Physical segmentation advantages

[qw]

And again I wonder how a technical discussion can turn into a personal fight. What is wrong with you guys?

[Owen]

Wrong. You're ignoring all of the things that make segmentation bad (memory mapped files, swap space, physical address space de-fragmentation, etc) and focusing on a few negligible things in a deluded attempt at pretending segmentation doesn't suck badly.

Using paging to support those is ancient (70s?)

1. Memory mapped files - no reason why this cannot be implemented without paging

So when I memory map a 4TB file... you are proposing to try and allocate 4TB of physical memory?

I think neither of us have a machine capable of that

2. Swap space - you simply do not support such ancient constructs

OK, no swap space. What about paging out of unmodified file mappings (such as executable code)? Note that many programs contain considerable quantities of initialization code which can safely be discarded after startup

3. Physical address defragmentation - not much of a problem when computers have many GB of memory.

Still very much a problem. When my system has 7 of 8GB of RAM used, how much contiguous RAM do you think the operating system will be able to find? Hint: Probably chunks of no more than 1MB.

This is clearly useless if I want to allocate more than 1MB. For example, my application wants to create a full screen window and therefore would like a 1920x1080 framebuffer (~10MB)

Wrong. In fact it's the exact opposite - OSs use segmentation for backward compatibility, and inevitably break compatibility to remove segmentation in later versions (once they realise the pain of keeping it is worse than the pain of breaking compatibility).

Not so, popular OS designs (UNIX-compability) requires paging as the whole concept is constructed on top of paging.

You may find hints of the exact opposite in the fact that the signal UNIX sends a process for most memory faults is SIGSEGV - segmentation violation. Its' just that the early machines had different notions of segments (and no notions of paging). Indeed, the vfork(2) system call was introduced in order to make the fork/exec paradigm efficient on such machines (because otherwise you had to copy the entire process' memory in order to discard it immediately after)

[OSwhatever]

3. Physical address defragmentation - not much of a problem when computers have many GB of memory. Still very much a problem. When my system has 7 of 8GB of RAM used, how much contiguous RAM do you think the operating system will be able to find? Hint: Probably chunks of no more than 1MB.

There are going to be problems with fragmentation unless a segmented OS has compaction of the memory, then you will have large chunks of memory again even if the available memory is low. Compaction is the key here. Also if you don't like compaction there are other allocation strategies to minimize fragmentation. In paged OSes, there are problems with fragmentation there as well, how are they dealing with it?

Most then stuff you do with paging, can be achieved with segmentation as well. User mode programs are likely to request blocks of memories in power of two sizes anyway in order to simplify the allocation and that you don't need to visit the kernel for every little memory allocation in user space.

I just think that translation is just unnecessary, it's just an extra layer and indirection that we don't really need. It's like fetching water from the other side of the river. Also with paging you have to plan your virtual address space. With segments this is not necessary as they are a segment with a base address somewhere in the physical memory and you don't deal with addresses at all, only offsets within your segment. So in a sense segmentation is also some kind of translation but I think it gives you more freedom to choose, the layout, structures and sizes of your segments.

[Brendan]

Hi,

3. Physical address defragmentation - not much of a problem when computers have many GB of memory. Still very much a problem. When my system has 7 of 8GB of RAM used, how much contiguous RAM do you think the operating system will be able to find? Hint: Probably chunks of no more than 1MB.

There are going to be problems with fragmentation unless a segmented OS has compaction of the memory, then you will have large chunks of memory again even if the available memory is low. Compaction is the key here. Also if you don't like compaction there are other allocation strategies to minimize fragmentation. In paged OSes, there are problems with fragmentation there as well, how are they dealing with it?

See this post.

Basically, both paging and segmentation have overhead, but the overhead of segmentation is worse.

Most then stuff you do with paging, can be achieved with segmentation as well.

Segmentation advocates like to say that. Then you ask them how to do basic things (e.g. a large memory mapped file with pieces in the middle still on disk) and they can't think of a sane way. It's like they think that if they say it often enough it might magically become true one day.

