Shared Memory
Shared Memory
12.1
Introduction
Shared memory is the fastest form of IPC available. Once the memory is mapped into the address space of the processes that are sharing the memory region, no kernel involvement occurs in passing data between the processes. What is normally required, however, is some form of synchronization between the processes that are storing and fetching information to and from the shared memory region. In Part 3, we discussed various forms of synchronization: mutexes, condition variables, readwrite locks, record locks, and semaphores.
What we mean by no kernel involvement is that the processes do not execute any system calls into the kernel to pass the data. Obviously, the kernel must establish the memory mappings that allow the processes to share the memory, and then manage this memory over time (handle page faults, and the like).
Consider the normal steps involved in the clientserver le copying program that we used as an example for the various types of message passing (Figure 4.1). The server reads from the input le. The le data is read by the kernel into its memory and then copied from the kernel to the process. The server writes this data in a message, using a pipe, FIFO, or message queue. These forms of IPC normally require the data to be copied from the process to the kernel.
We use the qualier normally because Posix message queues can be implemented using memory-mapped I/O (the mmap function that we describe in this chapter), as we showed in Section 5.8 and as we show in the solution to Exercise 12.2. In Figure 12.1, we assume
303
304
Chapter 12
that Posix message queues are implemented within the kernel, which is another possibility. But pipes, FIFOs, and System V message queues all involve copying the data from the process to the kernel for a write or msgsnd, or copying the data from the kernel to the process for a read or msgrcv.
The client reads the data from the IPC channel, normally requiring the data to be copied from the kernel to the process. Finally, the data is copied from the clients buffer, the second argument to the write function, to the output le. A total of four copies of the data are normally required. Additionally, these four copies are done between the kernel and a process, often an expensive copy (more expensive than copying data within the kernel, or copying data within a single process). Figure 12.1 depicts this movement of the data between the client and server, through the kernel.
output le
IPC (pipe, FIFO, or message queue) Figure 12.1 Flow of le data from server to client.
input le
The problem with these forms of IPCpipes, FIFOs, and message queuesis that for two processes to exchange information, the information has to go through the kernel. Shared memory provides a way around this by letting two or more processes share a region of memory. The processes must, of course, coordinate or synchronize the use of the shared memory among themselves. (Sharing a common piece of memory is similar to sharing a disk le, such as the sequence number le used in all the le locking examples.) Any of the techniques described in Part 3 can be used for this synchronization. The steps for the clientserver example are now as follows: The server gets access to a shared memory object using (say) a semaphore. The server reads from the input le into the shared memory object. The second argument to the read, the address of the data buffer, points into the shared memory object. When the read is complete, the server noties the client, using a semaphore. The client writes the data from the shared memory object to the output le.
Section 12.1
Introduction
305
shared memory
server
output le
input le Figure 12.2 Copying le data from server to client using shared memory.
In this gure the data is copied only twicefrom the input le into shared memory and from shared memory to the output le. We draw one dashed box enclosing the client and the shared memory object, and another dashed box enclosing the server and the shared memory object, to reinforce that the shared memory object appears in the address space of both the client and the server. The concepts involved in using shared memory are similar for both the Posix interface and the System V interface. We describe the former in Chapter 13 and the latter in Chapter 14. In this chapter, we return to our sequence-number-increment example that we started in Chapter 9. But we now store the sequence number in memory instead of in a le. We rst reiterate that memory is not shared by default between a parent and child across a fork. The program in Figure 12.3 has a parent and child increment a global integer named count.
Create and initialize semaphore
1214
We create and initialize a semaphore that protects what we think is a shared variable (the global count). Since this assumption is false, this semaphore is not really needed. Notice that we remove the semaphore name from the system by calling sem_unlink, but although this removes the pathname, it has no effect on the semaphore that is already open. We do this so that the pathname is removed from the lesystem even if the program aborts.
Set standard output unbuffered and fork
15
1629
We set standard output unbuffered because both the parent and child will be writing to it. This prevents interleaving of the output from the two processes. The parent and child each execute a loop that increments the counter the specied number of times, being careful to increment the variable only when the semaphore is held.
