Protocols are nice, but there is no way for C code to verify whether
a type's "protocol" structure actually implements some particular
protocol. As a result, you can pass an object that implements the
"vfs" protocol to one that expects the "stream" protocol, and the
opposite of awesomeness ensues.
This patch adds an OPTIONAL (but enabled by default) protocol identifier
as the first member of any protocol structure. This identifier is
simply a unique QSTR chosen by the protocol designer and used by each
protocol implementer. When checking for protocol support, instead of
just checking whether the object's type has a non-NULL protocol field,
use `mp_proto_get` which implements the protocol check when possible.
The existing protocols are now named:
protocol_framebuf
protocol_i2c
protocol_pin
protocol_stream
protocol_spi
protocol_vfs
(most of these are unused in CP and are just inherited from MP; vfs and
stream are definitely used though)
I did not find any crashing examples, but here's one to give a flavor of what
is improved, using `micropython_coverage`. Before the change,
the vfs "ioctl" protocol is invoked, and the result is not intelligible
as json (but it could have resulted in a hard fault, potentially):
>>> import uos, ujson
>>> u = uos.VfsPosix('/tmp')
>>> ujson.load(u)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: syntax error in JSON
After the change, the vfs object is correctly detected as not supporting
the stream protocol:
>>> ujson.load(p)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
OSError: stream operation not supported
If a translation only has unicode code points 255 and below, the "values"
array can be 8 bits instead of 16 bits. This reclaims some code size,
e.g., in a local build, trinket_m0 / en_US reclaimed 112 bytes and de_DE
reclaimed 104 bytes. However, languages like zh_Latn_pinyin, which use
code points above 255, did not benefit.
By treating each unicode code-point as a single entity for huffman
compression, the overall compression rate can be somewhat improved
without changing the algorithm. On the decompression side, when
compressed values above 127 are encountered, they need to be
converted from a 16-bit Unicode code point into a UTF-8 byte
sequence.
Doing this returns approximately 1.5kB of flash storage with the
zh_Latn_pinyin translation. (292 -> 1768 bytes remaining in my build
of trinket_m0)
Other "more ASCII" translations benefit less, and in fact
zh_Latn_pinyin is no longer the most constrained translation!
(de_DE 1156 -> 1384 bytes free in flash, I didn't check others
before pushing for CI)
English is slightly pessimized, 2840 -> 2788 bytes, probably mostly
because the "values" array was changed from uint8_t to uint16_t,
which is strictly not required for an all-ASCII translation. This
could probably be avoided in this case, but as English is not the
most constrained translation it doesn't really matter.
Testing performed: built for feather nRF52840 express and trinket m0
in English and zh_Latn_pinyin; ran and verified the localized
messages such as
Àn xià rènhé jiàn jìnrù REPL. Shǐyòng CTRL-D chóngxīn jiāzài.
and
Press any key to enter the REPL. Use CTRL-D to reload.
were properly displayed.
When adding the ability for boards to turn on the `@micropython.native`, `viper`, and `asm_thumb` decorators it was pointed out that it's somewhat awkward to write libraries and drivers that can take advantage of this since the decorators raise `SyntaxErrors` if they aren't enabled. In the case of `viper` and `asm_thumb` this behavior makes sense as they require writing non-normative code. Drivers could have a normal and viper/thumb implementation and implement them as such:
```python
try:
import _viper_impl as _impl
except SyntaxError:
import _python_impl as _impl
def do_thing():
return _impl.do_thing()
```
For `native`, however, this behavior and the pattern to work around it is less than ideal. Since `native` code should also be valid Python code (although not necessarily the other way around) using the pattern above means *duplicating* the Python implementation and adding `@micropython.native` in the code. This is an unnecessary maintenance burden.
This commit *modifies* the behavior of the `@micropython.native` decorator. On boards with `CIRCUITPY_ENABLE_MPY_NATIVE` turned on it operates as usual. On boards with it turned off it does *nothing*- it doesn't raise a `SyntaxError` and doesn't apply optimizations. This means we can write our drivers/libraries once and take advantage of speedups on boards where they are enabled.
This improves performance of running python code by 34%, based
on the "pystone" benchmark on metro m4 express at 5000 passes
(1127.65 -> 1521.6 passes/second).
