The mouse signal SIGUSR2 was not turned off for normal termination and
in some other cases. Thus mouse signals arriving after the frame
buffer was unmapped always caused fatal traps. The fatal traps occurred
about 1 time in 5 if the mouse was wiggled while vgl is ending.
The screen switch signal SIGUSR1 was turned off after clearing the
flag that it sets. Unlike the mouse signal, this signal is handled
synchronously, but VGLEnd() does screen clearing which does the
synchronous handling. This race is harder to lose. I think it can
get vgl into deadlocked state (waiting in the screen switch handler
with SIGUSR1 to leave that state already turned off).
Turn off the mouse cursor before clearing the screen in VGLEnd().
Otherwise, clearing is careful to not clear the mouse cursor. Undrawing
an active mouse cursor uses a lot of state, so is dangerous for abnormal
termination, but so is clearing. Clearing is slow and is usually not
needed, since the kernel also does it (not quite right).
worked right for white interiors and black borders was used). Advertise
this by changing the default colors to a red interior and a white
border (the same as the kernel default). Add undocumented env variables
for changing these colors. Also change to the larger and better-shaped
16x10 cursor sometimes used in the kernel. The kernel choice is
fancier, but libvgl is closer to supporting the larger cursors needed
in newer modes.
The (n)and-or logic for the cursor doesn't work right for more than 2
colors. The (n)and part only masks out all color bits for the pixel
under the cursor when all bits are set in the And mask. With more
complicated logic, the non-masked bits could be used to implement
translucent cursors, but they actually just gave strange colors
(especially in packed and planar modes where the bits are indirect
through 1 or 2 palettes so it is hard to predict the final color).
They also gave a bug for writing pixels under the cursor. The
non-masked bits under the cursor were not combined in this case.
Drop support for combining with bits under the cursor by making any nonzero
value in the And mask mean all bits set.
Convert the Or mask (which is represented as a half-initialized 256-color
bitmap) to a fully initialized bitmap with the correct number of colors.
The 256-color representation must be as in 3:3:2 direct mode iff the final
bitmap has more than 256 colors. The conversion of colors is not very
efficient, so convert at initialization time.
code for reading from the frame buffer.
Reading from the frame buffer is usually much slower than writing to
the frame buffer. Typically 10 to 100 times slower. It old modes,
it takes many more PIOs, and in newer modes with no PIOs writes are
often write-combined while reads remain uncached.
Reading from the frame buffer is not very common, so this change doesn't
give speedups of 10 to 100 times. My main test case is a floodfill()
function that reads about as many pixels as it writes. The speedups
are typically a factor of 2 to 4.
Duplicating writes to the shadow buffer is slower when no reads from the
frame buffer are done, but reads are often done for the pixels under the
mouse cursor, and doing these reads from the shadow buffer more than
compensates for the overhead of writing the shadow buffer in at least the
slower modes. Management of the mouse cursor also becomes simpler.
The shadow buffer doesn't take any extra memory, except twice as much
in old 4-plane modes. A buffer for holding a copy of the frame buffer
was allocated up front for use in the screen switching signal handler.
This wasn't changed when the handler was made async-signal safe. Use
the same buffer the shadow (but make it twice as large in the 4-plane
modes), and remove large special code for writing it as well as large
special code for reading ut. It used to have a rawer format in the
4-plane modes. Now it has a bitmap format which takes twice as much
memory but can be written almost as fast without special code.
VIDBUFs that are not the whole frame buffer were never supported, and the
change depends on this. Check for invalid VIDBUFs in some places and do
nothing. The removed code did something not so good.
not doing any unnecessary PIO instructions or refusing to start when the
i/o privilege needed for these instructions cannot be acquired.
This turns off useless palette management in direct modes. Palette
management had no useful effect since the hardware palette is not used
in these modes.
This transiently acquires i/o privilege if possible as needed to give
VGLSetBorder() and VGLBlankDisplay() a chance of working. Neither has
much chance of working. I was going to drop support for them in direct
modes, but found that VGLBlankDisplay() still works with an old graphics
card on a not so old LCD monitor.
This has some good side effects: reduce glitches for managing the palette
for screen switches, and speed up and reduce async-signal-unsafeness in
mouse cursor drawing.
screen bitmap and within a single MEMBUF were broken when first source
line is before the first destination line and the sub-bitmaps overlap.
The fix just copies horizontal lines in reverse order when the first
source line is before the first destination line. This switches
directions unnecessarily in some cases, but the switch is about as
fast as doing a precise detection of overlaps. When the first lines
are the same, there can be undetected overlap in the horizontal
direction. The old code already handles this mostly accidentally by
using bcopy() for MEMBUFs and by copying through a temporary buffer
for the screen bitmap although the latter is sub-optimal in direct
modes.
with old kernels, by breaking the support for large frame buffers in the
same way as for current kernels.