306
Chapter 12
1 #include
"unpipc.h" "mysem"
shm / incr1.c
4 int 5 main(int argc, char **argv) 6 { 7 int i, nloop; 8 sem_t *mutex; 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 } if (argc != 2) err_quit("usage: incr1 <#loops>"); nloop = atoi(argv[1]); /* create, initialize, and unlink semaphore */ mutex = Sem_open(Px_ipc_name(SEM_NAME), O_CREAT | O_EXCL, FILE_MODE, 1); Sem_unlink(Px_ipc_name(SEM_NAME)); setbuf(stdout, NULL); /* stdout is unbuffered */ if (Fork() == 0) { /* child */ for (i = 0; i < nloop; i++) { Sem_wait(mutex); printf("child: %d\n", count++); Sem_post(mutex); } exit(0); } /* parent */ for (i = 0; i < nloop; i++) { Sem_wait(mutex); printf("parent: %d\n", count++); Sem_post(mutex); } exit(0);
shm / incr1.c
Figure 12.3 Parent and child both increment the same global.
If we run this program and look only at the output when the system switches between the parent and child, we have the following:
child: 0 child: 1 . . . child: 678 child: 679 parent: 0 parent: 1 . . . parent: 1220 parent: 1221 child: 680 child runs rst, counter starts at 0
Section 12.2
307
child: 681 . . . child: 2078 child: 2079 parent: 1222 parent: 1223
As we can see, both processes have their own copy of the global count. Each starts with the value of this variable as 0, and each increments its own copy of this variable. Figure 12.4 shows the parent before calling fork.
parent int count; parent executing here if (Fork() == 0) { /* child */ ... } /* parent */ ... Figure 12.4 Parent before calling fork.
When fork is called, the child starts with its own copy of the parents data space. Figure 12.5 shows the two processes after fork returns.
parent int count; if (Fork() == 0) { /* child */ ... } /* parent */ ... Figure 12.5 Parent and child after fork returns. child int count; if (Fork() == 0) { /* child */ ... } /* parent */ ...
We see that the parent and child each have their own copy of the variable count.
12.2
308
Chapter 12
1. with a regular le to provide memory-mapped I/O (Section 12.3), 2. with special les to provide anonymous memory mappings (Sections 12.4 and 12.5), and 3. with shm_open to provide Posix shared memory between unrelated processes (Chapter 13).
#include <sys/mman.h> void *mmap(void *addr, size_t len, int prot, int ags, int fd, off_t offset); Returns: starting address of mapped region if OK, MAP_FAILED on error
addr can specify the starting address within the process of where the descriptor fd should be mapped. Normally, this is specied as a null pointer, telling the kernel to choose the starting address. In any case, the return value of the function is the starting address of where the descriptor has been mapped. len is the number of bytes to map into the address space of the process, starting at offset bytes from the beginning of the le. Normally, offset is 0. Figure 12.6 shows this mapping.
address space of process high memory
The protection of the memory-mapped region is specied by the prot argument using the constants in Figure 12.7. The common value for this argument is PROT_READ | PROT_WRITE for readwrite access.
Section 12.2
309
Description data can be read data can be written data can be executed data cannot be accessed
Description changes are shared changes are private interpret the addr argument exactly
The ags are specied by the constants in Figure 12.8. Either the MAP_SHARED or the MAP_PRIVATE ag must be specied, optionally ORed with MAP_FIXED. If MAP_PRIVATE is specied, then modications to the mapped data by the calling process are visible only to that process and do not change the underlying object (either a le object or a shared memory object). If MAP_SHARED is specied, modications to the mapped data by the calling process are visible to all processes that are sharing the object, and these changes do modify the underlying object. For portability, MAP_FIXED should not be specied. If it is not specied, but addr is not a null pointer, then it is implementation dependent as to what the implementation does with addr. The nonnull value of addr is normally taken as a hint about where the memory should be located. Portable code should specify addr as a null pointer and should not specify MAP_FIXED. One way to share memory between a parent and child is to call mmap with MAP_SHARED before calling fork. Posix.1 then guarantees that memory mappings in the parent are retained in the child. Furthermore, changes made by the parent are visible to the child and vice versa. We show an example of this shortly. After mmap returns success, the fd argument can be closed. This has no effect on the mapping that was established by mmap. To remove a mapping from the address space of the process, we call munmap.