In addition, by instrumenting the tick function and monitoring on an
oscilloscope, the time actually spent in run_background_tasks() on
the metro m4 decreases from average 43% to 0.5%. (however, there's
some additional overhead that is moved around and not accounted for
in that "0.5%" figure, each time supervisor_run_background_tasks_if_tick
is called but no tick has occurred)
On the CPB, it increases pystone from 633 to 769, a smaller percentage
increase of 21%. I did not measure the time actually spent in
run_background_tasks() on CPB.
Testing performed: on metro m4 and cpb, run pystone adapted from python3.4
(change time.time to time.monotonic for sub-second resolution)
Besides running a 5000 pass test, I also ran a 50-pass test while
scoping how long an output pin was set. Average: 34.59ms or 1445/s on m4,
67.61ms or 739/s on cbp, both matching the other pystone result reasonably
well.
import pystone
import board
import digitalio
import time
d = digitalio.DigitalInOut(board.D13)
d.direction = digitalio.Direction.OUTPUT
while True:
d.value = 0
time.sleep(.01)
d.value = 1
pystone.main(50)
This code is shared by most parts, except where not all the #ifdefs
inside the tick function were present in all ports. This mostly would
have broken gamepad tick support on non-samd ports.
The "ms32" and "ms64" variants of the tick functions are introduced
because there is no 64-bit atomic read. Disabling interrupts avoids
a low probability bug where milliseconds could be off by ~49.5 days
once every ~49.5 days (2^32 ms).
Avoiding disabling interrupts when only the low 32 bits are needed is a minor
optimization.
Testing performed: on metro m4 express, USB still works and
time.monotonic_ns() still counts up
To benefit from gcc's "once-only headers" implementation, the
"wrapper-#ifndef" must be the first non-comment part of the file,
according to the manual for various gcc/cpp versions.
This PR refines the _bleio API. It was originally motivated by
the addition of a new CircuitPython service that enables reading
and modifying files on the device. Moving the BLE lifecycle outside
of the VM motivated a number of changes to remove heap allocations
in some APIs.
It also motivated unifying connection initiation to the Adapter class
rather than the Central and Peripheral classes which have been removed.
Adapter now handles the GAP portion of BLE including advertising, which
has moved but is largely unchanged, and scanning, which has been enhanced
to return an iterator of filtered results.
Once a connection is created (either by us (aka Central) or a remote
device (aka Peripheral)) it is represented by a new Connection class.
This class knows the current connection state and can discover and
instantiate remote Services along with their Characteristics and
Descriptors.
Relates to #586
If kw_args is NULL then memcpy() gets a NULL source argument.
This is undefined behavior under the C standard, even if 0 bytes
are being copied.
This problem was found using clang 7's scan-build static analyzer.
Left shift of negative numbers is undefined in the "C" standard. Multiplying
by 128 has the intended effect (in the absence of integer overflow, anyway),
can be implemented using the same shift instruction, but does not invoke
undefined behavior.
This problem was found using clang 7's scan-build static analyzer.
The remaining assignment was added in upstream micropython; the
deleted assignment was added in circuitpython as part of the long-lived
object area feature. During the merge, the redundant assignment
was not removed.
(since collected is a local variable and no pointers to it escape,
it doesn't seem possible for the placement of the assignment before
or after GC_ENTER() is important)
This diagnostic was found by clang 7's scan-build static analyzer.
Left shift of negative numbers is undefined in the "C" standard. Multiplying
by 256 has the intended effect (in the absence of integer overflow, anyway),
can be implemented using the same shift instruction, but does not invoke
undefined behavior.
This problem was found using clang 7's scan-build static analyzer.
While finding sources of clicks and buzzes in nrf i2sout, I identified
this site as one which could be long running. Reproducer code was to
play a 22.05kHz sample and repeatedly print `os.listdir('')`
While finding sources of clicks and buzzes in nrf i2sout, I identified
this site as one which could be long running. Reproducer code was to
play a 22.05kHz sample and repeatedly print `os.listdir('')`
Testing performed: That the shipped .mpy files on a PyPortal (CP 4.x)
still work (play audio) with this branch, instead of erroring because
`WaveFile` can't be found in `audioio`.
