CUDA's low-level virtual memory management mechanism involves:
Physical allocations
Virtual address range reservations
Mappings between the above
Conveniently, if you map a physical allocation to some address range - you can "free" the physical allocation and keep using it until it is unmapped.
Can we also do this for virtual address range reservations? i.e. will they be kept alive until the mappings are gone? The documentation doesn't say.
You can't early-free reserved address ranges.
Virtual memory address range reservations apparently don't have the reference-counting mechanism which physical allocations have; the driver expects them to be released when they are no longer in use, period. An attempt to free the reservation will fail with CUDA_ERROR_INVALID_VALUE.
Related
How is the Virtual address space greater than Physical address space ?
suppose a Virtual 0x7000 maps to physical address 0x8000, can another virtual address lets say
0x7500 map to the same physical location as 0x8000, if not then how can there be more virtual
address and limited physical memory since mapping has to convert to physical address?
Please help me understand this concept.
http://en.wikipedia.org/wiki/Virtual_memory.
Virtual Memory uses both physical ram and hard disk space to represent more memory than may physically exist, and provides an interface whereby each program can request memory resources without having to be concerned with the other programs existent on the machine and which memory addresses they may request.
The whole virtual address space does not have to be mapped to physical memory at the same time. That's what makes it "virtual". The contents of that virtual memory which is allocated but not currently mapped to physical memory reside on some form of external storage, typically disk.
It is the memory management system's job to move virtual memory pages into and out of physical memory as needed, and the requirement to do so is why virtual-memory computers can slow down overall when enough memory is allocated that it no longer all fits in physical memory at the same time.
I have come across a term - holes in the memory in Linux. I believe this is the memory that is I/O remapped. Is my understanding correct?
Holes in memory can mean different things:
1) It can refer to physical memory addressing: For historical and boot-strapping reasons, in the "standard PC" (x86) architecture, all of system RAM is not contiguous. There are "holes" in the address space where memory-mapped I/O resides. For example, from the earliest days, there has been an area reserved for boot ROM (BIOS) and video memory. Also, there is a large area of the address space which is reserved for dynamic assignment to PCI (and PCI-X or PCI-Express) peripherals. These areas are often mapped as needed into kernel virtual address space by device drivers (which may be referred to as "I/O remapping").
Memory controllers built-in to the motherboard allow the physical address of the RAM to be configured (this is typically handled by the BIOS in the standard PC architecture). Other [non-x86] architectures often have similar holes in the physical address space.
2.) The term can also refer to unassigned regions in the virtual address space. Both kernel virtual address space and user processes' virtual address space typically have "holes" in them. For example, linux doesn't map any physical memory corresponding to virtual address 0 (i.e. the first page of the address space never has valid memory) -- this allows null pointer references to be trapped.
In some kinds of memory allocations, the linux kernel maintains unmapped areas between properly allocated virtual memory regions in order to trap faulty memory references (i.e. that stray beyond the end of the allocated space).
Assuming x86, I'm starting to learn that addresses 0x0 thru 0x7FFFFFFF are for the process; whereas anything higher is reserved for the kernel.
I have three curiosities:
1) Does a process EVER call an address higher than 0x7FFFFFFF? I assume it will always result in some sort of access denied? How is that access denied enforced?
2) Does "shared memory" IPC work by mapping two processes virtual addresses to the same physical address range?
3) The amount of RAM in your machine can vary. You may have 2GB, or much more like 16GB. How does this affect the addressing of RAM? Does the kernel ever leave a bunch of RAM unused because it was reserved for itself, but doesn't need it? How can I see this?
I am not very sure but you will find the maximum in this MSDN doc about how it works:-
The range of virtual addresses that is available to a process is
called the virtual address space for the process. Each user-mode
process has its own private virtual address space. For a 32-bit
process, the virtual address space is usually the 2-gigabyte range
0x00000000 through 0x7FFFFFFF. For a 64-bit process, the virtual
address space is the 8-terabyte range 0x000'00000000 through
0x7FF'FFFFFFFF. A range of virtual addresses is sometimes called a
range of virtual memory.
