The method of operation of libctf when the dictionary is writable has
before now been that types that are added land in the dynamic type
section, which is a linked list and hash of IDs -> dynamic type
definitions (and, recently a hash of names): the DTDs are a bit of CTF
representing the ctf_type_t and ad hoc C structures representing the
vlen. Historically, libctf was unable to do anything with these types,
not even look them up by ID, let alone by name: if you wanted to do that
say if you were adding a type that depended on one you just added) you
called ctf_update, which serializes all the DTDs into a CTF file and
reopens it, copying its guts over the fp it's called with. The
ctf_updated types are then frozen in amber and unchangeable: all lookups
will return the types in the static portion in preference to the dynamic
portion, and we will refuse to re-add things that already exist in the
static portion (and, of late, in the dynamic portion too). The libctf
machinery remembers the boundary between static and dynamic types and
looks in the right portion for each type. Lots of things still don't
quite work with dynamic types (e.g. getting their size), but enough
works to do a bunch of additions and then a ctf_update, most of the
time.
Except it doesn't, because ctf_add_type finds it necessary to walk the
full dynamic type definition list looking for types with matching names,
so it gets slower and slower with every type you add: fixing this
requires calling ctf_update periodically for no other reason than to
avoid massively slowing things down.
This is all clunky and very slow but kind of works, until you consider
that it is in fact possible and indeed necessary to modify one sort of
type after it has been added: forwards. These are necessarily promoted
to structs, unions or enums, and when they do so *their type ID does not
change*. So all of a sudden we are changing types that already exist in
the static portion. ctf_update gets massively confused by this and
allocates space enough for the forward (with no members), but then emits
the new dynamic type (with all the members) into it. You get an
assertion failure after that, if you're lucky, or a coredump.
So this commit rejigs things a bit and arranges to exclusively use the
dynamic type definitions in writable dictionaries, and the static type
definitions in readable dictionaries: we don't at any time have a mixture
of static and dynamic types, and you don't need to call ctf_update to
make things "appear". The ctf_dtbyname hash I introduced a few months
ago, which maps things like "struct foo" to DTDs, is removed, replaced
instead by a change of type of the four dictionaries which track names.
Rather than just being (unresizable) ctf_hash_t's populated only at
ctf_bufopen time, they are now a ctf_names_t structure, which is a pair
of ctf_hash_t and ctf_dynhash_t, with the ctf_hash_t portion being used
in readonly dictionaries, and the ctf_dynhash_t being used in writable
ones. The decision as to which to use is centralized in the new
functions ctf_lookup_by_rawname (which takes a type kind) and
ctf_lookup_by_rawhash, which it calls (which takes a ctf_names_t *.)
This change lets us switch from using static to dynamic name hashes on
the fly across the entirety of libctf without complexifying anything: in
fact, because we now centralize the knowledge about how to map from type
kind to name hash, it actually simplifies things and lets us throw out
quite a lot of now-unnecessary complexity, from ctf_dtnyname (replaced
by the dynamic half of the name tables), through to ctf_dtnextid (now
that a dictionary's static portion is never referenced if the dictionary
is writable, we can just use ctf_typemax to indicate the maximum type:
dynamic or non-dynamic does not matter, and we no longer need to track
the boundary between the types). You can now ctf_rollback() as far as
you like, even past a ctf_update or for that matter a full writeout; all
the iteration functions work just as well on writable as on read-only
dictionaries; ctf_add_type no longer needs expensive duplicated code to
run over the dynamic types hunting for ones it might be interested in;
and the linker no longer needs a hack to call ctf_update so that calling
ctf_add_type is not impossibly expensive.
There is still a bit more complexity: some new code paths in ctf-types.c
need to know how to extract information from dynamic types. This
complexity will go away again in a few months when libctf acquires a
proper intermediate representation.
You can still call ctf_update if you like (it's public API, after all),
but its only effect now is to set the point to which ctf_discard rolls
back.
Obviously *something* still needs to serialize the CTF file before
writeout, and this job is done by ctf_serialize, which does everything
ctf_update used to except set the counter used by ctf_discard. It is
automatically called by the various functions that do CTF writeout:
nobody else ever needs to call it.