User mode programs are likely to request blocks of memories in power of two sizes anyway in order to simplify the allocation and that you don't need to visit the kernel for every little memory allocation in user space.

You can't let applications create their own GDT/LDT entries (otherwise they'll create a CPL=3 data segment to write wherever they like in kernel space) so you end up with a kernel API call for every single allocation regardless of how small it is. Normally (for paging) the application asks for larger chunks and divides it up with no security problems and much fewer kernel API calls.

I just think that translation is just unnecessary, it's just an extra layer and indirection that we don't really need. It's like fetching water from the other side of the river. Also with paging you have to plan your virtual address space. With segments this is not necessary as they are a segment with a base address somewhere in the physical memory and you don't deal with addresses at all, only offsets within your segment. So in a sense segmentation is also some kind of translation but I think it gives you more freedom to choose, the layout, structures and sizes of your segments.

I just think the translation (done by segmentation) is just unnecessary, it's just an extra layer of indirection that we don't really need. Also with segmentation you have to plan out your segments. With paging this is not necessary as they're just page/s anywhere in physical memory and you don't need to deal with segments at all. So in a sense, paging is also some kind of translation, but it's been repeatedly proven to give you more freedom to choose, the layout, structures and sizes of your virtual address space areas.


Cheers,

Brendan

[alexfru]

You can't let applications create their own GDT/LDT entries (otherwise they'll create a CPL=3 data segment to write wherever they like in kernel space) so you end up with a kernel API call for every single allocation regardless of how small it is. Normally (for paging) the application asks for larger chunks and divides it up with no security problems and much fewer kernel API calls.

Wait, if you have per-process virtual address spaces and non-user pages are properly marked as supervisor-only, user segments can be allowed to point anywhere in the address space and page translation will still catch user's disallowed page accesses. AFAIU, we have that already working in x86 Windows and Linux. You can have user code and data segments span the entire 4G without a problem.

[OSwhatever]

Hi,

Basically, both paging and segmentation have overhead, but the overhead of segmentation is worse.

Segmentation advocates like to say that. Then you ask them how to do basic things (e.g. a large memory mapped file with pieces in the middle still on disk) and they can't think of a sane way. It's like they think that if they say it often enough it might magically become true one day.

You can't let applications create their own GDT/LDT entries (otherwise they'll create a CPL=3 data segment to write wherever they like in kernel space) so you end up with a kernel API call for every single allocation regardless of how small it is. Normally (for paging) the application asks for larger chunks and divides it up with no security problems and much fewer kernel API calls.


Cheers,

Brendan

Brendan, you are very x86 oriented and often bring up the limitations of that segmentation. I'm talking about the general case, a CPU that does not exist but implement segmentation. There is a lot of things you can do differently with segmentation rather than the x86 case. For example:

Very large segment descriptor cache (like a TLB)
Allow user processes to create their own segment descriptors based on segments given by the kernel
Atomically and cache aware updating the base address of a segment.
Expand the number of allowed segments compared to the x86 case.

[Brendan]

Hi,

You can't let applications create their own GDT/LDT entries (otherwise they'll create a CPL=3 data segment to write wherever they like in kernel space) so you end up with a kernel API call for every single allocation regardless of how small it is. Normally (for paging) the application asks for larger chunks and divides it up with no security problems and much fewer kernel API calls.

Wait, if you have per-process virtual address spaces and non-user pages are properly marked as supervisor-only, user segments can be allowed to point anywhere in the address space and page translation will still catch user's disallowed page accesses. AFAIU, we have that already working in x86 Windows and Linux. You can have user code and data segments span the entire 4G without a problem.

Of course paging solves the problem. We were talking about segmentation without paging though.


