archive from temporary file";
goto err_no;
}
/* If the user wanted to know if we wrote BTF out, normally we ask this about
the entire archive, link-against dicts, we only wrote one dict out, so we
want to make sure that that dict alone is BTF. */
if (is_btf)
{
if (!(fp->ctf_link_flags & CTF_LINK_DEDUP_AGAINST_FIRST))
{
if ((*is_btf = ctf_link_output_is_btf (fp)) < 0)
{
errloc = "determining if archive is BTF";
goto err; /* errno is set for us. */
}
}
else
{
if ((*is_btf = ctf_serialize_output_dict_is_btf (arg.files[0])) < 0)
{
errloc = "determining if archive is BTF";
goto err; /* errno is set for us. */
}
}
}
/* Turn off the is-linking flag, and any other flags we flipped, on all the
dicts in this link. */
for (i = 0; i < arg.i; i++)
arg.files[i]->ctf_flags &= ~(LCTF_LINKING | LCTF_NO_STR_DEDUP);
*size = fsize;
free (arg.files);
free (transformed_name);
fclose (f);
return buf;
err_no:
ctf_set_errno (fp, errno);
err_set:
/* Turn off flags we flipped on all the dicts in this link, as above. */
for (i = 0; i < arg.i; i++)
arg.files[i]->ctf_flags &= ~(LCTF_LINKING | LCTF_NO_STR_DEDUP);
err:
free (buf);
if (f)
fclose (f);
free (arg.files);
free (transformed_name);
if (errloc)
ctf_err (err_locus (fp), 0, _("%s"), errloc);
else
ctf_err (err_locus (fp), 0, NULL);
return NULL;
}
Reference in a new issue View git blame Copy permalink