问题
I am interested in the prospect of using memory mapped IO, preferably exploiting the facilities in boost::interprocess for cross-platform support, to map non-contiguous system-page-size blocks in a file into a contiguous address space in memory.
A simplified concrete scenario:
I've a number of 'plain-old-data' structures, each of a fixed length (less than the system page size.) These structures are concatenated into a (very long) stream with the type & location of structures determined by the values of those structures that proceed them in the stream. I'm aiming to minimize latency and maximize throughput in a demanding concurrent environment.
I can read this data very effectively by memory-mapping it in blocks of at least twice the system-page-size... and establishing a new mapping immediately having read a structure extending beyond the penultimate system-page-boundary. This allows the code that interacts with the plain-old-data structures to be blissfully unaware that these structures are memory mapped... and, for example, could compare two different structures using memcmp() directly without having to care about page boundaries.
Where things get interesting is with respect to updating these data streams... while they're being (concurrently) read. The strategy I'd like to use is inspired by 'Copy On Write' on a system-page-size granularity... essentially writing 'overlay-pages' - allowing one process to read the old data while another reads the updated data.
While managing which overlay pages to use, and when, isn't necessarily trivial... that's not my main concern. My main concern is that I may have a structure spanning pages 4 and 5, then update a structure wholly contained in page 5... writing the new page in location 6... leaving page 5 to be 'garbage collected' when it is determined to be no-longer reachable. This means that, if I map page 4 into location M, I need to map page 6 into memory location M+page_size... in order to be able to reliably process structures that cross page boundaries using existing (non-memory-mapping-aware) functions.
I'm trying to establish the best strategy, and I'm hampered by documentation I feel is incomplete. Essentially, I need to decouple allocation of address space from memory mapping into that address space. With mmap(), I'm aware that I can use MAP_FIXED - if I wish to explicitly control the mapping location... but I'm unclear how I should reserve address space in order to do this safely. Can I map /dev/zero for two pages without MAP_FIXED, then use MAP_FIXED twice to map two pages into that allocated space at explicit VM addresses? If so, should I call munmap() three times too? Will it leak resources and/or have any other untoward overhead? To make the issue even more complex, I'd like comparable behaviour on Windows... is there any way to do this? Are there neat solutions if I were to compromise my cross-platform ambitions?
--
Thanks for your answer, Mahmoud... I've read, and think I've understood that code... I've compiled it under Linux and it behaves as you suggest.
My main concerns are with line 62 - using MAP_FIXED. It makes some assumptions about mmap, which I've been unable to confirm when I read the documentation I can find. You're mapping the 'update' page into the same address space as mmap() returned initially - I assume that this is 'correct' - i.e. not something that just happens to work on Linux? I'd also need to assume that it works cross-platform for file-mappings as well as anonymous mappings.
The sample definitely moves me forwards... documenting that what I ultimately need is probably achievable with mmap() on Linux - at least. What I'd really like is a pointer to documentation that shows that the MAP_FIXED line will work as the sample demonstrates... and, idealy, a transformation from the Linux/Unix specific mmap() to a platform independent (Boost::interprocess) approach.
回答1:
Your question is a little confusing. From what I understood, this code will do what you need:
#define PAGESIZE 4096
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/mman.h>
#include <errno.h>
#include <sys/types.h>
#include <fcntl.h>
#include <unistd.h>
#include <assert.h>
struct StoredObject
{
int IntVal;
char StrVal[25];
};
int main(int argc, char **argv)
{
int fd = open("mmapfile", O_RDWR | O_CREAT | O_TRUNC, (mode_t) 0600);
//Set the file to the size of our data (2 pages)
lseek(fd, PAGESIZE*2 - 1, SEEK_SET);
write(fd, "", 1); //The final byte
unsigned char *mapPtr = (unsigned char *) mmap(0, PAGESIZE * 2, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
struct StoredObject controlObject;
controlObject.IntVal = 12;
strcpy(controlObject.StrVal, "Mary had a little lamb.\n");
struct StoredObject *mary1;
mary1 = (struct StoredObject *)(mapPtr + PAGESIZE - 4); //Will fall on the boundary between first and second page
memcpy(mary1, &controlObject, sizeof(StoredObject));
printf("%d, %s", mary1->IntVal, mary1->StrVal);
//Should print "12, Mary had a little lamb.\n"
struct StoredObject *john1;
john1 = mary1 + 1; //Comes immediately after mary1 in memory; will start and end in the second page
memcpy(john1, &controlObject, sizeof(StoredObject));
john1->IntVal = 42;
strcpy(john1->StrVal, "John had a little lamb.\n");
printf("%d, %s", john1->IntVal, john1->StrVal);
//Should print "12, Mary had a little lamb.\n"
//Make sure the data's on the disk, as this is the initial, "read-only" data
msync(mapPtr, PAGESIZE * 2, MS_SYNC);
//This is the inital data set, now in memory, loaded across two pages
//At this point, someone could be reading from there. We don't know or care.
