libctf: symbol type linking support

This adds facilities to write out the function info and data object
sections, which efficiently map from entries in the symbol table to
types.  The write-side code is entirely new: the read-side code was
merely significantly changed and support for indexed tables added
(pointed to by the no-longer-unused cth_objtidxoff and cth_funcidxoff
header fields).

With this in place, you can use ctf_lookup_by_symbol to look up the
types of symbols of function and object type (and, as before, you can
use ctf_lookup_variable to look up types of file-scope variables not
present in the symbol table, as long as you know their name: but
variables that are also data objects are now found in the data object
section instead.)

(Compatible) file format change:

The CTF spec has always said that the function info section looks much
like the CTF_K_FUNCTIONs in the type section: an info word (including an
argument count) followed by a return type and N argument types. This
format is suboptimal: it means function symbols cannot be deduplicated
and it causes a lot of ugly code duplication in libctf.  But
conveniently the compiler has never emitted this!  Because it has always
emitted a rather different format that libctf has never accepted, we can
be sure that there are no instances of this function info section in the
wild, and can freely change its format without compatibility concerns or
a file format version bump.  (And since it has never been emitted in any
code that generated any older file format version, either, we need keep
no code to read the format as specified at all!)

So the function info section is now specified as an array of uint32_t,
exactly like the object data section: each entry is a type ID in the
type section which must be of kind CTF_K_FUNCTION, the prototype of
this function.

This allows function types to be deduplicated and also correctly encodes
the fact that all functions declared in C really are types available to
the program: so they should be stored in the type section like all other
types.  (In format v4, we will be able to represent the types of static
functions as well, but that really does require a file format change.)

We introduce a new header flag, CTF_F_NEWFUNCINFO, which is set if the
new function info format is in use.  A sufficiently new compiler will
always set this flag.  New libctf will always set this flag: old libctf
will refuse to open any CTF dicts that have this flag set.  If the flag
is not set on a dict being read in, new libctf will disregard the
function info section.  Format v4 will remove this flag (or, rather, the
flag has no meaning there and the bit position may be recycled for some
other purpose).

New API:

Symbol addition:
  ctf_add_func_sym: Add a symbol with a given name and type.  The
                    type must be of kind CTF_K_FUNCTION (a function
                    pointer).  Internally this adds a name -> type
                    mapping to the ctf_funchash in the ctf_dict.
  ctf_add_objt_sym: Add a symbol with a given name and type.  The type
                    kind can be anything, including function pointers.
		    This adds to ctf_objthash.

These both treat symbols as name -> type mappings: the linker associates
symbol names with symbol indexes via the ctf_link_shuffle_syms callback,
which sets up the ctf_dynsyms/ctf_dynsymidx/ctf_dynsymmax fields in the
ctf_dict.  Repeated relinks can add more symbols.

Variables that are also exposed as symbols are removed from the variable
section at serialization time.

CTF symbol type sections which have enough pads, defined by
CTF_INDEX_PAD_THRESHOLD (whether because they are in dicts with symbols
where most types are unknown, or in archive where most types are defined
in some child or parent dict, not in this specific dict) are sorted by
name rather than symidx and accompanied by an index which associates
each symbol type entry with a name: the existing ctf_lookup_by_symbol
will map symbol indexes to symbol names and look the names up in the
index automatically.  (This is currently ELF-symbol-table-dependent, but
there is almost nothing specific to ELF in here and we can add support
for other symbol table formats easily).

The compiler also uses index sections to communicate the contents of
object file symbol tables without relying on any specific ordering of
symbols: it doesn't need to sort them, and libctf will detect an
unsorted index section via the absence of the new CTF_F_IDXSORTED header
flag, and sort it if needed.

Iteration:
  ctf_symbol_next: Iterator which returns the types and names of symbols
                   one by one, either for function or data symbols.

This does not require any sorting: the ctf_link machinery uses it to
pull in all the compiler-provided symbols cheaply, but it is not
restricted to that use.

