The preceding change revealed a new bug: the string table is sorted for
better compression, so repeated serialization with type (or member)
additions in the middle can move strings around. But every
serialization flushes the set of refs (the memory locations that are
automatically updated with a final string offset when the strtab is
updated), so if we are not to have string offsets go stale, we must do
all ref additions within the serialization code (which walks the
complete set of types and symbols anyway). Unfortunately, we were adding
one ref in another place: the type name in the dynamic type definitions,
which has a ref added to it by ctf_add_generic.
So adding a type, serializing (via, say, one of the ctf_write
functions), adding another type with a name that sorts earlier, and
serializing again will corrupt the name of the first type because it no
longer had a ref pointing to its dtd entry's name when its string offset
was shifted later in the strtab to mae way for the other type.
To ensure that we don't miss strings, we also maintain a set of *pending
refs* that will be added later (during serialization), and remove
entries from that set when the ref is finally added. We always use
ctf_str_add_pending outside ctf-serialize.c, ensure that ctf_serialize
adds all strtab offsets as refs (even those in the dtds) on every
serialization, and mandate that no refs are live on entry to
ctf_serialize and that all pending refs are gone before strtab
finalization. (Of necessity ctf_serialize has to traverse all strtab
offsets in the dtds in order to serialize them, so adding them as refs
at the same time is easy.)
(Note that we still can't erase unused atoms when we roll back, though
we can erase unused refs: members and enums are still not removed by
rollbacks and might reference strings added after the snapshot.)
libctf/ChangeLog
2021-03-18 Nick Alcock <nick.alcock@oracle.com>
* ctf-hash.c (ctf_dynset_elements): New.
* ctf-impl.h (ctf_dynset_elements): Declare it.
(ctf_str_add_pending): Likewise.
(ctf_dict_t) <ctf_str_pending_ref>: New, set of refs that must be
added during serialization.
* ctf-string.c (ctf_str_create_atoms): Initialize it.
(CTF_STR_ADD_REF): New flag.
(CTF_STR_MAKE_PROVISIONAL): Likewise.
(CTF_STR_PENDING_REF): Likewise.
(ctf_str_add_ref_internal): Take a flags word rather than int
params. Populate, and clear out, ctf_str_pending_ref.
(ctf_str_add): Adjust accordingly.
(ctf_str_add_external): Likewise.
(ctf_str_add_pending): New.
(ctf_str_remove_ref): Also remove the potential ref if it is a
pending ref.
* ctf-serialize.c (ctf_serialize): Prohibit addition of strings
with ctf_str_add_ref before serialization. Ensure that the
ctf_str_pending_ref set is empty before strtab finalization.
(ctf_emit_type_sect): Add a ref to the ctt_name.
* ctf-create.c (ctf_add_generic): Add the ctt_name as a pending
ref.
* testsuite/libctf-writable/reserialize-strtab-corruption.*: New test.
The existing ctf_lookup_by_symbol and ctf_arc_lookup_symbol functions
suffice to look up the types of symbols if the caller already has a
symbol number. But the caller often doesn't have one of those and only
knows the name of the symbol: also, in object files, the caller might
not have a useful symbol number in any sense (and neither does libctf:
the 'symbol number' we use in that case literally starts at 0 for the
lexicographically first-sorted symbol in the symtypetab and counts those
symbols, so it corresponds to nothing useful).
This means that even though object files have a symtypetab (generated by
the compiler or by ld -r), the only way we can look up anything in it is
to iterate over all symbols in turn with ctf_symbol_next until we find
the one we want.
This is unhelpful and pointlessly inefficient.
So add a pair of functions to look up symbols by name in a dict and in a
whole archive: ctf_lookup_by_symbol_name and ctf_arc_lookup_symbol_name.
These are identical to the existing functions except that they take
symbol names rather than symbol numbers.
To avoid insane repetition, we do some refactoring in the process, so
that both ctf_lookup_by_symbol and ctf_arc_lookup_symbol turn into thin
wrappers around internal functions that do both lookup by symbol index
and lookup by name. This massively reduces code duplication because
even the existing lookup-by-index stuff wants to use a name sometimes
(when looking up in indexed sections), and the new lookup-by-name stuff
has to turn it into an index sometimes (when looking up in non-indexed
sections): doing it this way lets us share most of that.