With this in place, forwards that are promoted to non-forwards no longer
crash the link, even if it happens tens of thousands of types later.
v5: fix tabdamage.
libctf/
* ctf-impl.h (ctf_names_t): New.
(ctf_lookup_t) <ctf_hash>: Now a ctf_names_t, not a ctf_hash_t.
(ctf_file_t) <ctf_structs>: Likewise.
<ctf_unions>: Likewise.
<ctf_enums>: Likewise.
<ctf_names>: Likewise.
<ctf_lookups>: Improve comment.
<ctf_ptrtab_len>: New.
<ctf_prov_strtab>: New.
<ctf_str_prov_offset>: New.
<ctf_dtbyname>: Remove, redundant to the names hashes.
<ctf_dtnextid>: Remove, redundant to ctf_typemax.
(ctf_dtdef_t) <dtd_name>: Remove.
<dtd_data>: Note that the ctt_name is now populated.
(ctf_str_atom_t) <csa_offset>: This is now the strtab
offset for internal strings too.
<csa_external_offset>: New, the external strtab offset.
(CTF_INDEX_TO_TYPEPTR): Handle the LCTF_RDWR case.
(ctf_name_table): New declaration.
(ctf_lookup_by_rawname): Likewise.
(ctf_lookup_by_rawhash): Likewise.
(ctf_set_ctl_hashes): Likewise.
(ctf_serialize): Likewise.
(ctf_dtd_insert): Adjust.
(ctf_simple_open_internal): Likewise.
(ctf_bufopen_internal): Likewise.
(ctf_list_empty_p): Likewise.
(ctf_str_remove_ref): Likewise.
(ctf_str_add): Returns uint32_t now.
(ctf_str_add_ref): Likewise.
(ctf_str_add_external): Now returns a boolean (int).
* ctf-string.c (ctf_strraw_explicit): Check the ctf_prov_strtab
for strings in the appropriate range.
(ctf_str_create_atoms): Create the ctf_prov_strtab. Detect OOM
when adding the null string to the new strtab.
(ctf_str_free_atoms): Destroy the ctf_prov_strtab.
(ctf_str_add_ref_internal): Add make_provisional argument. If
make_provisional, populate the offset and fill in the
ctf_prov_strtab accordingly.
(ctf_str_add): Return the offset, not the string.
(ctf_str_add_ref): Likewise.
(ctf_str_add_external): Return a success integer.
(ctf_str_remove_ref): New, remove a single ref.
(ctf_str_count_strtab): Do not count the initial null string's
length or the existence or length of any unreferenced internal
atoms.
(ctf_str_populate_sorttab): Skip atoms with no refs.
(ctf_str_write_strtab): Populate the nullstr earlier. Add one
to the cts_len for the null string, since it is no longer done
in ctf_str_count_strtab. Adjust for csa_external_offset rename.
Populate the csa_offset for both internal and external cases.
Flush the ctf_prov_strtab afterwards, and reset the
ctf_str_prov_offset.
* ctf-create.c (ctf_grow_ptrtab): New.
(ctf_create): Call it. Initialize new fields rather than old
ones. Tell ctf_bufopen_internal that this is a writable dictionary.
Set the ctl hashes and data model.
(ctf_update): Rename to...
(ctf_serialize): ... this. Leave a compatibility function behind.
Tell ctf_simple_open_internal that this is a writable dictionary.
Pass the new fields along from the old dictionary. Drop
ctf_dtnextid and ctf_dtbyname. Use ctf_strraw, not dtd_name.
Do not zero out the DTD's ctt_name.
(ctf_prefixed_name): Rename to...
(ctf_name_table): ... this. No longer return a prefixed name: return
the applicable name table instead.
(ctf_dtd_insert): Use it, and use the right name table. Pass in the
kind we're adding. Migrate away from dtd_name.
(ctf_dtd_delete): Adjust similarly. Remove the ref to the
deleted ctt_name.
(ctf_dtd_lookup_type_by_name): Remove.
