* lib/Makefile.am (libcoreutils_a_SOURCES): Remove xmemxfrm.c,
xmemxfrm.h.
* lib/memxfrm.c, lib/memxfrm.h, lib/xmemxfrm.c, lib/xmemxfrm.h: Remove.
* m4/memxfrm.m4: Likewise.
* m4/prereq.m4 (gl_PREREQ): Remove gl_MEMXFRM.
* po/POTFILES.in: Remove lib/xmemxfrm.c.
* src/sort.c: Don't include xmemxfrm.h.
(cmp_hashes): Remove.
(xstrxfrm): New function.
(compare_random): If a line contains NULs, don't create a big
buffer that contains the strxfrm output of each string in the line.
Instead, accumulate checksums and differences as we go, so that
at any one time we have to store at most the output of a single
strxfrm call when processing the line. This removes the need for
an memxfrm function.
* src/sort.c (debug_width): New function, which does not stop
counting tabs at \0, and also invokes mbsnwidth. Stamp out strnlen!
(count_tabs): Remove.
(debug_key): Use debug_width instead of mbsnwidth and count_tabs.
* tests/misc/sort-debug-keys: Check that \0 and \t intermix.
* NEWS: Document changes to sort -h, which are now minor with
respect to the pre-July-30th version.
* doc/coreutils.texi (sort invocation): Likewise. The
documentation now describes how -h comparison is done rather than
being vague with border cases.
* src/sort.c (long_double, strtold): Move back to general_numcompare.
(LD, compute_human): Remove.
(find_unit_order): Remove THOU_SEP parameter, since thousands
separators are now allowed by all callers. Revert to previous
behavior of sorting by suffix, and returning the order rather than
2 * order + binary, since we no longer care whether binary powers
are being used. However, treat all zeros the same, instead of
sorting 0M before 0G; this is more consistent with the desired
behavior of sorting -1G before -1M.
* tests/misc/sort (h1, h3, h6): Adjust to match mostly-reverted
behavior. However, check that all zeros sort together.
* tests/misc/sort-debug-keys: Omit a "_", since the trailing "i"
in "1234Gi" is no longer part of the key.
* NEWS: Document changes to sort -h.
* doc/coreutils.texi (sort invocation): Likewise.
* src/sort.c (long_double, strtold): Move to prelude, since they're
now used by multiple functions.
(LD): New macro.
(struct keyfield.iec_present): Remove this member. All uses removed.
(check_mixed_SI_IEC): Remove. This code was busted in the presence
of multiple threads, as it had a race condition.
(find_unit_order): Remove arg KEY; add arg THOU_SEP; arg ENDPTR is
now char ** rather than char const **. Return an integer that
distinguishes decimal from binary powers. Parse the number
consistently with the intersection of strtold and strnumcmp.
Set *ENDPTR unconditionally.
(compute_human): New static function.
(human_numcompare): Remove arg KEY. Remove 'const' from other args.
Use strnumcmp if possible, but fall back on floating point if not.
(numcompare, general_numcompare): Arg EA is now char ** rather
than char const **.
(numcompare): Adjust to new find_unit_order signature and behavior.
(keycompare): Adjus to new human_numcompare signature.
* tests/misc/sort (h1, h3, h4, h6): Adjust to new behavior.
* tests/misc/sort-debug-keys: Likewise.
* src/sort.c (fillbuf): Don't append eol unless the line is nonempty.
This fixes a bug that was partly but not completely fixed by
the aadc67dfdb commit (dated July 15).
* tests/misc/sort (realloc-buf-2): New test, which catches this
bug on 64-bit hosts.
* src/sort.c (mergelines, queue_destroy, queue_init, queue_insert):
(queue_pop, write_unique, mergelines_node, check_insert):
(update_parent): No longer inline; these uses of "inline"
seemed unlikely to help performance much.
* gl/lib/heap.c (struct heap): Move this here...
* gl/lib/heap.h (struct heap): ... from here, as outside code no
longer needs to access any of these members.
* src/sort.c (queue_pop): Omit unnecessary unlock+lock after
pthread_cond_wait returns. Don't access "count" member of the
heap; any efficiency gains should be quite minor, the access
complicates this code, and "count" should be private anyway.
