From Linux PARISC Wiki
(Difference between revisions)
Jump to: navigation, search
m (Status & TODO 2015-12-25)
m (Status & TODO 2015-12-26)
Line 13: Line 13:
# glibc -- update patches for 2.21
# glibc -- update patches for 2.21
# libsoup2.4 -- testsuite error
# libsoup2.4 -- testsuite error
# nss - problem with glibc dladdr() - see:
# cupt
# cupt
# libqb - testsuite error
# libqb - testsuite error

Revision as of 08:03, 30 December 2015


PA-RISC Linux TODO (actually this is Helge's TODO list)

Status & TODO 2015-12-26

  1. STARTED: vlc - linker does not drop dependencies to unneeded shared libs,, problem arises due to __gmon_start__ symbol which then wrongly gets resolved e.g. by other shared libs. Fix for glibc is reported:
  2. elfutils -- testsuite error
  3. glib2.0 -- testsuite error
  4. gtk+3.0 -- testsuite error
  5. glibc -- update patches for 2.21
  6. libsoup2.4 -- testsuite error
  7. nss - problem with glibc dladdr() - see:
  8. cupt
  9. libqb - testsuite error
  10. capnproto - testsuite errors: 1) nan != nan ->quiet vs. signalling nans"? __builtin_nanf("") returns 0x7fbfffff, 2) info._sifields._rt.si_sigval.sival_ptr -> hppa behaves correctly. Testsuite is wrong.
  11. vtk6 - dependency issues (other platforms have same problem)
  • glibc: Ask to change the shm_nattch member of shmid64_ds from long to int (for 64bit userspace)
  • convert irq handling to use CONFIG_IRQ_DOMAIN_HIERARCHY (serial port problem, action handler msising when registering irq)
  • Due to very bad performance running multiple copies of dot program (Dave):
    • Another issue that I have been wondering about is the performance of our kernel spinlock code. PA 8800 and PA 8900 processors have a cache line length of 128 bytes. I believe thereare some places where we assume 64 bytes. We also allocate spinlocks on 16 byte boundaries. We might get better performance if we allocated on 32 (PA 1.X) and 128 (PA 2.0). Helge: change: __PA_LDCW_ALIGNMENT
    • I've also seen that better performance may be obtained by doing a byte store to spinlock prior to ldcw. This dirties the cache line and allows operation in cache in certain situations. I think this mainly applies to PA 1.X but I'm not sure. I have some example code somewhere. Helge: change struct arch_spinlock_t; in spinlock_types.h
  • MOSTLY DONE: push gcc patch for variable CAS (gcc atomic ops, __sync_XXX) upstream (Kernel 3.17, - packages needing rebuild: libatomic-ops (gauche and libgc have copies of libatomic-ops included), x265, ceph, lizardfs, firebird2.5, gauche, idba, mapsembler2, padevchooser, filezilla, idba,
  • add wiki page about kernel cross-compiling
  • Kernel: still missing Linux kernel development (no patches available yet):
    • Fix FTRACE:
    • optimize flush_dcache_page() implementation like sparc. See Documentation/cachetlb.txt, e.g. used in aio subsystem
    • fix wrong _init notation in kernel
    • NO_IRQ -> should be converted to (0) instead of (-1) -> Helge (check assembly, check in drivers/parisc/*)
    • implement alternative coding for UP/SMP. See arch/x86/include/asm/alternative.h
    • check assembly in drivers/parisc/*
    • use pdc_add_valid() in kern_addr_valid(), protect against crashable acesses (like for superio-pc-drivers-detection in lp), add pdc_add_valid() to create a probe_kernel_read() implementation.
    • arch/parisc/kernel/sys_parisc.c -> use SYSCALL_ALIAS() -> maybe
    • on C8000, report 1 socket with 2 cores.
    • Review gettimeofday and clock_gettime() implementation, possibly use a light-weight-syscall and rely on cr16 and cpu speed for more accurate timing? similiar to commit
  • Kernel: Low priority patches needs further work
    • Fix exception table entries to be of type int (32bit), even on 64bit kernel -> saves space. - - partly done.
    • disable FP-register save/restore at kernel entry? -
    • enable 16k/64k kernel page sizes (still cacheflushing-issues with SCSI driver, additional work in glibc needed to deal with >4k page size, e.g. ld-loader alignments of data/code segments)
    • Superdome support - can someone test? Should work better now with latest kernels.
    • task_struct/thread_info split -- task_struct should not be visible in entry.S, we need to move some items into thread_info -- this includes pt_regs and maybe some of the flags (ptrace, etc)
    • Visualize FX support (will probably never happen!)
  • PROBABLY NOT POSSIBLE: implement BUG() and WARN() so that interrupt happens in delay branch of "or,<> %r0,cond,%r0". Not possible, because this gives "control reaches end of non-void function" warnings/errors.
  • Debian packages / Userspace work
    • IMPORTANT: fix OpenJDK - fixes required in zeroport (stack grows up!).
    • 64-bit userspace, 64-bit binutils needs to be fixed to get multiple stub section support.
    • gcc -> join 32- and 64bit compiler, make "-m64" work, get rid of hppa64-linux-gnu-gcc
  • palo (PA-RISC Boot Loader):
    • Modify (if possible), that palo can read (sequential) tftpbootable-vmlinuxes as well as the current boot-vmlinux images (simplifies build/handling)
    • move decompression of Kernel to Linux kernel code instead of palo

