Let's reuse the logic from scalecache and call cserve2
functions when the scalecache should be used.
So, now, cserve2 server will not scale any image... This is
too computationally intensive for the server's main thread.
This is not optimal but makes a hell of a lot more sense for
the moment. (since cserve2 manages the SHM segments)
High-level problem:
cserve2 does not support load_opts properly when opening an image.
As a result, when (pre)loading a JPEG file with specific load
options (eg. w, h, region and orientation), the image buffer might
have the wrong dimensions.
So, we need to use load_opts when computing file hash key.
And, pass these load options to the loader slave,
and use them while OPENING the image. This will set
properly the geometry.
Fixes test "Preload and Prescale" in elementary_test.
Pass around "animated" flag for images that can be animated.
Fallback to local cache if the image is animated.
Implementing support for animated images in cserve2 does
not seem to make a lot of sense considering each frame must
be requested independently in real time,... and to be honest
there doesn't seem to be any valid use case anyway :)
Images and Files can be discovered by the client scanner before they
are valid (aka. loaded or opened). We want clients to ignore all
shared objects that are not ready yet, as they are in an undefined
state (values and memory might be invalid).
Use tags such as INDX, FONT, IMAG, FILE.
These can help debugging and for robustness, as they will
be visible in the memory dump.
Update generation_id after repack as well
Also, print shared arrays usage as it is sent as first message
upon connection to the server.
Add fonts to LRU and flush LRU as needed.
Seems to work okay. Could probably be finetuned a bit.
- Create one Glyph_Data array per Font_Entry
- Copy Font_Source descriptors in shared array
- Copy font entries in shared index
- Send font index path over socket
- Merge Font_Source and Font_Data in shared memory
- Send font index path on client connect
- Repack font index after free
- Pass font shm name to clients
DBG and INF messages pollute the logs and performance reports
from valgrind, let's disable them at build-time for better
benchmarking.
Fix other compilation warnings and clean code a bit
- Send paths to all connected clients
- Send paths upon connection from a new client
- Client side:
- Fast find opened images
- Fast find loaded images
- Optimize load by skipping socket wait
Also, remap indexes when needed (file size changed)
When index file grows, the clients are not notified, so they
must constantly check that their mapping matches the current array
size.
- Move required struct definitions to common evas_cs2.h
- Open/close shm and map arrays to memory
Pretty stupid for now as it will close & reopen the shm files
even when unchanged.
- Check shm size and array header consistency
- Implement message send
- Send messages from server when calling repack()
In its current model, cserve2 will reopen, reload and scale an
image whenever a new scaled version is requested by the client.
Instead, we can load the original image when too many scaled
versions are requested, and place the original in the LRU.
Remove Image_Load_Opts and use Evas_Image_Load_Opts for cs2 master to
slave communication.
2 things to note:
- The struct size is bigger (but at the same time we removed a memcpy)
- The scale_load field is not passed to the slave
When opening an image, OPEN and SETOPTS were always sent
consecutively. Merging these two messages may improve the
performance a bit (not measured), but also simplify the
whole thing.
Note: cserve2 debug tools have not been fixed yet.
Some hashtable lookups failed because the keys were added
either with eina_stringshare_add() or e_s_add_lenth() but
looked for with e_s_add_length() only.
Signed-off-by: Cedric Bail <cedric.bail@samsung.com>
glibc double free happens when shutting down cserve, while
requests are being processed. this is because the entry is
being deleted from the hash but the cancel request fails
(so it should be deleted again from the hash)
Another patch going to fix the double free.
The _request_failed() function is called by error responses from slaves,
and iterating over references of a entry and removing each of them must
be done with EINA_LIST_FOREACH_SAFE(), since _entry_free_cb() calls
_entry_reference_del() which then removes the reference that is used in
the next iteration in for-loop from _request_failed().
Signed-off-by: Paulo Alcantara <pcacjr@profusion.mobi>
SVN revision: 81580
I've tested it and I can confirm that it's working fine with the new
scalecache support on Cserve2. So, let's use it again.
Signed-off-by: Paulo Alcantara <pcacjr@profusion.mobi>
Patch by: Paulo Alcantara <pcacjr@profusion.mobi>
SVN revision: 80177
I've tested make -j 3 install and it works nicely
I've tested expedite with software and opengl xlib,
and it works. Not tested other engines, so please
report any problems (engines or other) on the ML.
TODO: examples and tests, I'll add them later
ISSUE: Eina_Unicode size check. It indirectly depends on
eina_config.h, which is created at the end of the
configure script. So its size is always 0. I don't
know how that size is used, so I can't do a lot,
for now.
SVN revision: 78895
Jean-Philippe Andre
Carsten Haitzler
Paulo Alcantara
Vincent Torri