* src/sort.c (lock_node, unlock_node, queue_destroy, queue_init):
(queue_pop):
Omit 'restrict'; it shouldn't help here, as these functions have just
one pointer parameter and don't access static storage.
(queue_insert, check_insert, update_parent): Omit 'restrict', as
the pointer types differ, and are not char * or unsigned char *,
and therefore can't alias.
(write_unique): Omit 'restrict', as the pointer types are all
read-only.
(merge_loop, sortlines): Omit 'restrict', as any performance
advantages are extremely unlikely and it's not worth cluttering
the code for that.
(struct thread_args): Omit 'restrict': this seems to be incorrect.
It's unlikely for 'restrict' to be correct inside a typedef.
* src/sort.c (inittables, general_numcompare, compare_nodes):
(queue_init, queue_pop): Omit casts that are not needed, typically
because they are between void * and some other pointer type.
* src/du.c (process_file): Avoid recalculation of hashes
and of file-exclusion for directories. Do not descend into
the same directory more than once, unless -l is given; this is faster.
Calculate stat buffer lazily, since it
need not be computed at all for excluded files.
Count space if FTS_ERR, since stat buffer is always valid then.
No need for 'print' local variable.
(main): Use FTS_NOSTAT. Use FTS_TIGHT_CYCLE_CHECK only when not
hashing everything, since process_file finds cycles on its own
when hashing everything.
* tests/du/deref: Add test cases for -L bugs.
* tests/misc/sort-merge-fdlimit: This test was written assuming that
-R typically opens /dev/urandom, but that's no longer the case.
Redo test to specify a random source; this resurrects the point of
checking for file descriptor exhaustion. Also try plain -R, since
that implementation may change in the future too.
* gl/lib/rand-isaac.c: Remove the I/O; this belongs elsewhere.
Add support for ISAAC64. Port to hosts with padding bits.
Add self to author list. Include <limits.h>, for CHAR_BIT.
Don't include string.h, sys/time.h, unistd.h.
(min, just): New functions.
(IF32): New macros.
(ind, ISAAC_STEP, isaac_refill, mix, isaac_init, isaac_seed):
Add support for ISAAC64. Port to hosts with padding bits.
(ind): Now an inline function rather than a macro; no need for it
to be a macro with modern compilers.
(ISAAC_STEP): Renamed from isaac_step, since it's not function-like.
Don't bother to pass args that are always the same. All uses changed.
(ISAAC_STEP, ISAAC_SEED): Move to inside the only function body
that can use it.
(ISAAC_MIX): Renamed from isaac_mix, since it's now a macro and is
no longer function-like. Don't bother saving and restoring state;
no longer needed now that we're not a function. All uses changed.
(isaac_seed_start, isaac_seed_data, isaac_seed_finish): Remove.
(isaac_seed): Take just the one arg; the caller now sets s->m.
* gl/lib/rand-isaac.h: Use _GL_RAND_ISAAC_H to protect, instead
of RAND_ISAAC_H. Try out " #" rather than "# " for indenting.
(ISAAC_BITS_LOG, ISAAC_BITS): New macros.
(ISAAC_WORDS_LOG): Renamed from ISAAC_LOG.
(isaac_word): New type. All uses of uint32_t changed to isaac_word,
to support ISAAC64.
(struct isaac_state): Rename member MM to M, and make it public.
(isaac_seed, isaac_refill): Adjust to new API.
* gl/lib/randread.c: Include sys/time.h.
(get_nonce): New function, containing the nonce stuff that used
to be in rand-isaac.c but better belongs here.
(randread_new): Use it.
* gl/modules/randread (Depends-on): Add inline.
* gl/modules/randread-tests: New file.
* gl/tests/test-rand-isaac.c: New file.
* bootstrap.conf: Include the new module
* gl/lib/fadvise.c: Provide a simpler interface to posix_fadvise.
(fadvise): Provide hint to the whole file associated with a stream.
(fdadvise): Provide hint to the specific portion of a file
associated with a file descriptor.
* gl/lib/fadvise.h: Redefine POSIX_FADV_* to FADVISE_* enums.
* gl/modules/fadvise: New file.
* m4/jm-macros.m4: Remove the no longer needed posix_fadvise check.
* .x-sc_program_name: Exclude test-fadvise.c from this check.
* gl/tests/test-fadvise (main): New test program.
* gl/modules/fadvise-testss: A new index to reference the tests.
