External app crash

[parisseb]

Describe the bug

I'm trying to add sha256+rsa2048 signature authentification to the KhiCAS launcher external app (like I did for the Casio Graph 90/fxcg50)
https://www-fourier.univ-grenoble-alpes.fr/~parisse/numworks/launch.tgz
but if I enable signature check then there is a crash (reset) immediatly after the end of check_signature.

To Reproduce

Build the launcher app and install khicas.nwa on your calc, it runs OK (assuming KhiCAS is installed in slot B).
Enable signature check by uncommenting #define WITH_SIGNATURE at the beginning of src/main.cpp, uncomment #define DEBUG as well to see the debug messages, build again, now it crashes immediatly after leaving the function check_signature.
Observe that the crash does not happen during crypto operations, the values displayed are correct inside check_signature. These operations should work OK since it's the same code as on the Casio (sha code is from MicroPython, and I tried with 2 big integers libraries, tommath and GMP, GMP is working since it is used inside KhiCAS).
But between the return instruction in check_signature and the next instruction in main, it crashes and resets. I have no idea what's happening and how to further investigate the issue. Could it be a stack overflow? Or something related to the "unusual" link process?

Environment

N0120 device with Epsilon 24 beta

[yannis300307]

I don't know if my issue is the same as yours, but when I transfer an app to my N0120 with Epsilon 23.2.6, it works the first time, and the app works properly. But when I update the app by transferring a new version, the transfer does work, but the entire calculator crashes and reboots when I run the app again. This issue happens with every program I tried. I don't think that it's an issue with the program itself because it should not have memory issues as I tested it with a Rust app. Also, I don't have this issue with my N0110 with Epsilon 23.2.3.

[parisseb]

I experience something similar when I upgrade KhiCAS without upgrading the launcher, a crash/reset happens, I send again the same launcher app and everything works again. I have no idea why this happens.
We have 3 different unexplained crashes, I believe it's now time for Numworks to give us their advice! And also give us a roadmap for improved external app support, we are still waiting for an Epsilon that does not require reinstalling an external app each time a crash/reset/exam mode cycle happens. Nobody except the author and maybe a few enthusiasts will use external app with the current constraints.

[parisseb]

Is there some caching mechanism for the external flash? If so and if the cache is not invalidated after a program is flashed on the Numworks, this could explain some crashes.

[parisseb]

BTW, I fixed one bug in my code, external app authentification is now working.
But the other independant reported crashs remain, and might be related to cache.

[Yaya-Cout]

There are two caches: ICache (for instructions, but I didn't find any function call to the function used to clear it) and DCache (for data). DCache is cleared after writing data, like the sync command on Linux). It's cleared after writing exam mode state (https://github.com/numworks/epsilon/blob/version-18/ion/src/device/kernel/drivers/persisting_bytes.cpp#L76) and on older versions before jumping to the DFU subfirmware after copying it to RAM (https://github.com/UpsilonNumworks/Upsilon/blob/upsilon-dev/ion/src/device/shared/usb/dfu_relocated.cpp#L58).

Unfortunately, this function is in the kernel in newer version, so I don't have the code to check. I couldn't find any SVC call to call this function from the userland/external app (https://github.com/numworks/epsilon/blob/master/ion/src/device/shared/drivers/svcall.h).

The cache doesn't seems to be cleared after DFU on Omega/Upsilon, so it's perhaps not the source of your bug.

[adri1numworks]

For now, there won't be a version which does not remove external apps after a reset, crash or exam mode. This is not only for technical reasons but also in order to ensure a safe exam mode, which cannot be bypassed.

This being said, we are improving our tutorials for the installation of third-party apps. From V24, you'll also be able to debug your external apps directly on the simulator. This should greatly improve your experience!

[yannis300307]

I'm really happy that you've added the app debugging in the simulator. Thanks.

However, I think that my crash is due to the OS or the hardware and not my app. Maybe it's a temporary bug in that specific version, and I should update it to the latest beta version.

