Re: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix message allocation

Re: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix message allocation

[Daniel Drake]

nickcheng wrote:
> Hi,
> I definitely agree it is in atomic context but why is the memory not for
> DMA?
> Would you please show me why?

It would probably be easier if you could explain where you believe the 
memory IS used for DMA :)

Anyway, looking at ARCMSR_MESSAGE_READ_RQBUFFER
current code does this:
		ver_addr = kmalloc(1032, GFP_ATOMIC);

Here are the cases when that buffer is used:

checking for successful malloc: not DMA
		if (!ver_addr) {

copying the address: not DMA
		ptmpQbuffer = ver_addr;

memcpying 1 byte into the buffer: not DMA
			memcpy(ptmpQbuffer, pQbuffer, 1);

incrementing address: not DMA
			ptmpQbuffer++;

memcpying from the buffer: not DMA
		memcpy(pcmdmessagefld->messagedatabuffer, ver_addr, allxfer_len);

freeing the buffer: not DMA
		kfree(ver_addr);

Daniel

RE: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix message allocation

[nickcheng]

Sorry, maybe I did not ask distinctly enough.
I mean if I would like to allocate a memory space from ZONE_DMA for atomic
context, why can I not use kmalloc(1032, GFP_ATOMIC|GFP_DMA)?
In case of lack of GFP_DMA, kmalloc would grab the memory from ZONE_HIGH or
ZONE_HIGHMEM, isn't it?(I read it from the textbook of Linux Kernel
Development by Robert Love)
Or the basic is that you don't think it is necessary to allocate a memory
space from DMA area?
Please give me some comments.
Thank you very much, :-）

-----Original Message-----
From: Daniel Drake [mailto:dsd@gentoo.org] 
Sent: Monday, February 25, 2008 6:15 PM
To: nick.cheng@areca.com.tw
Cc: 'James Bottomley'; linux-scsi@vger.kernel.org
Subject: Re: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix message
allocation

nickcheng wrote:
> Hi,
> I definitely agree it is in atomic context but why is the memory not for
> DMA?
> Would you please show me why?

It would probably be easier if you could explain where you believe the 
memory IS used for DMA :)

Anyway, looking at ARCMSR_MESSAGE_READ_RQBUFFER
current code does this:
		ver_addr = kmalloc(1032, GFP_ATOMIC);

Here are the cases when that buffer is used:

checking for successful malloc: not DMA
		if (!ver_addr) {

copying the address: not DMA
		ptmpQbuffer = ver_addr;

memcpying 1 byte into the buffer: not DMA
			memcpy(ptmpQbuffer, pQbuffer, 1);

incrementing address: not DMA
			ptmpQbuffer++;

memcpying from the buffer: not DMA
		memcpy(pcmdmessagefld->messagedatabuffer, ver_addr,
allxfer_len);

freeing the buffer: not DMA
		kfree(ver_addr);

Daniel

-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

RE: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix message allocation

[James Bottomley]

On Mon, 2008-02-25 at 18:47 +0800, nickcheng wrote:
> Sorry, maybe I did not ask distinctly enough.
> I mean if I would like to allocate a memory space from ZONE_DMA for atomic
> context, why can I not use kmalloc(1032, GFP_ATOMIC|GFP_DMA)?
> In case of lack of GFP_DMA, kmalloc would grab the memory from ZONE_HIGH or
> ZONE_HIGHMEM, isn't it?(I read it from the textbook of Linux Kernel
> Development by Robert Love)

Um, no that's not true at all.  GFP_DMA only allocates memory from
ZONE_DMA and fails otherwise.  You only need memory from ZONE_DMA if you
cannot address physical memory above 24 bits (the old ISA restriction).
This only applies if you're a standard ISA device and you're going
actually to DMA to the memory in question.  Neither of which applies in
the arcmsr case, since you're only using the memory as a copy buffer
within the kernel (it never sees an actual DMA transfer), and arcmsr
doesn't have the ISA restrictions.

> Or the basic is that you don't think it is necessary to allocate a memory
> space from DMA area?
> Please give me some comments.

It's unnecessary because you never DMA to it, but even if you did, since
arcmsr is a non-ISA device (with 64 bit DMA mask falling back to 32) you
can just use ordinary kmalloc'd memory for that (provided you obey the
coherence requirements).

James

RE: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix messageallocation

[nickcheng]

James,
Appreciate for your answer.
One more question: do you think there are any compatibility issues on
kmalloc/kfree pair?
I just trace the kernel code. It looks like it would not.
But I am not quite sure absolutely.
May I have your comments?
Thank you,

-----Original Message-----
From: James Bottomley [mailto:James.Bottomley@HansenPartnership.com] 
Sent: Monday, February 25, 2008 10:39 PM
To: nick.cheng@areca.com.tw
Cc: 'Daniel Drake'; linux-scsi@vger.kernel.org; 'erich'
Subject: RE: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix
messageallocation

On Mon, 2008-02-25 at 18:47 +0800, nickcheng wrote:
> Sorry, maybe I did not ask distinctly enough.
> I mean if I would like to allocate a memory space from ZONE_DMA for atomic
> context, why can I not use kmalloc(1032, GFP_ATOMIC|GFP_DMA)?
> In case of lack of GFP_DMA, kmalloc would grab the memory from ZONE_HIGH
or
> ZONE_HIGHMEM, isn't it?(I read it from the textbook of Linux Kernel
> Development by Robert Love)

Um, no that's not true at all.  GFP_DMA only allocates memory from
ZONE_DMA and fails otherwise.  You only need memory from ZONE_DMA if you
cannot address physical memory above 24 bits (the old ISA restriction).
This only applies if you're a standard ISA device and you're going
actually to DMA to the memory in question.  Neither of which applies in
the arcmsr case, since you're only using the memory as a copy buffer
within the kernel (it never sees an actual DMA transfer), and arcmsr
doesn't have the ISA restrictions.

> Or the basic is that you don't think it is necessary to allocate a memory
> space from DMA area?
> Please give me some comments.

It's unnecessary because you never DMA to it, but even if you did, since
arcmsr is a non-ISA device (with 64 bit DMA mask falling back to 32) you
can just use ordinary kmalloc'd memory for that (provided you obey the
coherence requirements).

James

RE: Patch added to scsi-rc-fixes-2.6: [SCSI] arcmsr: fix messageallocation

[James Bottomley]

On Tue, 2008-02-26 at 12:29 +0800, nickcheng wrote:
> James,
> Appreciate for your answer.
> One more question: do you think there are any compatibility issues on
> kmalloc/kfree pair?
> I just trace the kernel code. It looks like it would not.
> But I am not quite sure absolutely.
> May I have your comments?

You mean does kmalloc/kfree obey the same rules as the coherent
allocators in all regards except coherence ... pretty much, yes.  There
are some technical differences at the bottom in terms of page alignment,
but nothing you need to worry about.

James