#include <sys/mman.h> int munmap(void *addr, size_t len); Returns: 0 if OK, 1 on error
The addr argument is the address that was returned by mmap, and the len is the size of that mapped region. Further references to these addresses result in the generation of a SIGSEGV signal to the process (assuming, of course, that a later call to mmap does not reuse this portion of the address space).
310
Chapter 12
If the mapped region was mapped using MAP_PRIVATE, the changes made are discarded. In Figure 12.6, the kernels virtual memory algorithm keeps the memory-mapped le (typically on disk) synchronized with the memory-mapped region in memory, assuming a MAP_SHARED segment. That is, if we modify a location in memory that is memory-mapped to a le, then at some time later the kernel will update the le accordingly. But sometimes, we want to make certain that the le on disk corresponds to what is in the memory-mapped region, and we call msync to perform this synchronization.
#include <sys/mman.h> int msync(void *addr, size_t len, int ags); Returns: 0 if OK, 1 on error
The addr and len arguments normally refer to the entire memory-mapped region of memory, although subsets of this region can also be specied. The ags argument is formed from the combination of constants shown in Figure 12.9.
Constant MS_ASYNC MS_SYNC MS_INVALIDATE Description perform asynchronous writes perform synchronous writes invalidate cached data
One of the two constants MS_ASYNC and MS_SYNC must be specied, but not both. The difference in these two is that MS_ASYNC returns once the write operations are queued by the kernel, whereas MS_SYNC returns only after the write operations are complete. If MS_INVALIDATE is also specied, all in-memory copies of the le data that are inconsistent with the le data are invalidated. Subsequent references will obtain data from the le. Why Use mmap? Our description of mmap so far has implied a memory-mapped le: some le that we open and then map into our address space by calling mmap. The nice feature in using a memory-mapped le is that all the I/O is done under the covers by the kernel, and we just write code that fetches and stores values in the memory-mapped region. We never call read, write, or lseek. Often, this can simplify our code.
Recall our implementation of Posix message queues using mmap and the storing of values into a msg_hdr structure in Figure 5.30 and the fetching of values from a msg_hdr structure in Figure 5.32.
Section 12.3
311
Beware of some caveats, however, in that not all les can be memory mapped. Trying to map a descriptor that refers to a terminal or a socket, for example, generates an error return from mmap. These types of descriptors must be accessed using read and write (or variants thereof ). Another use of mmap is to provide shared memory between unrelated processes. In this case, the actual contents of the le become the initial contents of the memory that is shared, and any changes made by the processes to this shared memory are then copied back to the le (providing lesystem persistence). This assumes that MAP_SHARED is specied, which is required to share the memory between processes.
Details on the implementation of mmap and its relationship to the kernels virtual memory algorithms are provided in [McKusick et al. 1996] for 4.4BSD and in [Vahalia 1996] and [Goodheart and Cox 1994] for SVR4.
12.3
1114
We have a new command-line argument that is the name of a le that will be memory mapped. We open the le for reading and writing, creating the le if it does not exist, and then write an integer with a value of 0 to the le.
mmap then close descriptor
1516
We call mmap to map the le that was just opened into the memory of this process. The rst argument is a null pointer, telling the system to pick the starting address. The length is the size of an integer, and we specify readwrite access. By specifying a fourth argument of MAP_SHARED, any changes made by the parent will be seen by the child, and vice versa. The return value is the starting address of the memory region that will be shared, and we store it in ptr.
fork
2034
We set standard output unbuffered and call fork. The parent and child both increment the integer counter pointed to by ptr. Memory-mapped les are handled specially by fork, in that memory mappings created by the parent before calling fork are shared by the child. Therefore, what we have done by opening the le and calling mmap with the MAP_SHARED ag is provide a piece of memory that is shared between the parent and child. Furthermore, since the shared memory is a memory-mapped le, any changes to the shared memory (the piece of memory pointed to by ptr of size sizeof(int)) are also reected in the actual le (whose name was specied by the command-line argument).
