Corinna Vinschen [Tue, 22 Jan 2019 14:23:05 +0000 (15:23 +0100)]
Cygwin: posix timers: reimplement using OS timer
- Rename files timer.* to posix_timer.*.
- Reimplement using an OS timer rather than a handcrafted wait loop.
- Use a Slim R/W Lock for synchronization.
- Drop timer chaining. It doesn't server a purpose since all timers
are local only.
- Rename ttstart to itimer_tracker to better reflect its purpose.
It's not the anchor for a timer chain anymore anyway.
- Drop fixup_timers_after_fork. Everything is process-local, nothing
gets inherited.
- Rename timer_tracker::disarm_event to disarm_overrun_event for
better readability.
Corinna Vinschen [Tue, 22 Jan 2019 14:06:51 +0000 (15:06 +0100)]
Cygwin: timerfd: another overrun computation fix and drop useless variable
- When correcting the next expiration timestamp, the number of
expirations gets computed correctly, just the expiration timestamp
itself is then only incremented by a single interval, rather than
the just computed expired intervals. Fix that.
- drop the local clock variable in timerfd_tracker::create. It doesn't
serve any purpose.
Corinna Vinschen [Mon, 21 Jan 2019 11:26:51 +0000 (12:26 +0100)]
Cygwin: timerfd: rename overrun_count to expiration_count
The value returned by reading from a timerfd is not an overrun
count in the same sense as for posix timers, it's an expiry counter.
Reflect that in the name.
Corinna Vinschen [Mon, 21 Jan 2019 10:14:16 +0000 (11:14 +0100)]
Cygwin: timerfd: fix overrun computation
- Drop erroneous initial computation of overrun count in settime
for absolute non-realtime clocks. It's repeated in thread_func
and thus counted twice.
- Fix overrun computation for timestamp offsets being a multiple of
the timer interval. The timestamp has to be corrected after the
first offset, otherwise the correction loop counts the intervals
again.
Corinna Vinschen [Sun, 20 Jan 2019 21:18:17 +0000 (22:18 +0100)]
Cygwin: timerfd: fix read(2) running wild
- On systems with inexact realtime clock, the current timestamp may
be fractionally smaller than the desired timestamp. This breaks the
computation for incrementing overrun_count so overrun_count may end
up as 0. Expiring the timer with an overrun_count of 0 is a no-go.
Make sure we always increment overrun_count by at least one after
timer expiry.
- Do not expire the timer when another process deletes its timer_tracker.
This, too, may result in a 0 overrun_count.
Corinna Vinschen [Sat, 19 Jan 2019 18:53:48 +0000 (19:53 +0100)]
Cygwin: timerfd: reimplement from scratch
Using posix timers "timer_tracker" as base class for timerfd was flawed.
Posix timers are not inherited by child processes and don't survive
execve. The method used by posix timers didn't allow to share timers
between processes. The timers were still per-process timers and worked
entirely separate from each other. Reading from these timers via
different descriptors was only synchronized within the same process.
This does not reflect the timerfd semantics in Linux: The per-file
timers can be dup'ed and survive fork and execve. They are still just
descriptors pointing to the same timer object originally created by
timerfd_create. Synchronization is performed between all descriptor
instances of the same timer, system-wide.
Thus, reimplement timerfd using a timer instance in shared memory,
a kernel timer, and a handful of sync objects.
Every process maintains a per-process timerfd struct on the cygheap
maintaining a per-process thread. Every process sharing the same
timerfd will run this thread checking the state of the timer, similar
to the posix timer thread, just working on the shared objects and
synchronizing its job with each other thread.
Drop the timerfd implementation in the posix timer code and move the
public API to fhandler_timerfd.c. Keep the ttstart timer_tracker
anchor out of "NO_COPY" since the fixup_after_fork code should run to
avoid memory leakage.
Corinna Vinschen [Wed, 16 Jan 2019 14:33:15 +0000 (15:33 +0100)]
Cygwin: timerfd: implement fork semantics
- Puzzeling: Commit ec98d19a08c2e4678e8a6f40fea0c9bbeaa4a2c7
changed ttstart to NO_COPY but kept all the code to handle
fixup after fork. Revert to not-NO_COPY and make timerfd
fork work.
- On fixup_after_fork, keep timerfd timers and restart thread
if they were armed in the parent.
- Move timerfd timer_trackers to cygheap. Overload timer_tracker
new and delete methods to handle timers accordingly. This is not
exactly required for fork, but exec will be grateful.
- Give up on TFD_TIMER_CANCEL_ON_SET for now. There's no easy way
to recognize a discontinuous change in a clock.
Corinna Vinschen [Tue, 15 Jan 2019 23:11:03 +0000 (00:11 +0100)]
Cygwin: posix timers: fix overrun computation
- Drop initial overrun computation from timer_tracker::settimer.
It's performed in timer_tracker::thread_func anyway.
- Fix regression in returning correct overrun count narrowed down to
int from timer_getoverrun. This has been introduced by changing
overrun_count_curr to LONG64.
