Sound userspace drivers (fishing for insight)
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
This isn't really a serious question, but I just encountered some stuff about userspace sound drivers for Linux and was like O_o so now I'm curious as to the design aspect here. I can't think of a good way to do it (where 'good' means something more than 'ugly hack').
So here's my input on this, and if anyone cares to enlighten me maybe you can bounce some responses back. It'll be a good learning experience for me.
- General transport and context switches
In general, transporting sound data in and out of kernel space is a horrid thought. Consider the latencies, which all real-time audio people will quickly get angry about. Write sound; context switch; sound in driver; context switch back. This over and over? Now we all know better than that.
So my first impulse would be using the generic kernel-user transport facilities like NETLNK. But at a glance this doesn't seem to do anything for me; this would still be writing to a socket, which causes a syscall into kernel space. In short, you're stuck with context switches.
On the up-side, in any scheme the user space program has to shift data into the kernel; with alsa it's written to a /proc interface, which causes a context switch into kernel space AFAIK because it calls write() to an fd. Still, we're wasting time copying to another process' memory space.
- Ring-1, Ring-2
I have heard that these mysterious privilege levels allow bitmaps to give direct hardware access, i.e. to PCI devices. Through this mystical magic, a driver could directly access a sound card. This opens up new possibilities. . or does it?
First off how the heck complex is userspace PCI control, and what gains do you really get? Won't every syscall() become a sudden context switch? Or (as I suspect and would like to find out) does the "overhead of context switching" mainly imply the copy_from_user() function and the process of rewriting a lot of PTEs into a new privilege level (kernel instead of user)?
Assuming that you can implement direct hardware access or at least direct access to memory to avoid the context switch (in the manner I'm assuming the latency comes from), this may become actually faster. Still you're copying memory from Ring-3 (user) to Ring-1 or Ring-2, so there are no gains here.
- Shared mapping?
If my assumption on how this works is correct, then copy_from_user() is difficult because you have to walk the range of passed VM addresses and change/copy affected PTEs. The old PTEs become non-writable (copy-on-write trigger) and the new PTEs are created (non-writable CoW and of course Ring-3 privilege) for the kernel.
This assumption brings an interesting thought: what about a shared mapping? An area of memory writable by userspace and readable by the kernel? This seems silly. Under the assumptions I've made, this would still require CoWing existing PTEs or flat out copying things over. Obviously I lack understanding of the problem; none of this solves anything, and this is typical overhead whenever you allocate memory, even from userspace.
So we come to an interesting thought: I've found yet another interesting bit of kernel design that I don't understand, and am curious of. I understand basic memory management concepts and preemptive multitasking just because of pestering the LKML from time to time; might as well go for a little more and see if you guys bite again. Any takers?
- --
All content of all messages exchanged herein are left in the Public Domain, unless otherwise explicitly stated.
Creative brains are a valuable, limited resource. They shouldn't be wasted on re-inventing the wheel when there are so many fascinating new problems waiting out there.
-- Eric Steven Raymond
We will enslave their women, eat their children and rape their cattle!