312
Chapter 12
1 #include
"unpipc.h" "mysem"
shm / incr2.c
2 #define SEM_NAME
3 int 4 main(int argc, char **argv) 5 { 6 int fd, i, nloop, zero = 0; 7 int *ptr; 8 sem_t *mutex; 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 } if (argc != 3) err_quit("usage: incr2 <pathname> <#loops>"); nloop = atoi(argv[2]); /* open file, initialize to 0, map into memory */ fd = Open(argv[1], O_RDWR | O_CREAT, FILE_MODE); Write(fd, &zero, sizeof(int)); ptr = Mmap(NULL, sizeof(int), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); Close(fd); /* create, initialize, and unlink semaphore */ mutex = Sem_open(Px_ipc_name(SEM_NAME), O_CREAT | O_EXCL, FILE_MODE, 1); Sem_unlink(Px_ipc_name(SEM_NAME)); setbuf(stdout, NULL); /* stdout is unbuffered */ if (Fork() == 0) { /* child */ for (i = 0; i < nloop; i++) { Sem_wait(mutex); printf("child: %d\n", (*ptr)++); Sem_post(mutex); } exit(0); } /* parent */ for (i = 0; i < nloop; i++) { Sem_wait(mutex); printf("parent: %d\n", (*ptr)++); Sem_post(mutex); } exit(0);
shm / incr2.c
Figure 12.10 Parent and child incrementing a counter in shared memory.
If we execute this program, we see that the memory pointed to by ptr is indeed shared between the parent and child. We show only the values when the kernel switches between the two processes.
solaris % incr2 /tmp/temp.1 10000 child: 0 child starts rst child: 1 . . . child: 128 child: 129 parent: 130 child is stopped, parent starts
Section 12.3
313
parent: 131 . . . parent: 636 parent: 637 child: 638 parent is stopped, child starts child: 639 . . . child: 1517 child: 1518 parent: 1519 child is stopped, parent starts parent: 1520 . . . parent: 19999 nal line of output solaris % od -D /tmp/temp.1 0000000 0000020000 0000004
Since the le was memory mapped, we can look at the le after the program terminates with the od program and see that the nal value of the counter (20,000) is indeed stored in the le. Figure 12.11 is a modication of Figure 12.5 showing the shared memory, and showing that the semaphore is also shared. We show the semaphore as being in the kernel, but as we mentioned with Posix semaphores, this is not a requirement. Whatever implementation is used, the semaphore must have at least kernel persistence. The semaphore could be stored as another memory-mapped le, as we demonstrated in Section 10.15.
shared memory count parent int *ptr; if (Fork() == 0) { /* child */ ... } /* parent */ ... child int *ptr; if (Fork() == 0) { /* child */ ... } /* parent */ ... process kernel semaphore: 0 or 1
We show that the parent and child each have their own copy of the pointer ptr, but each copy points to the same integer in shared memory: the counter that both processes increment.
314
Chapter 12
We now modify our program from Figure 12.10 to use a Posix memory-based semaphore instead of a Posix named semaphore, and store this semaphore in the shared memory. Figure 12.12 is the new program.
1 #include "unpipc.h"
shm / incr3.c
2 struct shared { 3 sem_t mutex; 4 int count; 5 } shared; 6 int 7 main(int argc, char **argv) 8 { 9 int fd, i, nloop; 10 struct shared *ptr; 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 }
if (argc != 3) err_quit("usage: incr3 <pathname> <#loops>"); nloop = atoi(argv[2]); /* open file, initialize to 0, map into memory */ fd = Open(argv[1], O_RDWR | O_CREAT, FILE_MODE); Write(fd, &shared, sizeof(struct shared)); ptr = Mmap(NULL, sizeof(struct shared), PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); Close(fd); /* initialize semaphore that is shared between processes */ Sem_init(&ptr->mutex, 1, 1); setbuf(stdout, NULL); /* stdout is unbuffered */ if (Fork() == 0) { /* child */ for (i = 0; i < nloop; i++) { Sem_wait(&ptr->mutex); printf("child: %d\n", ptr->count++); Sem_post(&ptr->mutex); } exit(0); } /* parent */ for (i = 0; i < nloop; i++) { Sem_wait(&ptr->mutex); printf("parent: %d\n", ptr->count++); Sem_post(&ptr->mutex); } exit(0);
shm / incr3.c
Figure 12.12 Counter and semaphore are both in shared memory.
