The CTF file format has always supported "external strtabs", which
internally are strtab offsets with their MSB on: such refs
get their strings from the strtab passed in at CTF file open time:
this is usually intended to be the ELF strtab, and that's what this
implementation is meant to support, though in theory the external
strtab could come from anywhere.
This commit adds support for these external strings in the ctf-string.c
strtab tracking layer. It's quite easy: we just add a field csa_offset
to the atoms table that tracks all strings: this field tracks the offset
of the string in the ELF strtab (with its MSB already on, courtesy of a
new macro CTF_SET_STID), and adds a new function that sets the
csa_offset to the specified offset (plus MSB). Then we just need to
avoid writing out strings to the internal strtab if they have csa_offset
set, and note that the internal strtab is shorter than it might
otherwise be.
(We could in theory save a little more time here by eschewing sorting
such strings, since we never actually write the strings out anywhere,
but that would mean storing them separately and it's just not worth the
complexity cost until profiling shows it's worth doing.)
We also have to go through a bit of extra effort at variable-sorting
time. This was previously using direct references to the internal
strtab: it couldn't use ctf_strptr or ctf_strraw because the new strtab
is not yet ready to put in its usual field (in a ctf_file_t that hasn't
even been allocated yet at this stage): but now we're using the external
strtab, this will no longer do because it'll be looking things up in the
wrong strtab, with disastrous results. Instead, pass the new internal
strtab in to a new ctf_strraw_explicit function which is just like
ctf_strraw except you can specify a ne winternal strtab to use.
But even now that it is using a new internal strtab, this is not quite
enough: it can't look up strings in the external strtab because ld
hasn't written it out yet, and when it does will write it straight to
disk. Instead, when we write the internal strtab, note all the offset
-> string mappings that we have noted belong in the *external* strtab to
a new "synthetic external strtab" dynhash, ctf_syn_ext_strtab, and look
in there at ctf_strraw time if it is set. This uses minimal extra
memory (because only strings in the external strtab that we actually use
are stored, and even those come straight out of the atoms table), but
let both variable sorting and name interning when ctf_bufopen is next
called work fine. (This also means that we don't need to filter out
spurious ECTF_STRTAB warnings from ctf_bufopen but can pass them back to
the caller, once we wrap ctf_bufopen so that we have a new internal
variant of ctf_bufopen etc that we can pass the synthetic external
strtab to. That error has been filtered out since the days of Solaris
libctf, which didn't try to handle the problem of getting external
strtabs right at construction time at all.)
v3: add the synthetic strtab and all associated machinery.
v5: fix tabdamage.
include/
* ctf.h (CTF_SET_STID): New.
libctf/
* ctf-impl.h (ctf_str_atom_t) <csa_offset>: New field.
(ctf_file_t) <ctf_syn_ext_strtab>: Likewise.
(ctf_str_add_ref): Name the last arg.
(ctf_str_add_external) New.
(ctf_str_add_strraw_explicit): Likewise.
(ctf_simple_open_internal): Likewise.
(ctf_bufopen_internal): Likewise.
* ctf-string.c (ctf_strraw_explicit): Split from...
(ctf_strraw): ... here, with new support for ctf_syn_ext_strtab.
(ctf_str_add_ref_internal): Return the atom, not the
string.
(ctf_str_add): Adjust accordingly.
(ctf_str_add_ref): Likewise. Move up in the file.
(ctf_str_add_external): New: update the csa_offset.
(ctf_str_count_strtab): Only account for strings with no csa_offset
in the internal strtab length.
(ctf_str_write_strtab): If the csa_offset is set, update the
string's refs without writing the string out, and update the
ctf_syn_ext_strtab. Make OOM handling less ugly.
* ctf-create.c (struct ctf_sort_var_arg_cb): New.
(ctf_update): Handle failure to populate the strtab. Pass in the
new ctf_sort_var arg. Adjust for ctf_syn_ext_strtab addition.
Call ctf_simple_open_internal, not ctf_simple_open.
(ctf_sort_var): Call ctf_strraw_explicit rather than looking up
strings by hand.
* ctf-hash.c (ctf_hash_insert_type): Likewise (but using
ctf_strraw). Adjust to diagnose ECTF_STRTAB nonetheless.
* ctf-open.c (init_types): No longer filter out ECTF_STRTAB.
(ctf_file_close): Destroy the ctf_syn_ext_strtab.
(ctf_simple_open): Rename to, and reimplement as a wrapper around...
(ctf_simple_open_internal): ... this new function, which calls
ctf_bufopen_internal.
(ctf_bufopen): Rename to, and reimplement as a wrapper around...
(ctf_bufopen_internal): ... this new function, which sets
ctf_syn_ext_strtab.
No code handles these yet, but our latest GCC patches are generating
them, so we have to be ready for them or erroneously conclude that we
have file corruption.
(This simultaneously fixes a longstanding bug, concealed because nothing
was generating anything in the object or function info sections, where
the end of the section was being tested against the wrong thing: it
would have walked over the entire contents of the variable section and
treated them as part of the function info section. This had to change
now anyway because the new sections have landed in between.)
include/
* ctf.h: Add object index and function index sections. Describe
them. Improve the description of the variable section and clarify
the constraints on backward-pointing type nodes.