(ctf_dynamic_type): Always return NULL on read-only dictionaries.
No longer check ctf_dtnextid: check ctf_typemax instead.
(ctf_snapshot): No longer use ctf_dtnextid: use ctf_typemax instead.
(ctf_rollback): Likewise. No longer fail with ECTF_OVERROLLBACK. Use
ctf_name_table and the right name table, and migrate away from
dtd_name as in ctf_dtd_delete.
(ctf_add_generic): Pass in the kind explicitly and pass it to
ctf_dtd_insert. Use ctf_typemax, not ctf_dtnextid. Migrate away
from dtd_name to using ctf_str_add_ref to populate the ctt_name.
Grow the ptrtab if needed.
(ctf_add_encoded): Pass in the kind.
(ctf_add_slice): Likewise.
(ctf_add_array): Likewise.
(ctf_add_function): Likewise.
(ctf_add_typedef): Likewise.
(ctf_add_reftype): Likewise. Initialize the ctf_ptrtab, checking
ctt_name rather than dtd_name.
(ctf_add_struct_sized): Pass in the kind. Use
ctf_lookup_by_rawname, not ctf_hash_lookup_type /
ctf_dtd_lookup_type_by_name.
(ctf_add_union_sized): Likewise.
(ctf_add_enum): Likewise.
(ctf_add_enum_encoded): Likewise.
(ctf_add_forward): Likewise.
(ctf_add_type): Likewise.
(ctf_compress_write): Call ctf_serialize: adjust for ctf_size not
being initialized until after the call.
(ctf_write_mem): Likewise.
(ctf_write): Likewise.
* ctf-archive.c (arc_write_one_ctf): Likewise.
* ctf-lookup.c (ctf_lookup_by_name): Use ctf_lookuup_by_rawhash, not
ctf_hash_lookup_type.
(ctf_lookup_by_id): No longer check the readonly types if the
dictionary is writable.
* ctf-open.c (init_types): Assert that this dictionary is not
writable. Adjust to use the new name hashes, ctf_name_table,
and ctf_ptrtab_len. GNU style fix for the final ptrtab scan.
(ctf_bufopen_internal): New 'writable' parameter. Flip on LCTF_RDWR
if set. Drop out early when dictionary is writable. Split the
ctf_lookups initialization into...
(ctf_set_cth_hashes): ... this new function.
(ctf_simple_open_internal): Adjust. New 'writable' parameter.
(ctf_simple_open): Adjust accordingly.
(ctf_bufopen): Likewise.
(ctf_file_close): Destroy the appropriate name hashes. No longer
destroy ctf_dtbyname, which is gone.
(ctf_getdatasect): Remove spurious "extern".
* ctf-types.c (ctf_lookup_by_rawname): New, look up types in the
specified name table, given a kind.
(ctf_lookup_by_rawhash): Likewise, given a ctf_names_t *.
(ctf_member_iter): Add support for iterating over the
dynamic type list.
(ctf_enum_iter): Likewise.
(ctf_variable_iter): Likewise.
(ctf_type_rvisit): Likewise.
(ctf_member_info): Add support for types in the dynamic type list.
(ctf_enum_name): Likewise.
(ctf_enum_value): Likewise.
(ctf_func_type_info): Likewise.
(ctf_func_type_args): Likewise.
* ctf-link.c (ctf_accumulate_archive_names): No longer call
ctf_update.
(ctf_link_write): Likewise.
(ctf_link_intern_extern_string): Adjust for new
ctf_str_add_external return value.
(ctf_link_add_strtab): Likewise.
* ctf-util.c (ctf_list_empty_p): New.
GCC can emit references to type 0 to indicate that this type is one that
is not representable in the version of CTF it emits (for instance,
version 3 cannot encode vector types). Type 0 is already used in the
function section to indicate padding inserted to skip functions we do
not want to encode the type of, so using zero in this way is a good
extension of the format: but libctf reports such types as ECTF_BADID,
which is indistinguishable from file corruption via links to truly
nonexistent types with IDs like 0xDEADBEEF etc, which we really do want
to stop for.