The actual name->index lookup is done by ctf_lookup_symbol_idx. We do
not anticipate this lookup to be as heavily used as ld.so symbol lookup
by many orders of magnitude, so using the ELF symbol hashes would
probably take more time to read them than is saved by using the hashes,
and it adds a lot of complexity. Instead, do a linear search for the
symbol name, caching all the name -> index mappings as we go, so that
future searches are likely to hit in the cache. To avoid having to
repeat this search over and over in a CTF archive when
ctf_arc_lookup_symbol_name is used, have cached archive lookups (the
sort done by ctf_arc_lookup_symbol* and the ctf_archive_next iterator)
pick out the first dict they cache in a given archive and store it in a
new ctf_archive field, ctfi_crossdict_cache. This can be used to store
cross-dictionary cached state that depends on things like the ELF symbol
table rather than the contents of any one dict. ctf_lookup_symbol_idx
then caches its name->index mappings in the dictionary named in the
crossdict cache, if any, so that ctf_lookup_symbol_idx in other dicts
in the same archive benefit from the previous linear search, and the
symtab only needs to be scanned at most once.
(Note that if you call ctf_lookup_by_symbol_name in one specific dict,
and then follow it with a ctf_arc_lookup_symbol_name, the former will
not use the crossdict cache because it's only populated by the dict
opens in ctf_arc_lookup_symbol_name. This is harmless except for a small
one-off waste of memory and time: it's only a cache, after all. We can
fix this later by using the archive caching machinery more
aggressively.)
In ctf-archive, we do similar things, turning ctf_arc_lookup_symbol into
a wrapper around a new function that does both index -> ID and name ->
ID lookups across all dicts in an archive. We add a new
ctfi_symnamedicts cache that maps symbol names to the ctf_dict_t * that
it was found in (so that linear searches for symbols don't need to be
repeated): but we also *remove* a cache, the ctfi_syms cache that was
memoizing the actual ctf_id_t returned from every call to
ctf_arc_lookup_symbol. This is pointless: all it saves is one call to
ctf_lookup_by_symbol, and that's basically an array lookup and nothing
more so isn't worth caching. (Equally, given that symbol -> index
mappings are cached by ctf_lookup_by_symbol_name, those calls are nearly
free after the first call, so there's no point caching the ctf_id_t in
that case either.)
We fix up one test that was doing manual symbol lookup to use
ctf_arc_lookup_symbol instead, and enhance it to check that the caching
layer is not totally broken: we also add a new test to do lookups in a
.o file, and another to do lookups in an archive with conflicted types
and make sure that sort of multi-dict lookup is actually working.
include/ChangeLog
2021-02-17 Nick Alcock <nick.alcock@oracle.com>
* ctf-api.h (ctf_arc_lookup_symbol_name): New.
(ctf_lookup_by_symbol_name): Likewise.
libctf/ChangeLog
2021-02-17 Nick Alcock <nick.alcock@oracle.com>
* ctf-impl.h (ctf_dict_t) <ctf_symhash>: New.
<ctf_symhash_latest>: Likewise.
(struct ctf_archive_internal) <ctfi_crossdict_cache>: New.
<ctfi_symnamedicts>: New.
<ctfi_syms>: Remove.
(ctf_lookup_symbol_name): Remove.
* ctf-lookup.c (ctf_lookup_symbol_name): Propagate errors from
parent properly. Make static.
(ctf_lookup_symbol_idx): New, linear search for the symbol name,
cached in the crossdict cache's ctf_symhash (if available), or
this dict's (otherwise).
(ctf_try_lookup_indexed): Allow the symname to be passed in.
(ctf_lookup_by_symbol): Turn into a wrapper around...
(ctf_lookup_by_sym_or_name): ... this, supporting name lookup too,
using ctf_lookup_symbol_idx in non-writable dicts. Special-case
name lookup in dynamic dicts without reported symbols, which have
no symtab or dynsymidx but where name lookup should still work.
(ctf_lookup_by_symbol_name): New, another wrapper.
* ctf-archive.c (enosym): Note that this is present in
ctfi_symnamedicts too.
(ctf_arc_close): Adjust for removal of ctfi_syms. Free the
ctfi_symnamedicts.
(ctf_arc_flush_caches): Likewise.
(ctf_dict_open_cached): Memoize the first cached dict in the
crossdict cache.
(ctf_arc_lookup_symbol): Turn into a wrapper around...
(ctf_arc_lookup_sym_or_name): ... this. No longer cache
ctf_id_t lookups: just call ctf_lookup_by_symbol as needed (but
still cache the dicts those lookups succeed in). Add
lookup-by-name support, with dicts of successful lookups cached in
ctfi_symnamedicts. Refactor the caching code a bit.
(ctf_arc_lookup_symbol_name): New, another wrapper.
* ctf-open.c (ctf_dict_close): Free the ctf_symhash.
* libctf.ver (LIBCTF_1.2): New version. Add
ctf_lookup_by_symbol_name, ctf_arc_lookup_symbol_name.
* testsuite/libctf-lookup/enum-symbol.c (main): Use
ctf_arc_lookup_symbol rather than looking up the name ourselves.