//We want to modify john1, but don't want to write over the existing data
//Easy as pie.
//This is the shadow map. COW-like optimization will take place:
//we'll map the entire address space from the shared source, then overlap with a new map to modify
//This is mapped anywhere, letting the system decide what address we'll be using for the new data pointer
unsigned char *mapPtr2 = (unsigned char *) mmap(0, PAGESIZE * 2, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
//Map the second page on top of the first mapping; this is the one that we're modifying. It is *not* backed by disk
unsigned char *temp = (unsigned char *) mmap(mapPtr2 + PAGESIZE, PAGESIZE, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_FIXED | MAP_ANON, 0, 0);
if (temp == MAP_FAILED)
{
printf("Fixed map failed. %s", strerror(errno));
}
assert(temp == mapPtr2 + PAGESIZE);
//Make a copy of the old data that will later be changed
memcpy(mapPtr2 + PAGESIZE, mapPtr + PAGESIZE, PAGESIZE);
//The two address spaces should still be identical until this point
assert(memcmp(mapPtr, mapPtr2, PAGESIZE * 2) == 0);
//We can now make our changes to the second page as needed
struct StoredObject *mary2 = (struct StoredObject *)(((unsigned char *)mary1 - mapPtr) + mapPtr2);
struct StoredObject *john2 = (struct StoredObject *)(((unsigned char *)john1 - mapPtr) + mapPtr2);
john2->IntVal = 52;
strcpy(john2->StrVal, "Mike had a little lamb.\n");
//Test that everything worked OK
assert(memcmp(mary1, mary2, sizeof(struct StoredObject)) == 0);
printf("%d, %s", john2->IntVal, john2->StrVal);
//Should print "52, Mike had a little lamb.\n"
//Now assume our garbage collection routine has detected that no one is using the original copy of the data
munmap(mapPtr, PAGESIZE * 2);
mapPtr = mapPtr2;
//Now we're done with all our work and want to completely clean up
munmap(mapPtr2, PAGESIZE * 2);
close(fd);
return 0;
}
My modified answer should address your safety concerns.
Only use MAP_FIXED on the second mmap call (like I have above). The cool thing about MAP_FIXED is that it lets you overwrite an existing mmap address section. It'll unload the range you're overlapping and replace it with your new mapped content:
MAP_FIXED
[...] If the memory
region specified by addr and len overlaps pages of any existing
mapping(s), then the overlapped part of the existing mapping(s) will be
discarded. [...]
This way, you let the OS take care of finding a contiguous memory block of hundreds of megs for you (never call MAP_FIXED on address you don't know for sure isn't available). Then you call MAP_FIXED on a subsection of that now-mapped huge space with the data that you will be modifying. Tada.