-- Evil alien overlord from Blasto
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2.1 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
iQIVAwUBRA5XQws1xW0HCTEFAQLr7g//Yqnj5hIbANQR/SsiirkLhYAcM7DC3lGe vi50fu7YZSwaDNJ/NKWjHVmiuI1UWLehEe6n8BWHYR9tq0JzNzwwpEPUWUxIv7tp TsKKWR4uP8YYbFoFoNo8zK/kJQiV0ynVvuBCVCYm3OUqBi8hIqKut0GgyZc4qo7M aOmq3mFlw5zITfYV2pYVq9ttTU2ZJ36pXU4gJJi0+V5KVO51ZuAXDxDB6oaLewq7 csayDae/yy9Aene44fSSuO+bfS6JPNq9uTXS42v0bcYkZm+DppSTaZcPetj1cMcU qM1d/0jh6bQwVpDRb7A7xumP+TfQM7J8/5rR2TFTpgc63qA1Fyzu5qd/hJ9xz28A pum/36ZbF/fiLR5qwMB3tlwN9z6f1MrEmEzOKtCap/Mq3IklSnyc429mpTs8smjZ WzbRaLU2rNxgL2KFDOZvhrfN0PP4nu+DDSauTmYdaSHfdNQ0tlvZAapq8SMQy5LU VAY3ZDEMP4qAIGMfXTRR49JAuM0OLWpV/fXeE9Uwv0KIyomfw/TJvW20O2vblN06 Ryj4HGwELeaZpc+D8E6eWXSllX/16qp5zgXAFnNgvBAXfRnaKPzSnhNOdEozNhxM oFk3Kf9PoyEn3fqRFXJi+f9o02VvDAEY/6EOwsALUOAWADqbW+D/ku7MKEMt7pFT Jo/dyODrxfA=
=YKVj
-----END PGP SIGNATURE-----
Quoted 5 lines · John Richard Moser
In general, transporting sound data in and out of kernel space is a horrid thought. Consider the latencies, which all real-time audio people will quickly get angry about. Write sound; context switch; sound in driver; context switch back. This over and over? Now we all know better than that.
Nope. If you have a cache coherent architecture like x86, ALSA mmaps directly the DMA buffer to application and even pointers to this buffer are mmaped, thus there is no content switch when apps are doing r/w and there is "zero data copy". The only contents switch is when apps call poll(), but you should do it, otherwise you'll eat all cpu time.
Quoted 1 line
into the kernel; with alsa it's written to a /proc interface, which
Using /proc? Where? I've not noticed it :-)
Jaroslav
-----
Jaroslav Kysela <[email protected]>
Linux Kernel Sound Maintainer
ALSA Project, SUSE Labs
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Quoted 8 lines · Jaroslav Kysela
Quoted 5 lines · John Richard Moser
In general, transporting sound data in and out of kernel space is a horrid thought. Consider the latencies, which all real-time audio people will quickly get angry about. Write sound; context switch; sound in driver; context switch back. This over and over? Now we all know better than that.
Quoted 6 lines
Nope. If you have a cache coherent architecture like x86, ALSA mmaps directly the DMA buffer to application and even pointers to this buffer are mmaped, thus there is no content switch when apps are doing r/w and there is "zero data copy". The only contents switch is when apps call
poll(), but you should do it, otherwise you'll eat all cpu time.
Awesome. Nice design :) I hath learned today and I woke up 10 minutes ago.
Quoted 3 lines
Quoted 1 line
into the kernel; with alsa it's written to a /proc interface, which
Using /proc? Where? I've not noticed it :-)
I thought that was what /proc/asound/card0/ was for? :)
Quoted 8 lines
Jaroslav
-----
Jaroslav Kysela <[email protected]>
Linux Kernel Sound Maintainer
ALSA Project, SUSE Labs
- --
All content of all messages exchanged herein are left in the Public Domain, unless otherwise explicitly stated.
Creative brains are a valuable, limited resource. They shouldn't be wasted on re-inventing the wheel when there are so many fascinating new problems waiting out there.
-- Eric Steven Raymond
We will enslave their women, eat their children and rape their cattle!