In particular, this stops all traversals of types dead at this point,
preventing us from even dumping CTF files containing unrepresentable
types to see what's going on!
So add a new error, ECTF_NONREPRESENTABLE, which is returned by
recursive type resolution when a reference to a zero type is found. (No
zero type is ever emitted into the CTF file by GCC, only references to
one). We can't do much with types that are ultimately nonrepresentable,
but we can do enough to keep functioning.
Adjust ctf_add_type to ensure that top-level types of type zero and
structure and union members of ultimate type zero are simply skipped
without reporting an error, so we can copy structures and unions that
contain nonrepresentable members (skipping them and leaving a hole where
they would be, so no consumers downstream of the linker need to worry
about this): adjust the dumper so that we dump members of
nonrepresentable types in a simple form that indicates
nonrepresentability rather than terminating the dump, and do not falsely
assume all errors to be -ENOMEM: adjust the linker so that types that
fail to get added are simply skipped, so that both nonrepresentable
types and outright errors do not terminate the type addition, which
could skip many valid types and cause further errors when variables of
those types are added.
In future, when we gain the ability to call back to the linker to report
link-time type resolution errors, we should report failures to add all
but nonrepresentable types. But we can't do that yet.
v5: Fix tabdamage.
include/
* ctf-api.h (ECTF_NONREPRESENTABLE): New.
libctf/
* ctf-types.c (ctf_type_resolve): Return ECTF_NONREPRESENTABLE on
type zero.
* ctf-create.c (ctf_add_type): Detect and skip nonrepresentable
members and types.
(ctf_add_variable): Likewise for variables pointing to them.
* ctf-link.c (ctf_link_one_type): Do not warn for nonrepresentable
type link failure, but do warn for others.
* ctf-dump.c (ctf_dump_format_type): Likewise. Do not assume all
errors to be ENOMEM.
(ctf_dump_member): Likewise.
(ctf_dump_type): Likewise.
(ctf_dump_header_strfield): Do not assume all errors to be ENOMEM.
(ctf_dump_header_sectfield): Do not assume all errors to be ENOMEM.
(ctf_dump_header): Likewise.
(ctf_dump_label): likewise.
(ctf_dump_objts): likewise.
(ctf_dump_funcs): likewise.
(ctf_dump_var): likewise.
(ctf_dump_str): Likewise.
Once the deduplicator is capable of actually detecting conflicting types
with the same name (i.e., not yet) we will place such conflicting types,
and types that depend on them, into CTF dictionaries that are the child
of the main dictionary we usually emit: currently, this will lead to the
.ctf section becoming a CTF archive rather than a single dictionary,
with the default-named archive member (_CTF_SECTION, or NULL) being the
main shared dictionary with most of the types in it.
By default, the sections are named after the compilation unit they come
from (complete path and all), with the cuname field in the CTF header
providing further evidence of the name without requiring the caller to
engage in tiresome parsing. But some callers may not wish the mapping
from input CU to output sub-dictionary to be purely CU-based.
The machinery here allows this to be freely changed, in two ways:
- callers can call ctf_link_add_cu_mapping to specify that a single
input compilation unit should have its types placed in some other CU
if they conflict: the CU will always be created, even if empty, so
the consuming program can depend on its existence. You can map
multiple input CUs to one output CU to force all their types to be
merged together: if some of *those* types conflict, the behaviour is
currently unspecified (the new deduplicator will specify it).
- callers can call ctf_link_set_memb_name_changer to provide a function
which is passed every CTF sub-dictionary name in turn (including
_CTF_SECTION) and can return a new name, or NULL if no change is
desired. The mapping from input to output names should not map two
input names to the same output name: if this happens, the two are not
merged but will result in an archive with two members with the same
name (technically valid, but it's hard to access the second
same-named member: you have to do an iteration over archive members).
This is used by the kernel's ctfarchive machinery (not yet upstream) to
encode CTF under member names like {module name}.ctf rather than
.ctf.CU, but it is anticipated that other large projects may wish to
have their own storage for CTF outside of .ctf sections and may wish to
have new naming schemes that suit their special-purpose consumers.