(Compatible) changes in API:
  ctf_lookup_by_symbol: can now be called for object and function
                        symbols: never returns ECTF_NOTDATA (which is
			now not thrown by anything, but is kept for
                        compatibility and because it is a plausible
                        error that we might start throwing again at some
                        later date).

Internally we also have changes to the ctf-string functionality so that
"external" strings (those where we track a string -> offset mapping, but
only write out an offset) can be consulted via the usual means
(ctf_strptr) before the strtab is written out.  This is important
because ctf_link_add_linker_symbol can now be handed symbols named via
strtab offsets, and ctf_link_shuffle_syms must figure out their actual
names by looking in the external symtab we have just been fed by the
ctf_link_add_strtab callback, long before that strtab is written out.

include/ChangeLog
2020-11-20  Nick Alcock  <nick.alcock@oracle.com>

	* ctf-api.h (ctf_symbol_next): New.
	(ctf_add_objt_sym): Likewise.
	(ctf_add_func_sym): Likewise.
	* ctf.h: Document new function info section format.
	(CTF_F_NEWFUNCINFO): New.
	(CTF_F_IDXSORTED): New.
	(CTF_F_MAX): Adjust accordingly.

libctf/ChangeLog
2020-11-20  Nick Alcock  <nick.alcock@oracle.com>

	* ctf-impl.h (CTF_INDEX_PAD_THRESHOLD): New.
	(_libctf_nonnull_): Likewise.
	(ctf_in_flight_dynsym_t): New.
	(ctf_dict_t) <ctf_funcidx_names>: Likewise.
	<ctf_objtidx_names>: Likewise.
	<ctf_nfuncidx>: Likewise.
	<ctf_nobjtidx>: Likewise.
	<ctf_funcidx_sxlate>: Likewise.
	<ctf_objtidx_sxlate>: Likewise.
	<ctf_objthash>: Likewise.
	<ctf_funchash>: Likewise.
	<ctf_dynsyms>: Likewise.
	<ctf_dynsymidx>: Likewise.
	<ctf_dynsymmax>: Likewise.
	<ctf_in_flight_dynsym>: Likewise.
	(struct ctf_next) <u.ctn_next>: Likewise.
	(ctf_symtab_skippable): New prototype.
	(ctf_add_funcobjt_sym): Likewise.
	(ctf_dynhash_sort_by_name): Likewise.
	(ctf_sym_to_elf64): Rename to...
	(ctf_elf32_to_link_sym): ... this, and...
	(ctf_elf64_to_link_sym): ... this.
	* ctf-open.c (init_symtab): Check for lack of CTF_F_NEWFUNCINFO
	flag, and presence of index sections.  Refactor out
	ctf_symtab_skippable and ctf_elf*_to_link_sym, and use them.  Use
	ctf_link_sym_t, not Elf64_Sym.  Skip initializing objt or func
	sxlate sections if corresponding index section is present.  Adjust
	for new func info section format.
	(ctf_bufopen_internal): Add ctf_err_warn to corrupt-file error
	handling.  Report incorrect-length index sections.  Always do an
	init_symtab, even if there is no symtab section (there may be index
	sections still).
	(flip_objts): Adjust comment: func and objt sections are actually
	identical in structure now, no need to caveat.
	(ctf_dict_close):  Free newly-added data structures.
	* ctf-create.c (ctf_create): Initialize them.
	(ctf_symtab_skippable): New, refactored out of
	init_symtab, with st_nameidx_set check added.
	(ctf_add_funcobjt_sym): New, add a function or object symbol to the
	ctf_objthash or ctf_funchash, by name.
	(ctf_add_objt_sym): Call it.
	(ctf_add_func_sym): Likewise.
	(symtypetab_delete_nonstatic_vars): New, delete vars also present as
	data objects.
	(CTF_SYMTYPETAB_EMIT_FUNCTION): New flag to symtypetab emitters:
	this is a function emission, not a data object emission.
	(CTF_SYMTYPETAB_EMIT_PAD): New flag to symtypetab emitters: emit