Flash usage grew by 28 bytes. (I expected 24, there must be some other
effect on size/alignment that I didn't predict)
In #2013, @danh says:
My choice of where to put the semicolon is deliberate,
so that we can say
RUN_BACKGROUND_TASKS;
not have a redundant semicolon, and not confuse C code formatting.
.. such as namedtuple and attrtuple objects. This is the same
predicate used elsewhere in the file to check for adequate compatibility
between the types.
This was discovered due to crashing `time.time()` on the nrf port.
Closes: #2052
It is possible for this routine to expand some inputs, and in fact
it does for certan strings in the proposed Korean translation of
CircuitPython (#1858). I did not determine what the maximum
expansion is -- it's probably modest, like len()/7+2 bytes or
something -- so I tried to just make enc[] an adequate
over-allocation, and then ensured that all the strings in the
proposed ko.po now worked. The worst actual expansion seems to be a
string that goes from 65 UTF-8-encoded bytes to 68 compressed bytes
(+4.6%). Only a few out of all strings are reported as
non-compressed.
When nrf pwm audio is introduced, it will be called `audiopwmio`. To
enable code sharing with the existing (dac-based) `audioio`, factor
the sample and mixer types to `audiocore`.
INCOMPATIBLE CHANGE: Now, `Mixer`, `RawSample` and `WaveFile` must
be imported from `audiocore`, not `audioio`.
This also improves Palette so it stores the original RGB888 colors.
Lastly, it adds I2CDisplay as a display bus to talk over I2C. Particularly
useful for the SSD1306.
Fixes#1828. Fixes#1956
If a native displayio object is accessed before it's super().__init__()
has been called, then a placeholder is given that will cause a crash if
accessed. This is tricky to get right so we detect this case and raise
a NotInplementedError instead of crashing.
Fixes#1881
For both small and long integers, raise an exception if calling
struct.pack, adding an element to an array.array, or formatting an int
with int.to_bytes would overflow the requested size.
* Update pybadge pins and flash for rev D
* TileGrid now validates the type of the pixel_shader.
* Display actually handles incoming subclass objects.
* MicroPython will inspect native parents to see if special
accessors are used.
This fixes a crash on boards with built-in displays which statically
allocate the display bus. When the pointer is provided to never
free, it tries to allocate on the non-existant heap and crashes.
This feature is controlled at compile time by MICROPY_PY_URE_SUB, disabled
by default.
Thanks to @dmazzella for the original patch for this feature; see #3770.
This feature is controlled at compile time by
MICROPY_PY_URE_MATCH_SPAN_START_END, disabled by default.
Thanks to @dmazzella for the original patch for this feature; see #3770.
This feature is controlled at compile time by MICROPY_PY_URE_MATCH_GROUPS,
disabled by default.
Thanks to @dmazzella for the original patch for this feature; see #3770.
This changes a number of things in displayio:
* Introduces BuiltinFont and Glyph so the built in font can be used by libraries. For boards with
a font it is available as board.TERMINAL_FONT. Fixes#1172
* Remove _load_row from Bitmap in favor of bitmap[] access. Index can be x/y tuple or overall index. Fixes#1191
* Add width and height properties to Bitmap.
* Add insert and [] access to Group. Fixes#1518
* Add index param to pop on Group.
* Terminal no longer takes unicode character info. It takes a BuiltinFont instead.
* Fix Terminal's handling of [###D vt100 commands used when up arrowing into repl history.
* Add x and y positions to Group plus scale as well.
* Add bitmap accessor for BuiltinFont
This creates a common safe mode mechanic that ports can share.
As a result, the nRF52 now has safe mode support as well.
The common safe mode adds a 700ms delay at startup where a reset
during that window will cause a reset into safe mode. This window
is designated by a yellow status pixel and flashing the single led
three times.
A couple NeoPixel fixes are included for the nRF52 as well.
Fixes#1034. Fixes#990. Fixes#615.
The backtrace cannot be given because it relies on the validity
of the qstr data structures on the heap which may have been
corrupted.
In fact, it still can crash hard when the bytecode itself is
overwritten. To fix, we'd need a way to skip gathering the
backtrace completely.
This also increases the default stack size on M4s so it can
accomodate the stack needed by ASF4s nvm API.
This adds support for the OSError attributes : errno, strerror, filename and filename2.
CPython only sets errno if 2 arguments has been passed in. This has not been implemented here.