(ctf_header): Add cth_objtidxoff, cth_funcidxoff.
libctf/
* ctf-open.c (init_symtab): Check for overflow against the right
section.
(upgrade_header): Set cth_objtidxoff, cth_funcidxoff to zero-length.
(upgrade_types_v1): Note that these sections are not checked.
(flip_header): Endian-swap the header fields.
(flip_ctf): Endian-swap the sections.
(flip_objts): Update comment.
(ctf_bufopen): Check header offsets and alignment for validity.
libctf supports dynamic upgrading of the type table as file format
versions change, but before now has not supported changes to the CTF
header. Doing this is complicated by the baroque storage method used:
the CTF header is kept prepended to the rest of the CTF data, just as
when read from the file, and written out from there, and is
endian-flipped in place.
This makes accessing it needlessly hard and makes it almost impossible
to make the header larger if we add fields. The general storage
machinery around the malloced ctf pointer (the 'ctf_base') is also
overcomplicated: the pointer is sometimes malloced locally and sometimes
assigned from a parameter, so freeing it requires checking to see if
that parameter was used, needlessly coupling ctf_bufopen and
ctf_file_close together.
So split the header out into a new ctf_file_t.ctf_header, which is
written out explicitly: squeeze it out of the CTF buffer whenever we
reallocate it, and use ctf_file_t.ctf_buf to skip past the header when
we do not need to reallocate (when no upgrading or endian-flipping is
required). We now track whether the CTF base can be freed explicitly
via a new ctf_dynbase pointer which is non-NULL only when freeing is
possible.
With all this done, we can upgrade the header on the fly and add new
fields as desired, via a new upgrade_header function in ctf-open.
As with other forms of upgrading, libctf upgrades older headers
automatically to the latest supported version at open time.
For a first use of this field, we add a new string field cth_cuname, and
a corresponding setter/getter pair ctf_cuname_set and ctf_cuname: this
is used by debuggers to determine whether a CTF section's types relate
to a single compilation unit, or to all compilation units in the
program. (Types with ambiguous definitions in different CUs have only
one of these types placed in the top-level shared .ctf container: the
rest are placed in much smaller per-CU containers, which have the shared
container as their parent. Since CTF must be useful in the absence of
DWARF, we store the names of the relevant CUs ourselves, so the debugger
can look them up.)
v5: fix tabdamage.
include/
* ctf-api.h (ctf_cuname): New function.
(ctf_cuname_set): Likewise.
* ctf.h: Improve comment around upgrading, no longer
implying that v2 is the target of upgrades (it is v3 now).
(ctf_header_v2_t): New, old-format header for backward
compatibility.
(ctf_header_t): Add cth_cuname: this is the first of several
header changes in format v3.
libctf/
* ctf-impl.h (ctf_file_t): New fields ctf_header, ctf_dynbase,
ctf_cuname, ctf_dyncuname: ctf_base and ctf_buf are no longer const.
* ctf-open.c (ctf_set_base): Preserve the gap between ctf_buf and
ctf_base: do not assume that it is always sizeof (ctf_header_t).
Print out ctf_cuname: only print out ctf_parname if set.
(ctf_free_base): Removed, ctf_base is no longer freed: free
ctf_dynbase instead.
(ctf_set_version): Fix spacing.
(upgrade_header): New, in-place header upgrading.
(upgrade_types): Rename to...
(upgrade_types_v1): ... this. Free ctf_dynbase, not ctf_base. No
longer track old and new headers separately. No longer allow for
header sizes explicitly: squeeze the headers out on upgrade (they
are preserved in fp->ctf_header). Set ctf_dynbase, ctf_base and
ctf_buf explicitly. Use ctf_free, not ctf_free_base.
(upgrade_types): New, also handle ctf_parmax updating.
(flip_header): Flip ctf_cuname.
(flip_types): Flip BUF explicitly rather than deriving BUF from
BASE.
(ctf_bufopen): Store the header in fp->ctf_header. Correct minimum
required alignment of objtoff and funcoff. No longer store it in
the ctf_buf unless that buf is derived unmodified from the input.
Set ctf_dynbase where ctf_base is dynamically allocated. Drop locals
that duplicate fields in ctf_file: move allocation of ctf_file
further up instead. Call upgrade_header as needed. Move
version-specific ctf_parmax initialization into upgrade_types. More
concise error handling.
(ctf_file_close): No longer test for null pointers before freeing.
Free ctf_dyncuname, ctf_dynbase, and ctf_header. Do not call
ctf_free_base.
(ctf_cuname): New.
(ctf_cuname_set): New.
* ctf-create.c (ctf_update): Populate ctf_cuname.
(ctf_gzwrite): Write out the header explicitly. Remove obsolescent
comment.
(ctf_write): Likewise.
(ctf_compress_write): Get the header from ctf_header, not ctf_base.
Fix the compression length: fp->ctf_size never counted the CTF
header. Simplify the compress call accordingly.
