Skip to contentSkip to conversation

Combining Reserve and Map into a new IOCTL

RFC Ameya Palande 5 participants 6 messages Sep 1, 2009 – Sep 3, 2009

lore
Share & tools
lore
#0 lore raw

Hi,

Current DSPBridge MPU side API provides following IOCTLs which are related to reserving and mapping DSP address space:

1. DSPProcessor_ReserveMemory(): Reserve a virtually contiguous region of DSP address space.
2. DSPProcessor_Map(): Maps an MPU buffer to the DSP virtual address space.
3. DSPProcessor_UnMap(): Remove an MPU buffer mapping from the DSP virtual address space.
4. DSPProcessor_UnReserveMemory(): Frees a previously reserved region of the DSP virtual address space.

Typical call sequence is:

DSPProcessor_ReserveMemory()
DSPProcessor_Map()
DSPProcessor_UnMap()
DSPProcessor_UnReserveMemory()

Current approach has following problems:
1. Caller has to perform 4 system calls in order to map and unmap a buffer.
2. Kernel has no idea about the type of buffer (input/output). So depending on buffer type caller has to explicitly call DSPProcessor_FlushMemory() or DSPProcessor_InvalidateMemory().

Proposed approach:
Introduce 2 new IOCTLs which combine (reserve, map) and (unmap, unreserve). Caller should also specify buffer type (input/output) attribute as a parameter to new mapping IOCTL.

Benefits of new approach:
1. Saves 2 system calls per map and unmap pair.
2. By implementing lazy unreserve we can introduce cache of reserved mappings, which can skip reserve, unreserve operations.
3. Kernel can take care of flushing/invalidating cache depending on buffer type, which saves system call overhead and removed explicit cache control from user space.

These IOCTLs can be added to the current set of API which doesn't break compatibility with old applications.

Waiting for comments!

Ideas proposed in this document are from:
1. Hiroshi Doyu <[email protected]>
2. Felipe Contreras <[email protected]>

Cheers,
Ameya.

#1 lore raw
Quoted 1 line · Ameya Palande

Current approach has following problems:

0. It is brain-dead.

Quoted 18 lines

1. Caller has to perform 4 system calls in order to map and unmap a buffer.
2. Kernel has no idea about the type of buffer (input/output). So depending on buffer type caller has to explicitly call DSPProcessor_FlushMemory() or DSPProcessor_InvalidateMemory().

Proposed approach:
Introduce 2 new IOCTLs which combine (reserve, map) and (unmap, unreserve). Caller should also specify buffer type (input/output) attribute as a parameter to new mapping IOCTL.

Benefits of new approach:
1. Saves 2 system calls per map and unmap pair.
2. By implementing lazy unreserve we can introduce cache of reserved mappings, which can skip reserve, unreserve operations.
3. Kernel can take care of flushing/invalidating cache depending on buffer type, which saves system call overhead and removed explicit cache control from user space.

Quoted 8 lines

These IOCTLs can be added to the current set of API which doesn't break compatibility with old applications.

Waiting for comments!

Ideas proposed in this document are from:
1. Hiroshi Doyu<[email protected]>
2. Felipe Contreras<[email protected]>

Sounds good for an non-expert.

Signature · 6 lines
Best Regards,
Artem Bityutskiy (Артём Битюцкий)
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#2 lore raw
In reply to Ameya Palande
Quoted 39 lines · Ameya Palande

Hi,

Current DSPBridge MPU side API provides following IOCTLs which are related to reserving and mapping DSP address space:

1. DSPProcessor_ReserveMemory(): Reserve a virtually contiguous region of DSP address space.
2. DSPProcessor_Map(): Maps an MPU buffer to the DSP virtual address space.
3. DSPProcessor_UnMap(): Remove an MPU buffer mapping from the DSP virtual address space.
4. DSPProcessor_UnReserveMemory(): Frees a previously reserved region of the DSP virtual address space.

Typical call sequence is:

DSPProcessor_ReserveMemory()
DSPProcessor_Map()
DSPProcessor_UnMap()
DSPProcessor_UnReserveMemory()

Current approach has following problems:
1. Caller has to perform 4 system calls in order to map and unmap a buffer.
2. Kernel has no idea about the type of buffer (input/output). So depending on buffer type caller has to explicitly call DSPProcessor_FlushMemory() or DSPProcessor_InvalidateMemory().

Proposed approach:
Introduce 2 new IOCTLs which combine (reserve, map) and (unmap, unreserve). Caller should also specify buffer type (input/output) attribute as a parameter to new mapping IOCTL.

Benefits of new approach:
1. Saves 2 system calls per map and unmap pair.
2. By implementing lazy unreserve we can introduce cache of reserved mappings, which can skip reserve, unreserve operations.
3. Kernel can take care of flushing/invalidating cache depending on buffer type, which saves system call overhead and removed explicit cache control from user space.