CPython OSError.args is capped at 2 items for backward compatibility reasons. This has not been
implemented here.
MICROPY_CPYTHON_COMPAT has to be enabled to get these attributes.
mp_common_errno_to_str() has been extended to check mp_errno_to_str() as well. This is done to ease
reuse for the strerror argument.
Add traceback chain to sys.exec_info()[2].
No actual frame info is added, but just enough to recreate the printed
exception traceback.
Used by the unittest module which collects errors and failures and prints
them at the end.
This gives access to the function underlying the bound method.
Used in the converted CPython stdlib logging.Formatter class to handle
overrriding a default converter method bound to a class variable.
The method becomes bound when accessed from an instance of that class.
I didn't investigate why CircuitPython turns it into a bound method.
This reverts commit 869024dd6e.
Ctrl-C stopped producing KeyboardInterrupt with this change on CircuitPython.
The Unix and stm32 ports handles Ctrl-C differently with a handler which is
probably why they where not affected.
Fixes#1092
This saves code space in builds which use link-time optimization.
The optimization drops the untranslated strings and replaces them
with a compressed_string_t struct. It can then be decompressed to
a c string.
Builds without LTO work as well but include both untranslated
strings and compressed strings.
This work could be expanded to include QSTRs and loaded strings if
a compress method is added to C. Its tracked in #531.
Imports generate a lot of garbage so cleaning it up immediately
reduces the likelihood longer lived data structures don't end up in
the middle of the heap.
Fixes#856
The existing mp_get_stream_raise() helper does explicit checks that the
input object is a real pointer object, has a non-NULL stream protocol, and
has the desired stream C method (read/write/ioctl). In most cases it is
not necessary to do these checks because it is guaranteed that the input
object has the stream protocol and desired C methods. For example, native
objects that use the stream wrappers (eg mp_stream_readinto_obj) in their
locals dict always have the stream protocol (or else they shouldn't have
these wrappers in their locals dict).
This patch introduces an efficient mp_get_stream() which doesn't do any
checks and just extracts the stream protocol struct. This should be used
in all cases where the argument object is known to be a stream. The
existing mp_get_stream_raise() should be used primarily to verify that an
object does have the correct stream protocol methods.
All uses of mp_get_stream_raise() in py/stream.c have been converted to use
mp_get_stream() because the argument is guaranteed to be a proper stream
object.
This patch improves efficiency of stream operations and reduces code size.
This patch changes dupterm to call the native C stream methods on the
connected stream objects, instead of calling the Python readinto/write
methods. This is much more efficient for native stream objects like UART
and webrepl and doesn't require allocating a special dupterm array.
This change is a minor breaking change from the user's perspective because
dupterm no longer accepts pure user stream objects to duplicate on. But
with the recent addition of uio.IOBase it is possible to still create such
classes just by inheriting from uio.IOBase, for example:
import uio, uos
class MyStream(uio.IOBase):
def write(self, buf):
# existing write implementation
def readinto(self, buf):
# existing readinto implementation
uos.dupterm(MyStream())
Via the config value MICROPY_PY_UHASHLIB_SHA256. Default to enabled to
keep backwards compatibility.
Also add default value for the sha1 class, to at least document its
existence.
A user class derived from IOBase and implementing readinto/write/ioctl can
now be used anywhere a native stream object is accepted.
The mapping from C to Python is:
stream_p->read --> readinto(buf)
stream_p->write --> write(buf)
stream_p->ioctl --> ioctl(request, arg)
Among other things it allows the user to:
- create an object which can be passed as the file argument to print:
print(..., file=myobj), and then print will pass all the data to the
object via the objects write method (same as CPython)
- pass a user object to uio.BufferedWriter to buffer the writes (same as
CPython)
- use select.select on a user object
- register user objects with select.poll, in particular so user objects can
be used with uasyncio
- create user files that can be returned from user filesystems, and import
can import scripts from these user files
For example:
class MyOut(io.IOBase):
def write(self, buf):
print('write', repr(buf))
return len(buf)
print('hello', file=MyOut())
The feature is enabled via MICROPY_PY_IO_IOBASE which is disabled by
default.
This patch adds the gc_sweep_all() function which does a garbage collection
without tracing any root pointers, so frees all the memory, and most
importantly runs any remaining finalisers.
