As of Linux 6.5.2-arch1-1 real addressing seems broken

[VorpalBlade]

For some reason, it seems sysctl vm.mmap_min_addr=0 is no longer enough to make SheepShaver work on kernel 6.5.2-arch1-1 (from Arch Linux).

I now keep getting ERROR: Cannot map Low Memory Globals: Cannot allocate memory. though this worked fine yesterday.

In addition it segfaults if I try to build with direct addressing mode instead (I will investigate that and possibly open a separate bug about it).

[VorpalBlade]

Looking further into this it seems to happen because of this code:

#if REAL_ADDRESSING && HAVE_LINKER_SCRIPT
    if (vm_mac_acquire_fixed(0, RAMSize) == 0) {
        D(bug("Could allocate RAM from 0x0000\n"));
        RAMBase = 0;
        RAMBaseHost = Mac2HostAddr(RAMBase);
        memory_mapped_from_zero = true;
    }
#endif
    if (!memory_mapped_from_zero) {
#ifndef PAGEZERO_HACK
        // Create Low Memory area (0x0000..0x3000)
        if (vm_mac_acquire_fixed(0, 0x3000) < 0) {
            sprintf(str, GetString(STR_LOW_MEM_MMAP_ERR), strerror(errno));
            ErrorAlert(str);
            goto quit;
        }
        lm_area_mapped = true;
#endif

It appears that on my system REAL_ADDRESSING and HAVE_LINKER_SCRIPT is true (so the first block is executed). That creates a mapping up to ram size successfully. However, the call to vm_protect in vm_alloc.cpp then fails for some odd reason. This makes the whole call fail. Which in turn causes the second branch (!memory_mapped_from_zero in main_unix.cpp) to be entered incorrectly.

[VorpalBlade]

I have created a bug report in Arch Linux against the arch kernel, but I will need to figure out if this is an upstream kernel regression also.

[JoshuaPettus]

Any word on this? It's doesn't seem distro dependent as it's still a problem in fedora 39 against kernel 6.6.12

[VorpalBlade]

It was a kernel regression. It was fixed in 6.5.7. I haven't used sheepshaver in a while now, are you saying it is broken again? If so, you should probably show it can be reproduced on an upstream kernel and report it.

[JoshuaPettus]

Unfortunately for me it was never fixed. It has been a problem since 6.5. I've been testing it periodically since. Perhaps whatever patch was made never made it into fedora?

[VorpalBlade]

I believe it was fixed upstream, and Arch Linux uses close to upstream kernels.

[thomas-covenant]

current ubuntu 22.04.03 is using 6.5.0-15-generic kernel and is not fixed. i try to use 6.5.11 and see if it is fixed. but that kernel version might not run with 22.04.3

edit: the later kernels 6.5.11 / 6.6 do not work and create an error when installing. so the issue remains. pity as sheepshaver is much better performing than qemu.

[mikemocha]

This is not only an Arch Linux problem but also occurring on other distros. I have the same problem on an Ubuntu 23.10 variant. uname -a output

6.5.0-14-generic #14-Ubuntu SMP PREEMPT_DYNAMIC Tue Nov 14 14:59:49 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux

[JoshuaPettus]

Yeah I can see that being a problem on more "stable" distros for a while.

I had found out a couple days ago that when I run Sheepshaver in terminal I no longer get the "ERROR: Cannot map Low Memory Globals: Cannot allocate memory." So maybe I did get the patch.

The issue is when I open sheepshaver I still get a message that I can't not load extensions because there is no virtual ram. Really odd as I don't think that was ever emulated properly in sheepshaver. I'm a little disappointed because I started having that issue when this issue popped up so I was sure it was the same problem.. I didn't make any other changes but the normal updates