Testing of the first code to generate CTF_K_SLICEs on big-endian
revealed a bunch of new problems in this area. Most importantly, the
trick we did earlier to avoid wasting two bytes on padding in the
ctf_slice_t is best avoided: because it leads to the whole file after
that point no longer being naturally aligned, all multibyte accesses
from then on must use memmove() to avoid unaligned access on platforms
where that is fatal. In future, this is planned, but for now we are
still doing direct access in many places, so we must revert to making
ctf_slice_t properly aligned for storage in an array.
Rather than wasting bytes on padding, we boost the size of cts_offset
and cts_bits. This is still a waste of space (we cannot have offsets or
bits in bitfields > 256) but it cannot be avoided for now, and slices
are not so common that this will be a serious problem.
A possibly-worse endianness problem fixed at the same time involves
a codepath used only for foreign-endian, uncompressed CTF files, where
we were not copying the actual CTF data into the buffer, leading to
libctf reading only zeroes (or, possibly, uninitialized garbage).
Finally, when we read in a CTF file, we copy the header and work from
the copy. We were flipping the endianness of the header copy, and of
the body of the file buffer, but not of the header in the file buffer
itself: so if we write the file back out again we end up with an
unreadable frankenfile with header and body of different endiannesses.
Fix by flipping both copies of the header.
include/
* ctf.h (ctf_slice_t): Make cts_offset and cts_bits unsigned
short, so following structures are properly aligned.
libctf/
* ctf-open.c (get_vbytes_common): Return the new slice size.
(ctf_bufopen): Flip the endianness of the CTF-section header copy.
Remember to copy in the CTF data when opening an uncompressed
foreign-endian CTF file. Prune useless variable manipulation.
This stops the file format from depending on the size of the host int.
(It does mean that we cannot encode enums with a value > 2^32 on
platforms with an int > 2^32: this will be fixed in the next format
revision.)
include/
* ctf.h (ctf_enum.cte_value): Fix type to int32_t.
If you need to store a large number of CTF containers somewhere, this
provides a dedicated facility for doing so: an mmappable archive format
like a very simple tar or ar without all the system-dependent format
horrors or need for heavy file copying, with built-in compression of
files above a particular size threshold.
libctf automatically mmap()s uncompressed elements of these archives, or
uncompresses them, as needed. (If the platform does not support mmap(),
copying into dynamically-allocated buffers is used.)
Archive iteration operations are partitioned into raw and non-raw
forms. Raw operations pass thhe raw archive contents to the callback:
non-raw forms open each member with ctf_bufopen() and pass the resulting
ctf_file_t to the iterator instead. This lets you manipulate the raw
data in the archive, or the contents interpreted as a CTF file, as
needed.
It is not yet known whether we will store CTF archives in a linked ELF
object in one of these (akin to debugdata) or whether they'll get one
section per TU plus one parent container for types shared between them.
(In the case of ELF objects with very large numbers of TUs, an archive
of all of them would seem preferable, so we might just use an archive,
and add lzma support so you can assume that .gnu_debugdata and .ctf are
compressed using the same algorithm if both are present.)
To make usage easier, the ctf_archive_t is not the on-disk
representation but an abstraction over both ctf_file_t's and archives of
many ctf_file_t's: users see both CTF archives and raw CTF files as
ctf_archive_t's upon opening, the only difference being that a raw CTF
file has only a single "archive member", named ".ctf" (the default if a
null pointer is passed in as the name). The next commit will make use
of this facility, in addition to providing the public interface to
actually open archives. (In the future, it should be possible to have
all CTF sections in an ELF file appear as an "archive" in the same
fashion.)
This machinery is also used to allow library-internal creators of
ctf_archive_t's (such as the next commit) to stash away an ELF string
and symbol table, so that all opens of members in a given archive will
use them. This lets CTF archives exploit the ELF string and symbol
table just like raw CTF files can.
(All this leads to somewhat confusing type naming. The ctf_archive_t is
a typedef for the opaque internal type, struct ctf_archive_internal: the
non-internal "struct ctf_archive" is the on-disk structure meant for
other libraries manipulating CTF files. It is probably clearest to use
the struct name for struct ctf_archive_internal inside the program, and
the typedef names outside.)
libctf/
* ctf-archive.c: New.
* ctf-impl.h (ctf_archive_internal): New type.
(ctf_arc_open_internal): New declaration.
(ctf_arc_bufopen): Likewise.
(ctf_arc_close_internal): Likewise.
include/
* ctf.h (CTFA_MAGIC): New.
(struct ctf_archive): New.
(struct ctf_archive_modent): Likewise.
* ctf-api.h (ctf_archive_member_f): New.
(ctf_archive_raw_member_f): Likewise.
(ctf_arc_write): Likewise.
(ctf_arc_close): Likewise.
(ctf_arc_open_by_name): Likewise.
(ctf_archive_iter): Likewise.
(ctf_archive_raw_iter): Likewise.
(ctf_get_arc): Likewise.
The data structures and macros in this header can be used, if desired,
to access or create CTF files directly, without going through libctf,
though this should rarely be necessary in practice.
libctf relies on this header as its description of the CTF file format.
include/
* ctf.h: New file.
Alan Modra
Nick Alcock