We dene a structure containing the integer counter and a semaphore to protect it. This structure will be stored in the shared memory object.
Section 12.4
315
We create the le that will be mapped, and write a structure of 0 to the le. All we are doing is initializing the counter, because the value of the semaphore will be initialized by the call to sem_init. Nevertheless, writing an entire structure of 0 is simpler than to try to write only an integer of 0.
Initialize semaphore
2021
We are now using a memory-based semaphore, instead of a named semaphore, so we call sem_init to initialize its value to 1. The second argument must be nonzero, to indicate that the semaphore is being shared between processes. Figure 12.13 is a modication of Figure 12.11, noting the change that the semaphore has moved from the kernel into shared memory.
shared memory count & semaphore parent struct shared *ptr; if (Fork() == 0) { /* child */ ... } /* parent */ ... child struct shared *ptr; if (Fork() == 0) { /* child */ ... } /* parent */ ...
12.4
316
Chapter 12
(Many Berkeley-derived implementations, such as SunOS 4.1.x and BSD/OS 3.1, also support /dev/zero.) Figure 12.14 shows the only portion of Figure 12.10 that changes when we use 4.4BSD anonymous memory mapping.
3 int 4 main(int argc, char **argv) 5 { 6 int i, nloop; 7 int *ptr; 8 sem_t *mutex; 9 10 11 12 13 14 if (argc != 2) err_quit("usage: incr_map_anon <#loops>"); nloop = atoi(argv[1]); /* map into memory */ ptr = Mmap(NULL, sizeof(int), PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0);
Figure 12.14 4.4BSD anonymous memory mapping.
611
shm / incr_map_anon.c
shm / incr_map_anon.c
1214
The automatic variables fd and zero are gone, as is the command-line argument that specied the pathname that was created. We no longer open a le. The MAP_ANON ag is specied in the call to mmap, and the fth argument (the descriptor) is 1.
12.5
shm / incr_dev_zero.c
shm / incr_dev_zero.c
Section 12.6
317
611
1215
The automatic variable zero is gone, as is the command-line argument that specied the pathname that was created. We open /dev/zero, and the descriptor is then used in the call to mmap. We are guaranteed that the memory-mapped region is initialized to 0.
12.6
shm / test1.c
2 int 3 main(int argc, char **argv) 4 { 5 int fd, i; 6 char *ptr; 7 size_t filesize, mmapsize, pagesize; 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 } if (argc != 4) err_quit("usage: test1 <pathname> <filesize> <mmapsize>"); filesize = atoi(argv[2]); mmapsize = atoi(argv[3]); /* open file: create or truncate; set file size */ fd = Open(argv[1], O_RDWR | O_CREAT | O_TRUNC, FILE_MODE); Lseek(fd, filesize - 1, SEEK_SET); Write(fd, "", 1); ptr = Mmap(NULL, mmapsize, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); Close(fd); pagesize = Sysconf(_SC_PAGESIZE); printf("PAGESIZE = %ld\n", (long) pagesize); for (i = 0; i < max(filesize, mmapsize); i += pagesize) { printf("ptr[%d] = %d\n", i, ptr[i]); ptr[i] = 1; printf("ptr[%d] = %d\n", i + pagesize - 1, ptr[i + pagesize - 1]); ptr[i + pagesize - 1] = 1; } printf("ptr[%d] = %d\n", i, ptr[i]); exit(0);
shm / test1.c
Figure 12.16 Memory mapping when mmap equals le size.
318
Chapter 12
Command-line arguments
811
The command-line arguments specify the pathname of the le that will be created and memory mapped, the size to which that le is set, and the size of the memory mapping.
Create, open, truncate le; set le size
1215
The le being opened is created if it does not exist, or truncated to a size of 0 if it already exists. The size of the le is then set to the specied size by seeking to that size minus 1 byte and writing 1 byte.
Memory map le
1617
The le is memory mapped, using the size specied as the nal command-line argument. The descriptor is then closed.
Print page size
1819
The page size of the implementation is obtained using sysconf and printed.