On Windows, something like this should work (I'm on a Mac at the moment, so untested):
int main(int argc, char **argv)
{
HANDLE hFile = CreateFile(L"mmapfile", GENERIC_READ | GENERIC_WRITE, FILE_SHARE_READ, NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
//Set the file to the size of our data (2 pages)
SetFilePointer(hFile, PAGESIZE*2 - 1, 0, FILE_BEGIN);
DWORD bytesWritten = -1;
WriteFile(hFile, "", 1, &bytesWritten, NULL);
HANDLE hMap = CreateFileMapping(hFile, NULL, PAGE_READWRITE, 0, PAGESIZE * 2, NULL);
unsigned char *mapPtr = (unsigned char *) MapViewOfFile(hMap, FILE_MAP_READ | FILE_MAP_WRITE, 0, 0, PAGESIZE * 2);
struct StoredObject controlObject;
controlObject.IntVal = 12;
strcpy(controlObject.StrVal, "Mary had a little lamb.\n");
struct StoredObject *mary1;
mary1 = (struct StoredObject *)(mapPtr + PAGESIZE - 4); //Will fall on the boundary between first and second page
memcpy(mary1, &controlObject, sizeof(StoredObject));
printf("%d, %s", mary1->IntVal, mary1->StrVal);
//Should print "12, Mary had a little lamb.\n"
struct StoredObject *john1;
john1 = mary1 + 1; //Comes immediately after mary1 in memory; will start and end in the second page
memcpy(john1, &controlObject, sizeof(StoredObject));
john1->IntVal = 42;
strcpy(john1->StrVal, "John had a little lamb.\n");
printf("%d, %s", john1->IntVal, john1->StrVal);
//Should print "12, Mary had a little lamb.\n"
//Make sure the data's on the disk, as this is the initial, "read-only" data
//msync(mapPtr, PAGESIZE * 2, MS_SYNC);
//This is the inital data set, now in memory, loaded across two pages
//At this point, someone could be reading from there. We don't know or care.
//We want to modify john1, but don't want to write over the existing data
//Easy as pie.
//This is the shadow map. COW-like optimization will take place:
//we'll map the entire address space from the shared source, then overlap with a new map to modify
//This is mapped anywhere, letting the system decide what address we'll be using for the new data pointer
unsigned char *reservedMem = (unsigned char *) VirtualAlloc(NULL, PAGESIZE * 2, MEM_RESERVE, PAGE_READWRITE);
HANDLE hMap2 = CreateFileMapping(hFile, NULL, PAGE_READWRITE, 0, PAGESIZE, NULL);
unsigned char *mapPtr2 = (unsigned char *) MapViewOfFileEx(hMap2, FILE_MAP_READ | FILE_MAP_WRITE, 0, 0, PAGESIZE, reservedMem);
//Map the second page on top of the first mapping; this is the one that we're modifying. It is *not* backed by disk
unsigned char *temp = (unsigned char *) MapViewOfFileEx(hMap2, FILE_MAP_READ | FILE_MAP_WRITE, 0, 0, PAGESIZE, reservedMem + PAGESIZE);
if (temp == NULL)
{
printf("Fixed map failed. 0x%x\n", GetLastError());
return -1;
}
assert(temp == mapPtr2 + PAGESIZE);
//Make a copy of the old data that will later be changed
memcpy(mapPtr2 + PAGESIZE, mapPtr + PAGESIZE, PAGESIZE);
//The two address spaces should still be identical until this point
assert(memcmp(mapPtr, mapPtr2, PAGESIZE * 2) == 0);
//We can now make our changes to the second page as needed
struct StoredObject *mary2 = (struct StoredObject *)(((unsigned char *)mary1 - mapPtr) + mapPtr2);
struct StoredObject *john2 = (struct StoredObject *)(((unsigned char *)john1 - mapPtr) + mapPtr2);
john2->IntVal = 52;
strcpy(john2->StrVal, "Mike had a little lamb.\n");
//Test that everything worked OK
assert(memcmp(mary1, mary2, sizeof(struct StoredObject)) == 0);
printf("%d, %s", john2->IntVal, john2->StrVal);
//Should print "52, Mike had a little lamb.\n"
//Now assume our garbage collection routine has detected that no one is using the original copy of the data
//munmap(mapPtr, PAGESIZE * 2);
mapPtr = mapPtr2;
//Now we're done with all our work and want to completely clean up
//munmap(mapPtr2, PAGESIZE * 2);
//close(fd);
return 0;
}
回答2:
but I'm unclear how I should reserve address space in order to do this safely
That's going to vary by OS, but a little digging on msdn for mmap (I started with "xp mmap" on the msdn search) shows Microsoft have their usual VerboseAndHelpfullyCapitalizedNames for (the many) functions that implement pieces of mmap. Both the file- and anonymous- mappers can handle fixed-address requests just the same as any POSIX-2001 system can, i.e. if something else in your address space is talking to the kernel, you get to sort it out. No way I'm going to touch "safely", there's no such thing as "safely" with code you're wanting to port to unspecified platforms. You're going to have to build your own pool of pre-mapped anonymous memory that you can unmap and parcel out later under your own control.