	pads for symbols with no type (only set for unindexed sections).
	(CTF_SYMTYPETAB_FORCE_INDEXED): New flag to symtypetab emitters:
	always emit indexed.
	(symtypetab_density): New, figure out section sizes.
	(emit_symtypetab): New, emit a symtypetab.
	(emit_symtypetab_index): New, emit a symtypetab index.
	(ctf_serialize): Call them, emitting suitably sorted symtypetab
	sections and indexes.  Set suitable header flags.  Copy over new
	fields.
	* ctf-hash.c (ctf_dynhash_sort_by_name): New, used to impose an
	order on symtypetab index sections.
	* ctf-link.c (ctf_add_type_mapping): Delete erroneous comment
	relating to code that was never committed.
	(ctf_link_one_variable): Improve variable name.
	(check_sym): New, symtypetab analogue of check_variable.
	(ctf_link_deduplicating_one_symtypetab): New.
	(ctf_link_deduplicating_syms): Likewise.
	(ctf_link_deduplicating): Call them.
	(ctf_link_deduplicating_per_cu): Note that we don't call them in
	this case (yet).
	(ctf_link_add_strtab): Set the error on the fp correctly.
	(ctf_link_add_linker_symbol): New (no longer a do-nothing stub), add
	a linker symbol to the in-flight list.
	(ctf_link_shuffle_syms): New (no longer a do-nothing stub), turn the
	in-flight list into a mapping we can use, now its names are
	resolvable in the external strtab.
	* ctf-string.c (ctf_str_rollback_atom): Don't roll back atoms with
	external strtab offsets.
	(ctf_str_rollback): Adjust comment.
	(ctf_str_write_strtab): Migrate ctf_syn_ext_strtab population from
	writeout time...
	(ctf_str_add_external): ... to string addition time.
	* ctf-lookup.c (ctf_lookup_var_key_t): Rename to...
	(ctf_lookup_idx_key_t): ... this, now we use it for syms too.
	<clik_names>: New member, a name table.
	(ctf_lookup_var): Adjust accordingly.
	(ctf_lookup_variable): Likewise.
	(ctf_lookup_by_id): Shuffle further up in the file.
	(ctf_symidx_sort_arg_cb): New, callback for...
	(sort_symidx_by_name): ... this new function to sort a symidx
	found to be unsorted (likely originating from the compiler).
	(ctf_symidx_sort): New, sort a symidx.
	(ctf_lookup_symbol_name): Support dynamic symbols with indexes
	provided by the linker.  Use ctf_link_sym_t, not Elf64_Sym.
	Check the parent if a child lookup fails.
	(ctf_lookup_by_symbol): Likewise.  Work for function symbols too.
	(ctf_symbol_next): New, iterate over symbols with types (without
	sorting).
	(ctf_lookup_idx_name): New, bsearch for symbol names in indexes.
	(ctf_try_lookup_indexed): New, attempt an indexed lookup.
	(ctf_func_info): Reimplement in terms of ctf_lookup_by_symbol.
	(ctf_func_args): Likewise.
	(ctf_get_dict): Move...
	* ctf-types.c (ctf_get_dict): ... here.
	* ctf-util.c (ctf_sym_to_elf64): Re-express as...
	(ctf_elf64_to_link_sym): ... this.  Add new st_symidx field, and
	st_nameidx_set (always 0, so st_nameidx can be ignored).  Look in
	the ELF strtab for names.
	(ctf_elf32_to_link_sym): Likewise, for Elf32_Sym.
	(ctf_next_destroy): Destroy ctf_next_t.u.ctn_next if need be.
	* libctf.ver: Add ctf_symbol_next, ctf_add_objt_sym and
	ctf_add_func_sym.

1 files changed, 25 insertions, 15 deletions

diff --git a/include/ctf.h b/include/ctf.h
index d0a21f1..c7a1e43 100644
--- a/include/ctf.h
+++ b/include/ctf.h
@@ -73,18 +73,20 @@ extern "C"
    the data types section.  Each label is accompanied by a type ID i.  A given
    label refers to the group of types whose IDs are in the range [0, i].
 