New in v3.
v4: check for strdup failure.
v5: fix tabdamage.
include/
* ctf-api.h (ctf_link_add_cu_mapping): New.
(ctf_link_memb_name_changer_f): New.
(ctf_link_set_memb_name_changer): New.
libctf/
* ctf-impl.h (ctf_file_t) <ctf_link_cu_mappping>: New.
<ctf_link_memb_name_changer>: Likewise.
<ctf_link_memb_name_changer_arg>: Likewise.
* ctf-create.c (ctf_update): Update accordingly.
* ctf-open.c (ctf_file_close): Likewise.
* ctf-link.c (ctf_create_per_cu): Apply the cu mapping.
(ctf_link_add_cu_mapping): New.
(ctf_link_set_memb_name_changer): Likewise.
(ctf_change_parent_name): New.
(ctf_name_list_accum_cb_arg_t) <dynames>: New, storage for names
allocated by the caller's ctf_link_memb_name_changer.
<ndynames>: Likewise.
(ctf_accumulate_archive_names): Call the ctf_link_memb_name_changer.
(ctf_link_write): Likewise (for _CTF_SECTION only): also call
ctf_change_parent_name. Free any resulting names.
The compiler describes the name and type of all file-scope variables in
this section. Merging it at link time requires using the type mapping
added in the previous commit to determine the appropriate type for the
variable in the output, given its type in the input: we check the shared
container first, and if the type doesn't exist there, it must be a
conflicted type in the per-CU child, and the variable should go there
too. We also put the variable in the per-CU child if a variable with
the same name but a different type already exists in the parent: we
ignore any such conflict in the child because CTF cannot represent such
things, nor can they happen unless a third-party linking program has
overridden the mapping of CU to CTF archive member name (using machinery
added in a later commit).
v3: rewritten using an algorithm that actually works in the case of
conflicting names. Some code motion from the next commit. Set
the per-CU parent name.
v4: check for strdup failure.
v5: fix tabdamage.
include/
* ctf-api.h (ECTF_INTERNAL): New.
libctf/
* ctf-link.c (ctf_create_per_cu): New, refactored out of...
(ctf_link_one_type): ... here, with parent-name setting added.
(check_variable): New.
(ctf_link_one_variable): Likewise.
(ctf_link_one_input_archive_member): Call it.
* ctf-error.c (_ctf_errlist): Updated with new errors.
This lets you call ctf_type_mapping (dest_fp, src_fp, src_type_id)
and get told what type ID the corresponding type has in the target
ctf_file_t. This works even if it was added by a recursive call, and
because it is stored in the target ctf_file_t it works even if we
had to add one type to multiple ctf_file_t's as part of conflicting
type handling.
We empty out this mapping after every archive is linked: because it maps
input to output fps, and we only visit each input fp once, its contents
are rendered entirely useless every time the source fp changes.
v3: add several missing mapping additions. Add ctf_dynhash_empty, and
empty after every input archive.
v5: fix tabdamage.
libctf/
* ctf-impl.h (ctf_file_t): New field ctf_link_type_mapping.
(struct ctf_link_type_mapping_key): New.
(ctf_hash_type_mapping_key): Likewise.
(ctf_hash_eq_type_mapping_key): Likewise.
(ctf_add_type_mapping): Likewise.
(ctf_type_mapping): Likewise.
(ctf_dynhash_empty): Likewise.
* ctf-open.c (ctf_file_close): Update accordingly.
* ctf-create.c (ctf_update): Likewise.
(ctf_add_type): Populate the mapping.
* ctf-hash.c (ctf_hash_type_mapping_key): Hash a type mapping key.
(ctf_hash_eq_type_mapping_key): Check the key for equality.
(ctf_dynhash_insert): Fix comment typo.
(ctf_dynhash_empty): New.
* ctf-link.c (ctf_add_type_mapping): New.
(ctf_type_mapping): Likewise.
(empty_link_type_mapping): New.
(ctf_link_one_input_archive): Call it.
