At Wed, 17 Jul 2013 09:30:21 +0930, Rusty Russell wrote:
Lucas De Marchi lucas.de.marchi@gmail.com writes:
On Tue, Jul 16, 2013 at 5:28 AM, Philipp Hahn pmhahn@pmhahn.de wrote:
Hello,
Am Dienstag 16 Juli 2013, 08:43:36 schrieb Takashi Iwai:
At Tue, 16 Jul 2013 15:11:51 +0930, Rusty Russell wrote:
Philipp Matthias Hahn pmhahn@pmhahn.de writes:
My x86_64 systems has some trouble loading some ALSA snd-* modules since the upgrade to 3.10.[01]: Several automatic modprobe calls hang, but loading snd-intel-hda and snd-audio-usb by hand still works.
Not a known problem to me, at least. Perhaps it's a dep loop somehow.
I remember that someone reported it being Debian specific snd-seq-oss loading stuff.
FYI: "oss-compat" is installed.
... 1071 ? S 0:00 sh -c /sbin/modprobe --ignore-install snd && { /sbin/modprobe --quiet snd 1080 ? D 0:00 _ /sbin/modprobe --quiet snd-seq
This was first, and it's waiting. Which means it must be doing something weird, because snd_seq_oss is loading now:
snd_seq_oss 33717 1 - Loading 0xffffffffa041b000
Perhaps in the tangle of modprobe install commands somewhere this gets invoked?
Likely, but I wonder what triggered a bug suddenly on 3.10. There is absolutely no change in sound/core/seq/*, and through a quick look, I couldn't find any suspicious change in kernel/module.c that may lead to this problem between 3.9 and 3.10.
Philipp, can you get a sysrq-T trace while the stall?
I've finally been able to reducte the number of processes to get a full trace; see attached file.
Please note that in this case the proprietary "nvidia" module was loaded, since I currently onyl have remove access to the machine. The original trace from yesterday happend without the nvidia module ever being loaded.
Am Dienstag 16 Juli 2013, 08:42:35 schrieb Lucas De Marchi:
On Tue, Jul 16, 2013 at 2:41 AM, Rusty Russell rusty@rustcorp.com.au wrote: First thing to check is the /etc/modprobe.d/*.conf file that contains these install commands. Did they change besides the kernel upgrade?
Not that I know of: Booting 3.9.9 works fine, 3.10.x shows the problem.
Then one possible explanation is that in 3.9.9 you don't have some of the modules causing this problem
...
Philipp, which kernel are you upgrading from?
I just upgraded from 3.9.9 to 3.10.0; also tried 3.10.1 which did not improve the situation.
I don't see anything to blame in the changes for the past releases. Any chance a bad entry in your .conf was added too? You may want to paste the output of modprobe -c, at least until "# End of configuration files. Dumping indexes now:"
blacklist snd_pcsp blacklist arkfb blacklist aty128fb blacklist atyfb blacklist radeonfb blacklist cirrusfb blacklist cyber2000fb blacklist gx1fb blacklist gxfb blacklist kyrofb blacklist matroxfb_base blacklist mb862xxfb blacklist neofb blacklist nvidiafb blacklist pm2fb blacklist pm3fb blacklist s3fb blacklist savagefb blacklist sisfb blacklist tdfxfb blacklist tridentfb blacklist viafb blacklist vt8623fb blacklist garmin_gps blacklist nouveau install binfmt_0000 /bin/true install sound_slot_0 /sbin/modprobe snd-card-0 install sound_slot_1 /sbin/modprobe snd-card-1 install sound_slot_2 /sbin/modprobe snd-card-2 install sound_slot_3 /sbin/modprobe snd-card-3 install sound_slot_4 /sbin/modprobe snd-card-4 install sound_slot_5 /sbin/modprobe snd-card-5 install sound_slot_6 /sbin/modprobe snd-card-6 install sound_slot_7 /sbin/modprobe snd-card-7 install snd /sbin/modprobe --ignore-install snd && { /sbin/modprobe --quiet snd-ioctl32 ; /sbin/modprobe --quiet snd-seq ; : ; } install snd_rawmidi /sbin/modprobe --ignore-install snd-rawmidi && { /sbin/modprobe --quiet snd-seq-midi ; : ; } install snd_emu10k1 /sbin/modprobe --ignore-install snd-emu10k1 && { /sbin/modprobe --quiet snd-emu10k1-synth ; : ; } alias net_pf_16_proto_1 wire alias net_pf_16_proto_3 ip_queue alias net_pf_16_proto_8 scsi_transport_iscsi alias net_pf_16_proto_9 audit alias net_pf_16_proto_11 cn alias net_pf_16_proto_13 ip6_queue alias binfmt_204 binfmt_aout alias binfmt_263 binfmt_aout alias binfmt_264 binfmt_aout alias binfmt_267 binfmt_aout alias binfmt_387 binfmt_aout alias block_major_3_* ide_generic alias block_major_22_* ide_generic alias block_major_33_* ide_generic alias block_major_34_* ide_generic