The diagram shows the virtual address spaces for two 64-bit processes:
Notepad.exe and MyApp.exe. Each process has its own virtual address
space that goes from 0x000'0000000 through 0x7FF'FFFFFFFF. Each shaded
block represents one page (4 kilobytes in size) of virtual or physical
memory. Notice that the Notepad process uses three contiguous pages of
virtual addresses, starting at 0x7F7'93950000. But those three
contiguous pages of virtual addresses are mapped to noncontiguous
pages in physical memory. Also notice that both processes use a page
of virtual memory beginning at 0x7F7'93950000, but those virtual pages
are mapped to different pages of physical memory.
I am not able to exactly difference between kernel logical address and virtual address. In Linux device driver book it says that all logical address are kernel virtual address, and virtual address doesn't have any linear mapping. But logically wise when we say it is logical and when we say virtual and in which situation we use these two ?
The Linux kernel maps most of the virtual address space that belongs to the kernel to perform 1:1 mapping with an offset of the first part of physical memory. (slightly less then for 1Gb for 32bit x86, can be different for other processors or configurations). For example, for kernel code on x86 address 0xc00000001 is mapped to physical address 0x1.
This is called logical mapping - a 1:1 mapping (with an offset) that allows the kernel to access most of the physical memory of the machine.
But this is not enough - sometime we have more then 1Gb physical memory on a 32bit machine, sometime we want to reference non contiguous physical memory blocks as contiguous to make thing simple, sometime we want to map memory mapped IO regions which are not RAM.
For this, the kernel keeps a region at the top of its virtual address space where it does a "random" page to page mapping. The mapping there do not follow the 1:1 pattern of the logical mapping area. This is what we call the virtual mapping.
It is important to add that on many platforms (x86 is an example), both the logical and virtual mapping are done using the same hardware mechanism (TLB controlling virtual memory). In many cases, the "logical mapping" is actually done using virtual memory facility of the processor, so this can be a little confusing. The difference therefore is the pattern according to which the mapping is done: 1:1 for logical, something random for virtual.
Basically there are 3 kinds of addressing, namely
Logical Addressing : Address is formed by base and offset. This is nothing but segmented addressing, where the address (or offset) in the program is always used with the base value in the segment descriptor
Linear Addressing : Also called virtual address. Here adresses are contigous, but the physical address are not. Paging is used to implement this.
Physical addressing : The actual address on the Main Memory!
Now, in linux, Kernel memory (in address space) is beyond 3 GB ( 3GB to 4GB), i.e. 0xc000000..The addresses used by Kernel are not Physical addresses. To map the virtual address it uses PAGE_OFFSET. Care must be taken that no page translation is involved. i.e. these addresses are contiguous in nature. However there is a limit to this, i.e. 896 MB on x86. Beyond which paging is used for translation. When you use vmalloc, these addresses are returned to access the allocated memory.
In short, when someone refers to Virtual Memory in context of User Space, then it is through Paging. If Kernel Virtual Memory is mentioned then it is either PAGE_OFFSETed or vmalloced address.
(Reference - Understanding Linux Kernel - 2.6 based )
Shash
Kernel logical addresses are mappings accessible to kernel code through normal CPU memory access functions. On 32-bit systems, only 4GB of kernel logical address space exists, even if more physical memory than that is in use. Logical address space backed by physical memory can be allocated with kmalloc.
Virtual addresses do not necessarily have corresponding logical addresses. You can allocate physical memory with vmalloc and get back a virtual address that has no corresponding logical address (on 32-bit systems with PAE, for example). You can then use kmap to assign a logical address to that virtual address.
Simply speaking, virtual address would include "high memory", which doesn't do the 1:1 mapping for the physical address,if your RAM size is more than the address range of kernel(typically,For 1G/3G in X86,your RAM is 3G but your kernel addressing range is 1G) and also the address return from kmap() and vmalloc(), which requires the kernel to establish page table for the memory mapping. since logic address is always memory mapped by the kernel(1:1 mapping), you don't need to explicitly call kernel API,like set_pte to set up the page table entry for the particular page.
so virtual address can't be logic address all the time.
I'm a little confused by the meaning of "Aliasing" between CPU-cache and Physical address.