Large frame buffers may be too large to map into kva, and the kernel
(syscons) only uses the first screen page anyway, so r203535, r205557
and 248799 limit the buffer size in VESA modes to the first screen
page, apparently without noticing that this breaks applications by
using the same limit for user mappings as for kernel mappings. In
vgl, this makes the virtual screen the same as the physical screen.
However, this is almost a feature since clearing and switching large
(usually mostly unused) frame buffers takes too long. E.g., on a 16
year old low-end AGP card it takes about 12 seconds to clear the 128MB
frame buffer in old kernels that map it all and also map it with slow
attributes (e.g., uncacheable). Older PCI cards are even slower, but
usually have less memory. Newer PCIe cards are faster, but may have
many GB of memory. Also, vgl malloc()s a shadow buffer with the same
size as the frame buffer, so large frame buffers are even more wasteful
in applications than in the kernel.
Use the same limit in vgl as in newer kernels.
Virtual screens and panning still work in non-VESA modes that have
more than 1 page. The reduced buffer size in the kernel also breaks
mmap() of the last physical page in modes where the reduced size is
not a multiple of the physical page size. The same reduction in vgl
only reduces the virtual screen size.
depths > 8. Add some smaller optimizations for these depths. Use a
more generic method for all depths >= 8, although this gives tiny
pessimizations for these depths.
For clearing the whole frame buffer, avoid the same pessimizations
for depths > 8. Add some larger optimizations for these depths. Use
an even more generic method for all depths >= 8 to give the optimizations
for depths > 8 and a tiny pessimization for depth 8.
The main pessimization was that old versions of bcopy() copy 1 byte at a
time for all trailing bytes. (i386 still does this. amd64 now pessimizzes
large sizes instead of small ones if the CPU supports ERMS. dev/fb gets
this wrong by mostly not using the bcopy() family or the technically correct
bus space functions but by mostly copying 2 bytes at a time using an
unoptimized loop without even volatile declarations to prevent the compiler
rewriting it.)
The sizes here are 1, 2, 3 or 4 bytes, so depths 9-16 were up to twice as
slow as necessary and depths 17-24 were up to 3 times slower than necessary.
Fix this (except depths 17-24 are still up to 2 times slower than necessary)
by using (builtin) memcpy() instead of bcopy() and reorganizing so that the
complier can see the small constant sizes. Reduce special cases while
reorganizing although this is slightly slower than adding special cases.
The compiler inlining (and even -O2 vs -O0) makes little difference compared
with reducing the number of accesses except on modern hardware it gives a
small improvement.
Clearing was also pessimized mainly by the extra accesses. Fix it quite
differently by creating a MEMBUF containing 1 line (in fast memory using
a slow method) and copying this. This is only slightly slower than reducing
everything to efficient memset()s and bcopy()s, but simpler, especially
for the segmented case. This works for planar modes too, but don't use it
then since the old method was actually optimal for planar modes (it works
by moving the slow i/o instructions out of inner loops), while for direct
modes the slow instructions were all in the invisible inner loop in bcopy().
Use htole32() and le32toh() and some type puns instead of unoptimized
functions for converting colors. This optimization is mostly in the noise.
libvgl is only supported on x86, so it could hard-code the assumption that
the byte order is le32, but the old conversion functions didn't hard-code
this.
the same code as the VIDBUF8 case, so it only worked for depths <= 8.
The 2 directions for copying between VIDBUFs and MEMBUFs worked by using
a Read/Write organization which makes the destination a VIDBUF so the
MEMBUF case was not reached, and the VIDBUF cases have already been fixed.
Fix this by removing "optimizations" for the VIDBUF8 case so that the
MEMBUF case can fall through to the general (non-segmented) case. The
optimizations were to duplicate code for the VIDBUF8 case so as to
avoid 2 multiplications by 1 at runtime. This optimization is not useful
since the multiplications are not in the inner loop.
Remove the same "optimization" for the VIDBUF8S case. It was even less
useful there since it duplicated more to do relatively less.
Correct restoring was only attempted for mode 258 (800x600x4 P). (This
was the only useful graphics mode supported in the kernel until 10-15
years ago, and is still the only one explicitly documented in the man
page). The comment says that it is the geometry (subscreen size) that
is restored, but it seems to only be necessary to restore the font
size, with the geometry only needed since it is set by the same ioctl.
The font size was not restored for this mode, but was forced to 16.
For other graphics modes, the font size was clobbered to 0. This
confuses but doesn't crash the kernel (font size 0 gives null text).