Fish it out repeatedly, to make sure that symbol caching isn't
broken.
(symidx_64): Remove.
(symidx_32): Remove.
* testsuite/libctf-lookup/enum-symbol-obj.lk: Test symbol lookup
in an unlinked object file (indexed symtypetab sections only).
* testsuite/libctf-writable/symtypetab-nonlinker-writeout.c
(try_maybe_reporting): Check symbol types via
ctf_lookup_by_symbol_name as well as ctf_symbol_next.
* testsuite/libctf-lookup/conflicting-type-syms.*: New test of
lookups in a multi-dict archive.
The variable section in a CTF dict is meant to contain the types of
variables that do not appear in the symbol table (mostly file-scope
static declarations). We implement this by having the compiler emit
all potential data symbols into both sections, then delete those
symbols from the variable section that correspond to data symbols the
linker has reported.
Unfortunately, the check for this in ctf_serialize is wrong: rather than
checking the set of linker-reported symbols, we check the set of names
in the data object symtypetab section: if the linker has reported no
symbols at all (usually if ld -r has been run, or if a non-linker
program that does not use symbol tables is calling ctf_link) this will
include every single symbol, emptying the variable section completely.
Worse, when ld -r is in use, we want to force writeout of every
symtypetab entry on the inputs, in an indexed section, whether or not
the linker has reported them, since this isn't a final link yet and the
symbol table is not finalized (and may grow more symbols than the linker
has yet reported). But the check for this is flawed too: we were
relying on ctf_link_shuffle_syms not having been called if no symbols
exist, but that function is *always* called by ld even when ld -r is in
use: ctf_link_add_linker_symbol is the one that's not called when there
are no symbols.
We clearly need to rethink this. Using the emptiness of the set of
reported symbols as a test for ld -r is just ugly: the linker already
knows if ld -r is underway and can just tell us. So add a new linker
flag CTF_LINK_NO_FILTER_REPORTED_SYMS that is set to stop the linker
filtering the symbols in the symtypetab sections using the set that the
linker has reported: use the presence or absence of this flag to
determine whether to emit unindexed symtabs: we only remove entries from
the variable section when filtering symbols, and we only remove them if
they are in the reported symbol set, fixing the case where no symbols
are reported by the linker at all.
(The negative sense of the new CTF_LINK flag is intentional: the common
case, both for ld and for simple tools that want to do a ctf_link with
no ELF symbol table in sight, is probably to filter out symbols that no
linker has reported: i.e., for the simple tools, all of them.)
There's another wrinkle, though. It is quite possible for a non-linker
to add symbols to a dict via ctf_add_*_sym and then write it out via the
ctf_write APIs: perhaps it's preparing a dict for a later linker
invocation. Right now this would not lead to anything terribly
meaningful happening: ctf_serialize just assumes it was called via
ctf_link if symbols are present. So add an (internal-to-libctf) flag
that indicates that a writeout is happening via ctf_link_write, and set
it there (propagating it to child dicts as needed). ctf_serialize can
then spot when it is not being called by a linker, and arrange to always
write out an indexed, sorted symtypetab for fastest possible future
symbol lookup by name in that case. (The writeouts done by ld -r are
unsorted, because the only thing likely to use those symtabs is the
linker, which doesn't benefit from symtypetab sorting.)
Tests added for all three linking cases (ld -r, ld -shared, ld), with a
bit of testsuite framework enhancement to stop it unconditionally
linking the CTF to be checked by the lookup program with -shared, so
tests can now examine CTF linked with -r or indeed with no flags at all,
though the output filename is still foo.so even in this case.
Another test added for the non-linker case that endeavours to determine
whether the symtypetab is sorted by examining the order of entries
returned from ctf_symbol_next: nobody outside libctf should rely on
this ordering, but this test is not outside libctf :)
include/ChangeLog
2021-01-26 Nick Alcock <nick.alcock@oracle.com>
* ctf-api.h (CTF_LINK_NO_FILTER_REPORTED_SYMS): New.
ld/ChangeLog
2021-01-26 Nick Alcock <nick.alcock@oracle.com>
* ldlang.c (lang_merge_ctf): Set CTF_LINK_NO_FILTER_REPORTED_SYMS
when appropriate.
libctf/ChangeLog
2021-01-27 Nick Alcock <nick.alcock@oracle.com>
* ctf-impl.c (_libctf_nonnull_): Add parameters.
(LCTF_LINKING): New flag.
(ctf_dict_t) <ctf_link_flags>: Mention it.
* ctf-link.c (ctf_link): Keep LCTF_LINKING set across call.
(ctf_write): Likewise, including in child dictionaries.