Jeff Johnston [Tue, 15 Jan 2019 15:36:50 +0000 (10:36 -0500)]
AMD GCN Port contributed by Andrew Stubbs <ams@codesourcery.com>
Add support for the AMD GCN GPU architecture. This is primarily intended for
use with OpenMP and OpenACC offloading. It can also be used for stand-alone
programs, but this is intended mostly for testing the compiler and is not
expected to be useful in general.
The GPU architecture is highly parallel, and therefore Newlib must be
configured to use dynamic re-entrancy, and thread-safe malloc.
The only I/O available is a via a shared-memory interface provided by libgomp
and the gcn-run tool included with GCC. At this time this is limited to
stdout, argc/argv, and the return code.
Corinna Vinschen [Mon, 14 Jan 2019 19:38:24 +0000 (20:38 +0100)]
Cygwin: wincap: split has_posix_file_info
While FileRenameInformationEx is defined starting with Windows
10 1709 per MSDN, it only starts working in W10 1809, apparently.
Users of 1803 report "Function not implemented".
Introduce wincap_10_1809 and change the version check in
wincapc::init accordingly. Split has_posix_file_info into
has_posix_unlink_semantics and has_posix_rename_semantics.
Enable the latter only starting with W10 1809.
Corinna Vinschen [Mon, 14 Jan 2019 16:19:37 +0000 (17:19 +0100)]
Cygwin: signalfd: implement non-polling select
Allow the signal thread to recognize we're called in consequence of
select on a signalfd. If the signal is part of the wait mask, don't
call any signal handler and don't remove the signal from the queue,
so a subsequent read (or sigwaitinfo/sigtimedwait) still gets the
signal. Instead, just signal the event object at
_cygtls::signalfd_select_wait for the thread running select.
The addition of signalfd_select_wait to _cygtls unearthed the alignment
problem of the context member again. To make sure this doesn't get lost,
improve the related comment in the header file so that this (hopefully)
doesn't get lost (again).
Corinna Vinschen [Sun, 13 Jan 2019 22:13:33 +0000 (23:13 +0100)]
Cygwin: signal: implement signalfd
First cut of a signalfd implementation.
Still TODO: Non-polling select.
This should mostly work as on Linux except for missing support
for some members of struct signalfd_siginfo, namely ssi_fd,
ssi_band (both SIGIO/SIGPOLL, not fully implemented) and ssi_trapno
(HW exception, required HW support).
Corinna Vinschen [Sun, 13 Jan 2019 21:43:52 +0000 (22:43 +0100)]
Cygwin: select: fix overwriting fd sets if poll returns no fd
There's a long-standing bug in select. If we have poll-only
descriptors in the fd set, select overwrites the incoming
fd sets with the polling result. If none of the fds is ready,
select has to loop again. But now the fd sets are set to all
zero and select hangs.
Fix this by utilizing the local fd sets r, w, e as storage for
the incoming fd sets and use them to initialize select_stuff.
If we have to loop, overwritung the incoming fd sets doesn't matter.
While at it, rename r, w, e to readfds_in, writefds_in, exceptfds_in.
Cygwin: try_to_bin: don't check recycler filename all the time
So far we check the recycler name all the time, and the last interation
also only managed to handle two ways to write the recycler. However,
an adventurous user might change the case of the recycler arbitrarily.
Fix this problem by keeping track of the name in a somewhat relaxed
fashion. Use camel back on drive C by default, all upper case elsewhere.
Only if the rename op fails do we fix the recycler name on the fly
when trying to create it, and it turns out it already existed.
Cygwin: try_to_bin: fix rootdir handle after reopening
If the first rename fails, we reopen the rootdir for creating a subdir.
The rootdir handle can change its value at this point, but the code
doesn't take this into account. The subsequent rename then fails with
STATUS_INVALID_HANDLE. Fix this by copying the new rootdir value to
pfri->RootDirectory.
So far we reopened the file if it was opened case sensitive to
workaround the problem that the recycler could be named in
camel back or all upper case, depending on who created it.
That's a problem for O_TMPFILE on pre-W10. As soon as the
original HANDLE gets closed, delete-on-close is converted to full
delete disposition and all useful operations on the file cease to
work (STATUS_ACCESS_DENIED or STATUS_FILE_DELETED).
To avoid that problem drop the reopen code and check for the exact
recycler filename, either $Recycle.Bin or $RECYCLE.BIN, if the file
has been opened case sensitive.
Cygwin: introduce fhandler_process_fd and add stat(2) handling
move special fd symlink code into own fhandler_process_fd class
to simplify further additions to /proc/PID/fd/DESCRIPTOR symlink
handling.
Add a method to handle stat(2) on such a proc fd symlink by handle.
This allows correct reply from stat(2) if the target file has been
deleted. This eventually fixes `awk -f /dev/fd/3 3<<eof'.
Cygwin: path_conv: add PATH_RESOLVE_PROCFD path_types flag
path_conv now sets the PATH_RESOLVE_PROCFD flag in path_flags if
the PC_SYM_NOFOLLOW_PROCFD pathconv_arg flag has been set on input
*and* the file is actually a proc fd symlink.
Add matching path_conv::follow_fd_symlink method for checking and
use it in open(2).