This confuses and crashes vidcontrol. The only way to recover was to
use vidcontrol to set the mode to any text mode on the way back to the
original graphics mode.
vidcontrol gets this wrong in the opposite way when backing out of
changes after an error. It restores the font size correctly, but
forces the geometry to the full screen size.
r80270 has the usual wrong fix for unsafe signal handling -- just set
a flag and return to let an event loop check the flag and do safe
handling. This never works for signals like SIGBUS and SIGSEGV that
repeat and works poorly for others unless the application has an event
loop designed to support this.
For these signals, clean up unsafely as before, except for arranging that
nested signals are fatal and forcing a nested signal if the cleanup doesn't
cause one.
method as in /bin/sh.
We still do technically undefined things in the signal handler, but it
is safe in practice to access state that is protected by INTOFF/INTON.
In a recent commit, I sprinkled VGLMouseFrozen++/-- operations in
places that need INTOFF/INTON. This prevented clobbering of pixels
under the mouse, but left mouse signals deferred for too long. It is
necessary to call the signal handler when the count goes to 0. Old
versions did this in the unfreeze function, but didn't block actual
signals, so the signal handler raced itself. The sprinkled operations
reduced the races, but when then worked to block a race they left
signals deferred for too long.
Use INTOFF/INTON to fix complete loss of mouse signals while reading
the mouse status. Clobbering of the state was prevented by SIG_IGN'ing
mouse signals, but that has a high overhead and broke more than it
fixed by losing mouse signals completely. sigprocmask() works to block
signals without losing them completely, but its overhead is also too
high.
libvgl's mouse signal handling is often worse than none. Applications
can't block waiting for a mouse or keyboard or other event, but have
to busy-wait. The SIG_IGN's lost about half of all mouse events while
busy-waiting for mouse events.
It started truncating its color arg to 8 bits using plot() in r229415.
The version in r229415 is also more than 3 times slower in segmented
modes, by doing more syscalls to move the window.
segmented modes.
Also fix some style bugs in the 2 changed lines. libvgl uses a very non-KNF
style with 2-column indentation with no tabs except for regressions.
Reading of single pixels didn't look under the cursor.
Copying of 1x1 bitmaps didn't look under the cursor for either reading
or writing.
Copying of larger bitmaps looked under the cursor for at most the
destination.
Copying of larger bitmaps looked under a garbage cursor (for the Display
bitmap) when the destination is a MEMBUF. The results are not used, so
this only wasted time and flickered the cursor.
Writing of single pixels looked under a garbage cursor for MEMBUF
destinations, as above except this clobbered the current cursor and
didn't update the MEMBUF. Writing of single pixels is not implemented
yet in depths > 8. Otherwise, writing of single pixels worked. It was
the only working case for accessing pixels under the cursor.
Clearing of MEMBUFs wasted time freezing the cursor in the Display bitmap.
The fixes abuse the top bits in the color arg to the cursor freezing
function to control the function. Also clear the top 8 bits so that
applications can't clobber the control bits or create 256 aliases for
every 24-bit pixel value in depth 32.
Races fixed:
Showing and hiding the cursor only tried to avoid races with the mouse
event signal handler for internal operations. There are still many
shorter races from not using volatile or sig_atomic_t for the variable
to control this. This variable also controls freezes, and has more
complicated states than before.
The internal operation of unfreezing the cursor opened a race window
by unsetting the signal/freeze variable before showing the cursor.
depths for the source and target are not supported. The bits for higher
numbered planes (mostly for red) were either not copied or were copied to
lower numbered planes for nearby pixels.
Quick fix for creation of mouse cursor bitmaps in all depths. This fix is
only complete for the default lightwhite cursor with a black frame.
Even the lightwhite and black colors are hard to find. The templates
use 0xff for lightwhite, but that means brightblue in the simplest mode
(Truecolor depth 24). Other modes are even more complicated -- they are
singly or doubly indirect throught palette(s) and changing of the palettes
by applications is supported.
Details:
Replicate the template value for Truecolor modes to fill out the target
depth (and more for depths not a multiple of 8). Do this for every
drawing of the cursor so that it sort of works for mouse cursor bitmaps
set by applications.
Use 0xf for lightwhite in most other modes. Only do this for the
default cursor so that it doesn't affect mouse cursor bitmaps set by
applications. 0xf mostly works because it was originally for CGA
lightwhite and is emulated using 1 or 2 indirections on EGA and VGA.
0x3f (EGA white) and 0xff (VGA black) direct palette indexes mostly
don't work since backwards compatibility inhibits or prevents them
representing lightwhite. But 0x3f (EGA white) must be used for mode
37 (VGA_MODEX) (320x240x8 V) since this mode is closer to EGA than VGA.
was not taken modulo the window size in VGLClear().
Segmented modes also need a kernel fix to almost work. The ioctl to set
the window origin is broken.