This helps primarily for soft reset: it will close any open files, any open
sockets, and help to get the system back to a clean state upon soft reset.
This patch is a code optimisation, trading text bytes for speed. On
pyboard it's an increase of 0.06% in code size for a gain (in pystone
performance) of roughly 6.5%.
The patch optimises load/store/delete of attributes in user defined classes
by not looking up special accessors (@property, __get__, __delete__,
__set__, __setattr__ and __getattr_) if they are guaranteed not to exist in
the class.
Currently, if you do my_obj.foo() then the runtime has to do a few checks
to see if foo is a property or has __get__, and if so delegate the call.
And for stores things like my_obj.foo = 1 has to first check if foo is a
property or has __set__ defined on it.
Doing all those checks each and every time the attribute is accessed has a
performance penalty. This patch eliminates all those checks for cases when
it's guaranteed that the checks will always fail, ie no attributes are
properties nor have any special accessor methods defined on them.
To make this guarantee it checks all attributes of a user-defined class
when it is first created. If any of the attributes of the user class are
properties or have special accessors, or any of the base classes of the
user class have them, then it sets a flag in the class to indicate that
special accessors must be checked for. Then in the load/store/delete code
it checks this flag to see if it can take the shortcut and optimise the
lookup.
It's an optimisation that's pretty widely applicable because it improves
lookup performance for all methods of user defined classes, and stores of
attributes, at least for those that don't have special accessors. And, it
allows to enable descriptors with minimal additional runtime overhead if
they are not used for a particular user class.
There is one restriction on dynamic class creation that has been introduced
by this patch: a user-defined class cannot go from zero special accessors
to one special accessor (or more) after that class has been subclassed. If
the script attempts this an AttributeError is raised (see addition to
tests/misc/non_compliant.py for an example of this case).
The cost in code space bytes for the optimisation in this patch is:
unix x64: +528
unix nanbox: +508
stm32: +192
cc3200: +200
esp8266: +332
esp32: +244
Performance tests that were done:
- on unix x86-64, pystone improved by about 5%
- on pyboard, pystone improved by about 6.5%, from 1683 up to 1794
- on pyboard, bm_chaos (from CPython benchmark suite) improved by about 5%
- on esp32, pystone improved by about 30% (but there are caching effects)
- on esp32, bm_chaos improved by about 11%
This VFS component allows to mount a host POSIX filesystem within the uPy
VFS sub-system. All traditional POSIX file access then goes through the
VFS, allowing to sandbox a uPy process to a certain sub-dir of the host
system, as well as mount other filesystem types alongside the host
filesystem.
Since a long time now, mp_obj_type_t no longer refers explicitly to
mp_stream_p_t but rather to an abstract "const void *protocol". So there's
no longer any need to define mp_stream_p_t in obj.h and it can go with all
its associated definitions in stream.h. Pretty much all users of this type
will already include the stream header.
We can provide a basic version of mp_errno_to_str even if the uerrno
module won't be provided. Rather than looking errno names up in the
uerrno module's globals dict, we'll just rely on a simple mapping in the
function itself.
The code_state.old_globals variable is there to save the globals state so
should be used for this purpose, to avoid the need for additional local
variables on the C stack.
Without this, if GC threshold is hit and there is not enough memory left to
satisfy the request, gc_collect() will run a second time and the search for
memory will happen again and will fail again.
Thanks to @adritium for pointing out this issue, see #3786.
Under ubsan, when evaluating hash(-0.) the following diagnostic occurs:
../../py/objfloat.c:102:15: runtime error: negation of
-9223372036854775808 cannot be represented in type 'mp_int_t' (aka
'long'); cast to an unsigned type to negate this value to itself
So do just that, to tell the compiler that we want to perform this
operation using modulo arithmetic rules.
Before this, ubsan would detect a problem when executing
hash(006699999999999999999999999999999999999999999999999999999999999999999999)
../../py/mpz.c:1539:20: runtime error: left shift of 1067371580458 by
32 places cannot be represented in type 'mp_int_t' (aka 'long')
When the overflow does occur it now happens as defined by the rules of
unsigned arithmetic.
When computing e.g. hash(0.4e3) with ubsan enabled, a diagnostic like the
following would occur:
../../py/objfloat.c:91:30: runtime error: shift exponent 44 is too
large for 32-bit type 'int'
By casting constant "1" to the right type the intended value is preserved.
