(xfopen): Use stdout if *how != 'r'.
(mergefps): Remove FPS arg.
Open all input files, and close all files when done.
If OFP is null, open the output file (but after opening input files).
All callers changed.
(first_same_file): New function.
(sort, merge): Remove arg OFP; we now open the output file as needed.
All callers changed.
(merge): New arg MAX_MERGE. All callers changed.
(sort): For "sort F -o F", close the input before opening the output.
(main): Do not use close_stdout; 'sort' and 'merge' now close stdout.
(This also fixes a close-stdout-twice bug.)
Remove test for overlapping input and output files, as 'sort' no longer
needs to worry about overlap, and 'merge' checks for overlap itself.
Use first_same_file to inform 'merge' about how much to merge at
the top level, to avoid overlap.
(create_temp_file, xfopen, xfclose, write_bytes, sort_buffer_size,
fillbuf, main): Use it to regularize error messages. The only change
in behavior is that write_bytes and the final close used to say "write
error" but now give just the output file name, which should be enough.
Use fopen_safer, not fopen, to avoid subtle bugs when fopen returns
stdin, stdout, or stderr.
(xfclose): stdout is no longer a special case.
(main): Close output file, don't just flush it; there might be
an error on the close.
representing temp files. This improved CPU performance of
'sort -S 1 *.[ch]' by 17% on my host.
(struct tempnode): name member now uses struct hack.
(temphead): Now a pointer, not a structure. All uses changed.
(create_temp_file): Allocate node using struct hack.
(zaptemp): Free node using struct hack. Use pointer comparison, not
string comparison.
Defend against a DoS attack where someone else creates a
temporary file with the same name as ours. Use mkstemp to do
this, supplying our own mkstemp if the system doesn't have one.
Also, fix a race condition during cleanup on hosts without
sigaction.
(NAME_MAX_IN_DIR): Remove.
(sigprocmask, sigset_t): New macros, defined only on older hosts.
(caught_signals): New var.
(xtmpfopen, tempname): Removed.
(create_temp_file): New function, combining the functions of the old
xtmpfopen and tempname. All callers changed.
Use mkstemp to create the file.
(sighandler): On hosts without sigaction, ignore signals while
cleaning up, instead of letting them interrupt cleanup.
(main): Initialize caught_signals. On hosts with sigaction, block all
caught signals while handling one. Remove duplicate code.
Now it prints a trailing `@'.
(gobble_file): Don't clobber lstat stats of command line
arguments when using -d and -F (--directory and --classify) options.
(struct tempnode.next): Now volatile.
(zaptemp): Free the file name after removing it from the temp list,
not before, because a signal can arrive between the two actions
and cleanup () traverses the list.
Also check for memory quotas.
Revamp storage management so that line tables and character data are
taken from the same buffer. Line tables are now in reverse order,
since they grow down while the character data grow up.
(<sys/resource.h>): Include if HAVE_SYS_RESOURCE_H.
(struct rlimit, getrlimit): Define a replacement if RLIMIT_DATA
is not defined.
(RLIMIT_AS): Define to RLIMIT_DATA if not defined.
(struct lines): Remove.
(struct buffer): New members nlines, line_bytes, eof.
Remove member newline_free; no longer needed, since the code no longer
runs out of line table space.
(SORTALLOC_MIN, SORTALLOC_DEFAULT_MIN): Remove.
(sort_size): Renamed from sortalloc; now applies to the sum of the
character data and the line table, not just the character data.
(MIN_SORT_SIZE, INPUT_FILE_SIZE_GUESS): New macros.
(linelength): remove.
(specify_sort_size): Don't worry about the distinction between the
character data and the line table; that is now the caller's
responsibility.
(default_sort_size): Return the value, instead of being executed for
side effect. Return half of available memory, or 1/16 of total memory,
whichever is greater; except do not exceed 1/2 of quota.
(sort_buffer_size): New function.
(initbuf): New arg LINE_BYTES. Ensure that the line array is properly
aligned. Initialize the new set of struct buffer members.
(buffer_linelim): New function.
(fillbuf): Return int, not size_t, since the callers merely care
whether the result is nonzero. New arg FILE so that error messages
can report the file name. Keep track of eof. Initialize the line
table too, taking its memory from the input buffer's memory; this
subsumes the old findlines function and removes the need for worrying
about running out of line table entries.
(checkfp, mergefps, sortlines, merge, sort): Adjust to the new storage
management regime, in particular the fact that line tables are now
filled in by fillbuf and are in reverse order.
(checkfp): Now takes char *, not const char *, since subroutines
require that now. Rewrite to avoid lint and duplicate code.
If line length alloc calculation overflows,
simply allocate enough memory to hold the line.
(mergefps): New arg FILES, used for buffer size calculation and error
messages. Rewrite to avoid lint. Do not loop if savealloc*2
overflows.
(mergefps, merge): Zap temporary files eagerly rather than lazily;
this is needed because we now pass FILES to mergefps.
(sortlines): Args now point at end of arrays, not at beginnings.
(sort): Do not allocate temporary line array for sortlines;
instead, take the space from the same buffer.
(main): Adjust to sort_size and default_sort_size changes.
(wheel_tab): New global table.
(WHEEL_START, WHEEL_END): Define.
(factor): Remove the loop that special-cased `2'.
Instead of incrementing by `2', use the offsets from the wheel table.
From Michael Steffens.