* src/sort.c (stream_open): Use the new interface.
* src/dd.c (iwrite): Likewise.
* configure.ac (optional_pkglib_progs): Only update
after the main programs have been selected, so that
libstdbuf.so can be excluded if stdbuf also is.
* gl/lib/rand-isaac.c (isaac_seed_start): New arg SEEDED.
(isaac_seed): New args FD and BYTES_BOUND. Read from FD if possible.
Don't bother with low-quality sources if FD has enough bytes.
* gl/lib/rand-isaac.h: New size_t arg for isaac_seed.
* gl/lib/randread.c: Include fcntl.h, unistd.h.
(NAME_OF_NONCE_DEVICE): New #define.
(nonce_device): New static var.
(randread_new): Use nonce device if available.
* src/sort.c (random_md5_state): New static var.
(random_md5_state_init): New function, to initialize random_md5_state.
(random_state, randread_source): Remove.
(cmp_hashes): Use random_md5_state rather than random_state.
Break ties using memcmp, not by getting more randomness.
If MD5 collisions turn into a problem in practice, we should
simply use a better checksum.
(main): If -R is given, call random_md5_state_init rather than
going single-threaded.
Programs like 'sort' were linking to -lrt in order to get
clock_gettime, but this was misguided: it wasted considerable
resources while gaining at most 10 bits of entropy. Almost nobody
needs the entropy, and there are better ways to get much better
entropy for people who do need it.
* gl/lib/rand-isaac.c (isaac_seed): Include <sys/time.h> not
"gethrxtime.h".
(isaac_seed): Use gettimeofday rather than gethrxtime.
* gl/modules/randread (Depends-on): Depend on gettimeofday
and not gethrxtime.
* src/Makefile.am (mktemp_LDADD, shred_LDADD, shuf_LDADD, sort_LDADD):
(tac_LDADD): Omit $(LIB_GETHRXTIME); no longer needed.
* src/sort.c (keycompare): Use xmemcoll0, as it avoids
a couple of stores.
(write_bytes): Leave the buffer the way we found it,
as it might be used again for a later comparison,
if -u is used.
Don't write NUL after the comparison buffers on each compare,
which increases performance by about 3% for short lines
on a pentium-m with gcc-4.4.1
* src/sort.c: (fillbuf): Delimit input items with NUL.
(write_bytes): Restore the item delimiter char which was
replaced with NUL in fillbuf().
* src/Makefile.am (printf_LDADD, seq_LDADD, sleep_LDADD, sort_LDADD):
(tail_LDADD, uptime_LDADD): Omit $(POW_LIB), as it's no longer
needed due to recent gnulib changes, where the strtod module no
longer uses the pow function. strtold needs pow only because it's
sometimes aliased to strtod. See
http://lists.gnu.org/archive/html/bug-gnulib/2010-07/msg00076.html
* NEWS: Add another blank line before the previous version.
(Bug fixes): Move to the start.
(Changes in behavior): Add the item about the du mem usage change
from the "New features" section.
* gl/lib/heap.c (heap_alloc): Use the fact that the xalloc
routines will not return NULL. Also remove the redundant
temporary variables.
(heap_insert): From Jim Meyering, use x2nrealloc() which
is simpler while handling overflow and increasing the
size more efficiently. This reallocation is currently
unused by coreutils in any case as it preallocates enough.
This patch is by Gene Auyeung, Chris Dickens, Chen Guo, and Mike
Nichols, based off of a patch by Paul Eggert, Glen Lenker, et. al.,
with a basic heap implementation based off of the GDSL heap,
originally by Nicolas Darnis.
The number of sorts done in parallel is limited to the number
of available processors by default, or can be further restricted
with the --parallel option.
On a dual-die, 8 core Intel Xeon, results show sorting with
8 threads is almost 4 times faster than using a single thread.
Timings when sorting a 96MB file:
THREADS TIME (s)
1 5.10
2 2.87
4 1.75
8 1.31
Single threaded sorting has also been improved,
especially for cheaper comparison operations:
COMMAND BEFORE (s) AFTER (s)
sort 8.822 8.716
sort -g 10.336 10.222
sort -n 3.077 2.961
LANG=C sort 2.169 2.066
* bootstrap.conf: Add heap, pthread.
* coreutils.texi (sort): Describe the new --parallel option.