Fuzz testing combined with the undefined behavior sanitizer found that
parsing unreasonable float literals like 1e+9999999999999 resulted in
undefined behavior due to overflow in signed integer arithmetic, and a
wrong result being returned.
There is no need to use the mp_int_t type which may be 64-bits wide, there
is enough bit-width in a normal int to parse reasonable exponents. Using
int helps to reduce code size for 64-bit ports, especially nan-boxing
builds. (Similarly for the "dig" variable which is now an unsigned int.)
Calling memset(NULL, value, 0) is not standards compliant so we must add an
explicit check that emit->label_offsets is indeed not NULL before calling
memset (this pointer will be NULL on the first pass of the parse tree and
it's more logical / safer to check this pointer rather than check that the
pass is not the first one).
Code sanitizers will warn if NULL is passed as the first value to memset,
and compilers may optimise the code based on the knowledge that any pointer
passed to memset is guaranteed not to be NULL.
Before this patch:
>>> print(')
... ')
Traceback (most recent call last):
File "<stdin>", line 1
SyntaxError: invalid syntax
After this patch:
>>> print(')
Traceback (most recent call last):
File "<stdin>", line 1
SyntaxError: invalid syntax
This matches CPython and prevents getting stuck in REPL continuation when a
1-quote is unmatched.
Before this patch, when using the switch statement for dispatch in the VM
(not computed goto) a pending exception check was done after each opcode.
This is not necessary and this patch makes the pending exception check only
happen when explicitly requested by certain opcodes, like jump. This
improves performance of the VM by about 2.5% when using the switch.
This patch fixes the macro so you can pass any name in, and the macro will
make more sense if you're reading it on its own. It worked previously
because n_state is always passed in as n_state_out_var.
gcc 8.0 supports the naked attribute for x86 systems so it can now be used
here. And in fact it is necessary to use this for nlr_push because gcc 8.0
no longer generates a prelude for this function (even without the naked
attribute).
This patch moves the start of the root pointer section in mp_state_ctx_t
so that it skips entries that are not pointers and don't need scanning.
Previously, the start of the root pointer section was at the very beginning
of the mp_state_ctx_t struct (which is the beginning of mp_state_thread_t).
This was the original assembler version of the NLR code was hard-coded to
have the nlr_top pointer at the start of this state structure. But now
that the NLR code is partially written in C there is no longer this
restriction on the location of nlr_top (and a comment to this effect has
been removed in this patch).
So now the root pointer section starts part way through the
mp_state_thread_t structure, after the entries which are not root pointers.
This patch also moves the non-pointer entries for MICROPY_ENABLE_SCHEDULER
outside the root pointer section.
Moving non-pointer entries out of the root pointer section helps to make
the GC more precise and should help to prevent some cases of collectable
garbage being kept.
This patch also has a measurable improvement in performance of the
pystone.py benchmark: on unix x86-64 and stm32 there was an improvement of
roughly 0.6% (tested with both gcc 7.3 and gcc 8.1).
This patch changes 2 things in the endianness detection:
1. Don't assume that __BYTE_ORDER__ not being __ORDER_LITTLE_ENDIAN__ means
that the machine is big endian, so add an explicit check that this macro
is indeed __ORDER_BIG_ENDIAN__ (same with __BYTE_ORDER, __LITTLE_ENDIAN
and __BIG_ENDIAN). A machine could have PDP endianness.
2. Remove the checks which base their autodetection decision on whether any
little or big endian macros are defined (eg __LITTLE_ENDIAN__ or
__BIG_ENDIAN__). Just because a system defines these does not mean it
has that endianness.
See issue #3760.
Dan Halbert
Jeff Epler
Scott Shawcroft
Scott Shawcroft
Roy Hooper
Thea Flowers
David Grimes
Jerry Needell
Hierophect
sommersoft
Elvis Pfützenreuter
Matt Wozniski
Terri Oda
scottbelden
Kathryn Lingel
Radomir Dopieralski
Damiano Mazzella
Lionel Debroux
Noralf Trønnes
Damien George
Craig Forbes
Uri Shaked
Yonatan Goldschmidt
Matt Wozniski
Jeff Epler
Jan Klusacek
Li Weiwei
Tom Collins
bildzeitung