Read and store the memory-mapped region
2026
The memory-mapped region is read (the rst byte of each page and the last byte of each page), and the values printed. We expect the values to all be 0. We also set the rst and last bytes of the page to 1. We expect one of the references to generate a signal eventually, which will terminate the program. When the for loop terminates, we print the rst byte of the next page, expecting this to fail (assuming that the program has not already failed). The rst scenario that we show is when the le size equals the memory-mapped size, but this size is not a multiple of the page size.
solaris % ls -l foo foo: No such file or directory solaris % test1 foo 5000 5000 PAGESIZE = 4096 ptr[0] = 0 ptr[4095] = 0 ptr[4096] = 0 ptr[8191] = 0 Segmentation Fault(coredump) solaris % ls -l foo -rw-r--r-1 rstevens other1 solaris % od -b -A d foo 0000000 001 000 000 000 000 000 0000016 000 000 000 000 000 000 * 0004080 000 000 000 000 000 000 0004096 001 000 000 000 000 000 0004112 000 000 000 000 000 000 * 0005000
5000 Mar 20 17:18 foo 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 001 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
The page size is 4096 bytes, and we are able to read the entire second page (indexes 4096 through 8191), but a reference to the third page (index 8192) generates SIGSEGV, which
Section 12.6
319
the shell prints as Segmentation Fault. Even though we set ptr[8191] to 1, this value is not written to the le, and the les size remains 5000. The kernel lets us read and write that portion of the nal page beyond our mapping (since the kernels memory protection works with pages), but anything that we write to this extension is not written to the le. The other 3 bytes that we set to 1, indexes 0, 4095, and 4096, are copied back to the le, which we verify with the od command. (The -b option says to print the bytes in octal, and the -A d option says to print the addresses in decimal.) Figure 12.17 depicts this example.
le size
references OK
mmap() size Figure 12.17 Memory mapping when mmap size equals le size.
If we run our example under Digital Unix, we see similar results, but the page size is now 8192.
alpha % ls -l foo foo not found alpha % test1 foo 5000 5000 PAGESIZE = 8192 ptr[0] = 0 ptr[8191] = 0 Memory fault(coredump) alpha % ls -l foo -rw-r--r-1 rstevens operator
We are still able to reference beyond the end of our memory-mapped region but within that page of memory (indexes 5000 through 8191). Referencing ptr[8192] generates SIGSEGV, as we expect. In our next example with Figure 12.16, we specify a memory mapping (15000 bytes) that is larger than the le size (5000 bytes).
solaris % rm foo solaris % test1 foo 5000 15000 PAGESIZE = 4096 ptr[0] = 0 ptr[4095] = 0 ptr[4096] = 0
320
Chapter 12
The results are similar to our earlier example when the le size and the memory map size were the same (both 5000). This example generates SIGBUS (which the shell prints as Bus Error ), whereas the previous example generated SIGSEGV. The difference is that SIGBUS means we have referenced within our memory-mapped region but beyond the size of the underlying object. The SIGSEGV in the previous example meant we had referenced beyond the end of our memory-mapped region. What we have shown here is that the kernel knows the size of the underlying object that is mapped (the le foo in this case), even though we have closed the descriptor for that object. The kernel allows us to specify a size to mmap that is larger than the size of this object, but we cannot reference beyond its end (except for the bytes within the nal page that are beyond the end of the object, indexes 5000 through 8191). Figure 12.18 depicts this example.
le size
le offset: 0 index: 0 4999 4999 5000 8191 8192 mmap() size 14999
Our next program is shown in Figure 12.19. It shows a common technique for handling a le that is growing: specify a memory-map size that is larger than the le, keep track of the les current size (making certain not to reference beyond the current endof-le), and then just let the les size increase as more data is written to the le.
Open le
911
We open a le, creating it if it does not exist or truncating it if it already exists. The le is memory mapped with a size of 32768, even though the les current size is 0.
Increase le size
1216
We increase the size of the le, 4096 bytes at a time, by calling ftruncate (Section 13.3), and fetch the byte that is now the nal byte of the le.