* gl/lib/heap.c: New file. Very basic heap implementation.
* gl/lib/heap.h: New file.
* gl/modules/heap: New file.
* src/Makefile.am: Add LIB_PTHREAD.
* src/sort.c: Include heap.h, nproc.h, pthread.h.
(MAX_MERGE): New macro.
(SUBTHREAD_LINES_HEURISTIC, PARALLEL_OPTION): New constants.
(MERGE_END, MERGE_ROOT): New constants.
(struct merge_node): New struct.
(struct merge_node_queue): New struct.
(sortlines temp): Remove declaration.
(usage, long_options, main): New option, --parallel.
(specify_nthreads): New function.
(mergelines): New signature, to emphasize the fact that the HI area
must be part of the destination. All callers changed.
(sequential_sort): New function, renamed from sortlines. Merge in
the functionality of sortlines_temp.
(compare_nodes): New function.
(lock_node, unlock_node): New functions.
(queue_destroy): New function.
(queue_init): New function.
(queue_insert): New function.
(queue_pop): New function.
(write_unique): New function.
(mergelines_node): New function.
(check_insert): New function.
(update_parent): New function.
(merge_loop): New function.
(sortlines): Rewrite to support and use parallelism, with a new
signature. All callers changed.
(struct thread_args): New struct.
(sortlines_thread): New function.
(sortlines_temp): Remove.
(sort): New argument NTHREADS. All uses changed. Output moved to
mergelines_node.
(main): disable threading if we are sorting at random.
* tests/Makefile.am (TESTS): Add misc/sort-benchmark-random.
* tests/misc/sort-benchmark-random: New file.
Signed-off-by: Pádraig Brady <P@draigBrady.com>
* gl/lib/ino-map.c (ino_hash): Declare "i" as unsigned int.
Use an intermediate variable for the for-loop upper bound,
so it's a little more readable. Adjust comment.
* gl/lib/di-set.c (di_ent_hash): Likewise.
* src/dd.c (dd_copy): Use requested blocksize (not adjusted) in
diagnostic, to forestall user complaints that the numbers don't
match exactly. Report both exact and human-readable sizes, using
a message format that is consistent with both "BBBB bytes (N XB)
copied" in dd.c and "memory exhausted" in lib/xmalloc.c.
Problem reported by Jim Meyering in:
http://debbugs.gnu.org/cgi/bugreport.cgi?bug=6524#74
* gl/lib/di-set.c (di_ent_hash): Rework so that the compiler does
not incorrectly warn about shifting by 64-bits in unreachable code.
* gl/lib/ino-map.c (ino_hash): Likewise.
* gl/lib/dev-map.c, gl/lib/dev-map.h, gl/modules/dev-map: Remove.
* gl/lib/ino-map.c, gl/lib/ino-map.h, gl/modules/ino-map: New files.
* gl/modules/dev-map-tests, gl/tests/test-dev-map.c: Remove.
* gl/modules/ino-map-tests, gl/tests/test-ino-map.c: New files.
* gl/lib/di-set.h (struct di_set): Renamed from struct di_set_state,
and now private. All uses changed.
(_ATTRIBUTE_NONNULL_): Don't assume C99.
(di_set_alloc): Renamed from di_set_init, with no size arg.
Now allocates the object rather than initializing it.
For now, this no longer takes an initial size; we can put this
back later if it is needed.
* gl/lib/di-set.c: Include hash.h, ino-map.h, and limits.h instead of
stdio.h, assert.h, stdint.h, sys/types.h (di-set.h includes that
now), sys/stat.h, and verify.h.
(N_DEV_BITS_4, N_INO_BITS_4, N_DEV_BITS_8, N_INO_BITS_8): Remove.
(struct dev_ino_4, struct dev_ino_8, struct dev_ino_full): Remove.
(enum di_mode): Remove.
(hashint): New typedef.
(HASHINT_MAX, LARGE_INO_MIN): New macros.
(struct di_ent): Now maps a dev_t to a inode set, instead of
containing a union.
(struct dev_map_ent): Remove.
(struct di_set): New type.
(is_encoded_ptr, decode_ptr, di_ent_create): Remove.
(di_ent_hash, di_ent_compare, di_ent_free, di_set_alloc, di_set_free):
(di_set_insert): Adjust to new representation.