回答3:
I tested the windows code from @Mahmoud, well actually I tested the following similar code, and it doesn't work (the Linux code works.) If you uncomment VirtualFree, it will work. As mentioned in my comment above, on windows you can reserve the address space with VirtualAlloc, but you can't use MapViewOfFileEx with an already mapped address, so you need to VirtualFree it first. Then there's a race condition where another thread can grab the memory address before you do, so you have to do everything in a loop, e.g. try up to 1000 times and then give up.
package main
import (
"fmt"
"os"
"syscall"
)
func main() {
const size = 1024 * 1024
file, err := os.Create("foo.dat")
if err != nil {
panic(err)
}
if err := file.Truncate(size); err != nil {
panic(err)
}
const MEM_COMMIT = 0x1000
addr, err := virtualAlloc(0, size, MEM_COMMIT, protReadWrite)
if err != nil {
panic(err)
}
fd, err := syscall.CreateFileMapping(
syscall.Handle(file.Fd()),
nil,
uint32(protReadWrite),
0,
uint32(size),
nil,
)
//if err := virtualFree(addr); err != nil {
// panic(err)
//}
base, err := mapViewOfFileEx(fd, syscall.FILE_MAP_READ|syscall.FILE_MAP_WRITE, 0, 0, size, addr)
if base == 0 {
panic("mapViewOfFileEx returned 0")
}
if err != nil {
panic(err)
}
fmt.Println("success!")
}
type memProtect uint32
const (
protReadOnly memProtect = 0x02
protReadWrite memProtect = 0x04
protExecute memProtect = 0x20
protAll memProtect = 0x40
)
var (
modkernel32 = syscall.MustLoadDLL("kernel32.dll")
procMapViewOfFileEx = modkernel32.MustFindProc("MapViewOfFileEx")
procVirtualAlloc = modkernel32.MustFindProc("VirtualAlloc")
procVirtualFree = modkernel32.MustFindProc("VirtualFree")
procVirtualProtect = modkernel32.MustFindProc("VirtualProtect")
)
func mapViewOfFileEx(handle syscall.Handle, prot memProtect, offsetHigh uint32, offsetLow uint32, length uintptr, target uintptr) (addr uintptr, err error) {
r0, _, e1 := syscall.Syscall6(procMapViewOfFileEx.Addr(), 6, uintptr(handle), uintptr(prot), uintptr(offsetHigh), uintptr(offsetLow), length, target)
addr = uintptr(r0)
if addr == 0 {
if e1 != 0 {
err = error(e1)
} else {
err = syscall.EINVAL
}
}
return addr, nil
}
func virtualAlloc(addr, size uintptr, allocType uint32, prot memProtect) (mem uintptr, err error) {
r0, _, e1 := syscall.Syscall6(procVirtualAlloc.Addr(), 4, addr, size, uintptr(allocType), uintptr(prot), 0, 0)
mem = uintptr(r0)
if e1 != 0 {
return 0, error(e1)
}
return mem, nil
}
func virtualFree(addr uintptr) error {
const MEM_RELEASE = 0x8000
_, _, e1 := syscall.Syscall(procVirtualFree.Addr(), 3, addr, 0, MEM_RELEASE)
if e1 != 0 {
return error(e1)
}
return nil
}
来源：https://stackoverflow.com/questions/10454964/mapping-non-contiguous-blocks-from-a-file-into-contiguous-memory-addresses