This is the start of work on the core of the linking mechanism for CTF
sections. This commit handles the type and string sections.
The linker calls these functions in sequence:
ctf_link_add_ctf: to add each CTF section in the input in turn to a
newly-created ctf_file_t (which will appear in the output, and which
itself will become the shared parent that contains types that all
TUs have in common (in all link modes) and all types that do not
have conflicting definitions between types (by default). Input files
that are themselves products of ld -r are supported, though this is
not heavily tested yet.
ctf_link: called once all input files are added to merge the types in
all the input containers into the output container, eliminating
duplicates.
ctf_link_add_strtab: called once the ELF string table is finalized and
all its offsets are known, this calls a callback provided by the
linker which returns the string content and offset of every string in
the ELF strtab in turn: all these strings which appear in the input
CTF strtab are eliminated from it in favour of the ELF strtab:
equally, any strings that only appear in the input strtab will
reappear in the internal CTF strtab of the output.
ctf_link_shuffle_syms (not yet implemented): called once the ELF symtab
is finalized, this calls a callback provided by the linker which
returns information on every symbol in turn as a ctf_link_sym_t. This
is then used to shuffle the function info and data object sections in
the CTF section into symbol table order, eliminating the index
sections which map those sections to symbol names before that point.
Currently just returns ECTF_NOTYET.
ctf_link_write: Returns a buffer containing either a serialized
ctf_file_t (if there are no types with conflicting definitions in the
object files in the link) or a ctf_archive_t containing a large
ctf_file_t (the common types) and a bunch of small ones named after
individual CUs in which conflicting types are found (containing the
conflicting types, and all types that reference them). A threshold
size above which compression takes place is passed as one parameter.
(Currently, only gzip compression is supported, but I hope to add lzma
as well.)
Lifetime rules for this are simple: don't close the input CTF files
until you've called ctf_link for the last time. We do not assume
that symbols or strings passed in by the callback outlast the
call to ctf_link_add_strtab or ctf_link_shuffle_syms.
Right now, the duplicate elimination mechanism is the one already
present as part of the ctf_add_type function, and is not particularly
good: it misses numerous actual duplicates, and the conflicting-types
detection hardly ever reports that types conflict, even when they do
(one of them just tends to get silently dropped): it is also very slow.
This will all be fixed in the next few weeks, but the fix hardly touches
any of this code, and the linker does work without it, just not as
well as it otherwise might. (And when no CTF section is present,
there is no effect on performance, of course. So only people using
a trunk GCC with not-yet-committed patches will even notice. By the
time it gets upstream, things should be better.)
v3: Fix error handling.
v4: check for strdup failure.
v5: fix tabdamage.
include/
* ctf-api.h (struct ctf_link_sym): New, a symbol in flight to the
libctf linking machinery.
(CTF_LINK_SHARE_UNCONFLICTED): New.
(CTF_LINK_SHARE_DUPLICATED): New.
(ECTF_LINKADDEDLATE): New, replacing ECTF_UNUSED.
(ECTF_NOTYET): New, a 'not yet implemented' message.
(ctf_link_add_ctf): New, add an input file's CTF to the link.
(ctf_link): New, merge the type and string sections.
(ctf_link_strtab_string_f): New, callback for feeding strtab info.
(ctf_link_iter_symbol_f): New, callback for feeding symtab info.
(ctf_link_add_strtab): New, tell the CTF linker about the ELF
strtab's strings.
(ctf_link_shuffle_syms): New, ask the CTF linker to shuffle its
symbols into symtab order.
(ctf_link_write): New, ask the CTF linker to write the CTF out.
libctf/
* ctf-link.c: New file, linking of the string and type sections.
* Makefile.am (libctf_a_SOURCES): Add it.
* Makefile.in: Regenerate.
* ctf-impl.h (ctf_file_t): New fields ctf_link_inputs,
ctf_link_outputs.
* ctf-create.c (ctf_update): Update accordingly.
* ctf-open.c (ctf_file_close): Likewise.
* ctf-error.c (_ctf_errlist): Updated with new errors.
Alan Modra
Nick Alcock