(ctf_link_shuffle_syms): Make sure ctf_dynsyms is NULL if there
are no reported symbols.
* ctf-create.c (symtypetab_delete_nonstatic_vars): Make sure
the variable has been reported as a symbol by the linker.
(symtypetab_skippable): Mention relationship between SYMFP and the
flags.
(symtypetab_density): Adjust nonnullity. Exit early if no symbols
were reported and force-indexing is off (i.e., we are doing a
final link).
(ctf_serialize): Handle the !LCTF_LINKING case by writing out an
indexed, sorted symtypetab (and allow SYMFP to be NULL in this
case). Turn sorting off if this is a non-final link. Only delete
nonstatic vars if we are filtering symbols and the linker has
reported some.
* testsuite/libctf-regression/nonstatic-var-section-ld-r*:
New test of variable and symtypetab section population when
ld -r is used.
* testsuite/libctf-regression/nonstatic-var-section-ld-executable.lk:
Likewise, when ld of an executable is used.
* testsuite/libctf-regression/nonstatic-var-section-ld.lk:
Likewise, when ld -shared alone is used.
* testsuite/libctf-regression/nonstatic-var-section-ld*.c:
Lookup programs for the above.
* testsuite/libctf-writable/symtypetab-nonlinker-writeout.*: New
test, testing survival of symbols across ctf_write paths.
* testsuite/lib/ctf-lib.exp (run_lookup_test): New option,
nonshared, suppressing linking of the SOURCE with -shared.
When you look up a type by name using ctf_lookup_by_name, in most cases
libctf can just strip off any qualifiers and look for the name, but for
pointer types this doesn't work, since the caller will want the pointer
type itself. But pointer types are nameless, and while they cite the
types they point to, looking up a type by name requires a link going the
*other way*, from the type pointed to to the pointer type that points to
it.
libctf has always built this up at open time: ctf_ptrtab is an array of
type indexes pointing from the index of every type to the index of the
type that points to it. But because it is built up at open time (and
because it uses type indexes and not type IDs) it is restricted to
working within a single dict and ignoring parent/child
relationships. This is normally invisible, unless you manage to get a
dict with a type in the parent but the only pointer to it in a child.
The ctf_ptrtab will not track this relationship, so lookups of this
pointer type by name will fail. Since which type is in the parent and
which in the child is largely opaque to the user (which goes where is up
to the deduplicator, and it can and does reshuffle things to save
space), this leads to a very bad user experience, with an
obviously-visible pointer type which ctf_lookup_by_name claims doesn't
exist.
The fix is to have another array, ctf_pptrtab, which is populated in
child dicts: like the parent's ctf_ptrtab, it has one element per type
in the parent, but is all zeroes except for those types which are
pointed to by types in the child: so it maps parent dict indices to
child dict indices. The array is grown, and new child types scanned,
whenever a lookup happens and new types have been added to the child
since the last time a lookup happened that might need the pptrtab.
(So for non-writable dicts, this only happens once, since new types
cannot be added to non-writable dicts at all.)
Since this introduces new complexity (involving updating only part of
the ctf_pptrtab) which is only seen when a writable dict is in use, we
introduce a new libctf-writable testsuite that contains lookup tests
with no corresponding CTF-containing .c files (which can thus be run
even on platforms with no .ctf-section support in the linker yet), and
add a test to check that creation of pointers in children to types in
parents and a following lookup by name works as expected. The non-
writable case is tested in a new libctf-regression testsuite which is
used to track now-fixed outright bugs in libctf.
libctf/ChangeLog
2021-01-05 Nick Alcock <nick.alcock@oracle.com>
* ctf-impl.h (ctf_dict_t) <ctf_pptrtab>: New.
<ctf_pptrtab_len>: New.
<ctf_pptrtab_typemax>: New.
* ctf-create.c (ctf_serialize): Update accordingly.
(ctf_add_reftype): Note that we don't need to update pptrtab here,
despite updating ptrtab.
* ctf-open.c (ctf_dict_close): Destroy the pptrtab.
(ctf_import): Likewise.
(ctf_import_unref): Likewise.
* ctf-lookup.c (grow_pptrtab): New.
(refresh_pptrtab): New, update a pptrtab.
(ctf_lookup_by_name): Turn into a wrapper around (and rename to)...
(ctf_lookup_by_name_internal): ... this: construct the pptrtab, and
use it in addition to the parent's ptrtab when parent dicts are
searched.
* testsuite/libctf-regression/regression.exp: New testsuite for
regression tests.
* testsuite/libctf-regression/pptrtab*: New test.
* testsuite/libctf-writable/writable.exp: New testsuite for tests of
writable CTF dicts.
* testsuite/libctf-writable/pptrtab*: New test.
Nick Alcock