Kernel patches

Debian kernel and userspace programs

  • Debian kernel
    • DONE: send patches to remove UP from debian kernels, use sil680 instead of siimage, see -> fixed in 3.12 / 3.13
    • DONE: Enable BMC in c8000 default kernel config - ipmi driver (CONFIG_IPMI_HANDLER=y) -
    • DONE: remove HIL RTC from debian kernel - it creates a backtrace at boot because we have already another rtc0 active.
    • DONE: add ppdev driver, needed by CUPS, systemd reports missing module load (3.17 & 3.16 kernel)
    • DONE: Add crc-modules, event, isofs, jfs, mouse, sata, sound, squashfs, udf, uinput, scsi-common, scsi-extra, loop, serial and usb-serial debian-installer packages
    • DONE: Kernel build problem: a) disable -mfast-indirect-calls in Makefile, b) gcc compiler bug in > gcc-4.8.3-3
    • DONE: Drop HPUX-support from kernel (Kernel 4.0)

ABI breakages (HP-UX support) Homepage and Debian build server infrastructure work

Updates for next newsletter

  • fixed 64bit SMP kernel & sticon on machines with more than 4GB RAM
  • Dave found O_SYNC bug in glibc (Typo in hppa __O_SYNC define in <bits/fcntl.h>)
  • lots of syscall fixes regarding compat kernel
  • Migration
  • access
  • 3.13 with all patches in (coherency)
  • buildds running
  • wiki migration finished (05 Feb 2014)
  • old parisc-linux kernel mailinglist imported into - - 23 Nov 2014
  • parisc-linux domain is now owned by Helge (August 2014)
  • Debian 8.0 installer images available.

Testsuites, kernel and userspace testing tools

  • GLIBC issues:
  • We also have a problem with poll in rpcbind on 32-bit kernels. Runs forever. Not analyzed yet.
  • glib2.0 atomic bug: Dave: I see the problem building various packages -- apt is is one. It occurs when multiple dot processes are started by graph when package documentation is being generated. The number of dot processes spawned depends on the number of processors on the build system. However, the faults occur even on a UP system. I debugged one of these segvs enough to see that the problem occurred in a glib2.0 library. It's using the GCC "sync" interface.
  • From Dave: gcc thread issues including thread cancellation, glib2.0 problem, maybe __tls_get_addr clobber too?

TODO Drivers

  • IPMI support on rp-class / Support graceful shutdown from BMC on rp servers (`PC -GRACEFUL`)
  • write Lasi floppy driver
  • write Suckyio floppy driver
  • write spifi driver (rbrad)
  • modify ncr53c8xx driver for Outfield (735 & 755)
  • write GSC FDDI driver
  • write Timi ASIC (74x) support
  • EISA DMA support

Review all the todo entries below!

  • ns87415 dma doesn't work reliably on suckyio-systems
  • investigate not putting in extable entries for put_kernel_asm; will probably reduce kernel size
Personal tools