These bugs are rarely problems since non-VESA modes only need
segmentation to support multiple pages but libvgl doesn't support
multiple pages and treats these modes as non-segmented, and VESA modes
are usually mapped linearly except on old hardware so they really are
non-segmented.
to match the change in its declaration. Change the declaration back to
"byte color" since setting of the border color is not supported for more
than 256 colors.
VGLBitmapString() and VGLSetBorder() so as to not truncate to 8 bits.
Complete the corresponding fix for VGLGetXY() and VGLPutXY() (parts
of the man page were out of date).
Since the font format is undocumented, it is unclear how non-multiples
of 8 should be padded to bytes in the font file. Use the same
representation as bdf text format (big- endian, with padding in the
lower bits).
There seems to be no alternative to reading each plane independently using
3 slow i/o's per plane (this delivers 8 nearby pixels, but we don't buffer
the results so run 8 times slower than necessary.
All the code for this was there, but it was ifdefed out and replaced by
simpler code that cannot work in planar modes. The ifdefed out code
was correct except it was missing a volatile declaration, so compilers
optimized the multiple dummy reads in it to a single read.
Support for 16-bit and 32-bit Truecolor modes was supposed to be
complete in r70991 of main.c and in nearby revisions for other files, but
it was broken by the overruns in most cases (all cases were the mouse
is enabled, and most cases where bitmaps are used). r70991 also
uninintentionally added support for depths 9-15, 17-23 and 25-31.
Depth 24 was more obviously broken and its support is ifdefed out. In
the other ranges, only depth 15 is common. It was broken by buffer
overruns in all cases.
bitmap.c:
- the static buffer was used even when it was too small (but it was
large enough to often work accidentally in depth 16)
- the size of the dynamically allocated buffer was too small
- the sizing info bitmap->PixelBytes was not inititialzed in the bitmap
constructor. It often ended up as 0 for MEMBUFs, so using it in more
places gave more null pointer accesses. (It is per-bitmap, but since
conversion between bitmaps of different depths is not supported (except
from 4 bits by padding to 8), it would work better if it were global.)
main.c:
- depths were rounded down instead of up to a multiple of 8, so PixelBytes
was 1 too small for depths above 8 except 16, 24 and 32.
- PixelBytes was not initialized for 4-bit planar modes. It isn't really
used for frame buffer accesses in these modes, but needs to be 1 in
MEMBUF images.
mouse.c:
- the mouse cursor buffers were too small.
vgl.h:
- PixelBytes was not initialized in the static bitmap constructor. It
should be initialized to the value for the current mode, but that is
impossible in a static constructor. Initialize it to -1 so as to
fail if it is used without further initialization.
All modes that are supposed to be supported now don't crash in
nontrivial tests, and almost work. Missing uses of PixelBytes now
give in-bounds wrong pointers instead of overruns. Misconversions of
bitmaps give multiple miscolored mouse cursors instead of 1 white one,
and similarly for bitmaps copied through a MEMBUF.
args (neither MAP_PRIVATE nor MAP_SHARED). It was broken in r271635
and/or r271724 by stricter checking. The compatibility code in r271724
doesn't work for my old binaries (actually new binaries with old
libraries).
PR: needed to test the fix for PR 162373
Mainly focus on files that use BSD 2-Clause license, however the tool I
was using mis-identified many licenses so this was mostly a manual - error
prone - task.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
This both avoids some dependencies on xinstall.host and allows
bootstrapping on older releases to work due to lack of at least 'install -l'
support.
Sponsored by: EMC / Isilon Storage Division
Off by default, build behaves normally.
WITH_META_MODE we get auto objdir creation, the ability to
start build from anywhere in the tree.
Still need to add real targets under targets/ to build packages.
Differential Revision: D2796
Reviewed by: brooks imp
http://www.graphicsgems.org/
At the time it claimed to be 3-4 times faster than the traditional
algorithm.
PR: 18769
Approved by: jhb (mentor)
MFC after: 2 weeks
'inline' to 'static inline'. Otherwise, a C99 compiler (such as clang)
will output an undefined symbol for those functions in the resulting
object file. (Even gcc will do this, when you use "-std=c99".)
This should fix the "undefined reference to `set4pixels'" errors that
some people were seeing during ports building, when their world was
compiled with clang.
Approved by: rpaulo (mentor)
Similar to libexec/, do the same with lib/. Make WARNS=6 the norm and
lower it when needed.
I'm setting WARNS?=0 for secure/. It seems secure/ includes the
Makefile.inc provided by lib/. I'm not going to touch that directory.
Most of the code there is contributed anyway.
preparation for 8.0-RELEASE. Add the previous version of those
libraries to ObsoleteFiles.inc and bump __FreeBSD_Version.
Reviewed by: kib
Approved by: re (rwatson)