alias block_major_37_* ide_tape alias block_major_44_* ftl alias block_major_46_* pcd alias block_major_47_* pf alias block_major_56_* ide_generic alias block_major_57_* ide_generic alias block_major_58_* lvm_mod alias block_major_88_* ide_generic alias block_major_89_* ide_generic alias block_major_90_* ide_generic alias block_major_91_* ide_generic alias block_major_93_* nftl alias block_major_97_* pg alias char_major_10_1 psmouse alias char_major_10_139 openprom alias char_major_10_157 applicom alias char_major_10_181 toshiba alias char_major_10_183 hw_random alias char_major_10_187 irnet alias char_major_10_189 ussp alias char_major_10_250 hci_vhci alias char_major_13_0 joydev alias char_major_13_1 joydev alias char_major_13_2 joydev alias char_major_13_3 joydev alias char_major_13_32 mousedev alias char_major_13_33 mousedev alias char_major_13_34 mousedev alias char_major_13_35 mousedev alias char_major_13_63 mousedev alias char_major_13_64 evdev alias char_major_13_65 evdev alias char_major_13_66 evdev alias char_major_13_67 evdev alias char_major_19_* cyclades alias char_major_20_* cyclades alias char_major_22_* pcxx alias char_major_23_* pcxx alias char_major_27_* ftape alias char_major_34_* scc alias char_major_35_* tclmidi alias char_major_48_* riscom8 alias char_major_49_* riscom8 alias char_major_57_* esp alias char_major_58_* esp alias char_major_63_* kdebug alias char_major_67_* coda alias char_major_75_* specialix alias char_major_76_* specialix alias char_major_81_* videodev alias char_major_83_* vtx alias char_major_89_* i2c_dev alias char_major_90_* mtdchar alias char_major_96_* pt alias char_major_97_* pg alias char_major_107_* 3dfx alias char_major_109_* lvm_mod alias char_major_166_* cdc_acm alias char_major_171_0 raw1394 alias char_major_171_1 video1394 alias char_major_171_2 dv1394 alias char_major_171_3 amdtp alias char_major_180_* usbcore alias char_major_195_* nvidia alias char_major_200_* vxspec alias char_major_202_* msr alias char_major_203_* cpuid alias char_major_206_* osst alias char_major_208_* ussp alias char_major_227_* tub3270 alias bt_proto_7 avdtp alias cipcb0 cipcb alias cipcb1 cipcb alias cipcb2 cipcb alias cipcb3 cipcb alias dummy0 dummy alias dummy1 dummy alias plip0 plip alias plip1 plip alias slip0 slip alias slip1 slip alias tunl0 ipip alias gre0 ip_gre alias usbdevfs usbcore alias char_major_195* nvidia options snd_pcsp index=-2 options snd_usb_audio index=-2 options bt87x index=-2 options cx88_alsa index=-2 options snd_atiixp_modem index=-2 options snd_intel8x0m index=-2 options snd_via82xx_modem index=-2 options snd_hda_intel model=6stack-dig index=0 options snd_usb_audio index=1 options dvb_ttpci adapter_nr=1 options budget_ci adapter_nr=0 options nbd max_part=15 options nvidia NVreg_DeviceFileUID=0 NVreg_DeviceFileGID=44 NVreg_DeviceFileMode=0660 options libata force=noncq options systemd log_target=kmsg softdep uhci_hcd pre: ehci-hcd softdep ohci_hcd pre: ehci-hcd softdep snd_pcm post: snd-pcm-oss softdep snd_mixer post: snd-mixer-oss softdep snd_seq post: snd-seq-midi snd-seq-oss
hum... it looks like a loop between the modprobe calls and the request_module(), done in snd_seq's init. The request_module() call will end up trying to load snd_seq again and it will remain waiting on kernel/module.c:add_unformed_module().
In kmod we don't trust a COMING state on sysfs to avoid calling (f)init_module() because a) the previous call may fail and b) it's racy.
Yes, add_unformed_module makes the same call. I think the answer is simply "don't do that".
Changing the request_module() in the module to be async would solve the problem (what Takashi Iwai did)... but this has been a little controversial in the past. Rusty, what do you think? Maybe in kmod we can take the COMING state into consideration for *dependencies*? Or is moving that call to a worker acceptable?
I thought about adding a post_init() call for modules for this kind of thing, but it's actually pretty rare. Open-coding it seems fine.
The conversion to async call would be a problem if it needs to change the code flow largely. In this particular case, however, it's no big problem (as little as the patch size is), so I'm going to apply my patch as is for now.
thanks,
Takashi