NetSurf 3.11 has been released!

@Chris

Quote:

Now that we know that bsdsocket.libary and Roadshow “beta” cause these problems I'm not sure if debug protocols are useful for them anymore.

I have entered it as you have explained it so that I can rule out any errors on my part, I will leave you a screenshot.




log:

The last site I visited before the problems started was Amigans, you can see it in the log...otherwise NetSurf is stable there have been no crashes.

@salass00
Quote:At least 15-20 years ago, when I was still working on the AmigaOS port of OWB, the pthreads based threaded resolver didn't work at all.
IIRC I used libadns for OWB instead.

No idea about Odyssey as that's something completely different, it's not based on the AmigaOS 4.x port of OWB but on the the MorphOS/AROS ones instead.

Before libcurl 7.85, I was using a patch for hostip4.c that kas1e sent me (although I thought he got it from Joerg?), that used the method salass00 describes in post #160.

Since 7.85 with Futaura's patches for AmiSSL v5 in the official repo, that patch is in amigaos.c, but with the difference that it will try to use gethostbyname_r if it's available. It should be easy to just bypass that check or return false.

@joerg

I was referring to the threaded resolver in libcurl. I was (same as MickJT) under the impression that this was your work, or at least based on it.

I just added getaddrinfo() support to my test program and it causes an immediate DSI crash in bsdsocket.library.

What I'm currently finding most puzzling about it is that the stack trace goes directly from my program to bsdsocket.library even though I know that it's being called through the wrapper function in newlib.library (I added some debug output there before and after), and the newlib function is definitely creating a stack frame and it is using bctrl, which updates the link register, to call bsdsocket (I checked both of these things by disassembling the object code generated by gcc).

Edit: If I get Ranger to do a stack trace of the crashed process it correctly shows the newlib.library entry in between. Not sure why it isn't showing in the Grim Reaper stack trace.

My changelog on OS4Depot for libcurl credits both kas1e & Joerg for the threaded resolver fix in my 7.46.0 port back in January 2016. My recollection is that kas1e credited Joerg when he sent it to me. I don't know how long he had it for. The Odyssey changelog mentions a threaded resolver back in 2010 (from OWB history it seems), but I don't know what implementation that would have been.

@MickJT
Yeah, it was Joerg explaining me how to do so and how he did in his OWB years ago, and i made the same in Odyssey as well (so every libcurl i port for usage with Odyssey had this change). I remember i even tested newlib/roadshow way, and it werent working correctly, so i wrote about to Olaf, and while i have answer, nothing come up from as i can see from this thread

@kas1e

I can't remember if it was MUI-OWB or Odyssey back then (2016). I think Odyssey.

Edit: You had issues? I vaguely remember a time from before/after there was a threaded resolver, but not issues with it. My memory could be wrong though! Maybe it wasn't using a threaded resolver and I only thought it was?

@MickJT
Yes there were same issues with Odyssey too (of course not Odyssey per se, but curl+os4 differences), but not sure if i release those broken versions (but if, i then fix it fast, thanks to Joerg)

@kas1e

https://forum.hyperion-entertainment.c ... 0fdb89d2063434923dd56b8a3

I found that thread but it's from November 2012, so probably before you had the hostip4.c fix (now part of amigaos.c and used when gethostbyname_r is unavailable).

Edit: So are you saying it was ok after the hostip4.c fix? Or you still had issues? I should have clarified that I only meant after modifying hostip4.c (back then).

@MickJT
Last time i build curl for latest Odyssey, were enough to fix hostip4.c, but i use OpenSSL which port myself too, not AmiSSL. The fix in hostip4.c just:



Then also at the end:



And that was all. (Also file.c as always a bit changed, but that not related to threaded resolver).

See there for example: https://github.com/kas1e/Odyssey/blob/master/SDK.txt

I even noted there:

curl - v7.68.0 added amigaos4 native threading resolver, fixed file:// part

I.e. those hostip4.c and file.c changes

@salass00

Quote:

After writing my gethostbyname_r() fallback implementation and noting to my surprise that it exhibits the same flaw I suddenly realized that the erroneous reported IPs are down to my bad test code and not bsdsocket.library.

The reason is that I use inet_ntoa() to convert the IPs to string format which is not thread safe (the gethostbyname() test works fine because I use isock->INet_NtoA() there).

This still doesn't explain why gethostbyname_r() hangs though, or why getaddrinfo() crashes.

@kas1e

And that fix has worked fine? That's the one I was using before libcurl 7.85 (when that fix became part of the official source, although only when gethostbyname_r is not available in bsdsocket.library).

libcurl 7.68.0 came out in January 2020, but you'd sent me that hostip4.c file sometime in or before January 2016 when I used it for libcurl 7.46.0

@salass00

So there's definitely still a problem with gethostbyname_r then. I'll make sure to avoid it in the future.

I just noticed to my surprise that I also have the same 4.321 version of bsdsocket.library on my Sam460 beta installation. It exhibits the same random hanging in gethostbyname_r() and crashing in getaddrinfo() so these problems are not Tabor/A1222 specific.

Given that this version 4.321 is quite old (2016) it is possible that these problems are already fixed in the latest roadshow.