-- Evil alien overlord from Blasto
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2.1 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
iQIVAwUBRA8DlQs1xW0HCTEFAQKeqg//SunhyEf8o4MGpQkrrWImJJ5HOaYFragk pgr++6lXTxNTJflpKRO+1iblYCrkakOdLaUn01UZqgaxzquaFEuw9N7bNUpI8Fjn MzS11rYsttZMpfUyT5PiAbMhkiguvBzVv24wcsWwsg83RsH73OZtGGu9BC8SNrw/ jxaJv+B893R6LVIvhMMbRSYAizAEBqfiatttr64G2wNdQ8Hn58fJjStNTyYHFSWo EwsZe19xOW3XgTlRb3S12e6pw9kxDM46/hxpceXB0PwqUtq8kbD8I9Jeb838M3qf XM6+HgXYtSpbJ7zMYVPFI//vo140oHSvK8so0Ifz5NQ561AnqIteDPPKqwJTr02D BqcDZkrXNYoY3qP3rDnTobMciVJycSF6OJp+GHrDxoieV/kKnzjLbvcA4xYU8LmV PZv0/zNX2ZulrKpnVE8MitwurXnhEVZ8NG82ie7V3COu3JyNzAx+uvQCjUIRDef3 ne2icJcyY6kOuWFL0+FslCW1DwyNt9169PEzvJf5Ns590vEwmaxH9u5GfMsjmPHC jmh6zmxNyqSJ05XpWGAeM3iPVgeD0aTqi5dnbXIB7AUF/hmhoCf3co7rQ+y7404g YR7z0vaoK551wCJbkSLa4lnl7oo6lpt8LdS12wIx0c5RGwICTHzCG9uCfAZk+RqX TedLwHW7cfQ=
=BSFO
-----END PGP SIGNATURE-----
At Wed, 08 Mar 2006 11:17:26 -0500,
Quoted 6 lines · John Richard Moser
Quoted 4 lines
Quoted 2 lines
into the kernel; with alsa it's written to a /proc interface, which
Using /proc? Where? I've not noticed it :-)
I thought that was what
/proc/asound/card0/was for? :)
/proc is not referred from alsa-lib, i.e. programs don't access them. It's just for users who want to read some information.
Takashi
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Quoted 9 lines · Takashi Iwai
At Wed, 08 Mar 2006 11:17:26 -0500,
Quoted 3 lines · John Richard Moser
Quoted 2 lines
Quoted 1 line
into the kernel; with alsa it's written to a /proc interface, which
Using /proc? Where? I've not noticed it :-)
I thought that was what
/proc/asound/card0/was for? :)/proc is not referred from alsa-lib, i.e. programs don't access them. It's just for users who want to read some information.
This makes me wonder then how stuff is set up. There's no device, is there a sound card syscall now?
Quoted 3 lines
Takashi
- --
All content of all messages exchanged herein are left in the Public Domain, unless otherwise explicitly stated.
Creative brains are a valuable, limited resource. They shouldn't be wasted on re-inventing the wheel when there are so many fascinating new problems waiting out there.
-- Eric Steven Raymond
We will enslave their women, eat their children and rape their cattle!
-- Evil alien overlord from Blasto
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2.1 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
iQIVAwUBRA8oxQs1xW0HCTEFAQKGuA//RX/shtqKKPYQwSdP8O+Ai5C7VnQV+GiN I/+OalJIjfK/zUiJP6CIsFcNxTFTPiUJOwF/MKguiwWHsMwr/CWu/MVqAQEDC/yP UO1hSLcGKTp8xHh5jSIyotjisXaGx8rSq0Ye8ZDGPgDlVXb27USkmqWwySBCO+OJ pFtzGy9GxuJ+1OWoiXd0ytTSVm6WPEX/zFNsQmY4AnFid5aTrlxjDjQpx2cxt4iq j1nc+kkt450m4y233Cuncb7UXI/WzVhUX89dJf4RkjMKs5QDBvtQhzTB4K3PPByg l13xQq8si/knA8jcYcEbtHrzh+bTlvSwa8hJfONG3tbShHedJNQA+kLm6zpwm/kw hGkW0/IYU/Kpxh6Hg8yqb67ibbLFKjNc0njwRTZ5CLkE4pUwT/NASY4j5IcXYMrp bOla+ASuy/g0dYngmh0wK4tqJJSfl+SdZt+AkeVm6x5XtYv2k2uEhfQsDWpqhp0z TXiauuLhdJZ3mle21ZcY634Ws9C2vLmH6/TBQAgb0azryo9W4FQvEfPszbPyln04 irNC/VkKQ9vDr3spjXlqW9hCCD176+uOoHsI79WnUiz1EUdQ02kZ4mLm/i4JuXS8 Y8rezaHx7iypylVlI7c5CaDuEnrsBfvX7i/FqUa3/E0yH0v2wU0EcmC2K1OF1d7M q0jA3aTHj+o=
=wEi9
-----END PGP SIGNATURE-----
At Wed, 08 Mar 2006 13:56:06 -0500,
Quoted 9 lines · John Richard Moser
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1Quoted 4 lines · Takashi Iwai
At Wed, 08 Mar 2006 11:17:26 -0500,
Quoted 2 lines · John Richard Moser
Quoted 2 lines
Quoted 1 line
into the kernel; with alsa it's written to a /proc interface, which
Using /proc? Where? I've not noticed it :-)
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
I thought that was what
/proc/asound/card0/was for? :)/proc is not referred from alsa-lib, i.e. programs don't access them. It's just for users who want to read some information.