First I found It's definition on Wikipedia :
However, VIVT suffers from aliasing problems, where several different virtual addresses may refer to the same physical address. Another problem is homonyms, where the same virtual address maps to several different physical addresses.
but after a while I saw a different definition on a presentation(ppt)
of DAC'05: "Energy-Efficient Physically Tagged Caches for Embedded Processors with
Virtual Memory"
Cache aliasing and synonyms:
Alias: Same virtual address from different contexts mapped to different physical addresses Synonym: Different virtual address mapped to the same physical address (data sharing)
As I'm not a native speaker, I don't know which is correct,
though I feel the Wiki's definition is correct.
Edit:
Concept of "aliasing" in CPU cache usually means "synonym", on the contrary is "homonym". In a more generic level, "aliasing" is "confusing" or "chaos" or something like that. So In my opinion, "aliasing" exactly means the mapping of (X->Y) is "not bijective", where
"X" = the subset of physical addresses units which has been cached. (each element is a line of byte)
"Y" = the set of valid cache lines. (elements a also "line")
You'd need to learn about Virtual Memory first, but basically it's this:
The memory addresses your program uses aren't the physical addresses that the RAM uses; they're virtual addresses mapped to physical addresses by the CPU.
Multiple virtual addressses can point to the same physical address.
That means that you can have two copies of the same data in separate parts of the cache without knowing it... and they wouldn't be updated correctly, so you'd get wrong results.
Edit:
Exerpt of reference:
Cache aliasing occurs when multiple mappings to a physical page of memory have conflicting caching states, such as cached and uncached. Due to these conflicting states, data in that physical page may become corrupted when the processor's cache is flushed. If that page is being used for DMA by a driver, this can lead to hardware stability problems and system lockups.
For those who are still unconvinced:
On ARMv4 and ARMv5 processors, cache is organized as a virtual-indexed, virtual-tagged (VIVT) cache in which both the index and the tag are based on the virtual address. The main advantage of this method is that cache lookups are faster because the translation look-aside buffer (TLB) is not involved in matching cache lines for a virtual address. However, this caching method does require more frequent cache flushing because of cache aliasing, in which the same physical address can be mapped to multiple virtual addresses.
#Wu yes you do need to understand virtual memory little to understand aliasing. Let me give you a few lines of explanation first:
Lets say I have a RAM (physical memory) of 1GB. I want to present my programmer with a view that I have 4GB memory then I use virtual memory. In virtual memory, the programmer thinks that he/she has 4GB and writes their program from that perspective. They do not need to know how much physical memory exists. The advantage is that program will run on computers with different amounts of RAM. Also, the program can run on a computer together with other programs (also consuming physical memory).
So here is how virtual memory is implement. I will give a simple 1-level virtual memory system (Intel has a 2/3-level system which just makes it complicated for explanation.
Our problem here is that the programmer has 4 Billion addresses and we only have 1 billion places to put those 4 billion addresses. So, addresses are from the virtual address space need to be mapped to physical address space. This is done using a simple index table called a Page Table. You access a Page Table with a virtual address and it gives you the physical address of that memory location.
Some details: Remember that physical space is only 1GB so the system only keeps the most recently accessed 1GB worth in physical memory and keeps the rest in system disk. When the program requests a particular address, we first check if it is already in physical memory. If so, it is returned to the program. If not, it brought from the disk and put into physical memory and then returned to the program. The latter is known as a Page Fault.
Coming back to aliasing in context of virtual memory: since there is mapping between virtual -> physical addresses, it is possible to make two virtual addresses to map to the same physical address. it is the same as saying that if I look at my page table for virtual
address X and Y, I will get the same physical address in BOTH cases.
I show below a simple example of a 8 entry Page Table. Say there are 8 vitual addresses and only 3 physical addresses. The page table looks as follows:
0: 1
1: On disk
2: 2
3: 1
4: On disk
5: On disk
6: On disk
7: 0
This mean that if virtual address 4 is accessed, you will get a page fault.
If virtual addresses 3 is accessed, you will get the physical address 1
In this case, virtual addresses 0 and 3 are aliasing to the same physical address 1 for both of them
NOTE: I used the terms physical and virtual addresses everywhere to simplify the concept. In a real system, the virtual-to-physical mapping is not on a per address basis . Instead, we map chunks of virtual space to physical space. Each chunk is called a Page (thats why the mapping table is called a page table) and the size of the chunk is a property of the ISA, e.g., Intel x86 has 4Kbyte pages.