[adri1numworks]

We'll check this also! @HugoNumworks you have any idea?

[parisseb]

For now, there won't be a version which does not remove external apps after a reset, crash or exam mode. This is not only for technical reasons but also in order to ensure a safe exam mode, which cannot be bypassed.

I don't understand your point : External apps are disabled in exam mode.
Removing external apps if the user press the RESET button is meaningfull for countries that do not enforce exam mode, it's an easy way to make sure that the user has no access to unofficial content. But removing external apps after a crash does not enter in this category: it's just the best way to discourage users to run 3rd party external apps. Quoting EmilieNumworks last year in #2167
"
The NumWorks team agrees with your assessment: some features need to be improved to facilitate the development of external apps.
We already have identified and included in our development trajectory two items that you mention:

Having different behaviors to handle an intentional reset (by pressing the back button) and a reset due to a crash. Erasing external apps following an intentional reset is feature we need to comply with some exam board requirements. However after a crash, we could keep the external apps as you suggest.
A better environment to develop external applications: an up-to-date tutorial, the ability to run external apps on simulators...
We intend to develop both soon!

"

At my University in the course I teach (2nd year students in physics), we allow calculators and do not enforce exam mode, so that students can run KhiCAS to check their answers or have access to some features that are usually not available on calculators (like Frenet frame, osculating circle, evolute, slopefields...) but are available in KhiCAS. We lend Casio Graph 90 for the students that do not have a compatible calculator. I first told students owning a Numworks that they did not need to switch to the Casio, they could run KhiCAS as an external app. But I won't again, because during an exam, they are clearly disadvantaged, if KhiCAS crashed on their Numworks, they lost all their data and moreover they lost access to KhiCAS. On the Casio, if a crash happens, the KhiCAS addin is still available and also the last session state that was saved.

BTW, KhiCAS is now compatible with exam mode on the Casio Graph 90, 35eii and on the ti83ce. It is also on old unlocked N0110, with a certification mechanism that can insure someone else that they are running the official Khi/KhiCAS version, cf. https://www-fourier.univ-grenoble-alpes.fr/~parisse/numworks/nws.html#exam). We are not trying to cheat,
I'm trying to give all students access to CAS where it's allowed in exams, and not to keep it reserved for students who can afford a TI Nspire CX2 CAS or HP Prime or Casio Classpad. That's the reason why I have added sha256/rsa2048 strong authentification for KhiCAS as an external app (now working), I did all the implementation work I can do myself, because I believe KhiCAS should also be available to N0115/N0120 users in France. Now, I can't do more, it's up to Numworks to authenticate KhiCAS in an exam mode designed for France. If Numworks refuses to do so, it's their legal right of course, but Numworks would be morally accountable of this, not me.

[yannis300307]

I have more info about my issue. To test if this is a memory issue, I wrote a little Python program that fill the memory with null bytes for i in range(10000): a=i*"\0" and it does fix my crash. Each time I transfer my app, I run this code, and it doesn't crash when I start the app. Maybe my program lets data in the ram that is not properly cleared by Epsilon.

Edit : Not all the time, finally...

[adri1numworks]

I think it's useful to clarify this discussion because there are actually 4 topics in one.

First, retrieving the external apps after a reset is not something we want to do because the reset is explicitly asked by the user. It'd be absurd not to do a real reset while one is requested.

Second, after a crash, users want to retrieve its external apps. This is something we'd like to improve but we currently do not have a satisfying solution to offer.

Third, for the external apps being retrieved after an exam mode, this is a similar issue to the second one. We don't have a satisfying solution to offer.

Fourth, and final, for the topic of keeping the external apps during an exam: NumWorks will not sign external apps because we are legally accountable for the safety of our exam mode. We will not keep non-official apps during exam mode.

[parisseb]

I'm surprised about point 4 for France, I don't know for other countries of course. I continue in French for France.