Cheers,

Brendan

[Brendan]

Hi,

Brendan, you are very x86 oriented and often bring up the limitations of that segmentation. I'm talking about the general case, a CPU that does not exist but implement segmentation. There is a lot of things you can do differently with segmentation rather than the x86 case. For example:

Very large segment descriptor cache (like a TLB)
Allow user processes to create their own segment descriptors based on segments given by the kernel
Atomically and cache aware updating the base address of a segment.
Expand the number of allowed segments compared to the x86 case.

The 3 main disadvantages of paging are TLB misses, TLB invalidation and "multi-CPU TLB shootdown". These are the problems that the typical "anti-paging" fools are trying to avoid. As soon as you add "a very large segment descriptor cache" you get (the equivalent of) these 3 problems - "segment descriptor cache misses", "segment descriptor cache invalidation" and "multi-CPU segment descriptor cache shootdown".

If you allow user processes to create their own segment descriptors based on segments given by the kernel; what happens when the kernel needs to defragment physical memory? Does it have to examine every "process owned descriptor table" and reshuffle it too? When do the protection checks happen (every time anything modifies some "processor owned descriptor table", or every time a segment is loaded into the segment register and/or cache)? What is the overhead of this (bad, very bad, or worse)?

You want to do "atomic and cache aware updates" of the base address of a segment? Do you want to check the entire "segment descriptor cache" in every single CPU, every time any CPU does any write? Alternatively, will every CPU auto-detect when a write modifies a segment base address somehow and then broadcast a special message to all other CPUs? What happens in aliasing cases (e.g. one CPU thinks an area is normal data while another CPU thinks the same area contains segment base addresses)?

You can't invent an imaginary "science fiction" future system, pretend that it's a realistic alternative for today's systems while ignoring all of the potential problems, then expect people to take you seriously.


Cheers,

Brendan

[linguofreak]

3. Physical address defragmentation - not much of a problem when computers have many GB of memory. Still very much a problem. When my system has 7 of 8GB of RAM used, how much contiguous RAM do you think the operating system will be able to find? Hint: Probably chunks of no more than 1MB.

There are going to be problems with fragmentation unless a segmented OS has compaction of the memory, then you will have large chunks of memory again even if the available memory is low. Compaction is the key here. Also if you don't like compaction there are other allocation strategies to minimize fragmentation. In paged OSes, there are problems with fragmentation there as well, how are they dealing with it?

Compaction is slow. Compaction is not the solution, compaction is the problem.

For a paged OS, the smallest chunk the OS ever allocates to a process is one page, and the page can be placed anywhere in the process's address space, regardless of its position in physical memory. So physical memory can become arbitrarily heavily fragmented, and it will never be a problem. A paged OS never needs to take time to compact the space allocated to programs, whereas a segmented OS without paging (or a hybrid segmentation/paging system running on a processor designed the same way that Intel designed the x86_32) has to. OS components and user programs on a paging system will occasionally have to compact their own heaps, but they'd have to do that on *any* system, and on a platform with segmentation but no paging, you end up having to do both types of compaction instead of just the one.

Segmentation has some enticing advantages on the front-end (the algorithmic structures available to user programs, as well as to kernel components that don't need to concern themselves with the actual way that the processor implements memory management), and I doubt microkernels will ever actually become mainstream as long as flat address spaces remain the norm. Paging, however, wins hands down on the back-end (for the parts of the kernel that actually do memory management), so the potential advantages of segmentation will never be realized on any processor that doesn't implement segmentation entirely in terms of paging.

[rdos]

You can't let applications create their own GDT/LDT entries (otherwise they'll create a CPL=3 data segment to write wherever they like in kernel space) so you end up with a kernel API call for every single allocation regardless of how small it is. Normally (for paging) the application asks for larger chunks and divides it up with no security problems and much fewer kernel API calls.

Not true. Segmentation can subdivide larger segments into chunks just like paging can. This of course is necessary with x86 segmentation as Intel didn't think clearly when they moved to 32-bit (they should have extended the segment registers to 32-bit as well).

Also with segmentation you have to plan out your segments.

Only with x86 segments because Intel didn't extend segment registers to 32-bit. Another implementation doesn't need to have these limitations.