This makes me wonder then how stuff is set up. There's no device, is there a sound card syscall now?
You must have /dev/snd/* files, and the access is via normal syscalls... Check strace of your program.
Takashi
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Quoted 11 lines · Takashi Iwai
At Wed, 08 Mar 2006 13:56:06 -0500,
Quoted 9 lines · John Richard Moser
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1Quoted 5 lines · Takashi Iwai
At Wed, 08 Mar 2006 11:17:26 -0500,
Quoted 3 lines · John Richard Moser
Quoted 2 lines
Quoted 1 line
into the kernel; with alsa it's written to a /proc interface, which
Using /proc? Where? I've not noticed it :-)
I thought that was what
/proc/asound/card0/was for? :)
Quoted 9 lines
Quoted 5 lines
Quoted 3 lines
/proc is not referred from alsa-lib, i.e. programs don't access them. It's just for users who want to read some information.
This makes me wonder then how stuff is set up. There's no device, is there a sound card syscall now?
You must have
/dev/snd/*files, and the access is via normal syscalls... Check strace of your program.
Ahh nice, I was always looking for /dev/dsp :>
Quoted 3 lines
Takashi
- --
All content of all messages exchanged herein are left in the Public Domain, unless otherwise explicitly stated.
Creative brains are a valuable, limited resource. They shouldn't be wasted on re-inventing the wheel when there are so many fascinating new problems waiting out there.
-- Eric Steven Raymond
We will enslave their women, eat their children and rape their cattle!
-- Evil alien overlord from Blasto
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2.1 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
iQIVAwUBRA86Pws1xW0HCTEFAQIr9w//at9VpczxSKrCUzQGMYe33SW2vrksVVIg dTE12O1cc3gw0PK5UaD/QiN9+uoHKXY1YSkBBeCQQ2We50txjuy30jZ0j8PhkE22 PmuMb2knq9isspAT6Mr0ErRLS1+2FEvVN52yWwustl8cMiGrq812bYq63YHhGTPe 1b587S9+/wrEmswfA5/dvfJ+iEr3rHM0X7JPzJ721ayU1oeisAHw2GQCiXxQfwXZ aqOCnDRr7VgiPgnMC0nu8JVbNEdvJru4Dwgkc2siXoLnYmFYOiLv8JzujDJH3NfP aVxJ/NaoryvDSWXTcLOhm2nXKO7hD6id6wmd4ncSLedieK2Pw9WyAnVl40JguYgG OrTzmwJsYcc++isyLcYoDyJi861Urhly4Umnx8GQRxhpzJJ31eX+6pnNaic1f9Rv CNKOKrhLYG1AjqvDn0A1gyarMbXuXWu9D8Cjtv89nISr3AbGhq2Hz4oG78zRTiOX S56wnL8432RwhahQOkMSyJVdrxszN56FYiraWwfclrpAdE8TujV1YPYPG8qo2EGH 9vk6O0ZVccZUiyt3WLnyl+r/PLdsEaiFzrG7l0HANWq6KcCES8FClhfrWueCFQlM 4e7yAofQSKzJ1GrofaghVLloEJjoAoCWrJcVHMv6gAkPa8m8pwK9tUul3FRdDMD/ 1EHU/9tOMmY=
=8+gV
-----END PGP SIGNATURE-----