(di_ino_hash, map_device, map_inode_number): New functions.
* gl/modules/di-set (Depends-on): Replace dev-map with ino-map.
Remove 'verify'.
* gl/tests/test-di-set.c: Adjust to the above changes to API.
* src/du.c (INITIAL_DI_SET_SIZE): Remove.
(hash_ins, main): Adjust to new di-set API.
Add comments and adjust interfaces to allow low-level failure
to propagate out to callers.
* src/stat.c (out_file_context): Return bool, not void,
so we can tell callers about failure.
(print_statfs, print_stat, print_it): Propagate failure to caller.
(do_statfs): Propagate print_it failure to caller.
(do_stat): Likewise.
I nearly forgot to update do_stat to propagate print_it failure,
and it compiled just fine in spite of that. To prevent possibility
of a repeat, I've marked each function that returns non-void with
ATTRIBUTE_WARN_UNUSED_RESULT.
* tests/init.cfg: Introduce a retry_delay_() function to
repeatedly call a test function that requires a delay.
This delay can now be shorter for the common case on fast
systems, but will double until a configurable limit it reached
before failing on slower systems.
* tests/dd/reblock: Use retry_delay_.
* tests/misc/cat-buf: Likewise.
* tests/misc/stdbuf: Likewise.
* tests/tail-2/F-vs-rename: Likewise.
* tests/tail-2/flush-initial: Likewise.
* tests/tail-2/tail-n0f: Likewise.
* tests/tail-2/wait: Likewise.
* test/dd/misc: Comment that delay is needed to trigger failure.
* src/du.c (INITIAL_DI_SET_SIZE): Increase to the prime just under
1024. This gives a speed-up of about 2% when processing a tree
containing 100,000 files, each with a link count greater than 1,
all pointing to files in some other tree.
When processing a hard-linked file, du must keep track of the file's
device and inode numbers in order to avoid counting its storage
more than once. When du would process many hard linked files --
as are created by some backup tools -- the amount of memory required
for the supporting data structure could become prohibitively large.
This patch takes advantage of the fact that the amount of information
in the numbers of the typical dev,inode pair is far less than even
32 bits, and hence usually fits in the space of a pointer, be it
32 or 64 bits wide. A typical du traversal examines files on no
more than a handful of distinct devices, so the device number can
be encoded in just a few bits. Similarly, few inode numbers use
all of the high bits in an ino_t. Before, we would represent the
dev,inode pair using a naive struct, and allocate space for each.
Thus, an entry in the hash table consisted of a pointer (to that
struct) and a "next" pointer. With this change, we encode the
dev,inode information and put those bits in place of the pointer,
and thus do away with the need to allocate additional space for
each dev,inode pair.
* src/du.c: Include "di-set.h".
Don't include "hash.h"; it's no longer used.
(INITIAL_DI_SET_SIZE): Define.
(di_set): New global, to replace "htab".
(entry_hash, entry_compare, hash_init): Remove functions.
(hash_ins): Use di-set functions, rather than ones from the hash module.
(main): Likewise.
* bootstrap.conf (gnulib_modules): Add the new di-set module.
* NEWS (New features): Mention it.
* NEWS: Mention this.
* src/du.c (hash_all): New static var.
(process_file): Use it.
(main): Set it.
* tests/du/hard-link: Add a couple of test cases to help make
sure this bug stays squashed.
* tests/du/files0-from: Adjust existing tests to reflect
change in semantics with duplicate arguments.
* src/copy.c (copy_attr): A new function which merges copy_attr_by_fd
and copy_attr_by_name. Also display all errors when --attributes-only
* src/copy.c (copy_reg): Skip copying the file contents if specified.
Refactor the SELinux error handling code a little and display all
SELinux errors when only copying attributes.
* src/copy.h (struct cp_options): Add a data_copy_required boolean
* src/cp.c (main): Default to copying data but don't if specified
* src/install.c: Default to copying data
* src/mv.c: Likewise
tests/cp/reflink-perm: Add a test to check that --attributes-only
does not copy data
* tests/cp/acl: Likewise. Also refactor to remove redundant
acl manipulation
* doc/coreutils.texi (cp invocation): Describe the new option
* NEWS: Mention the new feature
Paul Eggert
Pádraig Brady
Mike Frysinger
Chen Guo
Jim Meyering