Quoted 8 lines

These IOCTLs can be added to the current set of API which doesn't break compatibility with old applications.

Waiting for comments!

Ideas proposed in this document are from:
1. Hiroshi Doyu <[email protected]>
2. Felipe Contreras <[email protected]>

The whole proposal looks good to me :)

However, sometimes you would still want to do some extra cache operations afer Map() there should be a way to do 'flush', 'clean' and 'inv'.

Cheers.

Signature · 1 line
Felipe Contreras
#3 lore raw

The initial idea behind DSPProcessor_ReserveMemory call was to provide the User's the option of doing DSP buffer management for a pool of memory by themselves. So, call Reserve one time, and do map/unmap multiple times within that region, and call Unreserve only once you are done using the Buffer. But obviously that's not how this is being used.

Thank you,
Best regards,
Hari

Quoted 83 lines

-----Original Message-----
From: [email protected] [mailto:linux-omap- [email protected]] On Behalf Of Felipe Contreras
Sent: Wednesday, September 02, 2009 10:38 AM
To: Palande Ameya (Nokia-D/Helsinki)
Cc: [email protected]; Ramirez Luna, Omar; Guzman Lugo, Fernando; Doyu Hiroshi (Nokia-D/Helsinki); Tereshonkov Roman (Nokia-D/Helsinki)
Subject: Re: [DSPBRIDGE RFC] Combining Reserve and Map into a new IOCTL

Quoted 3 lines · Ameya Palande

Hi,

Current DSPBridge MPU side API provides following IOCTLs which are

related to

Quoted 3 lines

reserving and mapping DSP address space:

1. DSPProcessor_ReserveMemory(): Reserve a virtually contiguous region

of DSP

Quoted 2 lines

address space.
2. DSPProcessor_Map(): Maps an MPU buffer to the DSP virtual address

space.

Quoted 1 line

3. DSPProcessor_UnMap(): Remove an MPU buffer mapping from the DSP

virtual

Quoted 2 lines

address space.
4. DSPProcessor_UnReserveMemory(): Frees a previously reserved region of

the

Quoted 11 lines

DSP virtual address space.

Typical call sequence is:

DSPProcessor_ReserveMemory()
DSPProcessor_Map()
DSPProcessor_UnMap()
DSPProcessor_UnReserveMemory()

Current approach has following problems:
1. Caller has to perform 4 system calls in order to map and unmap a

buffer.

Quoted 1 line

2. Kernel has no idea about the type of buffer (input/output). So

depending

Quoted 1 line

on buffer type caller has to explicitly call

DSPProcessor_FlushMemory() or

Quoted 4 lines

DSPProcessor_InvalidateMemory().

Proposed approach:
Introduce 2 new IOCTLs which combine (reserve, map) and (unmap,

unreserve).

Quoted 8 lines

Caller should also specify buffer type (input/output) attribute as a parameter to new mapping IOCTL.

Benefits of new approach:
1. Saves 2 system calls per map and unmap pair.
2. By implementing lazy unreserve we can introduce cache of reserved mappings, which can skip reserve, unreserve operations.
3. Kernel can take care of flushing/invalidating cache depending on

buffer

Quoted 1 line

type, which saves system call overhead and removed explicit cache

control

Quoted 10 lines

from user space.

These IOCTLs can be added to the current set of API which doesn't break compatibility with old applications.

Waiting for comments!

Ideas proposed in this document are from:
1. Hiroshi Doyu <[email protected]>
2. Felipe Contreras <[email protected]>

The whole proposal looks good to me :)

However, sometimes you would still want to do some extra cache operations afer Map() there should be a way to do 'flush', 'clean' and 'inv'.

Cheers.

--
Felipe Contreras
--
To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#4 lore raw
Quoted 6 lines · Hari Kanigeri

The initial idea behind DSPProcessor_ReserveMemory call was to provide the User's the option of doing DSP buffer management for a pool of memory by themselves. So, call Reserve one time, and do map/unmap multiple times within that region, and call Unreserve only once you are done using the Buffer. But obviously that's not how this is being used.

And they still would be able to do that because the old ioctls are not be removed (yet).

Signature · 1 line
Felipe Contreras
#5 lore raw
In reply to Kanigeri, Hari

Hi Hari,

From: "ext Kanigeri, Hari" <[email protected]>
Subject: RE: [DSPBRIDGE RFC] Combining Reserve and Map into a new IOCTL
Date: Wed, 2 Sep 2009 18:16:26 +0200

Quoted 6 lines

The initial idea behind DSPProcessor_ReserveMemory call was to provide the User's the option of doing DSP buffer management for a pool of memory by themselves. So, call Reserve one time, and do map/unmap multiple times within that region, and call Unreserve only once you are done using the Buffer. But obviously that's not how this is being used.

Would it be possible to explan a little bit more about the advantage of above use case, where it reserves onece, and maps/unmaps multiple times?