-   Data object and function records are stored in the same order as they appear
-   in the corresponding symbol table, except that symbols marked SHN_UNDEF are
-   not stored and symbols that have no type data are padded out with zeroes.
-   For each data object, the type ID (a small integer) is recorded.  For each
-   function, the type ID of the return type and argument types is recorded.
+   Data object and function records (collectively, "symtypetabs") are stored in
+   the same order as they appear in the corresponding symbol table, except that
+   symbols marked SHN_UNDEF are not stored and symbols that have no type data
+   are padded out with zeroes.  For each entry in these tables, the type ID (a
+   small integer) is recorded.  (Functions get CTF_K_FUNCTION types, just like
+   data objects that are function pointers.)
 
    For situations in which the order of the symbols in the symtab is not known,
-   a pair of optional indexes follow the data object and function info sections:
-   each of these is an array of strtab indexes, mapped 1:1 to the corresponding
-   data object / function info section, giving each entry in those sections a
-   name so that the linker can correlate them with final symtab entries and
-   reorder them accordingly (dropping the indexes in the process).
+   or most symbols have no type in this dict and most entries would be
+   zero-pads, a pair of optional indexes follow the data object and function
+   info sections: each of these is an array of strtab indexes, mapped 1:1 to the
+   corresponding data object / function info section, giving each entry in those
+   sections a name so that the linker can correlate them with final symtab
+   entries and reorder them accordingly (dropping the indexes in the process).
 
    Variable records (as distinct from data objects) provide a modicum of support
    for non-ELF systems, mapping a variable name to a CTF type ID.  The variable
@@ -92,7 +94,8 @@ extern "C"
    not define how the consumer maps these variable names to addresses or
    anything else, or indeed what these names represent: they might be names
    looked up at runtime via dlsym() or names extracted at runtime by a debugger
-   or anything else the consumer likes.
+   or anything else the consumer likes.  Variable records with identically-
+   named entries in the data object section are removed.
 
    The data types section is a list of variable size records that represent each
    type, in order by their ID.  The types themselves form a directed graph,
@@ -104,9 +107,9 @@ extern "C"
    Strings are recorded as a string table ID (0 or 1) and a byte offset into the
    string table.  String table 0 is the internal CTF string table.  String table
    1 is the external string table, which is the string table associated with the
-   ELF symbol table for this object.  CTF does not record any strings that are
-   already in the symbol table, and the CTF string table does not contain any
-   duplicated strings.
+   ELF dynamic symbol table for this object.  CTF does not record any strings
+   that are already in the symbol table, and the CTF string table does not
+   contain any duplicated strings.
 
    If the CTF data has been merged with another parent CTF object, some outgoing
    edges may refer to type nodes that exist in another CTF object.  The debugger
@@ -199,9 +202,16 @@ typedef struct ctf_header
 #define CTF_VERSION_3 4
 #define CTF_VERSION CTF_VERSION_3 /* Current version.  */
 
+/* All of these flags bar CTF_F_COMPRESS and CTF_F_IDXSORTED are bug-workaround
+   flags and are valid only in format v3: in v2 and below they cannot occur and
+   in v4 and later, they will be recycled for other purposes.  */
+
 #define CTF_F_COMPRESS	0x1		/* Data buffer is compressed by libctf.  */
+#define CTF_F_NEWFUNCINFO 0x2		/* New v3 func info section format.  */
+#define CTF_F_IDXSORTED 0x4		/* Index sections already sorted.  */
 #define CTF_F_DYNSTR 0x8		/* Strings come from .dynstr.  */
-#define CTF_F_MAX (CTF_F_COMPRESS | CTF_F_DYNSTR)
+#define CTF_F_MAX (CTF_F_COMPRESS | CTF_F_NEWFUNCINFO | CTF_F_IDXSORTED	\
+		   | CTF_F_DYNSTR)
 
 typedef struct ctf_lblent
 {