Skip to contentSkip to conversation

Ramdisk for > 1GB / >2 GB

DISCUSSION Manfred Spraul 3 participants 3 messages Feb 1, 1999 – Feb 5, 1999 Memory management

lore
Share & tools
lore
#0 lore raw

I've written a ramdisk driver that can use physical, unmapped memory. I've posted a beta version this morning to linux-kernel.
Basically, it is a kernel patch that manages the memory (alloc_hugemem(), free_hugemem()), and a block device driver that can use this memory.

I'm new in the Linux MM, perhaps you could help me on these questions:

1) SMP:
I use a spinlock for every ramdisk, and one page for each drive as a window to the physical memory. Since only 1 processor uses this page, I can use __flush_tlb_one( == INVLPG only on the local processor) without any further synchronization.
Is that stable on SMP, and do you think that this parallel enough? Linus suggested using one 4MB pte for each processor, but I think that this would be to much overhead.
Another idea would be using a hash table (eg. 32 spinlocks, 32 pages) that is shared by all processors.

2) I can't make the driver a module because I use 'pgt_offset_k()' to traverse the page tables, but init_mm is not exported.
Is there any other way how I can find the pte_t that belongs to my page?

3) Is more than 2 GB memory a problem that only applies to the i386 architecture, or is there demand for that on PowerPC, Sparc32?

Regards,
Manfred

--
To unsubscribe, send a message with 'unsubscribe linux-mm my@address' in the body to [email protected]. For more info on Linux MM,
see: http://humbolt.geo.uu.nl/Linux-MM/

#1 lore raw
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"MS" == Manfred Spraul <[email protected]> writes:

MS> I've written a ramdisk driver that can use physical, unmapped memory. I've MS> posted a beta version this morning to linux-kernel.
MS> Basically, it is a kernel patch that manages the memory (alloc_hugemem(), MS> free_hugemem()), and a block device driver that can use this memory.

MS> I'm new in the Linux MM, perhaps you could help me on these questions:

MS> 1) SMP:
MS> I use a spinlock for every ramdisk, and one page for each drive as a window MS> to the physical memory. Since only 1 processor uses this page, I can use MS> __flush_tlb_one( == INVLPG only on the local processor) without any further MS> synchronization.

Sounds good. But it's not my area of expertise.

MS> Is that stable on SMP, and do you think that this parallel enough?

MS> Linus suggested using one 4MB pte for each processor, but I think that this MS> would be to much overhead.
MS> Another idea would be using a hash table (eg. 32 spinlocks, 32 pages) that MS> is shared by all processors.

Except for quantity of address space consumed a 4MB pte should be equal to a 4k pte.

MS> 3) Is more than 2 GB memory a problem that only applies to the i386 MS> architecture, or is there demand for that on PowerPC, Sparc32?

It's a problem for 32bit architectures. Most of the RISC processors (I believe) have 64bit extensions so it's less of an issue there.

Eric
--
To unsubscribe, send a message with 'unsubscribe linux-mm my@address' in the body to [email protected]. For more info on Linux MM,
see: http://humbolt.geo.uu.nl/Linux-MM/

#2 lore raw
In reply to Manfred Spraul
Quoted 2 lines · Manfred Spraul

3) Is more than 2 GB memory a problem that only applies to the i386 architecture, or is there demand for that on PowerPC, Sparc32?

The limit is even lower, 512mb for the MIPS32 kernel. This was so far not a problem since the so far supported machines had a lower maximum but the problem now showed up big time on the radar, so there will be a MIPS64 kernel with a 1TB maximum soon to solve the problem in the not far future.

Ralf
--
To unsubscribe, send a message with 'unsubscribe linux-mm my@address' in the body to [email protected]. For more info on Linux MM,
see: http://humbolt.geo.uu.nl/Linux-MM/