Je ne vois pas bien ce qui empêche de discuter avec la dgesco et d'autoriser KhiCAS pour un mode examen spécifique à la France, le logiciel étant placé sous ma responsabilité : je suis fonctionnaire (en tant qu'enseignant-chercheur, il m'est arrivé de présider un jury de bac), mon logiciel Giac est présent dans les concours de recrutement de l'Education Nationale (c'est le moteur CAS de Geogebra), et également sur les HP Prime, calculatrice autorisée aux examens en France.
On peut imaginer que cela soit réservé à un modèle vendu uniquement en France (pour éviter les tentatives de fraude dans des pays qui interdisent le CAS). TI vend bien la ti83 et 82 spécifiques à la France, Casio les 35eii/90/math+ et d'ailleurs Numworks a déjà fait ça pour le Portugal, alors que la taille du marché est bien inférieure à la France, environ 8 fois moins d'effectifs pour la classe d'age lycée. Numworks vend actuellement 2 modèles de calculatrice, la n0115 et la n0120, sans qu'on comprenne bien à qui est destiné l'une ou l'autre. Or vu les ventes sur amazon, il est assez clair que les ventes de Numworks se font majoritairement en France, et avec plus de 30 000 ventes rien que sur amazon.fr pour 2024/25, il y aurait largement de quoi justifier un modèle spécifique à la France.

Pour info, je viens de rédiger des indications d'utilisation de KhiCAS sur les 3 sujets de bac 2025 déjà parus (centres étrangers), pour aider les élèves à vérifier leurs calculs, en calcul formel mais aussi en géométrie analytique 3d:
https://tiplanet.org/forum/viewtopic.php?f=51&t=27051&p=278034#p278032
Sans KhiCAS, les calculatrices graphiques milieu de gamme apportent une aide sur environ 50% des 3 sujets (l'exercice de proba-stats et une partie de l'exercice suite récurrente, plus un peu d'aide sur une représentation graphique dans l'exo d'analyse). En ajoutant KhiCAS sur les Graph 90 ou sur les Numworks N0110 avec Epsilon<=15.5, on couvre 100% du sujet. On pourrait faire de même sur les N0110 avec Epsilon>=16, les N0115 et les N0120, mais pour ça il faut que KhiCAS fonctionne aussi en mode examen! Il y va aussi de l'équité entre candidats au bac, d'ailleurs cité dans le premier article du code de l'Education https://www.legifrance.gouv.fr/codes/article_lc/LEGIARTI000043982767.
Au passage, on observe sur ces 3 sujets que seuls deux d'entre eux ont une question d'algorithmique, facilement résolue sans utiliser l'appli Python, en déphasage assez étonnant entre l'utilisation marketing et ce qui en est réellement fait en examen, on verra si cela se confirme pour les autres sujets de cette année.

J'arrête là, si vous souhaitez en discuter en privé, n'hésitez pas à me contacter par email.

[HugoNumworks]

@yannis300307

when I transfer an app to my N0120 with Epsilon 23.2.6, it works the first time, and the app works properly. But when I update the app by transferring a new version, the transfer does work, but the entire calculator crashes and reboots when I run the app again.

I tried from a 23.2.6, N0120, installed a sample app (a version of https://github.com/numworks/epsilon-sample-app-c), and ran it.
Then I went back on the Home menu, and installed a slightly edited version of it.
It ran without issue.

Is there a step I missed ? On which app have you experienced the issue ?

[yannis300307]

@HugoNumworks
It happens on my Numcraft project. This project uses a lot of ram. I created it using the Rust sample app template. I'm not sure what exactly is the problem but when I transfer multiple times the exact same nwa using nwlink or the website, it works properly but the crash happens when I change something (even a single value). So it could be something related to a hash or cashing. For more details, I power off and wake up the calculator between each transfer to reopen the Connection menu.

[parisseb]

This is not related to nwlink, I get similar crashes if I link external apps without nwlink (by the way, nwlink does not spare flash, for the KhiCAS launcher the size is reduced by a factor greater than 2 if I link without nwlink).
My best guess is that it's related to wrong cache, perhaps leaving DFU does not invalidate some cache.