Section 12.6
321
shm / test2.c
4 int 5 main(int argc, char **argv) 6 { 7 int fd, i; 8 char *ptr; 9 10 11 12 13 14 15 16 17 18 } /* open: create or truncate; then mmap file */ fd = Open(FILE, O_RDWR | O_CREAT | O_TRUNC, FILE_MODE); ptr = Mmap(NULL, SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); for (i = 4096; i <= SIZE; i += 4096) { printf("setting file size to %d\n", i); Ftruncate(fd, i); printf("ptr[%d] = %d\n", i - 1, ptr[i - 1]); } exit(0);
shm / test2.c
Figure 12.19 Memory map example that lets the le size grow.
When we run this program, we see that as we increase the size of the le, we are able to reference the new data through our established memory map.
alpha % ls -l test.data test.data: No such file or directory alpha % test2 setting file size to 4096 ptr[4095] = 0 setting file size to 8192 ptr[8191] = 0 setting file size to 12288 ptr[12287] = 0 setting file size to 16384 ptr[16383] = 0 setting file size to 20480 ptr[20479] = 0 setting file size to 24576 ptr[24575] = 0 setting file size to 28672 ptr[28671] = 0 setting file size to 32768 ptr[32767] = 0 alpha % ls -l test.data -rw-r--r-1 rstevens other1 32768 Mar 20 17:53 test.data
322
Chapter 12
This example shows that the kernel keeps track of the size of the underlying object that is memory mapped (the le test.data in this example), and we are always able to reference bytes that are within the current le size that are also within our memory map. We obtain identical results under Solaris 2.6. This section has dealt with memory-mapped les and mmap. In Exercise 13.1, we modify our two programs to work with Posix shared memory and see the same results.
12.7
Summary
Shared memory is the fastest form of IPC available, because one copy of the data in the shared memory is available to all the threads or processes that share the memory. Some form of synchronization is normally required, however, to coordinate the various threads or processes that are sharing the memory. This chapter has focused on the mmap function and the mapping of regular les into memory, because this is one way to share memory between related or unrelated processes. Once we have memory mapped a le, we no longer use read, write, or lseek to access the le; instead, we just fetch or store the memory locations that have been mapped to the le by mmap. Changing explicit le I/O into fetches and stores of memory can often simplify our programs and sometimes increase performance. When the memory is to be shared across a subsequent fork, this can be simplied by not creating a regular le to map, but using anonymous memory mapping instead. This involves either a new ag of MAP_ANON (for Berkeley-derived kernels) or mapping /dev/zero (for SVR4-derived kernels). Our reason for covering mmap in such detail is both because memory mapping of les is a useful technique and because mmap is used for Posix shared memory, which is the topic of the next chapter. Also available are four additional functions (that we do not cover) dened by Posix dealing with memory management: mlockall causes all of the memory of the process to be memory resident. munlockall undoes this locking. mlock causes a specied range of addresses of the process to be memory resident, where the function arguments are a starting address and a number of bytes from that address. munlock unlocks a specied region of memory.
Exercises
12.1 12.2 What would happen in Figure 12.19 if we executed the code within the for loop one more time? Assume that we have two processes, a sender and a receiver, with the former just sending messages to the latter. Assume that System V message queues are used and draw a diagram of how the messages go from the sender to the receiver. Now assume that our
Chapter 12
Exercises
323
implementation of Posix message queues from Section 5.8 is used, and draw a diagram of the transfer of messages. 12.3 With mmap and MAP_SHARED, we said that the kernel virtual memory algorithm updates the actual le with any modications that are made to the memory image. Read the manual page for /dev/zero to determine what happens when the kernel writes the changes back to the le. Modify Figure 12.10 to specify MAP_PRIVATE instead of MAP_SHARED, and verify that the results are similar to the results from Figure 12.3. What are the contents of the le that is memory mapped? In Section 6.9, we mentioned that one way to select on a System V message queue is to create a piece of anonymous shared memory, create a child, and let the child block in its call to msgrcv, reading the message into shared memory. The parent also creates two pipes; one is used by the child to notify the parent that a message is ready in shared memory, and the other pipe is used by the parent to notify the child that the shared memory is now available. This allows the parent to select on the read end of the pipe, along with any other descriptors on which it wants to select. Code this solution. Call our my_shm function (Figure A.46) to allocate the anonymous shared memory object. Use our msgcreate and msgsnd programs from Section 6.6 to create the message queue, and then place records onto the queue. The parent should just print the size and type of each message that the child reads.
12.4
12.5