Cygwin: path_conv: decouple path_types from mount types
- Remove another unfortunate amalgamation: Mount flags (MOUNT_xxx)
are converted to path_types (PATH_xxx) and mixed with non-mount
path_types flags in the same storage, leading to a tangled,
pell-mell usage of mount flags and path flags in path_conv and
symlink_info.
- There's also the case of PC_NONULLEMPTY. It's used in exactly
one place with a path_conv constructor only used in this single
place, just to override the automatic PC_NULLEMPTY addition
when calling the other path_conv constructors. Crazily,
PC_NONULLEMPTY is a define, no path_types flag, despite its
name.
- It doesn't help that the binary flag exists as mount and path
flag, while the text flag only exists as path flag. This leads
to mount code using path flags to set text/binary. Very confusing
is the fact that a text mount/path flag is not actually required;
the mount code sets the text flag on non binary mounts anyway, so
there are only two states. However, to puzzle people a bit more,
path_conv::binary wrongly implies there's a third, non-binary/non-text
state.
Clean up this mess:
- Store path flags separately from mount flags in path_conv and
symlink_info classes and change all checks and testing inline
methods accordingly.
- Make PC_NONULLEMPTY a simple path_types flag and drop the
redundant path_check constructor.
- Clean up the definition of pathconv_arg, path_types, and mount flags.
Use _BIT expression, newly define in cygwin/bits.h.
Cygwin: path_conv: decouple pathconv_flags from path_types
There's an unfortunate amalgamation of caller-provided pathconv_arg
flags with path_types flags which in turn are mostly mount flags.
This leads to a confusion of flag values in sylink_info::pflags and,
in turn, in path_conv::path_flags.
This patch decouples pathconv_flags from the other flags by making
sure that a pathconv_flag is never copied into a variable used for
path_types flags. Also, remove PATH_NO_ACCESS_CHECK since it's
not necessary.
Cygwin: Move O_TMPFILE to bin and allow linkat by handle
Along the same lines as the previous patch: By reopening an
O_TMPFILE by handle, we can now move the file to the bin at
open time and thus free'ing up the parent dir and *still*
open the file as /proc/PID/fd/DESCRIPTOR by linkat(2).
Cygwin: implement /proc/PID/fd/DESCRIPTOR reopening by handle
Allows expressions along the lines of `cat /proc/self/fd/0 <<EOF'.
The problem here is that the temporary file used for the here script
has already been deleted by the shell. Opening by filename, as
implemented so far, doesn't work because the file has been moved
to the bin.
Allow reopening files by handle the same way from another process
as long as we have sufficient permissions on the foreign process.
Cygwin: pinfo: add method to send a serialized path_conv and HANDLE
To allow reopening a file open in another process by HANDLE, introduce
a matching file_pathconv method, taking a file descriptor as parameter.
The result is a serialized path_conv and a HANDLE value. The HANDLE is
valid in the foreign process and MUST be duplicated into the target
process before usage.
Cygwin: fhandler_base::open: allow to reopen file from handle
So far io_handle is NULL when calling fhandler_base::open to
open or create a file. Add a check for io_handle to allow
priming the fhandler with a HANDLE value so we can reopen a
file from a HANDLE on file systems supporting it. This allows
to open already deleted files for further action. This will
be used by open("/proc/PID/fd/DESCRIPTOR") scenarios.
There's no POSIX semantics allowing to link a file with a link
count of 0 and making it available in the file system again.
In fact, the Linux linkat extension AT_EMPTY_PATH explicitely
disallows to link a file descriptor to a file with a link count
of 0, except for O_TMPFILE without O_EXCL.
Wilco Dijkstra [Mon, 31 Dec 2018 18:01:52 +0000 (18:01 +0000)]
Improve performance of memmem
This patch significantly improves performance of memmem using a novel
modified Horspool algorithm. Needles up to size 256 use a bad-character
table indexed by hashed pairs of characters to quickly skip past mismatches.
Long needles use a self-adapting filtering step to avoid comparing the whole
needle repeatedly.
By limiting the needle length to 256, the shift table only requires 8 bits
per entry, lowering preprocessing overhead and minimizing cache effects.
This limit also implies worst-case performance is linear.
Small needles up to size 2 use a dedicated linear search. Very long needles
use the Two-Way algorithm (to avoid increasing stack size inlining is now disabled).
The performance gain is 6.6 times on English text on AArch64 using random
needles with average size 8 (this is even faster than the recently improved strstr
algorithm, so I'll update that in the near future).
The size-optimized memmem has also been rewritten from scratch to get a
2.7x performance gain.
Tested against GLIBC testsuite and randomized tests.
Corinna Vinschen [Wed, 26 Dec 2018 20:51:13 +0000 (21:51 +0100)]
Cygwin: open(2): Change comment in disabled O_TMPFILE POSIX unlink code
- Turns out, the definition of POSIX unlink semantics is half-hearted
so far: It's not possible to link an open file HANDLE if it has
been deleted with POSIX semantics, nor is it possible to remove
the delete disposition. This breaks linkat on an O_TMPFILE.